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, 2014, 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;
49 struct kmem_cache *ldlm_interval_tree_slab;
51 int ldlm_srv_namespace_nr = 0;
52 int ldlm_cli_namespace_nr = 0;
54 struct mutex ldlm_srv_namespace_lock;
55 struct list_head ldlm_srv_namespace_list;
57 struct mutex ldlm_cli_namespace_lock;
58 /* Client Namespaces that have active resources in them.
59 * Once all resources go away, ldlm_poold moves such namespaces to the
61 struct list_head ldlm_cli_active_namespace_list;
62 /* Client namespaces that don't have any locks in them */
63 struct list_head ldlm_cli_inactive_namespace_list;
65 static struct proc_dir_entry *ldlm_type_proc_dir;
66 static struct proc_dir_entry *ldlm_ns_proc_dir;
67 struct proc_dir_entry *ldlm_svc_proc_dir;
69 /* during debug dump certain amount of granted locks for one resource to avoid
71 static unsigned int ldlm_dump_granted_max = 256;
75 lprocfs_dump_ns_seq_write(struct file *file, const char __user *buffer,
76 size_t count, loff_t *off)
78 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
79 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
82 LPROC_SEQ_FOPS_WO_TYPE(ldlm, dump_ns);
84 LPROC_SEQ_FOPS_RW_TYPE(ldlm_rw, uint);
85 LPROC_SEQ_FOPS_RO_TYPE(ldlm, uint);
87 int ldlm_proc_setup(void)
90 struct lprocfs_vars list[] = {
91 { .name = "dump_namespaces",
92 .fops = &ldlm_dump_ns_fops,
94 { .name = "dump_granted_max",
95 .fops = &ldlm_rw_uint_fops,
96 .data = &ldlm_dump_granted_max },
97 { .name = "cancel_unused_locks_before_replay",
98 .fops = &ldlm_rw_uint_fops,
99 .data = &ldlm_cancel_unused_locks_before_replay },
102 LASSERT(ldlm_ns_proc_dir == NULL);
104 ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
107 if (IS_ERR(ldlm_type_proc_dir)) {
108 CERROR("LProcFS failed in ldlm-init\n");
109 rc = PTR_ERR(ldlm_type_proc_dir);
113 ldlm_ns_proc_dir = lprocfs_register("namespaces",
116 if (IS_ERR(ldlm_ns_proc_dir)) {
117 CERROR("LProcFS failed in ldlm-init\n");
118 rc = PTR_ERR(ldlm_ns_proc_dir);
122 ldlm_svc_proc_dir = lprocfs_register("services",
125 if (IS_ERR(ldlm_svc_proc_dir)) {
126 CERROR("LProcFS failed in ldlm-init\n");
127 rc = PTR_ERR(ldlm_svc_proc_dir);
131 rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
133 CERROR("LProcFS failed in ldlm-init\n");
140 lprocfs_remove(&ldlm_svc_proc_dir);
142 lprocfs_remove(&ldlm_ns_proc_dir);
144 lprocfs_remove(&ldlm_type_proc_dir);
146 ldlm_svc_proc_dir = NULL;
150 void ldlm_proc_cleanup(void)
152 if (ldlm_svc_proc_dir)
153 lprocfs_remove(&ldlm_svc_proc_dir);
155 if (ldlm_ns_proc_dir)
156 lprocfs_remove(&ldlm_ns_proc_dir);
158 if (ldlm_type_proc_dir)
159 lprocfs_remove(&ldlm_type_proc_dir);
162 static int lprocfs_ns_resources_seq_show(struct seq_file *m, void *v)
164 struct ldlm_namespace *ns = m->private;
166 struct cfs_hash_bd bd;
169 /* result is not strictly consistant */
170 cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
171 res += cfs_hash_bd_count_get(&bd);
172 return lprocfs_u64_seq_show(m, &res);
