4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2012, Intel Corporation.
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
43 # include <lustre_dlm.h>
45 #include <lustre_fid.h>
46 #include <obd_class.h>
47 #include "ldlm_internal.h"
49 struct kmem_cache *ldlm_resource_slab, *ldlm_lock_slab;
51 atomic_t ldlm_srv_namespace_nr = ATOMIC_INIT(0);
52 atomic_t ldlm_cli_namespace_nr = ATOMIC_INIT(0);
54 struct mutex ldlm_srv_namespace_lock;
55 LIST_HEAD(ldlm_srv_namespace_list);
57 struct mutex ldlm_cli_namespace_lock;
58 LIST_HEAD(ldlm_cli_namespace_list);
60 proc_dir_entry_t *ldlm_type_proc_dir = NULL;
61 proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
62 proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
64 extern unsigned int ldlm_cancel_unused_locks_before_replay;
66 /* during debug dump certain amount of granted locks for one resource to avoid
68 unsigned int ldlm_dump_granted_max = 256;
71 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
72 unsigned long count, void *data)
74 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
75 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
79 int ldlm_proc_setup(void)
82 struct lprocfs_vars list[] = {
83 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
85 lprocfs_rd_uint, lprocfs_wr_uint,
86 &ldlm_dump_granted_max, NULL },
87 { "cancel_unused_locks_before_replay",
88 lprocfs_rd_uint, lprocfs_wr_uint,
89 &ldlm_cancel_unused_locks_before_replay, NULL },
92 LASSERT(ldlm_ns_proc_dir == NULL);
94 ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
97 if (IS_ERR(ldlm_type_proc_dir)) {
98 CERROR("LProcFS failed in ldlm-init\n");
99 rc = PTR_ERR(ldlm_type_proc_dir);
103 ldlm_ns_proc_dir = lprocfs_register("namespaces",
106 if (IS_ERR(ldlm_ns_proc_dir)) {
107 CERROR("LProcFS failed in ldlm-init\n");
108 rc = PTR_ERR(ldlm_ns_proc_dir);
112 ldlm_svc_proc_dir = lprocfs_register("services",
115 if (IS_ERR(ldlm_svc_proc_dir)) {
116 CERROR("LProcFS failed in ldlm-init\n");
117 rc = PTR_ERR(ldlm_svc_proc_dir);
121 rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
126 lprocfs_remove(&ldlm_ns_proc_dir);
128 lprocfs_remove(&ldlm_type_proc_dir);
130 ldlm_svc_proc_dir = NULL;
134 void ldlm_proc_cleanup(void)
136 if (ldlm_svc_proc_dir)
137 lprocfs_remove(&ldlm_svc_proc_dir);
139 if (ldlm_ns_proc_dir)
140 lprocfs_remove(&ldlm_ns_proc_dir);
142 if (ldlm_type_proc_dir)
143 lprocfs_remove(&ldlm_type_proc_dir);
146 static int lprocfs_rd_ns_resources(char *page, char **start, off_t off,
147 int count, int *eof, void *data)
149 struct ldlm_namespace *ns = data;
154 /* result is not strictly consistant */
155 cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
156 res += cfs_hash_bd_count_get(&bd);
157 return lprocfs_rd_u64(page, start, off, count, eof, &res);
160 static int lprocfs_rd_ns_locks(char *page, char **start, off_t off,
161 int count, int *eof, void *data)
163 struct ldlm_namespace *ns = data;
166 locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
167 LPROCFS_FIELDS_FLAGS_SUM);
168 return lprocfs_rd_u64(page, start, off, count, eof, &locks);
171 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
172 int count, int *eof, void *data)
174 struct ldlm_namespace *ns = data;
175 __u32 *nr = &ns->ns_max_unused;
177 if (ns_connect_lru_resize(ns))
178 nr = &ns->ns_nr_unused;
179 return lprocfs_rd_uint(page, start, off, count, eof, nr);
182 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
183 unsigned long count, void *data)
185 struct ldlm_namespace *ns = data;
186 char dummy[MAX_STRING_SIZE + 1], *end;
190 dummy[MAX_STRING_SIZE] = '\0';
191 if (copy_from_user(dummy, buffer, MAX_STRING_SIZE))
194 if (strncmp(dummy, "clear", 5) == 0) {
196 "dropping all unused locks from namespace %s\n",
198 if (ns_connect_lru_resize(ns)) {
199 int canceled, unused = ns->ns_nr_unused;
201 /* Try to cancel all @ns_nr_unused locks. */
202 canceled = ldlm_cancel_lru(ns, unused, 0,
204 if (canceled < unused) {
206 "not all requested locks are canceled, "
207 "requested: %d, canceled: %d\n", unused,
212 tmp = ns->ns_max_unused;
213 ns->ns_max_unused = 0;
214 ldlm_cancel_lru(ns, 0, 0, LDLM_CANCEL_PASSED);
215 ns->ns_max_unused = tmp;
220 tmp = simple_strtoul(dummy, &end, 0);
222 CERROR("invalid value written\n");
225 lru_resize = (tmp == 0);
227 if (ns_connect_lru_resize(ns)) {
229 ns->ns_max_unused = (unsigned int)tmp;
231 if (tmp > ns->ns_nr_unused)
232 tmp = ns->ns_nr_unused;
