Whamcloud - gitweb
LU-17680 ldlm: fix ldlm_res_hop_hash() argument
[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  *
31  * lustre/ldlm/ldlm_resource.c
32  *
33  * Author: Phil Schwan <phil@clusterfs.com>
34  * Author: Peter Braam <braam@clusterfs.com>
35  */
36
37 #define DEBUG_SUBSYSTEM S_LDLM
38 #include <lustre_dlm.h>
39 #include <lustre_fid.h>
40 #include <obd_class.h>
41 #include <libcfs/linux/linux-hash.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;
49 int ldlm_cli_namespace_nr;
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. Once all resources go
56  * away, ldlm_poold moves such namespaces to the inactive list
57  */
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 /* For debug dump, amount of granted locks for one resource to avoid DDOS. */
67 static unsigned int ldlm_dump_granted_max = 256;
68
69 static ssize_t ldebugfs_dump_ns_seq_write(struct file *file,
70                                           const char __user *buffer,
71                                           size_t count, loff_t *off)
72 {
73         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
74         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
75         RETURN(count);
76 }
77
78 LDEBUGFS_FOPS_WR_ONLY(ldlm, dump_ns);
79
80 static int ldlm_rw_uint_seq_show(struct seq_file *m, void *v)
81 {
82         seq_printf(m, "%u\n", *(unsigned int *)m->private);
83         return 0;
84 }
85
86 static ssize_t
87 ldlm_rw_uint_seq_write(struct file *file, const char __user *buffer,
88                        size_t count, loff_t *off)
89 {
90         struct seq_file *seq = file->private_data;
91
92         if (!count)
93                 return 0;
94
95         return kstrtouint_from_user(buffer, count, 0,
96                                     (unsigned int *)seq->private);
97 }
98
99 LDEBUGFS_SEQ_FOPS(ldlm_rw_uint);
100
101 #ifdef HAVE_SERVER_SUPPORT
102
103 static int seq_watermark_show(struct seq_file *m, void *data)
104 {
105         seq_printf(m, "%llu\n", *(__u64 *)m->private);
106         return 0;
107 }
108
109 static ssize_t seq_watermark_write(struct file *file,
110                                    const char __user *buffer, size_t count,
111                                    loff_t *off)
112 {
113         struct seq_file *m = file->private_data;
114         u64 value;
115         __u64 watermark;
116         __u64 *data = m->private;
117         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
118         char kernbuf[22] = "";
119         int rc;
120
121         if (count >= sizeof(kernbuf))
122                 return -EINVAL;
123
124         if (copy_from_user(kernbuf, buffer, count))
125                 return -EFAULT;
126         kernbuf[count] = 0;
127
128         rc = sysfs_memparse(kernbuf, count, &value, "MiB");
129         if (rc < 0) {
130                 CERROR("Failed to set %s, rc = %d.\n",
131                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
132                        rc);
133                 return rc;
134         } else if (value != 0 && value < (1 << 20)) {
135                 CERROR("%s should be greater than 1MB.\n",
136                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
137                 return -EINVAL;
138         }
139         watermark = value >> 20;
140
141         if (wm_low) {
142                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
143                         CERROR("lock_reclaim_threshold_mb must be smaller than lock_limit_mb.\n");
144                         return -EINVAL;
145                 }
146
147                 *data = watermark;
148                 if (watermark != 0) {
149                         watermark <<= 20;
150                         do_div(watermark, sizeof(struct ldlm_lock));
151                 }
152                 ldlm_reclaim_threshold = watermark;
153         } else {
154                 if (ldlm_reclaim_threshold_mb != 0 &&
155                     watermark < ldlm_reclaim_threshold_mb) {
156                         CERROR("lock_limit_mb must be greater than "
157                                "lock_reclaim_threshold_mb.\n");
158                         return -EINVAL;
159                 }
160
161                 *data = watermark;
162                 if (watermark != 0) {
163                         watermark <<= 20;
164                         do_div(watermark, sizeof(struct ldlm_lock));
165                 }
166                 ldlm_lock_limit = watermark;
167         }
168
169         return count;
170 }
171
172 static int seq_watermark_open(struct inode *inode, struct file *file)
173 {
174         return single_open(file, seq_watermark_show, inode->i_private);
175 }
176
177 static const struct file_operations ldlm_watermark_fops = {
178         .owner          = THIS_MODULE,
179         .open           = seq_watermark_open,
180         .read           = seq_read,
181         .write          = seq_watermark_write,
182         .llseek         = seq_lseek,
183         .release        = lprocfs_single_release,
184 };
185
186 static int seq_granted_show(struct seq_file *m, void *data)
187 {
188         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
189                    (struct percpu_counter *)m->private));
190         return 0;
191 }
192
193 static int seq_granted_open(struct inode *inode, struct file *file)
194 {
195         return single_open(file, seq_granted_show, inode->i_private);
196 }
197
198 static const struct file_operations ldlm_granted_fops = {
199         .owner  = THIS_MODULE,
200         .open   = seq_granted_open,
201         .read   = seq_read,
202         .llseek = seq_lseek,
203         .release = seq_release,
204 };
205
206 #endif /* HAVE_SERVER_SUPPORT */
207
208 static struct ldebugfs_vars ldlm_debugfs_list[] = {
209         { .name =       "dump_namespaces",
210           .fops =       &ldlm_dump_ns_fops,
211           .proc_mode =  0222 },
212         { .name =       "dump_granted_max",
213           .fops =       &ldlm_rw_uint_fops,
214           .data =       &ldlm_dump_granted_max },
215 #ifdef HAVE_SERVER_SUPPORT
216         { .name =       "lock_reclaim_threshold_mb",
217           .fops =       &ldlm_watermark_fops,
218           .data =       &ldlm_reclaim_threshold_mb },
219         { .name =       "lock_limit_mb",
220           .fops =       &ldlm_watermark_fops,
221           .data =       &ldlm_lock_limit_mb },
222         { .name =       "lock_granted_count",
223           .fops =       &ldlm_granted_fops,
224           .data =       &ldlm_granted_total },
225 #endif
226         { NULL }
227 };
228
229 int ldlm_debugfs_setup(void)
230 {
231         ENTRY;
232         ldlm_debugfs_dir = debugfs_create_dir(OBD_LDLM_DEVICENAME,
233                                              debugfs_lustre_root);
234         ldlm_ns_debugfs_dir = debugfs_create_dir("namespaces",
235                                                  ldlm_debugfs_dir);
236         ldlm_svc_debugfs_dir = debugfs_create_dir("services",
237                                                   ldlm_debugfs_dir);
238
239         ldebugfs_add_vars(ldlm_debugfs_dir, ldlm_debugfs_list, NULL);
240
241         RETURN(0);
242 }
243
244 void ldlm_debugfs_cleanup(void)
245 {
246         debugfs_remove_recursive(ldlm_debugfs_dir);
247
248         ldlm_svc_debugfs_dir = NULL;
249         ldlm_ns_debugfs_dir = NULL;
250         ldlm_debugfs_dir = NULL;
251 }
252
253 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
254                                    char *buf)
255 {
256         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
257                                                  ns_kobj);
258         __u64                   res = 0;
259         int                     i;
260
261         /* result is not strictly consistant */
262         for (i = 0; i < (1 << ns->ns_bucket_bits); i++)
263                 res += atomic_read(&ns->ns_rs_buckets[i].nsb_count);
264         return sprintf(buf, "%lld\n", res);
265 }
266 LUSTRE_RO_ATTR(resource_count);
267
268 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
269                                char *buf)
270 {
271         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
272                                                  ns_kobj);
273         __u64                   locks;
