1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
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
44 # include <lustre_dlm.h>
46 # include <liblustre.h>
49 #include <obd_class.h>
50 #include "ldlm_internal.h"
52 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
54 cfs_atomic_t ldlm_srv_namespace_nr = CFS_ATOMIC_INIT(0);
55 cfs_atomic_t ldlm_cli_namespace_nr = CFS_ATOMIC_INIT(0);
57 cfs_semaphore_t ldlm_srv_namespace_lock;
58 CFS_LIST_HEAD(ldlm_srv_namespace_list);
60 cfs_semaphore_t ldlm_cli_namespace_lock;
61 CFS_LIST_HEAD(ldlm_cli_namespace_list);
63 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
64 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
65 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
68 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
69 unsigned long count, void *data)
71 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
72 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
76 int ldlm_proc_setup(void)
79 struct lprocfs_vars list[] = {
80 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
83 LASSERT(ldlm_ns_proc_dir == NULL);
85 ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
88 if (IS_ERR(ldlm_type_proc_dir)) {
89 CERROR("LProcFS failed in ldlm-init\n");
90 rc = PTR_ERR(ldlm_type_proc_dir);
94 ldlm_ns_proc_dir = lprocfs_register("namespaces",
97 if (IS_ERR(ldlm_ns_proc_dir)) {
98 CERROR("LProcFS failed in ldlm-init\n");
99 rc = PTR_ERR(ldlm_ns_proc_dir);
103 ldlm_svc_proc_dir = lprocfs_register("services",
106 if (IS_ERR(ldlm_svc_proc_dir)) {
107 CERROR("LProcFS failed in ldlm-init\n");
108 rc = PTR_ERR(ldlm_svc_proc_dir);
112 rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
117 lprocfs_remove(&ldlm_ns_proc_dir);
119 lprocfs_remove(&ldlm_type_proc_dir);
121 ldlm_svc_proc_dir = NULL;
125 void ldlm_proc_cleanup(void)
127 if (ldlm_svc_proc_dir)
128 lprocfs_remove(&ldlm_svc_proc_dir);
130 if (ldlm_ns_proc_dir)
131 lprocfs_remove(&ldlm_ns_proc_dir);
133 if (ldlm_type_proc_dir)
134 lprocfs_remove(&ldlm_type_proc_dir);
137 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
138 int count, int *eof, void *data)
140 struct ldlm_namespace *ns = data;
141 __u32 *nr = &ns->ns_max_unused;
143 if (ns_connect_lru_resize(ns))
144 nr = &ns->ns_nr_unused;
145 return lprocfs_rd_uint(page, start, off, count, eof, nr);
148 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
149 unsigned long count, void *data)
151 struct ldlm_namespace *ns = data;
152 char dummy[MAX_STRING_SIZE + 1], *end;
156 dummy[MAX_STRING_SIZE] = '\0';
157 if (cfs_copy_from_user(dummy, buffer, MAX_STRING_SIZE))
160 if (strncmp(dummy, "clear", 5) == 0) {
162 "dropping all unused locks from namespace %s\n",
164 if (ns_connect_lru_resize(ns)) {
165 int canceled, unused = ns->ns_nr_unused;
167 /* Try to cancel all @ns_nr_unused locks. */
168 canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
170 if (canceled < unused) {
172 "not all requested locks are canceled, "
173 "requested: %d, canceled: %d\n", unused,
178 tmp = ns->ns_max_unused;
179 ns->ns_max_unused = 0;
180 ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
181 ns->ns_max_unused = tmp;
186 tmp = simple_strtoul(dummy, &end, 0);
188 CERROR("invalid value written\n");
191 lru_resize = (tmp == 0);
193 if (ns_connect_lru_resize(ns)) {
195 ns->ns_max_unused = (unsigned int)tmp;
197 if (tmp > ns->ns_nr_unused)
198 tmp = ns->ns_nr_unused;
199 tmp = ns->ns_nr_unused - tmp;
202 "changing namespace %s unused locks from %u to %u\n",
203 ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
204 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
208 "disable lru_resize for namespace %s\n",
210 ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
214 "changing namespace %s max_unused from %u to %u\n",
215 ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
216 ns->ns_max_unused = (unsigned int)tmp;
217 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
219 /* Make sure that originally lru resize was supported before
220 * turning it on here. */
222 (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
224 "enable lru_resize for namespace %s\n",
226 ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
233 void ldlm_proc_namespace(struct ldlm_namespace *ns)
235 struct lprocfs_vars lock_vars[2];
236 char lock_name[MAX_STRING_SIZE + 1];
239 LASSERT(ns->ns_name != NULL);
241 lock_name[MAX_STRING_SIZE] = '\0';