233 tmp = ns->ns_nr_unused - tmp;
236 "changing namespace %s unused locks from %u to %u\n",
237 ldlm_ns_name(ns), ns->ns_nr_unused,
239 ldlm_cancel_lru(ns, tmp, LCF_ASYNC, LDLM_CANCEL_PASSED);
243 "disable lru_resize for namespace %s\n",
245 ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
249 "changing namespace %s max_unused from %u to %u\n",
250 ldlm_ns_name(ns), ns->ns_max_unused,
252 ns->ns_max_unused = (unsigned int)tmp;
253 ldlm_cancel_lru(ns, 0, LCF_ASYNC, LDLM_CANCEL_PASSED);
255 /* Make sure that LRU resize was originally supported before
256 * turning it on here. */
258 (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
260 "enable lru_resize for namespace %s\n",
262 ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
269 static int lprocfs_rd_elc(char *page, char **start, off_t off,
270 int count, int *eof, void *data)
272 struct ldlm_namespace *ns = data;
273 unsigned int supp = ns_connect_cancelset(ns);
275 return lprocfs_rd_uint(page, start, off, count, eof, &supp);
278 static int lprocfs_wr_elc(struct file *file, const char *buffer,
279 unsigned long count, void *data)
281 struct ldlm_namespace *ns = data;
282 unsigned int supp = -1;
285 rc = lprocfs_wr_uint(file, buffer, count, &supp);
290 ns->ns_connect_flags &= ~OBD_CONNECT_CANCELSET;
291 else if (ns->ns_orig_connect_flags & OBD_CONNECT_CANCELSET)
292 ns->ns_connect_flags |= OBD_CONNECT_CANCELSET;
296 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
298 struct proc_dir_entry *dir;
300 dir = lprocfs_srch(ldlm_ns_proc_dir, ldlm_ns_name(ns));
302 CERROR("dlm namespace %s has no procfs dir?\n",
305 lprocfs_remove(&dir);
308 if (ns->ns_stats != NULL)
309 lprocfs_free_stats(&ns->ns_stats);
312 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
314 struct lprocfs_vars lock_vars[2];
315 char lock_name[MAX_STRING_SIZE + 1];
318 LASSERT(ns->ns_rs_hash != NULL);
320 ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
321 if (ns->ns_stats == NULL)
324 lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
325 LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
327 lock_name[MAX_STRING_SIZE] = '\0';
329 memset(lock_vars, 0, sizeof(lock_vars));
330 lock_vars[0].name = lock_name;
332 snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count",
334 lock_vars[0].data = ns;
335 lock_vars[0].read_fptr = lprocfs_rd_ns_resources;
336 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
338 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count",
340 lock_vars[0].data = ns;
341 lock_vars[0].read_fptr = lprocfs_rd_ns_locks;
342 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
344 if (ns_is_client(ns)) {
345 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
347 lock_vars[0].data = &ns->ns_nr_unused;
348 lock_vars[0].read_fptr = lprocfs_rd_uint;
349 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
351 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
353 lock_vars[0].data = ns;
354 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
355 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
356 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
358 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
360 lock_vars[0].data = &ns->ns_max_age;
361 lock_vars[0].read_fptr = lprocfs_rd_uint;
362 lock_vars[0].write_fptr = lprocfs_wr_uint;
363 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
365 snprintf(lock_name, MAX_STRING_SIZE, "%s/early_lock_cancel",
367 lock_vars[0].data = ns;
368 lock_vars[0].read_fptr = lprocfs_rd_elc;
369 lock_vars[0].write_fptr = lprocfs_wr_elc;
370 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
372 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
374 lock_vars[0].data = &ns->ns_ctime_age_limit;
375 lock_vars[0].read_fptr = lprocfs_rd_uint;
376 lock_vars[0].write_fptr = lprocfs_wr_uint;
377 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
379 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
381 lock_vars[0].data = &ns->ns_timeouts;
382 lock_vars[0].read_fptr = lprocfs_rd_uint;
383 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
385 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
387 lock_vars[0].data = &ns->ns_max_nolock_size;
388 lock_vars[0].read_fptr = lprocfs_rd_uint;
389 lock_vars[0].write_fptr = lprocfs_wr_uint;
390 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
392 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
394 lock_vars[0].data = &ns->ns_contention_time;
