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) 2011, 2012, Whamcloud, Inc.
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 <lustre_fid.h>
50 #include <obd_class.h>
51 #include "ldlm_internal.h"
53 cfs_mem_cache_t *ldlm_resource_slab, *ldlm_lock_slab;
55 cfs_atomic_t ldlm_srv_namespace_nr = CFS_ATOMIC_INIT(0);
56 cfs_atomic_t ldlm_cli_namespace_nr = CFS_ATOMIC_INIT(0);
58 cfs_mutex_t ldlm_srv_namespace_lock;
59 CFS_LIST_HEAD(ldlm_srv_namespace_list);
61 cfs_mutex_t ldlm_cli_namespace_lock;
62 CFS_LIST_HEAD(ldlm_cli_namespace_list);
64 cfs_proc_dir_entry_t *ldlm_type_proc_dir = NULL;
65 cfs_proc_dir_entry_t *ldlm_ns_proc_dir = NULL;
66 cfs_proc_dir_entry_t *ldlm_svc_proc_dir = NULL;
68 extern unsigned int ldlm_cancel_unused_locks_before_replay;
70 /* during debug dump certain amount of granted locks for one resource to avoid
72 unsigned int ldlm_dump_granted_max = 256;
75 static int ldlm_proc_dump_ns(struct file *file, const char *buffer,
76 unsigned long count, void *data)
78 ldlm_dump_all_namespaces(LDLM_NAMESPACE_SERVER, D_DLMTRACE);
79 ldlm_dump_all_namespaces(LDLM_NAMESPACE_CLIENT, D_DLMTRACE);
83 int ldlm_proc_setup(void)
86 struct lprocfs_vars list[] = {
87 { "dump_namespaces", NULL, ldlm_proc_dump_ns, NULL },
89 lprocfs_rd_uint, lprocfs_wr_uint,
90 &ldlm_dump_granted_max, NULL },
91 { "cancel_unused_locks_before_replay",
92 lprocfs_rd_uint, lprocfs_wr_uint,
93 &ldlm_cancel_unused_locks_before_replay, NULL },
96 LASSERT(ldlm_ns_proc_dir == NULL);
98 ldlm_type_proc_dir = lprocfs_register(OBD_LDLM_DEVICENAME,
101 if (IS_ERR(ldlm_type_proc_dir)) {
102 CERROR("LProcFS failed in ldlm-init\n");
103 rc = PTR_ERR(ldlm_type_proc_dir);
107 ldlm_ns_proc_dir = lprocfs_register("namespaces",
110 if (IS_ERR(ldlm_ns_proc_dir)) {
111 CERROR("LProcFS failed in ldlm-init\n");
112 rc = PTR_ERR(ldlm_ns_proc_dir);
116 ldlm_svc_proc_dir = lprocfs_register("services",
119 if (IS_ERR(ldlm_svc_proc_dir)) {
120 CERROR("LProcFS failed in ldlm-init\n");
121 rc = PTR_ERR(ldlm_svc_proc_dir);
125 rc = lprocfs_add_vars(ldlm_type_proc_dir, list, NULL);
130 lprocfs_remove(&ldlm_ns_proc_dir);
132 lprocfs_remove(&ldlm_type_proc_dir);
134 ldlm_svc_proc_dir = NULL;
138 void ldlm_proc_cleanup(void)
140 if (ldlm_svc_proc_dir)
141 lprocfs_remove(&ldlm_svc_proc_dir);
143 if (ldlm_ns_proc_dir)
144 lprocfs_remove(&ldlm_ns_proc_dir);
146 if (ldlm_type_proc_dir)
147 lprocfs_remove(&ldlm_type_proc_dir);
150 static int lprocfs_rd_ns_resources(char *page, char **start, off_t off,
151 int count, int *eof, void *data)
153 struct ldlm_namespace *ns = data;
158 /* result is not strictly consistant */
159 cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, i)
160 res += cfs_hash_bd_count_get(&bd);
161 return lprocfs_rd_u64(page, start, off, count, eof, &res);
164 static int lprocfs_rd_ns_locks(char *page, char **start, off_t off,
165 int count, int *eof, void *data)
167 struct ldlm_namespace *ns = data;
170 locks = lprocfs_stats_collector(ns->ns_stats, LDLM_NSS_LOCKS,
171 LPROCFS_FIELDS_FLAGS_SUM);
172 return lprocfs_rd_u64(page, start, off, count, eof, &locks);
175 static int lprocfs_rd_lru_size(char *page, char **start, off_t off,
176 int count, int *eof, void *data)
178 struct ldlm_namespace *ns = data;
179 __u32 *nr = &ns->ns_max_unused;
181 if (ns_connect_lru_resize(ns))
182 nr = &ns->ns_nr_unused;
183 return lprocfs_rd_uint(page, start, off, count, eof, nr);
186 static int lprocfs_wr_lru_size(struct file *file, const char *buffer,
187 unsigned long count, void *data)
189 struct ldlm_namespace *ns = data;
190 char dummy[MAX_STRING_SIZE + 1], *end;
194 dummy[MAX_STRING_SIZE] = '\0';
195 if (cfs_copy_from_user(dummy, buffer, MAX_STRING_SIZE))
198 if (strncmp(dummy, "clear", 5) == 0) {
200 "dropping all unused locks from namespace %s\n",
202 if (ns_connect_lru_resize(ns)) {
203 int canceled, unused = ns->ns_nr_unused;
205 /* Try to cancel all @ns_nr_unused locks. */
206 canceled = ldlm_cancel_lru(ns, unused, LDLM_SYNC,
208 if (canceled < unused) {
210 "not all requested locks are canceled, "
211 "requested: %d, canceled: %d\n", unused,
