1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
41 * Author: Nikita Danilov <nikita.danilov@sun.com>
44 #define DEBUG_SUBSYSTEM S_CLASS
46 # define EXPORT_SYMTAB
49 #include <obd_class.h>
50 #include <obd_support.h>
51 #include <lustre_fid.h>
52 #include <libcfs/list.h>
53 /* lu_time_global_{init,fini}() */
56 #include <cl_object.h>
57 #include "cl_internal.h"
59 /** Lock class of cl_lock::cll_guard */
60 static cfs_lock_class_key_t cl_lock_guard_class;
61 static cfs_mem_cache_t *cl_lock_kmem;
63 static struct lu_kmem_descr cl_lock_caches[] = {
65 .ckd_cache = &cl_lock_kmem,
66 .ckd_name = "cl_lock_kmem",
67 .ckd_size = sizeof (struct cl_lock)
75 * Basic lock invariant that is maintained at all times. Caller either has a
76 * reference to \a lock, or somehow assures that \a lock cannot be freed.
78 * \see cl_lock_invariant()
80 static int cl_lock_invariant_trusted(const struct lu_env *env,
81 const struct cl_lock *lock)
83 return ergo(lock->cll_state == CLS_FREEING, lock->cll_holds == 0) &&
84 cfs_atomic_read(&lock->cll_ref) >= lock->cll_holds &&
85 lock->cll_holds >= lock->cll_users &&
86 lock->cll_holds >= 0 &&
87 lock->cll_users >= 0 &&
92 * Stronger lock invariant, checking that caller has a reference on a lock.
94 * \see cl_lock_invariant_trusted()
96 static int cl_lock_invariant(const struct lu_env *env,
97 const struct cl_lock *lock)
101 result = cfs_atomic_read(&lock->cll_ref) > 0 &&
102 cl_lock_invariant_trusted(env, lock);
103 if (!result && env != NULL)
104 CL_LOCK_DEBUG(D_ERROR, env, lock, "invariant broken");
109 * Returns lock "nesting": 0 for a top-lock and 1 for a sub-lock.
111 static enum clt_nesting_level cl_lock_nesting(const struct cl_lock *lock)
113 return cl_object_header(lock->cll_descr.cld_obj)->coh_nesting;
117 * Returns a set of counters for this lock, depending on a lock nesting.
119 static struct cl_thread_counters *cl_lock_counters(const struct lu_env *env,
120 const struct cl_lock *lock)
122 struct cl_thread_info *info;
123 enum clt_nesting_level nesting;
125 info = cl_env_info(env);
126 nesting = cl_lock_nesting(lock);
127 LASSERT(nesting < ARRAY_SIZE(info->clt_counters));
128 return &info->clt_counters[nesting];
131 static void cl_lock_trace0(int level, const struct lu_env *env,
132 const char *prefix, const struct cl_lock *lock,
133 const char *func, const int line)
135 struct cl_object_header *h = cl_object_header(lock->cll_descr.cld_obj);
136 CDEBUG(level, "%s: %p@(%d %p %d %d %d %d %d %lx)"
137 "(%p/%d/%d) at %s():%d\n",
138 prefix, lock, cfs_atomic_read(&lock->cll_ref),
139 lock->cll_guarder, lock->cll_depth,
140 lock->cll_state, lock->cll_error, lock->cll_holds,
141 lock->cll_users, lock->cll_flags,
142 env, h->coh_nesting, cl_lock_nr_mutexed(env),
145 #define cl_lock_trace(level, env, prefix, lock) \
146 cl_lock_trace0(level, env, prefix, lock, __FUNCTION__, __LINE__)
148 #define RETIP ((unsigned long)__builtin_return_address(0))
150 #ifdef CONFIG_LOCKDEP
151 static cfs_lock_class_key_t cl_lock_key;
153 static void cl_lock_lockdep_init(struct cl_lock *lock)
155 lockdep_set_class_and_name(lock, &cl_lock_key, "EXT");
158 static void cl_lock_lockdep_acquire(const struct lu_env *env,
159 struct cl_lock *lock, __u32 enqflags)
161 cl_lock_counters(env, lock)->ctc_nr_locks_acquired++;
162 #ifdef HAVE_LOCK_MAP_ACQUIRE
163 lock_map_acquire(&lock->dep_map);
164 #else /* HAVE_LOCK_MAP_ACQUIRE */
165 lock_acquire(&lock->dep_map, !!(enqflags & CEF_ASYNC),
166 /* try: */ 0, lock->cll_descr.cld_mode <= CLM_READ,
167 /* check: */ 2, RETIP);
168 #endif /* HAVE_LOCK_MAP_ACQUIRE */
171 static void cl_lock_lockdep_release(const struct lu_env *env,
172 struct cl_lock *lock)
174 cl_lock_counters(env, lock)->ctc_nr_locks_acquired--;
175 lock_release(&lock->dep_map, 0, RETIP);
178 #else /* !CONFIG_LOCKDEP */
180 static void cl_lock_lockdep_init(struct cl_lock *lock)
182 static void cl_lock_lockdep_acquire(const struct lu_env *env,
183 struct cl_lock *lock, __u32 enqflags)
185 static void cl_lock_lockdep_release(const struct lu_env *env,
186 struct cl_lock *lock)
189 #endif /* !CONFIG_LOCKDEP */
192 * Adds lock slice to the compound lock.
194 * This is called by cl_object_operations::coo_lock_init() methods to add a
195 * per-layer state to the lock. New state is added at the end of
196 * cl_lock::cll_layers list, that is, it is at the bottom of the stack.
198 * \see cl_req_slice_add(), cl_page_slice_add(), cl_io_slice_add()
200 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
201 struct cl_object *obj,
202 const struct cl_lock_operations *ops)
205 slice->cls_lock = lock;
206 cfs_list_add_tail(&slice->cls_linkage, &lock->cll_layers);
207 slice->cls_obj = obj;
208 slice->cls_ops = ops;
211 EXPORT_SYMBOL(cl_lock_slice_add);
214 * Returns true iff a lock with the mode \a has provides at least the same
215 * guarantees as a lock with the mode \a need.
217 int cl_lock_mode_match(enum cl_lock_mode has, enum cl_lock_mode need)
219 LINVRNT(need == CLM_READ || need == CLM_WRITE ||
220 need == CLM_PHANTOM || need == CLM_GROUP);
221 LINVRNT(has == CLM_READ || has == CLM_WRITE ||
222 has == CLM_PHANTOM || has == CLM_GROUP);
223 CLASSERT(CLM_PHANTOM < CLM_READ);
224 CLASSERT(CLM_READ < CLM_WRITE);
225 CLASSERT(CLM_WRITE < CLM_GROUP);
227 if (has != CLM_GROUP)
232 EXPORT_SYMBOL(cl_lock_mode_match);
235 * Returns true iff extent portions of lock descriptions match.
