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);
164 ns->ns_max_unused = (unsigned int)tmp;
166 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].name = lock_name;
184 snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count", ns->ns_name);
185 lock_vars[0].data = &ns->ns_refcount;
186 lock_vars[0].read_fptr = lprocfs_rd_atomic;
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 lock_vars[0].read_fptr = lprocfs_rd_u64;
192 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
195 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
197 lock_vars[0].data = &ns->ns_nr_unused;
198 lock_vars[0].read_fptr = lprocfs_uint_rd;
199 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
201 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
203 lock_vars[0].data = ns;
204 lock_vars[0].read_fptr = lprocfs_read_lru_size;
205 lock_vars[0].write_fptr = lprocfs_write_lru_size;
206 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
209 #undef MAX_STRING_SIZE
211 #define ldlm_proc_namespace(ns) do {} while (0)
214 struct ldlm_namespace *ldlm_namespace_new(char *name, __u32 client)
216 struct ldlm_namespace *ns = NULL;
217 struct list_head *bucket;
223 CERROR("ldlm_get_ref failed: %d\n", rc);
227 OBD_ALLOC(ns, sizeof(*ns));
231 OBD_VMALLOC(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
235 OBD_ALLOC(ns->ns_name, strlen(name) + 1);
237 GOTO(out_hash, NULL);
239 strcpy(ns->ns_name, name);
241 INIT_LIST_HEAD(&ns->ns_root_list);
242 l_lock_init(&ns->ns_lock);
243 init_waitqueue_head(&ns->ns_refcount_waitq);
244 atomic_set(&ns->ns_refcount, 0);
245 ns->ns_client = client;
246 spin_lock_init(&ns->ns_counter_lock);
249 for (bucket = ns->ns_hash + RES_HASH_SIZE - 1; bucket >= ns->ns_hash;
251 INIT_LIST_HEAD(bucket);
253 INIT_LIST_HEAD(&ns->ns_unused_list);
254 ns->ns_nr_unused = 0;
255 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
257 down(&ldlm_namespace_lock);
258 list_add(&ns->ns_list_chain, &ldlm_namespace_list);
259 up(&ldlm_namespace_lock);
260 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 if (!ldlm_resource_putref(res)) {
358 CERROR("Namespace %s resource refcount %d "
359 "after lock cleanup\n",
361 atomic_read(&res->lr_refcount));
365 l_unlock(&ns->ns_lock);
370 /* Cleanup, but also free, the namespace */
371 int ldlm_namespace_free(struct ldlm_namespace *ns, int force)
376 down(&ldlm_namespace_lock);
377 list_del(&ns->ns_list_chain);
378 up(&ldlm_namespace_lock);
380 /* At shutdown time, don't call the cancellation callback */
381 ldlm_namespace_cleanup(ns, 0);
385 struct proc_dir_entry *dir;
386 dir = lprocfs_srch(ldlm_ns_proc_dir, ns->ns_name);
388 CERROR("dlm namespace %s has no procfs dir?\n",
396 if (atomic_read(&ns->ns_refcount) > 0) {
397 struct l_wait_info lwi = LWI_INTR(NULL, NULL);
400 "dlm namespace %s free waiting on refcount %d\n",
401 ns->ns_name, atomic_read(&ns->ns_refcount));
402 rc = l_wait_event(ns->ns_refcount_waitq,
403 atomic_read(&ns->ns_refcount) == 0, &lwi);
404 if (atomic_read(&ns->ns_refcount)) {
405 LCONSOLE_ERROR("Lock manager: wait for %s namespace "
406 "cleanup aborted with %d resources in "
407 "use. (%d)\nI'm going to try to clean "
408 "up anyway, but I might need a reboot "
409 "of this node.\n", ns->ns_name,
410 atomic_read(&ns->ns_refcount), rc);
413 "dlm namespace %s free done waiting\n", ns->ns_name);
416 POISON(ns->ns_hash, 0x5a, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
417 OBD_VFREE(ns->ns_hash, sizeof(*ns->ns_hash) * RES_HASH_SIZE);
418 OBD_FREE(ns->ns_name, strlen(ns->ns_name) + 1);
419 OBD_FREE(ns, sizeof(*ns));
426 static __u32 ldlm_hash_fn(struct ldlm_resource *parent, struct ldlm_res_id name)
