4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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.
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).
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
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2010, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
31 * lustre/ldlm/ldlm_resource.c
33 * Author: Phil Schwan <phil@clusterfs.com>
34 * Author: Peter Braam <braam@clusterfs.com>
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"
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;
48 int ldlm_srv_namespace_nr = 0;
49 int ldlm_cli_namespace_nr = 0;
51 DEFINE_MUTEX(ldlm_srv_namespace_lock);
52 LIST_HEAD(ldlm_srv_namespace_list);
54 DEFINE_MUTEX(ldlm_cli_namespace_lock);
55 /* Client Namespaces that have active resources in them.
56 * Once all resources go away, ldlm_poold moves such namespaces to the
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);
62 static struct dentry *ldlm_debugfs_dir;
63 static struct dentry *ldlm_ns_debugfs_dir;
64 struct dentry *ldlm_svc_debugfs_dir;
66 /* during debug dump certain amount of granted locks for one resource to avoid
68 static unsigned int ldlm_dump_granted_max = 256;
70 static ssize_t ldebugfs_dump_ns_seq_write(struct file *file,
71 const char __user *buffer,
72 size_t count, loff_t *off)
74 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
75 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
79 LDEBUGFS_FOPS_WR_ONLY(ldlm, dump_ns);
81 static int ldlm_rw_uint_seq_show(struct seq_file *m, void *v)
83 seq_printf(m, "%u\n", *(unsigned int *)m->private);
88 ldlm_rw_uint_seq_write(struct file *file, const char __user *buffer,
89 size_t count, loff_t *off)
91 struct seq_file *seq = file->private_data;
96 return kstrtouint_from_user(buffer, count, 0,
97 (unsigned int *)seq->private);
100 LDEBUGFS_SEQ_FOPS(ldlm_rw_uint);
102 #ifdef HAVE_SERVER_SUPPORT
104 static int seq_watermark_show(struct seq_file *m, void *data)
106 seq_printf(m, "%llu\n", *(__u64 *)m->private);
110 static ssize_t seq_watermark_write(struct file *file,
111 const char __user *buffer, size_t count,
114 struct seq_file *m = file->private_data;
117 __u64 *data = m->private;
118 bool wm_low = (data == &ldlm_reclaim_threshold_mb) ? true : false;
119 char kernbuf[22] = "";
122 if (count >= sizeof(kernbuf))
125 if (copy_from_user(kernbuf, buffer, count))
129 rc = sysfs_memparse(kernbuf, count, &value, "MiB");
131 CERROR("Failed to set %s, rc = %d.\n",
132 wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb",
135 } else if (value != 0 && value < (1 << 20)) {
136 CERROR("%s should be greater than 1MB.\n",
137 wm_low ? "lock_reclaim_threshold_mb" : "lock_limit_mb");
140 watermark = value >> 20;
143 if (ldlm_lock_limit_mb != 0 && watermark > ldlm_lock_limit_mb) {
144 CERROR("lock_reclaim_threshold_mb must be smaller than "
150 if (watermark != 0) {
152 do_div(watermark, sizeof(struct ldlm_lock));
154 ldlm_reclaim_threshold = watermark;
156 if (ldlm_reclaim_threshold_mb != 0 &&
157 watermark < ldlm_reclaim_threshold_mb) {
158 CERROR("lock_limit_mb must be greater than "
159 "lock_reclaim_threshold_mb.\n");
164 if (watermark != 0) {
166 do_div(watermark, sizeof(struct ldlm_lock));
168 ldlm_lock_limit = watermark;
174 static int seq_watermark_open(struct inode *inode, struct file *file)
176 return single_open(file, seq_watermark_show, inode->i_private);
179 static const struct file_operations ldlm_watermark_fops = {
180 .owner = THIS_MODULE,
181 .open = seq_watermark_open,
183 .write = seq_watermark_write,
185 .release = lprocfs_single_release,
188 static int seq_granted_show(struct seq_file *m, void *data)
190 seq_printf(m, "%llu\n", percpu_counter_sum_positive(
191 (struct percpu_counter *)m->private));
195 static int seq_granted_open(struct inode *inode, struct file *file)
197 return single_open(file, seq_granted_show, inode->i_private);
200 static const struct file_operations ldlm_granted_fops = {
201 .owner = THIS_MODULE,
202 .open = seq_granted_open,
205 .release = seq_release,
208 #endif /* HAVE_SERVER_SUPPORT */
210 static struct ldebugfs_vars ldlm_debugfs_list[] = {
211 { .name = "dump_namespaces",
212 .fops = &ldlm_dump_ns_fops,
214 { .name = "dump_granted_max",
215 .fops = &ldlm_rw_uint_fops,
216 .data = &ldlm_dump_granted_max },
217 #ifdef HAVE_SERVER_SUPPORT
218 { .name = "lock_reclaim_threshold_mb",
219 .fops = &ldlm_watermark_fops,
220 .data = &ldlm_reclaim_threshold_mb },
221 { .name = "lock_limit_mb",
222 .fops = &ldlm_watermark_fops,
223 .data = &ldlm_lock_limit_mb },
224 { .name = "lock_granted_count",
225 .fops = &ldlm_granted_fops,
226 .data = &ldlm_granted_total },
231 int ldlm_debugfs_setup(void)
234 ldlm_debugfs_dir = debugfs_create_dir(OBD_LDLM_DEVICENAME,
235 debugfs_lustre_root);
236 ldlm_ns_debugfs_dir = debugfs_create_dir("namespaces",
238 ldlm_svc_debugfs_dir = debugfs_create_dir("services",