395 lock_vars[0].read_fptr = lprocfs_rd_uint;
396 lock_vars[0].write_fptr = lprocfs_wr_uint;
397 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
399 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
401 lock_vars[0].data = &ns->ns_contended_locks;
402 lock_vars[0].read_fptr = lprocfs_rd_uint;
403 lock_vars[0].write_fptr = lprocfs_wr_uint;
404 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
406 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_parallel_ast",
408 lock_vars[0].data = &ns->ns_max_parallel_ast;
409 lock_vars[0].read_fptr = lprocfs_rd_uint;
410 lock_vars[0].write_fptr = lprocfs_wr_uint;
411 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
415 #undef MAX_STRING_SIZE
418 #define ldlm_namespace_proc_unregister(ns) ({;})
419 #define ldlm_namespace_proc_register(ns) ({0;})
423 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
424 const void *key, unsigned mask)
426 const struct ldlm_res_id *id = key;
430 for (i = 0; i < RES_NAME_SIZE; i++)
435 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
436 const void *key, unsigned mask)
438 const struct ldlm_res_id *id = key;
443 fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
444 fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
445 fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
447 hash = fid_flatten32(&fid);
448 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
449 if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
450 val = id->name[LUSTRE_RES_ID_HSH_OFF];
451 hash += (val >> 5) + (val << 11);
455 hash = cfs_hash_long(hash, hs->hs_bkt_bits);
456 /* give me another random factor */
457 hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3);
459 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
460 hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
465 static void *ldlm_res_hop_key(struct hlist_node *hnode)
467 struct ldlm_resource *res;
469 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
470 return &res->lr_name;
473 static int ldlm_res_hop_keycmp(const void *key, struct hlist_node *hnode)
475 struct ldlm_resource *res;
477 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
478 return ldlm_res_eq((const struct ldlm_res_id *)key,
479 (const struct ldlm_res_id *)&res->lr_name);
482 static void *ldlm_res_hop_object(struct hlist_node *hnode)
484 return hlist_entry(hnode, struct ldlm_resource, lr_hash);
487 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, struct hlist_node *hnode)
489 struct ldlm_resource *res;
491 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
492 ldlm_resource_getref(res);
495 static void ldlm_res_hop_put_locked(cfs_hash_t *hs, struct hlist_node *hnode)
497 struct ldlm_resource *res;
499 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
500 /* cfs_hash_for_each_nolock is the only chance we call it */
501 ldlm_resource_putref_locked(res);
504 static void ldlm_res_hop_put(cfs_hash_t *hs, struct hlist_node *hnode)
506 struct ldlm_resource *res;
508 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
509 ldlm_resource_putref(res);
512 cfs_hash_ops_t ldlm_ns_hash_ops = {
513 .hs_hash = ldlm_res_hop_hash,
514 .hs_key = ldlm_res_hop_key,
515 .hs_keycmp = ldlm_res_hop_keycmp,
517 .hs_object = ldlm_res_hop_object,
518 .hs_get = ldlm_res_hop_get_locked,
519 .hs_put_locked = ldlm_res_hop_put_locked,
520 .hs_put = ldlm_res_hop_put
523 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
524 .hs_hash = ldlm_res_hop_fid_hash,
525 .hs_key = ldlm_res_hop_key,
526 .hs_keycmp = ldlm_res_hop_keycmp,
528 .hs_object = ldlm_res_hop_object,
529 .hs_get = ldlm_res_hop_get_locked,
530 .hs_put_locked = ldlm_res_hop_put_locked,
531 .hs_put = ldlm_res_hop_put
535 ldlm_ns_type_t nsd_type;
536 /** hash bucket bits */
537 unsigned nsd_bkt_bits;
539 unsigned nsd_all_bits;
540 /** hash operations */
541 cfs_hash_ops_t *nsd_hops;
542 } ldlm_ns_hash_def_t;
544 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
547 .nsd_type = LDLM_NS_TYPE_MDC,
550 .nsd_hops = &ldlm_ns_fid_hash_ops,
553 .nsd_type = LDLM_NS_TYPE_MDT,
556 .nsd_hops = &ldlm_ns_fid_hash_ops,
559 .nsd_type = LDLM_NS_TYPE_OSC,
562 .nsd_hops = &ldlm_ns_hash_ops,
565 .nsd_type = LDLM_NS_TYPE_OST,
568 .nsd_hops = &ldlm_ns_hash_ops,
571 .nsd_type = LDLM_NS_TYPE_MGC,
574 .nsd_hops = &ldlm_ns_hash_ops,
577 .nsd_type = LDLM_NS_TYPE_MGT,
580 .nsd_hops = &ldlm_ns_hash_ops,
583 .nsd_type = LDLM_NS_TYPE_UNKNOWN,
588 * Create and initialize new empty namespace.