431 for (i = 0; i < RES_NAME_SIZE; i++)
432 hash += name.name[i];
434 hash += (__u32)((unsigned long)parent >> 4);
436 return (hash & RES_HASH_MASK);
439 static struct ldlm_resource *ldlm_resource_new(void)
441 struct ldlm_resource *res;
443 OBD_SLAB_ALLOC(res, ldlm_resource_slab, SLAB_NOFS, sizeof *res);
447 memset(res, 0, sizeof(*res));
449 INIT_LIST_HEAD(&res->lr_children);
450 INIT_LIST_HEAD(&res->lr_childof);
451 INIT_LIST_HEAD(&res->lr_granted);
452 INIT_LIST_HEAD(&res->lr_converting);
453 INIT_LIST_HEAD(&res->lr_waiting);
454 sema_init(&res->lr_lvb_sem, 1);
455 atomic_set(&res->lr_refcount, 1);
460 /* Args: locked namespace
461 * Returns: newly-allocated, referenced, unlocked resource */
462 static struct ldlm_resource *
463 ldlm_resource_add(struct ldlm_namespace *ns, struct ldlm_resource *parent,
464 struct ldlm_res_id name, ldlm_type_t type)
466 struct list_head *bucket;
467 struct ldlm_resource *res;
470 LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
473 res = ldlm_resource_new();
477 l_lock(&ns->ns_lock);
478 memcpy(&res->lr_name, &name, sizeof(res->lr_name));
479 res->lr_namespace = ns;
480 atomic_inc(&ns->ns_refcount);
483 res->lr_most_restr = LCK_NL;
485 bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
486 list_add(&res->lr_hash, bucket);
488 if (parent == NULL) {
489 list_add(&res->lr_childof, &ns->ns_root_list);
491 res->lr_parent = parent;
492 list_add(&res->lr_childof, &parent->lr_children);
494 l_unlock(&ns->ns_lock);
500 /* Args: unlocked namespace
501 * Locks: takes and releases ns->ns_lock and res->lr_lock
502 * Returns: referenced, unlocked ldlm_resource or NULL */
503 struct ldlm_resource *
504 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
505 struct ldlm_res_id name, ldlm_type_t type, int create)
507 struct list_head *bucket, *tmp;
508 struct ldlm_resource *res = NULL;
512 LASSERT(ns->ns_hash != NULL);
513 LASSERT(name.name[0] != 0);
515 l_lock(&ns->ns_lock);
516 bucket = ns->ns_hash + ldlm_hash_fn(parent, name);
518 list_for_each(tmp, bucket) {
519 res = list_entry(tmp, struct ldlm_resource, lr_hash);
521 if (memcmp(&res->lr_name, &name, sizeof(res->lr_name)) == 0) {
522 ldlm_resource_getref(res);
523 l_unlock(&ns->ns_lock);
529 res = ldlm_resource_add(ns, parent, name, type);
536 if (create && ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
539 /* Although this is technically a lock inversion risk (lvb_sem
540 * should be taken before DLM lock), this resource was just
541 * created, so nobody else can take the lvb_sem yet. -p */
542 down(&res->lr_lvb_sem);
543 /* Drop the dlm lock, because lvbo_init can touch the disk */
544 l_unlock(&ns->ns_lock);
545 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
546 rc = ns->ns_lvbo->lvbo_init(res);
547 up(&res->lr_lvb_sem);
549 CERROR("lvbo_init failed for resource "LPU64"/"LPU64
550 ": rc %d\n", name.name[0], name.name[1], rc);
553 l_unlock(&ns->ns_lock);
559 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
561 LASSERT(res != NULL);
562 LASSERT(res != LP_POISON);
563 atomic_inc(&res->lr_refcount);
564 CDEBUG(D_DLMTRACE, "getref res: %p count: %d\n", res,
565 atomic_read(&res->lr_refcount));
569 /* Returns 1 if the resource was freed, 0 if it remains. */
570 int ldlm_resource_putref(struct ldlm_resource *res)
575 CDEBUG(D_DLMTRACE, "putref res: %p count: %d\n", res,
576 atomic_read(&res->lr_refcount) - 1);
577 LASSERT(atomic_read(&res->lr_refcount) > 0);
578 LASSERT(atomic_read(&res->lr_refcount) < LI_POISON);
580 if (atomic_dec_and_test(&res->lr_refcount)) {
581 struct ldlm_namespace *ns = res->lr_namespace;
584 l_lock(&ns->ns_lock);
