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 Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 #define DEBUG_SUBSYSTEM S_LDLM
26 # include <linux/lustre_dlm.h>
28 # include <liblustre.h>
31 #include <linux/obd_class.h>
32 #include "ldlm_internal.h"
34 kmem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
36 DECLARE_MUTEX(ldlm_namespace_lock);
37 struct list_head ldlm_namespace_list = LIST_HEAD_INIT(ldlm_namespace_list);
38 struct proc_dir_entry *ldlm_type_proc_dir = NULL;
39 struct proc_dir_entry *ldlm_ns_proc_dir = NULL;
40 struct proc_dir_entry *ldlm_svc_proc_dir = NULL;
43 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
44 unsigned long count, void *data)
46 ldlm_dump_all_namespaces(D_DLMTRACE);
50 int ldlm_proc_setup(void)
53 struct lprocfs_vars list[] = {
54 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
57 LASSERT(ldlm_ns_proc_dir == NULL);
59 ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
62 if (IS_ERR(ldlm_type_proc_dir)) {
63 CERROR("LProcFS failed in ldlm-init\n");
64 rc = PTR_ERR(ldlm_type_proc_dir);
68 ldlm_ns_proc_dir = lprocfs_register("namespaces",
71 if (IS_ERR(ldlm_ns_proc_dir)) {
72 CERROR("LProcFS failed in ldlm-init\n");
73 rc = PTR_ERR(ldlm_ns_proc_dir);
77 ldlm_svc_proc_dir = lprocfs_register("services",
80 if (IS_ERR(ldlm_svc_proc_dir)) {
81 CERROR("LProcFS failed in ldlm-init\n");
82 rc = PTR_ERR(ldlm_svc_proc_dir);
86 rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
91 lprocfs_remove(ldlm_ns_proc_dir);
93 lprocfs_remove(ldlm_type_proc_dir);
95 ldlm_type_proc_dir = NULL;
96 ldlm_ns_proc_dir = NULL;
97 ldlm_svc_proc_dir = NULL;
101 void ldlm_proc_cleanup(void)
103 if (ldlm_svc_proc_dir) {
104 lprocfs_remove(ldlm_svc_proc_dir);
105 ldlm_svc_proc_dir = NULL;
108 if (ldlm_ns_proc_dir) {
109 lprocfs_remove(ldlm_ns_proc_dir);
110 ldlm_ns_proc_dir = NULL;
113 if (ldlm_type_proc_dir) {
114 lprocfs_remove(ldlm_type_proc_dir);
115 ldlm_type_proc_dir = NULL;
119 static int lprocfs_uint_rd(char *page, char **start, off_t off,
120 int count, int *eof, void *data)
122 unsigned int *temp = (unsigned int *)data;
123 return snprintf(page, count, "%u\n", *temp);
126 static int lprocfs_read_lru_size(char *page, char **start, off_t off,
127 int count, int *eof, void *data)
129 struct ldlm_namespace *ns = data;
130 return lprocfs_uint_rd(page, start, off, count, eof,
134 #define MAX_STRING_SIZE 128
135 static int lprocfs_write_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;
142 dummy[MAX_STRING_SIZE] = '\0';
143 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 tmp = ns->ns_max_unused;
150 ns->ns_max_unused = 0;
151 ldlm_cancel_lru(ns, LDLM_SYNC);
152 ns->ns_max_unused = tmp;
156 tmp = simple_strtoul(dummy, &end, 0);
157 if (tmp == 0 && *end) {
158 CERROR("invalid value written\n");
162 CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
163 ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
165 ns->ns_max_unused = (unsigned int)tmp;
166 ldlm_cancel_lru(ns, LDLM_ASYNC);
170 void ldlm_proc_namespace(struct ldlm_namespace *ns)
172 struct lprocfs_vars lock_vars[2];
173 char lock_name[MAX_STRING_SIZE + 1];
176 LASSERT(ns->ns_name != NULL);
178 lock_name[MAX_STRING_SIZE] = '\0';
180 memset(lock_vars, 0, sizeof(lock_vars));
181 lock_vars[0].read_fptr = lprocfs_rd_u64;
182 lock_vars[0].name = lock_name;
184 snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
185 lock_vars[0].data = &ns->ns_resources;
186 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
188 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
189 lock_vars[0].data = &ns->ns_locks;
190 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
193 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
195 lock_vars[0].data = &ns->ns_nr_unused;
196 lock_vars[0].read_fptr = lprocfs_uint_rd;
197 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
199 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
201 lock_vars[0].data = ns;
202 lock_vars[0].read_fptr = lprocfs_read_lru_size;
203 lock_vars[0].write_fptr = lprocfs_write_lru_size;
204 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
208 #undef MAX_STRING_SIZE
