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) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
39 * Author: Jinshan Xiong <jinshan.xiong@intel.com>
42 #define DEBUG_SUBSYSTEM S_CLASS
44 #include <libcfs/libcfs.h>
45 #include <obd_class.h>
46 #include <obd_support.h>
47 #include <lustre_fid.h>
48 #include <libcfs/list.h>
49 #include <cl_object.h>
50 #include "cl_internal.h"
52 /** Lock class of cl_lock::cll_guard */
53 static struct lock_class_key cl_lock_guard_class;
54 static struct kmem_cache *cl_lock_kmem;
56 static struct lu_kmem_descr cl_lock_caches[] = {
58 .ckd_cache = &cl_lock_kmem,
59 .ckd_name = "cl_lock_kmem",
60 .ckd_size = sizeof (struct cl_lock)
67 #ifdef CONFIG_DEBUG_PAGESTATE_TRACKING
68 #define CS_LOCK_INC(o, item) \
69 atomic_inc(&cl_object_site(o)->cs_locks.cs_stats[CS_##item])
70 #define CS_LOCK_DEC(o, item) \
71 atomic_dec(&cl_object_site(o)->cs_locks.cs_stats[CS_##item])
72 #define CS_LOCKSTATE_INC(o, state) \
73 atomic_inc(&cl_object_site(o)->cs_locks_state[state])
74 #define CS_LOCKSTATE_DEC(o, state) \
75 atomic_dec(&cl_object_site(o)->cs_locks_state[state])
77 #define CS_LOCK_INC(o, item)
78 #define CS_LOCK_DEC(o, item)
79 #define CS_LOCKSTATE_INC(o, state)
80 #define CS_LOCKSTATE_DEC(o, state)
84 * Basic lock invariant that is maintained at all times. Caller either has a
85 * reference to \a lock, or somehow assures that \a lock cannot be freed.
87 * \see cl_lock_invariant()
89 static int cl_lock_invariant_trusted(const struct lu_env *env,
90 const struct cl_lock *lock)
92 return ergo(lock->cll_state == CLS_FREEING, lock->cll_holds == 0) &&
93 atomic_read(&lock->cll_ref) >= lock->cll_holds &&
94 lock->cll_holds >= lock->cll_users &&
95 lock->cll_holds >= 0 &&
96 lock->cll_users >= 0 &&
101 * Stronger lock invariant, checking that caller has a reference on a lock.
103 * \see cl_lock_invariant_trusted()
105 static int cl_lock_invariant(const struct lu_env *env,
106 const struct cl_lock *lock)
110 result = atomic_read(&lock->cll_ref) > 0 &&
111 cl_lock_invariant_trusted(env, lock);
112 if (!result && env != NULL)
113 CL_LOCK_DEBUG(D_ERROR, env, lock, "invariant broken\n");
118 * Returns lock "nesting": 0 for a top-lock and 1 for a sub-lock.
120 static enum clt_nesting_level cl_lock_nesting(const struct cl_lock *lock)
122 return cl_object_header(lock->cll_descr.cld_obj)->coh_nesting;
126 * Returns a set of counters for this lock, depending on a lock nesting.
128 static struct cl_thread_counters *cl_lock_counters(const struct lu_env *env,
129 const struct cl_lock *lock)
131 struct cl_thread_info *info;
132 enum clt_nesting_level nesting;
134 info = cl_env_info(env);
135 nesting = cl_lock_nesting(lock);
136 LASSERT(nesting < ARRAY_SIZE(info->clt_counters));
137 return &info->clt_counters[nesting];
140 static void cl_lock_trace0(int level, const struct lu_env *env,
141 const char *prefix, const struct cl_lock *lock,
142 const char *func, const int line)
144 struct cl_object_header *h = cl_object_header(lock->cll_descr.cld_obj);
145 CDEBUG(level, "%s: %p@(%d %p %d %d %d %d %d %lx)"
146 "(%p/%d/%d) at %s():%d\n",
147 prefix, lock, atomic_read(&lock->cll_ref),
148 lock->cll_guarder, lock->cll_depth,
149 lock->cll_state, lock->cll_error, lock->cll_holds,
150 lock->cll_users, lock->cll_flags,
151 env, h->coh_nesting, cl_lock_nr_mutexed(env),
154 #define cl_lock_trace(level, env, prefix, lock) \
155 cl_lock_trace0(level, env, prefix, lock, __FUNCTION__, __LINE__)
157 #define RETIP ((unsigned long)__builtin_return_address(0))
159 #ifdef CONFIG_LOCKDEP
160 static struct lock_class_key cl_lock_key;
162 static void cl_lock_lockdep_init(struct cl_lock *lock)
164 lockdep_set_class_and_name(lock, &cl_lock_key, "EXT");
167 static void cl_lock_lockdep_acquire(const struct lu_env *env,
168 struct cl_lock *lock, __u32 enqflags)
170 cl_lock_counters(env, lock)->ctc_nr_locks_acquired++;
171 lock_map_acquire(&lock->dep_map);
174 static void cl_lock_lockdep_release(const struct lu_env *env,
175 struct cl_lock *lock)
177 cl_lock_counters(env, lock)->ctc_nr_locks_acquired--;
178 lock_map_release(&lock->dep_map);
181 #else /* !CONFIG_LOCKDEP */
183 static void cl_lock_lockdep_init(struct cl_lock *lock)
185 static void cl_lock_lockdep_acquire(const struct lu_env *env,
186 struct cl_lock *lock, __u32 enqflags)
188 static void cl_lock_lockdep_release(const struct lu_env *env,
189 struct cl_lock *lock)
192 #endif /* !CONFIG_LOCKDEP */
195 * Adds lock slice to the compound lock.
197 * This is called by cl_object_operations::coo_lock_init() methods to add a
198 * per-layer state to the lock. New state is added at the end of
199 * cl_lock::cll_layers list, that is, it is at the bottom of the stack.
201 * \see cl_req_slice_add(), cl_page_slice_add(), cl_io_slice_add()
203 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
204 struct cl_object *obj,
205 const struct cl_lock_operations *ops)
208 slice->cls_lock = lock;
209 list_add_tail(&slice->cls_linkage, &lock->cll_layers);
210 slice->cls_obj = obj;
211 slice->cls_ops = ops;
214 EXPORT_SYMBOL(cl_lock_slice_add);
217 * Returns true iff a lock with the mode \a has provides at least the same
218 * guarantees as a lock with the mode \a need.
220 int cl_lock_mode_match(enum cl_lock_mode has, enum cl_lock_mode need)
222 LINVRNT(need == CLM_READ || need == CLM_WRITE ||
223 need == CLM_PHANTOM || need == CLM_GROUP);
224 LINVRNT(has == CLM_READ || has == CLM_WRITE ||
225 has == CLM_PHANTOM || has == CLM_GROUP);
226 CLASSERT(CLM_PHANTOM < CLM_READ);
227 CLASSERT(CLM_READ < CLM_WRITE);
228 CLASSERT(CLM_WRITE < CLM_GROUP);
230 if (has != CLM_GROUP)
235 EXPORT_SYMBOL(cl_lock_mode_match);
238 * Returns true iff extent portions of lock descriptions match.
240 int cl_lock_ext_match(const struct cl_lock_descr *has,
241 const struct cl_lock_descr *need)
244 has->cld_start <= need->cld_start &&
245 has->cld_end >= need->cld_end &&
246 cl_lock_mode_match(has->cld_mode, need->cld_mode) &&
247 (has->cld_mode != CLM_GROUP || has->cld_gid == need->cld_gid);
249 EXPORT_SYMBOL(cl_lock_ext_match);
252 * Returns true iff a lock with the description \a has provides at least the
253 * same guarantees as a lock with the description \a need.
255 int cl_lock_descr_match(const struct cl_lock_descr *has,
256 const struct cl_lock_descr *need)
259 cl_object_same(has->cld_obj, need->cld_obj) &&
260 cl_lock_ext_match(has, need);
262 EXPORT_SYMBOL(cl_lock_descr_match);
264 static void cl_lock_free(const struct lu_env *env, struct cl_lock *lock)
266 struct cl_object *obj = lock->cll_descr.cld_obj;
268 LINVRNT(!cl_lock_is_mutexed(lock));
271 cl_lock_trace(D_DLMTRACE, env, "free lock", lock);
272 while (!list_empty(&lock->cll_layers)) {
273 struct cl_lock_slice *slice;
275 slice = list_entry(lock->cll_layers.next,
276 struct cl_lock_slice, cls_linkage);
277 list_del_init(lock->cll_layers.next);
278 slice->cls_ops->clo_fini(env, slice);
280 CS_LOCK_DEC(obj, total);
281 CS_LOCKSTATE_DEC(obj, lock->cll_state);
282 lu_object_ref_del_at(&obj->co_lu, &lock->cll_obj_ref, "cl_lock", lock);
283 cl_object_put(env, obj);
284 lu_ref_fini(&lock->cll_reference);
285 lu_ref_fini(&lock->cll_holders);
286 mutex_destroy(&lock->cll_guard);
287 OBD_SLAB_FREE_PTR(lock, cl_lock_kmem);
292 * Releases a reference on a lock.
