Whamcloud - gitweb
LU-14291 lustre: only include nrs headers when needed
[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 <libcfs/linux/linux-hash.h>
43 #include "ldlm_internal.h"
44
45 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
46 struct kmem_cache *ldlm_interval_tree_slab;
47 struct kmem_cache *ldlm_inodebits_slab;
48
49 int ldlm_srv_namespace_nr = 0;
50 int ldlm_cli_namespace_nr = 0;
51
52 DEFINE_MUTEX(ldlm_srv_namespace_lock);
53 LIST_HEAD(ldlm_srv_namespace_list);
54
55 DEFINE_MUTEX(ldlm_cli_namespace_lock);
56 /* Client Namespaces that have active resources in them.
57  * Once all resources go away, ldlm_poold moves such namespaces to the
58  * inactive list */
59 LIST_HEAD(ldlm_cli_active_namespace_list);
60 /* Client namespaces that don't have any locks in them */
61 LIST_HEAD(ldlm_cli_inactive_namespace_list);
62
63 static struct dentry *ldlm_debugfs_dir;
64 static struct dentry *ldlm_ns_debugfs_dir;
65 struct dentry *ldlm_svc_debugfs_dir;
66
67 /* during debug dump certain amount of granted locks for one resource to avoid
68  * DDOS. */
69 static unsigned int ldlm_dump_granted_max = 256;
70
71 static ssize_t ldebugfs_dump_ns_seq_write(struct file *file,
72                                           const char __user *buffer,
73                                           size_t count, loff_t *off)
74 {
75         ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
76         ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
77         RETURN(count);
78 }
79
80 LDEBUGFS_FOPS_WR_ONLY(ldlm, dump_ns);
81
82 static int ldlm_rw_uint_seq_show(struct seq_file *m, void *v)
83 {
84         seq_printf(m, "%u\n", *(unsigned int *)m->private);
85         return 0;
86 }
87
88 static ssize_t
89 ldlm_rw_uint_seq_write(struct file *file, const char __user *buffer,
90                        size_t count, loff_t *off)
91 {
92         struct seq_file *seq = file->private_data;
93
94         if (!count)
95                 return 0;
96
97         return kstrtouint_from_user(buffer, count, 0,
98                                     (unsigned int *)seq->private);
99 }
100
101 LDEBUGFS_SEQ_FOPS(ldlm_rw_uint);
102
103 #ifdef HAVE_SERVER_SUPPORT
104
105 static int seq_watermark_show(struct seq_file *m, void *data)
106 {
107         seq_printf(m, "%llu\n", *(__u64 *)m->private);
108         return 0;
109 }
110
111 static ssize_t seq_watermark_write(struct file *file,
112                                    const char __user *buffer, size_t count,
113                                    loff_t *off)
114 {
115         struct seq_file *m = file->private_data;
116         u64 value;
117         __u64 watermark;
118         __u64 *data = m->private;
119         bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
120         char kernbuf[22] = "";
121         int rc;
122
123         if (count >= sizeof(kernbuf))
124                 return -EINVAL;
125
126         if (copy_from_user(kernbuf, buffer, count))
127                 return -EFAULT;
128         kernbuf[count] = 0;
129
130         rc = sysfs_memparse(kernbuf, count, &value, "MiB");
131         if (rc < 0) {
132                 CERROR("Failed to set %s, rc = %d.\n",
133                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
134                        rc);
135                 return rc;
136         } else if (value != 0 && value < (1 << 20)) {
137                 CERROR("%s should be greater than 1MB.\n",
138                        wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
139                 return -EINVAL;
140         }
141         watermark = value >> 20;
142
143         if (wm_low) {
144                 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
145                         CERROR("lock_reclaim_threshold_mb must be smaller than "
146                                "lock_limit_mb.\n");
147                         return -EINVAL;
148                 }
149
150                 *data = watermark;
151                 if (watermark != 0) {
152                         watermark <<= 20;
153                         do_div(watermark, sizeof(struct ldlm_lock));
154                 }
155                 ldlm_reclaim_threshold = watermark;
156         } else {
157                 if (ldlm_reclaim_threshold_mb != 0 &&
158                     watermark < ldlm_reclaim_threshold_mb) {
159                         CERROR("lock_limit_mb must be greater than "
160                                "lock_reclaim_threshold_mb.\n");
161                         return -EINVAL;
162                 }
163
164                 *data = watermark;
165                 if (watermark != 0) {
166                         watermark <<= 20;
167                         do_div(watermark, sizeof(struct ldlm_lock));
168                 }
169                 ldlm_lock_limit = watermark;
170         }
171
172         return count;
173 }
174
175 static int seq_watermark_open(struct inode *inode, struct file *file)
176 {
177         return single_open(file, seq_watermark_show, inode->i_private);
178 }
179
180 static const struct file_operations ldlm_watermark_fops = {
181         .owner          = THIS_MODULE,
182         .open           = seq_watermark_open,
183         .read           = seq_read,
184         .write          = seq_watermark_write,
185         .llseek         = seq_lseek,
186         .release        = lprocfs_single_release,
187 };
188
189 static int seq_granted_show(struct seq_file *m, void *data)
190 {
191         seq_printf(m, "%llu\n", percpu_counter_sum_positive(
192                    (struct percpu_counter *)m->private));
193         return 0;
194 }
195
196 static int seq_granted_open(struct inode *inode, struct file *file)
197 {
198         return single_open(file, seq_granted_show, inode->i_private);
199 }
200
201 static const struct file_operations ldlm_granted_fops = {
202         .owner  = THIS_MODULE,
203         .open   = seq_granted_open,
204         .read   = seq_read,
205         .llseek = seq_lseek,
206         .release = seq_release,
207 };
208
209 #endif /* HAVE_SERVER_SUPPORT */
210
211 static struct ldebugfs_vars ldlm_debugfs_list[] = {
212         { .name =       "dump_namespaces",
213           .fops =       &ldlm_dump_ns_fops,
214           .proc_mode =  0222 },
215         { .name =       "dump_granted_max",
216           .fops =       &ldlm_rw_uint_fops,
217           .data =       &ldlm_dump_granted_max },
218 #ifdef HAVE_SERVER_SUPPORT
219         { .name =       "lock_reclaim_threshold_mb",
220           .fops =       &ldlm_watermark_fops,
221           .data =       &ldlm_reclaim_threshold_mb },
222         { .name =       "lock_limit_mb",
223           .fops =       &ldlm_watermark_fops,
224           .data =       &ldlm_lock_limit_mb },
225         { .name =       "lock_granted_count",
226           .fops =       &ldlm_granted_fops,
227           .data =       &ldlm_granted_total },
228 #endif
229         { NULL }
230 };
231
232 int ldlm_debugfs_setup(void)
233 {
234         ENTRY;
235         ldlm_debugfs_dir = debugfs_create_dir(OBD_LDLM_DEVICENAME,
236                                              debugfs_lustre_root);
237         ldlm_ns_debugfs_dir = debugfs_create_dir("namespaces",
238                                                  ldlm_debugfs_dir);
239         ldlm_svc_debugfs_dir = debugfs_create_dir("services",
240                                                   ldlm_debugfs_dir);
241
242         ldebugfs_add_vars(ldlm_debugfs_dir, ldlm_debugfs_list, NULL);
243
244         RETURN(0);
245 }
246
247 void ldlm_debugfs_cleanup(void)
248 {
249         debugfs_remove_recursive(ldlm_debugfs_dir);
250
251         ldlm_svc_debugfs_dir = NULL;
252         ldlm_ns_debugfs_dir = NULL;
253         ldlm_debugfs_dir = NULL;
254 }
255
256 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
257                                    char *buf)
258 {
259         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
260                                                  ns_kobj);
261         __u64                   res = 0;
262         int                     i;
263
264         /* result is not strictly consistant */
265         for (i = 0; i < (1 << ns->ns_bucket_bits); i++)
