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 struct list_head ldlm_srv_namespace_list =
44 CFS_LIST_HEAD_INIT(ldlm_srv_namespace_list);
46 struct semaphore ldlm_cli_namespace_lock;
47 struct list_head ldlm_cli_namespace_list =
48 CFS_LIST_HEAD_INIT(ldlm_cli_namespace_list);
50 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
51 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
52 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
55 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
56 unsigned long count, void *data)
58 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
59 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
63 int ldlm_proc_setup(void)
66 struct lprocfs_vars list[] = {
67 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
70 LASSERT(ldlm_ns_proc_dir == NULL);
72 ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
75 if (IS_ERR(ldlm_type_proc_dir)) {
76 CERROR("LProcFS failed in ldlm-init\n");
77 rc = PTR_ERR(ldlm_type_proc_dir);
81 ldlm_ns_proc_dir = lprocfs_register("namespaces",
84 if (IS_ERR(ldlm_ns_proc_dir)) {
85 CERROR("LProcFS failed in ldlm-init\n");
86 rc = PTR_ERR(ldlm_ns_proc_dir);
90 ldlm_svc_proc_dir = lprocfs_register("services",
93 if (IS_ERR(ldlm_svc_proc_dir)) {
94 CERROR("LProcFS failed in ldlm-init\n");
95 rc = PTR_ERR(ldlm_svc_proc_dir);
99 rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
104 lprocfs_remove(&ldlm_ns_proc_dir);
106 lprocfs_remove(&ldlm_type_proc_dir);
108 ldlm_svc_proc_dir = NULL;
112 void ldlm_proc_cleanup(void)
114 if (ldlm_svc_proc_dir)
115 lprocfs_remove(&ldlm_svc_proc_dir);
117 if (ldlm_ns_proc_dir)
118 lprocfs_remove(&ldlm_ns_proc_dir);
120 if (ldlm_type_proc_dir)
121 lprocfs_remove(&ldlm_type_proc_dir);
124 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
125 int count, int *eof, void *data)
127 struct ldlm_namespace *ns = data;
128 __u32 *nr = &ns->ns_max_unused;
130 if (ns_connect_lru_resize(ns))
131 nr = &ns->ns_nr_unused;
132 return lprocfs_rd_uint(page, start, off, count, eof, nr);
135 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
136 unsigned long count, void *data)
138 struct ldlm_namespace *ns = data;
139 char dummy[MAX_STRING_SIZE + 1], *end;
143 dummy[MAX_STRING_SIZE] = '\0';
144 if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
147 if (count == 6 && memcmp(dummy, "clear", 5) == 0) {
149 "dropping all unused locks from namespace %s\n",
151 if (ns_connect_lru_resize(ns)) {
152 int canceled, unused = ns->ns_nr_unused;
154 /* Try to cancel all @ns_nr_unused locks. */
155 canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
157 if (canceled < unused) {
158 CERROR("not all requested locks are canceled, "
159 "requested: %d, canceled: %d\n", unused,
164 tmp = ns->ns_max_unused;
165 ns->ns_max_unused = 0;
166 ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
167 ns->ns_max_unused = tmp;
172 tmp = simple_strtoul(dummy, &end, 0);
174 CERROR("invalid value written\n");
177 lru_resize = (tmp == 0);
179 if (ns_connect_lru_resize(ns)) {
181 ns->ns_max_unused = (unsigned int)tmp;
183 if (tmp > ns->ns_nr_unused)
184 tmp = ns->ns_nr_unused;
185 tmp = ns->ns_nr_unused - tmp;
187 CDEBUG(D_DLMTRACE, "changing namespace %s unused locks from %u to %u\n",
188 ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
189 ldlm_cancel_lru(ns, (unsigned int)tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
192 CDEBUG(D_DLMTRACE, "disable lru_resize for namespace %s\n",
194 ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
197 CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
198 ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
199 ns->ns_max_unused = (unsigned int)tmp;
200 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
202 /* Make sure that originally lru resize was supported before
203 * turning it on here. */
205 (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
206 CDEBUG(D_DLMTRACE, "enable lru_resize for namespace %s\n",
208 ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
214 void ldlm_proc_namespace(struct ldlm_namespace *ns)
216 struct lprocfs_vars lock_vars[2];
217 char lock_name[MAX_STRING_SIZE + 1];