216 tmp = ns->ns_max_unused;
217 ns->ns_max_unused = 0;
218 ldlm_cancel_lru(ns, 0, LDLM_SYNC, LDLM_CANCEL_PASSED);
219 ns->ns_max_unused = tmp;
224 tmp = simple_strtoul(dummy, &end, 0);
226 CERROR("invalid value written\n");
229 lru_resize = (tmp == 0);
231 if (ns_connect_lru_resize(ns)) {
233 ns->ns_max_unused = (unsigned int)tmp;
235 if (tmp > ns->ns_nr_unused)
236 tmp = ns->ns_nr_unused;
237 tmp = ns->ns_nr_unused - tmp;
240 "changing namespace %s unused locks from %u to %u\n",
241 ldlm_ns_name(ns), ns->ns_nr_unused,
243 ldlm_cancel_lru(ns, tmp, LDLM_ASYNC, LDLM_CANCEL_PASSED);
247 "disable lru_resize for namespace %s\n",
249 ns->ns_connect_flags &= ~OBD_CONNECT_LRU_RESIZE;
253 "changing namespace %s max_unused from %u to %u\n",
254 ldlm_ns_name(ns), ns->ns_max_unused,
256 ns->ns_max_unused = (unsigned int)tmp;
257 ldlm_cancel_lru(ns, 0, LDLM_ASYNC, LDLM_CANCEL_PASSED);
259 /* Make sure that originally lru resize was supported before
260 * turning it on here. */
262 (ns->ns_orig_connect_flags & OBD_CONNECT_LRU_RESIZE)) {
264 "enable lru_resize for namespace %s\n",
266 ns->ns_connect_flags |= OBD_CONNECT_LRU_RESIZE;
273 void ldlm_namespace_proc_unregister(struct ldlm_namespace *ns)
275 struct proc_dir_entry *dir;
277 dir = lprocfs_srch(ldlm_ns_proc_dir, ldlm_ns_name(ns));
279 CERROR("dlm namespace %s has no procfs dir?\n",
282 lprocfs_remove(&dir);
285 if (ns->ns_stats != NULL)
286 lprocfs_free_stats(&ns->ns_stats);
289 int ldlm_namespace_proc_register(struct ldlm_namespace *ns)
291 struct lprocfs_vars lock_vars[2];
292 char lock_name[MAX_STRING_SIZE + 1];
295 LASSERT(ns->ns_rs_hash != NULL);
297 ns->ns_stats = lprocfs_alloc_stats(LDLM_NSS_LAST, 0);
298 if (ns->ns_stats == NULL)
301 lprocfs_counter_init(ns->ns_stats, LDLM_NSS_LOCKS,
302 LPROCFS_CNTR_AVGMINMAX, "locks", "locks");
304 lock_name[MAX_STRING_SIZE] = '\0';
306 memset(lock_vars, 0, sizeof(lock_vars));
307 lock_vars[0].name = lock_name;
309 snprintf(lock_name, MAX_STRING_SIZE, "%s/resource_count",
311 lock_vars[0].data = ns;
312 lock_vars[0].read_fptr = lprocfs_rd_ns_resources;
313 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
315 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_count",
317 lock_vars[0].data = ns;
318 lock_vars[0].read_fptr = lprocfs_rd_ns_locks;
319 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
321 if (ns_is_client(ns)) {
322 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_unused_count",
324 lock_vars[0].data = &ns->ns_nr_unused;
325 lock_vars[0].read_fptr = lprocfs_rd_uint;
326 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
328 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_size",
330 lock_vars[0].data = ns;
331 lock_vars[0].read_fptr = lprocfs_rd_lru_size;
332 lock_vars[0].write_fptr = lprocfs_wr_lru_size;
333 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
335 snprintf(lock_name, MAX_STRING_SIZE, "%s/lru_max_age",
337 lock_vars[0].data = &ns->ns_max_age;
338 lock_vars[0].read_fptr = lprocfs_rd_uint;
339 lock_vars[0].write_fptr = lprocfs_wr_uint;
340 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
342 snprintf(lock_name, MAX_STRING_SIZE, "%s/ctime_age_limit",
344 lock_vars[0].data = &ns->ns_ctime_age_limit;
345 lock_vars[0].read_fptr = lprocfs_rd_uint;
346 lock_vars[0].write_fptr = lprocfs_wr_uint;
347 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
349 snprintf(lock_name, MAX_STRING_SIZE, "%s/lock_timeouts",
351 lock_vars[0].data = &ns->ns_timeouts;
352 lock_vars[0].read_fptr = lprocfs_rd_uint;
353 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
355 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_nolock_bytes",
357 lock_vars[0].data = &ns->ns_max_nolock_size;
358 lock_vars[0].read_fptr = lprocfs_rd_uint;
359 lock_vars[0].write_fptr = lprocfs_wr_uint;
360 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
362 snprintf(lock_name, MAX_STRING_SIZE, "%s/contention_seconds",
364 lock_vars[0].data = &ns->ns_contention_time;
365 lock_vars[0].read_fptr = lprocfs_rd_uint;
366 lock_vars[0].write_fptr = lprocfs_wr_uint;
367 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