266                 res += atomic_read(&ns->ns_rs_buckets[i].nsb_count);
267         return sprintf(buf, "%lld\n", res);
268 }
269 LUSTRE_RO_ATTR(resource_count);
270
271 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
272                                char *buf)
273 {
274         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
275                                                  ns_kobj);
276         __u64                   locks;
277
278         locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
279                                         LPROCFS_FIELDS_FLAGS_SUM);
280         return sprintf(buf, "%lld\n", locks);
281 }
282 LUSTRE_RO_ATTR(lock_count);
283
284 static ssize_t lock_unused_count_show(struct kobject *kobj,
285                                       struct attribute *attr,
286                                       char *buf)
287 {
288         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
289                                                  ns_kobj);
290
291         return sprintf(buf, "%d\n", ns->ns_nr_unused);
292 }
293 LUSTRE_RO_ATTR(lock_unused_count);
294
295 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
296                              char *buf)
297 {
298         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
299                                                  ns_kobj);
300         __u32 *nr = &ns->ns_max_unused;
301
302         if (ns_connect_lru_resize(ns))
303                 nr = &ns->ns_nr_unused;
304         return sprintf(buf, "%u\n", *nr);
305 }
306
307 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
308                               const char *buffer, size_t count)
309 {
310         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
311                                                  ns_kobj);
312         unsigned long tmp;
313         int lru_resize;
314         int err;
315
316         if (strncmp(buffer, "clear", 5) == 0) {
317                 CDEBUG(D_DLMTRACE,
318                        "dropping all unused locks from namespace %s\n",
319                        ldlm_ns_name(ns));
320                 /* Try to cancel all @ns_nr_unused locks. */
321                 ldlm_cancel_lru(ns, INT_MAX, 0, LDLM_LRU_FLAG_CLEANUP);
322                 return count;
323         }
324
325         err = kstrtoul(buffer, 10, &tmp);
326         if (err != 0) {
327                 CERROR("lru_size: invalid value written\n");
328                 return -EINVAL;
329         }
330         lru_resize = (tmp == 0);
331
332         if (ns_connect_lru_resize(ns)) {
333                 if (!lru_resize)
334                         ns->ns_max_unused = (unsigned int)tmp;
335
336                 if (tmp > ns->ns_nr_unused)
337                         tmp = ns->ns_nr_unused;
338                 tmp = ns->ns_nr_unused - tmp;
339
340                 CDEBUG(D_DLMTRACE,
341                        "changing namespace %s unused locks from %u to %u\n",
342                        ldlm_ns_name(ns), ns->ns_nr_unused,
343                        (unsigned int)tmp);
344
345                 if (!lru_resize) {
346                         CDEBUG(D_DLMTRACE,
347                                "disable lru_resize for namespace %s\n",
348                                ldlm_ns_name(ns));
349                         ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
350                 }
351                 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, 0);
352         } else {
353                 CDEBUG(D_DLMTRACE,
354                        "changing namespace %s max_unused from %u to %u\n",
355                        ldlm_ns_name(ns), ns->ns_max_unused,
356                        (unsigned int)tmp);
357
358                 /* Make sure that LRU resize was originally supported before
359                  * turning it on here.
360                  */
361                 if (lru_resize &&
362                     (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
363                         CDEBUG(D_DLMTRACE,
364                                "enable lru_resize for namespace %s\n",
365                                ldlm_ns_name(ns));
366                         ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
367                 }
368                 ns->ns_max_unused = (unsigned int)tmp;
369                 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
370         }
371
372         return count;
373 }
374 LUSTRE_RW_ATTR(lru_size);
375
376 static ssize_t lru_cancel_batch_show(struct kobject *kobj,
377                                  struct attribute *attr, char *buf)
378 {
379         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
380                                                  ns_kobj);
381
382         return snprintf(buf, sizeof(buf) - 1, "%u\n", ns->ns_cancel_batch);
383 }
384
385 static ssize_t lru_cancel_batch_store(struct kobject *kobj,
386                                   struct attribute *attr,
387                                   const char *buffer, size_t count)
388 {
389         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
390                                                  ns_kobj);
391         unsigned long tmp;
392
393         if (kstrtoul(buffer, 10, &tmp))
394                 return -EINVAL;
395
396         ns->ns_cancel_batch = (unsigned int)tmp;
397
398         return count;
399 }
400 LUSTRE_RW_ATTR(lru_cancel_batch);
401
402 static ssize_t ns_recalc_pct_show(struct kobject *kobj,
403                                   struct attribute *attr, char *buf)
404 {
405         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
406                                                  ns_kobj);
407
408         return snprintf(buf, sizeof(buf) - 1, "%u\n", ns->ns_recalc_pct);
409 }
410
411 static ssize_t ns_recalc_pct_store(struct kobject *kobj,
412                                    struct attribute *attr,
413                                    const char *buffer, size_t count)
414 {
415         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
416                                                  ns_kobj);
417         unsigned long tmp;
418
419         if (kstrtoul(buffer, 10, &tmp))
420                 return -EINVAL;
421
422         if (tmp > 100)
423                 return -ERANGE;
424
425         ns->ns_recalc_pct = (unsigned int)tmp;
426
427         return count;
428 }
429 LUSTRE_RW_ATTR(ns_recalc_pct);
430
431 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
432                                 char *buf)
433 {
434         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
435                                                  ns_kobj);
436
437         return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
438 }
439
440 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
441                                  const char *buffer, size_t count)
442 {
443         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
444                                                  ns_kobj);
445         int scale = NSEC_PER_MSEC;
446         unsigned long long tmp;
447         char *buf;
448
449         /* Did the user ask in seconds or milliseconds. Default is in ms */
450         buf = strstr(buffer, "ms");
451         if (!buf) {
452                 buf = strchr(buffer, 's');
453                 if (buf)
454                         scale = NSEC_PER_SEC;
455         }
456
457         if (buf)
458                 *buf = '\0';
459
460         if (kstrtoull(buffer, 10, &tmp))
461                 return -EINVAL;
462
463         ns->ns_max_age = ktime_set(0, tmp * scale);
464
465         return count;
466 }
467 LUSTRE_RW_ATTR(lru_max_age);
468
469 static ssize_t early_lock_cancel_show(struct kobject *kobj,
470                                       struct attribute *attr,
471                                       char *buf)
472 {
473         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
474                                                  ns_kobj);
475
476         return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
477 }
478
479 static ssize_t early_lock_cancel_store(struct kobject *kobj,
480                                        struct attribute *attr,