243 memset(lock_vars, 0, sizeof(lock_vars));
244 lock_vars[0].name = lock_name;
246 snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
247 lock_vars[0].data = &ns->ns_refcount;
248 lock_vars[0].read_fptr = lprocfs_rd_atomic;
249 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
251 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
252 lock_vars[0].data = &ns->ns_locks;
253 lock_vars[0].read_fptr = lprocfs_rd_atomic;
254 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
256 if (ns_is_client(ns)) {
257 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
259 lock_vars[0].data = &ns->ns_nr_unused;
260 lock_vars[0].read_fptr = lprocfs_rd_uint;
261 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
263 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
265 lock_vars[0].data = ns;
266 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
267 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
268 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
270 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
272 lock_vars[0].data = &ns->ns_max_age;
273 lock_vars[0].read_fptr = lprocfs_rd_uint;
274 lock_vars[0].write_fptr = lprocfs_wr_uint;
275 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
277 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
279 lock_vars[0].data = &ns->ns_ctime_age_limit;
280 lock_vars[0].read_fptr = lprocfs_rd_uint;
281 lock_vars[0].write_fptr = lprocfs_wr_uint;
282 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
284 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
286 lock_vars[0].data = &ns->ns_timeouts;
287 lock_vars[0].read_fptr = lprocfs_rd_uint;
288 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
290 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
292 lock_vars[0].data = &ns->ns_max_nolock_size;
293 lock_vars[0].read_fptr = lprocfs_rd_uint;
294 lock_vars[0].write_fptr = lprocfs_wr_uint;
295 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
297 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
299 lock_vars[0].data = &ns->ns_contention_time;
300 lock_vars[0].read_fptr = lprocfs_rd_uint;
301 lock_vars[0].write_fptr = lprocfs_wr_uint;
302 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
304 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
306 lock_vars[0].data = &ns->ns_contended_locks;
307 lock_vars[0].read_fptr = lprocfs_rd_uint;
308 lock_vars[0].write_fptr = lprocfs_wr_uint;
309 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
312 #undef MAX_STRING_SIZE
314 #define ldlm_proc_namespace(ns) do {} while (0)
317 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
318 ldlm_side_t client, ldlm_appetite_t apt)
320 struct ldlm_namespace *ns = NULL;
322 int rc, idx, namelen;
327 CERROR("ldlm_get_ref failed: %d\n", rc);
335 OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
339 ns->ns_appetite = apt;
341 LASSERT(obd != NULL);
344 namelen = strlen(name);
345 OBD_ALLOC(ns->ns_name, namelen + 1);
347 GOTO(out_hash, NULL);
349 strcpy(ns->ns_name, name);
351 CFS_INIT_LIST_HEAD(&ns->ns_root_list);
352 CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
354 ns->ns_client = client;
355 cfs_spin_lock_init(&ns->ns_hash_lock);
356 cfs_atomic_set(&ns->ns_locks, 0);
357 ns->ns_resources = 0;
358 cfs_waitq_init(&ns->ns_waitq);
359 ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
360 ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
361 ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
363 for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
365 CFS_INIT_LIST_HEAD(bucket);
367 CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
368 ns->ns_nr_unused = 0;
369 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
370 ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
371 ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
373 cfs_spin_lock_init(&ns->ns_unused_lock);
374 ns->ns_orig_connect_flags = 0;
375 ns->ns_connect_flags = 0;
376 ldlm_proc_namespace(ns);
378 idx = cfs_atomic_read(ldlm_namespace_nr(client));
379 rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
381 CERROR("Can't initialize lock pool, rc %d\n", rc);
385 at_init(&ns->ns_at_estimate, ldlm_enqueue_min, 0);
387 ldlm_namespace_register(ns, client);
390 ldlm_namespace_cleanup(ns, 0);
391 OBD_FREE(ns->ns_name, namelen + 1);
393 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
