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, 2013, 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>
44 #include <lustre_fid.h>
45 #include <obd_class.h>
46 #include "ldlm_internal.h"
48 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
50 int ldlm_srv_namespace_nr = 0;
51 int ldlm_cli_namespace_nr = 0;
53 struct mutex ldlm_srv_namespace_lock;
54 struct list_head ldlm_srv_namespace_list;
56 struct mutex ldlm_cli_namespace_lock;
57 /* Client Namespaces that have active resources in them.
58 * Once all resources go away, ldlm_poold moves such namespaces to the
60 struct list_head ldlm_cli_active_namespace_list;
61 /* Client namespaces that don't have any locks in them */
62 struct list_head ldlm_cli_inactive_namespace_list;
64 struct proc_dir_entry *ldlm_type_proc_dir = NULL;
65 struct proc_dir_entry *ldlm_ns_proc_dir = NULL;
66 struct proc_dir_entry *ldlm_svc_proc_dir = NULL;
68 extern unsigned int ldlm_cancel_unused_locks_before_replay;
70 /* during debug dump certain amount of granted locks for one resource to avoid
72 unsigned int ldlm_dump_granted_max = 256;
76 lprocfs_dump_ns_seq_write(struct file *file, const char __user *buffer,
77 size_t count, loff_t *off)
79 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
80 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
83 LPROC_SEQ_FOPS_WO_TYPE(ldlm, dump_ns);
85 LPROC_SEQ_FOPS_RW_TYPE(ldlm_rw, uint);
86 LPROC_SEQ_FOPS_RO_TYPE(ldlm, uint);
88 int ldlm_proc_setup(void)
91 struct lprocfs_seq_vars list[] = {
92 { .name = "dump_namespaces",
93 .fops = &ldlm_dump_ns_fops,
95 { .name = "dump_granted_max",
96 .fops = &ldlm_rw_uint_fops,
97 .data = &ldlm_dump_granted_max },
98 { .name = "cancel_unused_locks_before_replay",
99 .fops = &ldlm_rw_uint_fops,
100 .data = &ldlm_cancel_unused_locks_before_replay },
103 LASSERT(ldlm_ns_proc_dir == NULL);
105 ldlm_type_proc_dir = lprocfs_seq_register(OBD_LDLM_DEVICENAME,
108 if (IS_ERR(ldlm_type_proc_dir)) {
109 CERROR("LProcFS failed in ldlm-init\n");
110 rc = PTR_ERR(ldlm_type_proc_dir);
114 ldlm_ns_proc_dir = lprocfs_seq_register("namespaces",
117 if (IS_ERR(ldlm_ns_proc_dir)) {
118 CERROR("LProcFS failed in ldlm-init\n");
119 rc = PTR_ERR(ldlm_ns_proc_dir);
123 ldlm_svc_proc_dir = lprocfs_seq_register("services",
126 if (IS_ERR(ldlm_svc_proc_dir)) {
127 CERROR("LProcFS failed in ldlm-init\n");
128 rc = PTR_ERR(ldlm_svc_proc_dir);
132 rc = lprocfs_seq_add_vars(ldlm_type_proc_dir, list, NULL);
134 CERROR("LProcFS failed in ldlm-init\n");
141 lprocfs_remove(&ldlm_svc_proc_dir);
143 lprocfs_remove(&ldlm_ns_proc_dir);
145 lprocfs_remove(&ldlm_type_proc_dir);
147 ldlm_svc_proc_dir = NULL;
151 void ldlm_proc_cleanup(void)
153 if (ldlm_svc_proc_dir)
154 lprocfs_remove(&ldlm_svc_proc_dir);
156 if (ldlm_ns_proc_dir)
157 lprocfs_remove(&ldlm_ns_proc_dir);
159 if (ldlm_type_proc_dir)
160 lprocfs_remove(&ldlm_type_proc_dir);
163 static int lprocfs_ns_resources_seq_show(struct seq_file *m, void *v)
165 struct ldlm_namespace *ns = m->private;
170 /* result is not strictly consistant */
171 cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
172 res += cfs_hash_bd_count_get(&bd);
173 return lprocfs_u64_seq_show(m, &res);
175 LPROC_SEQ_FOPS_RO(lprocfs_ns_resources);
177 static int lprocfs_ns_locks_seq_show(struct seq_file *m, void *v)
179 struct ldlm_namespace *ns = m->private;
182 locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
183 LPROCFS_FIELDS_FLAGS_SUM);
184 return lprocfs_u64_seq_show(m, &locks);
186 LPROC_SEQ_FOPS_RO(lprocfs_ns_locks);
188 static int lprocfs_lru_size_seq_show(struct seq_file *m, void *v)
190 struct ldlm_namespace *ns = m->private;