481                                        const char *buffer,
482                                        size_t count)
483 {
484         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
485                                                  ns_kobj);
486         unsigned long supp = -1;
487         int rc;
488
489         rc = kstrtoul(buffer, 10, &supp);
490         if (rc < 0)
491                 return rc;
492
493         if (supp == 0)
494                 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
495         else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
496                 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
497         return count;
498 }
499 LUSTRE_RW_ATTR(early_lock_cancel);
500
501 static ssize_t dirty_age_limit_show(struct kobject *kobj,
502                                     struct attribute *attr, char *buf)
503 {
504         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
505                                                  ns_kobj);
506
507         return snprintf(buf, PAGE_SIZE, "%llu\n",
508                         ktime_divns(ns->ns_dirty_age_limit, NSEC_PER_SEC));
509 }
510
511 static ssize_t dirty_age_limit_store(struct kobject *kobj,
512                                      struct attribute *attr,
513                                      const char *buffer, size_t count)
514 {
515         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
516                                                  ns_kobj);
517         unsigned long long tmp;
518
519         if (kstrtoull(buffer, 10, &tmp))
520                 return -EINVAL;
521
522         ns->ns_dirty_age_limit = ktime_set(tmp, 0);
523
524         return count;
525 }
526 LUSTRE_RW_ATTR(dirty_age_limit);
527
528 #ifdef HAVE_SERVER_SUPPORT
529 static ssize_t ctime_age_limit_show(struct kobject *kobj,
530                                     struct attribute *attr, char *buf)
531 {
532         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
533                                                  ns_kobj);
534
535         return snprintf(buf, PAGE_SIZE, "%u\n", ns->ns_ctime_age_limit);
536 }
537
538 static ssize_t ctime_age_limit_store(struct kobject *kobj,
539                                      struct attribute *attr,
540                                      const char *buffer, size_t count)
541 {
542         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
543                                                  ns_kobj);
544         unsigned long tmp;
545
546         if (kstrtoul(buffer, 10, &tmp))
547                 return -EINVAL;
548
549         ns->ns_ctime_age_limit = tmp;
550
551         return count;
552 }
553 LUSTRE_RW_ATTR(ctime_age_limit);
554
555 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
556                                   char *buf)
557 {
558         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
559                                                  ns_kobj);
560
561         return sprintf(buf, "%d\n", ns->ns_timeouts);
562 }
563 LUSTRE_RO_ATTR(lock_timeouts);
564
565 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
566                                      struct attribute *attr, char *buf)
567 {
568         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
569                                                  ns_kobj);
570
571         return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
572 }
573
574 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
575                                       struct attribute *attr,
576                                       const char *buffer, size_t count)
577 {
578         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
579                                                  ns_kobj);
580         unsigned long tmp;
581         int err;
582
583         err = kstrtoul(buffer, 10, &tmp);
584         if (err != 0)
585                 return -EINVAL;
586
587         ns->ns_max_nolock_size = tmp;
588
589         return count;
590 }
591 LUSTRE_RW_ATTR(max_nolock_bytes);
592
593 static ssize_t contention_seconds_show(struct kobject *kobj,
594                                        struct attribute *attr, char *buf)
595 {
596         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
597                                                  ns_kobj);
598
599         return scnprintf(buf, PAGE_SIZE, "%d\n", ns->ns_contention_time);
600 }
601
602 static ssize_t contention_seconds_store(struct kobject *kobj,
603                                         struct attribute *attr,
604                                         const char *buffer, size_t count)
605 {
606         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
607                                                  ns_kobj);
608         unsigned int tmp;
609
610         if (kstrtouint(buffer, 10, &tmp))
611                 return -EINVAL;
612
613         ns->ns_contention_time = tmp;
614
615         return count;
616 }
617 LUSTRE_RW_ATTR(contention_seconds);
618
619 static ssize_t contended_locks_show(struct kobject *kobj,
620                                     struct attribute *attr, char *buf)
621 {
622         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
623                                                  ns_kobj);
624
625         return sprintf(buf, "%u\n", ns->ns_contended_locks);
626 }
627
628 static ssize_t contended_locks_store(struct kobject *kobj,
629                                      struct attribute *attr,
630                                      const char *buffer, size_t count)
631 {
632         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
633                                                  ns_kobj);
634         unsigned long tmp;
635         int err;
636
637         err = kstrtoul(buffer, 10, &tmp);
638         if (err != 0)
639                 return -EINVAL;
640
641         ns->ns_contended_locks = tmp;
642
643         return count;
644 }
645 LUSTRE_RW_ATTR(contended_locks);
646
647 static ssize_t max_parallel_ast_show(struct kobject *kobj,
648                                      struct attribute *attr, char *buf)
649 {
650         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
651                                                  ns_kobj);
652
653         return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
654 }
655
656 static ssize_t max_parallel_ast_store(struct kobject *kobj,
657                                       struct attribute *attr,
658                                       const char *buffer, size_t count)
659 {
660         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
661                                                  ns_kobj);
662         unsigned long tmp;
663         int err;
664
665         err = kstrtoul(buffer, 10, &tmp);
666         if (err != 0)
667                 return -EINVAL;
668
669         ns->ns_max_parallel_ast = tmp;
670
671         return count;
672 }
673 LUSTRE_RW_ATTR(max_parallel_ast);
674
675 #endif /* HAVE_SERVER_SUPPORT */
676
677 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
678 static struct attribute *ldlm_ns_attrs[] = {
679         &lustre_attr_resource_count.attr,
680         &lustre_attr_lock_count.attr,
681         &lustre_attr_lock_unused_count.attr,
682         &lustre_attr_ns_recalc_pct.attr,
683         &lustre_attr_lru_size.attr,
684         &lustre_attr_lru_cancel_batch.attr,
685         &lustre_attr_lru_max_age.attr,
686         &lustre_attr_early_lock_cancel.attr,
687         &lustre_attr_dirty_age_limit.attr,
688 #ifdef HAVE_SERVER_SUPPORT
689         &lustre_attr_ctime_age_limit.attr,
690         &lustre_attr_lock_timeouts.attr,
691         &lustre_attr_max_nolock_bytes.attr,
692         &lustre_attr_contention_seconds.attr,
693         &lustre_attr_contended_locks.attr,
694         &lustre_attr_max_parallel_ast.attr,
695 #endif
696         NULL,
697 };
698
699 static void ldlm_ns_release(struct kobject *kobj)
700 {
701         struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
702                                                  ns_kobj);
703         complete(&ns->ns_kobj_unregister);
704 }
705
706 static struct kobj_type ldlm_ns_ktype = {
707         .default_attrs  = ldlm_ns_attrs,