401 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
403 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
404 * This is currently only used for recovery, and we make certain assumptions
405 * as a result--notably, that we shouldn't cancel locks with refs. -phil
407 * Called with the ns_lock held. */
408 static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q,
412 int rc = 0, client = ns_is_client(res->lr_namespace);
413 int local_only = (flags & LDLM_FL_LOCAL_ONLY);
418 struct ldlm_lock *lock = NULL;
420 /* first, we look for non-cleaned-yet lock
421 * all cleaned locks are marked by CLEANED flag */
423 cfs_list_for_each(tmp, q) {
424 lock = cfs_list_entry(tmp, struct ldlm_lock,
426 if (lock->l_flags & LDLM_FL_CLEANED) {
431 lock->l_flags |= LDLM_FL_CLEANED;
440 /* Set CBPENDING so nothing in the cancellation path
441 * can match this lock */
442 lock->l_flags |= LDLM_FL_CBPENDING;
443 lock->l_flags |= LDLM_FL_FAILED;
444 lock->l_flags |= flags;
446 /* ... without sending a CANCEL message for local_only. */
448 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
450 if (local_only && (lock->l_readers || lock->l_writers)) {
451 /* This is a little bit gross, but much better than the
452 * alternative: pretend that we got a blocking AST from
453 * the server, so that when the lock is decref'd, it
454 * will go away ... */
456 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
457 if (lock->l_completion_ast)
458 lock->l_completion_ast(lock, 0, NULL);
459 LDLM_LOCK_RELEASE(lock);
464 struct lustre_handle lockh;
467 ldlm_lock2handle(lock, &lockh);
468 rc = ldlm_cli_cancel(&lockh);
470 CERROR("ldlm_cli_cancel: %d\n", rc);
472 ldlm_resource_unlink_lock(lock);
474 LDLM_DEBUG(lock, "Freeing a lock still held by a "
476 ldlm_lock_destroy(lock);
478 LDLM_LOCK_RELEASE(lock);
484 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
490 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
494 for (i = 0; i < RES_HASH_SIZE; i++) {
495 cfs_spin_lock(&ns->ns_hash_lock);
496 tmp = ns->ns_hash[i].next;
497 while (tmp != &(ns->ns_hash[i])) {
498 struct ldlm_resource *res;
499 res = cfs_list_entry(tmp, struct ldlm_resource,
501 ldlm_resource_getref(res);
502 cfs_spin_unlock(&ns->ns_hash_lock);
503 LDLM_RESOURCE_ADDREF(res);
505 cleanup_resource(res, &res->lr_granted, flags);
506 cleanup_resource(res, &res->lr_converting, flags);
507 cleanup_resource(res, &res->lr_waiting, flags);
509 cfs_spin_lock(&ns->ns_hash_lock);
512 /* XXX: former stuff caused issues in case of race
513 * between ldlm_namespace_cleanup() and lockd() when
514 * client gets blocking ast when lock gets distracted by
515 * server. This is 1_4 branch solution, let's see how
517 LDLM_RESOURCE_DELREF(res);
518 if (!ldlm_resource_putref_locked(res)) {
519 CERROR("Namespace %s resource refcount nonzero "
520 "(%d) after lock cleanup; forcing "
523 cfs_atomic_read(&res->lr_refcount));
524 CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
525 LPU64") (rc: %d)\n", res,
526 res->lr_name.name[0],
527 res->lr_name.name[1],
528 res->lr_name.name[2],
529 res->lr_name.name[3],
530 cfs_atomic_read(&res->lr_refcount));
533 cfs_spin_unlock(&ns->ns_hash_lock);
539 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
543 /* At shutdown time, don't call the cancellation callback */
544 ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
546 if (ns->ns_refcount > 0) {
547 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
550 "dlm namespace %s free waiting on refcount %d\n",
551 ns->ns_name, ns->ns_refcount);
554 lwi = LWI_TIMEOUT(obd_timeout * CFS_HZ / 4, NULL, NULL);
556 rc = l_wait_event(ns->ns_waitq,
557 ns->ns_refcount == 0, &lwi);
559 /* Forced cleanups should be able to reclaim all references,
560 * so it's safe to wait forever... we can't leak locks... */
561 if (force && rc == -ETIMEDOUT) {
562 LCONSOLE_ERROR("Forced cleanup waiting for %s "
563 "namespace with %d resources in use, "
564 "(rc=%d)\n", ns->ns_name,
565 ns->ns_refcount, rc);
566 GOTO(force_wait, rc);
569 if (ns->ns_refcount) {
570 LCONSOLE_ERROR("Cleanup waiting for %s namespace "
571 "with %d resources in use, (rc=%d)\n",
573 ns->ns_refcount, rc);
574 RETURN(ELDLM_NAMESPACE_EXISTS);
577 "dlm namespace %s free done waiting\n", ns->ns_name);
584 * Performs various cleanups for passed \a ns to make it drop refc and be ready
585 * for freeing. Waits for refc == 0.
