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 if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
146 if (count == 6 && memcmp(dummy, "clear", 5) == 0) {
148 "dropping all unused locks from namespace %s\n",
150 tmp = ns->ns_max_unused;
151 ns->ns_max_unused = 0;
152 ldlm_cancel_lru(ns, LDLM_SYNC);
153 ns->ns_max_unused = tmp;
157 tmp = simple_strtoul(dummy, &end, 0);
158 if (tmp == 0 && *end) {
159 CERROR("invalid value written\n");
163 CDEBUG(D_DLMTRACE, "changing namespace %s max_unused from %u to %u\n",
164 ns->ns_name, ns->ns_max_unused, (unsigned int)tmp);
166 ns->ns_max_unused = (unsigned int)tmp;
167 ldlm_cancel_lru(ns, LDLM_ASYNC);
171 void ldlm_proc_namespace(struct ldlm_namespace *ns)
173 struct lprocfs_vars lock_vars[2];
174 char lock_name[MAX_STRING_SIZE + 1];
177 LASSERT(ns->ns_name != NULL);
179 lock_name[MAX_STRING_SIZE] = '\0';
181 memset(lock_vars, 0, sizeof(lock_vars));
182 lock_vars[0].read_fptr = lprocfs_rd_u64;
183 lock_vars[0].name = lock_name;
185 snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
186 lock_vars[0].data = &ns->ns_resources;
187 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
189 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count", ns->ns_name);
190 lock_vars[0].data = &ns->ns_locks;
191 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
194 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
196 lock_vars[0].data = &ns->ns_nr_unused;
197 lock_vars[0].read_fptr = lprocfs_uint_rd;
198 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
200 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
202 lock_vars[0].data = ns;
203 lock_vars[0].read_fptr = lprocfs_read_lru_size;
204 lock_vars[0].write_fptr = lprocfs_write_lru_size;
205 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
209 #undef MAX_STRING_SIZE
211 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
213 struct ldlm_namespace *ns = NULL;
214 struct list_head *bucket;
220 CERROR("ldlm_get_ref failed: %d\n", rc);
224 OBD_ALLOC(ns, sizeof(*ns));
228 OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
232 OBD_ALLOC(ns->ns_name, strlen(name) + 1);
234 GOTO(out_hash, NULL);
236 strcpy(ns->ns_name, name);
238 INIT_LIST_HEAD(&ns->ns_root_list);
239 l_lock_init(&ns->ns_lock);
241 ns->ns_client = client;
242 spin_lock_init(&ns->ns_counter_lock);
244 ns->ns_resources = 0;
245 init_waitqueue_head(&ns->ns_waitq);
247 for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
249 INIT_LIST_HEAD(bucket);
251 INIT_LIST_HEAD(&ns->ns_unused_list);
252 ns->ns_nr_unused = 0;
253 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
255 down(&ldlm_namespace_lock);
256 list_add(&ns->ns_list_chain, &ldlm_namespace_list);
257 up(&ldlm_namespace_lock);
259 ldlm_proc_namespace(ns);
264 POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
265 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
267 OBD_FREE(ns, sizeof(*ns));
273 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
275 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
276 * This is currently only used for recovery, and we make certain assumptions
277 * as a result--notably, that we shouldn't cancel locks with refs. -phil
279 * Called with the ns_lock held. */
280 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
283 struct list_head *tmp, *pos;
284 int rc = 0, client = res->lr_namespace->ns_client;
285 int local_only = (flags & LDLM_FL_LOCAL_ONLY);
288 list_for_each_safe(tmp, pos, q) {
289 struct ldlm_lock *lock;
290 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