241 ldebugfs_add_vars(ldlm_debugfs_dir, ldlm_debugfs_list, NULL);
246 void ldlm_debugfs_cleanup(void)
248 debugfs_remove_recursive(ldlm_debugfs_dir);
250 ldlm_svc_debugfs_dir = NULL;
251 ldlm_ns_debugfs_dir = NULL;
252 ldlm_debugfs_dir = NULL;
255 static ssize_t resource_count_show(struct kobject *kobj, struct attribute *attr,
258 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
263 /* result is not strictly consistant */
264 for (i = 0; i < (1 << ns->ns_bucket_bits); i++)
265 res += atomic_read(&ns->ns_rs_buckets[i].nsb_count);
266 return sprintf(buf, "%lld\n", res);
268 LUSTRE_RO_ATTR(resource_count);
270 static ssize_t lock_count_show(struct kobject *kobj, struct attribute *attr,
273 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
277 locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
278 LPROCFS_FIELDS_FLAGS_SUM);
279 return sprintf(buf, "%lld\n", locks);
281 LUSTRE_RO_ATTR(lock_count);
283 static ssize_t lock_unused_count_show(struct kobject *kobj,
284 struct attribute *attr,
287 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
290 return sprintf(buf, "%d\n", ns->ns_nr_unused);
292 LUSTRE_RO_ATTR(lock_unused_count);
294 static ssize_t lru_size_show(struct kobject *kobj, struct attribute *attr,
297 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
299 __u32 *nr = &ns->ns_max_unused;
301 if (ns_connect_lru_resize(ns))
302 nr = &ns->ns_nr_unused;
303 return sprintf(buf, "%u\n", *nr);
306 static ssize_t lru_size_store(struct kobject *kobj, struct attribute *attr,
307 const char *buffer, size_t count)
309 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
315 if (strncmp(buffer, "clear", 5) == 0) {
317 "dropping all unused locks from namespace %s\n",
319 /* Try to cancel all @ns_nr_unused locks. */
320 ldlm_cancel_lru(ns, INT_MAX, 0, LDLM_LRU_FLAG_CLEANUP);
324 err = kstrtoul(buffer, 10, &tmp);
326 CERROR("lru_size: invalid value written\n");
329 lru_resize = (tmp == 0);
331 if (ns_connect_lru_resize(ns)) {
333 ns->ns_max_unused = (unsigned int)tmp;
335 if (tmp > ns->ns_nr_unused)
336 tmp = ns->ns_nr_unused;
337 tmp = ns->ns_nr_unused - tmp;
340 "changing namespace %s unused locks from %u to %u\n",
341 ldlm_ns_name(ns), ns->ns_nr_unused,
346 "disable lru_resize for namespace %s\n",
348 ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
350 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, 0);
353 "changing namespace %s max_unused from %u to %u\n",
354 ldlm_ns_name(ns), ns->ns_max_unused,
357 /* Make sure that LRU resize was originally supported before
358 * turning it on here.
361 (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
363 "enable lru_resize for namespace %s\n",
365 ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
367 ns->ns_max_unused = (unsigned int)tmp;
368 ldlm_cancel_lru(ns, 0, LCF_ASYNC, 0);
373 LUSTRE_RW_ATTR(lru_size);
375 static ssize_t lru_cancel_batch_show(struct kobject *kobj,
376 struct attribute *attr, char *buf)
378 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
381 return snprintf(buf, sizeof(buf) - 1, "%u\n", ns->ns_cancel_batch);
384 static ssize_t lru_cancel_batch_store(struct kobject *kobj,
385 struct attribute *attr,
386 const char *buffer, size_t count)
388 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
392 if (kstrtoul(buffer, 10, &tmp))
395 ns->ns_cancel_batch = (unsigned int)tmp;
399 LUSTRE_RW_ATTR(lru_cancel_batch);
401 static ssize_t ns_recalc_pct_show(struct kobject *kobj,
402 struct attribute *attr, char *buf)
404 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
407 return snprintf(buf, sizeof(buf) - 1, "%u\n", ns->ns_recalc_pct);
410 static ssize_t ns_recalc_pct_store(struct kobject *kobj,
411 struct attribute *attr,
412 const char *buffer, size_t count)
414 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
418 if (kstrtoul(buffer, 10, &tmp))
424 ns->ns_recalc_pct = (unsigned int)tmp;
428 LUSTRE_RW_ATTR(ns_recalc_pct);
430 static ssize_t lru_max_age_show(struct kobject *kobj, struct attribute *attr,
433 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
436 return sprintf(buf, "%lld\n", ktime_to_ms(ns->ns_max_age));
439 static ssize_t lru_max_age_store(struct kobject *kobj, struct attribute *attr,
440 const char *buffer, size_t count)
442 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
444 int scale = NSEC_PER_MSEC;
445 unsigned long long tmp;
448 /* Did the user ask in seconds or milliseconds. Default is in ms */
449 buf = strstr(buffer, "ms");
451 buf = strchr(buffer, 's');
453 scale = NSEC_PER_SEC;
459 if (kstrtoull(buffer, 10, &tmp))
462 ns->ns_max_age = ktime_set(0, tmp * scale);
466 LUSTRE_RW_ATTR(lru_max_age);
468 static ssize_t early_lock_cancel_show(struct kobject *kobj,
469 struct attribute *attr,
472 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
475 return sprintf(buf, "%d\n", ns_connect_cancelset(ns));
478 static ssize_t early_lock_cancel_store(struct kobject *kobj,
