1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2002, 2003 Cluster File Systems, Inc.
5 * Author: Phil Schwan <phil@clusterfs.com>
6 * Author: Peter Braam <braam@clusterfs.com>
8 * This file is part of the Lustre file system, http://www.lustre.org
9 * Lustre is a trademark of Cluster File Systems, Inc.
11 * You may have signed or agreed to another license before downloading
12 * this software. If so, you are bound by the terms and conditions
13 * of that agreement, and the following does not apply to you. See the
14 * LICENSE file included with this distribution for more information.
16 * If you did not agree to a different license, then this copy of Lustre
17 * is open source software; you can redistribute it and/or modify it
18 * under the terms of version 2 of the GNU General Public License as
19 * published by the Free Software Foundation.
21 * In either case, Lustre is distributed in the hope that it will be
22 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
23 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
24 * license text for more details.
27 #define DEBUG_SUBSYSTEM S_LDLM
29 # include <lustre_dlm.h>
31 # include <liblustre.h>
34 #include <obd_class.h>
35 #include "ldlm_internal.h"
37 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
39 atomic_t ldlm_srv_namespace_nr = ATOMIC_INIT(0);
40 atomic_t ldlm_cli_namespace_nr = ATOMIC_INIT(0);
42 struct semaphore ldlm_srv_namespace_lock;
43 CFS_LIST_HEAD(ldlm_srv_namespace_list);
45 struct semaphore ldlm_cli_namespace_lock;
46 CFS_LIST_HEAD(ldlm_cli_namespace_list);
48 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
49 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
50 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
53 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
54 unsigned long count, void *data)
56 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
57 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
61 int ldlm_proc_setup(void)
64 struct lprocfs_vars list[] = {
65 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
68 LASSERT(ldlm_ns_proc_dir == NULL);
70 ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
73 if (IS_ERR(ldlm_type_proc_dir)) {
74 CERROR("LProcFS failed in ldlm-init\n");
75 rc = PTR_ERR(ldlm_type_proc_dir);
79 ldlm_ns_proc_dir = lprocfs_register("namespaces",
82 if (IS_ERR(ldlm_ns_proc_dir)) {
83 CERROR("LProcFS failed in ldlm-init\n");
84 rc = PTR_ERR(ldlm_ns_proc_dir);
88 ldlm_svc_proc_dir = lprocfs_register("services",
91 if (IS_ERR(ldlm_svc_proc_dir)) {
92 CERROR("LProcFS failed in ldlm-init\n");
93 rc = PTR_ERR(ldlm_svc_proc_dir);
97 rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
102 lprocfs_remove(&ldlm_ns_proc_dir);
104 lprocfs_remove(&ldlm_type_proc_dir);
106 ldlm_svc_proc_dir = NULL;
110 void ldlm_proc_cleanup(void)
112 if (ldlm_svc_proc_dir)
113 lprocfs_remove(&ldlm_svc_proc_dir);
115 if (ldlm_ns_proc_dir)
116 lprocfs_remove(&ldlm_ns_proc_dir);
118 if (ldlm_type_proc_dir)
119 lprocfs_remove(&ldlm_type_proc_dir);
122 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
123 int count, int *eof, void *data)
125 struct ldlm_namespace *ns = data;
126 __u32 *nr = &ns->ns_max_unused;
128 if (ns_connect_lru_resize(ns))
129 nr = &ns->ns_nr_unused;
130 return lprocfs_rd_uint(page, start, off, count, eof, nr);
133 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
134 unsigned long count, void *data)
136 struct ldlm_namespace *ns = data;
137 char dummy[MAX_STRING_SIZE + 1], *end;
141 dummy[MAX_STRING_SIZE] = '\0';
142 if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
145 if (count == 6 && memcmp(dummy, "clear", 5) == 0) {
147 "dropping all unused locks from namespace %s\n",
149 if (ns_connect_lru_resize(ns)) {
150 int canceled, unused = ns->ns_nr_unused;
152 /* Try to cancel all @ns_nr_unused locks. */
153 canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
155 if (canceled < unused) {
156 CERROR("not all requested locks are canceled, "