369 snprintf(lock_name, MAX_STRING_SIZE, "%s/contended_locks",
371 lock_vars[0].data = &ns->ns_contended_locks;
372 lock_vars[0].read_fptr = lprocfs_rd_uint;
373 lock_vars[0].write_fptr = lprocfs_wr_uint;
374 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
376 snprintf(lock_name, MAX_STRING_SIZE, "%s/max_parallel_ast",
378 lock_vars[0].data = &ns->ns_max_parallel_ast;
379 lock_vars[0].read_fptr = lprocfs_rd_uint;
380 lock_vars[0].write_fptr = lprocfs_wr_uint;
381 lprocfs_add_vars(ldlm_ns_proc_dir, lock_vars, 0);
385 #undef MAX_STRING_SIZE
388 #define ldlm_namespace_proc_unregister(ns) ({;})
389 #define ldlm_namespace_proc_register(ns) ({0;})
393 static unsigned ldlm_res_hop_hash(cfs_hash_t *hs,
394 const void *key, unsigned mask)
396 const struct ldlm_res_id *id = key;
400 for (i = 0; i < RES_NAME_SIZE; i++)
405 static unsigned ldlm_res_hop_fid_hash(cfs_hash_t *hs,
406 const void *key, unsigned mask)
408 const struct ldlm_res_id *id = key;
413 fid.f_seq = id->name[LUSTRE_RES_ID_SEQ_OFF];
414 fid.f_oid = (__u32)id->name[LUSTRE_RES_ID_VER_OID_OFF];
415 fid.f_ver = (__u32)(id->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32);
417 hash = fid_flatten32(&fid);
418 hash += (hash >> 4) + (hash << 12); /* mixing oid and seq */
419 if (id->name[LUSTRE_RES_ID_HSH_OFF] != 0) {
420 val = id->name[LUSTRE_RES_ID_HSH_OFF];
421 hash += (val >> 5) + (val << 11);
425 hash = cfs_hash_long(hash, hs->hs_bkt_bits);
426 /* give me another random factor */
427 hash -= cfs_hash_long((unsigned long)hs, val % 11 + 3);
429 hash <<= hs->hs_cur_bits - hs->hs_bkt_bits;
430 hash |= ldlm_res_hop_hash(hs, key, CFS_HASH_NBKT(hs) - 1);
435 static void *ldlm_res_hop_key(cfs_hlist_node_t *hnode)
437 struct ldlm_resource *res;
439 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
440 return &res->lr_name;
443 static int ldlm_res_hop_keycmp(const void *key, cfs_hlist_node_t *hnode)
445 struct ldlm_resource *res;
447 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
448 return ldlm_res_eq((const struct ldlm_res_id *)key,
449 (const struct ldlm_res_id *)&res->lr_name);
452 static void *ldlm_res_hop_object(cfs_hlist_node_t *hnode)
454 return cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
457 static void ldlm_res_hop_get_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
459 struct ldlm_resource *res;
461 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
462 ldlm_resource_getref(res);
463 LDLM_RESOURCE_ADDREF(res);
466 static void ldlm_res_hop_put_locked(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
468 struct ldlm_resource *res;
470 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
471 /* cfs_hash_for_each_nolock is the only chance we call it */
472 LDLM_RESOURCE_DELREF(res);
473 ldlm_resource_putref_locked(res);
476 static void ldlm_res_hop_put(cfs_hash_t *hs, cfs_hlist_node_t *hnode)
478 struct ldlm_resource *res;
480 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
481 LDLM_RESOURCE_DELREF(res);
482 ldlm_resource_putref(res);
485 cfs_hash_ops_t ldlm_ns_hash_ops = {
486 .hs_hash = ldlm_res_hop_hash,
487 .hs_key = ldlm_res_hop_key,
488 .hs_keycmp = ldlm_res_hop_keycmp,
490 .hs_object = ldlm_res_hop_object,
491 .hs_get = ldlm_res_hop_get_locked,
492 .hs_put_locked = ldlm_res_hop_put_locked,
493 .hs_put = ldlm_res_hop_put
496 cfs_hash_ops_t ldlm_ns_fid_hash_ops = {
497 .hs_hash = ldlm_res_hop_fid_hash,
498 .hs_key = ldlm_res_hop_key,
499 .hs_keycmp = ldlm_res_hop_keycmp,
501 .hs_object = ldlm_res_hop_object,
502 .hs_get = ldlm_res_hop_get_locked,
503 .hs_put_locked = ldlm_res_hop_put_locked,
504 .hs_put = ldlm_res_hop_put
508 ldlm_ns_type_t nsd_type;
509 /** hash bucket bits */
510 unsigned nsd_bkt_bits;
512 unsigned nsd_all_bits;
513 /** hash operations */
514 cfs_hash_ops_t *nsd_hops;
515 } ldlm_ns_hash_def_t;
517 ldlm_ns_hash_def_t ldlm_ns_hash_defs[] =
520 .nsd_type = LDLM_NS_TYPE_MDC,
523 .nsd_hops = &ldlm_ns_fid_hash_ops,
526 .nsd_type = LDLM_NS_TYPE_MDT,
529 .nsd_hops = &ldlm_ns_fid_hash_ops,
532 .nsd_type = LDLM_NS_TYPE_OSC,
535 .nsd_hops = &ldlm_ns_hash_ops,