191 __u32 *nr = &ns->ns_max_unused;
193 if (ns_connect_lru_resize(ns))
194 nr = &ns->ns_nr_unused;
195 return lprocfs_uint_seq_show(m, nr);
198 static ssize_t lprocfs_lru_size_seq_write(struct file *file,
199 const char __user *buffer,
200 size_t count, loff_t *off)
202 struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
203 char dummy[MAX_STRING_SIZE + 1], *end;
207 dummy[MAX_STRING_SIZE] = '\0';
208 if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
211 if (strncmp(dummy, "clear", 5) == 0) {
213 "dropping all unused locks from namespace %s\n",
215 if (ns_connect_lru_resize(ns)) {
216 int canceled, unused = ns->ns_nr_unused;
218 /* Try to cancel all @ns_nr_unused locks. */
219 canceled = ldlm_cancel_lru(ns, unused, 0,
221 if (canceled < unused) {
223 "not all requested locks are canceled, "
224 "requested: %d, canceled: %d\n", unused,
229 tmp = ns->ns_max_unused;
230 ns->ns_max_unused = 0;
231 ldlm_cancel_lru(ns, 0, 0, LDLM_CANCEL_PASSED);
232 ns->ns_max_unused = tmp;
237 tmp = simple_strtoul(dummy, &end, 0);
239 CERROR("invalid value written\n");
242 lru_resize = (tmp == 0);
244 if (ns_connect_lru_resize(ns)) {
246 ns->ns_max_unused = (unsigned int)tmp;
248 if (tmp > ns->ns_nr_unused)
249 tmp = ns->ns_nr_unused;
250 tmp = ns->ns_nr_unused - tmp;
253 "changing namespace %s unused locks from %u to %u\n",
254 ldlm_ns_name(ns), ns->ns_nr_unused,
256 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_CANCEL_PASSED);
260 "disable lru_resize for namespace %s\n",
262 ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
266 "changing namespace %s max_unused from %u to %u\n",
267 ldlm_ns_name(ns), ns->ns_max_unused,
269 ns->ns_max_unused = (unsigned int)tmp;
270 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_CANCEL_PASSED);
272 /* Make sure that LRU resize was originally supported before
273 * turning it on here. */
275 (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
277 "enable lru_resize for namespace %s\n",
279 ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
285 LPROC_SEQ_FOPS(lprocfs_lru_size);
287 static int lprocfs_elc_seq_show(struct seq_file *m, void *v)
289 struct ldlm_namespace *ns = m->private;
290 unsigned int supp = ns_connect_cancelset(ns);
292 return lprocfs_uint_seq_show(m, &supp);
295 static ssize_t lprocfs_elc_seq_write(struct file *file,
296 const char __user *buffer,
297 size_t count, loff_t *off)
299 struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
300 unsigned int supp = -1;
303 rc = lprocfs_wr_uint(file, buffer, count, &supp);
308 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
309 else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
310 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
313 LPROC_SEQ_FOPS(lprocfs_elc);
315 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
317 if (ns->ns_proc_dir_entry == NULL)
318 CERROR("dlm namespace %s has no procfs dir?\n",
321 lprocfs_remove(&ns->ns_proc_dir_entry);
323 if (ns->ns_stats != NULL)
324 lprocfs_free_stats(&ns->ns_stats);
327 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
329 struct lprocfs_seq_vars lock_vars[2];
330 char lock_name[MAX_STRING_SIZE + 1];
331 struct proc_dir_entry *ns_pde;
334 LASSERT(ns->ns_rs_hash != NULL);
336 if (ns->ns_proc_dir_entry != NULL) {
337 ns_pde = ns->ns_proc_dir_entry;
339 ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir);
342 ns->ns_proc_dir_entry = ns_pde;
345 ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
346 if (ns->ns_stats == NULL)
349 lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
350 LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