220 LASSERT(ns->ns_name != NULL);
222 lock_name[MAX_STRING_SIZE] = '\0';
224 memset(lock_vars, 0, sizeof(lock_vars));
225 lock_vars[0].name = lock_name;
227 snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
228 lock_vars[0].data = &ns->ns_refcount;
229 lock_vars[0].read_fptr = lprocfs_rd_atomic;
230 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
232 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
233 lock_vars[0].data = &ns->ns_locks;
234 lock_vars[0].read_fptr = lprocfs_rd_atomic;
235 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
237 if (ns_is_client(ns)) {
238 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
240 lock_vars[0].data = &ns->ns_nr_unused;
241 lock_vars[0].read_fptr = lprocfs_rd_uint;
242 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
244 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
246 lock_vars[0].data = ns;
247 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
248 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
249 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
251 snprintf(lock_name, MAX_STRING_SIZE, "%s/shrink_thumb",
253 lock_vars[0].data = ns;
254 lock_vars[0].read_fptr = lprocfs_rd_uint;
255 lock_vars[0].write_fptr = lprocfs_wr_uint;
256 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
258 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
260 lock_vars[0].data = &ns->ns_max_age;
261 lock_vars[0].read_fptr = lprocfs_rd_uint;
262 lock_vars[0].write_fptr = lprocfs_wr_uint;
263 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
265 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
267 lock_vars[0].data = &ns->ns_max_nolock_size;
268 lock_vars[0].read_fptr = lprocfs_rd_uint;
269 lock_vars[0].write_fptr = lprocfs_wr_uint;
270 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
272 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
274 lock_vars[0].data = &ns->ns_contention_time;
275 lock_vars[0].read_fptr = lprocfs_rd_uint;
276 lock_vars[0].write_fptr = lprocfs_wr_uint;
277 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
279 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
281 lock_vars[0].data = &ns->ns_contended_locks;
282 lock_vars[0].read_fptr = lprocfs_rd_uint;
283 lock_vars[0].write_fptr = lprocfs_wr_uint;
284 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
287 #undef MAX_STRING_SIZE
289 #define ldlm_proc_namespace(ns) do {} while (0)
292 struct ldlm_namespace *
293 ldlm_namespace_new(struct obd_device *obd, char *name,
294 ldlm_side_t client, ldlm_appetite_t apt)
296 struct ldlm_namespace *ns = NULL;
297 struct list_head *bucket;
298 int rc, idx, namelen;
303 CERROR("ldlm_get_ref failed: %d\n", rc);
311 OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
315 namelen = strlen(name);
316 OBD_ALLOC(ns->ns_name, namelen + 1);
318 GOTO(out_hash, NULL);
320 ns->ns_shrink_thumb = LDLM_LOCK_SHRINK_THUMB;
321 ns->ns_appetite = apt;
323 LASSERT(obd != NULL);
326 strcpy(ns->ns_name, name);
328 CFS_INIT_LIST_HEAD(&ns->ns_root_list);
329 CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
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 CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
346 ns->ns_nr_unused = 0;
347 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
348 ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
349 spin_lock_init(&ns->ns_unused_lock);
350 ns->ns_orig_connect_flags = 0;
351 ns->ns_connect_flags = 0;
353 ldlm_proc_namespace(ns);
355 idx = atomic_read(ldlm_namespace_nr(client));
357 rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
359 CERROR("Can't initialize lock pool, rc %d\n", rc);
363 at_init(&ns->ns_at_estimate, ldlm_enqueue_min, 0);
365 ldlm_namespace_register(ns, 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 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
507 /* At shutdown time, don't call the cancellation callback */
508 ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
510 if (ns->ns_refcount > 0) {
511 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
514 "dlm namespace %s free waiting on refcount %d\n",
515 ns->ns_name, ns->ns_refcount);