274
275         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
276                                         LPROCFS_FIELDS_FLAGS_SUM);
277         return sprintf(buf, "%lld\n", locks);
278 }
279 LUSTRE_RO_ATTR(lock_count);
280
281 static ssize_t lock_unused_count_show(struct kobject *kobj,
282                                       struct attribute *attr,
283                                       char *buf)
284 {
285         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
286                                                  ns_kobj);
287
288         return sprintf(buf, "%d\n", ns->ns_nr_unused);
289 }
290 LUSTRE_RO_ATTR(lock_unused_count);
291
292 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
293                              char *buf)
294 {
295         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
296                                                  ns_kobj);
297         __u32 *nr = &ns->ns_max_unused;
298
299         if (ns_connect_lru_resize(ns))
300                 nr = &ns->ns_nr_unused;
301         return sprintf(buf, "%u\n", *nr);
302 }
303
304 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
305                               const char *buffer, size_t count)
306 {
307         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
308                                                  ns_kobj);
309         unsigned long tmp;
310         int lru_resize;
311         int err;
312
313         if (strncmp(buffer, "clear", 5) == 0) {
314                 CDEBUG(D_DLMTRACE,
315                        "dropping all unused locks from namespace %s\n",
316                        ldlm_ns_name(ns));
317                 /* Try to cancel all @ns_nr_unused locks. */
318                 ldlm_cancel_lru(ns, INT_MAX, 0, LDLM_LRU_FLAG_CLEANUP);
319                 return count;
320         }
321
322         err = kstrtoul(buffer, 10, &tmp);
323         if (err != 0) {
324                 CERROR("lru_size: invalid value written\n");
325                 return -EINVAL;
326         }
327         lru_resize = (tmp == 0);
328
329         if (ns_connect_lru_resize(ns)) {
330                 if (!lru_resize)
331                         ns->ns_max_unused = (unsigned int)tmp;
332
333                 if (tmp > ns->ns_nr_unused)
334                         tmp = ns->ns_nr_unused;
335                 tmp = ns->ns_nr_unused - tmp;
336
337                 CDEBUG(D_DLMTRACE,
338                        "changing namespace %s unused locks from %u to %u\n",
339                        ldlm_ns_name(ns), ns->ns_nr_unused,
340                        (unsigned int)tmp);
341
342                 if (!lru_resize) {
343                         CDEBUG(D_DLMTRACE,
344                                "disable lru_resize for namespace %s\n",
345                                ldlm_ns_name(ns));
346                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
347                 }
348                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, 0);
349         } else {
350                 CDEBUG(D_DLMTRACE,
351                        "changing namespace %s max_unused from %u to %u\n",
352                        ldlm_ns_name(ns), ns->ns_max_unused,
353                        (unsigned int)tmp);
354
355                 /* Make sure that LRU resize was originally supported before
356                  * turning it on here.
357                  */
358                 if (lru_resize &&
359                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
360                         CDEBUG(D_DLMTRACE,
361                                "enable lru_resize for namespace %s\n",
362                                ldlm_ns_name(ns));
363                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
364                 }
365                 ns->ns_max_unused = (unsigned int)tmp;
366                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
367         }
368
369         return count;
370 }
371 LUSTRE_RW_ATTR(lru_size);
372
373 static ssize_t lru_cancel_batch_show(struct kobject *kobj,
374                                  struct attribute *attr, char *buf)
375 {
376         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
377                                                  ns_kobj);
378
379         return snprintf(buf, sizeof(buf) - 1, "%u\n", ns->ns_cancel_batch);
380 }
381
382 static ssize_t lru_cancel_batch_store(struct kobject *kobj,
383                                   struct attribute *attr,
384                                   const char *buffer, size_t count)
385 {
386         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
387                                                  ns_kobj);
388         unsigned long tmp;
389
390         if (kstrtoul(buffer, 10, &tmp))
391                 return -EINVAL;
392
393         ns->ns_cancel_batch = (unsigned int)tmp;
394
395         return count;
396 }
397 LUSTRE_RW_ATTR(lru_cancel_batch);
398
399 static ssize_t ns_recalc_pct_show(struct kobject *kobj,
400                                   struct attribute *attr, char *buf)
401 {
402         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
403                                                  ns_kobj);
404
405         return snprintf(buf, sizeof(buf) - 1, "%u\n", ns->ns_recalc_pct);
406 }
407
408 static ssize_t ns_recalc_pct_store(struct kobject *kobj,
409                                    struct attribute *attr,
410                                    const char *buffer, size_t count)
411 {
412         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
413                                                  ns_kobj);
414         unsigned long tmp;
415
416         if (kstrtoul(buffer, 10, &tmp))
417                 return -EINVAL;
418
419         if (tmp > 100)
420                 return -ERANGE;
421
422         ns->ns_recalc_pct = (unsigned int)tmp;
423
424         return count;
425 }
426 LUSTRE_RW_ATTR(ns_recalc_pct);
427
428 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
429                                 char *buf)
430 {
431         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
432                                                  ns_kobj);
433
434         return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
435 }
436
437 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
438                                  const char *buffer, size_t count)
439 {
440         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
441                                                  ns_kobj);
442         int scale = NSEC_PER_MSEC;
443         unsigned long long tmp;
444         char *buf;
445
446         /* Did the user ask in seconds or milliseconds. Default is in ms */
447         buf = strstr(buffer, "ms");
448         if (!buf) {
449                 buf = strchr(buffer, 's');
450                 if (buf)
451                         scale = NSEC_PER_SEC;
452         }
453
454         if (buf)
455                 *buf = '\0';
456
457         if (kstrtoull(buffer, 10, &tmp))
458                 return -EINVAL;
459
460         ns->ns_max_age = ktime_set(0, tmp * scale);
461
462         return count;
463 }
464 LUSTRE_RW_ATTR(lru_max_age);
465
466 static ssize_t early_lock_cancel_show(struct kobject *kobj,
467                                       struct attribute *attr,
468                                       char *buf)
469 {
470         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
471                                                  ns_kobj);
472
473         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
474 }
475
476 static ssize_t early_lock_cancel_store(struct kobject *kobj,
477                                        struct attribute *attr,
478                                        const char *buffer,
479                                        size_t count)
480 {
481         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
482                                                  ns_kobj);
483         unsigned long supp = -1;
484         int rc;
485
486         rc = kstrtoul(buffer, 10, &supp);
487         if (rc < 0)
488                 return rc;
489
490         if (supp == 0)
491                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
492         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