590 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
593 ldlm_ns_type_t ns_type)
595 struct ldlm_namespace *ns = NULL;
596 struct ldlm_ns_bucket *nsb;
597 ldlm_ns_hash_def_t *nsd;
603 LASSERT(obd != NULL);
607 CERROR("ldlm_get_ref failed: %d\n", rc);
611 for (idx = 0;;idx++) {
612 nsd = &ldlm_ns_hash_defs[idx];
613 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
614 CERROR("Unknown type %d for ns %s\n", ns_type, name);
618 if (nsd->nsd_type == ns_type)
626 ns->ns_rs_hash = cfs_hash_create(name,
627 nsd->nsd_all_bits, nsd->nsd_all_bits,
628 nsd->nsd_bkt_bits, sizeof(*nsb),
634 CFS_HASH_SPIN_BKTLOCK |
635 CFS_HASH_NO_ITEMREF);
636 if (ns->ns_rs_hash == NULL)
639 cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
640 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
641 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
642 nsb->nsb_namespace = ns;
646 ns->ns_appetite = apt;
647 ns->ns_client = client;
649 INIT_LIST_HEAD(&ns->ns_list_chain);
650 INIT_LIST_HEAD(&ns->ns_unused_list);
651 spin_lock_init(&ns->ns_lock);
652 atomic_set(&ns->ns_bref, 0);
653 init_waitqueue_head(&ns->ns_waitq);
655 ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
656 ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
657 ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
659 ns->ns_max_parallel_ast = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
660 ns->ns_nr_unused = 0;
661 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
662 ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
663 ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
665 ns->ns_orig_connect_flags = 0;
666 ns->ns_connect_flags = 0;
668 rc = ldlm_namespace_proc_register(ns);
670 CERROR("Can't initialize ns proc, rc %d\n", rc);
674 idx = atomic_read(ldlm_namespace_nr(client));
675 rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
677 CERROR("Can't initialize lock pool, rc %d\n", rc);
681 ldlm_namespace_register(ns, client);
684 ldlm_namespace_proc_unregister(ns);
685 ldlm_namespace_cleanup(ns, 0);
687 cfs_hash_putref(ns->ns_rs_hash);
694 EXPORT_SYMBOL(ldlm_namespace_new);
696 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
699 * Cancel and destroy all locks on a resource.
701 * If flags contains FL_LOCAL_ONLY, don't try to tell the server, just
702 * clean up. This is currently only used for recovery, and we make
703 * certain assumptions as a result--notably, that we shouldn't cancel
706 static void cleanup_resource(struct ldlm_resource *res, struct list_head *q,
709 struct list_head *tmp;
710 int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
711 bool local_only = !!(flags & LDLM_FL_LOCAL_ONLY);
714 struct ldlm_lock *lock = NULL;
716 /* First, we look for non-cleaned-yet lock
717 * all cleaned locks are marked by CLEANED flag. */
719 list_for_each(tmp, q) {
720 lock = list_entry(tmp, struct ldlm_lock,
722 if (lock->l_flags & LDLM_FL_CLEANED) {
727 lock->l_flags |= LDLM_FL_CLEANED;
736 /* Set CBPENDING so nothing in the cancellation path
737 * can match this lock. */
738 lock->l_flags |= LDLM_FL_CBPENDING;
739 lock->l_flags |= LDLM_FL_FAILED;
740 lock->l_flags |= flags;
742 /* ... without sending a CANCEL message for local_only. */
744 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
746 if (local_only && (lock->l_readers || lock->l_writers)) {
747 /* This is a little bit gross, but much better than the
748 * alternative: pretend that we got a blocking AST from
749 * the server, so that when the lock is decref'd, it
750 * will go away ... */
752 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
753 if (lock->l_completion_ast)
754 lock->l_completion_ast(lock, 0, NULL);
755 LDLM_LOCK_RELEASE(lock);
760 struct lustre_handle lockh;
763 ldlm_lock2handle(lock, &lockh);
764 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
766 CERROR("ldlm_cli_cancel: %d\n", rc);
768 ldlm_resource_unlink_lock(lock);
770 LDLM_DEBUG(lock, "Freeing a lock still held by a "
772 ldlm_lock_destroy(lock);
774 LDLM_LOCK_RELEASE(lock);