293 /* Set CBPENDING so nothing in the cancellation path
294 * can match this lock */
295 lock->l_flags |= LDLM_FL_CBPENDING;
296 lock->l_flags |= LDLM_FL_FAILED;
297 lock->l_flags |= flags;
299 if (local_only && (lock->l_readers || lock->l_writers)) {
300 /* This is a little bit gross, but much better than the
301 * alternative: pretend that we got a blocking AST from
302 * the server, so that when the lock is decref'd, it
303 * will go away ... */
304 /* ... without sending a CANCEL message. */
305 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
306 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
307 if (lock->l_completion_ast)
308 lock->l_completion_ast(lock, 0, NULL);
314 struct lustre_handle lockh;
315 ldlm_lock2handle(lock, &lockh);
317 rc = ldlm_cli_cancel(&lockh);
319 CERROR("ldlm_cli_cancel: %d\n", rc);
321 /* Force local cleanup on errors, too. */
322 if (local_only || rc != ELDLM_OK)
323 ldlm_lock_cancel(lock);
325 LDLM_DEBUG(lock, "Freeing a lock still held by a "
328 ldlm_resource_unlink_lock(lock);
329 ldlm_lock_destroy(lock);
336 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
341 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
345 l_lock(&ns->ns_lock);
346 for (i = 0; i < RES_HASH_SIZE; i++) {
347 struct list_head *tmp, *pos;
348 list_for_each_safe(tmp, pos, &(ns->ns_hash[i])) {
349 struct ldlm_resource *res;
350 res = list_entry(tmp, struct ldlm_resource, lr_hash);
351 ldlm_resource_getref(res);
353 cleanup_resource(res, &res->lr_granted, flags);
354 cleanup_resource(res, &res->lr_converting, flags);
355 cleanup_resource(res, &res->lr_waiting, flags);
357 /* XXX what a mess: don't force cleanup if we're
358 * local_only (which is only used by recovery). In that
359 * case, we probably still have outstanding lock refs
360 * which reference these resources. -phil */
361 if (!ldlm_resource_putref(res) &&
362 !(flags & LDLM_FL_LOCAL_ONLY)) {
363 CERROR("Resource refcount nonzero (%d) after "
364 "lock cleanup; forcing cleanup.\n",
365 atomic_read(&res->lr_refcount));
366 ldlm_resource_dump(D_ERROR, res);
367 atomic_set(&res->lr_refcount, 1);
368 ldlm_resource_putref(res);
372 l_unlock(&ns->ns_lock);
377 /* Cleanup, but also free, the namespace */
378 int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
383 down(&ldlm_namespace_lock);
384 list_del(&ns->ns_list_chain);
385 up(&ldlm_namespace_lock);
387 /* At shutdown time, don't call the cancellation callback */
388 ldlm_namespace_cleanup(ns, 0);
392 struct proc_dir_entry *dir;
393 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
395 CERROR("dlm namespace %s has no procfs dir?\n",
403 POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
404 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
405 OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
406 OBD_FREE(ns, sizeof(*ns));
413 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, struct ldlm_res_id name)
418 for (i = 0; i < RES_NAME_SIZE; i++)
419 hash += name.name[i];
421 hash += (__u32)((unsigned long)parent >> 4);
423 return (hash & RES_HASH_MASK);
426 static struct ldlm_resource *ldlm_resource_new(void)
428 struct ldlm_resource *res;
430 OBD_SLAB_ALLOC(res, ldlm_resource_slab, SLAB_NOFS, sizeof *res);
434 memset(res, 0, sizeof(*res));
436 INIT_LIST_HEAD(&res->lr_children);
437 INIT_LIST_HEAD(&res->lr_childof);
438 INIT_LIST_HEAD(&res->lr_granted);
439 INIT_LIST_HEAD(&res->lr_converting);
440 INIT_LIST_HEAD(&res->lr_waiting);
441 sema_init(&res->lr_lvb_sem, 1);
442 atomic_set(&res->lr_refcount, 1);
447 /* Args: locked namespace