237 int cl_lock_ext_match(const struct cl_lock_descr *has,
238 const struct cl_lock_descr *need)
241 has->cld_start <= need->cld_start &&
242 has->cld_end >= need->cld_end &&
243 cl_lock_mode_match(has->cld_mode, need->cld_mode) &&
244 (has->cld_mode != CLM_GROUP || has->cld_gid == need->cld_gid);
246 EXPORT_SYMBOL(cl_lock_ext_match);
249 * Returns true iff a lock with the description \a has provides at least the
250 * same guarantees as a lock with the description \a need.
252 int cl_lock_descr_match(const struct cl_lock_descr *has,
253 const struct cl_lock_descr *need)
256 cl_object_same(has->cld_obj, need->cld_obj) &&
257 cl_lock_ext_match(has, need);
259 EXPORT_SYMBOL(cl_lock_descr_match);
261 static void cl_lock_free(const struct lu_env *env, struct cl_lock *lock)
263 struct cl_object *obj = lock->cll_descr.cld_obj;
265 LINVRNT(!cl_lock_is_mutexed(lock));
268 cl_lock_trace(D_DLMTRACE, env, "free lock", lock);
270 while (!cfs_list_empty(&lock->cll_layers)) {
271 struct cl_lock_slice *slice;
273 slice = cfs_list_entry(lock->cll_layers.next,
274 struct cl_lock_slice, cls_linkage);
275 cfs_list_del_init(lock->cll_layers.next);
276 slice->cls_ops->clo_fini(env, slice);
278 cfs_atomic_dec(&cl_object_site(obj)->cs_locks.cs_total);
279 cfs_atomic_dec(&cl_object_site(obj)->cs_locks_state[lock->cll_state]);
280 lu_object_ref_del_at(&obj->co_lu, lock->cll_obj_ref, "cl_lock", lock);
281 cl_object_put(env, obj);
282 lu_ref_fini(&lock->cll_reference);
283 lu_ref_fini(&lock->cll_holders);
284 cfs_mutex_destroy(&lock->cll_guard);
285 OBD_SLAB_FREE_PTR(lock, cl_lock_kmem);
290 * Releases a reference on a lock.
292 * When last reference is released, lock is returned to the cache, unless it
293 * is in cl_lock_state::CLS_FREEING state, in which case it is destroyed
296 * \see cl_object_put(), cl_page_put()
298 void cl_lock_put(const struct lu_env *env, struct cl_lock *lock)
300 struct cl_object *obj;
301 struct cl_site *site;
303 LINVRNT(cl_lock_invariant(env, lock));
305 obj = lock->cll_descr.cld_obj;
306 LINVRNT(obj != NULL);
307 site = cl_object_site(obj);
309 CDEBUG(D_TRACE, "releasing reference: %d %p %lu\n",
310 cfs_atomic_read(&lock->cll_ref), lock, RETIP);
312 if (cfs_atomic_dec_and_test(&lock->cll_ref)) {
313 if (lock->cll_state == CLS_FREEING) {
314 LASSERT(cfs_list_empty(&lock->cll_linkage));
315 cl_lock_free(env, lock);
317 cfs_atomic_dec(&site->cs_locks.cs_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 cfs_atomic_read(&lock->cll_ref), lock, RETIP);
336 cfs_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 struct cl_site *site = cl_object_site(lock->cll_descr.cld_obj);
353 CDEBUG(D_TRACE, "acquiring trusted reference: %d %p %lu\n",
354 cfs_atomic_read(&lock->cll_ref), lock, RETIP);
355 if (cfs_atomic_inc_return(&lock->cll_ref) == 1)
356 cfs_atomic_inc(&site->cs_locks.cs_busy);
358 EXPORT_SYMBOL(cl_lock_get_trust);
361 * Helper function destroying the lock that wasn't completely initialized.
363 * Other threads can acquire references to the top-lock through its
364 * sub-locks. Hence, it cannot be cl_lock_free()-ed immediately.
366 static void cl_lock_finish(const struct lu_env *env, struct cl_lock *lock)
368 cl_lock_mutex_get(env, lock);
369 cl_lock_cancel(env, lock);
370 cl_lock_delete(env, lock);
371 cl_lock_mutex_put(env, lock);
372 cl_lock_put(env, lock);
375 static struct cl_lock *cl_lock_alloc(const struct lu_env *env,
376 struct cl_object *obj,
377 const struct cl_io *io,
378 const struct cl_lock_descr *descr)
380 struct cl_lock *lock;
381 struct lu_object_header *head;
382 struct cl_site *site = cl_object_site(obj);
385 OBD_SLAB_ALLOC_PTR_GFP(lock, cl_lock_kmem, CFS_ALLOC_IO);
387 cfs_atomic_set(&lock->cll_ref, 1);
388 lock->cll_descr = *descr;
389 lock->cll_state = CLS_NEW;
391 lock->cll_obj_ref = lu_object_ref_add(&obj->co_lu,
393 CFS_INIT_LIST_HEAD(&lock->cll_layers);
394 CFS_INIT_LIST_HEAD(&lock->cll_linkage);
395 CFS_INIT_LIST_HEAD(&lock->cll_inclosure);
396 lu_ref_init(&lock->cll_reference);
397 lu_ref_init(&lock->cll_holders);
398 cfs_mutex_init(&lock->cll_guard);
399 cfs_lockdep_set_class(&lock->cll_guard, &cl_lock_guard_class);
400 cfs_waitq_init(&lock->cll_wq);
401 head = obj->co_lu.lo_header;
402 cfs_atomic_inc(&site->cs_locks_state[CLS_NEW]);
403 cfs_atomic_inc(&site->cs_locks.cs_total);
404 cfs_atomic_inc(&site->cs_locks.cs_created);
405 cl_lock_lockdep_init(lock);
406 cfs_list_for_each_entry(obj, &head->loh_layers,
410 err = obj->co_ops->coo_lock_init(env, obj, lock, io);
412 cl_lock_finish(env, lock);
418 lock = ERR_PTR(-ENOMEM);
423 * Transfer the lock into INTRANSIT state and return the original state.
