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(char *name, ldlm_side_t client,
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;
322 namelen = strlen(name);
323 OBD_ALLOC(ns->ns_name, namelen + 1);
325 GOTO(out_hash, NULL);
327 strcpy(ns->ns_name, name);
329 CFS_INIT_LIST_HEAD(&ns->ns_root_list);
331 ns->ns_client = client;
332 spin_lock_init(&ns->ns_hash_lock);
333 atomic_set(&ns->ns_locks, 0);
334 ns->ns_resources = 0;
335 cfs_waitq_init(&ns->ns_waitq);
336 ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
337 ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
338 ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
340 for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
342 CFS_INIT_LIST_HEAD(bucket);
344 CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
345 ns->ns_nr_unused = 0;
346 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
347 ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
348 ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
349 spin_lock_init(&ns->ns_unused_lock);
350 ns->ns_orig_connect_flags = 0;
351 ns->ns_connect_flags = 0;
352 ldlm_proc_namespace(ns);
354 idx = atomic_read(ldlm_namespace_nr(client));
355 rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
357 CERROR("Can't initialize lock pool, rc %d\n", rc);
361 mutex_down(ldlm_namespace_lock(client));
362 list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
363 atomic_inc(ldlm_namespace_nr(client));
364 mutex_up(ldlm_namespace_lock(client));
368 ldlm_namespace_cleanup(ns, 0);
369 OBD_FREE(ns->ns_name, namelen + 1);
371 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
379 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
381 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
382 * This is currently only used for recovery, and we make certain assumptions
383 * as a result--notably, that we shouldn't cancel locks with refs. -phil
385 * Called with the ns_lock held. */
386 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
389 struct list_head *tmp;
390 int rc = 0, client = ns_is_client(res->lr_namespace);
391 int local_only = (flags & LDLM_FL_LOCAL_ONLY);
396 struct ldlm_lock *lock = NULL;
398 /* first, we look for non-cleaned-yet lock
399 * all cleaned locks are marked by CLEANED flag */
401 list_for_each(tmp, q) {
402 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
403 if (lock->l_flags & LDLM_FL_CLEANED) {
408 lock->l_flags |= LDLM_FL_CLEANED;
417 /* Set CBPENDING so nothing in the cancellation path
418 * can match this lock */
419 lock->l_flags |= LDLM_FL_CBPENDING;
420 lock->l_flags |= LDLM_FL_FAILED;
421 lock->l_flags |= flags;
423 /* ... without sending a CANCEL message for local_only. */
425 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
427 if (local_only && (lock->l_readers || lock->l_writers)) {
428 /* This is a little bit gross, but much better than the
429 * alternative: pretend that we got a blocking AST from
430 * the server, so that when the lock is decref'd, it
431 * will go away ... */
433 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
434 if (lock->l_completion_ast)
435 lock->l_completion_ast(lock, 0, NULL);
441 struct lustre_handle lockh;
444 ldlm_lock2handle(lock, &lockh);
445 rc = ldlm_cli_cancel(&lockh);
447 CERROR("ldlm_cli_cancel: %d\n", rc);
449 ldlm_resource_unlink_lock(lock);
451 LDLM_DEBUG(lock, "Freeing a lock still held by a "
453 ldlm_lock_destroy(lock);
461 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
463 struct list_head *tmp;
467 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
471 for (i = 0; i < RES_HASH_SIZE; i++) {
472 spin_lock(&ns->ns_hash_lock);
473 tmp = ns->ns_hash[i].next;
474 while (tmp != &(ns->ns_hash[i])) {
475 struct ldlm_resource *res;
476 res = list_entry(tmp, struct ldlm_resource, lr_hash);
477 ldlm_resource_getref(res);
478 spin_unlock(&ns->ns_hash_lock);
480 cleanup_resource(res, &res->lr_granted, flags);
481 cleanup_resource(res, &res->lr_converting, flags);
482 cleanup_resource(res, &res->lr_waiting, flags);
484 spin_lock(&ns->ns_hash_lock);
487 /* XXX: former stuff caused issues in case of race
488 * between ldlm_namespace_cleanup() and lockd() when
489 * client gets blocking ast when lock gets distracted by
490 * server. This is 1_4 branch solution, let's see how
492 if (!ldlm_resource_putref_locked(res))
494 "Namespace %s resource refcount nonzero "
495 "(%d) after lock cleanup; forcing cleanup.\n",
496 ns->ns_name, atomic_read(&res->lr_refcount));
498 spin_unlock(&ns->ns_hash_lock);
504 /* Cleanup, but also free, the namespace */
505 int ldlm_namespace_free_prior(struct ldlm_namespace *ns)
511 mutex_down(ldlm_namespace_lock(ns->ns_client));
513 * Some asserts and possibly other parts of code still using
514 * list_empty(&ns->ns_list_chain). This is why it is important
515 * to use list_del_init() here.