174 LPROC_SEQ_FOPS_RO(lprocfs_ns_resources);
176 static int lprocfs_ns_locks_seq_show(struct seq_file *m, void *v)
178 struct ldlm_namespace *ns = m->private;
181 locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
182 LPROCFS_FIELDS_FLAGS_SUM);
183 return lprocfs_u64_seq_show(m, &locks);
185 LPROC_SEQ_FOPS_RO(lprocfs_ns_locks);
187 static int lprocfs_lru_size_seq_show(struct seq_file *m, void *v)
189 struct ldlm_namespace *ns = m->private;
190 __u32 *nr = &ns->ns_max_unused;
192 if (ns_connect_lru_resize(ns))
193 nr = &ns->ns_nr_unused;
194 return lprocfs_uint_seq_show(m, nr);
197 static ssize_t lprocfs_lru_size_seq_write(struct file *file,
198 const char __user *buffer,
199 size_t count, loff_t *off)
201 struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
202 char dummy[MAX_STRING_SIZE + 1], *end;
206 dummy[MAX_STRING_SIZE] = '\0';
207 if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
210 if (strncmp(dummy, "clear", 5) == 0) {
212 "dropping all unused locks from namespace %s\n",
214 if (ns_connect_lru_resize(ns)) {
215 int canceled, unused = ns->ns_nr_unused;
217 /* Try to cancel all @ns_nr_unused locks. */
218 canceled = ldlm_cancel_lru(ns, unused, 0,
220 if (canceled < unused) {
222 "not all requested locks are canceled, "
223 "requested: %d, canceled: %d\n", unused,
228 tmp = ns->ns_max_unused;
229 ns->ns_max_unused = 0;
230 ldlm_cancel_lru(ns, 0, 0, LDLM_CANCEL_PASSED);
231 ns->ns_max_unused = tmp;
236 tmp = simple_strtoul(dummy, &end, 0);
238 CERROR("invalid value written\n");
241 lru_resize = (tmp == 0);
243 if (ns_connect_lru_resize(ns)) {
245 ns->ns_max_unused = (unsigned int)tmp;
247 if (tmp > ns->ns_nr_unused)
248 tmp = ns->ns_nr_unused;
249 tmp = ns->ns_nr_unused - tmp;
252 "changing namespace %s unused locks from %u to %u\n",
253 ldlm_ns_name(ns), ns->ns_nr_unused,
255 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_CANCEL_PASSED);
259 "disable lru_resize for namespace %s\n",
261 ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
265 "changing namespace %s max_unused from %u to %u\n",
266 ldlm_ns_name(ns), ns->ns_max_unused,
268 ns->ns_max_unused = (unsigned int)tmp;
269 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_CANCEL_PASSED);
271 /* Make sure that LRU resize was originally supported before
272 * turning it on here. */
274 (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
276 "enable lru_resize for namespace %s\n",
278 ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
284 LPROC_SEQ_FOPS(lprocfs_lru_size);
286 static int lprocfs_elc_seq_show(struct seq_file *m, void *v)
288 struct ldlm_namespace *ns = m->private;
289 unsigned int supp = ns_connect_cancelset(ns);
291 return lprocfs_uint_seq_show(m, &supp);
294 static ssize_t lprocfs_elc_seq_write(struct file *file,
295 const char __user *buffer,
296 size_t count, loff_t *off)
298 struct ldlm_namespace *ns = ((struct seq_file *)file->private_data)->private;
299 unsigned int supp = -1;
302 rc = lprocfs_wr_uint(file, buffer, count, &supp);
307 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