425 * \pre state: CLS_CACHED, CLS_HELD or CLS_ENQUEUED
426 * \post state: CLS_INTRANSIT
429 enum cl_lock_state cl_lock_intransit(const struct lu_env *env,
430 struct cl_lock *lock)
432 enum cl_lock_state state = lock->cll_state;
434 LASSERT(cl_lock_is_mutexed(lock));
435 LASSERT(state != CLS_INTRANSIT);
436 LASSERTF(state >= CLS_ENQUEUED && state <= CLS_CACHED,
437 "Malformed lock state %d.\n", state);
439 cl_lock_state_set(env, lock, CLS_INTRANSIT);
440 lock->cll_intransit_owner = cfs_current();
441 cl_lock_hold_add(env, lock, "intransit", cfs_current());
444 EXPORT_SYMBOL(cl_lock_intransit);
447 * Exit the intransit state and restore the lock state to the original state
449 void cl_lock_extransit(const struct lu_env *env, struct cl_lock *lock,
450 enum cl_lock_state state)
452 LASSERT(cl_lock_is_mutexed(lock));
453 LASSERT(lock->cll_state == CLS_INTRANSIT);
454 LASSERT(state != CLS_INTRANSIT);
455 LASSERT(lock->cll_intransit_owner == cfs_current());
457 lock->cll_intransit_owner = NULL;
458 cl_lock_state_set(env, lock, state);
459 cl_lock_unhold(env, lock, "intransit", cfs_current());
461 EXPORT_SYMBOL(cl_lock_extransit);
464 * Checking whether the lock is intransit state
466 int cl_lock_is_intransit(struct cl_lock *lock)
468 LASSERT(cl_lock_is_mutexed(lock));
469 return lock->cll_state == CLS_INTRANSIT &&
470 lock->cll_intransit_owner != cfs_current();
472 EXPORT_SYMBOL(cl_lock_is_intransit);
474 * Returns true iff lock is "suitable" for given io. E.g., locks acquired by
475 * truncate and O_APPEND cannot be reused for read/non-append-write, as they
476 * cover multiple stripes and can trigger cascading timeouts.
478 static int cl_lock_fits_into(const struct lu_env *env,
479 const struct cl_lock *lock,
480 const struct cl_lock_descr *need,
481 const struct cl_io *io)
483 const struct cl_lock_slice *slice;
485 LINVRNT(cl_lock_invariant_trusted(env, lock));
487 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
488 if (slice->cls_ops->clo_fits_into != NULL &&
489 !slice->cls_ops->clo_fits_into(env, slice, need, io))
495 static struct cl_lock *cl_lock_lookup(const struct lu_env *env,
496 struct cl_object *obj,
497 const struct cl_io *io,
498 const struct cl_lock_descr *need)
500 struct cl_lock *lock;
501 struct cl_object_header *head;
502 struct cl_site *site;
506 head = cl_object_header(obj);
507 site = cl_object_site(obj);
508 LINVRNT_SPIN_LOCKED(&head->coh_lock_guard);
509 cfs_atomic_inc(&site->cs_locks.cs_lookup);
510 cfs_list_for_each_entry(lock, &head->coh_locks, cll_linkage) {
513 matched = cl_lock_ext_match(&lock->cll_descr, need) &&
514 lock->cll_state < CLS_FREEING &&
515 lock->cll_error == 0 &&
516 !(lock->cll_flags & CLF_CANCELLED) &&
517 cl_lock_fits_into(env, lock, need, io);
518 CDEBUG(D_DLMTRACE, "has: "DDESCR"(%d) need: "DDESCR": %d\n",
519 PDESCR(&lock->cll_descr), lock->cll_state, PDESCR(need),
522 cl_lock_get_trust(lock);
523 cfs_atomic_inc(&cl_object_site(obj)->cs_locks.cs_hit);
531 * Returns a lock matching description \a need.
533 * This is the main entry point into the cl_lock caching interface. First, a
534 * cache (implemented as a per-object linked list) is consulted. If lock is
535 * found there, it is returned immediately. Otherwise new lock is allocated
536 * and returned. In any case, additional reference to lock is acquired.
538 * \see cl_object_find(), cl_page_find()
540 static struct cl_lock *cl_lock_find(const struct lu_env *env,
541 const struct cl_io *io,
542 const struct cl_lock_descr *need)
544 struct cl_object_header *head;
545 struct cl_object *obj;
546 struct cl_lock *lock;
547 struct cl_site *site;
552 head = cl_object_header(obj);
553 site = cl_object_site(obj);
555 cfs_spin_lock(&head->coh_lock_guard);
556 lock = cl_lock_lookup(env, obj, io, need);
557 cfs_spin_unlock(&head->coh_lock_guard);
560 lock = cl_lock_alloc(env, obj, io, need);
562 struct cl_lock *ghost;
564 cfs_spin_lock(&head->coh_lock_guard);
565 ghost = cl_lock_lookup(env, obj, io, need);
567 cfs_list_add_tail(&lock->cll_linkage,
569 cfs_spin_unlock(&head->coh_lock_guard);
570 cfs_atomic_inc(&site->cs_locks.cs_busy);
572 cfs_spin_unlock(&head->coh_lock_guard);
574 * Other threads can acquire references to the
575 * top-lock through its sub-locks. Hence, it
576 * cannot be cl_lock_free()-ed immediately.
578 cl_lock_finish(env, lock);
587 * Returns existing lock matching given description. This is similar to
588 * cl_lock_find() except that no new lock is created, and returned lock is
589 * guaranteed to be in enum cl_lock_state::CLS_HELD state.
591 struct cl_lock *cl_lock_peek(const struct lu_env *env, const struct cl_io *io,
592 const struct cl_lock_descr *need,
593 const char *scope, const void *source)
595 struct cl_object_header *head;
596 struct cl_object *obj;
597 struct cl_lock *lock;
601 head = cl_object_header(obj);
603 cfs_spin_lock(&head->coh_lock_guard);
604 lock = cl_lock_lookup(env, obj, io, need);
605 cfs_spin_unlock(&head->coh_lock_guard);
610 cl_lock_mutex_get(env, lock);
611 if (lock->cll_state == CLS_INTRANSIT)
612 cl_lock_state_wait(env, lock); /* Don't care return value. */
613 if (lock->cll_state == CLS_CACHED) {
615 result = cl_use_try(env, lock, 1);
617 cl_lock_error(env, lock, result);
619 ok = lock->cll_state == CLS_HELD;
621 cl_lock_hold_add(env, lock, scope, source);
622 cl_lock_user_add(env, lock);
623 cl_lock_put(env, lock);
625 cl_lock_mutex_put(env, lock);
627 cl_lock_put(env, lock);
633 EXPORT_SYMBOL(cl_lock_peek);
636 * Returns a slice within a lock, corresponding to the given layer in the
641 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
642 const struct lu_device_type *dtype)
644 const struct cl_lock_slice *slice;
646 LINVRNT(cl_lock_invariant_trusted(NULL, lock));
649 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
650 if (slice->cls_obj->co_lu.lo_dev->ld_type == dtype)
655 EXPORT_SYMBOL(cl_lock_at);
657 static void cl_lock_mutex_tail(const struct lu_env *env, struct cl_lock *lock)
659 struct cl_thread_counters *counters;
661 counters = cl_lock_counters(env, lock);
663 counters->ctc_nr_locks_locked++;
664 lu_ref_add(&counters->ctc_locks_locked, "cll_guard", lock);
665 cl_lock_trace(D_TRACE, env, "got mutex", lock);
669 * Locks cl_lock object.