586 if (atomic_read(&res->lr_refcount) != 0) {
587 /* We lost the race. */
588 l_unlock(&ns->ns_lock);
592 if (!list_empty(&res->lr_granted)) {
593 ldlm_resource_dump(D_ERROR, res);
597 if (!list_empty(&res->lr_converting)) {
598 ldlm_resource_dump(D_ERROR, res);
602 if (!list_empty(&res->lr_waiting)) {
603 ldlm_resource_dump(D_ERROR, res);
607 if (!list_empty(&res->lr_children)) {
608 ldlm_resource_dump(D_ERROR, res);
612 list_del_init(&res->lr_hash);
613 list_del_init(&res->lr_childof);
614 if (res->lr_lvb_data)
615 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
616 l_unlock(&ns->ns_lock);
618 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
620 if (atomic_dec_and_test(&ns->ns_refcount)) {
621 CDEBUG(D_DLMTRACE, "last ref on ns %s\n", ns->ns_name);
622 wake_up(&ns->ns_refcount_waitq);
632 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
633 struct ldlm_lock *lock)
635 l_lock(&res->lr_namespace->ns_lock);
637 ldlm_resource_dump(D_OTHER, res);
638 CDEBUG(D_OTHER, "About to add this lock:\n");
639 ldlm_lock_dump(D_OTHER, lock, 0);
641 if (lock->l_destroyed) {
642 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
646 LASSERT(list_empty(&lock->l_res_link));
648 list_add_tail(&lock->l_res_link, head);
650 l_unlock(&res->lr_namespace->ns_lock);
653 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
655 l_lock(&lock->l_resource->lr_namespace->ns_lock);
656 list_del_init(&lock->l_res_link);
657 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
660 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
662 desc->lr_type = res->lr_type;
663 memcpy(&desc->lr_name, &res->lr_name, sizeof(desc->lr_name));
666 void ldlm_dump_all_namespaces(int level)
668 struct list_head *tmp;
670 down(&ldlm_namespace_lock);
672 list_for_each(tmp, &ldlm_namespace_list) {
673 struct ldlm_namespace *ns;
674 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
675 ldlm_namespace_dump(level, ns);
678 up(&ldlm_namespace_lock);
681 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
683 struct list_head *tmp;
685 CDEBUG(level, "--- Namespace: %s (rc: %d, client: %d)\n", ns->ns_name,
686 atomic_read(&ns->ns_refcount), ns->ns_client);
688 l_lock(&ns->ns_lock);
689 if (time_after(jiffies, ns->ns_next_dump)) {
690 list_for_each(tmp, &ns->ns_root_list) {
691 struct ldlm_resource *res;
692 res = list_entry(tmp, struct ldlm_resource, lr_childof);
694 /* Once we have resources with children, this should
695 * really dump them recursively. */
696 ldlm_resource_dump(level, res);
698 ns->ns_next_dump = jiffies + 10 * HZ;
700 l_unlock(&ns->ns_lock);
703 void ldlm_resource_dump(int level, struct ldlm_resource *res)
705 struct list_head *tmp;
708 if (RES_NAME_SIZE != 4)
711 CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
712 ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
713 res->lr_name.name[2], res->lr_name.name[3],
714 atomic_read(&res->lr_refcount));
716 if (!list_empty(&res->lr_granted)) {
718 CDEBUG(level, "Granted locks:\n");
719 list_for_each(tmp, &res->lr_granted) {
720 struct ldlm_lock *lock;
721 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
722 ldlm_lock_dump(level, lock, ++pos);
725 if (!list_empty(&res->lr_converting)) {
727 CDEBUG(level, "Converting locks:\n");
728 list_for_each(tmp, &res->lr_converting) {
729 struct ldlm_lock *lock;
730 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
731 ldlm_lock_dump(level, lock, ++pos);
734 if (!list_empty(&res->lr_waiting)) {
736 CDEBUG(level, "Waiting locks:\n");
737 list_for_each(tmp, &res->lr_waiting) {
738 struct ldlm_lock *lock;
739 lock = list_entry(tmp, struct ldlm_lock, l_res_link);
740 ldlm_lock_dump(level, lock, ++pos);