210 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
212 struct ldlm_namespace *ns = NULL;
213 struct list_head *bucket;
219 CERROR("ldlm_get_ref failed: %d\n", rc);
223 OBD_ALLOC(ns, sizeof(*ns));
227 OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
231 OBD_ALLOC(ns->ns_name, strlen(name) + 1);
233 GOTO(out_hash, NULL);
235 strcpy(ns->ns_name, name);
237 INIT_LIST_HEAD(&ns->ns_root_list);
238 l_lock_init(&ns->ns_lock);
240 ns->ns_client = client;
241 spin_lock_init(&ns->ns_counter_lock);
243 ns->ns_resources = 0;
244 init_waitqueue_head(&ns->ns_waitq);
246 for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
248 INIT_LIST_HEAD(bucket);
250 INIT_LIST_HEAD(&ns->ns_unused_list);
251 ns->ns_nr_unused = 0;
252 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
254 down(&ldlm_namespace_lock);
255 list_add(&ns->ns_list_chain, &ldlm_namespace_list);
256 up(&ldlm_namespace_lock);
258 ldlm_proc_namespace(ns);
263 POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
264 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
266 OBD_FREE(ns, sizeof(*ns));
272 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
274 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
275 * This is currently only used for recovery, and we make certain assumptions
276 * as a result--notably, that we shouldn't cancel locks with refs. -phil
278 * Called with the ns_lock held. */
279 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
282 struct list_head *tmp, *pos;
283 int rc = 0, client = res->lr_namespace->ns_client;
284 int local_only = (flags & LDLM_FL_LOCAL_ONLY);
287 list_for_each_safe(tmp, pos, q) {
288 struct ldlm_lock *lock;
289 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
292 /* Set CBPENDING so nothing in the cancellation path
293 * can match this lock */
294 lock->l_flags |= LDLM_FL_CBPENDING;
295 lock->l_flags |= LDLM_FL_FAILED;
296 lock->l_flags |= flags;
298 if (local_only && (lock->l_readers || lock->l_writers)) {
299 /* This is a little bit gross, but much better than the
300 * alternative: pretend that we got a blocking AST from
301 * the server, so that when the lock is decref'd, it
302 * will go away ... */
303 /* ... without sending a CANCEL message. */
304 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
305 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
306 if (lock->l_completion_ast)
307 lock->l_completion_ast(lock, 0, NULL);
313 struct lustre_handle lockh;
314 ldlm_lock2handle(lock, &lockh);
316 rc = ldlm_cli_cancel(&lockh);
318 CERROR("ldlm_cli_cancel: %d\n", rc);
320 /* Force local cleanup on errors, too. */
321 if (local_only || rc != ELDLM_OK)
322 ldlm_lock_cancel(lock);
324 LDLM_DEBUG(lock, "Freeing a lock still held by a "
327 ldlm_resource_unlink_lock(lock);
328 ldlm_lock_destroy(lock);
335 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
340 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
344 l_lock(&ns->ns_lock);
345 for (i = 0; i < RES_HASH_SIZE; i++) {
346 struct list_head *tmp, *pos;
347 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
348 struct ldlm_resource *res;
349 res = list_entry(tmp, struct ldlm_resource, lr_hash);
350 ldlm_resource_getref(res);
352 cleanup_resource(res, &res->lr_granted, flags);
353 cleanup_resource(res, &res->lr_converting, flags);
354 cleanup_resource(res, &res->lr_waiting, flags);
356 /* XXX what a mess: don't force cleanup if we're
357 * local_only (which is only used by recovery). In that
358 * case, we probably still have outstanding lock refs
359 * which reference these resources. -phil */
360 if (!ldlm_resource_putref(res) &&
361 !(flags & LDLM_FL_LOCAL_ONLY)) {
362 CERROR("Resource refcount nonzero (%d) after "
363 "lock cleanup; forcing cleanup.\n",
364 atomic_read(&res->lr_refcount));
365 ldlm_resource_dump(D_ERROR, res);
366 atomic_set(&res->lr_refcount, 1);
367 ldlm_resource_putref(res);
371 l_unlock(&ns->ns_lock);
376 /* Cleanup, but also free, the namespace */
377 int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
382 down(&ldlm_namespace_lock);
383 list_del(&ns->ns_list_chain);
384 up(&ldlm_namespace_lock);