479 struct attribute *attr,
483 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
485 unsigned long supp = -1;
488 rc = kstrtoul(buffer, 10, &supp);
493 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
494 else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
495 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
498 LUSTRE_RW_ATTR(early_lock_cancel);
500 static ssize_t dirty_age_limit_show(struct kobject *kobj,
501 struct attribute *attr, char *buf)
503 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
506 return scnprintf(buf, PAGE_SIZE, "%llu\n",
507 ktime_divns(ns->ns_dirty_age_limit, NSEC_PER_SEC));
510 static ssize_t dirty_age_limit_store(struct kobject *kobj,
511 struct attribute *attr,
512 const char *buffer, size_t count)
514 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
516 unsigned long long tmp;
518 if (kstrtoull(buffer, 10, &tmp))
521 ns->ns_dirty_age_limit = ktime_set(tmp, 0);
525 LUSTRE_RW_ATTR(dirty_age_limit);
527 #ifdef HAVE_SERVER_SUPPORT
528 static ssize_t ctime_age_limit_show(struct kobject *kobj,
529 struct attribute *attr, char *buf)
531 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
534 return scnprintf(buf, PAGE_SIZE, "%u\n", ns->ns_ctime_age_limit);
537 static ssize_t ctime_age_limit_store(struct kobject *kobj,
538 struct attribute *attr,
539 const char *buffer, size_t count)
541 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
545 if (kstrtoul(buffer, 10, &tmp))
548 ns->ns_ctime_age_limit = tmp;
552 LUSTRE_RW_ATTR(ctime_age_limit);
554 static ssize_t lock_timeouts_show(struct kobject *kobj, struct attribute *attr,
557 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
560 return sprintf(buf, "%d\n", ns->ns_timeouts);
562 LUSTRE_RO_ATTR(lock_timeouts);
564 static ssize_t max_nolock_bytes_show(struct kobject *kobj,
565 struct attribute *attr, char *buf)
567 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
570 return sprintf(buf, "%u\n", ns->ns_max_nolock_size);
573 static ssize_t max_nolock_bytes_store(struct kobject *kobj,
574 struct attribute *attr,
575 const char *buffer, size_t count)
577 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
582 err = kstrtoul(buffer, 10, &tmp);
586 ns->ns_max_nolock_size = tmp;
590 LUSTRE_RW_ATTR(max_nolock_bytes);
592 static ssize_t contention_seconds_show(struct kobject *kobj,
593 struct attribute *attr, char *buf)
595 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
598 return scnprintf(buf, PAGE_SIZE, "%d\n", ns->ns_contention_time);
601 static ssize_t contention_seconds_store(struct kobject *kobj,
602 struct attribute *attr,
603 const char *buffer, size_t count)
605 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
609 if (kstrtouint(buffer, 10, &tmp))
612 ns->ns_contention_time = tmp;
616 LUSTRE_RW_ATTR(contention_seconds);
618 static ssize_t contended_locks_show(struct kobject *kobj,
619 struct attribute *attr, char *buf)
621 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
624 return sprintf(buf, "%u\n", ns->ns_contended_locks);
627 static ssize_t contended_locks_store(struct kobject *kobj,
628 struct attribute *attr,
629 const char *buffer, size_t count)
631 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
636 err = kstrtoul(buffer, 10, &tmp);
640 ns->ns_contended_locks = tmp;
644 LUSTRE_RW_ATTR(contended_locks);
646 static ssize_t max_parallel_ast_show(struct kobject *kobj,
647 struct attribute *attr, char *buf)
649 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
652 return sprintf(buf, "%u\n", ns->ns_max_parallel_ast);
655 static ssize_t max_parallel_ast_store(struct kobject *kobj,
656 struct attribute *attr,
657 const char *buffer, size_t count)
659 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
664 err = kstrtoul(buffer, 10, &tmp);
668 ns->ns_max_parallel_ast = tmp;
672 LUSTRE_RW_ATTR(max_parallel_ast);
674 #endif /* HAVE_SERVER_SUPPORT */
676 /* These are for namespaces in /sys/fs/lustre/ldlm/namespaces/ */
677 static struct attribute *ldlm_ns_attrs[] = {
678 &lustre_attr_resource_count.attr,
679 &lustre_attr_lock_count.attr,
680 &lustre_attr_lock_unused_count.attr,
681 &lustre_attr_ns_recalc_pct.attr,
682 &lustre_attr_lru_size.attr,
683 &lustre_attr_lru_cancel_batch.attr,
684 &lustre_attr_lru_max_age.attr,
685 &lustre_attr_early_lock_cancel.attr,
686 &lustre_attr_dirty_age_limit.attr,
687 #ifdef HAVE_SERVER_SUPPORT
688 &lustre_attr_ctime_age_limit.attr,
689 &lustre_attr_lock_timeouts.attr,
690 &lustre_attr_max_nolock_bytes.attr,
691 &lustre_attr_contention_seconds.attr,
692 &lustre_attr_contended_locks.attr,
693 &lustre_attr_max_parallel_ast.attr,
698 static void ldlm_ns_release(struct kobject *kobj)
700 struct ldlm_namespace *ns = container_of(kobj, struct ldlm_namespace,
702 complete(&ns->ns_kobj_unregister);
705 KOBJ_ATTRIBUTE_GROUPS(ldlm_ns);
707 static struct kobj_type ldlm_ns_ktype = {
708 .default_groups = KOBJ_ATTR_GROUPS(ldlm_ns),