308 else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
309 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
312 LPROC_SEQ_FOPS(lprocfs_elc);
314 static void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
316 if (ns->ns_proc_dir_entry == NULL)
317 CERROR("dlm namespace %s has no procfs dir?\n",
320 lprocfs_remove(&ns->ns_proc_dir_entry);
322 if (ns->ns_stats != NULL)
323 lprocfs_free_stats(&ns->ns_stats);
326 static int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
328 struct lprocfs_vars lock_vars[2];
329 char lock_name[MAX_STRING_SIZE + 1];
330 struct proc_dir_entry *ns_pde;
333 LASSERT(ns->ns_rs_hash != NULL);
335 if (ns->ns_proc_dir_entry != NULL) {
336 ns_pde = ns->ns_proc_dir_entry;
338 ns_pde = proc_mkdir(ldlm_ns_name(ns), ldlm_ns_proc_dir);
341 ns->ns_proc_dir_entry = ns_pde;
344 ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
345 if (ns->ns_stats == NULL)
348 lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
349 LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
351 lock_name[MAX_STRING_SIZE] = '\0';
353 memset(lock_vars, 0, sizeof(lock_vars));
354 lock_vars[0].name = lock_name;
356 ldlm_add_var(&lock_vars[0], ns_pde, "resource_count", ns,
357 &lprocfs_ns_resources_fops);
358 ldlm_add_var(&lock_vars[0], ns_pde, "lock_count", ns,
359 &lprocfs_ns_locks_fops);
361 if (ns_is_client(ns)) {
362 ldlm_add_var(&lock_vars[0], ns_pde, "lock_unused_count",
363 &ns->ns_nr_unused, &ldlm_uint_fops);
364 ldlm_add_var(&lock_vars[0], ns_pde, "lru_size", ns,
365 &lprocfs_lru_size_fops);
366 ldlm_add_var(&lock_vars[0], ns_pde, "lru_max_age",
367 &ns->ns_max_age, &ldlm_rw_uint_fops);
368 ldlm_add_var(&lock_vars[0], ns_pde, "early_lock_cancel",
369 ns, &lprocfs_elc_fops);
371 ldlm_add_var(&lock_vars[0], ns_pde, "ctime_age_limit",
372 &ns->ns_ctime_age_limit, &ldlm_rw_uint_fops);
373 ldlm_add_var(&lock_vars[0], ns_pde, "lock_timeouts",
374 &ns->ns_timeouts, &ldlm_uint_fops);
375 ldlm_add_var(&lock_vars[0], ns_pde, "max_nolock_bytes",
376 &ns->ns_max_nolock_size, &ldlm_rw_uint_fops);
377 ldlm_add_var(&lock_vars[0], ns_pde, "contention_seconds",
378 &ns->ns_contention_time, &ldlm_rw_uint_fops);
379 ldlm_add_var(&lock_vars[0], ns_pde, "contended_locks",
380 &ns->ns_contended_locks, &ldlm_rw_uint_fops);
381 ldlm_add_var(&lock_vars[0], ns_pde, "max_parallel_ast",
382 &ns->ns_max_parallel_ast, &ldlm_rw_uint_fops);
386 #undef MAX_STRING_SIZE
387 #else /* CONFIG_PROC_FS */
389 #define ldlm_namespace_proc_unregister(ns) ({;})
390 #define ldlm_namespace_proc_register(ns) ({0;})
392 #endif /* CONFIG_PROC_FS */
394 static unsigned ldlm_res_hop_hash(struct cfs_hash *hs,
395 const void *key, unsigned mask)
397 const struct ldlm_res_id *id = key;
401 for (i = 0; i < RES_NAME_SIZE; i++)
406 static unsigned ldlm_res_hop_fid_hash(struct cfs_hash *hs,
407 const void *key, unsigned mask)
409 const struct ldlm_res_id *id = key;
414 fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
415 fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
416 fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