493                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
494         return count;
495 }
496 LUSTRE_RW_ATTR(early_lock_cancel);
497
498 static ssize_t dirty_age_limit_show(struct kobject *kobj,
499                                     struct attribute *attr, char *buf)
500 {
501         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
502                                                  ns_kobj);
503
504         return scnprintf(buf, PAGE_SIZE, "%llu\n",
505                          ktime_divns(ns->ns_dirty_age_limit, NSEC_PER_SEC));
506 }
507
508 static ssize_t dirty_age_limit_store(struct kobject *kobj,
509                                      struct attribute *attr,
510                                      const char *buffer, size_t count)
511 {
512         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
513                                                  ns_kobj);
514         unsigned long long tmp;
515
516         if (kstrtoull(buffer, 10, &tmp))
517                 return -EINVAL;
518
519         ns->ns_dirty_age_limit = ktime_set(tmp, 0);
520
521         return count;
522 }
523 LUSTRE_RW_ATTR(dirty_age_limit);
524
525 #ifdef HAVE_SERVER_SUPPORT
526 static ssize_t ctime_age_limit_show(struct kobject *kobj,
527                                     struct attribute *attr, char *buf)
528 {
529         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
530                                                  ns_kobj);
531
532         return scnprintf(buf, PAGE_SIZE, "%u\n", ns->ns_ctime_age_limit);
533 }
534
535 static ssize_t ctime_age_limit_store(struct kobject *kobj,
536                                      struct attribute *attr,
537                                      const char *buffer, size_t count)
538 {
539         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
540                                                  ns_kobj);
541         unsigned long tmp;
542
543         if (kstrtoul(buffer, 10, &tmp))
544                 return -EINVAL;
545
546         ns->ns_ctime_age_limit = tmp;
547
548         return count;
549 }
550 LUSTRE_RW_ATTR(ctime_age_limit);
551
552 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
553                                   char *buf)
554 {
555         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
556                                                  ns_kobj);
557
558         return sprintf(buf, "%d\n", ns->ns_timeouts);
559 }
560 LUSTRE_RO_ATTR(lock_timeouts);
561
562 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
563                                      struct attribute *attr, char *buf)
564 {
565         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
566                                                  ns_kobj);
567
568         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
569 }
570
571 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
572                                       struct attribute *attr,
573                                       const char *buffer, size_t count)
574 {
575         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
576                                                  ns_kobj);
577         unsigned long tmp;
578         int err;
579
580         err = kstrtoul(buffer, 10, &tmp);
581         if (err != 0)
582                 return -EINVAL;
583
584         ns->ns_max_nolock_size = tmp;
585
586         return count;
587 }
588 LUSTRE_RW_ATTR(max_nolock_bytes);
589
590 static ssize_t contention_seconds_show(struct kobject *kobj,
591                                        struct attribute *attr, char *buf)
592 {
593         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
594                                                  ns_kobj);
595
596         return scnprintf(buf, PAGE_SIZE, "%d\n", ns->ns_contention_time);
597 }
598
599 static ssize_t contention_seconds_store(struct kobject *kobj,
600                                         struct attribute *attr,
601                                         const char *buffer, size_t count)
602 {
603         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
604                                                  ns_kobj);
605         unsigned int tmp;
606
607         if (kstrtouint(buffer, 10, &tmp))
608                 return -EINVAL;
609
610         ns->ns_contention_time = tmp;
611
612         return count;
613 }
614 LUSTRE_RW_ATTR(contention_seconds);
615
616 static ssize_t contended_locks_show(struct kobject *kobj,
617                                     struct attribute *attr, char *buf)
618 {
619         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
620                                                  ns_kobj);
621
622         return sprintf(buf, "%u\n", ns->ns_contended_locks);
623 }
624
625 static ssize_t contended_locks_store(struct kobject *kobj,
626                                      struct attribute *attr,
627                                      const char *buffer, size_t count)
628 {
629         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
630                                                  ns_kobj);
631         unsigned long tmp;
632         int err;
633
634         err = kstrtoul(buffer, 10, &tmp);
635         if (err != 0)
636                 return -EINVAL;
637
638         ns->ns_contended_locks = tmp;
639
640         return count;
641 }
642 LUSTRE_RW_ATTR(contended_locks);
643
644 static ssize_t max_parallel_ast_show(struct kobject *kobj,
645                                      struct attribute *attr, char *buf)
646 {
647         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
648                                                  ns_kobj);
649
650         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
651 }
652
653 static ssize_t max_parallel_ast_store(struct kobject *kobj,
654                                       struct attribute *attr,
655                                       const char *buffer, size_t count)
656 {
657         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
658                                                  ns_kobj);
659         unsigned long tmp;
660         int err;
661
662         err = kstrtoul(buffer, 10, &tmp);
663         if (err != 0)
664                 return -EINVAL;
665
666         ns->ns_max_parallel_ast = tmp;
667
668         return count;
669 }
670 LUSTRE_RW_ATTR(max_parallel_ast);
671
672 #endif /* HAVE_SERVER_SUPPORT */
673
674 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
675 static struct attribute *ldlm_ns_attrs[] = {
676         &lustre_attr_resource_count.attr,
677         &lustre_attr_lock_count.attr,
678         &lustre_attr_lock_unused_count.attr,
679         &lustre_attr_ns_recalc_pct.attr,
680         &lustre_attr_lru_size.attr,
681         &lustre_attr_lru_cancel_batch.attr,
682         &lustre_attr_lru_max_age.attr,
683         &lustre_attr_early_lock_cancel.attr,
684         &lustre_attr_dirty_age_limit.attr,
685 #ifdef HAVE_SERVER_SUPPORT
686         &lustre_attr_ctime_age_limit.attr,
687         &lustre_attr_lock_timeouts.attr,
688         &lustre_attr_max_nolock_bytes.attr,
689         &lustre_attr_contention_seconds.attr,
690         &lustre_attr_contended_locks.attr,
691         &lustre_attr_max_parallel_ast.attr,
692 #endif
693         NULL,
694 };
695
696 static void ldlm_ns_release(struct kobject *kobj)
697 {
698         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
699                                                  ns_kobj);
700         complete(&ns->ns_kobj_unregister);
701 }
702
703 KOBJ_ATTRIBUTE_GROUPS(ldlm_ns);
704
705 static struct kobj_type ldlm_ns_ktype = {
706         .default_groups = KOBJ_ATTR_GROUPS(ldlm_ns),
707         .sysfs_ops      = &lustre_sysfs_ops,
708         .release        = ldlm_ns_release,
709 };
710
711 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
712 {
713         if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
714                 CERROR("dlm namespace %s has no procfs dir?\n",
715                        ldlm_ns_name(ns));
716         else
717                 debugfs_remove_recursive(ns->ns_debugfs_entry);
718
719         if (ns->ns_stats != NULL)
720                 lprocfs_stats_free(&ns->ns_stats);
721 }
722
723 static void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
724 {