709 .sysfs_ops = &lustre_sysfs_ops,
710 .release = ldlm_ns_release,
713 static void ldlm_namespace_debugfs_unregister(struct ldlm_namespace *ns)
715 if (IS_ERR_OR_NULL(ns->ns_debugfs_entry))
716 CERROR("dlm namespace %s has no procfs dir?\n",
719 debugfs_remove_recursive(ns->ns_debugfs_entry);
721 if (ns->ns_stats != NULL)
722 lprocfs_stats_free(&ns->ns_stats);
725 void ldlm_namespace_sysfs_unregister(struct ldlm_namespace *ns)
727 kobject_put(&ns->ns_kobj);
728 wait_for_completion(&ns->ns_kobj_unregister);
731 int ldlm_namespace_sysfs_register(struct ldlm_namespace *ns)
735 ns->ns_kobj.kset = ldlm_ns_kset;
736 init_completion(&ns->ns_kobj_unregister);
737 err = kobject_init_and_add(&ns->ns_kobj, &ldlm_ns_ktype, NULL,
738 "%s", ldlm_ns_name(ns));
740 ns->ns_stats = lprocfs_stats_alloc(LDLM_NSS_LAST, 0);
742 kobject_put(&ns->ns_kobj);
746 lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
747 LPROCFS_CNTR_AVGMINMAX | LPROCFS_TYPE_LOCKS,
753 static int ldlm_namespace_debugfs_register(struct ldlm_namespace *ns)
755 struct dentry *ns_entry;
757 if (!IS_ERR_OR_NULL(ns->ns_debugfs_entry)) {
758 ns_entry = ns->ns_debugfs_entry;
760 ns_entry = debugfs_create_dir(ldlm_ns_name(ns),
761 ldlm_ns_debugfs_dir);
764 ns->ns_debugfs_entry = ns_entry;
769 #undef MAX_STRING_SIZE
771 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
772 const void *key, unsigned int mask)
774 const struct ldlm_res_id *id = key;
775 unsigned int val = 0;
778 for (i = 0; i < RES_NAME_SIZE; i++)
783 static unsigned int ldlm_res_hop_fid_hash(const struct ldlm_res_id *id, unsigned int bits)
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);
793 hash = fid_flatten32(&fid);
794 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
795 if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
796 val = id->name[LUSTRE_RES_ID_HSH_OFF];
800 hash += (val >> 5) + (val << 11);
801 return cfs_hash_32(hash, bits);
804 static void *ldlm_res_hop_key(struct hlist_node *hnode)
806 struct ldlm_resource *res;
808 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
809 return &res->lr_name;
812 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
814 struct ldlm_resource *res;
816 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
817 return ldlm_res_eq((const struct ldlm_res_id *)key,
818 (const struct ldlm_res_id *)&res->lr_name);
821 static void *ldlm_res_hop_object(struct hlist_node *hnode)
823 return hlist_entry(hnode, struct ldlm_resource, lr_hash);
827 ldlm_res_hop_get_locked(struct cfs_hash *hs, struct hlist_node *hnode)
829 struct ldlm_resource *res;
831 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
832 ldlm_resource_getref(res);
835 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
837 struct ldlm_resource *res;
839 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
840 ldlm_resource_putref(res);
843 static struct cfs_hash_ops ldlm_ns_hash_ops = {
844 .hs_hash = ldlm_res_hop_hash,
845 .hs_key = ldlm_res_hop_key,
846 .hs_keycmp = ldlm_res_hop_keycmp,
848 .hs_object = ldlm_res_hop_object,
849 .hs_get = ldlm_res_hop_get_locked,
850 .hs_put = ldlm_res_hop_put
854 /** hash bucket bits */
855 unsigned nsd_bkt_bits;
857 unsigned nsd_all_bits;
858 } ldlm_ns_hash_defs[] = {
859 [LDLM_NS_TYPE_MDC] = {
863 [LDLM_NS_TYPE_MDT] = {
867 [LDLM_NS_TYPE_OSC] = {
871 [LDLM_NS_TYPE_OST] = {
875 [LDLM_NS_TYPE_MGC] = {
879 [LDLM_NS_TYPE_MGT] = {
886 * Create and initialize new empty namespace.
888 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
889 enum ldlm_side client,
890 enum ldlm_appetite apt,
891 enum ldlm_ns_type ns_type)
893 struct ldlm_namespace *ns = NULL;
898 LASSERT(obd != NULL);
902 CERROR("%s: ldlm_get_ref failed: rc = %d\n", name, rc);
906 if (ns_type >= ARRAY_SIZE(ldlm_ns_hash_defs) ||
907 ldlm_ns_hash_defs[ns_type].nsd_bkt_bits == 0) {
909 CERROR("%s: unknown namespace type %d: rc = %d\n",
916 GOTO(out_ref, rc = -ENOMEM);
918 ns->ns_rs_hash = cfs_hash_create(name,
919 ldlm_ns_hash_defs[ns_type].nsd_all_bits,
920 ldlm_ns_hash_defs[ns_type].nsd_all_bits,
921 ldlm_ns_hash_defs[ns_type].nsd_bkt_bits,
928 CFS_HASH_SPIN_BKTLOCK |
929 CFS_HASH_NO_ITEMREF);
931 GOTO(out_ns, rc = -ENOMEM);
933 ns->ns_bucket_bits = ldlm_ns_hash_defs[ns_type].nsd_all_bits -
934 ldlm_ns_hash_defs[ns_type].nsd_bkt_bits;
936 OBD_ALLOC_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
937 if (!ns->ns_rs_buckets)
938 GOTO(out_hash, rc = -ENOMEM);
940 for (idx = 0; idx < (1 << ns->ns_bucket_bits); idx++) {
941 struct ldlm_ns_bucket *nsb = &ns->ns_rs_buckets[idx];
943 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
944 nsb->nsb_namespace = ns;
945 nsb->nsb_reclaim_start = 0;
946 atomic_set(&nsb->nsb_count, 0);
950 ns->ns_appetite = apt;
951 ns->ns_client = client;
952 ns->ns_name = kstrdup(name, GFP_KERNEL);
954 GOTO(out_hash, rc = -ENOMEM);
956 INIT_LIST_HEAD(&ns->ns_list_chain);
957 INIT_LIST_HEAD(&ns->ns_unused_list);