294 * When last reference is released, lock is returned to the cache, unless it
295 * is in cl_lock_state::CLS_FREEING state, in which case it is destroyed
298 * \see cl_object_put(), cl_page_put()
300 void cl_lock_put(const struct lu_env *env, struct cl_lock *lock)
302 struct cl_object *obj;
304 LINVRNT(cl_lock_invariant(env, lock));
306 obj = lock->cll_descr.cld_obj;
307 LINVRNT(obj != NULL);
309 CDEBUG(D_TRACE, "releasing reference: %d %p %lu\n",
310 atomic_read(&lock->cll_ref), lock, RETIP);
312 if (atomic_dec_and_test(&lock->cll_ref)) {
313 if (lock->cll_state == CLS_FREEING) {
314 LASSERT(list_empty(&lock->cll_linkage));
315 cl_lock_free(env, lock);
317 CS_LOCK_DEC(obj, busy);
321 EXPORT_SYMBOL(cl_lock_put);
324 * Acquires an additional reference to a lock.
326 * This can be called only by caller already possessing a reference to \a
329 * \see cl_object_get(), cl_page_get()
331 void cl_lock_get(struct cl_lock *lock)
333 LINVRNT(cl_lock_invariant(NULL, lock));
334 CDEBUG(D_TRACE, "acquiring reference: %d %p %lu\n",
335 atomic_read(&lock->cll_ref), lock, RETIP);
336 atomic_inc(&lock->cll_ref);
338 EXPORT_SYMBOL(cl_lock_get);
341 * Acquires a reference to a lock.
343 * This is much like cl_lock_get(), except that this function can be used to
344 * acquire initial reference to the cached lock. Caller has to deal with all
345 * possible races. Use with care!
347 * \see cl_page_get_trust()
349 void cl_lock_get_trust(struct cl_lock *lock)
351 CDEBUG(D_TRACE, "acquiring trusted reference: %d %p %lu\n",
352 atomic_read(&lock->cll_ref), lock, RETIP);
353 if (atomic_inc_return(&lock->cll_ref) == 1)
354 CS_LOCK_INC(lock->cll_descr.cld_obj, busy);
356 EXPORT_SYMBOL(cl_lock_get_trust);
359 * Helper function destroying the lock that wasn't completely initialized.
361 * Other threads can acquire references to the top-lock through its
362 * sub-locks. Hence, it cannot be cl_lock_free()-ed immediately.
364 static void cl_lock_finish(const struct lu_env *env, struct cl_lock *lock)
366 cl_lock_mutex_get(env, lock);
367 cl_lock_cancel(env, lock);
368 cl_lock_delete(env, lock);
369 cl_lock_mutex_put(env, lock);
370 cl_lock_put(env, lock);
373 static struct cl_lock *cl_lock_alloc(const struct lu_env *env,
374 struct cl_object *obj,
375 const struct cl_io *io,
376 const struct cl_lock_descr *descr)
378 struct cl_lock *lock;
379 struct lu_object_header *head;
382 OBD_SLAB_ALLOC_PTR_GFP(lock, cl_lock_kmem, GFP_NOFS);
384 atomic_set(&lock->cll_ref, 1);
385 lock->cll_descr = *descr;
386 lock->cll_state = CLS_NEW;
388 lu_object_ref_add_at(&obj->co_lu, &lock->cll_obj_ref, "cl_lock",
390 INIT_LIST_HEAD(&lock->cll_layers);
391 INIT_LIST_HEAD(&lock->cll_linkage);
392 INIT_LIST_HEAD(&lock->cll_inclosure);
393 lu_ref_init(&lock->cll_reference);
394 lu_ref_init(&lock->cll_holders);
395 mutex_init(&lock->cll_guard);
396 lockdep_set_class(&lock->cll_guard, &cl_lock_guard_class);
397 init_waitqueue_head(&lock->cll_wq);
398 head = obj->co_lu.lo_header;
399 CS_LOCKSTATE_INC(obj, CLS_NEW);
400 CS_LOCK_INC(obj, total);
401 CS_LOCK_INC(obj, create);
402 cl_lock_lockdep_init(lock);
403 list_for_each_entry(obj, &head->loh_layers, co_lu.lo_linkage) {
406 err = obj->co_ops->coo_lock_init(env, obj, lock, io);
408 cl_lock_finish(env, lock);
414 lock = ERR_PTR(-ENOMEM);
419 * Transfer the lock into INTRANSIT state and return the original state.
421 * \pre state: CLS_CACHED, CLS_HELD or CLS_ENQUEUED
422 * \post state: CLS_INTRANSIT
425 enum cl_lock_state cl_lock_intransit(const struct lu_env *env,
426 struct cl_lock *lock)
428 enum cl_lock_state state = lock->cll_state;
430 LASSERT(cl_lock_is_mutexed(lock));
431 LASSERT(state != CLS_INTRANSIT);
432 LASSERTF(state >= CLS_ENQUEUED && state <= CLS_CACHED,
433 "Malformed lock state %d.\n", state);
435 cl_lock_state_set(env, lock, CLS_INTRANSIT);
436 lock->cll_intransit_owner = current;
437 cl_lock_hold_add(env, lock, "intransit", current);
440 EXPORT_SYMBOL(cl_lock_intransit);
443 * Exit the intransit state and restore the lock state to the original state
445 void cl_lock_extransit(const struct lu_env *env, struct cl_lock *lock,
446 enum cl_lock_state state)
448 LASSERT(cl_lock_is_mutexed(lock));
449 LASSERT(lock->cll_state == CLS_INTRANSIT);
450 LASSERT(state != CLS_INTRANSIT);
451 LASSERT(lock->cll_intransit_owner == current);
453 lock->cll_intransit_owner = NULL;
454 cl_lock_state_set(env, lock, state);
455 cl_lock_unhold(env, lock, "intransit", current);
457 EXPORT_SYMBOL(cl_lock_extransit);
460 * Checking whether the lock is intransit state
462 int cl_lock_is_intransit(struct cl_lock *lock)
464 LASSERT(cl_lock_is_mutexed(lock));
465 return lock->cll_state == CLS_INTRANSIT &&
466 lock->cll_intransit_owner != current;
468 EXPORT_SYMBOL(cl_lock_is_intransit);
470 * Returns true iff lock is "suitable" for given io. E.g., locks acquired by
471 * truncate and O_APPEND cannot be reused for read/non-append-write, as they
472 * cover multiple stripes and can trigger cascading timeouts.
474 static int cl_lock_fits_into(const struct lu_env *env,
475 const struct cl_lock *lock,
476 const struct cl_lock_descr *need,
477 const struct cl_io *io)
479 const struct cl_lock_slice *slice;
481 LINVRNT(cl_lock_invariant_trusted(env, lock));
483 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
484 if (slice->cls_ops->clo_fits_into != NULL &&
485 !slice->cls_ops->clo_fits_into(env, slice, need, io))
491 static struct cl_lock *cl_lock_lookup(const struct lu_env *env,
492 struct cl_object *obj,
493 const struct cl_io *io,
494 const struct cl_lock_descr *need)
496 struct cl_lock *lock;
497 struct cl_object_header *head;
501 head = cl_object_header(obj);
502 assert_spin_locked(&head->coh_lock_guard);
503 CS_LOCK_INC(obj, lookup);
504 list_for_each_entry(lock, &head->coh_locks, cll_linkage) {
507 matched = cl_lock_ext_match(&lock->cll_descr, need) &&
508 lock->cll_state < CLS_FREEING &&
509 lock->cll_error == 0 &&
510 !(lock->cll_flags & CLF_CANCELLED) &&
511 cl_lock_fits_into(env, lock, need, io);
512 CDEBUG(D_DLMTRACE, "has: "DDESCR"(%d) need: "DDESCR": %d\n",
513 PDESCR(&lock->cll_descr), lock->cll_state, PDESCR(need),
516 cl_lock_get_trust(lock);
517 CS_LOCK_INC(obj, hit);
525 * Returns a lock matching description \a need.
