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 2008 Sun Microsystems, Inc. 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 atomic_t ldlm_srv_namespace_nr = ATOMIC_INIT(0);
55 atomic_t ldlm_cli_namespace_nr = ATOMIC_INIT(0);
57 struct semaphore ldlm_srv_namespace_lock;
58 struct list_head ldlm_srv_namespace_list =
59 CFS_LIST_HEAD_INIT(ldlm_srv_namespace_list);
61 struct semaphore ldlm_cli_namespace_lock;
62 struct list_head ldlm_cli_namespace_list =
63 CFS_LIST_HEAD_INIT(ldlm_cli_namespace_list);
65 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
66 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
67 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
69 extern unsigned int ldlm_cancel_unused_locks_before_replay;
72 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
73 unsigned long count, void *data)
75 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
76 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
80 int ldlm_proc_setup(void)
83 struct lprocfs_vars list[] = {
84 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
85 { "cancel_unused_locks_before_replay",
86 lprocfs_rd_uint, lprocfs_wr_uint,
87 &ldlm_cancel_unused_locks_before_replay, NULL },
90 LASSERT(ldlm_ns_proc_dir == NULL);
92 ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
95 if (IS_ERR(ldlm_type_proc_dir)) {
96 CERROR("LProcFS failed in ldlm-init\n");
97 rc = PTR_ERR(ldlm_type_proc_dir);
101 ldlm_ns_proc_dir = lprocfs_register("namespaces",
104 if (IS_ERR(ldlm_ns_proc_dir)) {
105 CERROR("LProcFS failed in ldlm-init\n");
106 rc = PTR_ERR(ldlm_ns_proc_dir);
110 ldlm_svc_proc_dir = lprocfs_register("services",
113 if (IS_ERR(ldlm_svc_proc_dir)) {
114 CERROR("LProcFS failed in ldlm-init\n");
115 rc = PTR_ERR(ldlm_svc_proc_dir);
119 rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
124 lprocfs_remove(&ldlm_ns_proc_dir);
126 lprocfs_remove(&ldlm_type_proc_dir);
128 ldlm_svc_proc_dir = NULL;
132 void ldlm_proc_cleanup(void)
134 if (ldlm_svc_proc_dir)
135 lprocfs_remove(&ldlm_svc_proc_dir);
137 if (ldlm_ns_proc_dir)
138 lprocfs_remove(&ldlm_ns_proc_dir);
140 if (ldlm_type_proc_dir)
141 lprocfs_remove(&ldlm_type_proc_dir);
144 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
145 int count, int *eof, void *data)
147 struct ldlm_namespace *ns = data;
148 __u32 *nr = &ns->ns_max_unused;
150 if (ns_connect_lru_resize(ns))
151 nr = &ns->ns_nr_unused;
152 return lprocfs_rd_uint(page, start, off, count, eof, nr);
155 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
156 unsigned long count, void *data)
158 struct ldlm_namespace *ns = data;
159 char dummy[MAX_STRING_SIZE + 1] = { '\0' }, *end;
163 if (count >= sizeof(dummy) || count == 0)
166 if (copy_from_user(dummy, buffer, count))
169 if (strncmp(dummy, "clear", 5) == 0) {
171 "dropping all unused locks from namespace %s\n",
173 if (ns_connect_lru_resize(ns)) {
174 int canceled, unused = ns->ns_nr_unused;
176 /* Try to cancel all @ns_nr_unused locks. */
177 canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
179 if (canceled < unused) {
181 "not all requested locks are canceled, "
182 "requested: %d, canceled: %d\n", unused,
187 tmp = ns->ns_max_unused;
188 ns->ns_max_unused = 0;
189 ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
190 ns->ns_max_unused = tmp;
195 tmp = simple_strtoul(dummy, &end, 0);
197 CERROR("invalid value written\n");
200 lru_resize = (tmp == 0);
202 if (ns_connect_lru_resize(ns)) {
204 ns->ns_max_unused = (unsigned int)tmp;
206 if (tmp > ns->ns_nr_unused)
207 tmp = ns->ns_nr_unused;
208 tmp = ns->ns_nr_unused - tmp;
211 "changing namespace %s unused locks from %u to %u\n",
212 ns->ns_name, ns->ns_nr_unused, (unsigned int)tmp);
213 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
217 "disable lru_resize for namespace %s\n",
219 ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
223 "changing namespace %s max_unused from %u to %u\n",
224 ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