418 hash = fid_flatten32(&fid);
419 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
420 if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
421 val = id->name[LUSTRE_RES_ID_HSH_OFF];
422 hash += (val >> 5) + (val << 11);
426 hash = hash_long(hash, hs->hs_bkt_bits);
427 /* give me another random factor */
428 hash -= hash_long((unsigned long)hs, val % 11 + 3);
430 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
431 hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
436 static void *ldlm_res_hop_key(struct hlist_node *hnode)
438 struct ldlm_resource *res;
440 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
441 return &res->lr_name;
444 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
446 struct ldlm_resource *res;
448 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
449 return ldlm_res_eq((const struct ldlm_res_id *)key,
450 (const struct ldlm_res_id *)&res->lr_name);
453 static void *ldlm_res_hop_object(struct hlist_node *hnode)
455 return hlist_entry(hnode, struct ldlm_resource, lr_hash);
459 ldlm_res_hop_get_locked(struct cfs_hash *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);
468 ldlm_res_hop_put_locked(struct cfs_hash *hs, struct hlist_node *hnode)
470 struct ldlm_resource *res;
472 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
473 /* cfs_hash_for_each_nolock is the only chance we call it */
474 ldlm_resource_putref_locked(res);
477 static void ldlm_res_hop_put(struct cfs_hash *hs, struct hlist_node *hnode)
479 struct ldlm_resource *res;
481 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
482 ldlm_resource_putref(res);
485 static struct cfs_hash_ops ldlm_ns_hash_ops = {
486 .hs_hash = ldlm_res_hop_hash,
487 .hs_key = ldlm_res_hop_key,
488 .hs_keycmp = ldlm_res_hop_keycmp,
490 .hs_object = ldlm_res_hop_object,
491 .hs_get = ldlm_res_hop_get_locked,
492 .hs_put_locked = ldlm_res_hop_put_locked,
493 .hs_put = ldlm_res_hop_put
496 static struct cfs_hash_ops ldlm_ns_fid_hash_ops = {
497 .hs_hash = ldlm_res_hop_fid_hash,
498 .hs_key = ldlm_res_hop_key,
499 .hs_keycmp = ldlm_res_hop_keycmp,
501 .hs_object = ldlm_res_hop_object,
502 .hs_get = ldlm_res_hop_get_locked,
503 .hs_put_locked = ldlm_res_hop_put_locked,
504 .hs_put = ldlm_res_hop_put
508 ldlm_ns_type_t nsd_type;
509 /** hash bucket bits */
510 unsigned nsd_bkt_bits;
512 unsigned nsd_all_bits;
513 /** hash operations */
514 struct cfs_hash_ops *nsd_hops;
515 } ldlm_ns_hash_def_t;
517 static ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
520 .nsd_type = LDLM_NS_TYPE_MDC,
523 .nsd_hops = &ldlm_ns_fid_hash_ops,
526 .nsd_type = LDLM_NS_TYPE_MDT,
529 .nsd_hops = &ldlm_ns_fid_hash_ops,
532 .nsd_type = LDLM_NS_TYPE_OSC,
535 .nsd_hops = &ldlm_ns_hash_ops,
538 .nsd_type = LDLM_NS_TYPE_OST,
541 .nsd_hops = &ldlm_ns_hash_ops,
544 .nsd_type = LDLM_NS_TYPE_MGC,
547 .nsd_hops = &ldlm_ns_hash_ops,
550 .nsd_type = LDLM_NS_TYPE_MGT,
553 .nsd_hops = &ldlm_ns_hash_ops,
556 .nsd_type = LDLM_NS_TYPE_UNKNOWN,
561 * Create and initialize new empty namespace.