778 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
779 struct hlist_node *hnode, void *arg)
781 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
782 __u64 flags = *(__u64 *)arg;
784 cleanup_resource(res, &res->lr_granted, flags);
785 cleanup_resource(res, &res->lr_converting, flags);
786 cleanup_resource(res, &res->lr_waiting, flags);
791 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
792 struct hlist_node *hnode, void *arg)
794 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
797 CERROR("Namespace %s resource refcount nonzero "
798 "(%d) after lock cleanup; forcing "
800 ldlm_ns_name(ldlm_res_to_ns(res)),
801 atomic_read(&res->lr_refcount) - 1);
803 CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
804 LPU64") (rc: %d)\n", res,
805 res->lr_name.name[0], res->lr_name.name[1],
806 res->lr_name.name[2], res->lr_name.name[3],
807 atomic_read(&res->lr_refcount) - 1);
809 ldlm_resource_dump(D_ERROR, res);
815 * Cancel and destroy all locks in the namespace.
817 * Typically used during evictions when server notified client that it was
818 * evicted and all of its state needs to be destroyed.
819 * Also used during shutdown.
821 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, __u64 flags)
824 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
828 cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean, &flags);
829 cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
832 EXPORT_SYMBOL(ldlm_namespace_cleanup);
835 * Attempts to free namespace.
837 * Only used when namespace goes away, like during an unmount.
839 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
843 /* At shutdown time, don't call the cancellation callback */
844 ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
846 if (atomic_read(&ns->ns_bref) > 0) {
847 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
850 "dlm namespace %s free waiting on refcount %d\n",
851 ldlm_ns_name(ns), atomic_read(&ns->ns_bref));
854 lwi = LWI_TIMEOUT(obd_timeout * HZ / 4, NULL, NULL);
856 rc = l_wait_event(ns->ns_waitq,
857 atomic_read(&ns->ns_bref) == 0, &lwi);
859 /* Forced cleanups should be able to reclaim all references,
860 * so it's safe to wait forever... we can't leak locks... */
861 if (force && rc == -ETIMEDOUT) {
862 LCONSOLE_ERROR("Forced cleanup waiting for %s "
863 "namespace with %d resources in use, "
864 "(rc=%d)\n", ldlm_ns_name(ns),
865 atomic_read(&ns->ns_bref), rc);
866 GOTO(force_wait, rc);
869 if (atomic_read(&ns->ns_bref)) {
870 LCONSOLE_ERROR("Cleanup waiting for %s namespace "
871 "with %d resources in use, (rc=%d)\n",
873 atomic_read(&ns->ns_bref), rc);
874 RETURN(ELDLM_NAMESPACE_EXISTS);
876 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
884 * Performs various cleanups for passed \a ns to make it drop refc and be
885 * ready for freeing. Waits for refc == 0.
887 * The following is done:
888 * (0) Unregister \a ns from its list to make inaccessible for potential
889 * users like pools thread and others;
890 * (1) Clear all locks in \a ns.
892 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
893 struct obd_import *imp,
903 spin_lock(&ns->ns_lock);
905 spin_unlock(&ns->ns_lock);
908 * Can fail with -EINTR when force == 0 in which case try harder.
910 rc = __ldlm_namespace_free(ns, force);
911 if (rc != ELDLM_OK) {
913 ptlrpc_disconnect_import(imp, 0);
914 ptlrpc_invalidate_import(imp);
918 * With all requests dropped and the import inactive
919 * we are gaurenteed all reference will be dropped.
921 rc = __ldlm_namespace_free(ns, 1);
928 * Performs freeing memory structures related to \a ns. This is only done
929 * when ldlm_namespce_free_prior() successfully removed all resources
930 * referencing \a ns and its refc == 0.
932 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
940 /* Make sure that nobody can find this ns in its list. */
941 ldlm_namespace_unregister(ns, ns->ns_client);
942 /* Fini pool _before_ parent proc dir is removed. This is important as
943 * ldlm_pool_fini() removes own proc dir which is child to @dir.