725         kobject_put(&ns->ns_kobj);
726         wait_for_completion(&ns->ns_kobj_unregister);
727 }
728
729 static int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
730 {
731         int err;
732
733         ns->ns_kobj.kset = ldlm_ns_kset;
734         init_completion(&ns->ns_kobj_unregister);
735         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
736                                    "%s", ldlm_ns_name(ns));
737
738         ns->ns_stats = lprocfs_stats_alloc(LDLM_NSS_LAST, 0);
739         if (!ns->ns_stats) {
740                 kobject_put(&ns->ns_kobj);
741                 return -ENOMEM;
742         }
743
744         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
745                              LPROCFS_CNTR_AVGMINMAX | LPROCFS_TYPE_LOCKS,
746                              "locks");
747
748         return err;
749 }
750
751 static int ldlm_namespace_debugfs_register(struct ldlm_namespace *ns)
752 {
753         struct dentry *ns_entry;
754
755         if (!IS_ERR_OR_NULL(ns->ns_debugfs_entry)) {
756                 ns_entry = ns->ns_debugfs_entry;
757         } else {
758                 ns_entry = debugfs_create_dir(ldlm_ns_name(ns),
759                                               ldlm_ns_debugfs_dir);
760                 if (!ns_entry)
761                         return -ENOMEM;
762                 ns->ns_debugfs_entry = ns_entry;
763         }
764
765         return 0;
766 }
767 #undef MAX_STRING_SIZE
768
769 static unsigned int ldlm_res_hop_hash(struct cfs_hash *hs,
770                                       const void *key, const unsigned int bits)
771 {
772         const struct ldlm_res_id *id = key;
773         unsigned int val = 0;
774         unsigned int i;
775
776         for (i = 0; i < RES_NAME_SIZE; i++)
777                 val += id->name[i];
778
779         return val & ((1UL << bits) - 1);
780 }
781
782 static unsigned int ldlm_res_hop_fid_hash(const struct ldlm_res_id *id,
783                                           const unsigned int bits)
784 {
785         struct lu_fid       fid;
786         __u32               hash;
787         __u32               val;
788
789         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
790         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
791         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
792
793         hash = fid_flatten32(&fid);
794         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
795
796         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0)
797                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
798         else
799                 val = fid_oid(&fid);
800
801         hash += (val >> 5) + (val << 11);
802         return cfs_hash_32(hash, bits);
803 }
804
805 static void *ldlm_res_hop_key(struct hlist_node *hnode)
806 {
807         struct ldlm_resource   *res;
808
809         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
810         return &res->lr_name;
811 }
812
813 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
814 {
815         struct ldlm_resource   *res;
816
817         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
818         return ldlm_res_eq((const struct ldlm_res_id *)key,
819                            (const struct ldlm_res_id *)&res->lr_name);
820 }
821
822 static void *ldlm_res_hop_object(struct hlist_node *hnode)
823 {
824         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
825 }
826
827 static void
828 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
829 {
830         struct ldlm_resource *res;
831
832         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
833         ldlm_resource_getref(res);
834 }
835
836 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
837 {
838         struct ldlm_resource *res;
839
840         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
841         ldlm_resource_putref(res);
842 }
843
844 static struct cfs_hash_ops ldlm_ns_hash_ops = {
845         .hs_hash        = ldlm_res_hop_hash,
846         .hs_key         = ldlm_res_hop_key,
847         .hs_keycmp      = ldlm_res_hop_keycmp,
848         .hs_keycpy      = NULL,
849         .hs_object      = ldlm_res_hop_object,
850         .hs_get         = ldlm_res_hop_get_locked,
851         .hs_put         = ldlm_res_hop_put
852 };
853
854 static struct {
855         /** hash bucket bits */
856         unsigned int            nsd_bkt_bits;
857         /** hash bits */
858         unsigned int            nsd_all_bits;
859 } ldlm_ns_hash_defs[] = {
860         [LDLM_NS_TYPE_MDC] = {
861                 .nsd_bkt_bits   = 11,
862                 .nsd_all_bits   = 16,
863         },
864         [LDLM_NS_TYPE_MDT] = {
865                 .nsd_bkt_bits   = 14,
866                 .nsd_all_bits   = 21,
867         },
868         [LDLM_NS_TYPE_OSC] = {
869                 .nsd_bkt_bits   = 8,
870                 .nsd_all_bits   = 12,
871         },
872         [LDLM_NS_TYPE_OST] = {
873                 .nsd_bkt_bits   = 11,
874                 .nsd_all_bits   = 17,
875         },
876         [LDLM_NS_TYPE_MGC] = {
877                 .nsd_bkt_bits   = 3,
878                 .nsd_all_bits   = 4,
879         },
880         [LDLM_NS_TYPE_MGT] = {
881                 .nsd_bkt_bits   = 3,
882                 .nsd_all_bits   = 4,
883         },
884 };
885
886 /**
887  * Create and initialize new empty namespace.
888  */
889 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
890                                           enum ldlm_side client,
891                                           enum ldlm_appetite apt,
892                                           enum ldlm_ns_type ns_type)
893 {
894         struct ldlm_namespace *ns = NULL;
895         int idx;
896         int rc;
897
898         ENTRY;
899         LASSERT(obd != NULL);
900
901         rc = ldlm_get_ref();
902         if (rc) {
903                 CERROR("%s: ldlm_get_ref failed: rc = %d\n", name, rc);
904                 RETURN(ERR_PTR(rc));
905         }
906
907         if (ns_type >= ARRAY_SIZE(ldlm_ns_hash_defs) ||
908             ldlm_ns_hash_defs[ns_type].nsd_bkt_bits == 0) {
909                 rc = -EINVAL;
910                 CERROR("%s: unknown namespace type %d: rc = %d\n",
911                        name, ns_type, rc);
912                 GOTO(out_ref, rc);
913         }
914
915         OBD_ALLOC_PTR(ns);
916         if (!ns)
917                 GOTO(out_ref, rc = -ENOMEM);
918
919         ns->ns_rs_hash = cfs_hash_create(name,
920                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
921                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
922                                          ldlm_ns_hash_defs[ns_type].nsd_bkt_bits,
923                                          0,
924                                          CFS_HASH_MIN_THETA,
925                                          CFS_HASH_MAX_THETA,
926                                          &ldlm_ns_hash_ops,
927                                          CFS_HASH_DEPTH |
928                                          CFS_HASH_BIGNAME |
929                                          CFS_HASH_SPIN_BKTLOCK |
930                                          CFS_HASH_NO_ITEMREF);
931         if (!ns->ns_rs_hash)
932                 GOTO(out_ns, rc = -ENOMEM);
933
934         ns->ns_bucket_bits = ldlm_ns_hash_defs[ns_type].nsd_all_bits -
935                              ldlm_ns_hash_defs[ns_type].nsd_bkt_bits;
936
937         OBD_ALLOC_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
938         if (!ns->ns_rs_buckets)
939                 GOTO(out_hash, rc = -ENOMEM);
940
941         for (idx = 0; idx < (1 << ns->ns_bucket_bits); idx++) {
942                 struct ldlm_ns_bucket *nsb = &ns->ns_rs_buckets[idx];
943
944                 at_init(&nsb->nsb_at_estimate, obd_get_ldlm_enqueue_min(obd), 0);
945                 nsb->nsb_namespace = ns;
946                 nsb->nsb_reclaim_start = 0;
947                 atomic_set(&nsb->nsb_count, 0);
948         }
949
950         ns->ns_obd = obd;
951         ns->ns_appetite = apt;
952         ns->ns_client = client;
953         ns->ns_name = kstrdup(name, GFP_KERNEL);
954         if (!ns->ns_name)
955                 GOTO(out_hash, rc = -ENOMEM);
956
957         INIT_LIST_HEAD(&ns->ns_list_chain);