225 ns->ns_max_unused = (unsigned int)tmp;
226 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
228 /* Make sure that originally lru resize was supported before
229 * turning it on here. */
231 (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
233 "enable lru_resize for namespace %s\n",
235 ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
241 void ldlm_proc_namespace(struct ldlm_namespace *ns)
243 struct lprocfs_vars lock_vars[2];
244 char lock_name[MAX_STRING_SIZE + 1];
247 LASSERT(ns->ns_name != NULL);
249 lock_name[MAX_STRING_SIZE] = '\0';
251 memset(lock_vars, 0, sizeof(lock_vars));
252 lock_vars[0].name = lock_name;
254 snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
255 lock_vars[0].data = &ns->ns_refcount;
256 lock_vars[0].read_fptr = lprocfs_rd_atomic;
257 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
259 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
260 lock_vars[0].data = &ns->ns_locks;
261 lock_vars[0].read_fptr = lprocfs_rd_atomic;
262 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
264 if (ns_is_client(ns)) {
265 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
267 lock_vars[0].data = &ns->ns_nr_unused;
268 lock_vars[0].read_fptr = lprocfs_rd_uint;
269 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
271 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
273 lock_vars[0].data = ns;
274 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
275 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
276 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
278 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
280 lock_vars[0].data = &ns->ns_max_age;
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/lock_timeouts",
287 lock_vars[0].data = &ns->ns_timeouts;
288 lock_vars[0].read_fptr = lprocfs_rd_uint;
289 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
291 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
293 lock_vars[0].data = &ns->ns_max_nolock_size;
294 lock_vars[0].read_fptr = lprocfs_rd_uint;
295 lock_vars[0].write_fptr = lprocfs_wr_uint;
296 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
298 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
300 lock_vars[0].data = &ns->ns_contention_time;
301 lock_vars[0].read_fptr = lprocfs_rd_uint;
302 lock_vars[0].write_fptr = lprocfs_wr_uint;
303 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
305 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
307 lock_vars[0].data = &ns->ns_contended_locks;
308 lock_vars[0].read_fptr = lprocfs_rd_uint;
309 lock_vars[0].write_fptr = lprocfs_wr_uint;
310 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
313 #undef MAX_STRING_SIZE
315 #define ldlm_proc_namespace(ns) do {} while (0)
318 struct ldlm_namespace *
319 ldlm_namespace_new(struct obd_device *obd, char *name,
320 ldlm_side_t client, ldlm_appetite_t apt)
322 struct ldlm_namespace *ns = NULL;
323 struct list_head *bucket;
324 int rc, idx, namelen;
329 CERROR("ldlm_get_ref failed: %d\n", rc);
337 OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
341 namelen = strlen(name);
342 OBD_ALLOC(ns->ns_name, namelen + 1);
344 GOTO(out_hash, NULL);
346 ns->ns_appetite = apt;
348 LASSERT(obd != NULL);
351 strcpy(ns->ns_name, name);
353 CFS_INIT_LIST_HEAD(&ns->ns_root_list);
354 CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
356 ns->ns_client = client;
357 spin_lock_init(&ns->ns_hash_lock);
358 atomic_set(&ns->ns_locks, 0);
359 ns->ns_resources = 0;
360 cfs_waitq_init(&ns->ns_waitq);
361 ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
362 ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
363 ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
365 for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
367 CFS_INIT_LIST_HEAD(bucket);
369 CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
370 CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
371 ns->ns_nr_unused = 0;
372 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