352 lock_name[MAX_STRING_SIZE] = '\0';
354 memset(lock_vars, 0, sizeof(lock_vars));
355 lock_vars[0].name = lock_name;
357 ldlm_add_var(&lock_vars[0], ns_pde, "resource_count", ns,
358 &lprocfs_ns_resources_fops);
359 ldlm_add_var(&lock_vars[0], ns_pde, "lock_count", ns,
360 &lprocfs_ns_locks_fops);
362 if (ns_is_client(ns)) {
363 ldlm_add_var(&lock_vars[0], ns_pde, "lock_unused_count",
364 &ns->ns_nr_unused, &ldlm_uint_fops);
365 ldlm_add_var(&lock_vars[0], ns_pde, "lru_size", ns,
366 &lprocfs_lru_size_fops);
367 ldlm_add_var(&lock_vars[0], ns_pde, "lru_max_age",
368 &ns->ns_max_age, &ldlm_rw_uint_fops);
369 ldlm_add_var(&lock_vars[0], ns_pde, "early_lock_cancel",
370 ns, &lprocfs_elc_fops);
372 ldlm_add_var(&lock_vars[0], ns_pde, "ctime_age_limit",
373 &ns->ns_ctime_age_limit, &ldlm_rw_uint_fops);
374 ldlm_add_var(&lock_vars[0], ns_pde, "lock_timeouts",
375 &ns->ns_timeouts, &ldlm_uint_fops);
376 ldlm_add_var(&lock_vars[0], ns_pde, "max_nolock_bytes",
377 &ns->ns_max_nolock_size, &ldlm_rw_uint_fops);
378 ldlm_add_var(&lock_vars[0], ns_pde, "contention_seconds",
379 &ns->ns_contention_time, &ldlm_rw_uint_fops);
380 ldlm_add_var(&lock_vars[0], ns_pde, "contended_locks",
381 &ns->ns_contended_locks, &ldlm_rw_uint_fops);
382 ldlm_add_var(&lock_vars[0], ns_pde, "max_parallel_ast",
383 &ns->ns_max_parallel_ast, &ldlm_rw_uint_fops);
387 #undef MAX_STRING_SIZE
390 #define ldlm_namespace_proc_unregister(ns) ({;})
391 #define ldlm_namespace_proc_register(ns) ({0;})
395 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
396 const void *key, unsigned mask)
398 const struct ldlm_res_id *id = key;
402 for (i = 0; i < RES_NAME_SIZE; i++)
407 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
408 const void *key, unsigned mask)
410 const struct ldlm_res_id *id = key;
415 fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
416 fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
417 fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
419 hash = fid_flatten32(&fid);
420 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
421 if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
422 val = id->name[LUSTRE_RES_ID_HSH_OFF];
423 hash += (val >> 5) + (val << 11);
427 hash = hash_long(hash, hs->hs_bkt_bits);
428 /* give me another random factor */
429 hash -= hash_long((unsigned long)hs, val % 11 + 3);
431 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
432 hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
437 static void *ldlm_res_hop_key(struct hlist_node *hnode)
439 struct ldlm_resource *res;
441 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
442 return &res->lr_name;
445 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
447 struct ldlm_resource *res;
449 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
450 return ldlm_res_eq((const struct ldlm_res_id *)key,
451 (const struct ldlm_res_id *)&res->lr_name);
454 static void *ldlm_res_hop_object(struct hlist_node *hnode)
456 return hlist_entry(hnode, struct ldlm_resource, lr_hash);
459 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, struct hlist_node *hnode)
461 struct ldlm_resource *res;
463 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
464 ldlm_resource_getref(res);
467 static void ldlm_res_hop_put_locked(cfs_hash_t *hs, struct hlist_node *hnode)
469 struct ldlm_resource *res;
471 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
472 /* cfs_hash_for_each_nolock is the only chance we call it */
473 ldlm_resource_putref_locked(res);
476 static void ldlm_res_hop_put(cfs_hash_t *hs, struct hlist_node *hnode)
478 struct ldlm_resource *res;
480 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