708         .sysfs_ops      = &lustre_sysfs_ops,
709         .release        = ldlm_ns_release,
710 };
711
712 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
713 {
714         if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
715                 CERROR("dlm namespace %s has no procfs dir?\n",
716                        ldlm_ns_name(ns));
717         else
718                 debugfs_remove_recursive(ns->ns_debugfs_entry);
719
720         if (ns->ns_stats != NULL)
721                 lprocfs_free_stats(&ns->ns_stats);
722 }
723
724 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
725 {
726         kobject_put(&ns->ns_kobj);
727         wait_for_completion(&ns->ns_kobj_unregister);
728 }
729
730 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
731 {
732         int err;
733
734         ns->ns_kobj.kset = ldlm_ns_kset;
735         init_completion(&ns->ns_kobj_unregister);
736         err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
737                                    "%s", ldlm_ns_name(ns));
738
739         ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
740         if (!ns->ns_stats) {
741                 kobject_put(&ns->ns_kobj);
742                 return -ENOMEM;
743         }
744
745         lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
746                              LPROCFS_CNTR_AVGMINMAX, "locks", "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 ldlm_res_hop_hash(struct cfs_hash *hs,
770                                   const void *key, unsigned int mask)
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         return val & mask;
779 }
780
781 static unsigned int ldlm_res_hop_fid_hash(const struct ldlm_res_id *id, unsigned int bits)
782 {
783         struct lu_fid       fid;
784         __u32               hash;
785         __u32               val;
786
787         fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
788         fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
789         fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
790
791         hash = fid_flatten32(&fid);
792         hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
793         if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
794                 val = id->name[LUSTRE_RES_ID_HSH_OFF];
795         } else {
796                 val = fid_oid(&fid);
797         }
798         hash += (val >> 5) + (val << 11);
799         return cfs_hash_32(hash, bits);
800 }
801
802 static void *ldlm_res_hop_key(struct hlist_node *hnode)
803 {
804         struct ldlm_resource   *res;
805
806         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
807         return &res->lr_name;
808 }
809
810 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
811 {
812         struct ldlm_resource   *res;
813
814         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
815         return ldlm_res_eq((const struct ldlm_res_id *)key,
816                            (const struct ldlm_res_id *)&res->lr_name);
817 }
818
819 static void *ldlm_res_hop_object(struct hlist_node *hnode)
820 {
821         return hlist_entry(hnode, struct ldlm_resource, lr_hash);
822 }
823
824 static void
825 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
826 {
827         struct ldlm_resource *res;
828
829         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
830         ldlm_resource_getref(res);
831 }
832
833 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
834 {
835         struct ldlm_resource *res;
836
837         res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
838         ldlm_resource_putref(res);
839 }
840
841 static struct cfs_hash_ops ldlm_ns_hash_ops = {
842         .hs_hash        = ldlm_res_hop_hash,
843         .hs_key         = ldlm_res_hop_key,
844         .hs_keycmp      = ldlm_res_hop_keycmp,
845         .hs_keycpy      = NULL,
846         .hs_object      = ldlm_res_hop_object,
847         .hs_get         = ldlm_res_hop_get_locked,
848         .hs_put         = ldlm_res_hop_put
849 };
850
851 static struct {
852         /** hash bucket bits */
853         unsigned                nsd_bkt_bits;
854         /** hash bits */
855         unsigned                nsd_all_bits;
856 } ldlm_ns_hash_defs[] = {
857         [LDLM_NS_TYPE_MDC] = {
858                 .nsd_bkt_bits   = 11,
859                 .nsd_all_bits   = 16,
860         },
861         [LDLM_NS_TYPE_MDT] = {
862                 .nsd_bkt_bits   = 14,
863                 .nsd_all_bits   = 21,
864         },
865         [LDLM_NS_TYPE_OSC] = {
866                 .nsd_bkt_bits   = 8,
867                 .nsd_all_bits   = 12,
868         },
869         [LDLM_NS_TYPE_OST] = {
870                 .nsd_bkt_bits   = 11,
871                 .nsd_all_bits   = 17,
872         },
873         [LDLM_NS_TYPE_MGC] = {
874                 .nsd_bkt_bits   = 3,
875                 .nsd_all_bits   = 4,
876         },
877         [LDLM_NS_TYPE_MGT] = {
878                 .nsd_bkt_bits   = 3,
879                 .nsd_all_bits   = 4,
880         },
881 };
882
883 /**
884  * Create and initialize new empty namespace.
885  */
886 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
887                                           enum ldlm_side client,
888                                           enum ldlm_appetite apt,
889                                           enum ldlm_ns_type ns_type)
890 {
891         struct ldlm_namespace *ns = NULL;
892         int idx;
893         int rc;
894
895         ENTRY;
896         LASSERT(obd != NULL);
897
898         rc = ldlm_get_ref();
899         if (rc) {
900                 CERROR("%s: ldlm_get_ref failed: rc = %d\n", name, rc);
901                 RETURN(ERR_PTR(rc));
902         }
903
904         if (ns_type >= ARRAY_SIZE(ldlm_ns_hash_defs) ||
905             ldlm_ns_hash_defs[ns_type].nsd_bkt_bits == 0) {
906                 rc = -EINVAL;
907                 CERROR("%s: unknown namespace type %d: rc = %d\n",
908                        name, ns_type, rc);
909                 GOTO(out_ref, rc);
910         }
911
912         OBD_ALLOC_PTR(ns);
913         if (!ns)
914                 GOTO(out_ref, rc = -ENOMEM);
915
916         ns->ns_rs_hash = cfs_hash_create(name,
917                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
918                                          ldlm_ns_hash_defs[ns_type].nsd_all_bits,
919                                          ldlm_ns_hash_defs[ns_type].nsd_bkt_bits,
920                                          0,
921                                          CFS_HASH_MIN_THETA,
922                                          CFS_HASH_MAX_THETA,
923                                          &ldlm_ns_hash_ops,
924                                          CFS_HASH_DEPTH |
925                                          CFS_HASH_BIGNAME |
926                                          CFS_HASH_SPIN_BKTLOCK |
927                                          CFS_HASH_NO_ITEMREF);
928         if (!ns->ns_rs_hash)
929                 GOTO(out_ns, rc = -ENOMEM);
930
931         ns->ns_bucket_bits = ldlm_ns_hash_defs[ns_type].nsd_all_bits -
932                              ldlm_ns_hash_defs[ns_type].nsd_bkt_bits;
933
934         OBD_ALLOC_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
935         if (!ns->ns_rs_buckets)
936                 GOTO(out_hash, rc = -ENOMEM);
937
938         for (idx = 0; idx < (1 << ns->ns_bucket_bits); idx++) {
939                 struct ldlm_ns_bucket *nsb = &ns->ns_rs_buckets[idx];
940
941                 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
942                 nsb->nsb_namespace = ns;
943                 nsb->nsb_reclaim_start = 0;
944                 atomic_set(&nsb->nsb_count, 0);
945         }
946
947         ns->ns_obd = obd;
948         ns->ns_appetite = apt;
949         ns->ns_client = client;
950         ns->ns_name = kstrdup(name, GFP_KERNEL);
951         if (!ns->ns_name)
952                 GOTO(out_hash, rc = -ENOMEM);
953
954         INIT_LIST_HEAD(&ns->ns_list_chain);
955         INIT_LIST_HEAD(&ns->ns_unused_list);
956         spin_lock_init(&ns->ns_lock);
957         atomic_set(&ns->ns_bref, 0);
958         init_waitqueue_head(&ns->ns_waitq);
959