373 ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
375 spin_lock_init(&ns->ns_unused_lock);
376 ns->ns_orig_connect_flags = 0;
377 ns->ns_connect_flags = 0;
379 ldlm_proc_namespace(ns);
381 idx = atomic_read(ldlm_namespace_nr(client));
383 rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
385 CERROR("Can't initialize lock pool, rc %d\n", rc);
389 at_init(&ns->ns_at_estimate, ldlm_enqueue_min, 0);
390 ldlm_namespace_register(ns, client);
393 ldlm_namespace_cleanup(ns, 0);
394 OBD_FREE(ns->ns_name, namelen + 1);
396 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
404 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
406 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
407 * This is currently only used for recovery, and we make certain assumptions
408 * as a result--notably, that we shouldn't cancel locks with refs. -phil
410 * Called with the ns_lock held. */
411 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
414 struct list_head *tmp;
415 int rc = 0, client = ns_is_client(res->lr_namespace);
416 int local_only = (flags & LDLM_FL_LOCAL_ONLY);
420 struct ldlm_lock *lock = NULL;
422 /* first, we look for non-cleaned-yet lock
423 * all cleaned locks are marked by CLEANED flag */
425 list_for_each(tmp, q) {
426 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
427 if (lock->l_flags & LDLM_FL_CLEANED) {
432 lock->l_flags |= LDLM_FL_CLEANED;
441 /* Set CBPENDING so nothing in the cancellation path
442 * can match this lock */
443 lock->l_flags |= LDLM_FL_CBPENDING;
444 lock->l_flags |= LDLM_FL_FAILED;
445 lock->l_flags |= flags;
447 /* ... without sending a CANCEL message for local_only. */
449 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
451 if (local_only && (lock->l_readers || lock->l_writers)) {
452 /* This is a little bit gross, but much better than the
453 * alternative: pretend that we got a blocking AST from
454 * the server, so that when the lock is decref'd, it
455 * will go away ... */
457 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
458 if (lock->l_completion_ast)
459 lock->l_completion_ast(lock, 0, NULL);
465 struct lustre_handle lockh;
468 ldlm_lock2handle(lock, &lockh);
469 rc = ldlm_cli_cancel(&lockh);
471 CERROR("ldlm_cli_cancel: %d\n", rc);
473 ldlm_resource_unlink_lock(lock);
475 LDLM_DEBUG(lock, "Freeing a lock still held by a "
477 ldlm_lock_destroy(lock);
485 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
487 struct list_head *tmp;
491 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
495 for (i = 0; i < RES_HASH_SIZE; i++) {
496 spin_lock(&ns->ns_hash_lock);
497 tmp = ns->ns_hash[i].next;
498 while (tmp != &(ns->ns_hash[i])) {
499 struct ldlm_resource *res;
500 res = list_entry(tmp, struct ldlm_resource, lr_hash);
501 ldlm_resource_getref(res);
502 spin_unlock(&ns->ns_hash_lock);
504 cleanup_resource(res, &res->lr_granted, flags);
505 cleanup_resource(res, &res->lr_converting, flags);
506 cleanup_resource(res, &res->lr_waiting, flags);
508 spin_lock(&ns->ns_hash_lock);
511 /* XXX: former stuff caused issues in case of race
512 * between ldlm_namespace_cleanup() and lockd() when
513 * client gets blocking ast when lock gets distracted by
514 * server. This is 1_4 branch solution, let's see how
516 if (!ldlm_resource_putref_locked(res)) {
517 CERROR("Namespace %s resource refcount nonzero "
518 "(%d) after lock cleanup; forcing cleanup.\n",
519 ns->ns_name, atomic_read(&res->lr_refcount));
520 CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
521 LPU64") (rc: %d)\n", res,
522 res->lr_name.name[0], res->lr_name.name[1],
523 res->lr_name.name[2], res->lr_name.name[3],
524 atomic_read(&res->lr_refcount));
527 spin_unlock(&ns->ns_hash_lock);
533 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
536 /* At shutdown time, don't call the cancellation callback */
537 ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
539 if (ns->ns_refcount > 0) {
540 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