671 * This is used to manipulate cl_lock fields, and to serialize state
672 * transitions in the lock state machine.
674 * \post cl_lock_is_mutexed(lock)
676 * \see cl_lock_mutex_put()
678 void cl_lock_mutex_get(const struct lu_env *env, struct cl_lock *lock)
680 LINVRNT(cl_lock_invariant(env, lock));
682 if (lock->cll_guarder == cfs_current()) {
683 LINVRNT(cl_lock_is_mutexed(lock));
684 LINVRNT(lock->cll_depth > 0);
686 struct cl_object_header *hdr;
687 struct cl_thread_info *info;
690 LINVRNT(lock->cll_guarder != cfs_current());
691 hdr = cl_object_header(lock->cll_descr.cld_obj);
693 * Check that mutices are taken in the bottom-to-top order.
695 info = cl_env_info(env);
696 for (i = 0; i < hdr->coh_nesting; ++i)
697 LASSERT(info->clt_counters[i].ctc_nr_locks_locked == 0);
698 cfs_mutex_lock_nested(&lock->cll_guard, hdr->coh_nesting);
699 lock->cll_guarder = cfs_current();
700 LINVRNT(lock->cll_depth == 0);
702 cl_lock_mutex_tail(env, lock);
704 EXPORT_SYMBOL(cl_lock_mutex_get);
707 * Try-locks cl_lock object.
709 * \retval 0 \a lock was successfully locked
711 * \retval -EBUSY \a lock cannot be locked right now
713 * \post ergo(result == 0, cl_lock_is_mutexed(lock))
715 * \see cl_lock_mutex_get()
717 int cl_lock_mutex_try(const struct lu_env *env, struct cl_lock *lock)
721 LINVRNT(cl_lock_invariant_trusted(env, lock));
725 if (lock->cll_guarder == cfs_current()) {
726 LINVRNT(lock->cll_depth > 0);
727 cl_lock_mutex_tail(env, lock);
728 } else if (cfs_mutex_trylock(&lock->cll_guard)) {
729 LINVRNT(lock->cll_depth == 0);
730 lock->cll_guarder = cfs_current();
731 cl_lock_mutex_tail(env, lock);
736 EXPORT_SYMBOL(cl_lock_mutex_try);
739 {* Unlocks cl_lock object.
741 * \pre cl_lock_is_mutexed(lock)
743 * \see cl_lock_mutex_get()
745 void cl_lock_mutex_put(const struct lu_env *env, struct cl_lock *lock)
747 struct cl_thread_counters *counters;
749 LINVRNT(cl_lock_invariant(env, lock));
750 LINVRNT(cl_lock_is_mutexed(lock));
751 LINVRNT(lock->cll_guarder == cfs_current());
752 LINVRNT(lock->cll_depth > 0);
754 counters = cl_lock_counters(env, lock);
755 LINVRNT(counters->ctc_nr_locks_locked > 0);
757 cl_lock_trace(D_TRACE, env, "put mutex", lock);
758 lu_ref_del(&counters->ctc_locks_locked, "cll_guard", lock);
759 counters->ctc_nr_locks_locked--;
760 if (--lock->cll_depth == 0) {
761 lock->cll_guarder = NULL;
762 cfs_mutex_unlock(&lock->cll_guard);
765 EXPORT_SYMBOL(cl_lock_mutex_put);
768 * Returns true iff lock's mutex is owned by the current thread.
770 int cl_lock_is_mutexed(struct cl_lock *lock)
772 return lock->cll_guarder == cfs_current();
774 EXPORT_SYMBOL(cl_lock_is_mutexed);
777 * Returns number of cl_lock mutices held by the current thread (environment).
779 int cl_lock_nr_mutexed(const struct lu_env *env)
781 struct cl_thread_info *info;
786 * NOTE: if summation across all nesting levels (currently 2) proves
787 * too expensive, a summary counter can be added to
788 * struct cl_thread_info.
790 info = cl_env_info(env);
791 for (i = 0, locked = 0; i < ARRAY_SIZE(info->clt_counters); ++i)
792 locked += info->clt_counters[i].ctc_nr_locks_locked;
795 EXPORT_SYMBOL(cl_lock_nr_mutexed);
797 static void cl_lock_cancel0(const struct lu_env *env, struct cl_lock *lock)
799 LINVRNT(cl_lock_is_mutexed(lock));
800 LINVRNT(cl_lock_invariant(env, lock));
802 if (!(lock->cll_flags & CLF_CANCELLED)) {
803 const struct cl_lock_slice *slice;
805 lock->cll_flags |= CLF_CANCELLED;
806 cfs_list_for_each_entry_reverse(slice, &lock->cll_layers,
808 if (slice->cls_ops->clo_cancel != NULL)
809 slice->cls_ops->clo_cancel(env, slice);
815 static void cl_lock_delete0(const struct lu_env *env, struct cl_lock *lock)
817 struct cl_object_header *head;
818 const struct cl_lock_slice *slice;
820 LINVRNT(cl_lock_is_mutexed(lock));
821 LINVRNT(cl_lock_invariant(env, lock));
824 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 cfs_spin_lock(&head->coh_lock_guard);
831 cfs_list_del_init(&lock->cll_linkage);
833 cfs_spin_unlock(&head->coh_lock_guard);
835 * From now on, no new references to this lock can be acquired
836 * by cl_lock_lookup().