960         ns->ns_max_nolock_size    = NS_DEFAULT_MAX_NOLOCK_BYTES;
961         ns->ns_contention_time    = NS_DEFAULT_CONTENTION_SECONDS;
962         ns->ns_contended_locks    = NS_DEFAULT_CONTENDED_LOCKS;
963
964         ns->ns_max_parallel_ast   = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
965         ns->ns_nr_unused          = 0;
966         ns->ns_max_unused         = LDLM_DEFAULT_LRU_SIZE;
967         ns->ns_cancel_batch       = LDLM_DEFAULT_LRU_SHRINK_BATCH;
968         ns->ns_recalc_pct         = LDLM_DEFAULT_SLV_RECALC_PCT;
969         ns->ns_max_age            = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
970         ns->ns_ctime_age_limit    = LDLM_CTIME_AGE_LIMIT;
971         ns->ns_dirty_age_limit    = ktime_set(LDLM_DIRTY_AGE_LIMIT, 0);
972         ns->ns_timeouts           = 0;
973         ns->ns_orig_connect_flags = 0;
974         ns->ns_connect_flags      = 0;
975         ns->ns_stopping           = 0;
976         ns->ns_reclaim_start      = 0;
977         ns->ns_last_pos           = &ns->ns_unused_list;
978         ns->ns_flags              = 0;
979
980         rc = ldlm_namespace_sysfs_register(ns);
981         if (rc) {
982                 CERROR("%s: cannot initialize ns sysfs: rc = %d\n", name, rc);
983                 GOTO(out_hash, rc);
984         }
985
986         rc = ldlm_namespace_debugfs_register(ns);
987         if (rc) {
988                 CERROR("%s: cannot initialize ns proc: rc = %d\n", name, rc);
989                 GOTO(out_sysfs, rc);
990         }
991
992         idx = ldlm_namespace_nr_read(client);
993         rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
994         if (rc) {
995                 CERROR("%s: cannot initialize lock pool, rc = %d\n", name, rc);
996                 GOTO(out_proc, rc);
997         }
998
999         ldlm_namespace_register(ns, client);
1000         RETURN(ns);
1001 out_proc:
1002         ldlm_namespace_debugfs_unregister(ns);
1003 out_sysfs:
1004         ldlm_namespace_sysfs_unregister(ns);
1005         ldlm_namespace_cleanup(ns, 0);
1006 out_hash:
1007         OBD_FREE_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
1008         kfree(ns->ns_name);
1009         cfs_hash_putref(ns->ns_rs_hash);
1010 out_ns:
1011         OBD_FREE_PTR(ns);
1012 out_ref:
1013         ldlm_put_ref();
1014         RETURN(ERR_PTR(rc));
1015 }
1016 EXPORT_SYMBOL(ldlm_namespace_new);
1017
1018 /**
1019  * Cancel and destroy all locks on a resource.
1020  *
1021  * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
1022  * clean up.  This is currently only used for recovery, and we make
1023  * certain assumptions as a result--notably, that we shouldn't cancel
1024  * locks with refs.
1025  */
1026 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
1027                              __u64 flags)
1028 {
1029         struct list_head *tmp;
1030         int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
1031         bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
1032
1033         do {
1034                 struct ldlm_lock *lock = NULL;
1035
1036                 /* First, we look for non-cleaned-yet lock
1037                  * all cleaned locks are marked by CLEANED flag. */
1038                 lock_res(res);
1039                 list_for_each(tmp, q) {
1040                         lock = list_entry(tmp, struct ldlm_lock,
1041                                           l_res_link);
1042                         if (ldlm_is_cleaned(lock)) {
1043                                 lock = NULL;
1044                                 continue;
1045                         }
1046                         LDLM_LOCK_GET(lock);
1047                         ldlm_set_cleaned(lock);
1048                         break;
1049                 }
1050
1051                 if (lock == NULL) {
1052                         unlock_res(res);
1053                         break;
1054                 }
1055
1056                 /* Set CBPENDING so nothing in the cancellation path
1057                  * can match this lock. */
1058                 ldlm_set_cbpending(lock);
1059                 ldlm_set_failed(lock);
1060                 lock->l_flags |= flags;
1061
1062                 /* ... without sending a CANCEL message for local_only. */
1063                 if (local_only)
1064                         ldlm_set_local_only(lock);
1065
1066                 if (local_only && (lock->l_readers || lock->l_writers)) {
1067                         /*
1068                          * This is a little bit gross, but much better than the
1069                          * alternative: pretend that we got a blocking AST from
1070                          * the server, so that when the lock is decref'd, it
1071                          * will go away ...
1072                          */
1073                         unlock_res(res);
1074                         LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1075                         if (lock->l_flags & LDLM_FL_FAIL_LOC)
1076                                 schedule_timeout_uninterruptible(
1077                                         cfs_time_seconds(4));
1078
1079                         if (lock->l_completion_ast)
1080                                 lock->l_completion_ast(lock,
1081                                                        LDLM_FL_FAILED, NULL);
1082                         LDLM_LOCK_RELEASE(lock);
1083                         continue;
1084                 }
1085
1086                 if (client) {
1087                         struct lustre_handle lockh;
1088
1089                         unlock_res(res);
1090                         ldlm_lock2handle(lock, &lockh);
1091                         rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1092                         if (rc)
1093                                 CERROR("ldlm_cli_cancel: %d\n", rc);
1094                 } else {
1095                         unlock_res(res);
1096                         LDLM_DEBUG(lock,
1097                                    "Freeing a lock still held by a client node");
1098                         ldlm_lock_cancel(lock);
1099                 }
1100                 LDLM_LOCK_RELEASE(lock);
1101         } while (1);
1102 }
1103
1104 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1105                                struct hlist_node *hnode, void *arg)
1106 {
1107         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1108         __u64 flags = *(__u64 *)arg;
1109
1110         cleanup_resource(res, &res->lr_granted, flags);
1111         cleanup_resource(res, &res->lr_waiting, flags);
1112
1113         return 0;
1114 }
1115
1116 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1117                                   struct hlist_node *hnode, void *arg)
1118 {
1119         struct ldlm_resource  *res = cfs_hash_object(hs, hnode);
1120
1121         lock_res(res);
1122         CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1123                "(%d) after lock cleanup; forcing cleanup.\n",
1124                ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1125                atomic_read(&res->lr_refcount) - 1);
1126
1127         /* Use D_NETERROR since it is in the default mask */
1128         ldlm_resource_dump(D_NETERROR, res);
1129         unlock_res(res);
1130         return 0;
1131 }
1132
1133 /**
1134  * Cancel and destroy all locks in the namespace.
1135  *
1136  * Typically used during evictions when server notified client that it was
1137  * evicted and all of its state needs to be destroyed.
1138  * Also used during shutdown.
1139  */
1140 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1141 {
1142         if (ns == NULL) {
1143                 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1144                 return ELDLM_OK;
1145         }
1146
1147         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1148                                  &flags, 0);
1149         cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1150                                  NULL, 0);
1151         return ELDLM_OK;
1152 }
1153 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1154
1155 /**
1156  * Attempts to free namespace.