543 "dlm namespace %s free waiting on refcount %d\n",
544 ns->ns_name, ns->ns_refcount);
547 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
549 rc = l_wait_event(ns->ns_waitq,
550 ns->ns_refcount == 0, &lwi);
552 /* Forced cleanups should be able to reclaim all references,
553 * so it's safe to wait forever... we can't leak locks... */
554 if (force && rc == -ETIMEDOUT) {
555 LCONSOLE_ERROR("Forced cleanup waiting for %s "
556 "namespace with %d resources in use, "
557 "(rc=%d)\n", ns->ns_name,
558 ns->ns_refcount, rc);
559 GOTO(force_wait, rc);
562 if (ns->ns_refcount) {
563 LCONSOLE_ERROR("Cleanup waiting for %s namespace "
564 "with %d resources in use, (rc=%d)\n",
566 ns->ns_refcount, rc);
567 RETURN(ELDLM_NAMESPACE_EXISTS);
570 "dlm namespace %s free done waiting\n", ns->ns_name);
576 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
577 struct obd_import *imp,
588 /* Can fail with -EINTR when force == 0 in which case try harder */
589 rc = __ldlm_namespace_free(ns, force);
590 if (rc != ELDLM_OK) {
592 ptlrpc_disconnect_import(imp, 0);
593 ptlrpc_invalidate_import(imp);
596 /* With all requests dropped and the import inactive
597 * we are gaurenteed all reference will be dropped. */
598 rc = __ldlm_namespace_free(ns, 1);
604 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
612 /* Make sure that nobody can find this ns in its list. */
613 ldlm_namespace_unregister(ns, ns->ns_client);
615 /* Fini pool _before_ parent proc dir is removed. This is important
616 * as ldlm_pool_fini() removes own proc dir which is child to @dir.
617 * Removing it after @dir may cause oops. */
618 ldlm_pool_fini(&ns->ns_pool);
622 struct proc_dir_entry *dir;
623 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
625 CERROR("dlm namespace %s has no procfs dir?\n",
628 lprocfs_remove(&dir);
632 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
633 OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
635 /* @ns should be not on list in this time, otherwise this will cause
636 * issues realted to using freed @ns in pools thread. */
637 LASSERT(list_empty(&ns->ns_list_chain));
643 /* Cleanup the resource, and free namespace.
646 * proc1: destroy import
647 * class_disconnect_export(grab cl_sem) ->
648 * -> ldlm_namespace_free ->
649 * -> lprocfs_remove(grab _lprocfs_lock).
650 * proc2: read proc info
651 * lprocfs_fops_read(grab _lprocfs_lock) ->
652 * -> osc_rd_active, etc(grab cl_sem).
654 * So that I have to split the ldlm_namespace_free into two parts - the first
655 * part ldlm_namespace_free_prior is used to cleanup the resource which is
656 * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
657 * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
660 void ldlm_namespace_free(struct ldlm_namespace *ns,
661 struct obd_import *imp,
664 ldlm_namespace_free_prior(ns, imp, force);
665 ldlm_namespace_free_post(ns);
668 void ldlm_namespace_get_locked(struct ldlm_namespace *ns)
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);
737 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, 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, struct ldlm_res_id name, __u32 hash)
788 struct list_head *bucket, *tmp;
789 struct ldlm_resource *res;
791 LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
792 bucket = ns->ns_hash + hash;
794 list_for_each(tmp, bucket) {
795 res = list_entry(tmp, struct ldlm_resource, lr_hash);
796 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0)
803 /* Args: locked namespace
804 * Returns: newly-allocated, referenced, unlocked resource */
805 static struct ldlm_resource *
806 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
807 struct ldlm_res_id name, __u32 hash, ldlm_type_t type)
809 struct list_head *bucket;
810 struct ldlm_resource *res, *old_res;
813 LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
816 res = ldlm_resource_new();
821 res->lr_namespace = ns;
823 res->lr_most_restr = LCK_NL;
825 spin_lock(&ns->ns_hash_lock);
826 old_res = ldlm_resource_find(ns, name, hash);