157 "requested: %d, canceled: %d\n", unused,
162 tmp = ns->ns_max_unused;
163 ns->ns_max_unused = 0;
164 ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
165 ns->ns_max_unused = tmp;
170 tmp = simple_strtoul(dummy, &end, 0);
172 CERROR("invalid value written\n");
175 lru_resize = (tmp == 0);
177 if (ns_connect_lru_resize(ns)) {
179 ns->ns_max_unused = (unsigned int)tmp;
181 if (tmp > ns->ns_nr_unused)
182 tmp = ns->ns_nr_unused;
183 tmp = ns->ns_nr_unused - tmp;
185 CDEBUG(D_DLMTRACE, "changing namespace %s unused locks from %u to %u\n",
186 ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
187 ldlm_cancel_lru(ns, (unsigned int)tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
190 CDEBUG(D_DLMTRACE, "disable lru_resize for namespace %s\n",
192 ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
195 CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
196 ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
197 ns->ns_max_unused = (unsigned int)tmp;
198 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
200 /* Make sure that originally lru resize was supported before
201 * turning it on here. */
203 (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
204 CDEBUG(D_DLMTRACE, "enable lru_resize for namespace %s\n",
206 ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
213 void ldlm_proc_namespace(struct ldlm_namespace *ns)
215 struct lprocfs_vars lock_vars[2];
216 char lock_name[MAX_STRING_SIZE + 1];
219 LASSERT(ns->ns_name != NULL);
221 lock_name[MAX_STRING_SIZE] = '\0';
223 memset(lock_vars, 0, sizeof(lock_vars));
224 lock_vars[0].name = lock_name;
226 snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
227 lock_vars[0].data = &ns->ns_refcount;
228 lock_vars[0].read_fptr = lprocfs_rd_atomic;
229 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
231 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
232 lock_vars[0].data = &ns->ns_locks;
233 lock_vars[0].read_fptr = lprocfs_rd_atomic;
234 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
236 if (ns_is_client(ns)) {
237 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
239 lock_vars[0].data = &ns->ns_nr_unused;
240 lock_vars[0].read_fptr = lprocfs_rd_uint;
241 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
243 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
245 lock_vars[0].data = ns;
246 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
247 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
248 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
250 snprintf(lock_name, MAX_STRING_SIZE, "%s/shrink_thumb",
252 lock_vars[0].data = ns;
253 lock_vars[0].read_fptr = lprocfs_rd_uint;
254 lock_vars[0].write_fptr = lprocfs_wr_uint;
255 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
257 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
259 lock_vars[0].data = &ns->ns_max_age;
260 lock_vars[0].read_fptr = lprocfs_rd_uint;
261 lock_vars[0].write_fptr = lprocfs_wr_uint;
262 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
264 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
266 lock_vars[0].data = &ns->ns_ctime_age_limit;
267 lock_vars[0].read_fptr = lprocfs_rd_uint;
268 lock_vars[0].write_fptr = lprocfs_wr_uint;
269 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
271 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
273 lock_vars[0].data = &ns->ns_max_nolock_size;
274 lock_vars[0].read_fptr = lprocfs_rd_uint;
275 lock_vars[0].write_fptr = lprocfs_wr_uint;
276 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
278 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
280 lock_vars[0].data = &ns->ns_contention_time;
281 lock_vars[0].read_fptr = lprocfs_rd_uint;
282 lock_vars[0].write_fptr = lprocfs_wr_uint;
283 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
285 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
287 lock_vars[0].data = &ns->ns_contended_locks;
288 lock_vars[0].read_fptr = lprocfs_rd_uint;
289 lock_vars[0].write_fptr = lprocfs_wr_uint;