1157  *
1158  * Only used when namespace goes away, like during an unmount.
1159  */
1160 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1161 {
1162         ENTRY;
1163
1164         /* At shutdown time, don't call the cancellation callback */
1165         ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1166
1167         if (atomic_read(&ns->ns_bref) > 0) {
1168                 int rc;
1169                 CDEBUG(D_DLMTRACE,
1170                        "dlm namespace %s free waiting on refcount %d\n",
1171                        ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1172 force_wait:
1173                 if (force)
1174                         rc = wait_event_idle_timeout(
1175                                 ns->ns_waitq,
1176                                 atomic_read(&ns->ns_bref) == 0,
1177                                 cfs_time_seconds(1) / 4);
1178                 else
1179                         rc = l_wait_event_abortable(
1180                                 ns->ns_waitq, atomic_read(&ns->ns_bref) == 0);
1181
1182                 /* Forced cleanups should be able to reclaim all references,
1183                  * so it's safe to wait forever... we can't leak locks... */
1184                 if (force && rc == 0) {
1185                         rc = -ETIMEDOUT;
1186                         LCONSOLE_ERROR("Forced cleanup waiting for %s "
1187                                        "namespace with %d resources in use, "
1188                                        "(rc=%d)\n", ldlm_ns_name(ns),
1189                                        atomic_read(&ns->ns_bref), rc);
1190                         GOTO(force_wait, rc);
1191                 }
1192
1193                 if (atomic_read(&ns->ns_bref)) {
1194                         LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1195                                        "with %d resources in use, (rc=%d)\n",
1196                                        ldlm_ns_name(ns),
1197                                        atomic_read(&ns->ns_bref), rc);
1198                         RETURN(ELDLM_NAMESPACE_EXISTS);
1199                 }
1200                 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1201                        ldlm_ns_name(ns));
1202         }
1203
1204         RETURN(ELDLM_OK);
1205 }
1206
1207 /**
1208  * Performs various cleanups for passed \a ns to make it drop refc and be
1209  * ready for freeing. Waits for refc == 0.
1210  *
1211  * The following is done:
1212  * (0) Unregister \a ns from its list to make inaccessible for potential
1213  * users like pools thread and others;
1214  * (1) Clear all locks in \a ns.
1215  */
1216 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1217                                struct obd_import *imp,
1218                                int force)
1219 {
1220         int rc;
1221
1222         ENTRY;
1223         if (!ns) {
1224                 EXIT;
1225                 return;
1226         }
1227
1228         spin_lock(&ns->ns_lock);
1229         ns->ns_stopping = 1;
1230         spin_unlock(&ns->ns_lock);
1231
1232         /*
1233          * Can fail with -EINTR when force == 0 in which case try harder.
1234          */
1235         rc = __ldlm_namespace_free(ns, force);
1236         if (rc != ELDLM_OK) {
1237                 if (imp) {
1238                         ptlrpc_disconnect_import(imp, 0);
1239                         ptlrpc_invalidate_import(imp);
1240                 }
1241
1242                 /*
1243                  * With all requests dropped and the import inactive
1244                  * we are gaurenteed all reference will be dropped.
1245                  */
1246                 rc = __ldlm_namespace_free(ns, 1);
1247                 LASSERT(rc == 0);
1248         }
1249         EXIT;
1250 }
1251 EXPORT_SYMBOL(ldlm_namespace_free_prior);
1252
1253 /**
1254  * Performs freeing memory structures related to \a ns. This is only done
1255  * when ldlm_namespce_free_prior() successfully removed all resources
1256  * referencing \a ns and its refc == 0.
1257  */
1258 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
1259 {
1260         ENTRY;
1261         if (!ns) {
1262                 EXIT;
1263                 return;
1264         }
1265
1266         /* Make sure that nobody can find this ns in its list. */
1267         ldlm_namespace_unregister(ns, ns->ns_client);
1268         /* Fini pool _before_ parent proc dir is removed. This is important as
1269          * ldlm_pool_fini() removes own proc dir which is child to @dir.
1270          * Removing it after @dir may cause oops. */
1271         ldlm_pool_fini(&ns->ns_pool);
1272
1273         ldlm_namespace_debugfs_unregister(ns);
1274         ldlm_namespace_sysfs_unregister(ns);
1275         cfs_hash_putref(ns->ns_rs_hash);
1276         OBD_FREE_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
1277         kfree(ns->ns_name);
1278         /* Namespace \a ns should be not on list at this time, otherwise
1279          * this will cause issues related to using freed \a ns in poold
1280          * thread.
1281          */
1282         LASSERT(list_empty(&ns->ns_list_chain));
1283         OBD_FREE_PTR(ns);
1284         ldlm_put_ref();
1285         EXIT;
1286 }
1287 EXPORT_SYMBOL(ldlm_namespace_free_post);
1288
1289 /**
1290  * Cleanup the resource, and free namespace.
1291  * bug 12864:
1292  * Deadlock issue:
1293  * proc1: destroy import
1294  *        class_disconnect_export(grab cl_sem) ->
1295  *              -> ldlm_namespace_free ->
1296  *              -> lprocfs_remove(grab _lprocfs_lock).
1297  * proc2: read proc info
1298  *        lprocfs_fops_read(grab _lprocfs_lock) ->
1299  *              -> osc_rd_active, etc(grab cl_sem).
1300  *
1301  * So that I have to split the ldlm_namespace_free into two parts - the first
1302  * part ldlm_namespace_free_prior is used to cleanup the resource which is
1303  * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1304  * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1305  * held.