587 * The following is done:
588 * (0) Unregister \a ns from its list to make inaccessible for potential users
589 * like pools thread and others;
590 * (1) Clear all locks in \a ns.
592 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
593 struct obd_import *imp,
605 * Can fail with -EINTR when force == 0 in which case try harder.
607 rc = __ldlm_namespace_free(ns, force);
608 if (rc != ELDLM_OK) {
610 ptlrpc_disconnect_import(imp, 0);
611 ptlrpc_invalidate_import(imp);
615 * With all requests dropped and the import inactive
616 * we are gaurenteed all reference will be dropped.
618 rc = __ldlm_namespace_free(ns, 1);
625 * Performs freeing memory structures related to \a ns. This is only done when
626 * ldlm_namespce_free_prior() successfully removed all resources referencing
627 * \a ns and its refc == 0.
629 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
639 * Make sure that nobody can find this ns in its list.
641 ldlm_namespace_unregister(ns, ns->ns_client);
643 * Fini pool _before_ parent proc dir is removed. This is important as
644 * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
645 * it after @dir may cause oops.
647 ldlm_pool_fini(&ns->ns_pool);
651 struct proc_dir_entry *dir;
652 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
654 CERROR("dlm namespace %s has no procfs dir?\n",
657 lprocfs_remove(&dir);
662 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
663 OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
666 * Namespace \a ns should be not on list in this time, otherwise this
667 * will cause issues realted to using freed \a ns in pools thread.
669 LASSERT(cfs_list_empty(&ns->ns_list_chain));
676 /* Cleanup the resource, and free namespace.
679 * proc1: destroy import
680 * class_disconnect_export(grab cl_sem) ->
681 * -> ldlm_namespace_free ->
682 * -> lprocfs_remove(grab _lprocfs_lock).
683 * proc2: read proc info
684 * lprocfs_fops_read(grab _lprocfs_lock) ->
685 * -> osc_rd_active, etc(grab cl_sem).
687 * So that I have to split the ldlm_namespace_free into two parts - the first
688 * part ldlm_namespace_free_prior is used to cleanup the resource which is
689 * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
690 * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
693 void ldlm_namespace_free(struct ldlm_namespace *ns,
694 struct obd_import *imp,
697 ldlm_namespace_free_prior(ns, imp, force);
698 ldlm_namespace_free_post(ns);
702 void ldlm_namespace_get_locked(struct ldlm_namespace *ns)
707 void ldlm_namespace_get(struct ldlm_namespace *ns)
709 cfs_spin_lock(&ns->ns_hash_lock);
710 ldlm_namespace_get_locked(ns);
711 cfs_spin_unlock(&ns->ns_hash_lock);
714 void ldlm_namespace_put_locked(struct ldlm_namespace *ns, int wakeup)
716 LASSERT(ns->ns_refcount > 0);
718 if (ns->ns_refcount == 0 && wakeup)
719 cfs_waitq_signal(&ns->ns_waitq);
722 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
724 cfs_spin_lock(&ns->ns_hash_lock);
725 ldlm_namespace_put_locked(ns, wakeup);
726 cfs_spin_unlock(&ns->ns_hash_lock);
729 /* Register @ns in the list of namespaces */
730 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
732 cfs_mutex_down(ldlm_namespace_lock(client));
733 LASSERT(cfs_list_empty(&ns->ns_list_chain));
734 cfs_list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
735 cfs_atomic_inc(ldlm_namespace_nr(client));
736 cfs_mutex_up(ldlm_namespace_lock(client));
739 /* Unregister @ns from the list of namespaces */
740 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
742 cfs_mutex_down(ldlm_namespace_lock(client));
743 LASSERT(!cfs_list_empty(&ns->ns_list_chain));
745 * Some asserts and possibly other parts of code still using
746 * list_empty(&ns->ns_list_chain). This is why it is important
747 * to use list_del_init() here.