944 * Removing it after @dir may cause oops. */
945 ldlm_pool_fini(&ns->ns_pool);
947 ldlm_namespace_proc_unregister(ns);
948 cfs_hash_putref(ns->ns_rs_hash);
949 /* Namespace \a ns should be not on list at this time, otherwise
950 * this will cause issues related to using freed \a ns in poold
952 LASSERT(list_empty(&ns->ns_list_chain));
959 * Cleanup the resource, and free namespace.
962 * proc1: destroy import
963 * class_disconnect_export(grab cl_sem) ->
964 * -> ldlm_namespace_free ->
965 * -> lprocfs_remove(grab _lprocfs_lock).
966 * proc2: read proc info
967 * lprocfs_fops_read(grab _lprocfs_lock) ->
968 * -> osc_rd_active, etc(grab cl_sem).
970 * So that I have to split the ldlm_namespace_free into two parts - the first
971 * part ldlm_namespace_free_prior is used to cleanup the resource which is
972 * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
973 * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
976 void ldlm_namespace_free(struct ldlm_namespace *ns,
977 struct obd_import *imp,
980 ldlm_namespace_free_prior(ns, imp, force);
981 ldlm_namespace_free_post(ns);
983 EXPORT_SYMBOL(ldlm_namespace_free);
985 void ldlm_namespace_get(struct ldlm_namespace *ns)
987 atomic_inc(&ns->ns_bref);
989 EXPORT_SYMBOL(ldlm_namespace_get);
991 void ldlm_namespace_put(struct ldlm_namespace *ns)
993 if (atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
994 wake_up(&ns->ns_waitq);
995 spin_unlock(&ns->ns_lock);
998 EXPORT_SYMBOL(ldlm_namespace_put);
1000 /** Register \a ns in the list of namespaces */
1001 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
1003 mutex_lock(ldlm_namespace_lock(client));
1004 LASSERT(list_empty(&ns->ns_list_chain));
1005 list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
1006 atomic_inc(ldlm_namespace_nr(client));
1007 mutex_unlock(ldlm_namespace_lock(client));
1010 /** Unregister \a ns from the list of namespaces. */
1011 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
1013 mutex_lock(ldlm_namespace_lock(client));
1014 LASSERT(!list_empty(&ns->ns_list_chain));
1015 /* Some asserts and possibly other parts of the code are still
1016 * using list_empty(&ns->ns_list_chain). This is why it is
1017 * important to use list_del_init() here. */
1018 list_del_init(&ns->ns_list_chain);
1019 atomic_dec(ldlm_namespace_nr(client));
1020 mutex_unlock(ldlm_namespace_lock(client));
1023 /** Should be called with ldlm_namespace_lock(client) taken. */
1024 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
1026 LASSERT(!list_empty(&ns->ns_list_chain));
1027 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1028 list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
1031 /** Should be called with ldlm_namespace_lock(client) taken. */
1032 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
1034 LASSERT(mutex_is_locked(ldlm_namespace_lock(client)));
1035 LASSERT(!list_empty(ldlm_namespace_list(client)));
1036 return container_of(ldlm_namespace_list(client)->next,
1037 struct ldlm_namespace, ns_list_chain);
1040 /** Create and initialize new resource. */
1041 static struct ldlm_resource *ldlm_resource_new(void)
1043 struct ldlm_resource *res;
1046 OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, __GFP_IO);
1050 INIT_LIST_HEAD(&res->lr_granted);
1051 INIT_LIST_HEAD(&res->lr_converting);
1052 INIT_LIST_HEAD(&res->lr_waiting);
1054 /* Initialize interval trees for each lock mode. */
1055 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1056 res->lr_itree[idx].lit_size = 0;
1057 res->lr_itree[idx].lit_mode = 1 << idx;
1058 res->lr_itree[idx].lit_root = NULL;
1061 atomic_set(&res->lr_refcount, 1);
1062 spin_lock_init(&res->lr_lock);
1063 lu_ref_init(&res->lr_reference);
1065 /* The creator of the resource must unlock the mutex after LVB
1066 * initialization. */
1067 mutex_init(&res->lr_lvb_mutex);
1068 mutex_lock(&res->lr_lvb_mutex);
1074 * Return a reference to resource with given name, creating it if necessary.