538 .nsd_type = LDLM_NS_TYPE_OST,
541 .nsd_hops = &ldlm_ns_hash_ops,
544 .nsd_type = LDLM_NS_TYPE_MGC,
547 .nsd_hops = &ldlm_ns_hash_ops,
550 .nsd_type = LDLM_NS_TYPE_MGT,
553 .nsd_hops = &ldlm_ns_hash_ops,
556 .nsd_type = LDLM_NS_TYPE_UNKNOWN,
560 struct ldlm_namespace *ldlm_namespace_new(struct obd_device *obd, char *name,
563 ldlm_ns_type_t ns_type)
565 struct ldlm_namespace *ns = NULL;
566 struct ldlm_ns_bucket *nsb;
567 ldlm_ns_hash_def_t *nsd;
573 LASSERT(obd != NULL);
577 CERROR("ldlm_get_ref failed: %d\n", rc);
581 for (idx = 0;;idx++) {
582 nsd = &ldlm_ns_hash_defs[idx];
583 if (nsd->nsd_type == LDLM_NS_TYPE_UNKNOWN) {
584 CERROR("Unknown type %d for ns %s\n", ns_type, name);
588 if (nsd->nsd_type == ns_type)
596 ns->ns_rs_hash = cfs_hash_create(name,
597 nsd->nsd_all_bits, nsd->nsd_all_bits,
598 nsd->nsd_bkt_bits, sizeof(*nsb),
604 CFS_HASH_SPIN_BKTLOCK |
605 CFS_HASH_NO_ITEMREF);
606 if (ns->ns_rs_hash == NULL)
609 cfs_hash_for_each_bucket(ns->ns_rs_hash, &bd, idx) {
610 nsb = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
611 at_init(&nsb->nsb_at_estimate, ldlm_enqueue_min, 0);
612 nsb->nsb_namespace = ns;
616 ns->ns_appetite = apt;
617 ns->ns_client = client;
619 CFS_INIT_LIST_HEAD(&ns->ns_list_chain);
620 CFS_INIT_LIST_HEAD(&ns->ns_unused_list);
621 cfs_spin_lock_init(&ns->ns_lock);
622 cfs_atomic_set(&ns->ns_bref, 0);
623 cfs_waitq_init(&ns->ns_waitq);
625 ns->ns_max_nolock_size = NS_DEFAULT_MAX_NOLOCK_BYTES;
626 ns->ns_contention_time = NS_DEFAULT_CONTENTION_SECONDS;
627 ns->ns_contended_locks = NS_DEFAULT_CONTENDED_LOCKS;
629 ns->ns_max_parallel_ast = LDLM_DEFAULT_PARALLEL_AST_LIMIT;
630 ns->ns_nr_unused = 0;
631 ns->ns_max_unused = LDLM_DEFAULT_LRU_SIZE;
632 ns->ns_max_age = LDLM_DEFAULT_MAX_ALIVE;
633 ns->ns_ctime_age_limit = LDLM_CTIME_AGE_LIMIT;
635 ns->ns_orig_connect_flags = 0;
636 ns->ns_connect_flags = 0;
638 rc = ldlm_namespace_proc_register(ns);
640 CERROR("Can't initialize ns proc, rc %d\n", rc);
644 idx = cfs_atomic_read(ldlm_namespace_nr(client));
645 rc = ldlm_pool_init(&ns->ns_pool, ns, idx, client);
647 CERROR("Can't initialize lock pool, rc %d\n", rc);
651 ldlm_namespace_register(ns, client);
654 ldlm_namespace_proc_unregister(ns);
655 ldlm_namespace_cleanup(ns, 0);
657 cfs_hash_putref(ns->ns_rs_hash);
665 extern struct ldlm_lock *ldlm_lock_get(struct ldlm_lock *lock);
667 /* If flags contains FL_LOCAL_ONLY, don't try to tell the server, just cleanup.
668 * This is currently only used for recovery, and we make certain assumptions
669 * as a result--notably, that we shouldn't cancel locks with refs. -phil */
670 static void cleanup_resource(struct ldlm_resource *res, cfs_list_t *q,
674 int rc = 0, client = ns_is_client(ldlm_res_to_ns(res));
675 int local_only = (flags & LDLM_FL_LOCAL_ONLY);
678 struct ldlm_lock *lock = NULL;
680 /* first, we look for non-cleaned-yet lock
681 * all cleaned locks are marked by CLEANED flag */
683 cfs_list_for_each(tmp, q) {
684 lock = cfs_list_entry(tmp, struct ldlm_lock,
686 if (lock->l_flags & LDLM_FL_CLEANED) {
691 lock->l_flags |= LDLM_FL_CLEANED;
700 /* Set CBPENDING so nothing in the cancellation path
701 * can match this lock */
702 lock->l_flags |= LDLM_FL_CBPENDING;
703 lock->l_flags |= LDLM_FL_FAILED;
704 lock->l_flags |= flags;
706 /* ... without sending a CANCEL message for local_only. */
708 lock->l_flags |= LDLM_FL_LOCAL_ONLY;
710 if (local_only && (lock->l_readers || lock->l_writers)) {
711 /* This is a little bit gross, but much better than the
712 * alternative: pretend that we got a blocking AST from
713 * the server, so that when the lock is decref'd, it
714 * will go away ... */
716 LDLM_DEBUG(lock, "setting FL_LOCAL_ONLY");
717 if (lock->l_completion_ast)
718 lock->l_completion_ast(lock, 0, NULL);
719 LDLM_LOCK_RELEASE(lock);
724 struct lustre_handle lockh;
727 ldlm_lock2handle(lock, &lockh);
728 rc = ldlm_cli_cancel(&lockh);
730 CERROR("ldlm_cli_cancel: %d\n", rc);
732 ldlm_resource_unlink_lock(lock);
734 LDLM_DEBUG(lock, "Freeing a lock still held by a "
736 ldlm_lock_destroy(lock);
738 LDLM_LOCK_RELEASE(lock);