290 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
293 #undef MAX_STRING_SIZE
295 #define ldlm_proc_namespace(ns) do {} while (0)
298 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
299 ldlm_side_t client, ldlm_appetite_t apt)
301 struct ldlm_namespace *ns = NULL;
302 struct list_head *bucket;
303 int rc, idx, namelen;
308 CERROR("ldlm_get_ref failed: %d\n", rc);
316 OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
320 ns->ns_shrink_thumb = LDLM_LOCK_SHRINK_THUMB;
321 ns->ns_appetite = apt;
323 LASSERT(obd != NULL);
326 namelen = strlen(name);
327 OBD_ALLOC(ns->ns_name, namelen + 1);
329 GOTO(out_hash, NULL);
331 strcpy(ns->ns_name, name);
333 CFS_INIT_LIST_HEAD(&ns->ns_root_list);
334 CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
336 ns->ns_client = client;
337 spin_lock_init(&ns->ns_hash_lock);
338 atomic_set(&ns->ns_locks, 0);
339 ns->ns_resources = 0;
340 cfs_waitq_init(&ns->ns_waitq);
341 ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
342 ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
343 ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
345 for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
347 CFS_INIT_LIST_HEAD(bucket);
349 CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
350 ns->ns_nr_unused = 0;
351 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
352 ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
353 ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
354 spin_lock_init(&ns->ns_unused_lock);
355 ns->ns_orig_connect_flags = 0;
356 ns->ns_connect_flags = 0;
357 ldlm_proc_namespace(ns);
359 idx = atomic_read(ldlm_namespace_nr(client));
360 rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
362 CERROR("Can't initialize lock pool, rc %d\n", rc);
366 ldlm_namespace_register(ns, client);
369 ldlm_namespace_cleanup(ns, 0);
370 OBD_FREE(ns->ns_name, namelen + 1);
372 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
380 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
382 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
383 * This is currently only used for recovery, and we make certain assumptions
384 * as a result--notably, that we shouldn't cancel locks with refs. -phil
386 * Called with the ns_lock held. */
387 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
390 struct list_head *tmp;
391 int rc = 0, client = ns_is_client(res->lr_namespace);
392 int local_only = (flags & LDLM_FL_LOCAL_ONLY);
397 struct ldlm_lock *lock = NULL;
399 /* first, we look for non-cleaned-yet lock
400 * all cleaned locks are marked by CLEANED flag */
402 list_for_each(tmp, q) {
403 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
404 if (lock->l_flags & LDLM_FL_CLEANED) {
409 lock->l_flags |= LDLM_FL_CLEANED;
418 /* Set CBPENDING so nothing in the cancellation path
419 * can match this lock */
420 lock->l_flags |= LDLM_FL_CBPENDING;
421 lock->l_flags |= LDLM_FL_FAILED;
422 lock->l_flags |= flags;
424 /* ... without sending a CANCEL message for local_only. */
426 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
428 if (local_only && (lock->l_readers || lock->l_writers)) {
429 /* This is a little bit gross, but much better than the
430 * alternative: pretend that we got a blocking AST from
431 * the server, so that when the lock is decref'd, it
432 * will go away ... */
434 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
435 if (lock->l_completion_ast)
436 lock->l_completion_ast(lock, 0, NULL);
442 struct lustre_handle lockh;
445 ldlm_lock2handle(lock, &lockh);
446 rc = ldlm_cli_cancel(&lockh);
448 CERROR("ldlm_cli_cancel: %d\n", rc);
450 ldlm_resource_unlink_lock(lock);
452 LDLM_DEBUG(lock, "Freeing a lock still held by a "
454 ldlm_lock_destroy(lock);
462 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
464 struct list_head *tmp;
468 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
472 for (i = 0; i < RES_HASH_SIZE; i++) {
473 spin_lock(&ns->ns_hash_lock);
474 tmp = ns->ns_hash[i].next;
475 while (tmp != &(ns->ns_hash[i])) {