749 cfs_list_del_init(&ns->ns_list_chain);
750 cfs_atomic_dec(ldlm_namespace_nr(client));
751 cfs_mutex_up(ldlm_namespace_lock(client));
754 /* Should be called under ldlm_namespace_lock(client) taken */
755 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
757 LASSERT(!cfs_list_empty(&ns->ns_list_chain));
758 LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
759 cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
762 /* Should be called under ldlm_namespace_lock(client) taken */
763 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
765 LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
766 LASSERT(!cfs_list_empty(ldlm_namespace_list(client)));
767 return container_of(ldlm_namespace_list(client)->next,
768 struct ldlm_namespace, ns_list_chain);
770 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
771 const struct ldlm_res_id *name)
776 for (i = 0; i < RES_NAME_SIZE; i++)
777 hash += name->name[i];
779 hash += (__u32)((unsigned long)parent >> 4);
781 return (hash & RES_HASH_MASK);
784 static struct ldlm_resource *ldlm_resource_new(void)
786 struct ldlm_resource *res;
789 OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, CFS_ALLOC_IO);
793 memset(res, 0, sizeof(*res));
795 CFS_INIT_LIST_HEAD(&res->lr_children);
796 CFS_INIT_LIST_HEAD(&res->lr_childof);
797 CFS_INIT_LIST_HEAD(&res->lr_granted);
798 CFS_INIT_LIST_HEAD(&res->lr_converting);
799 CFS_INIT_LIST_HEAD(&res->lr_waiting);
801 /* initialize interval trees for each lock mode*/
802 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
803 res->lr_itree[idx].lit_size = 0;
804 res->lr_itree[idx].lit_mode = 1 << idx;
805 res->lr_itree[idx].lit_root = NULL;
808 cfs_atomic_set(&res->lr_refcount, 1);
809 cfs_spin_lock_init(&res->lr_lock);
810 lu_ref_init(&res->lr_reference);
812 /* one who creates the resource must unlock
813 * the semaphore after lvb initialization */
814 cfs_init_mutex_locked(&res->lr_lvb_sem);
819 /* must be called with hash lock held */
820 static struct ldlm_resource *
821 ldlm_resource_find(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
824 cfs_list_t *bucket, *tmp;
825 struct ldlm_resource *res;
827 LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
828 bucket = ns->ns_hash + hash;
830 cfs_list_for_each(tmp, bucket) {
831 res = cfs_list_entry(tmp, struct ldlm_resource, lr_hash);
832 if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0)
839 /* Args: locked namespace
840 * Returns: newly-allocated, referenced, unlocked resource */
841 static struct ldlm_resource *
842 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
843 const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type)
846 struct ldlm_resource *res, *old_res;
849 LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
852 res = ldlm_resource_new();
856 res->lr_name = *name;
857 res->lr_namespace = ns;
859 res->lr_most_restr = LCK_NL;
861 cfs_spin_lock(&ns->ns_hash_lock);
862 old_res = ldlm_resource_find(ns, name, hash);
864 /* someone won the race and added the resource before */
865 ldlm_resource_getref(old_res);
866 cfs_spin_unlock(&ns->ns_hash_lock);
867 /* clean lu_ref for failed resource */
868 lu_ref_fini(&res->lr_reference);
869 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
870 /* synchronize WRT resource creation */
871 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
872 cfs_down(&old_res->lr_lvb_sem);
873 cfs_up(&old_res->lr_lvb_sem);
878 /* we won! let's add the resource */
879 bucket = ns->ns_hash + hash;
880 cfs_list_add(&res->lr_hash, bucket);
882 ldlm_namespace_get_locked(ns);
884 if (parent == NULL) {
885 cfs_list_add(&res->lr_childof, &ns->ns_root_list);
887 res->lr_parent = parent;
888 cfs_list_add(&res->lr_childof, &parent->lr_children);
890 cfs_spin_unlock(&ns->ns_hash_lock);