742 static int ldlm_resource_clean(cfs_hash_t *hs, cfs_hash_bd_t *bd,
743 cfs_hlist_node_t *hnode, void *arg)
745 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
746 int flags = (int)(unsigned long)arg;
748 cleanup_resource(res, &res->lr_granted, flags);
749 cleanup_resource(res, &res->lr_converting, flags);
750 cleanup_resource(res, &res->lr_waiting, flags);
755 static int ldlm_resource_complain(cfs_hash_t *hs, cfs_hash_bd_t *bd,
756 cfs_hlist_node_t *hnode, void *arg)
758 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
760 CERROR("Namespace %s resource refcount nonzero "
761 "(%d) after lock cleanup; forcing "
763 ldlm_ns_name(ldlm_res_to_ns(res)),
764 cfs_atomic_read(&res->lr_refcount) - 1);
766 CERROR("Resource: %p ("LPU64"/"LPU64"/"LPU64"/"
767 LPU64") (rc: %d)\n", res,
768 res->lr_name.name[0], res->lr_name.name[1],
769 res->lr_name.name[2], res->lr_name.name[3],
770 cfs_atomic_read(&res->lr_refcount) - 1);
774 int ldlm_namespace_cleanup(struct ldlm_namespace *ns, int flags)
777 CDEBUG(D_INFO, "NULL ns, skipping cleanup\n");
781 cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_clean,
782 (void *)(unsigned long)flags);
783 cfs_hash_for_each_nolock(ns->ns_rs_hash, ldlm_resource_complain, NULL);
787 static int __ldlm_namespace_free(struct ldlm_namespace *ns, int force)
791 /* At shutdown time, don't call the cancellation callback */
792 ldlm_namespace_cleanup(ns, force ? LDLM_FL_LOCAL_ONLY : 0);
794 if (cfs_atomic_read(&ns->ns_bref) > 0) {
795 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
798 "dlm namespace %s free waiting on refcount %d\n",
799 ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref));
802 lwi = LWI_TIMEOUT(obd_timeout * CFS_HZ / 4, NULL, NULL);
804 rc = l_wait_event(ns->ns_waitq,
805 cfs_atomic_read(&ns->ns_bref) == 0, &lwi);
807 /* Forced cleanups should be able to reclaim all references,
808 * so it's safe to wait forever... we can't leak locks... */
809 if (force && rc == -ETIMEDOUT) {
810 LCONSOLE_ERROR("Forced cleanup waiting for %s "
811 "namespace with %d resources in use, "
812 "(rc=%d)\n", ldlm_ns_name(ns),
813 cfs_atomic_read(&ns->ns_bref), rc);
814 GOTO(force_wait, rc);
817 if (cfs_atomic_read(&ns->ns_bref)) {
818 LCONSOLE_ERROR("Cleanup waiting for %s namespace "
819 "with %d resources in use, (rc=%d)\n",
821 cfs_atomic_read(&ns->ns_bref), rc);
822 RETURN(ELDLM_NAMESPACE_EXISTS);
824 CDEBUG(D_DLMTRACE, "dlm namespace %s free done waiting\n",
832 * Performs various cleanups for passed \a ns to make it drop refc and be ready
833 * for freeing. Waits for refc == 0.
835 * The following is done:
836 * (0) Unregister \a ns from its list to make inaccessible for potential users
837 * like pools thread and others;
838 * (1) Clear all locks in \a ns.
840 void ldlm_namespace_free_prior(struct ldlm_namespace *ns,
841 struct obd_import *imp,
851 cfs_spin_lock(&ns->ns_lock);
853 cfs_spin_unlock(&ns->ns_lock);
856 * Can fail with -EINTR when force == 0 in which case try harder.
858 rc = __ldlm_namespace_free(ns, force);
859 if (rc != ELDLM_OK) {
861 ptlrpc_disconnect_import(imp, 0);
862 ptlrpc_invalidate_import(imp);
866 * With all requests dropped and the import inactive
867 * we are gaurenteed all reference will be dropped.
869 rc = __ldlm_namespace_free(ns, 1);
876 * Performs freeing memory structures related to \a ns. This is only done when
877 * ldlm_namespce_free_prior() successfully removed all resources referencing
878 * \a ns and its refc == 0.
880 void ldlm_namespace_free_post(struct ldlm_namespace *ns)
890 * Make sure that nobody can find this ns in its list.
892 ldlm_namespace_unregister(ns, ns->ns_client);
894 * Fini pool _before_ parent proc dir is removed. This is important as
895 * ldlm_pool_fini() removes own proc dir which is child to @dir. Removing
896 * it after @dir may cause oops.
898 ldlm_pool_fini(&ns->ns_pool);
900 ldlm_namespace_proc_unregister(ns);
901 cfs_hash_putref(ns->ns_rs_hash);
903 * Namespace \a ns should be not on list in this time, otherwise this
904 * will cause issues realted to using freed \a ns in pools thread.