481 ldlm_resource_putref(res);
484 cfs_hash_ops_t ldlm_ns_hash_ops = {
485 .hs_hash = ldlm_res_hop_hash,
486 .hs_key = ldlm_res_hop_key,
487 .hs_keycmp = ldlm_res_hop_keycmp,
489 .hs_object = ldlm_res_hop_object,
490 .hs_get = ldlm_res_hop_get_locked,
491 .hs_put_locked = ldlm_res_hop_put_locked,
492 .hs_put = ldlm_res_hop_put
495 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
496 .hs_hash = ldlm_res_hop_fid_hash,
497 .hs_key = ldlm_res_hop_key,
498 .hs_keycmp = ldlm_res_hop_keycmp,
500 .hs_object = ldlm_res_hop_object,
501 .hs_get = ldlm_res_hop_get_locked,
502 .hs_put_locked = ldlm_res_hop_put_locked,
503 .hs_put = ldlm_res_hop_put
507 ldlm_ns_type_t nsd_type;
508 /** hash bucket bits */
509 unsigned nsd_bkt_bits;
511 unsigned nsd_all_bits;
512 /** hash operations */
513 cfs_hash_ops_t *nsd_hops;
514 } ldlm_ns_hash_def_t;
516 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
519 .nsd_type = LDLM_NS_TYPE_MDC,
522 .nsd_hops = &ldlm_ns_fid_hash_ops,
525 .nsd_type = LDLM_NS_TYPE_MDT,
528 .nsd_hops = &ldlm_ns_fid_hash_ops,
531 .nsd_type = LDLM_NS_TYPE_OSC,
534 .nsd_hops = &ldlm_ns_hash_ops,
537 .nsd_type = LDLM_NS_TYPE_OST,
540 .nsd_hops = &ldlm_ns_hash_ops,
543 .nsd_type = LDLM_NS_TYPE_MGC,
546 .nsd_hops = &ldlm_ns_hash_ops,
549 .nsd_type = LDLM_NS_TYPE_MGT,
552 .nsd_hops = &ldlm_ns_hash_ops,
555 .nsd_type = LDLM_NS_TYPE_UNKNOWN,
560 * Create and initialize new empty namespace.
562 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
565 ldlm_ns_type_t ns_type)
567 struct ldlm_namespace *ns = NULL;
568 struct ldlm_ns_bucket *nsb;
569 ldlm_ns_hash_def_t *nsd;
575 LASSERT(obd != NULL);
579 CERROR("ldlm_get_ref failed: %d\n", rc);
583 for (idx = 0;;idx++) {
584 nsd = &ldlm_ns_hash_defs[idx];
585 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
586 CERROR("Unknown type %d for ns %s\n", ns_type, name);
590 if (nsd->nsd_type == ns_type)
598 ns->ns_rs_hash = cfs_hash_create(name,
599 nsd->nsd_all_bits, nsd->nsd_all_bits,
600 nsd->nsd_bkt_bits, sizeof(*nsb),
606 CFS_HASH_SPIN_BKTLOCK |
607 CFS_HASH_NO_ITEMREF);
608 if (ns->ns_rs_hash == NULL)
611 cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
612 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
613 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
614 nsb->nsb_namespace = ns;
618 ns->ns_appetite = apt;
619 ns->ns_client = client;
621 INIT_LIST_HEAD(&ns->ns_list_chain);
622 INIT_LIST_HEAD(&ns->ns_unused_list);
623 spin_lock_init(&ns->ns_lock);
624 atomic_set(&ns->ns_bref, 0);
625 init_waitqueue_head(&ns->ns_waitq);
627 ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
628 ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
629 ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
631 ns->ns_max_parallel_ast = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
632 ns->ns_nr_unused = 0;
633 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
634 ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
635 ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
637 ns->ns_orig_connect_flags = 0;
638 ns->ns_connect_flags = 0;
640 rc = ldlm_namespace_proc_register(ns);
642 CERROR("Can't initialize ns proc, rc %d\n", rc);
646 idx = ldlm_namespace_nr_read(client);
647 rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
649 CERROR("Can't initialize lock pool, rc %d\n", rc);
653 ldlm_namespace_register(ns, client);
656 ldlm_namespace_proc_unregister(ns);
657 ldlm_namespace_cleanup(ns, 0);
659 cfs_hash_putref(ns->ns_rs_hash);
666 EXPORT_SYMBOL(ldlm_namespace_new);
668 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
671 * Cancel and destroy all locks on a resource.