476 struct ldlm_resource *res;
477 res = list_entry(tmp, struct ldlm_resource, lr_hash);
478 ldlm_resource_getref(res);
479 spin_unlock(&ns->ns_hash_lock);
481 cleanup_resource(res, &res->lr_granted, flags);
482 cleanup_resource(res, &res->lr_converting, flags);
483 cleanup_resource(res, &res->lr_waiting, flags);
485 spin_lock(&ns->ns_hash_lock);
488 /* XXX: former stuff caused issues in case of race
489 * between ldlm_namespace_cleanup() and lockd() when
490 * client gets blocking ast when lock gets distracted by
491 * server. This is 1_4 branch solution, let's see how
493 if (!ldlm_resource_putref_locked(res))
495 "Namespace %s resource refcount nonzero "
496 "(%d) after lock cleanup; forcing cleanup.\n",
497 ns->ns_name, atomic_read(&res->lr_refcount));
499 spin_unlock(&ns->ns_hash_lock);
505 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
509 /* At shutdown time, don't call the cancellation callback */
510 ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
512 if (ns->ns_refcount > 0) {
513 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
516 "dlm namespace %s free waiting on refcount %d\n",
517 ns->ns_name, ns->ns_refcount);
520 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
522 rc = l_wait_event(ns->ns_waitq,
523 ns->ns_refcount == 0, &lwi);
525 /* Forced cleanups should be able to reclaim all references,
526 * so it's safe to wait forever... we can't leak locks... */
527 if (force && rc == -ETIMEDOUT) {
528 LCONSOLE_ERROR("Forced cleanup waiting for %s "
529 "namespace with %d resources in use, "
530 "(rc=%d)\n", ns->ns_name,
531 ns->ns_refcount, rc);
532 GOTO(force_wait, rc);
535 if (ns->ns_refcount) {
536 LCONSOLE_ERROR("Cleanup waiting for %s namespace "
537 "with %d resources in use, (rc=%d)\n",
539 ns->ns_refcount, rc);
540 RETURN(ELDLM_NAMESPACE_EXISTS);
543 "dlm namespace %s free done waiting\n", ns->ns_name);
550 * Performs various cleanups for passed \a ns to make it drop refc and be ready
551 * for freeing. Waits for refc == 0.
553 * The following is done:
554 * (0) Unregister \a ns from its list to make inaccessible for potential users
555 * like pools thread and others;
556 * (1) Clear all locks in \a ns.
558 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
559 struct obd_import *imp,
570 * Make sure that nobody can find this ns in its list.
572 ldlm_namespace_unregister(ns, ns->ns_client);
575 * Can fail with -EINTR when force == 0 in which case try harder.
577 rc = __ldlm_namespace_free(ns, force);
578 if (rc != ELDLM_OK) {
580 ptlrpc_disconnect_import(imp, 0);
581 ptlrpc_invalidate_import(imp);
585 * With all requests dropped and the import inactive
586 * we are gaurenteed all reference will be dropped.
588 rc = __ldlm_namespace_free(ns, 1);
595 * Performs freeing memory structures related to \a ns. This is only done when
596 * ldlm_namespce_free_prior() successfully removed all resources referencing
597 * \a ns and its refc == 0.
599 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
608 * Fini pool _before_ parent proc dir is removed. This is important as
609 * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
610 * it after @dir may cause oops.
612 ldlm_pool_fini(&ns->ns_pool);
616 struct proc_dir_entry *dir;
617 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
619 CERROR("dlm namespace %s has no procfs dir?\n",
622 lprocfs_remove(&dir);
627 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
628 OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
631 * Namespace \a ns should be not on list in this time, otherwise this
632 * will cause issues realted to using freed \a ns in pools thread.
634 LASSERT(list_empty(&ns->ns_list_chain));
641 /* Cleanup the resource, and free namespace.
644 * proc1: destroy import
645 * class_disconnect_export(grab cl_sem) ->
646 * -> ldlm_namespace_free ->
647 * -> lprocfs_remove(grab _lprocfs_lock).
648 * proc2: read proc info
649 * lprocfs_fops_read(grab _lprocfs_lock) ->
650 * -> osc_rd_active, etc(grab cl_sem).