517 list_del_init(&ns->ns_list_chain);
518 atomic_dec(ldlm_namespace_nr(ns->ns_client));
519 ldlm_pool_fini(&ns->ns_pool);
520 mutex_up(ldlm_namespace_lock(ns->ns_client));
522 /* At shutdown time, don't call the cancellation callback */
523 ldlm_namespace_cleanup(ns, 0);
525 if (ns->ns_refcount > 0) {
526 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
529 "dlm namespace %s free waiting on refcount %d\n",
530 ns->ns_name, ns->ns_refcount);
531 rc = l_wait_event(ns->ns_waitq,
532 ns->ns_refcount == 0, &lwi);
534 LCONSOLE_ERROR_MSG(0x139, "Lock manager: wait for %s "
535 "namespace cleanup aborted with %d "
536 "resources in use. (%d)\nI'm going "
537 "to try to clean up anyway, but I "
538 "might need a reboot of this node.\n",
539 ns->ns_name, (int) ns->ns_refcount,
542 "dlm namespace %s free done waiting\n", ns->ns_name);
548 int ldlm_namespace_free_post(struct ldlm_namespace *ns, int force)
556 struct proc_dir_entry *dir;
557 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
559 CERROR("dlm namespace %s has no procfs dir?\n",
562 lprocfs_remove(&dir);
567 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
568 OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
570 * @ns should be not on list in this time, otherwise this will cause
571 * issues realted to using freed @ns in pools thread.
573 LASSERT(list_empty(&ns->ns_list_chain));
580 /* Cleanup the resource, and free namespace.
583 * proc1: destroy import
584 * class_disconnect_export(grab cl_sem) ->
585 * -> ldlm_namespace_free ->
586 * -> lprocfs_remove(grab _lprocfs_lock).
587 * proc2: read proc info
588 * lprocfs_fops_read(grab _lprocfs_lock) ->
589 * -> osc_rd_active, etc(grab cl_sem).
591 * So that I have to split the ldlm_namespace_free into two parts - the first
592 * part ldlm_namespace_free_prior is used to cleanup the resource which is
593 * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
594 * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
597 int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
599 ldlm_namespace_free_prior(ns);
600 ldlm_namespace_free_post(ns, force);
605 void ldlm_namespace_get_nolock(struct ldlm_namespace *ns)
607 LASSERT(ns->ns_refcount >= 0);
611 void ldlm_namespace_get(struct ldlm_namespace *ns)
613 spin_lock(&ns->ns_hash_lock);
614 ldlm_namespace_get_nolock(ns);
615 spin_unlock(&ns->ns_hash_lock);
618 void ldlm_namespace_put_nolock(struct ldlm_namespace *ns, int wakeup)
620 LASSERT(ns->ns_refcount > 0);
622 if (ns->ns_refcount == 0 && wakeup)
623 wake_up(&ns->ns_waitq);
626 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
628 spin_lock(&ns->ns_hash_lock);
629 ldlm_namespace_put_nolock(ns, wakeup);
630 spin_unlock(&ns->ns_hash_lock);
633 /* Should be called under ldlm_namespace_lock(client) taken */
634 void ldlm_namespace_move(struct ldlm_namespace *ns, ldlm_side_t client)
636 LASSERT(!list_empty(&ns->ns_list_chain));
637 LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
638 list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
641 /* Should be called under ldlm_namespace_lock(client) taken */
642 struct ldlm_namespace *ldlm_namespace_first(ldlm_side_t client)
644 LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
645 LASSERT(!list_empty(ldlm_namespace_list(client)));
646 return container_of(ldlm_namespace_list(client)->next,
647 struct ldlm_namespace, ns_list_chain);
649 static __u32 ldlm_hash_fn(struct ldlm_resource *parent,
650 const struct ldlm_res_id *name)
655 for (i = 0; i < RES_NAME_SIZE; i++)
656 hash += name->name[i];
658 hash += (__u32)((unsigned long)parent >> 4);
660 return (hash & RES_HASH_MASK);
663 static struct ldlm_resource *ldlm_resource_new(void)
665 struct ldlm_resource *res;
668 OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
672 memset(res, 0, sizeof(*res));
674 CFS_INIT_LIST_HEAD(&res->lr_children);
675 CFS_INIT_LIST_HEAD(&res->lr_childof);
676 CFS_INIT_LIST_HEAD(&res->lr_granted);
677 CFS_INIT_LIST_HEAD(&res->lr_converting);
678 CFS_INIT_LIST_HEAD(&res->lr_waiting);
680 /* initialize interval trees for each lock mode*/
681 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
682 res->lr_itree[idx].lit_size = 0;
683 res->lr_itree[idx].lit_mode = 1 << idx;
684 res->lr_itree[idx].lit_root = NULL;
687 atomic_set(&res->lr_refcount, 1);
688 spin_lock_init(&res->lr_lock);
690 /* one who creates the resource must unlock
691 * the semaphore after lvb initialization */
692 init_MUTEX_LOCKED(&res->lr_lvb_sem);
697 /* must be called with hash lock held */