563 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
566 ldlm_ns_type_t ns_type)
568 struct ldlm_namespace *ns = NULL;
569 struct ldlm_ns_bucket *nsb;
570 ldlm_ns_hash_def_t *nsd;
571 struct cfs_hash_bd bd;
576 LASSERT(obd != NULL);
580 CERROR("ldlm_get_ref failed: %d\n", rc);
584 for (idx = 0;;idx++) {
585 nsd = &ldlm_ns_hash_defs[idx];
586 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
587 CERROR("Unknown type %d for ns %s\n", ns_type, name);
591 if (nsd->nsd_type == ns_type)
599 ns->ns_rs_hash = cfs_hash_create(name,
600 nsd->nsd_all_bits, nsd->nsd_all_bits,
601 nsd->nsd_bkt_bits, sizeof(*nsb),
607 CFS_HASH_SPIN_BKTLOCK |
608 CFS_HASH_NO_ITEMREF);
609 if (ns->ns_rs_hash == NULL)
612 cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
613 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
614 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
615 nsb->nsb_namespace = ns;
619 ns->ns_appetite = apt;
620 ns->ns_client = client;
622 INIT_LIST_HEAD(&ns->ns_list_chain);
623 INIT_LIST_HEAD(&ns->ns_unused_list);
624 spin_lock_init(&ns->ns_lock);
625 atomic_set(&ns->ns_bref, 0);
626 init_waitqueue_head(&ns->ns_waitq);
628 ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
629 ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
630 ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
632 ns->ns_max_parallel_ast = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
633 ns->ns_nr_unused = 0;
634 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
635 ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
636 ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
638 ns->ns_orig_connect_flags = 0;
639 ns->ns_connect_flags = 0;
641 rc = ldlm_namespace_proc_register(ns);
643 CERROR("Can't initialize ns proc, rc %d\n", rc);
647 idx = ldlm_namespace_nr_read(client);
648 rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
650 CERROR("Can't initialize lock pool, rc %d\n", rc);
654 ldlm_namespace_register(ns, client);
657 ldlm_namespace_proc_unregister(ns);
658 ldlm_namespace_cleanup(ns, 0);
660 cfs_hash_putref(ns->ns_rs_hash);
667 EXPORT_SYMBOL(ldlm_namespace_new);
669 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
672 * Cancel and destroy all locks on a resource.
674 * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
675 * clean up. This is currently only used for recovery, and we make
676 * certain assumptions as a result--notably, that we shouldn't cancel
679 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
682 struct list_head *tmp;
683 int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
684 bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
687 struct ldlm_lock *lock = NULL;
689 /* First, we look for non-cleaned-yet lock
690 * all cleaned locks are marked by CLEANED flag. */
692 list_for_each(tmp, q) {
693 lock = list_entry(tmp, struct ldlm_lock,
695 if (ldlm_is_cleaned(lock)) {
700 ldlm_set_cleaned(lock);
709 /* Set CBPENDING so nothing in the cancellation path
710 * can match this lock. */
711 ldlm_set_cbpending(lock);
712 ldlm_set_failed(lock);
713 lock->l_flags |= flags;
715 /* ... without sending a CANCEL message for local_only. */
717 ldlm_set_local_only(lock);
719 if (local_only && (lock->l_readers || lock->l_writers)) {
720 /* This is a little bit gross, but much better than the
721 * alternative: pretend that we got a blocking AST from
722 * the server, so that when the lock is decref'd, it
723 * will go away ... */
725 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
726 if (lock->l_flags & LDLM_FL_FAIL_LOC) {
727 set_current_state(TASK_UNINTERRUPTIBLE);
728 schedule_timeout(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(struct cfs_hash *hs, struct cfs_hash_bd *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(struct cfs_hash *hs, struct cfs_hash_bd *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);
963 /* This is only for callers that care about refcount */
964 static int ldlm_namespace_get_return(struct ldlm_namespace *ns)
966 return atomic_inc_return(&ns->ns_bref);
969 void ldlm_namespace_put(struct ldlm_namespace *ns)
971 if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
972 wake_up(&ns->ns_waitq);
973 spin_unlock(&ns->ns_lock);
977 /** Register \a ns in the list of namespaces */
978 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
980 mutex_lock(ldlm_namespace_lock(client));
981 LASSERT(list_empty(&ns->ns_list_chain));
982 list_add(&ns->ns_list_chain, ldlm_namespace_inactive_list(client));