448 * Returns: newly-allocated, referenced, unlocked resource */
449 static struct ldlm_resource *
450 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
451 struct ldlm_res_id name, __u32 type)
453 struct list_head *bucket;
454 struct ldlm_resource *res;
457 LASSERTF(type >= LDLM_MIN_TYPE && type <= LDLM_MAX_TYPE,
460 res = ldlm_resource_new();
464 spin_lock(&ns->ns_counter_lock);
466 spin_unlock(&ns->ns_counter_lock);
468 l_lock(&ns->ns_lock);
469 memcpy(&res->lr_name, &name, sizeof(res->lr_name));
470 res->lr_namespace = ns;
474 res->lr_most_restr = LCK_NL;
476 bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
477 list_add(&res->lr_hash, bucket);
479 if (parent == NULL) {
480 list_add(&res->lr_childof, &ns->ns_root_list);
482 res->lr_parent = parent;
483 list_add(&res->lr_childof, &parent->lr_children);
485 l_unlock(&ns->ns_lock);
491 /* Args: unlocked namespace
492 * Locks: takes and releases ns->ns_lock and res->lr_lock
493 * Returns: referenced, unlocked ldlm_resource or NULL */
494 struct ldlm_resource *
495 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
496 struct ldlm_res_id name, __u32 type, int create)
498 struct ldlm_resource *res = NULL;
499 struct list_head *bucket, *tmp;
503 LASSERT(ns->ns_hash != NULL);
504 LASSERT(name.name[0] != 0);
506 l_lock(&ns->ns_lock);
507 bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
509 list_for_each(tmp, bucket) {
510 res = list_entry(tmp, struct ldlm_resource, lr_hash);
512 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0) {
513 ldlm_resource_getref(res);
514 l_unlock(&ns->ns_lock);
520 res = ldlm_resource_add(ns, parent, name, type);
527 if (create && ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
530 /* Although this is technically a lock inversion risk (lvb_sem
531 * should be taken before DLM lock), this resource was just
532 * created, so nobody else can take the lvb_sem yet. -p */
533 down(&res->lr_lvb_sem);
534 /* Drop the dlm lock, because lvbo_init can touch the disk */
535 l_unlock(&ns->ns_lock);
536 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
537 rc = ns->ns_lvbo->lvbo_init(res);
538 up(&res->lr_lvb_sem);
540 CERROR("lvbo_init failed for resource "
541 LPU64": rc %d\n", name.name[0], rc);
544 l_unlock(&ns->ns_lock);
550 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
552 LASSERT(res != NULL);
553 LASSERT(res != LP_POISON);
554 atomic_inc(&res->lr_refcount);
555 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
556 atomic_read(&res->lr_refcount));
560 /* Returns 1 if the resource was freed, 0 if it remains. */
561 int ldlm_resource_putref(struct ldlm_resource *res)
566 CDEBUG(D_INFO, "putref res: %p count: %d\n", res,
567 atomic_read(&res->lr_refcount) - 1);
568 LASSERT(atomic_read(&res->lr_refcount) > 0);
569 LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
571 if (atomic_dec_and_test(&res->lr_refcount)) {
572 struct ldlm_namespace *ns = res->lr_namespace;
575 l_lock(&ns->ns_lock);
577 if (atomic_read(&res->lr_refcount) != 0) {
578 /* We lost the race. */
579 l_unlock(&ns->ns_lock);
583 if (!list_empty(&res->lr_granted)) {
584 ldlm_resource_dump(D_ERROR, res);
588 if (!list_empty(&res->lr_converting)) {
589 ldlm_resource_dump(D_ERROR, res);
593 if (!list_empty(&res->lr_waiting)) {
594 ldlm_resource_dump(D_ERROR, res);
598 if (!list_empty(&res->lr_children)) {
599 ldlm_resource_dump(D_ERROR, res);
604 list_del_init(&res->lr_hash);
605 list_del_init(&res->lr_childof);
606 if (res->lr_lvb_data)
607 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
608 l_unlock(&ns->ns_lock);