652 * So that I have to split the ldlm_namespace_free into two parts - the first
653 * part ldlm_namespace_free_prior is used to cleanup the resource which is
654 * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
655 * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
658 void ldlm_namespace_free(struct ldlm_namespace *ns,
659 struct obd_import *imp,
662 ldlm_namespace_free_prior(ns, imp, force);
663 ldlm_namespace_free_post(ns);
667 void ldlm_namespace_get_locked(struct ldlm_namespace *ns)
669 LASSERT(ns->ns_refcount >= 0);
673 void ldlm_namespace_get(struct ldlm_namespace *ns)
675 spin_lock(&ns->ns_hash_lock);
676 ldlm_namespace_get_locked(ns);
677 spin_unlock(&ns->ns_hash_lock);
680 void ldlm_namespace_put_locked(struct ldlm_namespace *ns, int wakeup)
682 LASSERT(ns->ns_refcount > 0);
684 if (ns->ns_refcount == 0 && wakeup)
685 wake_up(&ns->ns_waitq);
688 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
690 spin_lock(&ns->ns_hash_lock);
691 ldlm_namespace_put_locked(ns, wakeup);
692 spin_unlock(&ns->ns_hash_lock);
695 /* Register @ns in the list of namespaces */
696 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
698 mutex_down(ldlm_namespace_lock(client));
699 LASSERT(list_empty(&ns->ns_list_chain));
700 list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
701 atomic_inc(ldlm_namespace_nr(client));
702 mutex_up(ldlm_namespace_lock(client));
705 /* Unregister @ns from the list of namespaces */
706 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
708 mutex_down(ldlm_namespace_lock(client));
709 LASSERT(!list_empty(&ns->ns_list_chain));
711 * Some asserts and possibly other parts of code still using
712 * list_empty(&ns->ns_list_chain). This is why it is important
713 * to use list_del_init() here.
715 list_del_init(&ns->ns_list_chain);
716 atomic_dec(ldlm_namespace_nr(client));
717 mutex_up(ldlm_namespace_lock(client));
720 /* Should be called under ldlm_namespace_lock(client) taken */
721 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
723 LASSERT(!list_empty(&ns->ns_list_chain));
724 LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
725 list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
728 /* Should be called under ldlm_namespace_lock(client) taken */
729 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
731 LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
732 LASSERT(!list_empty(ldlm_namespace_list(client)));
733 return container_of(ldlm_namespace_list(client)->next,
734 struct ldlm_namespace, ns_list_chain);
736 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
737 const struct ldlm_res_id *name)
742 for (i = 0; i < RES_NAME_SIZE; i++)
743 hash += name->name[i];
745 hash += (__u32)((unsigned long)parent >> 4);
747 return (hash & RES_HASH_MASK);
750 static struct ldlm_resource *ldlm_resource_new(void)
752 struct ldlm_resource *res;
755 OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
759 memset(res, 0, sizeof(*res));
761 CFS_INIT_LIST_HEAD(&res->lr_children);
762 CFS_INIT_LIST_HEAD(&res->lr_childof);
763 CFS_INIT_LIST_HEAD(&res->lr_granted);
764 CFS_INIT_LIST_HEAD(&res->lr_converting);
765 CFS_INIT_LIST_HEAD(&res->lr_waiting);
767 /* initialize interval trees for each lock mode*/
768 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
769 res->lr_itree[idx].lit_size = 0;
770 res->lr_itree[idx].lit_mode = 1 << idx;
771 res->lr_itree[idx].lit_root = NULL;
774 atomic_set(&res->lr_refcount, 1);
775 spin_lock_init(&res->lr_lock);
777 /* one who creates the resource must unlock
778 * the semaphore after lvb initialization */
779 init_MUTEX_LOCKED(&res->lr_lvb_sem);
784 /* must be called with hash lock held */
785 static struct ldlm_resource *
786 ldlm_resource_find(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
789 struct list_head *bucket, *tmp;
790 struct ldlm_resource *res;
792 LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
793 bucket = ns->ns_hash + hash;
795 list_for_each(tmp, bucket) {
796 res = list_entry(tmp, struct ldlm_resource, lr_hash);
797 if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0)
804 /* Args: locked namespace
805 * Returns: newly-allocated, referenced, unlocked resource */
806 static struct ldlm_resource *