527 * This is the main entry point into the cl_lock caching interface. First, a
528 * cache (implemented as a per-object linked list) is consulted. If lock is
529 * found there, it is returned immediately. Otherwise new lock is allocated
530 * and returned. In any case, additional reference to lock is acquired.
532 * \see cl_object_find(), cl_page_find()
534 static struct cl_lock *cl_lock_find(const struct lu_env *env,
535 const struct cl_io *io,
536 const struct cl_lock_descr *need)
538 struct cl_object_header *head;
539 struct cl_object *obj;
540 struct cl_lock *lock;
545 head = cl_object_header(obj);
547 spin_lock(&head->coh_lock_guard);
548 lock = cl_lock_lookup(env, obj, io, need);
549 spin_unlock(&head->coh_lock_guard);
552 lock = cl_lock_alloc(env, obj, io, need);
554 struct cl_lock *ghost;
556 spin_lock(&head->coh_lock_guard);
557 ghost = cl_lock_lookup(env, obj, io, need);
559 cl_lock_get_trust(lock);
560 list_add_tail(&lock->cll_linkage,
562 spin_unlock(&head->coh_lock_guard);
563 CS_LOCK_INC(obj, busy);
565 spin_unlock(&head->coh_lock_guard);
567 * Other threads can acquire references to the
568 * top-lock through its sub-locks. Hence, it
569 * cannot be cl_lock_free()-ed immediately.
571 cl_lock_finish(env, lock);
580 * Returns existing lock matching given description. This is similar to
581 * cl_lock_find() except that no new lock is created, and returned lock is
582 * guaranteed to be in enum cl_lock_state::CLS_HELD state.
584 struct cl_lock *cl_lock_peek(const struct lu_env *env, const struct cl_io *io,
585 const struct cl_lock_descr *need,
586 const char *scope, const void *source)
588 struct cl_object_header *head;
589 struct cl_object *obj;
590 struct cl_lock *lock;
593 head = cl_object_header(obj);
596 spin_lock(&head->coh_lock_guard);
597 lock = cl_lock_lookup(env, obj, io, need);
598 spin_unlock(&head->coh_lock_guard);
602 cl_lock_mutex_get(env, lock);
603 if (lock->cll_state == CLS_INTRANSIT)
604 /* Don't care return value. */
605 cl_lock_state_wait(env, lock);
606 if (lock->cll_state == CLS_FREEING) {
607 cl_lock_mutex_put(env, lock);
608 cl_lock_put(env, lock);
611 } while (lock == NULL);
613 cl_lock_hold_add(env, lock, scope, source);
614 cl_lock_user_add(env, lock);
615 if (lock->cll_state == CLS_CACHED)
616 cl_use_try(env, lock, 1);
617 if (lock->cll_state == CLS_HELD) {
618 cl_lock_mutex_put(env, lock);
619 cl_lock_lockdep_acquire(env, lock, 0);
620 cl_lock_put(env, lock);
622 cl_unuse_try(env, lock);
623 cl_lock_unhold(env, lock, scope, source);
624 cl_lock_mutex_put(env, lock);
625 cl_lock_put(env, lock);
631 EXPORT_SYMBOL(cl_lock_peek);
634 * Returns a slice within a lock, corresponding to the given layer in the
639 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
640 const struct lu_device_type *dtype)
642 const struct cl_lock_slice *slice;
644 LINVRNT(cl_lock_invariant_trusted(NULL, lock));
647 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
648 if (slice->cls_obj->co_lu.lo_dev->ld_type == dtype)
653 EXPORT_SYMBOL(cl_lock_at);
655 static void cl_lock_mutex_tail(const struct lu_env *env, struct cl_lock *lock)
657 struct cl_thread_counters *counters;
659 counters = cl_lock_counters(env, lock);
661 counters->ctc_nr_locks_locked++;
662 lu_ref_add(&counters->ctc_locks_locked, "cll_guard", lock);
663 cl_lock_trace(D_TRACE, env, "got mutex", lock);
667 * Locks cl_lock object.
669 * This is used to manipulate cl_lock fields, and to serialize state
670 * transitions in the lock state machine.
672 * \post cl_lock_is_mutexed(lock)
674 * \see cl_lock_mutex_put()
676 void cl_lock_mutex_get(const struct lu_env *env, struct cl_lock *lock)
678 LINVRNT(cl_lock_invariant(env, lock));
680 if (lock->cll_guarder == current) {
681 LINVRNT(cl_lock_is_mutexed(lock));
682 LINVRNT(lock->cll_depth > 0);
684 struct cl_object_header *hdr;
685 struct cl_thread_info *info;
688 LINVRNT(lock->cll_guarder != current);
689 hdr = cl_object_header(lock->cll_descr.cld_obj);
691 * Check that mutices are taken in the bottom-to-top order.
693 info = cl_env_info(env);
694 for (i = 0; i < hdr->coh_nesting; ++i)
695 LASSERT(info->clt_counters[i].ctc_nr_locks_locked == 0);
696 mutex_lock_nested(&lock->cll_guard, hdr->coh_nesting);
697 lock->cll_guarder = current;
698 LINVRNT(lock->cll_depth == 0);
700 cl_lock_mutex_tail(env, lock);
702 EXPORT_SYMBOL(cl_lock_mutex_get);
705 * Try-locks cl_lock object.
707 * \retval 0 \a lock was successfully locked
709 * \retval -EBUSY \a lock cannot be locked right now
711 * \post ergo(result == 0, cl_lock_is_mutexed(lock))
713 * \see cl_lock_mutex_get()
715 int cl_lock_mutex_try(const struct lu_env *env, struct cl_lock *lock)
719 LINVRNT(cl_lock_invariant_trusted(env, lock));
723 if (lock->cll_guarder == current) {
724 LINVRNT(lock->cll_depth > 0);
725 cl_lock_mutex_tail(env, lock);
726 } else if (mutex_trylock(&lock->cll_guard)) {
727 LINVRNT(lock->cll_depth == 0);
728 lock->cll_guarder = current;
729 cl_lock_mutex_tail(env, lock);
734 EXPORT_SYMBOL(cl_lock_mutex_try);
737 {* Unlocks cl_lock object.
739 * \pre cl_lock_is_mutexed(lock)
741 * \see cl_lock_mutex_get()
743 void cl_lock_mutex_put(const struct lu_env *env, struct cl_lock *lock)
745 struct cl_thread_counters *counters;
747 LINVRNT(cl_lock_invariant(env, lock));
748 LINVRNT(cl_lock_is_mutexed(lock));
749 LINVRNT(lock->cll_guarder == current);
750 LINVRNT(lock->cll_depth > 0);
752 counters = cl_lock_counters(env, lock);
753 LINVRNT(counters->ctc_nr_locks_locked > 0);
755 cl_lock_trace(D_TRACE, env, "put mutex", lock);
756 lu_ref_del(&counters->ctc_locks_locked, "cll_guard", lock);
757 counters->ctc_nr_locks_locked--;
758 if (--lock->cll_depth == 0) {
759 lock->cll_guarder = NULL;
760 mutex_unlock(&lock->cll_guard);
763 EXPORT_SYMBOL(cl_lock_mutex_put);
766 * Returns true iff lock's mutex is owned by the current thread.
768 int cl_lock_is_mutexed(struct cl_lock *lock)
770 return lock->cll_guarder == current;
772 EXPORT_SYMBOL(cl_lock_is_mutexed);
775 * Returns number of cl_lock mutices held by the current thread (environment).
777 int cl_lock_nr_mutexed(const struct lu_env *env)
779 struct cl_thread_info *info;
784 * NOTE: if summation across all nesting levels (currently 2) proves
785 * too expensive, a summary counter can be added to
786 * struct cl_thread_info.
788 info = cl_env_info(env);
789 for (i = 0, locked = 0; i < ARRAY_SIZE(info->clt_counters); ++i)
790 locked += info->clt_counters[i].ctc_nr_locks_locked;
793 EXPORT_SYMBOL(cl_lock_nr_mutexed);
795 static void cl_lock_cancel0(const struct lu_env *env, struct cl_lock *lock)
797 LINVRNT(cl_lock_is_mutexed(lock));
798 LINVRNT(cl_lock_invariant(env, lock));
800 if (!(lock->cll_flags & CLF_CANCELLED)) {
801 const struct cl_lock_slice *slice;
803 lock->cll_flags |= CLF_CANCELLED;
804 list_for_each_entry_reverse(slice, &lock->cll_layers,
806 if (slice->cls_ops->clo_cancel != NULL)
807 slice->cls_ops->clo_cancel(env, slice);
813 static void cl_lock_delete0(const struct lu_env *env, struct cl_lock *lock)
815 struct cl_object_header *head;
816 const struct cl_lock_slice *slice;
818 LINVRNT(cl_lock_is_mutexed(lock));
819 LINVRNT(cl_lock_invariant(env, lock));
822 if (lock->cll_state < CLS_FREEING) {
825 LASSERT(lock->cll_state != CLS_INTRANSIT);
826 cl_lock_state_set(env, lock, CLS_FREEING);
828 head = cl_object_header(lock->cll_descr.cld_obj);
830 spin_lock(&head->coh_lock_guard);
831 in_cache = !list_empty(&lock->cll_linkage);
833 list_del_init(&lock->cll_linkage);
834 spin_unlock(&head->coh_lock_guard);
836 if (in_cache) /* coh_locks cache holds a refcount. */
837 cl_lock_put(env, lock);
840 * From now on, no new references to this lock can be acquired
841 * by cl_lock_lookup().