610 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
612 spin_lock(&ns->ns_counter_lock);
614 if (ns->ns_resources == 0)
615 wake_up(&ns->ns_waitq);
616 spin_unlock(&ns->ns_counter_lock);
625 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
626 struct ldlm_lock *lock)
628 l_lock(&res->lr_namespace->ns_lock);
630 ldlm_resource_dump(D_OTHER, res);
631 CDEBUG(D_OTHER, "About to add this lock:\n");
632 ldlm_lock_dump(D_OTHER, lock, 0);
634 if (lock->l_destroyed) {
635 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
639 LASSERT(list_empty(&lock->l_res_link));
641 list_add_tail(&lock->l_res_link, head);
643 l_unlock(&res->lr_namespace->ns_lock);
646 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
647 struct ldlm_lock *new)
649 struct ldlm_resource *res = original->l_resource;
651 l_lock(&res->lr_namespace->ns_lock);
653 ldlm_resource_dump(D_OTHER, res);
654 CDEBUG(D_OTHER, "About to insert this lock after %p:\n", original);
655 ldlm_lock_dump(D_OTHER, new, 0);
657 if (new->l_destroyed) {
658 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
662 LASSERT(list_empty(&new->l_res_link));
664 list_add(&new->l_res_link, &original->l_res_link);
666 l_unlock(&res->lr_namespace->ns_lock);
669 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
671 l_lock(&lock->l_resource->lr_namespace->ns_lock);
672 list_del_init(&lock->l_res_link);
673 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
676 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
678 desc->lr_type = res->lr_type;
679 memcpy(&desc->lr_name, &res->lr_name, sizeof(desc->lr_name));
682 void ldlm_dump_all_namespaces(int level)
684 struct list_head *tmp;
686 down(&ldlm_namespace_lock);
688 list_for_each(tmp, &ldlm_namespace_list) {
689 struct ldlm_namespace *ns;
690 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
691 ldlm_namespace_dump(level, ns);
694 up(&ldlm_namespace_lock);
697 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
699 struct list_head *tmp;
701 CDEBUG(level, "--- Namespace: %s (rc: %d, client: %d)\n",
702 ns->ns_name, ns->ns_refcount, ns->ns_client);
704 l_lock(&ns->ns_lock);
705 if (time_after(jiffies, ns->ns_next_dump)) {
706 list_for_each(tmp, &ns->ns_root_list) {
707 struct ldlm_resource *res;
708 res = list_entry(tmp, struct ldlm_resource, lr_childof);
710 /* Once we have resources with children, this should
711 * really dump them recursively. */
712 ldlm_resource_dump(level, res);
714 ns->ns_next_dump = jiffies + 10 * HZ;
716 l_unlock(&ns->ns_lock);
719 void ldlm_resource_dump(int level, struct ldlm_resource *res)
721 struct list_head *tmp;
724 if (RES_NAME_SIZE != 4)
727 CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
728 ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
729 res->lr_name.name[2], res->lr_name.name[3],
730 atomic_read(&res->lr_refcount));
732 if (!list_empty(&res->lr_granted)) {
734 CDEBUG(level, "Granted locks:\n");
735 list_for_each(tmp, &res->lr_granted) {
736 struct ldlm_lock *lock;
737 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
738 ldlm_lock_dump(level, lock, ++pos);
741 if (!list_empty(&res->lr_converting)) {
743 CDEBUG(level, "Converting locks:\n");
744 list_for_each(tmp, &res->lr_converting) {
745 struct ldlm_lock *lock;
746 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
747 ldlm_lock_dump(level, lock, ++pos);
750 if (!list_empty(&res->lr_waiting)) {
752 CDEBUG(level, "Waiting locks:\n");
753 list_for_each(tmp, &res->lr_waiting) {
754 struct ldlm_lock *lock;
755 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
756 ldlm_lock_dump(level, lock, ++pos);