838 cfs_list_for_each_entry_reverse(slice, &lock->cll_layers,
840 if (slice->cls_ops->clo_delete != NULL)
841 slice->cls_ops->clo_delete(env, slice);
844 * From now on, no new references to this lock can be acquired
845 * by layer-specific means (like a pointer from struct
846 * ldlm_lock in osc, or a pointer from top-lock to sub-lock in
849 * Lock will be finally freed in cl_lock_put() when last of
850 * existing references goes away.
857 * Mod(ifie)s cl_lock::cll_holds counter for a given lock. Also, for a
858 * top-lock (nesting == 0) accounts for this modification in the per-thread
859 * debugging counters. Sub-lock holds can be released by a thread different
860 * from one that acquired it.
862 static void cl_lock_hold_mod(const struct lu_env *env, struct cl_lock *lock,
865 struct cl_thread_counters *counters;
866 enum clt_nesting_level nesting;
868 lock->cll_holds += delta;
869 nesting = cl_lock_nesting(lock);
870 if (nesting == CNL_TOP) {
871 counters = &cl_env_info(env)->clt_counters[CNL_TOP];
872 counters->ctc_nr_held += delta;
873 LASSERT(counters->ctc_nr_held >= 0);
878 * Mod(ifie)s cl_lock::cll_users counter for a given lock. See
879 * cl_lock_hold_mod() for the explanation of the debugging code.
881 static void cl_lock_used_mod(const struct lu_env *env, struct cl_lock *lock,
884 struct cl_thread_counters *counters;
885 enum clt_nesting_level nesting;
887 lock->cll_users += delta;
888 nesting = cl_lock_nesting(lock);
889 if (nesting == CNL_TOP) {
890 counters = &cl_env_info(env)->clt_counters[CNL_TOP];
891 counters->ctc_nr_used += delta;
892 LASSERT(counters->ctc_nr_used >= 0);
896 static void cl_lock_hold_release(const struct lu_env *env, struct cl_lock *lock,
897 const char *scope, const void *source)
899 LINVRNT(cl_lock_is_mutexed(lock));
900 LINVRNT(cl_lock_invariant(env, lock));
901 LASSERT(lock->cll_holds > 0);
904 cl_lock_trace(D_DLMTRACE, env, "hold release lock", lock);
905 lu_ref_del(&lock->cll_holders, scope, source);
906 cl_lock_hold_mod(env, lock, -1);
907 if (lock->cll_holds == 0) {
908 if (lock->cll_descr.cld_mode == CLM_PHANTOM ||
909 lock->cll_descr.cld_mode == CLM_GROUP)
911 * If lock is still phantom or grouplock when user is
912 * done with it---destroy the lock.
914 lock->cll_flags |= CLF_CANCELPEND|CLF_DOOMED;
915 if (lock->cll_flags & CLF_CANCELPEND) {
916 lock->cll_flags &= ~CLF_CANCELPEND;
917 cl_lock_cancel0(env, lock);
919 if (lock->cll_flags & CLF_DOOMED) {
920 /* no longer doomed: it's dead... Jim. */
921 lock->cll_flags &= ~CLF_DOOMED;
922 cl_lock_delete0(env, lock);
930 * Waits until lock state is changed.
932 * This function is called with cl_lock mutex locked, atomically releases
933 * mutex and goes to sleep, waiting for a lock state change (signaled by
934 * cl_lock_signal()), and re-acquires the mutex before return.
936 * This function is used to wait until lock state machine makes some progress
937 * and to emulate synchronous operations on top of asynchronous lock
940 * \retval -EINTR wait was interrupted
942 * \retval 0 wait wasn't interrupted
944 * \pre cl_lock_is_mutexed(lock)
946 * \see cl_lock_signal()
948 int cl_lock_state_wait(const struct lu_env *env, struct cl_lock *lock)
950 cfs_waitlink_t waiter;
951 cfs_sigset_t blocked;
955 LINVRNT(cl_lock_is_mutexed(lock));
956 LINVRNT(cl_lock_invariant(env, lock));
957 LASSERT(lock->cll_depth == 1);
958 LASSERT(lock->cll_state != CLS_FREEING); /* too late to wait */
960 cl_lock_trace(D_DLMTRACE, env, "state wait lock", lock);
961 result = lock->cll_error;
963 /* To avoid being interrupted by the 'non-fatal' signals
964 * (SIGCHLD, for instance), we'd block them temporarily.
966 blocked = cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
968 cfs_waitlink_init(&waiter);
969 cfs_waitq_add(&lock->cll_wq, &waiter);
970 cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
971 cl_lock_mutex_put(env, lock);
973 LASSERT(cl_lock_nr_mutexed(env) == 0);
974 cfs_waitq_wait(&waiter, CFS_TASK_INTERRUPTIBLE);
976 cl_lock_mutex_get(env, lock);
977 cfs_set_current_state(CFS_TASK_RUNNING);
978 cfs_waitq_del(&lock->cll_wq, &waiter);
979 result = cfs_signal_pending() ? -EINTR : 0;
981 /* Restore old blocked signals */
982 cfs_restore_sigs(blocked);
986 EXPORT_SYMBOL(cl_lock_state_wait);
988 static void cl_lock_state_signal(const struct lu_env *env, struct cl_lock *lock,
989 enum cl_lock_state state)
991 const struct cl_lock_slice *slice;
994 LINVRNT(cl_lock_is_mutexed(lock));
995 LINVRNT(cl_lock_invariant(env, lock));
997 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage)
998 if (slice->cls_ops->clo_state != NULL)
999 slice->cls_ops->clo_state(env, slice, state);
1000 cfs_waitq_broadcast(&lock->cll_wq);
1005 * Notifies waiters that lock state changed.
1007 * Wakes up all waiters sleeping in cl_lock_state_wait(), also notifies all
1008 * layers about state change by calling cl_lock_operations::clo_state()
1011 void cl_lock_signal(const struct lu_env *env, struct cl_lock *lock)
1014 cl_lock_trace(D_DLMTRACE, env, "state signal lock", lock);
1015 cl_lock_state_signal(env, lock, lock->cll_state);
1018 EXPORT_SYMBOL(cl_lock_signal);
1021 * Changes lock state.
1023 * This function is invoked to notify layers that lock state changed, possible
1024 * as a result of an asynchronous event such as call-back reception.