518 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
520 rc = l_wait_event(ns->ns_waitq,
521 ns->ns_refcount == 0, &lwi);
523 /* Forced cleanups should be able to reclaim all references,
524 * so it's safe to wait forever... we can't leak locks... */
525 if (force && rc == -ETIMEDOUT) {
526 LCONSOLE_ERROR("Forced cleanup waiting for %s "
527 "namespace with %d resources in use, "
528 "(rc=%d)\n", ns->ns_name,
529 ns->ns_refcount, rc);
530 GOTO(force_wait, rc);
533 if (ns->ns_refcount) {
534 LCONSOLE_ERROR("Cleanup waiting for %s namespace "
535 "with %d resources in use, (rc=%d)\n",
537 ns->ns_refcount, rc);
538 RETURN(ELDLM_NAMESPACE_EXISTS);
541 "dlm namespace %s free done waiting\n", ns->ns_name);
547 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
548 struct obd_import *imp,
558 /* Make sure that nobody can find this ns in its list. */
559 ldlm_namespace_unregister(ns, ns->ns_client);
561 /* Can fail with -EINTR when force == 0 in which case try harder */
562 rc = __ldlm_namespace_free(ns, force);
563 if (rc != ELDLM_OK) {
565 ptlrpc_disconnect_import(imp, 0);
566 ptlrpc_invalidate_import(imp);
569 /* With all requests dropped and the import inactive
570 * we are gaurenteed all reference will be dropped. */
571 rc = __ldlm_namespace_free(ns, 1);
577 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
585 /* Fini pool _before_ parent proc dir is removed. This is important as
586 * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
587 * it after @dir may cause oops. */
588 ldlm_pool_fini(&ns->ns_pool);
592 struct proc_dir_entry *dir;
593 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
595 CERROR("dlm namespace %s has no procfs dir?\n",
598 lprocfs_remove(&dir);
602 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
603 OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
605 /* @ns should be not on list in this time, otherwise this will cause
606 * issues realted to using freed @ns in pools thread. */
607 LASSERT(list_empty(&ns->ns_list_chain));
613 /* Cleanup the resource, and free namespace.
616 * proc1: destroy import
617 * class_disconnect_export(grab cl_sem) ->
618 * -> ldlm_namespace_free ->
619 * -> lprocfs_remove(grab _lprocfs_lock).
620 * proc2: read proc info
621 * lprocfs_fops_read(grab _lprocfs_lock) ->
622 * -> osc_rd_active, etc(grab cl_sem).
624 * So that I have to split the ldlm_namespace_free into two parts - the first
625 * part ldlm_namespace_free_prior is used to cleanup the resource which is
626 * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
627 * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
630 void ldlm_namespace_free(struct ldlm_namespace *ns,
631 struct obd_import *imp,
634 ldlm_namespace_free_prior(ns, imp, force);
635 ldlm_namespace_free_post(ns);
638 void ldlm_namespace_get_locked(struct ldlm_namespace *ns)
640 LASSERT(ns->ns_refcount >= 0);
644 void ldlm_namespace_get(struct ldlm_namespace *ns)
646 spin_lock(&ns->ns_hash_lock);
647 ldlm_namespace_get_locked(ns);
648 spin_unlock(&ns->ns_hash_lock);
651 void ldlm_namespace_put_locked(struct ldlm_namespace *ns, int wakeup)
653 LASSERT(ns->ns_refcount > 0);
655 if (ns->ns_refcount == 0 && wakeup)
656 wake_up(&ns->ns_waitq);
659 void ldlm_namespace_put(struct ldlm_namespace *ns, int wakeup)
661 spin_lock(&ns->ns_hash_lock);
662 ldlm_namespace_put_locked(ns, wakeup);
663 spin_unlock(&ns->ns_hash_lock);
666 /* Register @ns in the list of namespaces */
667 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
669 mutex_down(ldlm_namespace_lock(client));
670 LASSERT(list_empty(&ns->ns_list_chain));
671 list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
672 atomic_inc(ldlm_namespace_nr(client));
673 mutex_up(ldlm_namespace_lock(client));
676 /* Unregister @ns from the list of namespaces */
677 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
679 mutex_down(ldlm_namespace_lock(client));
680 LASSERT(!list_empty(&ns->ns_list_chain));
682 * Some asserts and possibly other parts of code still using
683 * list_empty(&ns->ns_list_chain). This is why it is important
684 * to use list_del_init() here.