673 * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
674 * clean up. This is currently only used for recovery, and we make
675 * certain assumptions as a result--notably, that we shouldn't cancel
678 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
681 struct list_head *tmp;
682 int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
683 bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
686 struct ldlm_lock *lock = NULL;
688 /* First, we look for non-cleaned-yet lock
689 * all cleaned locks are marked by CLEANED flag. */
691 list_for_each(tmp, q) {
692 lock = list_entry(tmp, struct ldlm_lock,
694 if (ldlm_is_cleaned(lock)) {
699 ldlm_set_cleaned(lock);
708 /* Set CBPENDING so nothing in the cancellation path
709 * can match this lock. */
710 ldlm_set_cbpending(lock);
711 ldlm_set_failed(lock);
712 lock->l_flags |= flags;
714 /* ... without sending a CANCEL message for local_only. */
716 ldlm_set_local_only(lock);
718 if (local_only && (lock->l_readers || lock->l_writers)) {
719 /* This is a little bit gross, but much better than the
720 * alternative: pretend that we got a blocking AST from
721 * the server, so that when the lock is decref'd, it
722 * will go away ... */
724 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
725 if (lock->l_flags & LDLM_FL_FAIL_LOC) {
726 schedule_timeout_and_set_state(
727 TASK_UNINTERRUPTIBLE,
728 cfs_time_seconds(4));
729 set_current_state(TASK_RUNNING);
731 if (lock->l_completion_ast)
732 lock->l_completion_ast(lock,
733 LDLM_FL_FAILED, NULL);
734 LDLM_LOCK_RELEASE(lock);
739 struct lustre_handle lockh;
742 ldlm_lock2handle(lock, &lockh);
743 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
745 CERROR("ldlm_cli_cancel: %d\n", rc);
747 ldlm_resource_unlink_lock(lock);
749 LDLM_DEBUG(lock, "Freeing a lock still held by a "
751 ldlm_lock_destroy(lock);
753 LDLM_LOCK_RELEASE(lock);
757 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
758 struct hlist_node *hnode, void *arg)
760 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
761 __u64 flags = *(__u64 *)arg;
763 cleanup_resource(res, &res->lr_granted, flags);
764 cleanup_resource(res, &res->lr_converting, flags);
765 cleanup_resource(res, &res->lr_waiting, flags);
770 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
771 struct hlist_node *hnode, void *arg)
773 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
776 CERROR("%s: namespace resource "DLDLMRES" (%p) refcount nonzero "
777 "(%d) after lock cleanup; forcing cleanup.\n",
778 ldlm_ns_name(ldlm_res_to_ns(res)), PLDLMRES(res), res,
779 atomic_read(&res->lr_refcount) - 1);
781 ldlm_resource_dump(D_ERROR, res);
787 * Cancel and destroy all locks in the namespace.
789 * Typically used during evictions when server notified client that it was
790 * evicted and all of its state needs to be destroyed.
791 * Also used during shutdown.
793 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
796 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
800 cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean, &flags);
801 cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
804 EXPORT_SYMBOL(ldlm_namespace_cleanup);
807 * Attempts to free namespace.
809 * Only used when namespace goes away, like during an unmount.
811 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
815 /* At shutdown time, don't call the cancellation callback */
816 ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
818 if (atomic_read(&ns->ns_bref) > 0) {
819 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
822 "dlm namespace %s free waiting on refcount %d\n",
823 ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
826 lwi = LWI_TIMEOUT(msecs_to_jiffies(obd_timeout *
827 MSEC_PER_SEC) / 4, NULL, NULL);
829 rc = l_wait_event(ns->ns_waitq,
830 atomic_read(&ns->ns_bref) == 0, &lwi);
832 /* Forced cleanups should be able to reclaim all references,
833 * so it's safe to wait forever... we can't leak locks... */
834 if (force && rc == -ETIMEDOUT) {
835 LCONSOLE_ERROR("Forced cleanup waiting for %s "
836 "namespace with %d resources in use, "
837 "(rc=%d)\n", ldlm_ns_name(ns),
838 atomic_read(&ns->ns_bref), rc);
839 GOTO(force_wait, rc);
842 if (atomic_read(&ns->ns_bref)) {
843 LCONSOLE_ERROR("Cleanup waiting for %s namespace "
844 "with %d resources in use, (rc=%d)\n",
846 atomic_read(&ns->ns_bref), rc);
847 RETURN(ELDLM_NAMESPACE_EXISTS);
849 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
857 * Performs various cleanups for passed \a ns to make it drop refc and be
858 * ready for freeing. Waits for refc == 0.
860 * The following is done:
861 * (0) Unregister \a ns from its list to make inaccessible for potential
862 * users like pools thread and others;
863 * (1) Clear all locks in \a ns.
865 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
866 struct obd_import *imp,
876 spin_lock(&ns->ns_lock);
878 spin_unlock(&ns->ns_lock);
881 * Can fail with -EINTR when force == 0 in which case try harder.
883 rc = __ldlm_namespace_free(ns, force);
884 if (rc != ELDLM_OK) {
886 ptlrpc_disconnect_import(imp, 0);
887 ptlrpc_invalidate_import(imp);
891 * With all requests dropped and the import inactive
892 * we are gaurenteed all reference will be dropped.