1026 * \post lock->cll_state == state
1028 * \see cl_lock_operations::clo_state()
1030 void cl_lock_state_set(const struct lu_env *env, struct cl_lock *lock,
1031 enum cl_lock_state state)
1033 struct cl_site *site = cl_object_site(lock->cll_descr.cld_obj);
1036 LASSERT(lock->cll_state <= state ||
1037 (lock->cll_state == CLS_CACHED &&
1038 (state == CLS_HELD || /* lock found in cache */
1039 state == CLS_NEW || /* sub-lock canceled */
1040 state == CLS_INTRANSIT)) ||
1041 /* lock is in transit state */
1042 lock->cll_state == CLS_INTRANSIT);
1044 if (lock->cll_state != state) {
1045 cfs_atomic_dec(&site->cs_locks_state[lock->cll_state]);
1046 cfs_atomic_inc(&site->cs_locks_state[state]);
1048 cl_lock_state_signal(env, lock, state);
1049 lock->cll_state = state;
1053 EXPORT_SYMBOL(cl_lock_state_set);
1055 static int cl_unuse_try_internal(const struct lu_env *env, struct cl_lock *lock)
1057 const struct cl_lock_slice *slice;
1063 LINVRNT(cl_lock_is_mutexed(lock));
1064 LINVRNT(cl_lock_invariant(env, lock));
1065 LASSERT(lock->cll_state == CLS_INTRANSIT);
1068 cfs_list_for_each_entry_reverse(slice, &lock->cll_layers,
1070 if (slice->cls_ops->clo_unuse != NULL) {
1071 result = slice->cls_ops->clo_unuse(env, slice);
1076 LASSERT(result != -ENOSYS);
1077 } while (result == CLO_REPEAT);
1083 * Yanks lock from the cache (cl_lock_state::CLS_CACHED state) by calling
1084 * cl_lock_operations::clo_use() top-to-bottom to notify layers.
1085 * @atomic = 1, it must unuse the lock to recovery the lock to keep the
1086 * use process atomic
1088 int cl_use_try(const struct lu_env *env, struct cl_lock *lock, int atomic)
1090 const struct cl_lock_slice *slice;
1092 enum cl_lock_state state;
1095 cl_lock_trace(D_DLMTRACE, env, "use lock", lock);
1097 LASSERT(lock->cll_state == CLS_CACHED);
1098 if (lock->cll_error)
1099 RETURN(lock->cll_error);
1102 state = cl_lock_intransit(env, lock);
1103 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1104 if (slice->cls_ops->clo_use != NULL) {
1105 result = slice->cls_ops->clo_use(env, slice);
1110 LASSERT(result != -ENOSYS);
1112 LASSERTF(lock->cll_state == CLS_INTRANSIT, "Wrong state %d.\n",
1118 if (result == -ESTALE) {
1120 * ESTALE means sublock being cancelled
1121 * at this time, and set lock state to
1122 * be NEW here and ask the caller to repeat.
1125 result = CLO_REPEAT;
1128 /* @atomic means back-off-on-failure. */
1131 rc = cl_unuse_try_internal(env, lock);
1132 /* Vet the results. */
1133 if (rc < 0 && result > 0)
1138 cl_lock_extransit(env, lock, state);
1141 EXPORT_SYMBOL(cl_use_try);
1144 * Helper for cl_enqueue_try() that calls ->clo_enqueue() across all layers
1147 static int cl_enqueue_kick(const struct lu_env *env,
1148 struct cl_lock *lock,
1149 struct cl_io *io, __u32 flags)
1152 const struct cl_lock_slice *slice;
1156 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1157 if (slice->cls_ops->clo_enqueue != NULL) {
1158 result = slice->cls_ops->clo_enqueue(env,
1164 LASSERT(result != -ENOSYS);
1169 * Tries to enqueue a lock.
1171 * This function is called repeatedly by cl_enqueue() until either lock is
1172 * enqueued, or error occurs. This function does not block waiting for
1173 * networking communication to complete.
1175 * \post ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1176 * lock->cll_state == CLS_HELD)
1178 * \see cl_enqueue() cl_lock_operations::clo_enqueue()
1179 * \see cl_lock_state::CLS_ENQUEUED
1181 int cl_enqueue_try(const struct lu_env *env, struct cl_lock *lock,
1182 struct cl_io *io, __u32 flags)
1187 cl_lock_trace(D_DLMTRACE, env, "enqueue lock", lock);
1191 LINVRNT(cl_lock_is_mutexed(lock));
1193 if (lock->cll_error != 0)
1195 switch (lock->cll_state) {
1197 cl_lock_state_set(env, lock, CLS_QUEUING);
1201 result = cl_enqueue_kick(env, lock, io, flags);
1203 cl_lock_state_set(env, lock, CLS_ENQUEUED);
1206 LASSERT(cl_lock_is_intransit(lock));
1210 /* yank lock from the cache. */
1211 result = cl_use_try(env, lock, 0);
1220 * impossible, only held locks with increased
1221 * ->cll_holds can be enqueued, and they cannot be
1226 } while (result == CLO_REPEAT);
1228 cl_lock_error(env, lock, result);
1229 RETURN(result ?: lock->cll_error);
1231 EXPORT_SYMBOL(cl_enqueue_try);
1234 * Cancel the conflicting lock found during previous enqueue.
1236 * \retval 0 conflicting lock has been canceled.
1237 * \retval -ve error code.
1239 int cl_lock_enqueue_wait(const struct lu_env *env,
1240 struct cl_lock *lock,
1243 struct cl_lock *conflict;
1247 LASSERT(cl_lock_is_mutexed(lock));
1248 LASSERT(lock->cll_state == CLS_QUEUING);
1249 LASSERT(lock->cll_conflict != NULL);
1251 conflict = lock->cll_conflict;
1252 lock->cll_conflict = NULL;
1254 cl_lock_mutex_put(env, lock);
1255 LASSERT(cl_lock_nr_mutexed(env) == 0);
1257 cl_lock_mutex_get(env, conflict);
1258 cl_lock_cancel(env, conflict);
1259 cl_lock_delete(env, conflict);
1261 while (conflict->cll_state != CLS_FREEING) {
1262 rc = cl_lock_state_wait(env, conflict);
1266 cl_lock_mutex_put(env, conflict);
1267 lu_ref_del(&conflict->cll_reference, "cancel-wait", lock);
1268 cl_lock_put(env, conflict);
1271 cl_lock_mutex_get(env, lock);
1276 EXPORT_SYMBOL(cl_lock_enqueue_wait);
1278 static int cl_enqueue_locked(const struct lu_env *env, struct cl_lock *lock,
1279 struct cl_io *io, __u32 enqflags)
1285 LINVRNT(cl_lock_is_mutexed(lock));
1286 LINVRNT(cl_lock_invariant(env, lock));
1287 LASSERT(lock->cll_holds > 0);
1289 cl_lock_user_add(env, lock);
1291 result = cl_enqueue_try(env, lock, io, enqflags);
1292 if (result == CLO_WAIT) {
1293 if (lock->cll_conflict != NULL)
1294 result = cl_lock_enqueue_wait(env, lock, 1);
1296 result = cl_lock_state_wait(env, lock);
1303 cl_lock_user_del(env, lock);
1304 cl_lock_error(env, lock, result);
1306 LASSERT(ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1307 lock->cll_state == CLS_HELD));
1314 * \pre current thread or io owns a hold on lock.
