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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ldlm/ldlm_resource.c
38 * Author: Phil Schwan <phil@clusterfs.com>
39 * Author: Peter Braam <braam@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_LDLM
43 # include <lustre_dlm.h>
45 #include <lustre_fid.h>
46 #include <obd_class.h>
47 #include "ldlm_internal.h"
49 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
51 atomic_t ldlm_srv_namespace_nr = ATOMIC_INIT(0);
52 atomic_t ldlm_cli_namespace_nr = ATOMIC_INIT(0);
54 struct mutex ldlm_srv_namespace_lock;
55 LIST_HEAD(ldlm_srv_namespace_list);
57 struct mutex ldlm_cli_namespace_lock;
58 LIST_HEAD(ldlm_cli_namespace_list);
60 proc_dir_entry_t *ldlm_type_proc_dir = NULL;
61 proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
62 proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
64 extern unsigned int ldlm_cancel_unused_locks_before_replay;
66 /* during debug dump certain amount of granted locks for one resource to avoid
68 unsigned int ldlm_dump_granted_max = 256;
71 static ssize_t lprocfs_wr_dump_ns(struct file *file, const char *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);
78 LPROC_SEQ_FOPS_WR_ONLY(ldlm, dump_ns);
80 LPROC_SEQ_FOPS_RW_TYPE(ldlm_rw, uint);
81 LPROC_SEQ_FOPS_RO_TYPE(ldlm, uint);
83 int ldlm_proc_setup(void)
86 struct lprocfs_vars list[] = {
87 { "dump_namespaces", &ldlm_dump_ns_fops, 0, 0222 },
88 { "dump_granted_max", &ldlm_rw_uint_fops,
89 &ldlm_dump_granted_max },
90 { "cancel_unused_locks_before_replay", &ldlm_rw_uint_fops,
91 &ldlm_cancel_unused_locks_before_replay },
94 LASSERT(ldlm_ns_proc_dir == NULL);
96 ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
99 if (IS_ERR(ldlm_type_proc_dir)) {
100 CERROR("LProcFS failed in ldlm-init\n");
101 rc = PTR_ERR(ldlm_type_proc_dir);
105 ldlm_ns_proc_dir = lprocfs_register("namespaces",
108 if (IS_ERR(ldlm_ns_proc_dir)) {
109 CERROR("LProcFS failed in ldlm-init\n");
110 rc = PTR_ERR(ldlm_ns_proc_dir);
114 ldlm_svc_proc_dir = lprocfs_register("services",
117 if (IS_ERR(ldlm_svc_proc_dir)) {
118 CERROR("LProcFS failed in ldlm-init\n");
119 rc = PTR_ERR(ldlm_svc_proc_dir);
123 rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
128 lprocfs_remove(&ldlm_ns_proc_dir);
130 lprocfs_remove(&ldlm_type_proc_dir);
132 ldlm_svc_proc_dir = NULL;
133 ldlm_type_proc_dir = NULL;
134 ldlm_ns_proc_dir = NULL;
138 void ldlm_proc_cleanup(void)
140 if (ldlm_svc_proc_dir)
141 lprocfs_remove(&ldlm_svc_proc_dir);
143 if (ldlm_ns_proc_dir)
144 lprocfs_remove(&ldlm_ns_proc_dir);
146 if (ldlm_type_proc_dir)
147 lprocfs_remove(&ldlm_type_proc_dir);
149 ldlm_svc_proc_dir = NULL;
150 ldlm_type_proc_dir = NULL;
151 ldlm_ns_proc_dir = NULL;
154 static int lprocfs_ns_resources_seq_show(struct seq_file *m, void *v)
156 struct ldlm_namespace *ns = m->private;
161 /* result is not strictly consistant */
162 cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
163 res += cfs_hash_bd_count_get(&bd);
164 return lprocfs_rd_u64(m, &res);
166 LPROC_SEQ_FOPS_RO(lprocfs_ns_resources);
168 static int lprocfs_ns_locks_seq_show(struct seq_file *m, void *v)
170 struct ldlm_namespace *ns = m->private;
173 locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
174 LPROCFS_FIELDS_FLAGS_SUM);
175 return lprocfs_rd_u64(m, &locks);
177 LPROC_SEQ_FOPS_RO(lprocfs_ns_locks);
179 static int lprocfs_lru_size_seq_show(struct seq_file *m, void *v)
181 struct ldlm_namespace *ns = m->private;
182 __u32 *nr = &ns->ns_max_unused;
184 if (ns_connect_lru_resize(ns))
185 nr = &ns->ns_nr_unused;
186 return lprocfs_rd_uint(m, nr);
189 static ssize_t lprocfs_lru_size_seq_write(struct file *file, const char *buffer,
190 size_t count, loff_t *off)
192 struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
193 char dummy[MAX_STRING_SIZE + 1], *end;
197 dummy[MAX_STRING_SIZE] = '\0';
198 if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
201 if (strncmp(dummy, "clear", 5) == 0) {