843 list_for_each_entry_reverse(slice, &lock->cll_layers,
845 if (slice->cls_ops->clo_delete != NULL)
846 slice->cls_ops->clo_delete(env, slice);
849 * From now on, no new references to this lock can be acquired
850 * by layer-specific means (like a pointer from struct
851 * ldlm_lock in osc, or a pointer from top-lock to sub-lock in
854 * Lock will be finally freed in cl_lock_put() when last of
855 * existing references goes away.
862 * Mod(ifie)s cl_lock::cll_holds counter for a given lock. Also, for a
863 * top-lock (nesting == 0) accounts for this modification in the per-thread
864 * debugging counters. Sub-lock holds can be released by a thread different
865 * from one that acquired it.
867 static void cl_lock_hold_mod(const struct lu_env *env, struct cl_lock *lock,
870 struct cl_thread_counters *counters;
871 enum clt_nesting_level nesting;
873 lock->cll_holds += delta;
874 nesting = cl_lock_nesting(lock);
875 if (nesting == CNL_TOP) {
876 counters = &cl_env_info(env)->clt_counters[CNL_TOP];
877 counters->ctc_nr_held += delta;
878 LASSERT(counters->ctc_nr_held >= 0);
883 * Mod(ifie)s cl_lock::cll_users counter for a given lock. See
884 * cl_lock_hold_mod() for the explanation of the debugging code.
886 static void cl_lock_used_mod(const struct lu_env *env, struct cl_lock *lock,
889 struct cl_thread_counters *counters;
890 enum clt_nesting_level nesting;
892 lock->cll_users += delta;
893 nesting = cl_lock_nesting(lock);
894 if (nesting == CNL_TOP) {
895 counters = &cl_env_info(env)->clt_counters[CNL_TOP];
896 counters->ctc_nr_used += delta;
897 LASSERT(counters->ctc_nr_used >= 0);
901 void cl_lock_hold_release(const struct lu_env *env, struct cl_lock *lock,
902 const char *scope, const void *source)
904 LINVRNT(cl_lock_is_mutexed(lock));
905 LINVRNT(cl_lock_invariant(env, lock));
906 LASSERT(lock->cll_holds > 0);
909 cl_lock_trace(D_DLMTRACE, env, "hold release lock", lock);
910 lu_ref_del(&lock->cll_holders, scope, source);
911 cl_lock_hold_mod(env, lock, -1);
912 if (lock->cll_holds == 0) {
913 CL_LOCK_ASSERT(lock->cll_state != CLS_HELD, env, lock);
914 if (lock->cll_descr.cld_mode == CLM_PHANTOM ||
915 lock->cll_descr.cld_mode == CLM_GROUP ||
916 lock->cll_state != CLS_CACHED)
918 * If lock is still phantom or grouplock when user is
919 * done with it---destroy the lock.
921 lock->cll_flags |= CLF_CANCELPEND|CLF_DOOMED;
922 if (lock->cll_flags & CLF_CANCELPEND) {
923 lock->cll_flags &= ~CLF_CANCELPEND;
924 cl_lock_cancel0(env, lock);
926 if (lock->cll_flags & CLF_DOOMED) {
927 /* no longer doomed: it's dead... Jim. */
928 lock->cll_flags &= ~CLF_DOOMED;
929 cl_lock_delete0(env, lock);
934 EXPORT_SYMBOL(cl_lock_hold_release);
937 * Waits until lock state is changed.
939 * This function is called with cl_lock mutex locked, atomically releases
940 * mutex and goes to sleep, waiting for a lock state change (signaled by
941 * cl_lock_signal()), and re-acquires the mutex before return.
943 * This function is used to wait until lock state machine makes some progress
944 * and to emulate synchronous operations on top of asynchronous lock
947 * \retval -EINTR wait was interrupted
949 * \retval 0 wait wasn't interrupted
951 * \pre cl_lock_is_mutexed(lock)
953 * \see cl_lock_signal()
955 int cl_lock_state_wait(const struct lu_env *env, struct cl_lock *lock)
962 LINVRNT(cl_lock_is_mutexed(lock));
963 LINVRNT(cl_lock_invariant(env, lock));
964 LASSERT(lock->cll_depth == 1);
965 LASSERT(lock->cll_state != CLS_FREEING); /* too late to wait */
967 cl_lock_trace(D_DLMTRACE, env, "state wait lock", lock);
968 result = lock->cll_error;
970 /* To avoid being interrupted by the 'non-fatal' signals
971 * (SIGCHLD, for instance), we'd block them temporarily.
973 blocked = cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
975 init_waitqueue_entry_current(&waiter);
976 add_wait_queue(&lock->cll_wq, &waiter);
977 set_current_state(TASK_INTERRUPTIBLE);
978 cl_lock_mutex_put(env, lock);
980 LASSERT(cl_lock_nr_mutexed(env) == 0);
982 /* Returning ERESTARTSYS instead of EINTR so syscalls
983 * can be restarted if signals are pending here */
984 result = -ERESTARTSYS;
985 if (likely(!OBD_FAIL_CHECK(OBD_FAIL_LOCK_STATE_WAIT_INTR))) {
986 waitq_wait(&waiter, TASK_INTERRUPTIBLE);
987 if (!cfs_signal_pending())
991 cl_lock_mutex_get(env, lock);
992 set_current_state(TASK_RUNNING);
993 remove_wait_queue(&lock->cll_wq, &waiter);
995 /* Restore old blocked signals */
996 cfs_restore_sigs(blocked);
1000 EXPORT_SYMBOL(cl_lock_state_wait);
1002 static void cl_lock_state_signal(const struct lu_env *env, struct cl_lock *lock,
1003 enum cl_lock_state state)
1005 const struct cl_lock_slice *slice;
1008 LINVRNT(cl_lock_is_mutexed(lock));
1009 LINVRNT(cl_lock_invariant(env, lock));
1011 list_for_each_entry(slice, &lock->cll_layers, cls_linkage)
1012 if (slice->cls_ops->clo_state != NULL)
1013 slice->cls_ops->clo_state(env, slice, state);
1014 wake_up_all(&lock->cll_wq);
1019 * Notifies waiters that lock state changed.
1021 * Wakes up all waiters sleeping in cl_lock_state_wait(), also notifies all
1022 * layers about state change by calling cl_lock_operations::clo_state()
1025 void cl_lock_signal(const struct lu_env *env, struct cl_lock *lock)
1028 cl_lock_trace(D_DLMTRACE, env, "state signal lock", lock);
1029 cl_lock_state_signal(env, lock, lock->cll_state);
1032 EXPORT_SYMBOL(cl_lock_signal);
1035 * Changes lock state.
1037 * This function is invoked to notify layers that lock state changed, possible
1038 * as a result of an asynchronous event such as call-back reception.
1040 * \post lock->cll_state == state
1042 * \see cl_lock_operations::clo_state()
1044 void cl_lock_state_set(const struct lu_env *env, struct cl_lock *lock,
1045 enum cl_lock_state state)
1048 LASSERT(lock->cll_state <= state ||
1049 (lock->cll_state == CLS_CACHED &&
1050 (state == CLS_HELD || /* lock found in cache */
1051 state == CLS_NEW || /* sub-lock canceled */
1052 state == CLS_INTRANSIT)) ||
1053 /* lock is in transit state */
1054 lock->cll_state == CLS_INTRANSIT);
1056 if (lock->cll_state != state) {
1057 CS_LOCKSTATE_DEC(lock->cll_descr.cld_obj, lock->cll_state);
1058 CS_LOCKSTATE_INC(lock->cll_descr.cld_obj, state);
1060 cl_lock_state_signal(env, lock, state);
1061 lock->cll_state = state;
1065 EXPORT_SYMBOL(cl_lock_state_set);
1067 static int cl_unuse_try_internal(const struct lu_env *env, struct cl_lock *lock)
1069 const struct cl_lock_slice *slice;
1075 LINVRNT(cl_lock_is_mutexed(lock));
1076 LINVRNT(cl_lock_invariant(env, lock));
1077 LASSERT(lock->cll_state == CLS_INTRANSIT);
1080 list_for_each_entry_reverse(slice, &lock->cll_layers,
1082 if (slice->cls_ops->clo_unuse != NULL) {
1083 result = slice->cls_ops->clo_unuse(env, slice);
1088 LASSERT(result != -ENOSYS);
1089 } while (result == CLO_REPEAT);
1095 * Yanks lock from the cache (cl_lock_state::CLS_CACHED state) by calling
1096 * cl_lock_operations::clo_use() top-to-bottom to notify layers.