983 ldlm_namespace_nr_inc(client);
984 mutex_unlock(ldlm_namespace_lock(client));
987 /** Unregister \a ns from the list of namespaces. */
988 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
990 mutex_lock(ldlm_namespace_lock(client));
991 LASSERT(!list_empty(&ns->ns_list_chain));
992 /* Some asserts and possibly other parts of the code are still
993 * using list_empty(&ns->ns_list_chain). This is why it is
994 * important to use list_del_init() here. */
995 list_del_init(&ns->ns_list_chain);
996 ldlm_namespace_nr_dec(client);
997 mutex_unlock(ldlm_namespace_lock(client));
1000 /** Should be called with ldlm_namespace_lock(client) taken. */
1001 void ldlm_namespace_move_to_active_locked(struct ldlm_namespace *ns,
1004 LASSERT(!list_empty(&ns->ns_list_chain));
1005 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1006 list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1009 /** Should be called with ldlm_namespace_lock(client) taken. */
1010 void ldlm_namespace_move_to_inactive_locked(struct ldlm_namespace *ns,
1013 LASSERT(!list_empty(&ns->ns_list_chain));
1014 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1015 list_move_tail(&ns->ns_list_chain,
1016 ldlm_namespace_inactive_list(client));
1019 /** Should be called with ldlm_namespace_lock(client) taken. */
1020 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
1022 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1023 LASSERT(!list_empty(ldlm_namespace_list(client)));
1024 return container_of(ldlm_namespace_list(client)->next,
1025 struct ldlm_namespace, ns_list_chain);
1028 /** Create and initialize new resource. */
1029 static struct ldlm_resource *ldlm_resource_new(ldlm_type_t type)
1031 struct ldlm_resource *res;
1034 OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, GFP_NOFS);
1038 if (type == LDLM_EXTENT) {
1039 OBD_SLAB_ALLOC(res->lr_itree, ldlm_interval_tree_slab,
1040 sizeof(*res->lr_itree) * LCK_MODE_NUM);
1041 if (res->lr_itree == NULL) {
1042 OBD_SLAB_FREE_PTR(res, ldlm_resource_slab);
1045 /* Initialize interval trees for each lock mode. */
1046 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1047 res->lr_itree[idx].lit_size = 0;
1048 res->lr_itree[idx].lit_mode = 1 << idx;
1049 res->lr_itree[idx].lit_root = NULL;
1053 INIT_LIST_HEAD(&res->lr_granted);
1054 INIT_LIST_HEAD(&res->lr_converting);
1055 INIT_LIST_HEAD(&res->lr_waiting);
1057 atomic_set(&res->lr_refcount, 1);
1058 spin_lock_init(&res->lr_lock);
1059 lu_ref_init(&res->lr_reference);
1061 /* Since LVB init can be delayed now, there is no longer need to
1062 * immediatelly acquire mutex here. */
1063 mutex_init(&res->lr_lvb_mutex);
1064 res->lr_lvb_initialized = false;
1070 * Return a reference to resource with given name, creating it if necessary.
1071 * Args: namespace with ns_lock unlocked
1072 * Locks: takes and releases NS hash-lock and res->lr_lock
1073 * Returns: referenced, unlocked ldlm_resource or NULL
1075 struct ldlm_resource *
1076 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1077 const struct ldlm_res_id *name, ldlm_type_t type, int create)
1079 struct hlist_node *hnode;
1080 struct ldlm_resource *res = NULL;
1081 struct cfs_hash_bd bd;
1083 int ns_refcount = 0;
1085 LASSERT(ns != NULL);
1086 LASSERT(parent == NULL);
1087 LASSERT(ns->ns_rs_hash != NULL);
1088 LASSERT(name->name[0] != 0);
1090 cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1091 hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1092 if (hnode != NULL) {
1093 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1097 version = cfs_hash_bd_version_get(&bd);
1098 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1101 return ERR_PTR(-ENOENT);
1103 LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1104 "type: %d\n", type);
1105 res = ldlm_resource_new(type);
1107 return ERR_PTR(-ENOMEM);
1109 res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1110 res->lr_name = *name;
1111 res->lr_type = type;
1113 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1114 hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1115 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1117 if (hnode != NULL) {
1118 /* Someone won the race and already added the resource. */
1119 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1120 /* Clean lu_ref for failed resource. */
1121 lu_ref_fini(&res->lr_reference);
1122 if (res->lr_itree != NULL)