686 list_del_init(&ns->ns_list_chain);
687 atomic_dec(ldlm_namespace_nr(client));
688 mutex_up(ldlm_namespace_lock(client));
691 /* Should be called under ldlm_namespace_lock(client) taken */
692 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
694 LASSERT(!list_empty(&ns->ns_list_chain));
695 LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
696 list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
699 /* Should be called under ldlm_namespace_lock(client) taken */
700 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
702 LASSERT_SEM_LOCKED(ldlm_namespace_lock(client));
703 LASSERT(!list_empty(ldlm_namespace_list(client)));
704 return container_of(ldlm_namespace_list(client)->next,
705 struct ldlm_namespace, ns_list_chain);
708 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, struct ldlm_res_id name)
713 for (i = 0; i < RES_NAME_SIZE; i++)
714 hash += name.name[i];
716 hash += (__u32)((unsigned long)parent >> 4);
718 return (hash & RES_HASH_MASK);
721 static struct ldlm_resource *ldlm_resource_new(void)
723 struct ldlm_resource *res;
726 OBD_SLAB_ALLOC(res, ldlm_resource_slab, CFS_ALLOC_IO, sizeof *res);
730 memset(res, 0, sizeof(*res));
732 CFS_INIT_LIST_HEAD(&res->lr_children);
733 CFS_INIT_LIST_HEAD(&res->lr_childof);
734 CFS_INIT_LIST_HEAD(&res->lr_granted);
735 CFS_INIT_LIST_HEAD(&res->lr_converting);
736 CFS_INIT_LIST_HEAD(&res->lr_waiting);
738 /* initialize interval trees for each lock mode*/
739 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
740 res->lr_itree[idx].lit_size = 0;
741 res->lr_itree[idx].lit_mode = 1 << idx;
742 res->lr_itree[idx].lit_root = NULL;
745 atomic_set(&res->lr_refcount, 1);
746 spin_lock_init(&res->lr_lock);
748 /* one who creates the resource must unlock
749 * the semaphore after lvb initialization */
750 init_MUTEX_LOCKED(&res->lr_lvb_sem);
755 /* must be called with hash lock held */
756 static struct ldlm_resource *
757 ldlm_resource_find(struct ldlm_namespace *ns, struct ldlm_res_id name, __u32 hash)
759 struct list_head *bucket, *tmp;
760 struct ldlm_resource *res;
762 LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
763 bucket = ns->ns_hash + hash;
765 list_for_each(tmp, bucket) {
766 res = list_entry(tmp, struct ldlm_resource, lr_hash);
767 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0)
774 /* Args: locked namespace
775 * Returns: newly-allocated, referenced, unlocked resource */
776 static struct ldlm_resource *
777 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
778 struct ldlm_res_id name, __u32 hash, ldlm_type_t type)
780 struct list_head *bucket;
781 struct ldlm_resource *res, *old_res;
784 LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
787 res = ldlm_resource_new();
792 res->lr_namespace = ns;
794 res->lr_most_restr = LCK_NL;
796 spin_lock(&ns->ns_hash_lock);
797 old_res = ldlm_resource_find(ns, name, hash);
799 /* someone won the race and added the resource before */
800 ldlm_resource_getref(old_res);
801 spin_unlock(&ns->ns_hash_lock);
802 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
803 /* synchronize WRT resource creation */
804 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
805 down(&old_res->lr_lvb_sem);
806 up(&old_res->lr_lvb_sem);
811 /* we won! let's add the resource */
812 bucket = ns->ns_hash + hash;
813 list_add(&res->lr_hash, bucket);
815 ldlm_namespace_get_locked(ns);
817 if (parent == NULL) {
818 list_add(&res->lr_childof, &ns->ns_root_list);
820 res->lr_parent = parent;
821 list_add(&res->lr_childof, &parent->lr_children);
823 spin_unlock(&ns->ns_hash_lock);