1097 * @atomic = 1, it must unuse the lock to recovery the lock to keep the
1098 * use process atomic
1100 int cl_use_try(const struct lu_env *env, struct cl_lock *lock, int atomic)
1102 const struct cl_lock_slice *slice;
1104 enum cl_lock_state state;
1107 cl_lock_trace(D_DLMTRACE, env, "use lock", lock);
1109 LASSERT(lock->cll_state == CLS_CACHED);
1110 if (lock->cll_error)
1111 RETURN(lock->cll_error);
1114 state = cl_lock_intransit(env, lock);
1115 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1116 if (slice->cls_ops->clo_use != NULL) {
1117 result = slice->cls_ops->clo_use(env, slice);
1122 LASSERT(result != -ENOSYS);
1124 LASSERTF(lock->cll_state == CLS_INTRANSIT, "Wrong state %d.\n",
1130 if (result == -ESTALE) {
1132 * ESTALE means sublock being cancelled
1133 * at this time, and set lock state to
1134 * be NEW here and ask the caller to repeat.
1137 result = CLO_REPEAT;
1140 /* @atomic means back-off-on-failure. */
1143 rc = cl_unuse_try_internal(env, lock);
1144 /* Vet the results. */
1145 if (rc < 0 && result > 0)
1150 cl_lock_extransit(env, lock, state);
1153 EXPORT_SYMBOL(cl_use_try);
1156 * Helper for cl_enqueue_try() that calls ->clo_enqueue() across all layers
1159 static int cl_enqueue_kick(const struct lu_env *env,
1160 struct cl_lock *lock,
1161 struct cl_io *io, __u32 flags)
1164 const struct cl_lock_slice *slice;
1168 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1169 if (slice->cls_ops->clo_enqueue != NULL) {
1170 result = slice->cls_ops->clo_enqueue(env,
1176 LASSERT(result != -ENOSYS);
1181 * Tries to enqueue a lock.
1183 * This function is called repeatedly by cl_enqueue() until either lock is
1184 * enqueued, or error occurs. This function does not block waiting for
1185 * networking communication to complete.
1187 * \post ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1188 * lock->cll_state == CLS_HELD)
1190 * \see cl_enqueue() cl_lock_operations::clo_enqueue()
1191 * \see cl_lock_state::CLS_ENQUEUED
1193 int cl_enqueue_try(const struct lu_env *env, struct cl_lock *lock,
1194 struct cl_io *io, __u32 flags)
1199 cl_lock_trace(D_DLMTRACE, env, "enqueue lock", lock);
1201 LINVRNT(cl_lock_is_mutexed(lock));
1203 result = lock->cll_error;
1207 switch (lock->cll_state) {
1209 cl_lock_state_set(env, lock, CLS_QUEUING);
1213 result = cl_enqueue_kick(env, lock, io, flags);
1214 /* For AGL case, the cl_lock::cll_state may
1215 * become CLS_HELD already. */
1216 if (result == 0 && lock->cll_state == CLS_QUEUING)
1217 cl_lock_state_set(env, lock, CLS_ENQUEUED);
1220 LASSERT(cl_lock_is_intransit(lock));
1224 /* yank lock from the cache. */
1225 result = cl_use_try(env, lock, 0);
1234 * impossible, only held locks with increased
1235 * ->cll_holds can be enqueued, and they cannot be
1240 } while (result == CLO_REPEAT);
1243 EXPORT_SYMBOL(cl_enqueue_try);
1246 * Cancel the conflicting lock found during previous enqueue.
1248 * \retval 0 conflicting lock has been canceled.
1249 * \retval -ve error code.
1251 int cl_lock_enqueue_wait(const struct lu_env *env,
1252 struct cl_lock *lock,
1255 struct cl_lock *conflict;
1259 LASSERT(cl_lock_is_mutexed(lock));
1260 LASSERT(lock->cll_state == CLS_QUEUING);
1261 LASSERT(lock->cll_conflict != NULL);
1263 conflict = lock->cll_conflict;
1264 lock->cll_conflict = NULL;
1266 cl_lock_mutex_put(env, lock);
1267 LASSERT(cl_lock_nr_mutexed(env) == 0);
1269 cl_lock_mutex_get(env, conflict);
1270 cl_lock_trace(D_DLMTRACE, env, "enqueue wait", conflict);
1271 cl_lock_cancel(env, conflict);
1272 cl_lock_delete(env, conflict);
1274 while (conflict->cll_state != CLS_FREEING) {
1275 rc = cl_lock_state_wait(env, conflict);
1279 cl_lock_mutex_put(env, conflict);
1280 lu_ref_del(&conflict->cll_reference, "cancel-wait", lock);
1281 cl_lock_put(env, conflict);
1284 cl_lock_mutex_get(env, lock);
1289 EXPORT_SYMBOL(cl_lock_enqueue_wait);
1291 static int cl_enqueue_locked(const struct lu_env *env, struct cl_lock *lock,
1292 struct cl_io *io, __u32 enqflags)
1298 LINVRNT(cl_lock_is_mutexed(lock));
1299 LINVRNT(cl_lock_invariant(env, lock));
1300 LASSERT(lock->cll_holds > 0);
1302 cl_lock_user_add(env, lock);
1304 result = cl_enqueue_try(env, lock, io, enqflags);
1305 if (result == CLO_WAIT) {
1306 if (lock->cll_conflict != NULL)
1307 result = cl_lock_enqueue_wait(env, lock, 1);
1309 result = cl_lock_state_wait(env, lock);
1316 cl_unuse_try(env, lock);
1317 LASSERT(ergo(result == 0 && !(enqflags & CEF_AGL),
1318 lock->cll_state == CLS_ENQUEUED ||
1319 lock->cll_state == CLS_HELD));
1326 * \pre current thread or io owns a hold on lock.
1328 * \post ergo(result == 0, lock->users increased)
1329 * \post ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1330 * lock->cll_state == CLS_HELD)
1332 int cl_enqueue(const struct lu_env *env, struct cl_lock *lock,
1333 struct cl_io *io, __u32 enqflags)
1339 cl_lock_lockdep_acquire(env, lock, enqflags);
1340 cl_lock_mutex_get(env, lock);
1341 result = cl_enqueue_locked(env, lock, io, enqflags);
1342 cl_lock_mutex_put(env, lock);
1344 cl_lock_lockdep_release(env, lock);
1345 LASSERT(ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1346 lock->cll_state == CLS_HELD));
1349 EXPORT_SYMBOL(cl_enqueue);
1352 * Tries to unlock a lock.
1354 * This function is called to release underlying resource:
1355 * 1. for top lock, the resource is sublocks it held;
1356 * 2. for sublock, the resource is the reference to dlmlock.
1358 * cl_unuse_try is a one-shot operation, so it must NOT return CLO_WAIT.
1360 * \see cl_unuse() cl_lock_operations::clo_unuse()
1361 * \see cl_lock_state::CLS_CACHED
1363 int cl_unuse_try(const struct lu_env *env, struct cl_lock *lock)
1366 enum cl_lock_state state = CLS_NEW;
1369 cl_lock_trace(D_DLMTRACE, env, "unuse lock", lock);
1371 if (lock->cll_users > 1) {
1372 cl_lock_user_del(env, lock);
1376 /* Only if the lock is in CLS_HELD or CLS_ENQUEUED state, it can hold
1377 * underlying resources. */
1378 if (!(lock->cll_state == CLS_HELD || lock->cll_state == CLS_ENQUEUED)) {
1379 cl_lock_user_del(env, lock);
1384 * New lock users (->cll_users) are not protecting unlocking
1385 * from proceeding. From this point, lock eventually reaches
1386 * CLS_CACHED, is reinitialized to CLS_NEW or fails into
1389 state = cl_lock_intransit(env, lock);
1391 result = cl_unuse_try_internal(env, lock);
1392 LASSERT(lock->cll_state == CLS_INTRANSIT);
1393 LASSERT(result != CLO_WAIT);
1394 cl_lock_user_del(env, lock);
1395 if (result == 0 || result == -ESTALE) {
1397 * Return lock back to the cache. This is the only
1398 * place where lock is moved into CLS_CACHED state.