1306  */
1307 void ldlm_namespace_free(struct ldlm_namespace *ns,
1308                          struct obd_import *imp,
1309                          int force)
1310 {
1311         ldlm_namespace_free_prior(ns, imp, force);
1312         ldlm_namespace_free_post(ns);
1313 }
1314 EXPORT_SYMBOL(ldlm_namespace_free);
1315
1316 void ldlm_namespace_get(struct ldlm_namespace *ns)
1317 {
1318         atomic_inc(&ns->ns_bref);
1319 }
1320
1321 /* This is only for callers that care about refcount */
1322 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1323 {
1324         return atomic_inc_return(&ns->ns_bref);
1325 }
1326
1327 void ldlm_namespace_put(struct ldlm_namespace *ns)
1328 {
1329         if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1330                 wake_up(&ns->ns_waitq);
1331                 spin_unlock(&ns->ns_lock);
1332         }
1333 }
1334
1335 /** Register \a ns in the list of namespaces */
1336 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1337 {
1338         mutex_lock(ldlm_namespace_lock(client));
1339         LASSERT(list_empty(&ns->ns_list_chain));
1340         list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1341         ldlm_namespace_nr_inc(client);
1342         mutex_unlock(ldlm_namespace_lock(client));
1343 }
1344
1345 /** Unregister \a ns from the list of namespaces. */
1346 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1347 {
1348         mutex_lock(ldlm_namespace_lock(client));
1349         LASSERT(!list_empty(&ns->ns_list_chain));
1350         /* Some asserts and possibly other parts of the code are still
1351          * using list_empty(&ns->ns_list_chain). This is why it is
1352          * important to use list_del_init() here. */
1353         list_del_init(&ns->ns_list_chain);
1354         ldlm_namespace_nr_dec(client);
1355         mutex_unlock(ldlm_namespace_lock(client));
1356 }
1357
1358 /** Should be called with ldlm_namespace_lock(client) taken. */
1359 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1360                                           enum ldlm_side client)
1361 {
1362         LASSERT(!list_empty(&ns->ns_list_chain));
1363         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1364         list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1365 }
1366
1367 /** Should be called with ldlm_namespace_lock(client) taken. */
1368 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1369                                             enum ldlm_side client)
1370 {
1371         LASSERT(!list_empty(&ns->ns_list_chain));
1372         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1373         list_move_tail(&ns->ns_list_chain,
1374                        ldlm_namespace_inactive_list(client));
1375 }
1376
1377 /** Should be called with ldlm_namespace_lock(client) taken. */
1378 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1379 {
1380         LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1381         LASSERT(!list_empty(ldlm_namespace_list(client)));
1382         return container_of(ldlm_namespace_list(client)->next,
1383                             struct ldlm_namespace, ns_list_chain);
1384 }
1385
1386 static bool ldlm_resource_extent_new(struct ldlm_resource *res)
1387 {
1388         int idx;
1389
1390         OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1391                        sizeof(*res->lr_itree) * LCK_MODE_NUM);
1392         if (res->lr_itree == NULL)
1393                 return false;
1394         /* Initialize interval trees for each lock mode. */
1395         for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1396                 res->lr_itree[idx].lit_size = 0;
1397                 res->lr_itree[idx].lit_mode = BIT(idx);
1398                 res->lr_itree[idx].lit_root = NULL;
1399         }
1400         return true;
1401 }
1402
1403 static bool ldlm_resource_inodebits_new(struct ldlm_resource *res)
1404 {
1405         int i;
1406
1407         OBD_ALLOC_PTR(res->lr_ibits_queues);
1408         if (res->lr_ibits_queues == NULL)
1409                 return false;
1410         for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
1411                 INIT_LIST_HEAD(&res->lr_ibits_queues->liq_waiting[i]);
1412         return true;
1413 }
1414
1415 /** Create and initialize new resource. */
1416 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1417 {
1418         struct ldlm_resource *res;
1419         bool rc;
1420
1421         OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1422         if (res == NULL)
1423                 return NULL;
1424
1425         switch (ldlm_type) {
1426         case LDLM_EXTENT:
1427                 rc = ldlm_resource_extent_new(res);
1428                 break;
1429         case LDLM_IBITS:
1430                 rc = ldlm_resource_inodebits_new(res);
1431                 break;
1432         default:
1433                 rc = true;
1434                 break;
1435         }
1436         if (!rc) {
1437                 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1438                 return NULL;
1439         }
1440
1441         INIT_LIST_HEAD(&res->lr_granted);
1442         INIT_LIST_HEAD(&res->lr_waiting);
1443
1444         atomic_set(&res->lr_refcount, 1);
1445         spin_lock_init(&res->lr_lock);
1446         lu_ref_init(&res->lr_reference);
1447
1448         /* Since LVB init can be delayed now, there is no longer need to
1449          * immediatelly acquire mutex here. */
1450         mutex_init(&res->lr_lvb_mutex);
1451         res->lr_lvb_initialized = false;
1452
1453         return res;
1454 }
1455
1456 static void ldlm_resource_free(struct ldlm_resource *res)
1457 {
1458         if (res->lr_type == LDLM_EXTENT) {
1459                 if (res->lr_itree != NULL)
1460                         OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1461                                       sizeof(*res->lr_itree) * LCK_MODE_NUM);
1462         } else if (res->lr_type == LDLM_IBITS) {
1463                 if (res->lr_ibits_queues != NULL)
1464                         OBD_FREE_PTR(res->lr_ibits_queues);
1465         }
1466
1467         OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1468 }
1469
1470 /**
1471  * Return a reference to resource with given name, creating it if necessary.
1472  * Args: namespace with ns_lock unlocked
1473  * Locks: takes and releases NS hash-lock and res->lr_lock
1474  * Returns: referenced, unlocked ldlm_resource or NULL
1475  */
1476 struct ldlm_resource *
1477 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1478                   const struct ldlm_res_id *name, enum ldlm_type type,
1479                   int create)
1480 {
1481         struct hlist_node       *hnode;
1482         struct ldlm_resource    *res = NULL;
1483         struct cfs_hash_bd              bd;
1484         __u64                   version;
1485         int                     ns_refcount = 0;
1486         int hash;
1487
1488         LASSERT(ns != NULL);
1489         LASSERT(parent == NULL);
1490         LASSERT(ns->ns_rs_hash != NULL);
1491         LASSERT(name->name[0] != 0);
1492
1493         cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1494         hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1495         if (hnode != NULL) {
1496                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1497                 GOTO(found, res);
1498         }
1499
1500         version = cfs_hash_bd_version_get(&bd);
1501         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1502
1503         if (create == 0)
1504                 return ERR_PTR(-ENOENT);
1505
1506         LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1507                  "type: %d\n", type);
1508         res = ldlm_resource_new(type);
1509         if (res == NULL)
1510                 return ERR_PTR(-ENOMEM);
1511
1512         hash = ldlm_res_hop_fid_hash(name, ns->ns_bucket_bits);
1513         res->lr_ns_bucket = &ns->ns_rs_buckets[hash];
1514         res->lr_name = *name;
1515         res->lr_type = type;
1516
1517         cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1518         hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1519                 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1520
1521         if (hnode != NULL) {
1522                 /* Someone won the race and already added the resource. */
1523                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1524                 /* Clean lu_ref for failed resource. */
1525                 lu_ref_fini(&res->lr_reference);
1526                 ldlm_resource_free(res);
1527 found:
1528                 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1529                 return res;
1530         }
1531         /* We won! Let's add the resource. */
1532         cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1533         if (atomic_inc_return(&res->lr_ns_bucket->nsb_count) == 1)
1534                 ns_refcount = ldlm_namespace_get_return(ns);
1535
1536         cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1537
1538         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1539
1540         /* Let's see if we happened to be the very first resource in this
1541          * namespace. If so, and this is a client namespace, we need to move
1542          * the namespace into the active namespaces list to be patrolled by
1543          * the ldlm_poold. */
1544         if (ns_is_client(ns) && ns_refcount == 1) {
1545                 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1546                 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1547                 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1548         }
1549
1550         return res;
1551 }
1552 EXPORT_SYMBOL(ldlm_resource_get);
1553
1554 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1555 {
1556         LASSERT(res != NULL);
1557         LASSERT(res != LP_POISON);
1558         atomic_inc(&res->lr_refcount);
1559         CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1560                atomic_read(&res->lr_refcount));
1561         return res;
1562 }
1563
1564 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1565                                          struct ldlm_resource *res)
1566 {
1567         struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1568
1569         if (!list_empty(&res->lr_granted)) {
1570                 ldlm_resource_dump(D_ERROR, res);
1571                 LBUG();
1572         }
1573
1574         if (!list_empty(&res->lr_waiting)) {
1575                 ldlm_resource_dump(D_ERROR, res);
1576                 LBUG();
1577         }
1578
1579         cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1580                                bd, &res->lr_hash);
1581         lu_ref_fini(&res->lr_reference);
1582         if (atomic_dec_and_test(&nsb->nsb_count))
1583                 ldlm_namespace_put(nsb->nsb_namespace);
1584 }
1585
1586 /* Returns 1 if the resource was freed, 0 if it remains. */
1587 int ldlm_resource_putref(struct ldlm_resource *res)
1588 {
1589         struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1590         struct cfs_hash_bd   bd;
1591
1592         LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1593         CDEBUG(D_INFO, "putref res: %p count: %d\n",
1594                res, atomic_read(&res->lr_refcount) - 1);
1595
1596         cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1597         if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1598                 __ldlm_resource_putref_final(&bd, res);
1599                 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1600                 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1601                         ns->ns_lvbo->lvbo_free(res);
1602                 ldlm_resource_free(res);
1603                 return 1;
1604         }
1605         return 0;
1606 }
1607 EXPORT_SYMBOL(ldlm_resource_putref);
1608
1609 static void __ldlm_resource_add_lock(struct ldlm_resource *res,
1610                                      struct list_head *head,
1611                                      struct ldlm_lock *lock,
1612                                      bool tail)
1613 {
1614         check_res_locked(res);
1615
1616         if (ldlm_is_destroyed(lock)) {
1617                 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1618                 return;
1619         }
1620
1621         LASSERT(list_empty(&lock->l_res_link));
1622
1623         if (tail)
1624                 list_add_tail(&lock->l_res_link, head);
1625         else
1626                 list_add(&lock->l_res_link, head);
1627
1628         if (res->lr_type == LDLM_IBITS)
1629                 ldlm_inodebits_add_lock(res, head, lock, tail);
1630
1631         ldlm_resource_dump(D_INFO, res);
1632 }
1633
1634 /**
1635  * Add a lock into a given resource into specified lock list.