825 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
828 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
829 rc = ns->ns_lvbo->lvbo_init(res);
831 CERROR("lvbo_init failed for resource "
832 LPU64": rc %d\n", name.name[0], rc);
833 /* we create resource with locked lr_lvb_sem */
834 up(&res->lr_lvb_sem);
840 /* Args: unlocked namespace
841 * Locks: takes and releases ns->ns_lock and res->lr_lock
842 * Returns: referenced, unlocked ldlm_resource or NULL */
843 struct ldlm_resource *
844 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
845 struct ldlm_res_id name, ldlm_type_t type, int create)
847 __u32 hash = ldlm_hash_fn(parent, name);
848 struct ldlm_resource *res = NULL;
852 LASSERT(ns->ns_hash != NULL);
853 LASSERT(name.name[0] != 0);
855 spin_lock(&ns->ns_hash_lock);
856 res = ldlm_resource_find(ns, name, hash);
858 ldlm_resource_getref(res);
859 spin_unlock(&ns->ns_hash_lock);
860 /* synchronize WRT resource creation */
861 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
862 down(&res->lr_lvb_sem);
863 up(&res->lr_lvb_sem);
867 spin_unlock(&ns->ns_hash_lock);
872 res = ldlm_resource_add(ns, parent, name, hash, type);
876 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
878 LASSERT(res != NULL);
879 LASSERT(res != LP_POISON);
880 atomic_inc(&res->lr_refcount);
881 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
882 atomic_read(&res->lr_refcount));
886 void __ldlm_resource_putref_final(struct ldlm_resource *res)
888 struct ldlm_namespace *ns = res->lr_namespace;
890 LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
892 if (!list_empty(&res->lr_granted)) {
893 ldlm_resource_dump(D_ERROR, res);
897 if (!list_empty(&res->lr_converting)) {
898 ldlm_resource_dump(D_ERROR, res);
902 if (!list_empty(&res->lr_waiting)) {
903 ldlm_resource_dump(D_ERROR, res);
907 if (!list_empty(&res->lr_children)) {
908 ldlm_resource_dump(D_ERROR, res);
912 /* Pass 0 as second argument to not wake up ->ns_waitq yet, will do it
914 ldlm_namespace_put_locked(ns, 0);
915 list_del_init(&res->lr_hash);
916 list_del_init(&res->lr_childof);
919 if (ns->ns_resources == 0)
920 wake_up(&ns->ns_waitq);
923 /* Returns 1 if the resource was freed, 0 if it remains. */
924 int ldlm_resource_putref(struct ldlm_resource *res)
926 struct ldlm_namespace *ns = res->lr_namespace;
930 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
931 atomic_read(&res->lr_refcount) - 1);
932 LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
933 atomic_read(&res->lr_refcount));
934 LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
935 atomic_read(&res->lr_refcount));
937 if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
938 __ldlm_resource_putref_final(res);
939 spin_unlock(&ns->ns_hash_lock);
940 if (res->lr_lvb_data)
941 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
942 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
949 /* Returns 1 if the resource was freed, 0 if it remains. */
950 int ldlm_resource_putref_locked(struct ldlm_resource *res)
955 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
956 atomic_read(&res->lr_refcount) - 1);
957 LASSERT(atomic_read(&res->lr_refcount) > 0);
958 LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
960 LASSERT(atomic_read(&res->lr_refcount) >= 0);
961 if (atomic_dec_and_test(&res->lr_refcount)) {
962 __ldlm_resource_putref_final(res);
963 if (res->lr_lvb_data)
964 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
965 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
972 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
973 struct ldlm_lock *lock)
975 check_res_locked(res);