1400 * If one of ->clo_unuse() methods returned -ESTALE, lock
1401 * cannot be placed into cache and has to be
1402 * re-initialized. This happens e.g., when a sub-lock was
1403 * canceled while unlocking was in progress.
1405 if (state == CLS_HELD && result == 0)
1409 cl_lock_extransit(env, lock, state);
1412 * Hide -ESTALE error.
1413 * If the lock is a glimpse lock, and it has multiple
1414 * stripes. Assuming that one of its sublock returned -ENAVAIL,
1415 * and other sublocks are matched write locks. In this case,
1416 * we can't set this lock to error because otherwise some of
1417 * its sublocks may not be canceled. This causes some dirty
1418 * pages won't be written to OSTs. -jay
1422 CERROR("result = %d, this is unlikely!\n", result);
1424 cl_lock_extransit(env, lock, state);
1426 RETURN(result ?: lock->cll_error);
1428 EXPORT_SYMBOL(cl_unuse_try);
1430 static void cl_unuse_locked(const struct lu_env *env, struct cl_lock *lock)
1435 result = cl_unuse_try(env, lock);
1437 CL_LOCK_DEBUG(D_ERROR, env, lock, "unuse return %d\n", result);
1445 void cl_unuse(const struct lu_env *env, struct cl_lock *lock)
1448 cl_lock_mutex_get(env, lock);
1449 cl_unuse_locked(env, lock);
1450 cl_lock_mutex_put(env, lock);
1451 cl_lock_lockdep_release(env, lock);
1454 EXPORT_SYMBOL(cl_unuse);
1457 * Tries to wait for a lock.
1459 * This function is called repeatedly by cl_wait() until either lock is
1460 * granted, or error occurs. This function does not block waiting for network
1461 * communication to complete.
1463 * \see cl_wait() cl_lock_operations::clo_wait()
1464 * \see cl_lock_state::CLS_HELD
1466 int cl_wait_try(const struct lu_env *env, struct cl_lock *lock)
1468 const struct cl_lock_slice *slice;
1472 cl_lock_trace(D_DLMTRACE, env, "wait lock try", lock);
1474 LINVRNT(cl_lock_is_mutexed(lock));
1475 LINVRNT(cl_lock_invariant(env, lock));
1476 LASSERTF(lock->cll_state == CLS_QUEUING ||
1477 lock->cll_state == CLS_ENQUEUED ||
1478 lock->cll_state == CLS_HELD ||
1479 lock->cll_state == CLS_INTRANSIT,
1480 "lock state: %d\n", lock->cll_state);
1481 LASSERT(lock->cll_users > 0);
1482 LASSERT(lock->cll_holds > 0);
1484 result = lock->cll_error;
1488 if (cl_lock_is_intransit(lock)) {
1493 if (lock->cll_state == CLS_HELD)
1498 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1499 if (slice->cls_ops->clo_wait != NULL) {
1500 result = slice->cls_ops->clo_wait(env, slice);
1505 LASSERT(result != -ENOSYS);
1507 LASSERT(lock->cll_state != CLS_INTRANSIT);
1508 cl_lock_state_set(env, lock, CLS_HELD);
1510 } while (result == CLO_REPEAT);
1513 EXPORT_SYMBOL(cl_wait_try);
1516 * Waits until enqueued lock is granted.
1518 * \pre current thread or io owns a hold on the lock
1519 * \pre ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1520 * lock->cll_state == CLS_HELD)
1522 * \post ergo(result == 0, lock->cll_state == CLS_HELD)
1524 int cl_wait(const struct lu_env *env, struct cl_lock *lock)
1529 cl_lock_mutex_get(env, lock);
1531 LINVRNT(cl_lock_invariant(env, lock));
1532 LASSERTF(lock->cll_state == CLS_ENQUEUED || lock->cll_state == CLS_HELD,
1533 "Wrong state %d \n", lock->cll_state);
1534 LASSERT(lock->cll_holds > 0);
1537 result = cl_wait_try(env, lock);
1538 if (result == CLO_WAIT) {
1539 result = cl_lock_state_wait(env, lock);
1546 cl_unuse_try(env, lock);
1547 cl_lock_lockdep_release(env, lock);
1549 cl_lock_trace(D_DLMTRACE, env, "wait lock", lock);
1550 cl_lock_mutex_put(env, lock);
1551 LASSERT(ergo(result == 0, lock->cll_state == CLS_HELD));
1554 EXPORT_SYMBOL(cl_wait);
1557 * Executes cl_lock_operations::clo_weigh(), and sums results to estimate lock
1560 unsigned long cl_lock_weigh(const struct lu_env *env, struct cl_lock *lock)
1562 const struct cl_lock_slice *slice;
1563 unsigned long pound;
1564 unsigned long ounce;
1567 LINVRNT(cl_lock_is_mutexed(lock));
1568 LINVRNT(cl_lock_invariant(env, lock));
1571 list_for_each_entry_reverse(slice, &lock->cll_layers, cls_linkage) {
1572 if (slice->cls_ops->clo_weigh != NULL) {
1573 ounce = slice->cls_ops->clo_weigh(env, slice);
1575 if (pound < ounce) /* over-weight^Wflow */
1581 EXPORT_SYMBOL(cl_lock_weigh);
1584 * Notifies layers that lock description changed.
1586 * The server can grant client a lock different from one that was requested
1587 * (e.g., larger in extent). This method is called when actually granted lock
1588 * description becomes known to let layers to accommodate for changed lock
1591 * \see cl_lock_operations::clo_modify()
1593 int cl_lock_modify(const struct lu_env *env, struct cl_lock *lock,
1594 const struct cl_lock_descr *desc)
1596 const struct cl_lock_slice *slice;
1597 struct cl_object *obj = lock->cll_descr.cld_obj;
1598 struct cl_object_header *hdr = cl_object_header(obj);
1602 cl_lock_trace(D_DLMTRACE, env, "modify lock", lock);
1603 /* don't allow object to change */
1604 LASSERT(obj == desc->cld_obj);
1605 LINVRNT(cl_lock_is_mutexed(lock));
1606 LINVRNT(cl_lock_invariant(env, lock));
1608 list_for_each_entry_reverse(slice, &lock->cll_layers, cls_linkage) {
1609 if (slice->cls_ops->clo_modify != NULL) {
1610 result = slice->cls_ops->clo_modify(env, slice, desc);
1615 CL_LOCK_DEBUG(D_DLMTRACE, env, lock, " -> "DDESCR"@"DFID"\n",
1616 PDESCR(desc), PFID(lu_object_fid(&desc->cld_obj->co_lu)));
1618 * Just replace description in place. Nothing more is needed for
1619 * now. If locks were indexed according to their extent and/or mode,
1620 * that index would have to be updated here.
1622 spin_lock(&hdr->coh_lock_guard);
1623 lock->cll_descr = *desc;
1624 spin_unlock(&hdr->coh_lock_guard);
1627 EXPORT_SYMBOL(cl_lock_modify);
1630 * Initializes lock closure with a given origin.
1632 * \see cl_lock_closure
1634 void cl_lock_closure_init(const struct lu_env *env,
1635 struct cl_lock_closure *closure,
1636 struct cl_lock *origin, int wait)
1638 LINVRNT(cl_lock_is_mutexed(origin));
1639 LINVRNT(cl_lock_invariant(env, origin));
1641 INIT_LIST_HEAD(&closure->clc_list);
1642 closure->clc_origin = origin;
1643 closure->clc_wait = wait;
1644 closure->clc_nr = 0;
1646 EXPORT_SYMBOL(cl_lock_closure_init);
1649 * Builds a closure of \a lock.
1651 * Building of a closure consists of adding initial lock (\a lock) into it,
1652 * and calling cl_lock_operations::clo_closure() methods of \a lock. These
1653 * methods might call cl_lock_closure_build() recursively again, adding more
1654 * locks to the closure, etc.
1656 * \see cl_lock_closure
1658 int cl_lock_closure_build(const struct lu_env *env, struct cl_lock *lock,
1659 struct cl_lock_closure *closure)
1661 const struct cl_lock_slice *slice;
1665 LINVRNT(cl_lock_is_mutexed(closure->clc_origin));
1666 LINVRNT(cl_lock_invariant(env, closure->clc_origin));
1668 result = cl_lock_enclosure(env, lock, closure);
1670 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1671 if (slice->cls_ops->clo_closure != NULL) {
1672 result = slice->cls_ops->clo_closure(env, slice,
1680 cl_lock_disclosure(env, closure);
1683 EXPORT_SYMBOL(cl_lock_closure_build);
1686 * Adds new lock to a closure.