894 rc = __ldlm_namespace_free(ns, 1);
901 * Performs freeing memory structures related to \a ns. This is only done
902 * when ldlm_namespce_free_prior() successfully removed all resources
903 * referencing \a ns and its refc == 0.
905 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
913 /* Make sure that nobody can find this ns in its list. */
914 ldlm_namespace_unregister(ns, ns->ns_client);
915 /* Fini pool _before_ parent proc dir is removed. This is important as
916 * ldlm_pool_fini() removes own proc dir which is child to @dir.
917 * Removing it after @dir may cause oops. */
918 ldlm_pool_fini(&ns->ns_pool);
920 ldlm_namespace_proc_unregister(ns);
921 cfs_hash_putref(ns->ns_rs_hash);
922 /* Namespace \a ns should be not on list at this time, otherwise
923 * this will cause issues related to using freed \a ns in poold
925 LASSERT(list_empty(&ns->ns_list_chain));
932 * Cleanup the resource, and free namespace.
935 * proc1: destroy import
936 * class_disconnect_export(grab cl_sem) ->
937 * -> ldlm_namespace_free ->
938 * -> lprocfs_remove(grab _lprocfs_lock).
939 * proc2: read proc info
940 * lprocfs_fops_read(grab _lprocfs_lock) ->
941 * -> osc_rd_active, etc(grab cl_sem).
943 * So that I have to split the ldlm_namespace_free into two parts - the first
944 * part ldlm_namespace_free_prior is used to cleanup the resource which is
945 * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
946 * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
949 void ldlm_namespace_free(struct ldlm_namespace *ns,
950 struct obd_import *imp,
953 ldlm_namespace_free_prior(ns, imp, force);
954 ldlm_namespace_free_post(ns);
956 EXPORT_SYMBOL(ldlm_namespace_free);
958 void ldlm_namespace_get(struct ldlm_namespace *ns)
960 atomic_inc(&ns->ns_bref);
962 EXPORT_SYMBOL(ldlm_namespace_get);
964 /* This is only for callers that care about refcount */
965 int ldlm_namespace_get_return(struct ldlm_namespace *ns)
967 return atomic_inc_return(&ns->ns_bref);
970 void ldlm_namespace_put(struct ldlm_namespace *ns)
972 if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
973 wake_up(&ns->ns_waitq);
974 spin_unlock(&ns->ns_lock);
977 EXPORT_SYMBOL(ldlm_namespace_put);
979 /** Register \a ns in the list of namespaces */
980 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
982 mutex_lock(ldlm_namespace_lock(client));
983 LASSERT(list_empty(&ns->ns_list_chain));
984 list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
985 ldlm_namespace_nr_inc(client);
986 mutex_unlock(ldlm_namespace_lock(client));
989 /** Unregister \a ns from the list of namespaces. */
990 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
992 mutex_lock(ldlm_namespace_lock(client));
993 LASSERT(!list_empty(&ns->ns_list_chain));
994 /* Some asserts and possibly other parts of the code are still
995 * using list_empty(&ns->ns_list_chain). This is why it is
996 * important to use list_del_init() here. */
997 list_del_init(&ns->ns_list_chain);
998 ldlm_namespace_nr_dec(client);
999 mutex_unlock(ldlm_namespace_lock(client));
1002 /** Should be called with ldlm_namespace_lock(client) taken. */
1003 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1006 LASSERT(!list_empty(&ns->ns_list_chain));
1007 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1008 list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1011 /** Should be called with ldlm_namespace_lock(client) taken. */
1012 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1015 LASSERT(!list_empty(&ns->ns_list_chain));
1016 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1017 list_move_tail(&ns->ns_list_chain,
1018 ldlm_namespace_inactive_list(client));
1021 /** Should be called with ldlm_namespace_lock(client) taken. */
1022 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
1024 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1025 LASSERT(!list_empty(ldlm_namespace_list(client)));
1026 return container_of(ldlm_namespace_list(client)->next,
1027 struct ldlm_namespace, ns_list_chain);
1030 /** Create and initialize new resource. */
1031 static struct ldlm_resource *ldlm_resource_new(void)
1033 struct ldlm_resource *res;
1036 OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1040 INIT_LIST_HEAD(&res->lr_granted);
1041 INIT_LIST_HEAD(&res->lr_converting);
1042 INIT_LIST_HEAD(&res->lr_waiting);
1044 /* Initialize interval trees for each lock mode. */
1045 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1046 res->lr_itree[idx].lit_size = 0;
1047 res->lr_itree[idx].lit_mode = 1 << idx;
1048 res->lr_itree[idx].lit_root = NULL;
1051 atomic_set(&res->lr_refcount, 1);
1052 spin_lock_init(&res->lr_lock);
1053 lu_ref_init(&res->lr_reference);
1055 /* Since LVB init can be delayed now, there is no longer need to
1056 * immediatelly acquire mutex here. */
1057 mutex_init(&res->lr_lvb_mutex);
1058 res->lr_lvb_initialized = false;
1064 * Return a reference to resource with given name, creating it if necessary.