1316 * \post ergo(result == 0, lock->users increased)
1317 * \post ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1318 * lock->cll_state == CLS_HELD)
1320 int cl_enqueue(const struct lu_env *env, struct cl_lock *lock,
1321 struct cl_io *io, __u32 enqflags)
1327 cl_lock_lockdep_acquire(env, lock, enqflags);
1328 cl_lock_mutex_get(env, lock);
1329 result = cl_enqueue_locked(env, lock, io, enqflags);
1330 cl_lock_mutex_put(env, lock);
1332 cl_lock_lockdep_release(env, lock);
1333 LASSERT(ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1334 lock->cll_state == CLS_HELD));
1337 EXPORT_SYMBOL(cl_enqueue);
1340 * Tries to unlock a lock.
1342 * This function is called repeatedly by cl_unuse() until either lock is
1343 * unlocked, or error occurs.
1344 * cl_unuse_try is a one-shot operation, so it must NOT return CLO_WAIT.
1346 * \pre lock->cll_state == CLS_HELD
1348 * \post ergo(result == 0, lock->cll_state == CLS_CACHED)
1350 * \see cl_unuse() cl_lock_operations::clo_unuse()
1351 * \see cl_lock_state::CLS_CACHED
1353 int cl_unuse_try(const struct lu_env *env, struct cl_lock *lock)
1356 enum cl_lock_state state = CLS_NEW;
1359 cl_lock_trace(D_DLMTRACE, env, "unuse lock", lock);
1361 LASSERT(lock->cll_state == CLS_HELD || lock->cll_state == CLS_ENQUEUED);
1362 if (lock->cll_users > 1) {
1363 cl_lock_user_del(env, lock);
1368 * New lock users (->cll_users) are not protecting unlocking
1369 * from proceeding. From this point, lock eventually reaches
1370 * CLS_CACHED, is reinitialized to CLS_NEW or fails into
1373 state = cl_lock_intransit(env, lock);
1375 result = cl_unuse_try_internal(env, lock);
1376 LASSERT(lock->cll_state == CLS_INTRANSIT);
1377 LASSERT(result != CLO_WAIT);
1378 cl_lock_user_del(env, lock);
1379 if (result == 0 || result == -ESTALE) {
1381 * Return lock back to the cache. This is the only
1382 * place where lock is moved into CLS_CACHED state.
1384 * If one of ->clo_unuse() methods returned -ESTALE, lock
1385 * cannot be placed into cache and has to be
1386 * re-initialized. This happens e.g., when a sub-lock was
1387 * canceled while unlocking was in progress.
1389 if (state == CLS_HELD && result == 0)
1393 cl_lock_extransit(env, lock, state);
1396 * Hide -ESTALE error.
1397 * If the lock is a glimpse lock, and it has multiple
1398 * stripes. Assuming that one of its sublock returned -ENAVAIL,
1399 * and other sublocks are matched write locks. In this case,
1400 * we can't set this lock to error because otherwise some of
1401 * its sublocks may not be canceled. This causes some dirty
1402 * pages won't be written to OSTs. -jay
1406 CERROR("result = %d, this is unlikely!\n", result);
1407 cl_lock_extransit(env, lock, state);
1410 result = result ?: lock->cll_error;
1412 cl_lock_error(env, lock, result);
1415 EXPORT_SYMBOL(cl_unuse_try);
1417 static void cl_unuse_locked(const struct lu_env *env, struct cl_lock *lock)
1422 result = cl_unuse_try(env, lock);
1424 CL_LOCK_DEBUG(D_ERROR, env, lock, "unuse return %d\n", result);
1432 void cl_unuse(const struct lu_env *env, struct cl_lock *lock)
1435 cl_lock_mutex_get(env, lock);
1436 cl_unuse_locked(env, lock);
1437 cl_lock_mutex_put(env, lock);
1438 cl_lock_lockdep_release(env, lock);
1441 EXPORT_SYMBOL(cl_unuse);
1444 * Tries to wait for a lock.
1446 * This function is called repeatedly by cl_wait() until either lock is
1447 * granted, or error occurs. This function does not block waiting for network
1448 * communication to complete.
1450 * \see cl_wait() cl_lock_operations::clo_wait()
1451 * \see cl_lock_state::CLS_HELD
1453 int cl_wait_try(const struct lu_env *env, struct cl_lock *lock)
1455 const struct cl_lock_slice *slice;
1459 cl_lock_trace(D_DLMTRACE, env, "wait lock try", lock);
1461 LINVRNT(cl_lock_is_mutexed(lock));
1462 LINVRNT(cl_lock_invariant(env, lock));
1463 LASSERT(lock->cll_state == CLS_ENQUEUED ||
1464 lock->cll_state == CLS_HELD ||
1465 lock->cll_state == CLS_INTRANSIT);
1466 LASSERT(lock->cll_users > 0);
1467 LASSERT(lock->cll_holds > 0);
1470 if (lock->cll_error != 0)
1473 if (cl_lock_is_intransit(lock)) {
1478 if (lock->cll_state == CLS_HELD)
1483 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1484 if (slice->cls_ops->clo_wait != NULL) {
1485 result = slice->cls_ops->clo_wait(env, slice);
1490 LASSERT(result != -ENOSYS);
1492 LASSERT(lock->cll_state != CLS_INTRANSIT);
1493 cl_lock_state_set(env, lock, CLS_HELD);
1495 } while (result == CLO_REPEAT);
1496 RETURN(result ?: lock->cll_error);
1498 EXPORT_SYMBOL(cl_wait_try);
1501 * Waits until enqueued lock is granted.