203 "dropping all unused locks from namespace %s\n",
205 if (ns_connect_lru_resize(ns)) {
206 int canceled, unused = ns->ns_nr_unused;
208 /* Try to cancel all @ns_nr_unused locks. */
209 canceled = ldlm_cancel_lru(ns, unused, 0,
211 if (canceled < unused) {
213 "not all requested locks are canceled, "
214 "requested: %d, canceled: %d\n", unused,
219 tmp = ns->ns_max_unused;
220 ns->ns_max_unused = 0;
221 ldlm_cancel_lru(ns, 0, 0, LDLM_CANCEL_PASSED);
222 ns->ns_max_unused = tmp;
227 tmp = simple_strtoul(dummy, &end, 0);
229 CERROR("invalid value written\n");
232 lru_resize = (tmp == 0);
234 if (ns_connect_lru_resize(ns)) {
236 ns->ns_max_unused = (unsigned int)tmp;
238 if (tmp > ns->ns_nr_unused)
239 tmp = ns->ns_nr_unused;
240 tmp = ns->ns_nr_unused - tmp;
243 "changing namespace %s unused locks from %u to %u\n",
244 ldlm_ns_name(ns), ns->ns_nr_unused,
246 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_CANCEL_PASSED);
250 "disable lru_resize for namespace %s\n",
252 ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
256 "changing namespace %s max_unused from %u to %u\n",
257 ldlm_ns_name(ns), ns->ns_max_unused,
259 ns->ns_max_unused = (unsigned int)tmp;
260 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_CANCEL_PASSED);
262 /* Make sure that LRU resize was originally supported before
263 * turning it on here. */
265 (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
267 "enable lru_resize for namespace %s\n",
269 ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
275 LPROC_SEQ_FOPS(lprocfs_lru_size);
277 static int lprocfs_elc_seq_show(struct seq_file *m, void *v)
279 struct ldlm_namespace *ns = m->private;
280 unsigned int supp = ns_connect_cancelset(ns);
282 return lprocfs_rd_uint(m, &supp);
285 static ssize_t lprocfs_elc_seq_write(struct file *file, const char *buffer,
286 size_t count, loff_t *off)
288 struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
289 unsigned int supp = -1;
292 rc = lprocfs_wr_uint(file, buffer, count, &supp);
297 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
298 else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
299 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
302 LPROC_SEQ_FOPS(lprocfs_elc);
304 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
306 if (ns->ns_proc_dir_entry == NULL)
307 CERROR("dlm namespace %s has no procfs dir?\n",
310 lprocfs_remove(&ns->ns_proc_dir_entry);
312 if (ns->ns_stats != NULL)
313 lprocfs_free_stats(&ns->ns_stats);
316 #define LDLM_NS_ADD_VAR(name, var, ops) \
318 snprintf(lock_name, MAX_STRING_SIZE, name); \
319 lock_vars[0].data = var; \
320 lock_vars[0].fops = ops; \
321 lprocfs_add_vars(ns_pde, lock_vars, 0); \
324 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
326 struct lprocfs_vars lock_vars[2];
327 char lock_name[MAX_STRING_SIZE + 1];
328 proc_dir_entry_t *ns_pde;
331 LASSERT(ns->ns_rs_hash != NULL);
333 if (ns->ns_proc_dir_entry != NULL) {
334 ns_pde = ns->ns_proc_dir_entry;
336 ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir);
339 ns->ns_proc_dir_entry = ns_pde;
342 ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
343 if (ns->ns_stats == NULL)
346 lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
347 LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
349 lock_name[MAX_STRING_SIZE] = '\0';
351 memset(lock_vars, 0, sizeof(lock_vars));
352 lock_vars[0].name = lock_name;
354 LDLM_NS_ADD_VAR("resource_count", ns, &lprocfs_ns_resources_fops);
355 LDLM_NS_ADD_VAR("lock_count", ns, &lprocfs_ns_locks_fops);
357 if (ns_is_client(ns)) {
358 LDLM_NS_ADD_VAR("lock_unused_count", &ns->ns_nr_unused,
360 LDLM_NS_ADD_VAR("lru_size", ns, &lprocfs_lru_size_fops);
361 LDLM_NS_ADD_VAR("lru_max_age", &ns->ns_max_age,
363 LDLM_NS_ADD_VAR("early_lock_cancel", ns, &lprocfs_elc_fops);
365 LDLM_NS_ADD_VAR("ctime_age_limit", &ns->ns_ctime_age_limit,
367 LDLM_NS_ADD_VAR("lock_timeouts", &ns->ns_timeouts,
369 LDLM_NS_ADD_VAR("max_nolock_bytes", &ns->ns_max_nolock_size,