386 /* At shutdown time, don't call the cancellation callback */
387 ldlm_namespace_cleanup(ns, 0);
391 struct proc_dir_entry *dir;
392 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
394 CERROR("dlm namespace %s has no procfs dir?\n",
402 POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
403 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
404 OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
405 OBD_FREE(ns, sizeof(*ns));
412 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, struct ldlm_res_id name)
417 for (i = 0; i < RES_NAME_SIZE; i++)
418 hash += name.name[i];
420 hash += (__u32)((unsigned long)parent >> 4);
422 return (hash & RES_HASH_MASK);
425 static struct ldlm_resource *ldlm_resource_new(void)
427 struct ldlm_resource *res;
429 OBD_SLAB_ALLOC(res, ldlm_resource_slab, SLAB_NOFS, sizeof *res);
433 memset(res, 0, sizeof(*res));
435 INIT_LIST_HEAD(&res->lr_children);
436 INIT_LIST_HEAD(&res->lr_childof);
437 INIT_LIST_HEAD(&res->lr_granted);
438 INIT_LIST_HEAD(&res->lr_converting);
439 INIT_LIST_HEAD(&res->lr_waiting);
440 sema_init(&res->lr_lvb_sem, 1);
441 atomic_set(&res->lr_refcount, 1);
446 /* Args: locked namespace
447 * Returns: newly-allocated, referenced, unlocked resource */
448 static struct ldlm_resource *
449 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
450 struct ldlm_res_id name, __u32 type)
452 struct list_head *bucket;
453 struct ldlm_resource *res;
456 LASSERTF(type >= LDLM_MIN_TYPE && type <= LDLM_MAX_TYPE,
459 res = ldlm_resource_new();
463 spin_lock(&ns->ns_counter_lock);
465 spin_unlock(&ns->ns_counter_lock);
467 l_lock(&ns->ns_lock);
468 memcpy(&res->lr_name, &name, sizeof(res->lr_name));
469 res->lr_namespace = ns;
473 res->lr_most_restr = LCK_NL;
475 bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
476 list_add(&res->lr_hash, bucket);
478 if (parent == NULL) {
479 list_add(&res->lr_childof, &ns->ns_root_list);
481 res->lr_parent = parent;
482 list_add(&res->lr_childof, &parent->lr_children);
484 l_unlock(&ns->ns_lock);
490 /* Args: unlocked namespace
491 * Locks: takes and releases ns->ns_lock and res->lr_lock
492 * Returns: referenced, unlocked ldlm_resource or NULL */
493 struct ldlm_resource *
494 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
495 struct ldlm_res_id name, __u32 type, int create)
497 struct list_head *bucket, *tmp;
498 struct ldlm_resource *res = NULL;
502 LASSERT(ns->ns_hash != NULL);
503 LASSERT(name.name[0] != 0);
505 l_lock(&ns->ns_lock);
506 bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
508 list_for_each(tmp, bucket) {
509 res = list_entry(tmp, struct ldlm_resource, lr_hash);
511 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0) {
512 ldlm_resource_getref(res);
513 l_unlock(&ns->ns_lock);
519 res = ldlm_resource_add(ns, parent, name, type);
526 if (create && ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
529 /* Although this is technically a lock inversion risk (lvb_sem
530 * should be taken before DLM lock), this resource was just
531 * created, so nobody else can take the lvb_sem yet. -p */
532 down(&res->lr_lvb_sem);
533 /* Drop the dlm lock, because lvbo_init can touch the disk */
534 l_unlock(&ns->ns_lock);
535 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
536 rc = ns->ns_lvbo->lvbo_init(res);
537 up(&res->lr_lvb_sem);
539 CERROR("lvbo_init failed for resource "LPU64": rc %d\n",
543 l_unlock(&ns->ns_lock);
549 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
551 LASSERT(res != NULL);
552 LASSERT(res != LP_POISON);
553 atomic_inc(&res->lr_refcount);
554 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
555 atomic_read(&res->lr_refcount));
559 /* Returns 1 if the resource was freed, 0 if it remains. */
560 int ldlm_resource_putref(struct ldlm_resource *res)
565 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
566 atomic_read(&res->lr_refcount) - 1);
567 LASSERT(atomic_read(&res->lr_refcount) > 0);
568 LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
570 if (atomic_dec_and_test(&res->lr_refcount)) {
571 struct ldlm_namespace *ns = res->lr_namespace;
574 l_lock(&ns->ns_lock);
576 if (atomic_read(&res->lr_refcount) != 0) {