958 spin_lock_init(&ns->ns_lock);
959 atomic_set(&ns->ns_bref, 0);
960 init_waitqueue_head(&ns->ns_waitq);
962 ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
963 ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
964 ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
966 ns->ns_max_parallel_ast = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
967 ns->ns_nr_unused = 0;
968 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
969 ns->ns_cancel_batch = LDLM_DEFAULT_LRU_SHRINK_BATCH;
970 ns->ns_recalc_pct = LDLM_DEFAULT_SLV_RECALC_PCT;
971 ns->ns_max_age = ktime_set(LDLM_DEFAULT_MAX_ALIVE, 0);
972 ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
973 ns->ns_dirty_age_limit = ktime_set(LDLM_DIRTY_AGE_LIMIT, 0);
975 ns->ns_orig_connect_flags = 0;
976 ns->ns_connect_flags = 0;
978 ns->ns_reclaim_start = 0;
979 ns->ns_last_pos = &ns->ns_unused_list;
982 rc = ldlm_namespace_sysfs_register(ns);
984 CERROR("%s: cannot initialize ns sysfs: rc = %d\n", name, rc);
988 rc = ldlm_namespace_debugfs_register(ns);
990 CERROR("%s: cannot initialize ns proc: rc = %d\n", name, rc);
994 idx = ldlm_namespace_nr_read(client);
995 rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
997 CERROR("%s: cannot initialize lock pool, rc = %d\n", name, rc);
1001 ldlm_namespace_register(ns, client);
1004 ldlm_namespace_debugfs_unregister(ns);
1006 ldlm_namespace_sysfs_unregister(ns);
1007 ldlm_namespace_cleanup(ns, 0);
1009 OBD_FREE_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
1011 cfs_hash_putref(ns->ns_rs_hash);
1016 RETURN(ERR_PTR(rc));
1018 EXPORT_SYMBOL(ldlm_namespace_new);
1021 * Cancel and destroy all locks on a resource.
1023 * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
1024 * clean up. This is currently only used for recovery, and we make
1025 * certain assumptions as a result--notably, that we shouldn't cancel
1028 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
1031 struct list_head *tmp;
1032 int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
1033 bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
1036 struct ldlm_lock *lock = NULL;
1038 /* First, we look for non-cleaned-yet lock
1039 * all cleaned locks are marked by CLEANED flag. */
1041 list_for_each(tmp, q) {
1042 lock = list_entry(tmp, struct ldlm_lock,
1044 if (ldlm_is_cleaned(lock)) {
1048 LDLM_LOCK_GET(lock);
1049 ldlm_set_cleaned(lock);
1058 /* Set CBPENDING so nothing in the cancellation path
1059 * can match this lock. */
1060 ldlm_set_cbpending(lock);
1061 ldlm_set_failed(lock);
1062 lock->l_flags |= flags;
1064 /* ... without sending a CANCEL message for local_only. */
1066 ldlm_set_local_only(lock);
1068 if (local_only && (lock->l_readers || lock->l_writers)) {
1070 * This is a little bit gross, but much better than the
1071 * alternative: pretend that we got a blocking AST from
1072 * the server, so that when the lock is decref'd, it
1076 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
1077 if (lock->l_flags & LDLM_FL_FAIL_LOC)
1078 schedule_timeout_uninterruptible(
1079 cfs_time_seconds(4));
1081 if (lock->l_completion_ast)
1082 lock->l_completion_ast(lock,
1083 LDLM_FL_FAILED, NULL);
1084 LDLM_LOCK_RELEASE(lock);
1089 struct lustre_handle lockh;
1092 ldlm_lock2handle(lock, &lockh);
1093 rc = ldlm_cli_cancel(&lockh, LCF_LOCAL);
1095 CERROR("ldlm_cli_cancel: %d\n", rc);
1099 "Freeing a lock still held by a client node");
1100 ldlm_lock_cancel(lock);
1102 LDLM_LOCK_RELEASE(lock);
1106 static int ldlm_resource_clean(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1107 struct hlist_node *hnode, void *arg)
1109 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1110 __u64 flags = *(__u64 *)arg;
1112 cleanup_resource(res, &res->lr_granted, flags);
1113 cleanup_resource(res, &res->lr_waiting, flags);
1118 static int ldlm_resource_complain(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1119 struct hlist_node *hnode, void *arg)
1121 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1124 CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
1125 "(%d) after lock cleanup; forcing cleanup.\n",
1126 ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
1127 atomic_read(&res->lr_refcount) - 1);
1129 /* Use D_NETERROR since it is in the default mask */
1130 ldlm_resource_dump(D_NETERROR, res);
1136 * Cancel and destroy all locks in the namespace.
1138 * Typically used during evictions when server notified client that it was
1139 * evicted and all of its state needs to be destroyed.
1140 * Also used during shutdown.
1142 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
1145 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
1149 cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
1151 cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain,
1155 EXPORT_SYMBOL(ldlm_namespace_cleanup);
1158 * Attempts to free namespace.
1160 * Only used when namespace goes away, like during an unmount.