371 LDLM_NS_ADD_VAR("contention_seconds", &ns->ns_contention_time,
373 LDLM_NS_ADD_VAR("contended_locks", &ns->ns_contended_locks,
375 LDLM_NS_ADD_VAR("max_parallel_ast", &ns->ns_max_parallel_ast,
380 #undef MAX_STRING_SIZE
383 #define ldlm_namespace_proc_unregister(ns) ({;})
384 #define ldlm_namespace_proc_register(ns) ({0;})
388 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
389 const void *key, unsigned mask)
391 const struct ldlm_res_id *id = key;
395 for (i = 0; i < RES_NAME_SIZE; i++)
400 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
401 const void *key, unsigned mask)
403 const struct ldlm_res_id *id = key;
408 fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
409 fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
410 fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
412 hash = fid_flatten32(&fid);
413 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
414 if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
415 val = id->name[LUSTRE_RES_ID_HSH_OFF];
416 hash += (val >> 5) + (val << 11);
420 hash = cfs_hash_long(hash, hs->hs_bkt_bits);
421 /* give me another random factor */
422 hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3);
424 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
425 hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
430 static void *ldlm_res_hop_key(struct hlist_node *hnode)
432 struct ldlm_resource *res;
434 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
435 return &res->lr_name;
438 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
440 struct ldlm_resource *res;
442 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
443 return ldlm_res_eq((const struct ldlm_res_id *)key,
444 (const struct ldlm_res_id *)&res->lr_name);
447 static void *ldlm_res_hop_object(struct hlist_node *hnode)
449 return hlist_entry(hnode, struct ldlm_resource, lr_hash);
452 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, struct hlist_node *hnode)
454 struct ldlm_resource *res;
456 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
457 ldlm_resource_getref(res);
460 static void ldlm_res_hop_put_locked(cfs_hash_t *hs, struct hlist_node *hnode)
462 struct ldlm_resource *res;
464 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
465 /* cfs_hash_for_each_nolock is the only chance we call it */
466 ldlm_resource_putref_locked(res);
469 static void ldlm_res_hop_put(cfs_hash_t *hs, struct hlist_node *hnode)
471 struct ldlm_resource *res;
473 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
474 ldlm_resource_putref(res);
477 cfs_hash_ops_t ldlm_ns_hash_ops = {
478 .hs_hash = ldlm_res_hop_hash,
479 .hs_key = ldlm_res_hop_key,
480 .hs_keycmp = ldlm_res_hop_keycmp,
482 .hs_object = ldlm_res_hop_object,
483 .hs_get = ldlm_res_hop_get_locked,
484 .hs_put_locked = ldlm_res_hop_put_locked,
485 .hs_put = ldlm_res_hop_put
488 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
489 .hs_hash = ldlm_res_hop_fid_hash,
490 .hs_key = ldlm_res_hop_key,
491 .hs_keycmp = ldlm_res_hop_keycmp,
493 .hs_object = ldlm_res_hop_object,
494 .hs_get = ldlm_res_hop_get_locked,
495 .hs_put_locked = ldlm_res_hop_put_locked,
496 .hs_put = ldlm_res_hop_put
500 ldlm_ns_type_t nsd_type;
501 /** hash bucket bits */
502 unsigned nsd_bkt_bits;
504 unsigned nsd_all_bits;
505 /** hash operations */
506 cfs_hash_ops_t *nsd_hops;
507 } ldlm_ns_hash_def_t;
509 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
512 .nsd_type = LDLM_NS_TYPE_MDC,
515 .nsd_hops = &ldlm_ns_fid_hash_ops,
518 .nsd_type = LDLM_NS_TYPE_MDT,
521 .nsd_hops = &ldlm_ns_fid_hash_ops,
524 .nsd_type = LDLM_NS_TYPE_OSC,
527 .nsd_hops = &ldlm_ns_hash_ops,
530 .nsd_type = LDLM_NS_TYPE_OST,
533 .nsd_hops = &ldlm_ns_hash_ops,
536 .nsd_type = LDLM_NS_TYPE_MGC,
539 .nsd_hops = &ldlm_ns_hash_ops,
542 .nsd_type = LDLM_NS_TYPE_MGT,
545 .nsd_hops = &ldlm_ns_hash_ops,
548 .nsd_type = LDLM_NS_TYPE_UNKNOWN,
553 * Create and initialize new empty namespace.