892 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
895 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
896 rc = ns->ns_lvbo->lvbo_init(res);
898 CERROR("lvbo_init failed for resource "
899 LPU64": rc %d\n", name->name[0], rc);
900 /* we create resource with locked lr_lvb_sem */
901 cfs_up(&res->lr_lvb_sem);
907 /* Args: unlocked namespace
908 * Locks: takes and releases ns->ns_lock and res->lr_lock
909 * Returns: referenced, unlocked ldlm_resource or NULL */
910 struct ldlm_resource *
911 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
912 const struct ldlm_res_id *name, ldlm_type_t type, int create)
914 __u32 hash = ldlm_hash_fn(parent, name);
915 struct ldlm_resource *res = NULL;
919 LASSERT(ns->ns_hash != NULL);
920 LASSERT(name->name[0] != 0);
922 cfs_spin_lock(&ns->ns_hash_lock);
923 res = ldlm_resource_find(ns, name, hash);
925 ldlm_resource_getref(res);
926 cfs_spin_unlock(&ns->ns_hash_lock);
927 /* synchronize WRT resource creation */
928 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
929 cfs_down(&res->lr_lvb_sem);
930 cfs_up(&res->lr_lvb_sem);
934 cfs_spin_unlock(&ns->ns_hash_lock);
939 res = ldlm_resource_add(ns, parent, name, hash, type);
943 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
945 LASSERT(res != NULL);
946 LASSERT(res != LP_POISON);
947 cfs_atomic_inc(&res->lr_refcount);
948 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
949 cfs_atomic_read(&res->lr_refcount));
953 void __ldlm_resource_putref_final(struct ldlm_resource *res)
955 struct ldlm_namespace *ns = res->lr_namespace;
957 LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
959 if (!cfs_list_empty(&res->lr_granted)) {
960 ldlm_resource_dump(D_ERROR, res);
964 if (!cfs_list_empty(&res->lr_converting)) {
965 ldlm_resource_dump(D_ERROR, res);
969 if (!cfs_list_empty(&res->lr_waiting)) {
970 ldlm_resource_dump(D_ERROR, res);
974 if (!cfs_list_empty(&res->lr_children)) {
975 ldlm_resource_dump(D_ERROR, res);
979 /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few
980 * lines below when all children are freed. */
981 ldlm_namespace_put_locked(ns, 0);
982 cfs_list_del_init(&res->lr_hash);
983 cfs_list_del_init(&res->lr_childof);
984 lu_ref_fini(&res->lr_reference);
987 if (ns->ns_resources == 0)
988 cfs_waitq_signal(&ns->ns_waitq);
991 /* Returns 1 if the resource was freed, 0 if it remains. */
992 int ldlm_resource_putref(struct ldlm_resource *res)
994 struct ldlm_namespace *ns = res->lr_namespace;
998 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
999 cfs_atomic_read(&res->lr_refcount) - 1);
1000 LASSERTF(cfs_atomic_read(&res->lr_refcount) > 0, "%d",
1001 cfs_atomic_read(&res->lr_refcount));
1002 LASSERTF(cfs_atomic_read(&res->lr_refcount) < LI_POISON, "%d",
1003 cfs_atomic_read(&res->lr_refcount));
1005 if (cfs_atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
1006 __ldlm_resource_putref_final(res);
1007 cfs_spin_unlock(&ns->ns_hash_lock);
1008 if (res->lr_lvb_data)
1009 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
1010 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1017 /* Returns 1 if the resource was freed, 0 if it remains. */
1018 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1023 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
1024 cfs_atomic_read(&res->lr_refcount) - 1);
1025 LASSERT(cfs_atomic_read(&res->lr_refcount) > 0);
1026 LASSERT(cfs_atomic_read(&res->lr_refcount) < LI_POISON);
1028 LASSERT(cfs_atomic_read(&res->lr_refcount) >= 0);
1029 if (cfs_atomic_dec_and_test(&res->lr_refcount)) {
1030 __ldlm_resource_putref_final(res);
1031 if (res->lr_lvb_data)
1032 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
1033 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1040 void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head,