906 LASSERT(cfs_list_empty(&ns->ns_list_chain));
913 /* Cleanup the resource, and free namespace.
916 * proc1: destroy import
917 * class_disconnect_export(grab cl_sem) ->
918 * -> ldlm_namespace_free ->
919 * -> lprocfs_remove(grab _lprocfs_lock).
920 * proc2: read proc info
921 * lprocfs_fops_read(grab _lprocfs_lock) ->
922 * -> osc_rd_active, etc(grab cl_sem).
924 * So that I have to split the ldlm_namespace_free into two parts - the first
925 * part ldlm_namespace_free_prior is used to cleanup the resource which is
926 * being used; the 2nd part ldlm_namespace_free_post is used to unregister the
927 * lprocfs entries, and then free memory. It will be called w/o cli->cl_sem
930 void ldlm_namespace_free(struct ldlm_namespace *ns,
931 struct obd_import *imp,
934 ldlm_namespace_free_prior(ns, imp, force);
935 ldlm_namespace_free_post(ns);
938 void ldlm_namespace_get(struct ldlm_namespace *ns)
940 cfs_atomic_inc(&ns->ns_bref);
943 void ldlm_namespace_put(struct ldlm_namespace *ns)
945 if (cfs_atomic_dec_and_lock(&ns->ns_bref, &ns->ns_lock)) {
946 cfs_waitq_signal(&ns->ns_waitq);
947 cfs_spin_unlock(&ns->ns_lock);
951 /* Register @ns in the list of namespaces */
952 void ldlm_namespace_register(struct ldlm_namespace *ns, ldlm_side_t client)
954 cfs_mutex_lock(ldlm_namespace_lock(client));
955 LASSERT(cfs_list_empty(&ns->ns_list_chain));
956 cfs_list_add(&ns->ns_list_chain, ldlm_namespace_list(client));
957 cfs_atomic_inc(ldlm_namespace_nr(client));
958 cfs_mutex_unlock(ldlm_namespace_lock(client));
961 /* Unregister @ns from the list of namespaces */
962 void ldlm_namespace_unregister(struct ldlm_namespace *ns, ldlm_side_t client)
964 cfs_mutex_lock(ldlm_namespace_lock(client));
965 LASSERT(!cfs_list_empty(&ns->ns_list_chain));
967 * Some asserts and possibly other parts of code still using
968 * list_empty(&ns->ns_list_chain). This is why it is important
969 * to use list_del_init() here.
971 cfs_list_del_init(&ns->ns_list_chain);
972 cfs_atomic_dec(ldlm_namespace_nr(client));
973 cfs_mutex_unlock(ldlm_namespace_lock(client));
976 /* Should be called under ldlm_namespace_lock(client) taken */
977 void ldlm_namespace_move_locked(struct ldlm_namespace *ns, ldlm_side_t client)
979 LASSERT(!cfs_list_empty(&ns->ns_list_chain));
980 LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client));
981 cfs_list_move_tail(&ns->ns_list_chain, ldlm_namespace_list(client));
984 /* Should be called under ldlm_namespace_lock(client) taken */
985 struct ldlm_namespace *ldlm_namespace_first_locked(ldlm_side_t client)
987 LASSERT_MUTEX_LOCKED(ldlm_namespace_lock(client));
988 LASSERT(!cfs_list_empty(ldlm_namespace_list(client)));
989 return container_of(ldlm_namespace_list(client)->next,
990 struct ldlm_namespace, ns_list_chain);
993 static struct ldlm_resource *ldlm_resource_new(void)
995 struct ldlm_resource *res;
998 OBD_SLAB_ALLOC_PTR_GFP(res, ldlm_resource_slab, CFS_ALLOC_IO);
1002 CFS_INIT_LIST_HEAD(&res->lr_granted);
1003 CFS_INIT_LIST_HEAD(&res->lr_converting);
1004 CFS_INIT_LIST_HEAD(&res->lr_waiting);
1006 /* initialize interval trees for each lock mode*/
1007 for (idx = 0; idx < LCK_MODE_NUM; idx++) {
1008 res->lr_itree[idx].lit_size = 0;
1009 res->lr_itree[idx].lit_mode = 1 << idx;
1010 res->lr_itree[idx].lit_root = NULL;
1013 cfs_atomic_set(&res->lr_refcount, 1);
1014 cfs_spin_lock_init(&res->lr_lock);
1015 lu_ref_init(&res->lr_reference);
1017 /* one who creates the resource must unlock
1018 * the mutex after lvb initialization */
1019 cfs_mutex_init(&res->lr_lvb_mutex);
1020 cfs_mutex_lock(&res->lr_lvb_mutex);
1025 /* Args: unlocked namespace
1026 * * Locks: takes and releases NS hash-lock and res->lr_lock
1027 * * Returns: referenced, unlocked ldlm_resource or NULL */
1028 struct ldlm_resource *
1029 ldlm_resource_get(struct ldlm_namespace *ns, struct ldlm_resource *parent,
1030 const struct ldlm_res_id *name, ldlm_type_t type, int create)
1032 cfs_hlist_node_t *hnode;
1033 struct ldlm_resource *res;
1037 LASSERT(ns != NULL);
1038 LASSERT(parent == NULL);
1039 LASSERT(ns->ns_rs_hash != NULL);
1040 LASSERT(name->name[0] != 0);
1042 cfs_hash_bd_get_and_lock(ns->ns_rs_hash, (void *)name, &bd, 0);
1043 hnode = cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1044 if (hnode != NULL) {
1045 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1046 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1047 /* synchronize WRT resource creation */