1688 * Try-locks \a lock and if succeeded, adds it to the closure (never more than
1689 * once). If try-lock failed, returns CLO_REPEAT, after optionally waiting
1690 * until next try-lock is likely to succeed.
1692 int cl_lock_enclosure(const struct lu_env *env, struct cl_lock *lock,
1693 struct cl_lock_closure *closure)
1697 cl_lock_trace(D_DLMTRACE, env, "enclosure lock", lock);
1698 if (!cl_lock_mutex_try(env, lock)) {
1700 * If lock->cll_inclosure is not empty, lock is already in
1703 if (list_empty(&lock->cll_inclosure)) {
1704 cl_lock_get_trust(lock);
1705 lu_ref_add(&lock->cll_reference, "closure", closure);
1706 list_add(&lock->cll_inclosure, &closure->clc_list);
1709 cl_lock_mutex_put(env, lock);
1712 cl_lock_disclosure(env, closure);
1713 if (closure->clc_wait) {
1714 cl_lock_get_trust(lock);
1715 lu_ref_add(&lock->cll_reference, "closure-w", closure);
1716 cl_lock_mutex_put(env, closure->clc_origin);
1718 LASSERT(cl_lock_nr_mutexed(env) == 0);
1719 cl_lock_mutex_get(env, lock);
1720 cl_lock_mutex_put(env, lock);
1722 cl_lock_mutex_get(env, closure->clc_origin);
1723 lu_ref_del(&lock->cll_reference, "closure-w", closure);
1724 cl_lock_put(env, lock);
1726 result = CLO_REPEAT;
1730 EXPORT_SYMBOL(cl_lock_enclosure);
1732 /** Releases mutices of enclosed locks. */
1733 void cl_lock_disclosure(const struct lu_env *env,
1734 struct cl_lock_closure *closure)
1736 struct cl_lock *scan;
1737 struct cl_lock *temp;
1739 cl_lock_trace(D_DLMTRACE, env, "disclosure lock", closure->clc_origin);
1740 list_for_each_entry_safe(scan, temp, &closure->clc_list,
1742 list_del_init(&scan->cll_inclosure);
1743 cl_lock_mutex_put(env, scan);
1744 lu_ref_del(&scan->cll_reference, "closure", closure);
1745 cl_lock_put(env, scan);
1748 LASSERT(closure->clc_nr == 0);
1750 EXPORT_SYMBOL(cl_lock_disclosure);
1752 /** Finalizes a closure. */
1753 void cl_lock_closure_fini(struct cl_lock_closure *closure)
1755 LASSERT(closure->clc_nr == 0);
1756 LASSERT(list_empty(&closure->clc_list));
1758 EXPORT_SYMBOL(cl_lock_closure_fini);
1761 * Destroys this lock. Notifies layers (bottom-to-top) that lock is being
1762 * destroyed, then destroy the lock. If there are holds on the lock, postpone
1763 * destruction until all holds are released. This is called when a decision is
1764 * made to destroy the lock in the future. E.g., when a blocking AST is
1765 * received on it, or fatal communication error happens.
1767 * Caller must have a reference on this lock to prevent a situation, when
1768 * deleted lock lingers in memory for indefinite time, because nobody calls
1769 * cl_lock_put() to finish it.
1771 * \pre atomic_read(&lock->cll_ref) > 0
1772 * \pre ergo(cl_lock_nesting(lock) == CNL_TOP,
1773 * cl_lock_nr_mutexed(env) == 1)
1774 * [i.e., if a top-lock is deleted, mutices of no other locks can be
1775 * held, as deletion of sub-locks might require releasing a top-lock
1778 * \see cl_lock_operations::clo_delete()
1779 * \see cl_lock::cll_holds
1781 void cl_lock_delete(const struct lu_env *env, struct cl_lock *lock)
1783 LINVRNT(cl_lock_is_mutexed(lock));
1784 LINVRNT(cl_lock_invariant(env, lock));
1785 LASSERT(ergo(cl_lock_nesting(lock) == CNL_TOP,
1786 cl_lock_nr_mutexed(env) == 1));
1789 cl_lock_trace(D_DLMTRACE, env, "delete lock", lock);
1790 if (lock->cll_holds == 0)
1791 cl_lock_delete0(env, lock);
1793 lock->cll_flags |= CLF_DOOMED;
1796 EXPORT_SYMBOL(cl_lock_delete);
1799 * Mark lock as irrecoverably failed, and mark it for destruction. This
1800 * happens when, e.g., server fails to grant a lock to us, or networking
1803 * \pre atomic_read(&lock->cll_ref) > 0
1805 * \see clo_lock_delete()
1806 * \see cl_lock::cll_holds
1808 void cl_lock_error(const struct lu_env *env, struct cl_lock *lock, int error)
1810 LINVRNT(cl_lock_is_mutexed(lock));
1811 LINVRNT(cl_lock_invariant(env, lock));
1814 if (lock->cll_error == 0 && error != 0) {
1815 cl_lock_trace(D_DLMTRACE, env, "set lock error", lock);
1816 lock->cll_error = error;
1817 cl_lock_signal(env, lock);
1818 cl_lock_cancel(env, lock);
1819 cl_lock_delete(env, lock);
1823 EXPORT_SYMBOL(cl_lock_error);
1826 * Cancels this lock. Notifies layers
1827 * (bottom-to-top) that lock is being cancelled, then destroy the lock. If
1828 * there are holds on the lock, postpone cancellation until
1829 * all holds are released.
1831 * Cancellation notification is delivered to layers at most once.
1833 * \see cl_lock_operations::clo_cancel()
1834 * \see cl_lock::cll_holds
1836 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock)
1838 LINVRNT(cl_lock_is_mutexed(lock));
1839 LINVRNT(cl_lock_invariant(env, lock));
1842 cl_lock_trace(D_DLMTRACE, env, "cancel lock", lock);
1843 if (lock->cll_holds == 0)
1844 cl_lock_cancel0(env, lock);
1846 lock->cll_flags |= CLF_CANCELPEND;
1849 EXPORT_SYMBOL(cl_lock_cancel);
1852 * Finds an existing lock covering given index and optionally different from a
1853 * given \a except lock.
1855 struct cl_lock *cl_lock_at_pgoff(const struct lu_env *env,
1856 struct cl_object *obj, pgoff_t index,
1857 struct cl_lock *except,
1858 int pending, int canceld)
1860 struct cl_object_header *head;
1861 struct cl_lock *scan;
1862 struct cl_lock *lock;
1863 struct cl_lock_descr *need;
1867 head = cl_object_header(obj);
1868 need = &cl_env_info(env)->clt_descr;
1871 need->cld_mode = CLM_READ; /* CLM_READ matches both READ & WRITE, but
1873 need->cld_start = need->cld_end = index;
1874 need->cld_enq_flags = 0;
1876 spin_lock(&head->coh_lock_guard);
1877 /* It is fine to match any group lock since there could be only one
1878 * with a uniq gid and it conflicts with all other lock modes too */
1879 list_for_each_entry(scan, &head->coh_locks, cll_linkage) {
1880 if (scan != except &&
1881 (scan->cll_descr.cld_mode == CLM_GROUP ||
1882 cl_lock_ext_match(&scan->cll_descr, need)) &&
1883 scan->cll_state >= CLS_HELD &&
1884 scan->cll_state < CLS_FREEING &&
1886 * This check is racy as the lock can be canceled right
1887 * after it is done, but this is fine, because page exists
1890 (canceld || !(scan->cll_flags & CLF_CANCELLED)) &&
1891 (pending || !(scan->cll_flags & CLF_CANCELPEND))) {
1892 /* Don't increase cs_hit here since this
1893 * is just a helper function. */
1894 cl_lock_get_trust(scan);
1899 spin_unlock(&head->coh_lock_guard);
1902 EXPORT_SYMBOL(cl_lock_at_pgoff);
1905 * Eliminate all locks for a given object.
1907 * Caller has to guarantee that no lock is in active use.