555 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
558 ldlm_ns_type_t ns_type)
560 struct ldlm_namespace *ns = NULL;
561 struct ldlm_ns_bucket *nsb;
562 ldlm_ns_hash_def_t *nsd;
568 LASSERT(obd != NULL);
572 CERROR("ldlm_get_ref failed: %d\n", rc);
576 for (idx = 0;;idx++) {
577 nsd = &ldlm_ns_hash_defs[idx];
578 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
579 CERROR("Unknown type %d for ns %s\n", ns_type, name);
583 if (nsd->nsd_type == ns_type)
591 ns->ns_rs_hash = cfs_hash_create(name,
592 nsd->nsd_all_bits, nsd->nsd_all_bits,
593 nsd->nsd_bkt_bits, sizeof(*nsb),
599 CFS_HASH_SPIN_BKTLOCK |
600 CFS_HASH_NO_ITEMREF);
601 if (ns->ns_rs_hash == NULL)
604 cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
605 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
606 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
607 nsb->nsb_namespace = ns;
611 ns->ns_appetite = apt;
612 ns->ns_client = client;
614 INIT_LIST_HEAD(&ns->ns_list_chain);
615 INIT_LIST_HEAD(&ns->ns_unused_list);
616 spin_lock_init(&ns->ns_lock);
617 atomic_set(&ns->ns_bref, 0);
618 init_waitqueue_head(&ns->ns_waitq);
620 ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
621 ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
622 ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
624 ns->ns_max_parallel_ast = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
625 ns->ns_nr_unused = 0;
626 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
627 ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
628 ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
630 ns->ns_orig_connect_flags = 0;
631 ns->ns_connect_flags = 0;
633 rc = ldlm_namespace_proc_register(ns);
635 CERROR("Can't initialize ns proc, rc %d\n", rc);
639 idx = atomic_read(ldlm_namespace_nr(client));
640 rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
642 CERROR("Can't initialize lock pool, rc %d\n", rc);
646 ldlm_namespace_register(ns, client);
649 ldlm_namespace_proc_unregister(ns);
650 ldlm_namespace_cleanup(ns, 0);
652 cfs_hash_putref(ns->ns_rs_hash);
659 EXPORT_SYMBOL(ldlm_namespace_new);
661 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
664 * Cancel and destroy all locks on a resource.