1065 * Args: namespace with ns_lock unlocked
1066 * Locks: takes and releases NS hash-lock and res->lr_lock
1067 * Returns: referenced, unlocked ldlm_resource or NULL
1069 struct ldlm_resource *
1070 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1071 const struct ldlm_res_id *name, ldlm_type_t type, int create)
1073 struct hlist_node *hnode;
1074 struct ldlm_resource *res = NULL;
1077 int ns_refcount = 0;
1079 LASSERT(ns != NULL);
1080 LASSERT(parent == NULL);
1081 LASSERT(ns->ns_rs_hash != NULL);
1082 LASSERT(name->name[0] != 0);
1084 cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1085 hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1086 if (hnode != NULL) {
1087 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1091 version = cfs_hash_bd_version_get(&bd);
1092 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1095 return ERR_PTR(-ENOENT);
1097 LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1098 "type: %d\n", type);
1099 res = ldlm_resource_new();
1101 return ERR_PTR(-ENOMEM);
1103 res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1104 res->lr_name = *name;
1105 res->lr_type = type;
1106 res->lr_most_restr = LCK_NL;
1108 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1109 hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1110 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1112 if (hnode != NULL) {
1113 /* Someone won the race and already added the resource. */
1114 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1115 /* Clean lu_ref for failed resource. */
1116 lu_ref_fini(&res->lr_reference);
1117 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1119 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1122 /* We won! Let's add the resource. */
1123 cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1124 if (cfs_hash_bd_count_get(&bd) == 1)
1125 ns_refcount = ldlm_namespace_get_return(ns);
1127 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1129 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1131 /* Let's see if we happened to be the very first resource in this
1132 * namespace. If so, and this is a client namespace, we need to move
1133 * the namespace into the active namespaces list to be patrolled by
1134 * the ldlm_poold. */
1135 if (ns_is_client(ns) && ns_refcount == 1) {
1136 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1137 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1138 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1143 EXPORT_SYMBOL(ldlm_resource_get);
1145 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1147 LASSERT(res != NULL);
1148 LASSERT(res != LP_POISON);
1149 atomic_inc(&res->lr_refcount);
1150 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1151 atomic_read(&res->lr_refcount));
1155 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1156 struct ldlm_resource *res)
1158 struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1160 if (!list_empty(&res->lr_granted)) {
1161 ldlm_resource_dump(D_ERROR, res);
1165 if (!list_empty(&res->lr_converting)) {
1166 ldlm_resource_dump(D_ERROR, res);
1170 if (!list_empty(&res->lr_waiting)) {
1171 ldlm_resource_dump(D_ERROR, res);
1175 cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1177 lu_ref_fini(&res->lr_reference);
1178 if (cfs_hash_bd_count_get(bd) == 0)
1179 ldlm_namespace_put(nsb->nsb_namespace);
1182 /* Returns 1 if the resource was freed, 0 if it remains. */
1183 int ldlm_resource_putref(struct ldlm_resource *res)
1185 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1188 LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1189 CDEBUG(D_INFO, "putref res: %p count: %d\n",
1190 res, atomic_read(&res->lr_refcount) - 1);
1192 cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1193 if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1194 __ldlm_resource_putref_final(&bd, res);
1195 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1196 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1197 ns->ns_lvbo->lvbo_free(res);
1198 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1203 EXPORT_SYMBOL(ldlm_resource_putref);
1205 /* Returns 1 if the resource was freed, 0 if it remains. */
1206 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1208 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1210 LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1211 CDEBUG(D_INFO, "putref res: %p count: %d\n",
1212 res, atomic_read(&res->lr_refcount) - 1);
1214 if (atomic_dec_and_test(&res->lr_refcount)) {
1217 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1218 &res->lr_name, &bd);
1219 __ldlm_resource_putref_final(&bd, res);
1220 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1221 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1222 * so we should never be here while calling cfs_hash_del,
1223 * cfs_hash_for_each_nolock is the only case we can get
1224 * here, which is safe to release cfs_hash_bd_lock.