1636  */
1637 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1638                             struct ldlm_lock *lock)
1639 {
1640         LDLM_DEBUG(lock, "About to add this lock");
1641
1642         __ldlm_resource_add_lock(res, head, lock, true);
1643 }
1644
1645 /**
1646  * Insert a lock into resource after specified lock.
1647  */
1648 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1649                                      struct ldlm_lock *new)
1650 {
1651         LASSERT(!list_empty(&original->l_res_link));
1652
1653         LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1654         __ldlm_resource_add_lock(original->l_resource,
1655                                  &original->l_res_link,
1656                                  new, false);
1657 }
1658
1659 /**
1660  * Insert a lock into resource before the specified lock.
1661  *
1662  * IBITS waiting locks are to be inserted to the ibit lists as well, and only
1663  * the insert-after operation is supported for them, because the set of bits
1664  * of the previous and the new locks must match. Therefore, get the previous
1665  * lock and insert after.
1666  */
1667 void ldlm_resource_insert_lock_before(struct ldlm_lock *original,
1668                                       struct ldlm_lock *new)
1669 {
1670         LASSERT(!list_empty(&original->l_res_link));
1671
1672         LDLM_DEBUG(new, "About to insert this lock before %p: ", original);
1673         __ldlm_resource_add_lock(original->l_resource,
1674                                  original->l_res_link.prev, new, false);
1675 }
1676
1677 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1678 {
1679         int type = lock->l_resource->lr_type;
1680
1681         check_res_locked(lock->l_resource);
1682         switch (type) {
1683         case LDLM_PLAIN:
1684                 ldlm_unlink_lock_skiplist(lock);
1685                 break;
1686         case LDLM_EXTENT:
1687                 ldlm_extent_unlink_lock(lock);
1688                 break;
1689         case LDLM_IBITS:
1690                 ldlm_inodebits_unlink_lock(lock);
1691                 break;
1692         }
1693         list_del_init(&lock->l_res_link);
1694 }
1695 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1696
1697 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1698 {
1699         desc->lr_type = res->lr_type;
1700         desc->lr_name = res->lr_name;
1701 }
1702
1703 /**
1704  * Print information about all locks in all namespaces on this node to debug
1705  * log.
1706  */
1707 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1708 {
1709         struct list_head *tmp;
1710
1711         if (!((libcfs_debug | D_ERROR) & level))
1712                 return;
1713
1714         mutex_lock(ldlm_namespace_lock(client));
1715
1716         list_for_each(tmp, ldlm_namespace_list(client)) {
1717                 struct ldlm_namespace *ns;
1718
1719                 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1720                 ldlm_namespace_dump(level, ns);
1721         }
1722
1723         mutex_unlock(ldlm_namespace_lock(client));
1724 }
1725
1726 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1727                               struct hlist_node *hnode, void *arg)
1728 {
1729         struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1730         int    level = (int)(unsigned long)arg;
1731
1732         lock_res(res);
1733         ldlm_resource_dump(level, res);
1734         unlock_res(res);
1735
1736         return 0;
1737 }
1738
1739 /**
1740  * Print information about all locks in this namespace on this node to debug
1741  * log.
1742  */
1743 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1744 {
1745         if (!((libcfs_debug | D_ERROR) & level))
1746                 return;
1747
1748         CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1749                ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1750                ns_is_client(ns) ? "client" : "server");
1751
1752         if (ktime_get_seconds() < ns->ns_next_dump)
1753                 return;
1754
1755         cfs_hash_for_each_nolock(ns->ns_rs_hash,
1756                                  ldlm_res_hash_dump,
1757                                  (void *)(unsigned long)level, 0);
1758         spin_lock(&ns->ns_lock);
1759         ns->ns_next_dump = ktime_get_seconds() + 10;
1760         spin_unlock(&ns->ns_lock);
1761 }
1762
1763 /**
1764  * Print information about all locks in this resource to debug log.
1765  */
1766 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1767 {
1768         struct ldlm_lock *lock;
1769         unsigned int granted = 0;
1770
1771         BUILD_BUG_ON(RES_NAME_SIZE != 4);
1772
1773         if (!((libcfs_debug | D_ERROR) & level))
1774                 return;
1775
1776         CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1777                PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1778
1779         if (!list_empty(&res->lr_granted)) {
1780                 CDEBUG(level, "Granted locks (in reverse order):\n");
1781                 list_for_each_entry_reverse(lock, &res->lr_granted,
1782                                                 l_res_link) {
1783                         LDLM_DEBUG_LIMIT(level, lock, "###");
1784                         if (!(level & D_CANTMASK) &&
1785                             ++granted > ldlm_dump_granted_max) {
1786                                 CDEBUG(level,
1787                                        "only dump %d granted locks to avoid DDOS.\n",
1788                                        granted);
1789                                 break;
1790                         }
1791                 }
1792         }
1793
1794         if (!list_empty(&res->lr_waiting)) {
1795                 CDEBUG(level, "Waiting locks:\n");
1796                 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1797                         LDLM_DEBUG_LIMIT(level, lock, "###");
1798         }
1799 }
1800 EXPORT_SYMBOL(ldlm_resource_dump);