1041 struct ldlm_lock *lock)
1043 check_res_locked(res);
1045 ldlm_resource_dump(D_INFO, res);
1046 CDEBUG(D_OTHER, "About to add this lock:\n");
1047 ldlm_lock_dump(D_OTHER, lock, 0);
1049 if (lock->l_destroyed) {
1050 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1054 LASSERT(cfs_list_empty(&lock->l_res_link));
1056 cfs_list_add_tail(&lock->l_res_link, head);
1059 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1060 struct ldlm_lock *new)
1062 struct ldlm_resource *res = original->l_resource;
1064 check_res_locked(res);
1066 ldlm_resource_dump(D_INFO, res);
1067 CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1068 ldlm_lock_dump(D_OTHER, new, 0);
1070 if (new->l_destroyed) {
1071 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1075 LASSERT(cfs_list_empty(&new->l_res_link));
1077 cfs_list_add(&new->l_res_link, &original->l_res_link);
1081 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1083 int type = lock->l_resource->lr_type;
1085 check_res_locked(lock->l_resource);
1086 if (type == LDLM_IBITS || type == LDLM_PLAIN)
1087 ldlm_unlink_lock_skiplist(lock);
1088 else if (type == LDLM_EXTENT)
1089 ldlm_extent_unlink_lock(lock);
1090 cfs_list_del_init(&lock->l_res_link);
1093 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1095 desc->lr_type = res->lr_type;
1096 desc->lr_name = res->lr_name;
1099 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1103 if (!((libcfs_debug | D_ERROR) & level))
1106 cfs_mutex_down(ldlm_namespace_lock(client));
1108 cfs_list_for_each(tmp, ldlm_namespace_list(client)) {
1109 struct ldlm_namespace *ns;
1110 ns = cfs_list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1111 ldlm_namespace_dump(level, ns);
1114 cfs_mutex_up(ldlm_namespace_lock(client));
1117 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1121 if (!((libcfs_debug | D_ERROR) & level))
1124 CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1125 ns->ns_name, ns->ns_refcount,
1126 ns_is_client(ns) ? "client" : "server");
1128 if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1131 cfs_spin_lock(&ns->ns_hash_lock);
1132 tmp = ns->ns_root_list.next;
1133 while (tmp != &ns->ns_root_list) {
1134 struct ldlm_resource *res;
1135 res = cfs_list_entry(tmp, struct ldlm_resource, lr_childof);
1137 ldlm_resource_getref(res);
1138 cfs_spin_unlock(&ns->ns_hash_lock);
1139 LDLM_RESOURCE_ADDREF(res);
1142 ldlm_resource_dump(level, res);
1145 LDLM_RESOURCE_DELREF(res);
1146 cfs_spin_lock(&ns->ns_hash_lock);
1148 ldlm_resource_putref_locked(res);
1150 ns->ns_next_dump = cfs_time_shift(10);
1151 cfs_spin_unlock(&ns->ns_hash_lock);
1154 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1159 CLASSERT(RES_NAME_SIZE == 4);
1161 if (!((libcfs_debug | D_ERROR) & level))
1164 CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1165 ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1166 res->lr_name.name[2], res->lr_name.name[3],
1167 cfs_atomic_read(&res->lr_refcount));
1169 if (!cfs_list_empty(&res->lr_granted)) {
1171 CDEBUG(level, "Granted locks:\n");
1172 cfs_list_for_each(tmp, &res->lr_granted) {
1173 struct ldlm_lock *lock;
1174 lock = cfs_list_entry(tmp, struct ldlm_lock,
1176 ldlm_lock_dump(level, lock, ++pos);
1179 if (!cfs_list_empty(&res->lr_converting)) {
1181 CDEBUG(level, "Converting locks:\n");
1182 cfs_list_for_each(tmp, &res->lr_converting) {
1183 struct ldlm_lock *lock;
1184 lock = cfs_list_entry(tmp, struct ldlm_lock,
1186 ldlm_lock_dump(level, lock, ++pos);
1189 if (!cfs_list_empty(&res->lr_waiting)) {
1191 CDEBUG(level, "Waiting locks:\n");
1192 cfs_list_for_each(tmp, &res->lr_waiting) {
1193 struct ldlm_lock *lock;
1194 lock = cfs_list_entry(tmp, struct ldlm_lock,
1196 ldlm_lock_dump(level, lock, ++pos);