666 * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
667 * clean up. This is currently only used for recovery, and we make
668 * certain assumptions as a result--notably, that we shouldn't cancel
671 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
674 struct list_head *tmp;
675 int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
676 bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
679 struct ldlm_lock *lock = NULL;
681 /* First, we look for non-cleaned-yet lock
682 * all cleaned locks are marked by CLEANED flag. */
684 list_for_each(tmp, q) {
685 lock = list_entry(tmp, struct ldlm_lock,
687 if (lock->l_flags & LDLM_FL_CLEANED) {
692 lock->l_flags |= LDLM_FL_CLEANED;
701 /* Set CBPENDING so nothing in the cancellation path
702 * can match this lock. */
703 lock->l_flags |= LDLM_FL_CBPENDING;
704 lock->l_flags |= LDLM_FL_FAILED;
705 lock->l_flags |= flags;
707 /* ... without sending a CANCEL message for local_only. */
709 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
711 if (local_only && (lock->l_readers || lock->l_writers)) {
712 /* This is a little bit gross, but much better than the
713 * alternative: pretend that we got a blocking AST from
714 * the server, so that when the lock is decref'd, it
715 * will go away ... */
717 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
718 if (lock->l_completion_ast)
719 lock->l_completion_ast(lock, 0, NULL);
720 LDLM_LOCK_RELEASE(lock);
725 struct lustre_handle lockh;
728 ldlm_lock2handle(lock, &lockh);
729 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
731 CERROR("ldlm_cli_cancel: %d\n", rc);
733 ldlm_resource_unlink_lock(lock);
735 LDLM_DEBUG(lock, "Freeing a lock still held by a "
737 ldlm_lock_destroy(lock);
739 LDLM_LOCK_RELEASE(lock);
743 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
744 struct hlist_node *hnode, void *arg)
746 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
747 __u64 flags = *(__u64 *)arg;
749 cleanup_resource(res, &res->lr_granted, flags);
750 cleanup_resource(res, &res->lr_converting, flags);
751 cleanup_resource(res, &res->lr_waiting, flags);
756 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
757 struct hlist_node *hnode, void *arg)
759 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
762 CERROR("Namespace %s resource refcount nonzero "
763 "(%d) after lock cleanup; forcing "
765 ldlm_ns_name(ldlm_res_to_ns(res)),
766 atomic_read(&res->lr_refcount) - 1);
768 CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
769 LPU64") (rc: %d)\n", res,
770 res->lr_name.name[0], res->lr_name.name[1],
771 res->lr_name.name[2], res->lr_name.name[3],
772 atomic_read(&res->lr_refcount) - 1);
774 ldlm_resource_dump(D_ERROR, res);
780 * Cancel and destroy all locks in the namespace.
782 * Typically used during evictions when server notified client that it was
783 * evicted and all of its state needs to be destroyed.
784 * Also used during shutdown.
786 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
789 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
793 cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean, &flags);
794 cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
797 EXPORT_SYMBOL(ldlm_namespace_cleanup);
800 * Attempts to free namespace.
802 * Only used when namespace goes away, like during an unmount.
804 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
808 /* At shutdown time, don't call the cancellation callback */
809 ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
811 if (atomic_read(&ns->ns_bref) > 0) {
812 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
815 "dlm namespace %s free waiting on refcount %d\n",
816 ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
819 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
821 rc = l_wait_event(ns->ns_waitq,
822 atomic_read(&ns->ns_bref) == 0, &lwi);
824 /* Forced cleanups should be able to reclaim all references,
825 * so it's safe to wait forever... we can't leak locks... */
826 if (force && rc == -ETIMEDOUT) {
827 LCONSOLE_ERROR("Forced cleanup waiting for %s "
828 "namespace with %d resources in use, "
829 "(rc=%d)\n", ldlm_ns_name(ns),
830 atomic_read(&ns->ns_bref), rc);
831 GOTO(force_wait, rc);
834 if (atomic_read(&ns->ns_bref)) {
835 LCONSOLE_ERROR("Cleanup waiting for %s namespace "
836 "with %d resources in use, (rc=%d)\n",
838 atomic_read(&ns->ns_bref), rc);
839 RETURN(ELDLM_NAMESPACE_EXISTS);
841 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
849 * Performs various cleanups for passed \a ns to make it drop refc and be
850 * ready for freeing. Waits for refc == 0.
852 * The following is done:
853 * (0) Unregister \a ns from its list to make inaccessible for potential
854 * users like pools thread and others;
855 * (1) Clear all locks in \a ns.
857 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
858 struct obd_import *imp,
868 spin_lock(&ns->ns_lock);
870 spin_unlock(&ns->ns_lock);
873 * Can fail with -EINTR when force == 0 in which case try harder.
875 rc = __ldlm_namespace_free(ns, force);
876 if (rc != ELDLM_OK) {
878 ptlrpc_disconnect_import(imp, 0);
879 ptlrpc_invalidate_import(imp);
883 * With all requests dropped and the import inactive
884 * we are gaurenteed all reference will be dropped.
886 rc = __ldlm_namespace_free(ns, 1);
893 * Performs freeing memory structures related to \a ns. This is only done
894 * when ldlm_namespce_free_prior() successfully removed all resources
895 * referencing \a ns and its refc == 0.
897 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
905 /* Make sure that nobody can find this ns in its list. */
906 ldlm_namespace_unregister(ns, ns->ns_client);
907 /* Fini pool _before_ parent proc dir is removed. This is important as
908 * ldlm_pool_fini() removes own proc dir which is child to @dir.