958         INIT_LIST_HEAD(&ns->ns_unused_list);
959         spin_lock_init(&ns->ns_lock);
960         atomic_set(&ns->ns_bref, 0);
961         init_waitqueue_head(&ns->ns_waitq);
962
963         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
964         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
965         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
966
967         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
968         ns->ns_nr_unused          = 0;
969         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
970         ns->ns_cancel_batch       = LDLM_DEFAULT_LRU_SHRINK_BATCH;
971         ns->ns_recalc_pct         = LDLM_DEFAULT_SLV_RECALC_PCT;
972         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
973         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
974         ns->ns_dirty_age_limit    = ktime_set(LDLM_DIRTY_AGE_LIMIT, 0);
975         ns->ns_timeouts           = 0;
976         ns->ns_orig_connect_flags = 0;
977         ns->ns_connect_flags      = 0;
978         ns->ns_stopping           = 0;
979         ns->ns_reclaim_start      = 0;
980         ns->ns_last_pos           = &ns->ns_unused_list;
981         ns->ns_flags              = 0;
982
983         rc = ldlm_namespace_sysfs_register(ns);
984         if (rc) {
985                 CERROR("%s: cannot initialize ns sysfs: rc = %d\n", name, rc);
986                 GOTO(out_hash, rc);
987         }
988
989         rc = ldlm_namespace_debugfs_register(ns);
990         if (rc) {
991                 CERROR("%s: cannot initialize ns proc: rc = %d\n", name, rc);
992                 GOTO(out_sysfs, rc);
993         }
994
995         idx = ldlm_namespace_nr_read(client);
996         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
997         if (rc) {
998                 CERROR("%s: cannot initialize lock pool, rc = %d\n", name, rc);
999                 GOTO(out_proc, rc);
1000         }
1001
1002         ldlm_namespace_register(ns, client);
1003         RETURN(ns);
1004 out_proc:
1005         ldlm_namespace_debugfs_unregister(ns);
1006 out_sysfs:
1007         ldlm_namespace_sysfs_unregister(ns);
1008         ldlm_namespace_cleanup(ns, 0);
1009 out_hash:
1010         OBD_FREE_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
1011         kfree(ns->ns_name);
1012         cfs_hash_putref(ns->ns_rs_hash);
1013 out_ns:
1014         OBD_FREE_PTR(ns);
1015 out_ref:
1016         ldlm_put_ref();
1017         RETURN(ERR_PTR(rc));
1018 }
1019 EXPORT_SYMBOL(ldlm_namespace_new);
1020
1021 /**
1022  * Cancel and destroy all locks on a resource.
1023  *
1024  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
1025  * clean up.  This is currently only used for recovery, and we make
1026  * certain assumptions as a result--notably, that we shouldn't cancel
1027  * locks with refs.
1028  */
1029 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
1030                              u64 flags)
1031 {
1032         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
1033         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
1034
1035         do {
1036                 struct ldlm_lock *lock = NULL, *tmp;
1037
1038                 /* First, we look for non-cleaned-yet lock. all cleaned locks
1039                  * are marked by CLEANED flag.
1040                  */
1041                 lock_res(res);
1042                 list_for_each_entry(tmp, q, l_res_link) {
1043                         if (ldlm_is_cleaned(tmp))
1044                                 continue;
1045
1046                         lock = tmp;
1047                         LDLM_LOCK_GET(lock);
1048                         ldlm_set_cleaned(lock);
1049                         break;
1050                 }
1051
1052                 if (lock == NULL) {
1053                         unlock_res(res);
1054                         break;
1055                 }
1056
1057                 /* Set CBPENDING so nothing is in the cancellation path
1058                  * can match this lock.
1059                  */
1060                 ldlm_set_cbpending(lock);
1061                 ldlm_set_failed(lock);
1062                 ldlm_clear_converting(lock);
1063                 lock->l_flags |= flags;
1064
1065                 /* ... without sending a CANCEL message for local_only. */
1066                 if (local_only)
1067                         ldlm_set_local_only(lock);
1068
1069                 if (local_only && (lock->l_readers || lock->l_writers)) {
1070                         /*
1071                          * This is a little bit gross, but much better than the
1072                          * alternative: pretend that we got a blocking AST from
1073                          * the server, so that when the lock is decref'd, it
1074                          * will go away ...
1075                          */
1076                         unlock_res(res);
1077                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1078                         if (lock->l_flags & LDLM_FL_FAIL_LOC)
1079                                 schedule_timeout_uninterruptible(
1080                                         cfs_time_seconds(4));
1081
1082                         if (lock->l_completion_ast)
1083                                 lock->l_completion_ast(lock,
1084                                                        LDLM_FL_FAILED, NULL);
1085                         LDLM_LOCK_RELEASE(lock);
1086                         continue;
1087                 }
1088
1089                 if (client) {
1090                         struct lustre_handle lockh;
1091
1092                         unlock_res(res);
1093                         ldlm_lock2handle(lock, &lockh);
1094                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1095                         if (rc)
1096                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1097                 } else {
1098                         unlock_res(res);
1099                         LDLM_DEBUG(lock,
1100                                    "Freeing a lock still held by a client node");
1101                         ldlm_lock_cancel(lock);
1102                 }
1103                 LDLM_LOCK_RELEASE(lock);
1104         } while (1);
1105 }
1106
1107 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1108                                struct hlist_node *hnode, void *arg)
1109 {
1110         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1111         __u64 flags = *(__u64 *)arg;
1112
1113         cleanup_resource(res, &res->lr_granted, flags);
1114         cleanup_resource(res, &res->lr_waiting, flags);
1115
1116         return 0;
1117 }
1118
1119 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1120                                   struct hlist_node *hnode, void *arg)
1121 {
1122         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1123
1124         lock_res(res);
1125         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1126                "(%d) after lock cleanup; forcing cleanup.\n",
1127                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1128                refcount_read(&res->lr_refcount) - 1);
1129
1130         /* Use D_NETERROR since it is in the default mask */
1131         ldlm_resource_dump(D_NETERROR, res);
1132         unlock_res(res);
1133         return 0;
1134 }
1135
1136 /**
1137  * Cancel and destroy all locks in the namespace.
1138  *
1139  * Typically used during evictions when server notified client that it was
1140  * evicted and all of its state needs to be destroyed.
1141  * Also used during shutdown.
1142  */
1143 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1144 {
1145         if (ns == NULL) {
1146                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1147                 return ELDLM_OK;
1148         }
1149
1150         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1151                                  &flags, 0);
1152         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1153                                  NULL, 0);
1154         return ELDLM_OK;
1155 }
1156 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1157
1158 /**
1159  * Attempts to free namespace.
1160  *
1161  * Only used when namespace goes away, like during an unmount.