698 static struct ldlm_resource *
699 ldlm_resource_find(struct ldlm_namespace *ns, const struct ldlm_res_id *name,
702 struct list_head *bucket, *tmp;
703 struct ldlm_resource *res;
705 LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
706 bucket = ns->ns_hash + hash;
708 list_for_each(tmp, bucket) {
709 res = list_entry(tmp, struct ldlm_resource, lr_hash);
710 if (memcmp(&res->lr_name, name, sizeof(res->lr_name)) == 0)
717 /* Args: locked namespace
718 * Returns: newly-allocated, referenced, unlocked resource */
719 static struct ldlm_resource *
720 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
721 const struct ldlm_res_id *name, __u32 hash, ldlm_type_t type)
723 struct list_head *bucket;
724 struct ldlm_resource *res, *old_res;
727 LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
730 res = ldlm_resource_new();
734 res->lr_name = *name;
735 res->lr_namespace = ns;
737 res->lr_most_restr = LCK_NL;
739 spin_lock(&ns->ns_hash_lock);
740 old_res = ldlm_resource_find(ns, name, hash);
742 /* someone won the race and added the resource before */
743 ldlm_resource_getref(old_res);
744 spin_unlock(&ns->ns_hash_lock);
745 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
746 /* synchronize WRT resource creation */
747 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
748 down(&old_res->lr_lvb_sem);
749 up(&old_res->lr_lvb_sem);
754 /* we won! let's add the resource */
755 bucket = ns->ns_hash + hash;
756 list_add(&res->lr_hash, bucket);
758 ldlm_namespace_get_nolock(ns);
760 if (parent == NULL) {
761 list_add(&res->lr_childof, &ns->ns_root_list);
763 res->lr_parent = parent;
764 list_add(&res->lr_childof, &parent->lr_children);
766 spin_unlock(&ns->ns_hash_lock);
768 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
771 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
772 rc = ns->ns_lvbo->lvbo_init(res);
774 CERROR("lvbo_init failed for resource "
775 LPU64": rc %d\n", name->name[0], rc);
776 /* we create resource with locked lr_lvb_sem */
777 up(&res->lr_lvb_sem);
783 /* Args: unlocked namespace
784 * Locks: takes and releases ns->ns_lock and res->lr_lock
785 * Returns: referenced, unlocked ldlm_resource or NULL */
786 struct ldlm_resource *
787 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
788 const struct ldlm_res_id *name, ldlm_type_t type, int create)
790 __u32 hash = ldlm_hash_fn(parent, name);
791 struct ldlm_resource *res = NULL;
795 LASSERT(ns->ns_hash != NULL);
796 LASSERT(name->name[0] != 0);
798 spin_lock(&ns->ns_hash_lock);
799 res = ldlm_resource_find(ns, name, hash);
801 ldlm_resource_getref(res);
802 spin_unlock(&ns->ns_hash_lock);
803 /* synchronize WRT resource creation */
804 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
805 down(&res->lr_lvb_sem);
806 up(&res->lr_lvb_sem);
810 spin_unlock(&ns->ns_hash_lock);
815 res = ldlm_resource_add(ns, parent, name, hash, type);
819 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
821 LASSERT(res != NULL);
822 LASSERT(res != LP_POISON);
823 atomic_inc(&res->lr_refcount);
824 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
825 atomic_read(&res->lr_refcount));
829 void __ldlm_resource_putref_final(struct ldlm_resource *res)
831 struct ldlm_namespace *ns = res->lr_namespace;
833 LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
835 if (!list_empty(&res->lr_granted)) {
836 ldlm_resource_dump(D_ERROR, res);
840 if (!list_empty(&res->lr_converting)) {
841 ldlm_resource_dump(D_ERROR, res);
845 if (!list_empty(&res->lr_waiting)) {
846 ldlm_resource_dump(D_ERROR, res);
850 if (!list_empty(&res->lr_children)) {
851 ldlm_resource_dump(D_ERROR, res);
855 /* Pass 0 here to not wake ->ns_waitq up yet, we will do it few
856 * lines below when all children are freed. */
857 ldlm_namespace_put_nolock(ns, 0);
858 list_del_init(&res->lr_hash);
859 list_del_init(&res->lr_childof);
862 if (ns->ns_resources == 0)
863 wake_up(&ns->ns_waitq);
866 /* Returns 1 if the resource was freed, 0 if it remains. */
867 int ldlm_resource_putref(struct ldlm_resource *res)
869 struct ldlm_namespace *ns = res->lr_namespace;
873 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
874 atomic_read(&res->lr_refcount) - 1);
875 LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
876 atomic_read(&res->lr_refcount));
877 LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
878 atomic_read(&res->lr_refcount));
880 if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
881 __ldlm_resource_putref_final(res);