909 * Removing it after @dir may cause oops. */
910 ldlm_pool_fini(&ns->ns_pool);
912 ldlm_namespace_proc_unregister(ns);
913 cfs_hash_putref(ns->ns_rs_hash);
914 /* Namespace \a ns should be not on list at this time, otherwise
915 * this will cause issues related to using freed \a ns in poold
917 LASSERT(list_empty(&ns->ns_list_chain));
924 * Cleanup the resource, and free namespace.
927 * proc1: destroy import
928 * class_disconnect_export(grab cl_sem) ->
929 * -> ldlm_namespace_free ->
930 * -> lprocfs_remove(grab _lprocfs_lock).
931 * proc2: read proc info
932 * lprocfs_fops_read(grab _lprocfs_lock) ->
933 * -> osc_rd_active, etc(grab cl_sem).
935 * So that I have to split the ldlm_namespace_free into two parts - the first
936 * part ldlm_namespace_free_prior is used to cleanup the resource which is
937 * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
938 * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
941 void ldlm_namespace_free(struct ldlm_namespace *ns,
942 struct obd_import *imp,
945 ldlm_namespace_free_prior(ns, imp, force);
946 ldlm_namespace_free_post(ns);
948 EXPORT_SYMBOL(ldlm_namespace_free);
950 void ldlm_namespace_get(struct ldlm_namespace *ns)
952 atomic_inc(&ns->ns_bref);
954 EXPORT_SYMBOL(ldlm_namespace_get);
956 void ldlm_namespace_put(struct ldlm_namespace *ns)
958 if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
959 wake_up(&ns->ns_waitq);
960 spin_unlock(&ns->ns_lock);
963 EXPORT_SYMBOL(ldlm_namespace_put);
965 /** Register \a ns in the list of namespaces */
966 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
968 mutex_lock(ldlm_namespace_lock(client));
969 LASSERT(list_empty(&ns->ns_list_chain));
970 list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
971 atomic_inc(ldlm_namespace_nr(client));
972 mutex_unlock(ldlm_namespace_lock(client));
975 /** Unregister \a ns from the list of namespaces. */
976 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
978 mutex_lock(ldlm_namespace_lock(client));
979 LASSERT(!list_empty(&ns->ns_list_chain));
980 /* Some asserts and possibly other parts of the code are still
981 * using list_empty(&ns->ns_list_chain). This is why it is
982 * important to use list_del_init() here. */
983 list_del_init(&ns->ns_list_chain);
984 atomic_dec(ldlm_namespace_nr(client));
985 mutex_unlock(ldlm_namespace_lock(client));
988 /** Should be called with ldlm_namespace_lock(client) taken. */
989 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
991 LASSERT(!list_empty(&ns->ns_list_chain));
992 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
993 list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
996 /** Should be called with ldlm_namespace_lock(client) taken. */
997 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
999 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1000 LASSERT(!list_empty(ldlm_namespace_list(client)));
1001 return container_of(ldlm_namespace_list(client)->next,
1002 struct ldlm_namespace, ns_list_chain);
1005 /** Create and initialize new resource. */
1006 static struct ldlm_resource *ldlm_resource_new(void)
1008 struct ldlm_resource *res;
1011 OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, __GFP_IO);
1015 INIT_LIST_HEAD(&res->lr_granted);
1016 INIT_LIST_HEAD(&res->lr_converting);
1017 INIT_LIST_HEAD(&res->lr_waiting);
1019 /* Initialize interval trees for each lock mode. */
1020 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1021 res->lr_itree[idx].lit_size = 0;
1022 res->lr_itree[idx].lit_mode = 1 << idx;
1023 res->lr_itree[idx].lit_root = NULL;
1026 atomic_set(&res->lr_refcount, 1);
1027 spin_lock_init(&res->lr_lock);
1028 lu_ref_init(&res->lr_reference);
1030 /* The creator of the resource must unlock the mutex after LVB
1031 * initialization. */
1032 mutex_init(&res->lr_lvb_mutex);
1033 mutex_lock(&res->lr_lvb_mutex);
1039 * Return a reference to resource with given name, creating it if necessary.