1162  */
1163 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1164 {
1165         ENTRY;
1166
1167         /* At shutdown time, don't call the cancellation callback */
1168         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1169
1170         if (atomic_read(&ns->ns_bref) > 0) {
1171                 int rc;
1172
1173                 CDEBUG(D_DLMTRACE,
1174                        "dlm namespace %s free waiting on refcount %d\n",
1175                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1176 force_wait:
1177                 if (force)
1178                         rc = wait_event_idle_timeout(
1179                                 ns->ns_waitq,
1180                                 atomic_read(&ns->ns_bref) == 0,
1181                                 cfs_time_seconds(1) / 4);
1182                 else
1183                         rc = l_wait_event_abortable(
1184                                 ns->ns_waitq, atomic_read(&ns->ns_bref) == 0);
1185
1186                 /* Forced cleanups should be able to reclaim all references,
1187                  * so it's safe to wait forever... we can't leak locks...
1188                  */
1189                 if (force && rc == 0) {
1190                         rc = -ETIMEDOUT;
1191                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1192                                        "namespace with %d resources in use, "
1193                                        "(rc=%d)\n", ldlm_ns_name(ns),
1194                                        atomic_read(&ns->ns_bref), rc);
1195                         GOTO(force_wait, rc);
1196                 }
1197
1198                 if (atomic_read(&ns->ns_bref)) {
1199                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1200                                        "with %d resources in use, (rc=%d)\n",
1201                                        ldlm_ns_name(ns),
1202                                        atomic_read(&ns->ns_bref), rc);
1203                         RETURN(ELDLM_NAMESPACE_EXISTS);
1204                 }
1205                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1206                        ldlm_ns_name(ns));
1207         }
1208
1209         RETURN(ELDLM_OK);
1210 }
1211
1212 /**
1213  * Performs various cleanups for passed \a ns to make it drop refc and be
1214  * ready for freeing. Waits for refc == 0.
1215  *
1216  * The following is done:
1217  * (0) Unregister \a ns from its list to make inaccessible for potential
1218  * users like pools thread and others;
1219  * (1) Clear all locks in \a ns.
1220  */
1221 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1222                                struct obd_import *imp, int force)
1223 {
1224         int rc;
1225
1226         ENTRY;
1227         if (!ns) {
1228                 EXIT;
1229                 return;
1230         }
1231
1232         spin_lock(&ns->ns_lock);
1233         ns->ns_stopping = 1;
1234         spin_unlock(&ns->ns_lock);
1235
1236         /* Can fail with -EINTR when force == 0 in which case try harder. */
1237         rc = __ldlm_namespace_free(ns, force);
1238         if (rc != ELDLM_OK) {
1239                 if (imp) {
1240                         ptlrpc_disconnect_import(imp, 0);
1241                         ptlrpc_invalidate_import(imp);
1242                 }
1243
1244                 /*
1245                  * With all requests dropped and the import inactive
1246                  * we are gaurenteed all reference will be dropped.
1247                  */
1248                 rc = __ldlm_namespace_free(ns, 1);
1249                 LASSERT(rc == 0);
1250         }
1251         EXIT;
1252 }
1253 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1254
1255 /**
1256  * Performs freeing memory structures related to \a ns. This is only done
1257  * when ldlm_namespce_free_prior() successfully removed all resources
1258  * referencing \a ns and its refc == 0.
1259  */
1260 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1261 {
1262         ENTRY;
1263         if (!ns) {
1264                 EXIT;
1265                 return;
1266         }
1267
1268         /* Make sure that nobody can find this ns in its list. */
1269         ldlm_namespace_unregister(ns, ns->ns_client);
1270         /* Fini pool _before_ parent proc dir is removed. This is important as
1271          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1272          * Removing it after @dir may cause oops.
1273          */
1274         ldlm_pool_fini(&ns->ns_pool);
1275
1276         ldlm_namespace_debugfs_unregister(ns);
1277         ldlm_namespace_sysfs_unregister(ns);
1278         cfs_hash_putref(ns->ns_rs_hash);
1279         OBD_FREE_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
1280         kfree(ns->ns_name);
1281         /* Namespace \a ns should be not on list at this time, otherwise
1282          * this will cause issues related to using freed \a ns in poold
1283          * thread.
1284          */
1285         LASSERT(list_empty(&ns->ns_list_chain));
1286         OBD_FREE_PTR(ns);
1287         ldlm_put_ref();
1288         EXIT;
1289 }
1290 EXPORT_SYMBOL(ldlm_namespace_free_post);
1291
1292 /**
1293  * Cleanup the resource, and free namespace.
1294  * bug 12864:
1295  * Deadlock issue:
1296  * proc1: destroy import
1297  *        class_disconnect_export(grab cl_sem) ->
1298  *              -> ldlm_namespace_free ->
1299  *              -> lprocfs_remove(grab _lprocfs_lock).
1300  * proc2: read proc info
1301  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1302  *              -> osc_rd_active, etc(grab cl_sem).
1303  *
1304  * So that I have to split the ldlm_namespace_free into two parts - the first
1305  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1306  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1307  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1308  * held.
1309  */
1310 void ldlm_namespace_free(struct ldlm_namespace *ns,
1311                          struct obd_import *imp,
1312                          int force)
1313 {
1314         ldlm_namespace_free_prior(ns, imp, force);
1315         ldlm_namespace_free_post(ns);
1316 }
1317 EXPORT_SYMBOL(ldlm_namespace_free);
1318
1319 void ldlm_namespace_get(struct ldlm_namespace *ns)
1320 {
1321         atomic_inc(&ns->ns_bref);
1322 }
1323
1324 /* This is only for callers that care about refcount */
1325 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1326 {
1327         return atomic_inc_return(&ns->ns_bref);
1328 }
1329
1330 void ldlm_namespace_put(struct ldlm_namespace *ns)
1331 {
1332         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1333                 wake_up(&ns->ns_waitq);
1334                 spin_unlock(&ns->ns_lock);
1335         }
1336 }
1337
1338 /** Register \a ns in the list of namespaces */
1339 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1340 {
1341         mutex_lock(ldlm_namespace_lock(client));
1342         LASSERT(list_empty(&ns->ns_list_chain));
1343         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1344         ldlm_namespace_nr_inc(client);
1345         mutex_unlock(ldlm_namespace_lock(client));
1346 }
1347
1348 /** Unregister \a ns from the list of namespaces. */
1349 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1350 {
1351         mutex_lock(ldlm_namespace_lock(client));
1352         LASSERT(!list_empty(&ns->ns_list_chain));
1353         /* Some asserts and possibly other parts of the code are still
1354          * using list_empty(&ns->ns_list_chain). This is why it is
1355          * important to use list_del_init() here.