882 spin_unlock(&ns->ns_hash_lock);
883 if (res->lr_lvb_data)
884 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
885 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
892 /* Returns 1 if the resource was freed, 0 if it remains. */
893 int ldlm_resource_putref_locked(struct ldlm_resource *res)
898 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
899 atomic_read(&res->lr_refcount) - 1);
900 LASSERT(atomic_read(&res->lr_refcount) > 0);
901 LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
903 LASSERT(atomic_read(&res->lr_refcount) >= 0);
904 if (atomic_dec_and_test(&res->lr_refcount)) {
905 __ldlm_resource_putref_final(res);
906 if (res->lr_lvb_data)
907 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
908 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
915 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
916 struct ldlm_lock *lock)
918 check_res_locked(res);
920 ldlm_resource_dump(D_OTHER, res);
921 CDEBUG(D_OTHER, "About to add this lock:\n");
922 ldlm_lock_dump(D_OTHER, lock, 0);
924 if (lock->l_destroyed) {
925 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
929 LASSERT(list_empty(&lock->l_res_link));
931 list_add_tail(&lock->l_res_link, head);
934 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
935 struct ldlm_lock *new)
937 struct ldlm_resource *res = original->l_resource;
939 check_res_locked(res);
941 ldlm_resource_dump(D_OTHER, res);
942 CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
943 ldlm_lock_dump(D_OTHER, new, 0);
945 if (new->l_destroyed) {
946 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
950 LASSERT(list_empty(&new->l_res_link));
952 list_add(&new->l_res_link, &original->l_res_link);
956 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
958 int type = lock->l_resource->lr_type;
960 check_res_locked(lock->l_resource);
961 if (type == LDLM_IBITS || type == LDLM_PLAIN)
962 ldlm_unlink_lock_skiplist(lock);
963 else if (type == LDLM_EXTENT)
964 ldlm_extent_unlink_lock(lock);
965 list_del_init(&lock->l_res_link);
968 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
970 desc->lr_type = res->lr_type;
971 desc->lr_name = res->lr_name;
974 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
976 struct list_head *tmp;
978 if (!((libcfs_debug | D_ERROR) & level))
981 mutex_down(ldlm_namespace_lock(client));
983 list_for_each(tmp, ldlm_namespace_list(client)) {
984 struct ldlm_namespace *ns;
985 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
986 ldlm_namespace_dump(level, ns);
989 mutex_up(ldlm_namespace_lock(client));
992 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
994 struct list_head *tmp;
996 if (!((libcfs_debug | D_ERROR) & level))
999 CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1000 ns->ns_name, ns->ns_refcount,
1001 ns_is_client(ns) ? "client" : "server");
1003 if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1006 spin_lock(&ns->ns_hash_lock);
1007 tmp = ns->ns_root_list.next;
1008 while (tmp != &ns->ns_root_list) {
1009 struct ldlm_resource *res;
1010 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1012 ldlm_resource_getref(res);
1013 spin_unlock(&ns->ns_hash_lock);
1016 ldlm_resource_dump(level, res);
1019 spin_lock(&ns->ns_hash_lock);
1021 ldlm_resource_putref_locked(res);
1023 ns->ns_next_dump = cfs_time_shift(10);
1024 spin_unlock(&ns->ns_hash_lock);
1027 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1029 struct list_head *tmp;
1032 CLASSERT(RES_NAME_SIZE == 4);
1034 if (!((libcfs_debug | D_ERROR) & level))
1037 CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1038 ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1039 res->lr_name.name[2], res->lr_name.name[3],
1040 atomic_read(&res->lr_refcount));
1042 if (!list_empty(&res->lr_granted)) {
1044 CDEBUG(level, "Granted locks:\n");
1045 list_for_each(tmp, &res->lr_granted) {
1046 struct ldlm_lock *lock;
1047 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1048 ldlm_lock_dump(level, lock, ++pos);
1051 if (!list_empty(&res->lr_converting)) {
1053 CDEBUG(level, "Converting locks:\n");
1054 list_for_each(tmp, &res->lr_converting) {
1055 struct ldlm_lock *lock;
1056 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1057 ldlm_lock_dump(level, lock, ++pos);
1060 if (!list_empty(&res->lr_waiting)) {
1062 CDEBUG(level, "Waiting locks:\n");
1063 list_for_each(tmp, &res->lr_waiting) {
1064 struct ldlm_lock *lock;
1065 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1066 ldlm_lock_dump(level, lock, ++pos);