1075 * Args: namespace with ns_lock unlocked
1076 * Locks: takes and releases NS hash-lock and res->lr_lock
1077 * Returns: referenced, unlocked ldlm_resource or NULL
1079 struct ldlm_resource *
1080 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1081 const struct ldlm_res_id *name, ldlm_type_t type, int create)
1083 struct hlist_node *hnode;
1084 struct ldlm_resource *res;
1088 LASSERT(ns != NULL);
1089 LASSERT(parent == NULL);
1090 LASSERT(ns->ns_rs_hash != NULL);
1091 LASSERT(name->name[0] != 0);
1093 cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1094 hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1095 if (hnode != NULL) {
1096 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1097 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1098 /* Synchronize with regard to resource creation. */
1099 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1100 mutex_lock(&res->lr_lvb_mutex);
1101 mutex_unlock(&res->lr_lvb_mutex);
1104 if (unlikely(res->lr_lvb_len < 0)) {
1105 ldlm_resource_putref(res);
1111 version = cfs_hash_bd_version_get(&bd);
1112 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1117 LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1118 "type: %d\n", type);
1119 res = ldlm_resource_new();
1123 res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1124 res->lr_name = *name;
1125 res->lr_type = type;
1126 res->lr_most_restr = LCK_NL;
1128 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1129 hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1130 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1132 if (hnode != NULL) {
1133 /* Someone won the race and already added the resource. */
1134 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1135 /* Clean lu_ref for failed resource. */
1136 lu_ref_fini(&res->lr_reference);
1137 /* We have taken lr_lvb_mutex. Drop it. */
1138 mutex_unlock(&res->lr_lvb_mutex);
1139 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1141 res = hlist_entry(hnode, struct ldlm_resource, lr_hash);
1142 /* Synchronize with regard to resource creation. */
1143 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1144 mutex_lock(&res->lr_lvb_mutex);
1145 mutex_unlock(&res->lr_lvb_mutex);
1148 if (unlikely(res->lr_lvb_len < 0)) {
1149 ldlm_resource_putref(res);
1154 /* We won! Let's add the resource. */
1155 cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1156 if (cfs_hash_bd_count_get(&bd) == 1)
1157 ldlm_namespace_get(ns);
1159 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1160 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1163 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1164 rc = ns->ns_lvbo->lvbo_init(res);
1166 CERROR("lvbo_init failed for resource "
1167 LPU64": rc %d\n", name->name[0], rc);
1168 if (res->lr_lvb_data) {
1169 OBD_FREE(res->lr_lvb_data, res->lr_lvb_len);
1170 res->lr_lvb_data = NULL;
1172 res->lr_lvb_len = rc;
1173 mutex_unlock(&res->lr_lvb_mutex);
1174 ldlm_resource_putref(res);
1179 /* We create resource with locked lr_lvb_mutex. */
1180 mutex_unlock(&res->lr_lvb_mutex);
1184 EXPORT_SYMBOL(ldlm_resource_get);
1186 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1188 LASSERT(res != NULL);
1189 LASSERT(res != LP_POISON);
1190 atomic_inc(&res->lr_refcount);
1191 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1192 atomic_read(&res->lr_refcount));
1196 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1197 struct ldlm_resource *res)
1199 struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1201 if (!list_empty(&res->lr_granted)) {
1202 ldlm_resource_dump(D_ERROR, res);
1206 if (!list_empty(&res->lr_converting)) {
1207 ldlm_resource_dump(D_ERROR, res);
1211 if (!list_empty(&res->lr_waiting)) {
1212 ldlm_resource_dump(D_ERROR, res);
1216 cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1218 lu_ref_fini(&res->lr_reference);
1219 if (cfs_hash_bd_count_get(bd) == 0)
1220 ldlm_namespace_put(nsb->nsb_namespace);
1223 /* Returns 1 if the resource was freed, 0 if it remains. */
1224 int ldlm_resource_putref(struct ldlm_resource *res)
1226 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1229 LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1230 CDEBUG(D_INFO, "putref res: %p count: %d\n",
1231 res, atomic_read(&res->lr_refcount) - 1);
1233 cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1234 if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1235 __ldlm_resource_putref_final(&bd, res);
1236 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1237 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1238 ns->ns_lvbo->lvbo_free(res);
1239 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1244 EXPORT_SYMBOL(ldlm_resource_putref);
1246 /* Returns 1 if the resource was freed, 0 if it remains. */
1247 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1249 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1251 LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1252 CDEBUG(D_INFO, "putref res: %p count: %d\n",
1253 res, atomic_read(&res->lr_refcount) - 1);
1255 if (atomic_dec_and_test(&res->lr_refcount)) {
1258 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1259 &res->lr_name, &bd);
1260 __ldlm_resource_putref_final(&bd, res);
1261 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1262 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1263 * so we should never be here while calling cfs_hash_del,
1264 * cfs_hash_for_each_nolock is the only case we can get
1265 * here, which is safe to release cfs_hash_bd_lock.