1356          */
1357         list_del_init(&ns->ns_list_chain);
1358         ldlm_namespace_nr_dec(client);
1359         mutex_unlock(ldlm_namespace_lock(client));
1360 }
1361
1362 /** Should be called with ldlm_namespace_lock(client) taken. */
1363 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1364                                           enum ldlm_side client)
1365 {
1366         LASSERT(!list_empty(&ns->ns_list_chain));
1367         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1368         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1369 }
1370
1371 /** Should be called with ldlm_namespace_lock(client) taken. */
1372 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1373                                             enum ldlm_side client)
1374 {
1375         LASSERT(!list_empty(&ns->ns_list_chain));
1376         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1377         list_move_tail(&ns->ns_list_chain,
1378                        ldlm_namespace_inactive_list(client));
1379 }
1380
1381 /** Should be called with ldlm_namespace_lock(client) taken. */
1382 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1383 {
1384         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1385         LASSERT(!list_empty(ldlm_namespace_list(client)));
1386         return container_of(ldlm_namespace_list(client)->next,
1387                             struct ldlm_namespace, ns_list_chain);
1388 }
1389
1390 static bool ldlm_resource_extent_new(struct ldlm_resource *res)
1391 {
1392         int idx;
1393
1394         OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1395                        sizeof(*res->lr_itree) * LCK_MODE_NUM);
1396         if (res->lr_itree == NULL)
1397                 return false;
1398         /* Initialize interval trees for each lock mode. */
1399         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1400                 res->lr_itree[idx].lit_size = 0;
1401                 res->lr_itree[idx].lit_mode = BIT(idx);
1402                 res->lr_itree[idx].lit_root = NULL;
1403         }
1404         return true;
1405 }
1406
1407 static bool ldlm_resource_inodebits_new(struct ldlm_resource *res)
1408 {
1409         int i;
1410
1411         OBD_ALLOC_PTR(res->lr_ibits_queues);
1412         if (res->lr_ibits_queues == NULL)
1413                 return false;
1414         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
1415                 INIT_LIST_HEAD(&res->lr_ibits_queues->liq_waiting[i]);
1416         return true;
1417 }
1418
1419 static bool ldlm_resource_flock_new(struct ldlm_resource *res)
1420 {
1421         res->lr_flock_node.lfn_needs_reprocess = false;
1422         res->lr_flock_node.lfn_root = NULL;
1423         atomic_set(&res->lr_flock_node.lfn_unlock_pending, 0);
1424
1425         return true;
1426 }
1427
1428 /** Create and initialize new resource. */
1429 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1430 {
1431         struct ldlm_resource *res;
1432         bool rc;
1433
1434         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1435         if (res == NULL)
1436                 return NULL;
1437
1438         switch (ldlm_type) {
1439         case LDLM_EXTENT:
1440                 rc = ldlm_resource_extent_new(res);
1441                 break;
1442         case LDLM_IBITS:
1443                 rc = ldlm_resource_inodebits_new(res);
1444                 break;
1445         case LDLM_FLOCK:
1446                 rc = ldlm_resource_flock_new(res);
1447                 break;
1448         default:
1449                 rc = true;
1450                 break;
1451         }
1452         if (!rc) {
1453                 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1454                 return NULL;
1455         }
1456
1457         INIT_LIST_HEAD(&res->lr_granted);
1458         INIT_LIST_HEAD(&res->lr_waiting);
1459
1460         refcount_set(&res->lr_refcount, 1);
1461         spin_lock_init(&res->lr_lock);
1462         lu_ref_init(&res->lr_reference);
1463
1464         /* Since LVB init can be delayed now, there is no longer need to
1465          * immediatelly acquire mutex here.
1466          */
1467         mutex_init(&res->lr_lvb_mutex);
1468         res->lr_lvb_initialized = false;
1469
1470         return res;
1471 }
1472
1473 static void __ldlm_resource_free(struct rcu_head *head)
1474 {
1475         struct ldlm_resource *res = container_of(head, struct ldlm_resource,
1476                                                  lr_rcu);
1477
1478         OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1479 }
1480
1481 static void ldlm_resource_free(struct ldlm_resource *res)
1482 {
1483         if (res->lr_type == LDLM_EXTENT) {
1484                 if (res->lr_itree != NULL)
1485                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1486                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1487         } else if (res->lr_type == LDLM_IBITS) {
1488                 if (res->lr_ibits_queues != NULL)
1489                         OBD_FREE_PTR(res->lr_ibits_queues);
1490         }
1491
1492         call_rcu(&res->lr_rcu, __ldlm_resource_free);
1493 }
1494
1495 /**
1496  * Return a reference to resource with given name, creating it if necessary.
1497  * Args: namespace with ns_lock unlocked
1498  * Locks: takes and releases NS hash-lock and res->lr_lock
1499  * Returns: referenced, unlocked ldlm_resource or ERR_PTR
1500  */
1501 struct ldlm_resource *
1502 ldlm_resource_get(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
1503                   enum ldlm_type type, int create)
1504 {
1505         struct hlist_node       *hnode;
1506         struct ldlm_resource    *res = NULL;
1507         struct cfs_hash_bd              bd;
1508         __u64                   version;
1509         int                     ns_refcount = 0;
1510         int hash;
1511
1512         LASSERT(ns != NULL);
1513         LASSERT(ns->ns_rs_hash != NULL);
1514         LASSERT(name->name[0] != 0);
1515
1516         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1517         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1518         if (hnode != NULL) {
1519                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1520                 GOTO(found, res);
1521         }
1522
1523         version = cfs_hash_bd_version_get(&bd);
1524         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1525
1526         if (create == 0)
1527                 return ERR_PTR(-ENOENT);
1528
1529         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1530                  "type: %d\n", type);
1531         res = ldlm_resource_new(type);
1532         if (res == NULL)
1533                 return ERR_PTR(-ENOMEM);
1534
1535         hash = ldlm_res_hop_fid_hash(name, ns->ns_bucket_bits);
1536         res->lr_ns_bucket = &ns->ns_rs_buckets[hash];
1537         res->lr_name = *name;
1538         res->lr_type = type;
1539
1540         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1541         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1542                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1543
1544         if (hnode != NULL) {
1545                 /* Someone won the race and already added the resource. */
1546                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1547                 /* Clean lu_ref for failed resource. */
1548                 lu_ref_fini(&res->lr_reference);
1549                 ldlm_resource_free(res);
1550 found:
1551                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1552                 return res;
1553         }
1554         /* We won! Let's add the resource. */
1555         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1556         if (atomic_inc_return(&res->lr_ns_bucket->nsb_count) == 1)
1557                 ns_refcount = ldlm_namespace_get_return(ns);
1558
1559         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1560
1561         CFS_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1562
1563         /* Let's see if we happened to be the very first resource in this
1564          * namespace. If so, and this is a client namespace, we need to move
1565          * the namespace into the active namespaces list to be patrolled by
1566          * the ldlm_poold.