1040 * Args: namespace with ns_lock unlocked
1041 * Locks: takes and releases NS hash-lock and res->lr_lock
1042 * Returns: referenced, unlocked ldlm_resource or NULL
1044 struct ldlm_resource *
1045 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1046 const struct ldlm_res_id *name, ldlm_type_t type, int create)
1048 struct hlist_node *hnode;
1049 struct ldlm_resource *res;
1053 LASSERT(ns != NULL);
1054 LASSERT(parent == NULL);
1055 LASSERT(ns->ns_rs_hash != NULL);
1056 LASSERT(name->name[0] != 0);
1058 cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1059 hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1060 if (hnode != NULL) {
1061 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1062 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1063 /* Synchronize with regard to resource creation. */
1064 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1065 mutex_lock(&res->lr_lvb_mutex);
1066 mutex_unlock(&res->lr_lvb_mutex);
1069 if (unlikely(res->lr_lvb_len < 0)) {
1070 ldlm_resource_putref(res);
1076 version = cfs_hash_bd_version_get(&bd);
1077 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1082 LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1083 "type: %d\n", type);
1084 res = ldlm_resource_new();
1088 res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1089 res->lr_name = *name;
1090 res->lr_type = type;
1091 res->lr_most_restr = LCK_NL;
1093 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1094 hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1095 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1097 if (hnode != NULL) {
1098 /* Someone won the race and already added the resource. */
1099 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1100 /* Clean lu_ref for failed resource. */
1101 lu_ref_fini(&res->lr_reference);
1102 /* We have taken lr_lvb_mutex. Drop it. */
1103 mutex_unlock(&res->lr_lvb_mutex);
1104 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1106 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1107 /* Synchronize with regard to resource creation. */
1108 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1109 mutex_lock(&res->lr_lvb_mutex);
1110 mutex_unlock(&res->lr_lvb_mutex);
1113 if (unlikely(res->lr_lvb_len < 0)) {
1114 ldlm_resource_putref(res);
1119 /* We won! Let's add the resource. */
1120 cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1121 if (cfs_hash_bd_count_get(&bd) == 1)
1122 ldlm_namespace_get(ns);
1124 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1125 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1128 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1129 rc = ns->ns_lvbo->lvbo_init(res);
1131 CERROR("lvbo_init failed for resource "
1132 LPU64": rc %d\n", name->name[0], rc);
1133 if (res->lr_lvb_data) {
1134 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
1135 res->lr_lvb_data = NULL;
1137 res->lr_lvb_len = rc;
1138 mutex_unlock(&res->lr_lvb_mutex);
1139 ldlm_resource_putref(res);
1144 /* We create resource with locked lr_lvb_mutex. */
1145 mutex_unlock(&res->lr_lvb_mutex);
1149 EXPORT_SYMBOL(ldlm_resource_get);
1151 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1153 LASSERT(res != NULL);
1154 LASSERT(res != LP_POISON);
1155 atomic_inc(&res->lr_refcount);
1156 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1157 atomic_read(&res->lr_refcount));
1161 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1162 struct ldlm_resource *res)
1164 struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1166 if (!list_empty(&res->lr_granted)) {
1167 ldlm_resource_dump(D_ERROR, res);
1171 if (!list_empty(&res->lr_converting)) {
1172 ldlm_resource_dump(D_ERROR, res);
1176 if (!list_empty(&res->lr_waiting)) {
1177 ldlm_resource_dump(D_ERROR, res);
1181 cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1183 lu_ref_fini(&res->lr_reference);
1184 if (cfs_hash_bd_count_get(bd) == 0)
1185 ldlm_namespace_put(nsb->nsb_namespace);
1188 /* Returns 1 if the resource was freed, 0 if it remains. */
1189 int ldlm_resource_putref(struct ldlm_resource *res)
1191 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1194 LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1195 CDEBUG(D_INFO, "putref res: %p count: %d\n",
1196 res, atomic_read(&res->lr_refcount) - 1);
1198 cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1199 if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1200 __ldlm_resource_putref_final(&bd, res);
1201 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1202 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1203 ns->ns_lvbo->lvbo_free(res);
1204 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1209 EXPORT_SYMBOL(ldlm_resource_putref);
1211 /* Returns 1 if the resource was freed, 0 if it remains. */
1212 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1214 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1216 LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1217 CDEBUG(D_INFO, "putref res: %p count: %d\n",
1218 res, atomic_read(&res->lr_refcount) - 1);
1220 if (atomic_dec_and_test(&res->lr_refcount)) {
1223 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1224 &res->lr_name, &bd);
1225 __ldlm_resource_putref_final(&bd, res);
1226 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1227 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1228 * so we should never be here while calling cfs_hash_del,
1229 * cfs_hash_for_each_nolock is the only case we can get
1230 * here, which is safe to release cfs_hash_bd_lock.