577 /* We lost the race. */
578 l_unlock(&ns->ns_lock);
582 if (!list_empty(&res->lr_granted)) {
583 ldlm_resource_dump(D_ERROR, res);
587 if (!list_empty(&res->lr_converting)) {
588 ldlm_resource_dump(D_ERROR, res);
592 if (!list_empty(&res->lr_waiting)) {
593 ldlm_resource_dump(D_ERROR, res);
597 if (!list_empty(&res->lr_children)) {
598 ldlm_resource_dump(D_ERROR, res);
603 list_del_init(&res->lr_hash);
604 list_del_init(&res->lr_childof);
605 if (res->lr_lvb_data)
606 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
607 l_unlock(&ns->ns_lock);
609 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
611 spin_lock(&ns->ns_counter_lock);
613 if (ns->ns_resources == 0)
614 wake_up(&ns->ns_waitq);
615 spin_unlock(&ns->ns_counter_lock);
624 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
625 struct ldlm_lock *lock)
627 l_lock(&res->lr_namespace->ns_lock);
629 ldlm_resource_dump(D_OTHER, res);
630 CDEBUG(D_OTHER, "About to add this lock:\n");
631 ldlm_lock_dump(D_OTHER, lock, 0);
633 if (lock->l_destroyed) {
634 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
638 LASSERT(list_empty(&lock->l_res_link));
640 list_add_tail(&lock->l_res_link, head);
642 l_unlock(&res->lr_namespace->ns_lock);
645 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
646 struct ldlm_lock *new)
648 struct ldlm_resource *res = original->l_resource;
650 l_lock(&res->lr_namespace->ns_lock);
652 ldlm_resource_dump(D_OTHER, res);
653 CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
654 ldlm_lock_dump(D_OTHER, new, 0);
656 if (new->l_destroyed) {
657 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
661 LASSERT(list_empty(&new->l_res_link));
663 list_add(&new->l_res_link, &original->l_res_link);
665 l_unlock(&res->lr_namespace->ns_lock);
668 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
670 l_lock(&lock->l_resource->lr_namespace->ns_lock);
671 list_del_init(&lock->l_res_link);
672 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
675 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
677 desc->lr_type = res->lr_type;
678 memcpy(&desc->lr_name, &res->lr_name, sizeof(desc->lr_name));
681 void ldlm_dump_all_namespaces(int level)
683 struct list_head *tmp;
685 down(&ldlm_namespace_lock);
687 list_for_each(tmp, &ldlm_namespace_list) {
688 struct ldlm_namespace *ns;
689 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
690 ldlm_namespace_dump(level, ns);
693 up(&ldlm_namespace_lock);
696 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
698 struct list_head *tmp;
700 CDEBUG(level, "--- Namespace: %s (rc: %d, client: %d)\n",
701 ns->ns_name, ns->ns_refcount, ns->ns_client);
703 l_lock(&ns->ns_lock);
704 if (time_after(jiffies, ns->ns_next_dump)) {
705 list_for_each(tmp, &ns->ns_root_list) {
706 struct ldlm_resource *res;
707 res = list_entry(tmp, struct ldlm_resource, lr_childof);
709 /* Once we have resources with children, this should
710 * really dump them recursively. */
711 ldlm_resource_dump(level, res);
713 ns->ns_next_dump = jiffies + 10 * HZ;
715 l_unlock(&ns->ns_lock);
718 void ldlm_resource_dump(int level, struct ldlm_resource *res)
720 struct list_head *tmp;
723 if (RES_NAME_SIZE != 4)
726 CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
727 ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
728 res->lr_name.name[2], res->lr_name.name[3],
729 atomic_read(&res->lr_refcount));
731 if (!list_empty(&res->lr_granted)) {
733 CDEBUG(level, "Granted locks:\n");
734 list_for_each(tmp, &res->lr_granted) {
735 struct ldlm_lock *lock;
736 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
737 ldlm_lock_dump(level, lock, ++pos);
740 if (!list_empty(&res->lr_converting)) {
742 CDEBUG(level, "Converting locks:\n");
743 list_for_each(tmp, &res->lr_converting) {
744 struct ldlm_lock *lock;
745 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
746 ldlm_lock_dump(level, lock, ++pos);
749 if (!list_empty(&res->lr_waiting)) {
751 CDEBUG(level, "Waiting locks:\n");
752 list_for_each(tmp, &res->lr_waiting) {
753 struct ldlm_lock *lock;
754 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
755 ldlm_lock_dump(level, lock, ++pos);