1162 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
1166 /* At shutdown time, don't call the cancellation callback */
1167 ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
1169 if (atomic_read(&ns->ns_bref) > 0) {
1172 "dlm namespace %s free waiting on refcount %d\n",
1173 ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
1176 rc = wait_event_idle_timeout(
1178 atomic_read(&ns->ns_bref) == 0,
1179 cfs_time_seconds(1) / 4);
1181 rc = l_wait_event_abortable(
1182 ns->ns_waitq, atomic_read(&ns->ns_bref) == 0);
1184 /* Forced cleanups should be able to reclaim all references,
1185 * so it's safe to wait forever... we can't leak locks... */
1186 if (force && rc == 0) {
1188 LCONSOLE_ERROR("Forced cleanup waiting for %s "
1189 "namespace with %d resources in use, "
1190 "(rc=%d)\n", ldlm_ns_name(ns),
1191 atomic_read(&ns->ns_bref), rc);
1192 GOTO(force_wait, rc);
1195 if (atomic_read(&ns->ns_bref)) {
1196 LCONSOLE_ERROR("Cleanup waiting for %s namespace "
1197 "with %d resources in use, (rc=%d)\n",
1199 atomic_read(&ns->ns_bref), rc);
1200 RETURN(ELDLM_NAMESPACE_EXISTS);
1202 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
1210 * Performs various cleanups for passed \a ns to make it drop refc and be
1211 * ready for freeing. Waits for refc == 0.
1213 * The following is done:
1214 * (0) Unregister \a ns from its list to make inaccessible for potential
1215 * users like pools thread and others;
1216 * (1) Clear all locks in \a ns.
1218 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
1219 struct obd_import *imp,
1230 spin_lock(&ns->ns_lock);
1231 ns->ns_stopping = 1;
1232 spin_unlock(&ns->ns_lock);
1235 * Can fail with -EINTR when force == 0 in which case try harder.
1237 rc = __ldlm_namespace_free(ns, force);
1238 if (rc != ELDLM_OK) {
1240 ptlrpc_disconnect_import(imp, 0);
1241 ptlrpc_invalidate_import(imp);
1245 * With all requests dropped and the import inactive
1246 * we are gaurenteed all reference will be dropped.
1248 rc = __ldlm_namespace_free(ns, 1);
1253 EXPORT_SYMBOL(ldlm_namespace_free_prior);
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.
1260 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
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 ldlm_pool_fini(&ns->ns_pool);
1275 ldlm_namespace_debugfs_unregister(ns);
1276 ldlm_namespace_sysfs_unregister(ns);
1277 cfs_hash_putref(ns->ns_rs_hash);
1278 OBD_FREE_PTR_ARRAY_LARGE(ns->ns_rs_buckets, 1 << ns->ns_bucket_bits);
1280 /* Namespace \a ns should be not on list at this time, otherwise
1281 * this will cause issues related to using freed \a ns in poold
1284 LASSERT(list_empty(&ns->ns_list_chain));
1289 EXPORT_SYMBOL(ldlm_namespace_free_post);
1292 * Cleanup the resource, and free namespace.
1295 * proc1: destroy import
1296 * class_disconnect_export(grab cl_sem) ->
1297 * -> ldlm_namespace_free ->
1298 * -> lprocfs_remove(grab _lprocfs_lock).
1299 * proc2: read proc info
1300 * lprocfs_fops_read(grab _lprocfs_lock) ->
1301 * -> osc_rd_active, etc(grab cl_sem).
1303 * So that I have to split the ldlm_namespace_free into two parts - the first
1304 * part ldlm_namespace_free_prior is used to cleanup the resource which is
1305 * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
1306 * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
1309 void ldlm_namespace_free(struct ldlm_namespace *ns,
1310 struct obd_import *imp,
1313 ldlm_namespace_free_prior(ns, imp, force);
1314 ldlm_namespace_free_post(ns);
1316 EXPORT_SYMBOL(ldlm_namespace_free);
1318 void ldlm_namespace_get(struct ldlm_namespace *ns)
1320 atomic_inc(&ns->ns_bref);
1323 /* This is only for callers that care about refcount */
1324 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
1326 return atomic_inc_return(&ns->ns_bref);
1329 void ldlm_namespace_put(struct ldlm_namespace *ns)
1331 if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
1332 wake_up(&ns->ns_waitq);
1333 spin_unlock(&ns->ns_lock);
1337 /** Register \a ns in the list of namespaces */
1338 void ldlm_namespace_register(struct ldlm_namespace *ns, enum ldlm_side client)
1340 mutex_lock(ldlm_namespace_lock(client));
1341 LASSERT(list_empty(&ns->ns_list_chain));
1342 list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
1343 ldlm_namespace_nr_inc(client);
1344 mutex_unlock(ldlm_namespace_lock(client));
1347 /** Unregister \a ns from the list of namespaces. */
1348 void ldlm_namespace_unregister(struct ldlm_namespace *ns, enum ldlm_side client)
1350 mutex_lock(ldlm_namespace_lock(client));
1351 LASSERT(!list_empty(&ns->ns_list_chain));
1352 /* Some asserts and possibly other parts of the code are still
1353 * using list_empty(&ns->ns_list_chain). This is why it is
1354 * important to use list_del_init() here. */
1355 list_del_init(&ns->ns_list_chain);
1356 ldlm_namespace_nr_dec(client);
1357 mutex_unlock(ldlm_namespace_lock(client));
1360 /** Should be called with ldlm_namespace_lock(client) taken. */
1361 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1362 enum ldlm_side client)
1364 LASSERT(!list_empty(&ns->ns_list_chain));
1365 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1366 list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1369 /** Should be called with ldlm_namespace_lock(client) taken. */
1370 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1371 enum ldlm_side client)
1373 LASSERT(!list_empty(&ns->ns_list_chain));
1374 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1375 list_move_tail(&ns->ns_list_chain,
1376 ldlm_namespace_inactive_list(client));