807 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
808 const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type)
810 struct list_head *bucket;
811 struct ldlm_resource *res, *old_res;
814 LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
817 res = ldlm_resource_new();
821 res->lr_name = *name;
822 res->lr_namespace = ns;
824 res->lr_most_restr = LCK_NL;
826 spin_lock(&ns->ns_hash_lock);
827 old_res = ldlm_resource_find(ns, name, hash);
829 /* someone won the race and added the resource before */
830 ldlm_resource_getref(old_res);
831 spin_unlock(&ns->ns_hash_lock);
832 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
833 /* synchronize WRT resource creation */
834 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
835 down(&old_res->lr_lvb_sem);
836 up(&old_res->lr_lvb_sem);
841 /* we won! let's add the resource */
842 bucket = ns->ns_hash + hash;
843 list_add(&res->lr_hash, bucket);
845 ldlm_namespace_get_locked(ns);
847 if (parent == NULL) {
848 list_add(&res->lr_childof, &ns->ns_root_list);
850 res->lr_parent = parent;
851 list_add(&res->lr_childof, &parent->lr_children);
853 spin_unlock(&ns->ns_hash_lock);
855 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
858 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
859 rc = ns->ns_lvbo->lvbo_init(res);
861 CERROR("lvbo_init failed for resource "
862 LPU64": rc %d\n", name->name[0], rc);
863 /* we create resource with locked lr_lvb_sem */
864 up(&res->lr_lvb_sem);
870 /* Args: unlocked namespace
871 * Locks: takes and releases ns->ns_lock and res->lr_lock
872 * Returns: referenced, unlocked ldlm_resource or NULL */
873 struct ldlm_resource *
874 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
875 const struct ldlm_res_id *name, ldlm_type_t type, int create)
877 __u32 hash = ldlm_hash_fn(parent, name);
878 struct ldlm_resource *res = NULL;
882 LASSERT(ns->ns_hash != NULL);
883 LASSERT(name->name[0] != 0);
885 spin_lock(&ns->ns_hash_lock);
886 res = ldlm_resource_find(ns, name, hash);
888 ldlm_resource_getref(res);
889 spin_unlock(&ns->ns_hash_lock);
890 /* synchronize WRT resource creation */
891 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
892 down(&res->lr_lvb_sem);
893 up(&res->lr_lvb_sem);
897 spin_unlock(&ns->ns_hash_lock);
902 res = ldlm_resource_add(ns, parent, name, hash, type);
906 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
908 LASSERT(res != NULL);
909 LASSERT(res != LP_POISON);
910 atomic_inc(&res->lr_refcount);
911 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
912 atomic_read(&res->lr_refcount));
916 void __ldlm_resource_putref_final(struct ldlm_resource *res)
918 struct ldlm_namespace *ns = res->lr_namespace;
920 LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
922 if (!list_empty(&res->lr_granted)) {
923 ldlm_resource_dump(D_ERROR, res);
927 if (!list_empty(&res->lr_converting)) {
928 ldlm_resource_dump(D_ERROR, res);
932 if (!list_empty(&res->lr_waiting)) {
933 ldlm_resource_dump(D_ERROR, res);
937 if (!list_empty(&res->lr_children)) {
938 ldlm_resource_dump(D_ERROR, res);
942 /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few
943 * lines below when all children are freed. */
944 ldlm_namespace_put_locked(ns, 0);
945 list_del_init(&res->lr_hash);
946 list_del_init(&res->lr_childof);
949 if (ns->ns_resources == 0)
950 wake_up(&ns->ns_waitq);
953 /* Returns 1 if the resource was freed, 0 if it remains. */
954 int ldlm_resource_putref(struct ldlm_resource *res)
956 struct ldlm_namespace *ns = res->lr_namespace;
960 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
961 atomic_read(&res->lr_refcount) - 1);
962 LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
963 atomic_read(&res->lr_refcount));
964 LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
965 atomic_read(&res->lr_refcount));
967 if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
968 __ldlm_resource_putref_final(res);
969 spin_unlock(&ns->ns_hash_lock);
970 if (res->lr_lvb_data)
971 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
972 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
979 /* Returns 1 if the resource was freed, 0 if it remains. */
980 int ldlm_resource_putref_locked(struct ldlm_resource *res)