1123 OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1124 sizeof(*res->lr_itree) * LCK_MODE_NUM);
1125 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1127 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1130 /* We won! Let's add the resource. */
1131 cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1132 if (cfs_hash_bd_count_get(&bd) == 1)
1133 ns_refcount = ldlm_namespace_get_return(ns);
1135 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1137 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1139 /* Let's see if we happened to be the very first resource in this
1140 * namespace. If so, and this is a client namespace, we need to move
1141 * the namespace into the active namespaces list to be patrolled by
1142 * the ldlm_poold. */
1143 if (ns_is_client(ns) && ns_refcount == 1) {
1144 mutex_lock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1145 ldlm_namespace_move_to_active_locked(ns, LDLM_NAMESPACE_CLIENT);
1146 mutex_unlock(ldlm_namespace_lock(LDLM_NAMESPACE_CLIENT));
1151 EXPORT_SYMBOL(ldlm_resource_get);
1153 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1155 LASSERT(res != NULL);
1156 LASSERT(res != LP_POISON);
1157 atomic_inc(&res->lr_refcount);
1158 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1159 atomic_read(&res->lr_refcount));
1163 static void __ldlm_resource_putref_final(struct cfs_hash_bd *bd,
1164 struct ldlm_resource *res)
1166 struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1168 if (!list_empty(&res->lr_granted)) {
1169 ldlm_resource_dump(D_ERROR, res);
1173 if (!list_empty(&res->lr_converting)) {
1174 ldlm_resource_dump(D_ERROR, res);
1178 if (!list_empty(&res->lr_waiting)) {
1179 ldlm_resource_dump(D_ERROR, res);
1183 cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1185 lu_ref_fini(&res->lr_reference);
1186 if (cfs_hash_bd_count_get(bd) == 0)
1187 ldlm_namespace_put(nsb->nsb_namespace);
1190 /* Returns 1 if the resource was freed, 0 if it remains. */
1191 int ldlm_resource_putref(struct ldlm_resource *res)
1193 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1194 struct cfs_hash_bd bd;
1196 LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1197 CDEBUG(D_INFO, "putref res: %p count: %d\n",
1198 res, atomic_read(&res->lr_refcount) - 1);
1200 cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1201 if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1202 __ldlm_resource_putref_final(&bd, res);
1203 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1204 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1205 ns->ns_lvbo->lvbo_free(res);
1206 if (res->lr_itree != NULL)
1207 OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1208 sizeof(*res->lr_itree) * LCK_MODE_NUM);
1209 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1214 EXPORT_SYMBOL(ldlm_resource_putref);
1216 /* Returns 1 if the resource was freed, 0 if it remains. */
1217 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1219 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1221 LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1222 CDEBUG(D_INFO, "putref res: %p count: %d\n",
1223 res, atomic_read(&res->lr_refcount) - 1);
1225 if (atomic_dec_and_test(&res->lr_refcount)) {
1226 struct cfs_hash_bd bd;
1228 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1229 &res->lr_name, &bd);
1230 __ldlm_resource_putref_final(&bd, res);
1231 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1232 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1233 * so we should never be here while calling cfs_hash_del,
1234 * cfs_hash_for_each_nolock is the only case we can get
1235 * here, which is safe to release cfs_hash_bd_lock.
1237 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1238 ns->ns_lvbo->lvbo_free(res);
1239 if (res->lr_itree != NULL)
1240 OBD_SLAB_FREE(res->lr_itree, ldlm_interval_tree_slab,
1241 sizeof(*res->lr_itree) * LCK_MODE_NUM);
1242 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1244 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1251 * Add a lock into a given resource into specified lock list.
1253 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1254 struct ldlm_lock *lock)
1256 check_res_locked(res);
1258 LDLM_DEBUG(lock, "About to add this lock:\n");
1260 if (ldlm_is_destroyed(lock)) {
1261 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1265 LASSERT(list_empty(&lock->l_res_link));
1267 list_add_tail(&lock->l_res_link, head);
1271 * Insert a lock into resource after specified lock.