1567          */
1568         if (ns_is_client(ns) && ns_refcount == 1) {
1569                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1570                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1571                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1572         }
1573
1574         return res;
1575 }
1576 EXPORT_SYMBOL(ldlm_resource_get);
1577
1578 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1579 {
1580         LASSERT(res != NULL);
1581         LASSERT(res != LP_POISON);
1582         refcount_inc(&res->lr_refcount);
1583         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1584                refcount_read(&res->lr_refcount));
1585         return res;
1586 }
1587
1588 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1589                                          struct ldlm_resource *res)
1590 {
1591         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1592
1593         if (!list_empty(&res->lr_granted)) {
1594                 ldlm_resource_dump(D_ERROR, res);
1595                 LBUG();
1596         }
1597
1598         if (!list_empty(&res->lr_waiting)) {
1599                 ldlm_resource_dump(D_ERROR, res);
1600                 LBUG();
1601         }
1602
1603         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1604                                bd, &res->lr_hash);
1605         lu_ref_fini(&res->lr_reference);
1606         if (atomic_dec_and_test(&nsb->nsb_count))
1607                 ldlm_namespace_put(nsb->nsb_namespace);
1608 }
1609
1610 /* Returns 1 if the resource was freed, 0 if it remains. */
1611 int ldlm_resource_putref(struct ldlm_resource *res)
1612 {
1613         struct ldlm_namespace *ns;
1614         struct cfs_hash_bd bd;
1615         int refcount;
1616
1617         if (refcount_dec_not_one(&res->lr_refcount))
1618                 return 0;
1619         ns = ldlm_res_to_ns(res);
1620         refcount = refcount_read(&res->lr_refcount);
1621         LASSERT(refcount < LI_POISON);
1622
1623         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1624                res, refcount_read(&res->lr_refcount) - 1);
1625
1626         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1627         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1628                 __ldlm_resource_putref_final(&bd, res);
1629                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1630                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1631                         ns->ns_lvbo->lvbo_free(res);
1632                 ldlm_resource_free(res);
1633                 return 1;
1634         }
1635         return 0;
1636 }
1637 EXPORT_SYMBOL(ldlm_resource_putref);
1638
1639 static void __ldlm_resource_add_lock(struct ldlm_resource *res,
1640                                      struct list_head *head,
1641                                      struct ldlm_lock *lock,
1642                                      bool tail)
1643 {
1644         check_res_locked(res);
1645
1646         if (ldlm_is_destroyed(lock)) {
1647                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1648                 return;
1649         }
1650
1651         LASSERT(list_empty(&lock->l_res_link));
1652
1653         if (tail)
1654                 list_add_tail(&lock->l_res_link, head);
1655         else
1656                 list_add(&lock->l_res_link, head);
1657
1658         if (res->lr_type == LDLM_IBITS)
1659                 ldlm_inodebits_add_lock(res, head, lock, tail);
1660
1661         ldlm_resource_dump(D_INFO, res);
1662 }
1663
1664 /* Add a lock into a given resource into specified lock list. */
1665 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1666                             struct ldlm_lock *lock)
1667 {
1668         LDLM_DEBUG(lock, "About to add this lock");
1669
1670         __ldlm_resource_add_lock(res, head, lock, true);
1671 }
1672
1673 /* Insert a lock into resource after specified lock. */
1674 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1675                                      struct ldlm_lock *new)
1676 {
1677         LASSERT(!list_empty(&original->l_res_link));
1678
1679         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1680         __ldlm_resource_add_lock(original->l_resource,
1681                                  &original->l_res_link,
1682                                  new, false);
1683 }
1684
1685 /**
1686  * Insert a lock into resource before the specified lock.
1687  *
1688  * IBITS waiting locks are to be inserted to the ibit lists as well, and only
1689  * the insert-after operation is supported for them, because the set of bits
1690  * of the previous and the new locks must match. Therefore, get the previous
1691  * lock and insert after.
1692  */
1693 void ldlm_resource_insert_lock_before(struct ldlm_lock *original,
1694                                       struct ldlm_lock *new)
1695 {
1696         LASSERT(!list_empty(&original->l_res_link));
1697
1698         LDLM_DEBUG(new, "About to insert this lock before %p: ", original);
1699         __ldlm_resource_add_lock(original->l_resource,
1700                                  original->l_res_link.prev, new, false);
1701 }
1702
1703 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1704 {
1705         int type = lock->l_resource->lr_type;
1706
1707         check_res_locked(lock->l_resource);
1708         switch (type) {
1709         case LDLM_PLAIN:
1710                 ldlm_unlink_lock_skiplist(lock);
1711                 break;
1712         case LDLM_EXTENT:
1713                 ldlm_extent_unlink_lock(lock);
1714                 break;
1715         case LDLM_IBITS:
1716                 ldlm_inodebits_unlink_lock(lock);
1717                 break;
1718         case LDLM_FLOCK:
1719                 ldlm_flock_unlink_lock(lock);
1720                 break;
1721         }
1722         list_del_init(&lock->l_res_link);
1723 }
1724 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1725
1726 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1727 {
1728         desc->lr_type = res->lr_type;
1729         desc->lr_name = res->lr_name;
1730 }
1731
1732 /* Print info about all locks in all namespaces on this node to debug log. */
1733 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1734 {
1735         struct ldlm_namespace *ns;
1736
1737         if (!((libcfs_debug | D_ERROR) & level))
1738                 return;
1739
1740         mutex_lock(ldlm_namespace_lock(client));
1741
1742         list_for_each_entry(ns, ldlm_namespace_list(client), ns_list_chain)
1743                 ldlm_namespace_dump(level, ns);
1744
1745         mutex_unlock(ldlm_namespace_lock(client));
1746 }
1747
1748 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1749                               struct hlist_node *hnode, void *arg)
1750 {
1751         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1752         int    level = (int)(unsigned long)arg;
1753
1754         lock_res(res);
1755         ldlm_resource_dump(level, res);
1756         unlock_res(res);
1757
1758         return 0;
1759 }
1760
1761 /* Print info about all locks in this namespace on this node to debug log. */
1762 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1763 {
1764         if (!((libcfs_debug | D_ERROR) & level))
1765                 return;
1766
1767         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1768                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1769                ns_is_client(ns) ? "client" : "server");
1770
1771         if (ktime_get_seconds() < ns->ns_next_dump)
1772                 return;
1773
1774         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1775                                  ldlm_res_hash_dump,
1776                                  (void *)(unsigned long)level, 0);
1777         spin_lock(&ns->ns_lock);
1778         ns->ns_next_dump = ktime_get_seconds() + 10;
1779         spin_unlock(&ns->ns_lock);
1780 }
1781
1782 /* Print information about all locks in this resource to debug log. */
1783 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1784 {
1785         struct ldlm_lock *lock;
1786         unsigned int granted = 0;
1787
1788         BUILD_BUG_ON(RES_NAME_SIZE != 4);
1789
1790         if (!((libcfs_debug | D_ERROR) & level))
1791                 return;
1792
1793         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1794                PLDLMRES(res), res, refcount_read(&res->lr_refcount));
1795
1796         if (!list_empty(&res->lr_granted)) {
1797                 CDEBUG(level, "Granted locks (in reverse order):\n");
1798                 list_for_each_entry_reverse(lock, &res->lr_granted,
1799                                                 l_res_link) {
1800                         LDLM_DEBUG_LIMIT(level, lock, "###");
1801                         if (!(level & D_CANTMASK) &&
1802                             ++granted > ldlm_dump_granted_max) {
1803                                 CDEBUG(level,
1804                                        "only dump %d granted locks to avoid DDOS.\n",
1805                                        granted);
1806                                 break;
1807                         }
1808                 }
1809         }
1810
1811         if (!list_empty(&res->lr_waiting)) {
1812                 CDEBUG(level, "Waiting locks:\n");
1813                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1814                         LDLM_DEBUG_LIMIT(level, lock, "###");
1815         }
1816 }
1817 EXPORT_SYMBOL(ldlm_resource_dump);