977 ldlm_resource_dump(D_OTHER, res);
978 CDEBUG(D_OTHER, "About to add this lock:\n");
979 ldlm_lock_dump(D_OTHER, lock, 0);
981 if (lock->l_destroyed) {
982 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
986 LASSERT(list_empty(&lock->l_res_link));
988 list_add_tail(&lock->l_res_link, head);
991 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
992 struct ldlm_lock *new)
994 struct ldlm_resource *res = original->l_resource;
996 check_res_locked(res);
998 ldlm_resource_dump(D_OTHER, res);
999 CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1000 ldlm_lock_dump(D_OTHER, new, 0);
1002 if (new->l_destroyed) {
1003 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1007 LASSERT(list_empty(&new->l_res_link));
1009 list_add(&new->l_res_link, &original->l_res_link);
1013 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1015 int type = lock->l_resource->lr_type;
1017 check_res_locked(lock->l_resource);
1018 if (type == LDLM_IBITS || type == LDLM_PLAIN)
1019 ldlm_unlink_lock_skiplist(lock);
1020 else if (type == LDLM_EXTENT)
1021 ldlm_extent_unlink_lock(lock);
1022 list_del_init(&lock->l_res_link);
1025 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1027 desc->lr_type = res->lr_type;
1028 desc->lr_name = res->lr_name;
1031 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1033 struct list_head *tmp;
1035 if (!((libcfs_debug | D_ERROR) & level))
1038 mutex_down(ldlm_namespace_lock(client));
1040 list_for_each(tmp, ldlm_namespace_list(client)) {
1041 struct ldlm_namespace *ns;
1042 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1043 ldlm_namespace_dump(level, ns);
1046 mutex_up(ldlm_namespace_lock(client));
1049 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1051 struct list_head *tmp;
1053 if (!((libcfs_debug | D_ERROR) & level))
1056 CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1057 ns->ns_name, ns->ns_refcount,
1058 ns_is_client(ns) ? "client" : "server");
1060 if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1063 spin_lock(&ns->ns_hash_lock);
1064 tmp = ns->ns_root_list.next;
1065 while (tmp != &ns->ns_root_list) {
1066 struct ldlm_resource *res;
1067 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1069 ldlm_resource_getref(res);
1070 spin_unlock(&ns->ns_hash_lock);
1073 ldlm_resource_dump(level, res);
1076 spin_lock(&ns->ns_hash_lock);
1078 ldlm_resource_putref_locked(res);
1080 ns->ns_next_dump = cfs_time_shift(10);
1081 spin_unlock(&ns->ns_hash_lock);
1084 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1086 struct list_head *tmp;
1089 CLASSERT(RES_NAME_SIZE == 4);
1091 if (!((libcfs_debug | D_ERROR) & level))
1094 CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1095 ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1096 res->lr_name.name[2], res->lr_name.name[3],
1097 atomic_read(&res->lr_refcount));
1099 if (!list_empty(&res->lr_granted)) {
1101 CDEBUG(level, "Granted locks:\n");
1102 list_for_each(tmp, &res->lr_granted) {
1103 struct ldlm_lock *lock;
1104 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1105 ldlm_lock_dump(level, lock, ++pos);
1108 if (!list_empty(&res->lr_converting)) {
1110 CDEBUG(level, "Converting locks:\n");
1111 list_for_each(tmp, &res->lr_converting) {
1112 struct ldlm_lock *lock;
1113 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1114 ldlm_lock_dump(level, lock, ++pos);
1117 if (!list_empty(&res->lr_waiting)) {
1119 CDEBUG(level, "Waiting locks:\n");
1120 list_for_each(tmp, &res->lr_waiting) {
1121 struct ldlm_lock *lock;
1122 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1123 ldlm_lock_dump(level, lock, ++pos);