1048 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1049 cfs_mutex_lock(&res->lr_lvb_mutex);
1050 cfs_mutex_unlock(&res->lr_lvb_mutex);
1055 version = cfs_hash_bd_version_get(&bd);
1056 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 0);
1061 LASSERTF(type >= LDLM_MIN_TYPE && type < LDLM_MAX_TYPE,
1062 "type: %d\n", type);
1063 res = ldlm_resource_new();
1067 res->lr_ns_bucket = cfs_hash_bd_extra_get(ns->ns_rs_hash, &bd);
1068 res->lr_name = *name;
1069 res->lr_type = type;
1070 res->lr_most_restr = LCK_NL;
1072 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1073 hnode = (version == cfs_hash_bd_version_get(&bd)) ? NULL :
1074 cfs_hash_bd_lookup_locked(ns->ns_rs_hash, &bd, (void *)name);
1076 if (hnode != NULL) {
1077 /* someone won the race and added the resource before */
1078 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1079 /* clean lu_ref for failed resource */
1080 lu_ref_fini(&res->lr_reference);
1081 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1083 res = cfs_hlist_entry(hnode, struct ldlm_resource, lr_hash);
1084 /* synchronize WRT resource creation */
1085 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1086 cfs_mutex_lock(&res->lr_lvb_mutex);
1087 cfs_mutex_unlock(&res->lr_lvb_mutex);
1091 /* we won! let's add the resource */
1092 cfs_hash_bd_add_locked(ns->ns_rs_hash, &bd, &res->lr_hash);
1093 if (cfs_hash_bd_count_get(&bd) == 1)
1094 ldlm_namespace_get(ns);
1096 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1097 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_init) {
1100 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_CREATE_RESOURCE, 2);
1101 rc = ns->ns_lvbo->lvbo_init(res);
1103 CERROR("lvbo_init failed for resource "
1104 LPU64": rc %d\n", name->name[0], rc);
1105 /* we create resource with locked lr_lvb_mutex */
1106 cfs_mutex_unlock(&res->lr_lvb_mutex);
1112 struct ldlm_resource *ldlm_resource_getref(struct ldlm_resource *res)
1114 LASSERT(res != NULL);
1115 LASSERT(res != LP_POISON);
1116 cfs_atomic_inc(&res->lr_refcount);
1117 CDEBUG(D_INFO, "getref res: %p count: %d\n", res,
1118 cfs_atomic_read(&res->lr_refcount));
1122 static void __ldlm_resource_putref_final(cfs_hash_bd_t *bd,
1123 struct ldlm_resource *res)
1125 struct ldlm_ns_bucket *nsb = res->lr_ns_bucket;
1127 if (!cfs_list_empty(&res->lr_granted)) {
1128 ldlm_resource_dump(D_ERROR, res);
1132 if (!cfs_list_empty(&res->lr_converting)) {
1133 ldlm_resource_dump(D_ERROR, res);
1137 if (!cfs_list_empty(&res->lr_waiting)) {
1138 ldlm_resource_dump(D_ERROR, res);
1142 cfs_hash_bd_del_locked(nsb->nsb_namespace->ns_rs_hash,
1144 lu_ref_fini(&res->lr_reference);
1145 if (cfs_hash_bd_count_get(bd) == 0)
1146 ldlm_namespace_put(nsb->nsb_namespace);
1149 /* Returns 1 if the resource was freed, 0 if it remains. */
1150 int ldlm_resource_putref(struct ldlm_resource *res)
1152 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1155 LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1156 CDEBUG(D_INFO, "putref res: %p count: %d\n",
1157 res, cfs_atomic_read(&res->lr_refcount) - 1);
1159 cfs_hash_bd_get(ns->ns_rs_hash, &res->lr_name, &bd);
1160 if (cfs_hash_bd_dec_and_lock(ns->ns_rs_hash, &bd, &res->lr_refcount)) {
1161 __ldlm_resource_putref_final(&bd, res);
1162 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1163 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1164 ns->ns_lvbo->lvbo_free(res);
1165 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1171 /* Returns 1 if the resource was freed, 0 if it remains. */
1172 int ldlm_resource_putref_locked(struct ldlm_resource *res)
1174 struct ldlm_namespace *ns = ldlm_res_to_ns(res);
1176 LASSERT_ATOMIC_GT_LT(&res->lr_refcount, 0, LI_POISON);
1177 CDEBUG(D_INFO, "putref res: %p count: %d\n",
1178 res, cfs_atomic_read(&res->lr_refcount) - 1);
1180 if (cfs_atomic_dec_and_test(&res->lr_refcount)) {
1183 cfs_hash_bd_get(ldlm_res_to_ns(res)->ns_rs_hash,
1184 &res->lr_name, &bd);
1185 __ldlm_resource_putref_final(&bd, res);
1186 cfs_hash_bd_unlock(ns->ns_rs_hash, &bd, 1);
1187 /* NB: ns_rs_hash is created with CFS_HASH_NO_ITEMREF,
1188 * so we should never be here while calling cfs_hash_del,
1189 * cfs_hash_for_each_nolock is the only case we can get
1190 * here, which is safe to release cfs_hash_bd_lock.