1226 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1227 ns->ns_lvbo->lvbo_free(res);
1228 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1230 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1237 * Add a lock into a given resource into specified lock list.
1239 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1240 struct ldlm_lock *lock)
1242 check_res_locked(res);
1244 LDLM_DEBUG(lock, "About to add this lock:\n");
1246 if (ldlm_is_destroyed(lock)) {
1247 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1251 LASSERT(list_empty(&lock->l_res_link));
1253 list_add_tail(&lock->l_res_link, head);
1257 * Insert a lock into resource after specified lock.
1259 * Obtain resource description from the lock we are inserting after.
1261 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1262 struct ldlm_lock *new)
1264 struct ldlm_resource *res = original->l_resource;
1266 check_res_locked(res);
1268 ldlm_resource_dump(D_INFO, res);
1269 LDLM_DEBUG(new, "About to insert this lock after %p:\n", original);
1271 if (ldlm_is_destroyed(new)) {
1272 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1276 LASSERT(list_empty(&new->l_res_link));
1278 list_add(&new->l_res_link, &original->l_res_link);
1282 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1284 int type = lock->l_resource->lr_type;
1286 check_res_locked(lock->l_resource);
1287 if (type == LDLM_IBITS || type == LDLM_PLAIN)
1288 ldlm_unlink_lock_skiplist(lock);
1289 else if (type == LDLM_EXTENT)
1290 ldlm_extent_unlink_lock(lock);
1291 list_del_init(&lock->l_res_link);
1293 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1295 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1297 desc->lr_type = res->lr_type;
1298 desc->lr_name = res->lr_name;
1302 * Print information about all locks in all namespaces on this node to debug
1305 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1307 struct list_head *tmp;
1309 if (!((libcfs_debug | D_ERROR) & level))
1312 mutex_lock(ldlm_namespace_lock(client));
1314 list_for_each(tmp, ldlm_namespace_list(client)) {
1315 struct ldlm_namespace *ns;
1316 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1317 ldlm_namespace_dump(level, ns);
1320 mutex_unlock(ldlm_namespace_lock(client));
1322 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1324 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1325 struct hlist_node *hnode, void *arg)
1327 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1328 int level = (int)(unsigned long)arg;
1331 ldlm_resource_dump(level, res);
1338 * Print information about all locks in this namespace on this node to debug
1341 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1343 if (!((libcfs_debug | D_ERROR) & level))
1346 CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1347 ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1348 ns_is_client(ns) ? "client" : "server");
1350 if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1353 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1355 (void *)(unsigned long)level);
1356 spin_lock(&ns->ns_lock);
1357 ns->ns_next_dump = cfs_time_shift(10);
1358 spin_unlock(&ns->ns_lock);
1360 EXPORT_SYMBOL(ldlm_namespace_dump);
1363 * Print information about all locks in this resource to debug log.
1365 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1367 struct ldlm_lock *lock;
1368 unsigned int granted = 0;
1370 CLASSERT(RES_NAME_SIZE == 4);
1372 if (!((libcfs_debug | D_ERROR) & level))
1375 CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1376 PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1378 if (!list_empty(&res->lr_granted)) {
1379 CDEBUG(level, "Granted locks (in reverse order):\n");
1380 list_for_each_entry_reverse(lock, &res->lr_granted,
1382 LDLM_DEBUG_LIMIT(level, lock, "###");
1383 if (!(level & D_CANTMASK) &&
1384 ++granted > ldlm_dump_granted_max) {
1385 CDEBUG(level, "only dump %d granted locks to "
1386 "avoid DDOS.\n", granted);
1391 if (!list_empty(&res->lr_converting)) {
1392 CDEBUG(level, "Converting locks:\n");
1393 list_for_each_entry(lock, &res->lr_converting, l_res_link)
1394 LDLM_DEBUG_LIMIT(level, lock, "###");
1396 if (!list_empty(&res->lr_waiting)) {
1397 CDEBUG(level, "Waiting locks:\n");
1398 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1399 LDLM_DEBUG_LIMIT(level, lock, "###");
1402 EXPORT_SYMBOL(ldlm_resource_dump);