1379 /** Should be called with ldlm_namespace_lock(client) taken. */
1380 struct ldlm_namespace *ldlm_namespace_first_locked(enum ldlm_side client)
1382 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1383 LASSERT(!list_empty(ldlm_namespace_list(client)));
1384 return container_of(ldlm_namespace_list(client)->next,
1385 struct ldlm_namespace, ns_list_chain);
1388 static bool ldlm_resource_extent_new(struct ldlm_resource *res)
1392 OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1393 sizeof(*res->lr_itree) * LCK_MODE_NUM);
1394 if (res->lr_itree == NULL)
1396 /* Initialize interval trees for each lock mode. */
1397 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1398 res->lr_itree[idx].lit_size = 0;
1399 res->lr_itree[idx].lit_mode = BIT(idx);
1400 res->lr_itree[idx].lit_root = NULL;
1405 static bool ldlm_resource_inodebits_new(struct ldlm_resource *res)
1409 OBD_ALLOC_PTR(res->lr_ibits_queues);
1410 if (res->lr_ibits_queues == NULL)
1412 for (i = 0; i < MDS_INODELOCK_NUMBITS; i++)
1413 INIT_LIST_HEAD(&res->lr_ibits_queues->liq_waiting[i]);
1417 static bool ldlm_resource_flock_new(struct ldlm_resource *res)
1419 res->lr_flock_node.lfn_needs_reprocess = false;
1420 atomic_set(&res->lr_flock_node.lfn_unlock_pending, 0);
1425 /** Create and initialize new resource. */
1426 static struct ldlm_resource *ldlm_resource_new(enum ldlm_type ldlm_type)
1428 struct ldlm_resource *res;
1431 OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1435 switch (ldlm_type) {
1437 rc = ldlm_resource_extent_new(res);
1440 rc = ldlm_resource_inodebits_new(res);
1443 rc = ldlm_resource_flock_new(res);
1450 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1454 INIT_LIST_HEAD(&res->lr_granted);
1455 INIT_LIST_HEAD(&res->lr_waiting);
1457 atomic_set(&res->lr_refcount, 1);
1458 spin_lock_init(&res->lr_lock);
1459 lu_ref_init(&res->lr_reference);
1461 /* Since LVB init can be delayed now, there is no longer need to
1462 * immediatelly acquire mutex here. */
1463 mutex_init(&res->lr_lvb_mutex);
1464 res->lr_lvb_initialized = false;
1469 static void __ldlm_resource_free(struct rcu_head *head)
1471 struct ldlm_resource *res = container_of(head, struct ldlm_resource,
1474 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1477 static void ldlm_resource_free(struct ldlm_resource *res)
1479 if (res->lr_type == LDLM_EXTENT) {
1480 if (res->lr_itree != NULL)
1481 OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1482 sizeof(*res->lr_itree) * LCK_MODE_NUM);
1483 } else if (res->lr_type == LDLM_IBITS) {
1484 if (res->lr_ibits_queues != NULL)
1485 OBD_FREE_PTR(res->lr_ibits_queues);
1488 call_rcu(&res->lr_rcu, __ldlm_resource_free);
1492 * Return a reference to resource with given name, creating it if necessary.
1493 * Args: namespace with ns_lock unlocked
1494 * Locks: takes and releases NS hash-lock and res->lr_lock
1495 * Returns: referenced, unlocked ldlm_resource or ERR_PTR
1497 struct ldlm_resource *
1498 ldlm_resource_get(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
1499 enum ldlm_type type, int create)
1501 struct hlist_node *hnode;
1502 struct ldlm_resource *res = NULL;
1503 struct cfs_hash_bd bd;
1505 int ns_refcount = 0;
1508 LASSERT(ns != NULL);
1509 LASSERT(ns->ns_rs_hash != NULL);
1510 LASSERT(name->name[0] != 0);
1512 cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1513 hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1514 if (hnode != NULL) {
1515 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1519 version = cfs_hash_bd_version_get(&bd);
1520 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1523 return ERR_PTR(-ENOENT);
1525 LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1526 "type: %d\n", type);
1527 res = ldlm_resource_new(type);
1529 return ERR_PTR(-ENOMEM);
1531 hash = ldlm_res_hop_fid_hash(name, ns->ns_bucket_bits);
1532 res->lr_ns_bucket = &ns->ns_rs_buckets[hash];
1533 res->lr_name = *name;
1534 res->lr_type = type;
1536 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1537 hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1538 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1540 if (hnode != NULL) {
1541 /* Someone won the race and already added the resource. */
1542 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1543 /* Clean lu_ref for failed resource. */
1544 lu_ref_fini(&res->lr_reference);
1545 ldlm_resource_free(res);
1547 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1550 /* We won! Let's add the resource. */
1551 cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1552 if (atomic_inc_return(&res->lr_ns_bucket->nsb_count) == 1)
1553 ns_refcount = ldlm_namespace_get_return(ns);
1555 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1557 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1559 /* Let's see if we happened to be the very first resource in this
1560 * namespace. If so, and this is a client namespace, we need to move
1561 * the namespace into the active namespaces list to be patrolled by
1562 * the ldlm_poold. */
1563 if (ns_is_client(ns) && ns_refcount == 1) {
1564 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1565 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1566 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1571 EXPORT_SYMBOL(ldlm_resource_get);
1573 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1575 LASSERT(res != NULL);
1576 LASSERT(res != LP_POISON);
1577 atomic_inc(&res->lr_refcount);
1578 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1579 atomic_read(&res->lr_refcount));