1192 if (ns->ns_lvbo && ns->ns_lvbo->lvbo_free)
1193 ns->ns_lvbo->lvbo_free(res);
1194 OBD_SLAB_FREE(res, ldlm_resource_slab, sizeof *res);
1196 cfs_hash_bd_lock(ns->ns_rs_hash, &bd, 1);
1202 void ldlm_resource_add_lock(struct ldlm_resource *res, cfs_list_t *head,
1203 struct ldlm_lock *lock)
1205 check_res_locked(res);
1207 LDLM_DEBUG(lock, "About to add this lock:\n");
1209 if (lock->l_destroyed) {
1210 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1214 LASSERT(cfs_list_empty(&lock->l_res_link));
1216 cfs_list_add_tail(&lock->l_res_link, head);
1219 void ldlm_resource_insert_lock_after(struct ldlm_lock *original,
1220 struct ldlm_lock *new)
1222 struct ldlm_resource *res = original->l_resource;
1224 check_res_locked(res);
1226 ldlm_resource_dump(D_INFO, res);
1227 LDLM_DEBUG(new, "About to insert this lock after %p:\n", original);
1229 if (new->l_destroyed) {
1230 CDEBUG(D_OTHER, "Lock destroyed, not adding to resource\n");
1234 LASSERT(cfs_list_empty(&new->l_res_link));
1236 cfs_list_add(&new->l_res_link, &original->l_res_link);
1240 void ldlm_resource_unlink_lock(struct ldlm_lock *lock)
1242 int type = lock->l_resource->lr_type;
1244 check_res_locked(lock->l_resource);
1245 if (type == LDLM_IBITS || type == LDLM_PLAIN)
1246 ldlm_unlink_lock_skiplist(lock);
1247 else if (type == LDLM_EXTENT)
1248 ldlm_extent_unlink_lock(lock);
1249 cfs_list_del_init(&lock->l_res_link);
1252 void ldlm_res2desc(struct ldlm_resource *res, struct ldlm_resource_desc *desc)
1254 desc->lr_type = res->lr_type;
1255 desc->lr_name = res->lr_name;
1258 void ldlm_dump_all_namespaces(ldlm_side_t client, int level)
1262 if (!((libcfs_debug | D_ERROR) & level))
1265 cfs_mutex_lock(ldlm_namespace_lock(client));
1267 cfs_list_for_each(tmp, ldlm_namespace_list(client)) {
1268 struct ldlm_namespace *ns;
1269 ns = cfs_list_entry(tmp, struct ldlm_namespace, ns_list_chain);
1270 ldlm_namespace_dump(level, ns);
1273 cfs_mutex_unlock(ldlm_namespace_lock(client));
1276 static int ldlm_res_hash_dump(cfs_hash_t *hs, cfs_hash_bd_t *bd,
1277 cfs_hlist_node_t *hnode, void *arg)
1279 struct ldlm_resource *res = cfs_hash_object(hs, hnode);
1280 int level = (int)(unsigned long)arg;
1283 ldlm_resource_dump(level, res);
1289 void ldlm_namespace_dump(int level, struct ldlm_namespace *ns)
1291 if (!((libcfs_debug | D_ERROR) & level))
1294 CDEBUG(level, "--- Namespace: %s (rc: %d, side: %s)\n",
1295 ldlm_ns_name(ns), cfs_atomic_read(&ns->ns_bref),
1296 ns_is_client(ns) ? "client" : "server");
1298 if (cfs_time_before(cfs_time_current(), ns->ns_next_dump))
1301 cfs_hash_for_each_nolock(ns->ns_rs_hash,
1303 (void *)(unsigned long)level);
1304 cfs_spin_lock(&ns->ns_lock);
1305 ns->ns_next_dump = cfs_time_shift(10);
1306 cfs_spin_unlock(&ns->ns_lock);
1309 void ldlm_resource_dump(int level, struct ldlm_resource *res)
1311 struct ldlm_lock *lock;
1312 unsigned int granted = 0;
1314 CLASSERT(RES_NAME_SIZE == 4);
1316 if (!((libcfs_debug | D_ERROR) & level))
1319 CDEBUG(level, "--- Resource: %p ("LPU64"/"LPU64"/"LPU64"/"LPU64
1320 ") (rc: %d)\n", res, res->lr_name.name[0], res->lr_name.name[1],
1321 res->lr_name.name[2], res->lr_name.name[3],
1322 cfs_atomic_read(&res->lr_refcount));
1324 if (!cfs_list_empty(&res->lr_granted)) {
1325 CDEBUG(level, "Granted locks (in reverse order):\n");
1326 cfs_list_for_each_entry_reverse(lock, &res->lr_granted,
1328 LDLM_DEBUG_LIMIT(level, lock, "###");
1329 if (!(level & D_CANTMASK) &&
1330 ++granted > ldlm_dump_granted_max) {
1331 CDEBUG(level, "only dump %d granted locks to "
1332 "avoid DDOS.\n", granted);
1337 if (!cfs_list_empty(&res->lr_converting)) {
1338 CDEBUG(level, "Converting locks:\n");
1339 cfs_list_for_each_entry(lock, &res->lr_converting, l_res_link)
1340 LDLM_DEBUG_LIMIT(level, lock, "###");
1342 if (!cfs_list_empty(&res->lr_waiting)) {
1343 CDEBUG(level, "Waiting locks:\n");
1344 cfs_list_for_each_entry(lock, &res->lr_waiting, l_res_link)
1345 LDLM_DEBUG_LIMIT(level, lock, "###");