1232 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1233 ns->ns_lvbo->lvbo_free(res);
1234 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1236 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1243 * Add a lock into a given resource into specified lock list.
1245 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1246 struct ldlm_lock *lock)
1248 check_res_locked(res);
1250 LDLM_DEBUG(lock, "About to add this lock:\n");
1252 if (lock->l_destroyed) {
1253 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1257 LASSERT(list_empty(&lock->l_res_link));
1259 list_add_tail(&lock->l_res_link, head);
1263 * Insert a lock into resource after specified lock.
1265 * Obtain resource description from the lock we are inserting after.
1267 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1268 struct ldlm_lock *new)
1270 struct ldlm_resource *res = original->l_resource;
1272 check_res_locked(res);
1274 ldlm_resource_dump(D_INFO, res);
1275 LDLM_DEBUG(new, "About to insert this lock after %p:\n", original);
1277 if (new->l_destroyed) {
1278 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1282 LASSERT(list_empty(&new->l_res_link));
1284 list_add(&new->l_res_link, &original->l_res_link);
1288 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1290 int type = lock->l_resource->lr_type;
1292 check_res_locked(lock->l_resource);
1293 if (type == LDLM_IBITS || type == LDLM_PLAIN)
1294 ldlm_unlink_lock_skiplist(lock);
1295 else if (type == LDLM_EXTENT)
1296 ldlm_extent_unlink_lock(lock);
1297 list_del_init(&lock->l_res_link);
1299 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1301 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1303 desc->lr_type = res->lr_type;
1304 desc->lr_name = res->lr_name;
1308 * Print information about all locks in all namespaces on this node to debug
1311 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1313 struct list_head *tmp;
1315 if (!((libcfs_debug | D_ERROR) & level))
1318 mutex_lock(ldlm_namespace_lock(client));
1320 list_for_each(tmp, ldlm_namespace_list(client)) {
1321 struct ldlm_namespace *ns;
1322 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1323 ldlm_namespace_dump(level, ns);
1326 mutex_unlock(ldlm_namespace_lock(client));
1328 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1330 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1331 struct hlist_node *hnode, void *arg)
1333 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1334 int level = (int)(unsigned long)arg;
1337 ldlm_resource_dump(level, res);
1344 * Print information about all locks in this namespace on this node to debug
1347 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1349 if (!((libcfs_debug | D_ERROR) & level))
1352 CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1353 ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1354 ns_is_client(ns) ? "client" : "server");
1356 if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1359 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1361 (void *)(unsigned long)level);
1362 spin_lock(&ns->ns_lock);
1363 ns->ns_next_dump = cfs_time_shift(10);
1364 spin_unlock(&ns->ns_lock);
1366 EXPORT_SYMBOL(ldlm_namespace_dump);
1369 * Print information about all locks in this resource to debug log.
1371 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1373 struct ldlm_lock *lock;
1374 unsigned int granted = 0;
1376 CLASSERT(RES_NAME_SIZE == 4);
1378 if (!((libcfs_debug | D_ERROR) & level))
1381 CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1382 ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1383 res->lr_name.name[2], res->lr_name.name[3],
1384 atomic_read(&res->lr_refcount));
1386 if (!list_empty(&res->lr_granted)) {
1387 CDEBUG(level, "Granted locks (in reverse order):\n");
1388 list_for_each_entry_reverse(lock, &res->lr_granted,
1390 LDLM_DEBUG_LIMIT(level, lock, "###");
1391 if (!(level & D_CANTMASK) &&
1392 ++granted > ldlm_dump_granted_max) {
1393 CDEBUG(level, "only dump %d granted locks to "
1394 "avoid DDOS.\n", granted);
1399 if (!list_empty(&res->lr_converting)) {
1400 CDEBUG(level, "Converting locks:\n");
1401 list_for_each_entry(lock, &res->lr_converting, l_res_link)
1402 LDLM_DEBUG_LIMIT(level, lock, "###");
1404 if (!list_empty(&res->lr_waiting)) {
1405 CDEBUG(level, "Waiting locks:\n");
1406 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1407 LDLM_DEBUG_LIMIT(level, lock, "###");