828 /* someone won the race and added the resource before */
829 ldlm_resource_getref(old_res);
830 spin_unlock(&ns->ns_hash_lock);
831 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
832 /* synchronize WRT resource creation */
833 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
834 down(&old_res->lr_lvb_sem);
835 up(&old_res->lr_lvb_sem);
840 /* we won! let's add the resource */
841 bucket = ns->ns_hash + hash;
842 list_add(&res->lr_hash, bucket);
844 ldlm_namespace_get_locked(ns);
846 if (parent == NULL) {
847 list_add(&res->lr_childof, &ns->ns_root_list);
849 res->lr_parent = parent;
850 list_add(&res->lr_childof, &parent->lr_children);
852 spin_unlock(&ns->ns_hash_lock);
854 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
857 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
858 rc = ns->ns_lvbo->lvbo_init(res);
860 CERROR("lvbo_init failed for resource "
861 LPU64": rc %d\n", name.name[0], rc);
862 /* we create resource with locked lr_lvb_sem */
863 up(&res->lr_lvb_sem);
869 /* Args: unlocked namespace
870 * Locks: takes and releases ns->ns_lock and res->lr_lock
871 * Returns: referenced, unlocked ldlm_resource or NULL */
872 struct ldlm_resource *
873 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
874 struct ldlm_res_id name, ldlm_type_t type, int create)
876 __u32 hash = ldlm_hash_fn(parent, name);
877 struct ldlm_resource *res = NULL;
881 LASSERT(ns->ns_hash != NULL);
882 LASSERT(name.name[0] != 0);
884 spin_lock(&ns->ns_hash_lock);
885 res = ldlm_resource_find(ns, name, hash);
887 ldlm_resource_getref(res);
888 spin_unlock(&ns->ns_hash_lock);
889 /* synchronize WRT resource creation */
890 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
891 down(&res->lr_lvb_sem);
892 up(&res->lr_lvb_sem);
896 spin_unlock(&ns->ns_hash_lock);
901 res = ldlm_resource_add(ns, parent, name, hash, type);
905 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
907 LASSERT(res != NULL);
908 LASSERT(res != LP_POISON);
909 atomic_inc(&res->lr_refcount);
910 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
911 atomic_read(&res->lr_refcount));
915 void __ldlm_resource_putref_final(struct ldlm_resource *res)
917 struct ldlm_namespace *ns = res->lr_namespace;
919 LASSERT_SPIN_LOCKED(&ns->ns_hash_lock);
921 if (!list_empty(&res->lr_granted)) {
922 ldlm_resource_dump(D_ERROR, res);
926 if (!list_empty(&res->lr_converting)) {
927 ldlm_resource_dump(D_ERROR, res);
931 if (!list_empty(&res->lr_waiting)) {
932 ldlm_resource_dump(D_ERROR, res);
936 if (!list_empty(&res->lr_children)) {
937 ldlm_resource_dump(D_ERROR, res);
941 /* Pass 0 as second argument to not wake up ->ns_waitq yet, will do it
943 ldlm_namespace_put_locked(ns, 0);
944 list_del_init(&res->lr_hash);
945 list_del_init(&res->lr_childof);
948 if (ns->ns_resources == 0)
949 wake_up(&ns->ns_waitq);
952 /* Returns 1 if the resource was freed, 0 if it remains. */
953 int ldlm_resource_putref(struct ldlm_resource *res)
955 struct ldlm_namespace *ns = res->lr_namespace;
959 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
960 atomic_read(&res->lr_refcount) - 1);
961 LASSERTF(atomic_read(&res->lr_refcount) > 0, "%d",
962 atomic_read(&res->lr_refcount));
963 LASSERTF(atomic_read(&res->lr_refcount) < LI_POISON, "%d",
964 atomic_read(&res->lr_refcount));
966 if (atomic_dec_and_lock(&res->lr_refcount, &ns->ns_hash_lock)) {
967 __ldlm_resource_putref_final(res);
968 spin_unlock(&ns->ns_hash_lock);
969 if (res->lr_lvb_data)
970 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
971 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
978 /* Returns 1 if the resource was freed, 0 if it remains. */
979 int ldlm_resource_putref_locked(struct ldlm_resource *res)
984 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
985 atomic_read(&res->lr_refcount) - 1);
986 LASSERT(atomic_read(&res->lr_refcount) > 0);
987 LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
989 LASSERT(atomic_read(&res->lr_refcount) >= 0);