1909 * \param cancel when this is set, cl_locks_prune() cancels locks before
1912 void cl_locks_prune(const struct lu_env *env, struct cl_object *obj, int cancel)
1914 struct cl_object_header *head;
1915 struct cl_lock *lock;
1918 head = cl_object_header(obj);
1920 spin_lock(&head->coh_lock_guard);
1921 while (!list_empty(&head->coh_locks)) {
1922 lock = container_of(head->coh_locks.next,
1923 struct cl_lock, cll_linkage);
1924 cl_lock_get_trust(lock);
1925 spin_unlock(&head->coh_lock_guard);
1926 lu_ref_add(&lock->cll_reference, "prune", current);
1929 cl_lock_mutex_get(env, lock);
1930 if (lock->cll_state < CLS_FREEING) {
1931 LASSERT(lock->cll_users <= 1);
1932 if (unlikely(lock->cll_users == 1)) {
1933 struct l_wait_info lwi = { 0 };
1935 cl_lock_mutex_put(env, lock);
1936 l_wait_event(lock->cll_wq,
1937 lock->cll_users == 0,
1943 cl_lock_cancel(env, lock);
1944 cl_lock_delete(env, lock);
1946 cl_lock_mutex_put(env, lock);
1947 lu_ref_del(&lock->cll_reference, "prune", current);
1948 cl_lock_put(env, lock);
1949 spin_lock(&head->coh_lock_guard);
1951 spin_unlock(&head->coh_lock_guard);
1954 EXPORT_SYMBOL(cl_locks_prune);
1956 static struct cl_lock *cl_lock_hold_mutex(const struct lu_env *env,
1957 const struct cl_io *io,
1958 const struct cl_lock_descr *need,
1959 const char *scope, const void *source)
1961 struct cl_lock *lock;
1966 lock = cl_lock_find(env, io, need);
1969 cl_lock_mutex_get(env, lock);
1970 if (lock->cll_state < CLS_FREEING &&
1971 !(lock->cll_flags & CLF_CANCELLED)) {
1972 cl_lock_hold_mod(env, lock, +1);
1973 lu_ref_add(&lock->cll_holders, scope, source);
1974 lu_ref_add(&lock->cll_reference, scope, source);
1977 cl_lock_mutex_put(env, lock);
1978 cl_lock_put(env, lock);
1984 * Returns a lock matching \a need description with a reference and a hold on
1987 * This is much like cl_lock_find(), except that cl_lock_hold() additionally
1988 * guarantees that lock is not in the CLS_FREEING state on return.
1990 struct cl_lock *cl_lock_hold(const struct lu_env *env, const struct cl_io *io,
1991 const struct cl_lock_descr *need,
1992 const char *scope, const void *source)
1994 struct cl_lock *lock;
1998 lock = cl_lock_hold_mutex(env, io, need, scope, source);
2000 cl_lock_mutex_put(env, lock);
2003 EXPORT_SYMBOL(cl_lock_hold);
2006 * Main high-level entry point of cl_lock interface that finds existing or
2007 * enqueues new lock matching given description.
2009 struct cl_lock *cl_lock_request(const struct lu_env *env, struct cl_io *io,
2010 const struct cl_lock_descr *need,
2011 const char *scope, const void *source)
2013 struct cl_lock *lock;
2015 __u32 enqflags = need->cld_enq_flags;
2019 lock = cl_lock_hold_mutex(env, io, need, scope, source);
2023 rc = cl_enqueue_locked(env, lock, io, enqflags);
2025 if (cl_lock_fits_into(env, lock, need, io)) {
2026 if (!(enqflags & CEF_AGL)) {
2027 cl_lock_mutex_put(env, lock);
2028 cl_lock_lockdep_acquire(env, lock,
2034 cl_unuse_locked(env, lock);
2036 cl_lock_trace(D_DLMTRACE, env,
2037 rc <= 0 ? "enqueue failed" : "agl succeed", lock);
2038 cl_lock_hold_release(env, lock, scope, source);
2039 cl_lock_mutex_put(env, lock);
2040 lu_ref_del(&lock->cll_reference, scope, source);
2041 cl_lock_put(env, lock);
2043 LASSERT(enqflags & CEF_AGL);
2045 } else if (rc != 0) {
2051 EXPORT_SYMBOL(cl_lock_request);
2054 * Adds a hold to a known lock.
2056 void cl_lock_hold_add(const struct lu_env *env, struct cl_lock *lock,
2057 const char *scope, const void *source)
2059 LINVRNT(cl_lock_is_mutexed(lock));
2060 LINVRNT(cl_lock_invariant(env, lock));
2061 LASSERT(lock->cll_state != CLS_FREEING);
2065 cl_lock_hold_mod(env, lock, +1);
2066 lu_ref_add(&lock->cll_holders, scope, source);
2067 lu_ref_add(&lock->cll_reference, scope, source);
2070 EXPORT_SYMBOL(cl_lock_hold_add);
2073 * Releases a hold and a reference on a lock, on which caller acquired a
2076 void cl_lock_unhold(const struct lu_env *env, struct cl_lock *lock,
2077 const char *scope, const void *source)
2079 LINVRNT(cl_lock_invariant(env, lock));
2081 cl_lock_hold_release(env, lock, scope, source);
2082 lu_ref_del(&lock->cll_reference, scope, source);
2083 cl_lock_put(env, lock);
2086 EXPORT_SYMBOL(cl_lock_unhold);
2089 * Releases a hold and a reference on a lock, obtained by cl_lock_hold().
2091 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock,
2092 const char *scope, const void *source)
2094 LINVRNT(cl_lock_invariant(env, lock));
2096 cl_lock_trace(D_DLMTRACE, env, "release lock", lock);
2097 cl_lock_mutex_get(env, lock);
2098 cl_lock_hold_release(env, lock, scope, source);
2099 cl_lock_mutex_put(env, lock);
2100 lu_ref_del(&lock->cll_reference, scope, source);
2101 cl_lock_put(env, lock);
2104 EXPORT_SYMBOL(cl_lock_release);
2106 void cl_lock_user_add(const struct lu_env *env, struct cl_lock *lock)
2108 LINVRNT(cl_lock_is_mutexed(lock));
2109 LINVRNT(cl_lock_invariant(env, lock));
2112 cl_lock_used_mod(env, lock, +1);
2115 EXPORT_SYMBOL(cl_lock_user_add);
2117 void cl_lock_user_del(const struct lu_env *env, struct cl_lock *lock)
2119 LINVRNT(cl_lock_is_mutexed(lock));
2120 LINVRNT(cl_lock_invariant(env, lock));
2121 LASSERT(lock->cll_users > 0);
2124 cl_lock_used_mod(env, lock, -1);
2125 if (lock->cll_users == 0)
2126 wake_up_all(&lock->cll_wq);
2129 EXPORT_SYMBOL(cl_lock_user_del);
2131 const char *cl_lock_mode_name(const enum cl_lock_mode mode)
2133 static const char *names[] = {
2134 [CLM_PHANTOM] = "P",
2139 CLASSERT(CLM_MAX == ARRAY_SIZE(names));
2142 EXPORT_SYMBOL(cl_lock_mode_name);
2145 * Prints human readable representation of a lock description.
2147 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2148 lu_printer_t printer,
2149 const struct cl_lock_descr *descr)
2151 const struct lu_fid *fid;
2153 fid = lu_object_fid(&descr->cld_obj->co_lu);
2154 (*printer)(env, cookie, DDESCR"@"DFID, PDESCR(descr), PFID(fid));
2156 EXPORT_SYMBOL(cl_lock_descr_print);
2159 * Prints human readable representation of \a lock to the \a f.
2161 void cl_lock_print(const struct lu_env *env, void *cookie,
2162 lu_printer_t printer, const struct cl_lock *lock)
2164 const struct cl_lock_slice *slice;
2165 (*printer)(env, cookie, "lock@%p[%d %d %d %d %d %08lx] ",
2166 lock, atomic_read(&lock->cll_ref),
2167 lock->cll_state, lock->cll_error, lock->cll_holds,
2168 lock->cll_users, lock->cll_flags);
2169 cl_lock_descr_print(env, cookie, printer, &lock->cll_descr);
2170 (*printer)(env, cookie, " {\n");
2172 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
2173 (*printer)(env, cookie, " %s@%p: ",
2174 slice->cls_obj->co_lu.lo_dev->ld_type->ldt_name,
2176 if (slice->cls_ops->clo_print != NULL)
2177 slice->cls_ops->clo_print(env, cookie, printer, slice);
2178 (*printer)(env, cookie, "\n");
2180 (*printer)(env, cookie, "} lock@%p\n", lock);
2182 EXPORT_SYMBOL(cl_lock_print);
2184 int cl_lock_init(void)
2186 return lu_kmem_init(cl_lock_caches);
2189 void cl_lock_fini(void)
2191 lu_kmem_fini(cl_lock_caches);