985 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
986 atomic_read(&res->lr_refcount) - 1);
987 LASSERT(atomic_read(&res->lr_refcount) > 0);
988 LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
990 LASSERT(atomic_read(&res->lr_refcount) >= 0);
991 if (atomic_dec_and_test(&res->lr_refcount)) {
992 __ldlm_resource_putref_final(res);
993 if (res->lr_lvb_data)
994 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
995 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1002 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1003 struct ldlm_lock *lock)
1005 check_res_locked(res);
1007 ldlm_resource_dump(D_OTHER, res);
1008 CDEBUG(D_OTHER, "About to add this lock:\n");
1009 ldlm_lock_dump(D_OTHER, lock, 0);
1011 if (lock->l_destroyed) {
1012 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1016 LASSERT(list_empty(&lock->l_res_link));
1018 list_add_tail(&lock->l_res_link, head);
1021 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1022 struct ldlm_lock *new)
1024 struct ldlm_resource *res = original->l_resource;
1026 check_res_locked(res);
1028 ldlm_resource_dump(D_OTHER, res);
1029 CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1030 ldlm_lock_dump(D_OTHER, new, 0);
1032 if (new->l_destroyed) {
1033 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1037 LASSERT(list_empty(&new->l_res_link));
1039 list_add(&new->l_res_link, &original->l_res_link);
1043 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1045 int type = lock->l_resource->lr_type;
1047 check_res_locked(lock->l_resource);
1048 if (type == LDLM_IBITS || type == LDLM_PLAIN)
1049 ldlm_unlink_lock_skiplist(lock);
1050 else if (type == LDLM_EXTENT)
1051 ldlm_extent_unlink_lock(lock);
1052 list_del_init(&lock->l_res_link);
1055 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1057 desc->lr_type = res->lr_type;
1058 desc->lr_name = res->lr_name;
1061 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1063 struct list_head *tmp;
1065 if (!((libcfs_debug | D_ERROR) & level))
1068 mutex_down(ldlm_namespace_lock(client));
1070 list_for_each(tmp, ldlm_namespace_list(client)) {
1071 struct ldlm_namespace *ns;
1072 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1073 ldlm_namespace_dump(level, ns);
1076 mutex_up(ldlm_namespace_lock(client));
1079 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1081 struct list_head *tmp;
1083 if (!((libcfs_debug | D_ERROR) & level))
1086 CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1087 ns->ns_name, ns->ns_refcount,
1088 ns_is_client(ns) ? "client" : "server");
1090 if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1093 spin_lock(&ns->ns_hash_lock);
1094 tmp = ns->ns_root_list.next;
1095 while (tmp != &ns->ns_root_list) {
1096 struct ldlm_resource *res;
1097 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1099 ldlm_resource_getref(res);
1100 spin_unlock(&ns->ns_hash_lock);
1103 ldlm_resource_dump(level, res);
1106 spin_lock(&ns->ns_hash_lock);
1108 ldlm_resource_putref_locked(res);
1110 ns->ns_next_dump = cfs_time_shift(10);
1111 spin_unlock(&ns->ns_hash_lock);
1114 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1116 struct list_head *tmp;
1119 CLASSERT(RES_NAME_SIZE == 4);
1121 if (!((libcfs_debug | D_ERROR) & level))
1124 CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1125 ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1126 res->lr_name.name[2], res->lr_name.name[3],
1127 atomic_read(&res->lr_refcount));
1129 if (!list_empty(&res->lr_granted)) {
1131 CDEBUG(level, "Granted locks:\n");
1132 list_for_each(tmp, &res->lr_granted) {
1133 struct ldlm_lock *lock;
1134 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1135 ldlm_lock_dump(level, lock, ++pos);
1138 if (!list_empty(&res->lr_converting)) {
1140 CDEBUG(level, "Converting locks:\n");
1141 list_for_each(tmp, &res->lr_converting) {
1142 struct ldlm_lock *lock;
1143 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1144 ldlm_lock_dump(level, lock, ++pos);
1147 if (!list_empty(&res->lr_waiting)) {
1149 CDEBUG(level, "Waiting locks:\n");
1150 list_for_each(tmp, &res->lr_waiting) {
1151 struct ldlm_lock *lock;
1152 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1153 ldlm_lock_dump(level, lock, ++pos);