990 if (atomic_dec_and_test(&res->lr_refcount)) {
991 __ldlm_resource_putref_final(res);
992 if (res->lr_lvb_data)
993 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
994 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1001 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1002 struct ldlm_lock *lock)
1004 check_res_locked(res);
1006 ldlm_resource_dump(D_INFO, res);
1007 CDEBUG(D_OTHER, "About to add this lock:\n");
1008 ldlm_lock_dump(D_OTHER, lock, 0);
1010 if (lock->l_destroyed) {
1011 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1015 LASSERT(list_empty(&lock->l_res_link));
1017 list_add_tail(&lock->l_res_link, head);
1020 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1021 struct ldlm_lock *new)
1023 struct ldlm_resource *res = original->l_resource;
1025 check_res_locked(res);
1027 ldlm_resource_dump(D_INFO, res);
1028 CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
1029 ldlm_lock_dump(D_OTHER, new, 0);
1031 if (new->l_destroyed) {
1032 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1036 LASSERT(list_empty(&new->l_res_link));
1038 list_add(&new->l_res_link, &original->l_res_link);
1042 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1044 int type = lock->l_resource->lr_type;
1046 check_res_locked(lock->l_resource);
1047 if (type == LDLM_IBITS || type == LDLM_PLAIN)
1048 ldlm_unlink_lock_skiplist(lock);
1049 else if (type == LDLM_EXTENT)
1050 ldlm_extent_unlink_lock(lock);
1051 list_del_init(&lock->l_res_link);
1054 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1056 desc->lr_type = res->lr_type;
1057 desc->lr_name = res->lr_name;
1060 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1062 struct list_head *tmp;
1064 if (!((libcfs_debug | D_ERROR) & level))
1067 mutex_down(ldlm_namespace_lock(client));
1069 list_for_each(tmp, ldlm_namespace_list(client)) {
1070 struct ldlm_namespace *ns;
1071 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1072 ldlm_namespace_dump(level, ns);
1075 mutex_up(ldlm_namespace_lock(client));
1078 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1080 struct list_head *tmp;
1082 if (!((libcfs_debug | D_ERROR) & level))
1085 CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1086 ns->ns_name, ns->ns_refcount,
1087 ns_is_client(ns) ? "client" : "server");
1089 if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1092 spin_lock(&ns->ns_hash_lock);
1093 tmp = ns->ns_root_list.next;
1094 while (tmp != &ns->ns_root_list) {
1095 struct ldlm_resource *res;
1096 res = list_entry(tmp, struct ldlm_resource, lr_childof);
1098 ldlm_resource_getref(res);
1099 spin_unlock(&ns->ns_hash_lock);
1102 ldlm_resource_dump(level, res);
1105 spin_lock(&ns->ns_hash_lock);
1107 ldlm_resource_putref_locked(res);
1109 ns->ns_next_dump = cfs_time_shift(10);
1110 spin_unlock(&ns->ns_hash_lock);
1113 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1115 struct list_head *tmp;
1118 CLASSERT(RES_NAME_SIZE == 4);
1120 if (!((libcfs_debug | D_ERROR) & level))
1123 CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1124 ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1125 res->lr_name.name[2], res->lr_name.name[3],
1126 atomic_read(&res->lr_refcount));
1128 if (!list_empty(&res->lr_granted)) {
1130 CDEBUG(level, "Granted locks:\n");
1131 list_for_each(tmp, &res->lr_granted) {
1132 struct ldlm_lock *lock;
1133 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1134 ldlm_lock_dump(level, lock, ++pos);
1137 if (!list_empty(&res->lr_converting)) {
1139 CDEBUG(level, "Converting locks:\n");
1140 list_for_each(tmp, &res->lr_converting) {
1141 struct ldlm_lock *lock;
1142 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1143 ldlm_lock_dump(level, lock, ++pos);
1146 if (!list_empty(&res->lr_waiting)) {
1148 CDEBUG(level, "Waiting locks:\n");
1149 list_for_each(tmp, &res->lr_waiting) {
1150 struct ldlm_lock *lock;
1151 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
1152 ldlm_lock_dump(level, lock, ++pos);