1583 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1584 struct ldlm_resource *res)
1586 struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1588 if (!list_empty(&res->lr_granted)) {
1589 ldlm_resource_dump(D_ERROR, res);
1593 if (!list_empty(&res->lr_waiting)) {
1594 ldlm_resource_dump(D_ERROR, res);
1598 cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1600 lu_ref_fini(&res->lr_reference);
1601 if (atomic_dec_and_test(&nsb->nsb_count))
1602 ldlm_namespace_put(nsb->nsb_namespace);
1605 /* Returns 1 if the resource was freed, 0 if it remains. */
1606 int ldlm_resource_putref(struct ldlm_resource *res)
1608 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1609 struct cfs_hash_bd bd;
1611 LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1612 CDEBUG(D_INFO, "putref res: %p count: %d\n",
1613 res, atomic_read(&res->lr_refcount) - 1);
1615 cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1616 if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1617 __ldlm_resource_putref_final(&bd, res);
1618 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1619 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1620 ns->ns_lvbo->lvbo_free(res);
1621 ldlm_resource_free(res);
1626 EXPORT_SYMBOL(ldlm_resource_putref);
1628 static void __ldlm_resource_add_lock(struct ldlm_resource *res,
1629 struct list_head *head,
1630 struct ldlm_lock *lock,
1633 check_res_locked(res);
1635 if (ldlm_is_destroyed(lock)) {
1636 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1640 LASSERT(list_empty(&lock->l_res_link));
1643 list_add_tail(&lock->l_res_link, head);
1645 list_add(&lock->l_res_link, head);
1647 if (res->lr_type == LDLM_IBITS)
1648 ldlm_inodebits_add_lock(res, head, lock, tail);
1650 ldlm_resource_dump(D_INFO, res);
1654 * Add a lock into a given resource into specified lock list.
1656 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1657 struct ldlm_lock *lock)
1659 LDLM_DEBUG(lock, "About to add this lock");
1661 __ldlm_resource_add_lock(res, head, lock, true);
1665 * Insert a lock into resource after specified lock.
1667 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1668 struct ldlm_lock *new)
1670 LASSERT(!list_empty(&original->l_res_link));
1672 LDLM_DEBUG(new, "About to insert this lock after %p: ", original);
1673 __ldlm_resource_add_lock(original->l_resource,
1674 &original->l_res_link,
1679 * Insert a lock into resource before the specified lock.
1681 * IBITS waiting locks are to be inserted to the ibit lists as well, and only
1682 * the insert-after operation is supported for them, because the set of bits
1683 * of the previous and the new locks must match. Therefore, get the previous
1684 * lock and insert after.
1686 void ldlm_resource_insert_lock_before(struct ldlm_lock *original,
1687 struct ldlm_lock *new)
1689 LASSERT(!list_empty(&original->l_res_link));
1691 LDLM_DEBUG(new, "About to insert this lock before %p: ", original);
1692 __ldlm_resource_add_lock(original->l_resource,
1693 original->l_res_link.prev, new, false);
1696 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1698 int type = lock->l_resource->lr_type;
1700 check_res_locked(lock->l_resource);
1703 ldlm_unlink_lock_skiplist(lock);
1706 ldlm_extent_unlink_lock(lock);
1709 ldlm_inodebits_unlink_lock(lock);
1712 list_del_init(&lock->l_res_link);
1714 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1716 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1718 desc->lr_type = res->lr_type;
1719 desc->lr_name = res->lr_name;
1723 * Print information about all locks in all namespaces on this node to debug
1726 void ldlm_dump_all_namespaces(enum ldlm_side client, int level)
1728 struct list_head *tmp;
1730 if (!((libcfs_debug | D_ERROR) & level))
1733 mutex_lock(ldlm_namespace_lock(client));
1735 list_for_each(tmp, ldlm_namespace_list(client)) {
1736 struct ldlm_namespace *ns;
1738 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1739 ldlm_namespace_dump(level, ns);
1742 mutex_unlock(ldlm_namespace_lock(client));
1745 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1746 struct hlist_node *hnode, void *arg)
1748 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1749 int level = (int)(unsigned long)arg;
1752 ldlm_resource_dump(level, res);
1759 * Print information about all locks in this namespace on this node to debug
1762 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1764 if (!((libcfs_debug | D_ERROR) & level))
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");
1771 if (ktime_get_seconds() < ns->ns_next_dump)
1774 cfs_hash_for_each_nolock(ns->ns_rs_hash,
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);
1783 * Print information about all locks in this resource to debug log.
1785 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1787 struct ldlm_lock *lock;
1788 unsigned int granted = 0;
1790 BUILD_BUG_ON(RES_NAME_SIZE != 4);
1792 if (!((libcfs_debug | D_ERROR) & level))
1795 CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1796 PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1798 if (!list_empty(&res->lr_granted)) {
1799 CDEBUG(level, "Granted locks (in reverse order):\n");
1800 list_for_each_entry_reverse(lock, &res->lr_granted,
1802 LDLM_DEBUG_LIMIT(level, lock, "###");
1803 if (!(level & D_CANTMASK) &&
1804 ++granted > ldlm_dump_granted_max) {
1806 "only dump %d granted locks to avoid DDOS.\n",
1813 if (!list_empty(&res->lr_waiting)) {
1814 CDEBUG(level, "Waiting locks:\n");
1815 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1816 LDLM_DEBUG_LIMIT(level, lock, "###");
1819 EXPORT_SYMBOL(ldlm_resource_dump);