1273 * Obtain resource description from the lock we are inserting after.
1275 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1276 struct ldlm_lock *new)
1278 struct ldlm_resource *res = original->l_resource;
1280 check_res_locked(res);
1282 ldlm_resource_dump(D_INFO, res);
1283 LDLM_DEBUG(new, "About to insert this lock after %p:\n", original);
1285 if (ldlm_is_destroyed(new)) {
1286 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1290 LASSERT(list_empty(&new->l_res_link));
1292 list_add(&new->l_res_link, &original->l_res_link);
1296 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1298 int type = lock->l_resource->lr_type;
1300 check_res_locked(lock->l_resource);
1301 if (type == LDLM_IBITS || type == LDLM_PLAIN)
1302 ldlm_unlink_lock_skiplist(lock);
1303 else if (type == LDLM_EXTENT)
1304 ldlm_extent_unlink_lock(lock);
1305 list_del_init(&lock->l_res_link);
1307 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1309 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1311 desc->lr_type = res->lr_type;
1312 desc->lr_name = res->lr_name;
1316 * Print information about all locks in all namespaces on this node to debug
1319 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1321 struct list_head *tmp;
1323 if (!((libcfs_debug | D_ERROR) & level))
1326 mutex_lock(ldlm_namespace_lock(client));
1328 list_for_each(tmp, ldlm_namespace_list(client)) {
1329 struct ldlm_namespace *ns;
1330 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1331 ldlm_namespace_dump(level, ns);
1334 mutex_unlock(ldlm_namespace_lock(client));
1337 static int ldlm_res_hash_dump(struct cfs_hash *hs, struct cfs_hash_bd *bd,
1338 struct hlist_node *hnode, void *arg)
1340 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1341 int level = (int)(unsigned long)arg;
1344 ldlm_resource_dump(level, res);
1351 * Print information about all locks in this namespace on this node to debug
1354 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1356 if (!((libcfs_debug | D_ERROR) & level))
1359 CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1360 ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1361 ns_is_client(ns) ? "client" : "server");
1363 if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1366 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1368 (void *)(unsigned long)level);
1369 spin_lock(&ns->ns_lock);
1370 ns->ns_next_dump = cfs_time_shift(10);
1371 spin_unlock(&ns->ns_lock);
1375 * Print information about all locks in this resource to debug log.
1377 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1379 struct ldlm_lock *lock;
1380 unsigned int granted = 0;
1382 CLASSERT(RES_NAME_SIZE == 4);
1384 if (!((libcfs_debug | D_ERROR) & level))
1387 CDEBUG(level, "--- Resource: "DLDLMRES" (%p) refcount = %d\n",
1388 PLDLMRES(res), res, atomic_read(&res->lr_refcount));
1390 if (!list_empty(&res->lr_granted)) {
1391 CDEBUG(level, "Granted locks (in reverse order):\n");
1392 list_for_each_entry_reverse(lock, &res->lr_granted,
1394 LDLM_DEBUG_LIMIT(level, lock, "###");
1395 if (!(level & D_CANTMASK) &&
1396 ++granted > ldlm_dump_granted_max) {
1397 CDEBUG(level, "only dump %d granted locks to "
1398 "avoid DDOS.\n", granted);
1403 if (!list_empty(&res->lr_converting)) {
1404 CDEBUG(level, "Converting locks:\n");
1405 list_for_each_entry(lock, &res->lr_converting, l_res_link)
1406 LDLM_DEBUG_LIMIT(level, lock, "###");
1408 if (!list_empty(&res->lr_waiting)) {
1409 CDEBUG(level, "Waiting locks:\n");
1410 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1411 LDLM_DEBUG_LIMIT(level, lock, "###");
1414 EXPORT_SYMBOL(ldlm_resource_dump);