1267 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1268 ns->ns_lvbo->lvbo_free(res);
1269 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1271 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1278 * Add a lock into a given resource into specified lock list.
1280 void ldlm_resource_add_lock(struct ldlm_resource *res, struct list_head *head,
1281 struct ldlm_lock *lock)
1283 check_res_locked(res);
1285 LDLM_DEBUG(lock, "About to add this lock:\n");
1287 if (lock->l_destroyed) {
1288 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1292 LASSERT(list_empty(&lock->l_res_link));
1294 list_add_tail(&lock->l_res_link, head);
1298 * Insert a lock into resource after specified lock.
1300 * Obtain resource description from the lock we are inserting after.
1302 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1303 struct ldlm_lock *new)
1305 struct ldlm_resource *res = original->l_resource;
1307 check_res_locked(res);
1309 ldlm_resource_dump(D_INFO, res);
1310 LDLM_DEBUG(new, "About to insert this lock after %p:\n", original);
1312 if (new->l_destroyed) {
1313 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1317 LASSERT(list_empty(&new->l_res_link));
1319 list_add(&new->l_res_link, &original->l_res_link);
1323 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1325 int type = lock->l_resource->lr_type;
1327 check_res_locked(lock->l_resource);
1328 if (type == LDLM_IBITS || type == LDLM_PLAIN)
1329 ldlm_unlink_lock_skiplist(lock);
1330 else if (type == LDLM_EXTENT)
1331 ldlm_extent_unlink_lock(lock);
1332 list_del_init(&lock->l_res_link);
1334 EXPORT_SYMBOL(ldlm_resource_unlink_lock);
1336 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1338 desc->lr_type = res->lr_type;
1339 desc->lr_name = res->lr_name;
1343 * Print information about all locks in all namespaces on this node to debug
1346 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1348 struct list_head *tmp;
1350 if (!((libcfs_debug | D_ERROR) & level))
1353 mutex_lock(ldlm_namespace_lock(client));
1355 list_for_each(tmp, ldlm_namespace_list(client)) {
1356 struct ldlm_namespace *ns;
1357 ns = list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1358 ldlm_namespace_dump(level, ns);
1361 mutex_unlock(ldlm_namespace_lock(client));
1363 EXPORT_SYMBOL(ldlm_dump_all_namespaces);
1365 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1366 struct hlist_node *hnode, void *arg)
1368 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1369 int level = (int)(unsigned long)arg;
1372 ldlm_resource_dump(level, res);
1379 * Print information about all locks in this namespace on this node to debug
1382 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1384 if (!((libcfs_debug | D_ERROR) & level))
1387 CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1388 ldlm_ns_name(ns), atomic_read(&ns->ns_bref),
1389 ns_is_client(ns) ? "client" : "server");
1391 if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1394 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1396 (void *)(unsigned long)level);
1397 spin_lock(&ns->ns_lock);
1398 ns->ns_next_dump = cfs_time_shift(10);
1399 spin_unlock(&ns->ns_lock);
1401 EXPORT_SYMBOL(ldlm_namespace_dump);
1404 * Print information about all locks in this resource to debug log.
1406 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1408 struct ldlm_lock *lock;
1409 unsigned int granted = 0;
1411 CLASSERT(RES_NAME_SIZE == 4);
1413 if (!((libcfs_debug | D_ERROR) & level))
1416 CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1417 ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1418 res->lr_name.name[2], res->lr_name.name[3],
1419 atomic_read(&res->lr_refcount));
1421 if (!list_empty(&res->lr_granted)) {
1422 CDEBUG(level, "Granted locks (in reverse order):\n");
1423 list_for_each_entry_reverse(lock, &res->lr_granted,
1425 LDLM_DEBUG_LIMIT(level, lock, "###");
1426 if (!(level & D_CANTMASK) &&
1427 ++granted > ldlm_dump_granted_max) {
1428 CDEBUG(level, "only dump %d granted locks to "
1429 "avoid DDOS.\n", granted);
1434 if (!list_empty(&res->lr_converting)) {
1435 CDEBUG(level, "Converting locks:\n");
1436 list_for_each_entry(lock, &res->lr_converting, l_res_link)
1437 LDLM_DEBUG_LIMIT(level, lock, "###");
1439 if (!list_empty(&res->lr_waiting)) {
1440 CDEBUG(level, "Waiting locks:\n");
1441 list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1442 LDLM_DEBUG_LIMIT(level, lock, "###");