1503 * \pre current thread or io owns a hold on the lock
1504 * \pre ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1505 * lock->cll_state == CLS_HELD)
1507 * \post ergo(result == 0, lock->cll_state == CLS_HELD)
1509 int cl_wait(const struct lu_env *env, struct cl_lock *lock)
1514 cl_lock_mutex_get(env, lock);
1516 LINVRNT(cl_lock_invariant(env, lock));
1517 LASSERTF(lock->cll_state == CLS_ENQUEUED || lock->cll_state == CLS_HELD,
1518 "Wrong state %d \n", lock->cll_state);
1519 LASSERT(lock->cll_holds > 0);
1522 result = cl_wait_try(env, lock);
1523 if (result == CLO_WAIT) {
1524 result = cl_lock_state_wait(env, lock);
1531 cl_lock_user_del(env, lock);
1532 cl_lock_error(env, lock, result);
1533 cl_lock_lockdep_release(env, lock);
1535 cl_lock_trace(D_DLMTRACE, env, "wait lock", lock);
1536 cl_lock_mutex_put(env, lock);
1537 LASSERT(ergo(result == 0, lock->cll_state == CLS_HELD));
1540 EXPORT_SYMBOL(cl_wait);
1543 * Executes cl_lock_operations::clo_weigh(), and sums results to estimate lock
1546 unsigned long cl_lock_weigh(const struct lu_env *env, struct cl_lock *lock)
1548 const struct cl_lock_slice *slice;
1549 unsigned long pound;
1550 unsigned long ounce;
1553 LINVRNT(cl_lock_is_mutexed(lock));
1554 LINVRNT(cl_lock_invariant(env, lock));
1557 cfs_list_for_each_entry_reverse(slice, &lock->cll_layers, cls_linkage) {
1558 if (slice->cls_ops->clo_weigh != NULL) {
1559 ounce = slice->cls_ops->clo_weigh(env, slice);
1561 if (pound < ounce) /* over-weight^Wflow */
1567 EXPORT_SYMBOL(cl_lock_weigh);
1570 * Notifies layers that lock description changed.
1572 * The server can grant client a lock different from one that was requested
1573 * (e.g., larger in extent). This method is called when actually granted lock
1574 * description becomes known to let layers to accommodate for changed lock
1577 * \see cl_lock_operations::clo_modify()
1579 int cl_lock_modify(const struct lu_env *env, struct cl_lock *lock,
1580 const struct cl_lock_descr *desc)
1582 const struct cl_lock_slice *slice;
1583 struct cl_object *obj = lock->cll_descr.cld_obj;
1584 struct cl_object_header *hdr = cl_object_header(obj);
1588 cl_lock_trace(D_DLMTRACE, env, "modify lock", lock);
1589 /* don't allow object to change */
1590 LASSERT(obj == desc->cld_obj);
1591 LINVRNT(cl_lock_is_mutexed(lock));
1592 LINVRNT(cl_lock_invariant(env, lock));
1594 cfs_list_for_each_entry_reverse(slice, &lock->cll_layers, cls_linkage) {
1595 if (slice->cls_ops->clo_modify != NULL) {
1596 result = slice->cls_ops->clo_modify(env, slice, desc);
1601 CL_LOCK_DEBUG(D_DLMTRACE, env, lock, " -> "DDESCR"@"DFID"\n",
1602 PDESCR(desc), PFID(lu_object_fid(&desc->cld_obj->co_lu)));
1604 * Just replace description in place. Nothing more is needed for
1605 * now. If locks were indexed according to their extent and/or mode,
1606 * that index would have to be updated here.
1608 cfs_spin_lock(&hdr->coh_lock_guard);
1609 lock->cll_descr = *desc;
1610 cfs_spin_unlock(&hdr->coh_lock_guard);
1613 EXPORT_SYMBOL(cl_lock_modify);
1616 * Initializes lock closure with a given origin.
1618 * \see cl_lock_closure
1620 void cl_lock_closure_init(const struct lu_env *env,
1621 struct cl_lock_closure *closure,
1622 struct cl_lock *origin, int wait)
1624 LINVRNT(cl_lock_is_mutexed(origin));
1625 LINVRNT(cl_lock_invariant(env, origin));
1627 CFS_INIT_LIST_HEAD(&closure->clc_list);
1628 closure->clc_origin = origin;
1629 closure->clc_wait = wait;
1630 closure->clc_nr = 0;
1632 EXPORT_SYMBOL(cl_lock_closure_init);
1635 * Builds a closure of \a lock.
1637 * Building of a closure consists of adding initial lock (\a lock) into it,
1638 * and calling cl_lock_operations::clo_closure() methods of \a lock. These
1639 * methods might call cl_lock_closure_build() recursively again, adding more
1640 * locks to the closure, etc.
1642 * \see cl_lock_closure
1644 int cl_lock_closure_build(const struct lu_env *env, struct cl_lock *lock,
1645 struct cl_lock_closure *closure)
1647 const struct cl_lock_slice *slice;
1651 LINVRNT(cl_lock_is_mutexed(closure->clc_origin));
1652 LINVRNT(cl_lock_invariant(env, closure->clc_origin));
1654 result = cl_lock_enclosure(env, lock, closure);
1656 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1657 if (slice->cls_ops->clo_closure != NULL) {
1658 result = slice->cls_ops->clo_closure(env, slice,
1666 cl_lock_disclosure(env, closure);
1669 EXPORT_SYMBOL(cl_lock_closure_build);
1672 * Adds new lock to a closure.
1674 * Try-locks \a lock and if succeeded, adds it to the closure (never more than
1675 * once). If try-lock failed, returns CLO_REPEAT, after optionally waiting
1676 * until next try-lock is likely to succeed.
1678 int cl_lock_enclosure(const struct lu_env *env, struct cl_lock *lock,
1679 struct cl_lock_closure *closure)
1683 cl_lock_trace(D_DLMTRACE, env, "enclosure lock", lock);
1684 if (!cl_lock_mutex_try(env, lock)) {
1686 * If lock->cll_inclosure is not empty, lock is already in
1689 if (cfs_list_empty(&lock->cll_inclosure)) {
1690 cl_lock_get_trust(lock);
1691 lu_ref_add(&lock->cll_reference, "closure", closure);
1692 cfs_list_add(&lock->cll_inclosure, &closure->clc_list);
1695 cl_lock_mutex_put(env, lock);
1698 cl_lock_disclosure(env, closure);
1699 if (closure->clc_wait) {
1700 cl_lock_get_trust(lock);
1701 lu_ref_add(&lock->cll_reference, "closure-w", closure);
1702 cl_lock_mutex_put(env, closure->clc_origin);
1704 LASSERT(cl_lock_nr_mutexed(env) == 0);
1705 cl_lock_mutex_get(env, lock);
1706 cl_lock_mutex_put(env, lock);
1708 cl_lock_mutex_get(env, closure->clc_origin);
1709 lu_ref_del(&lock->cll_reference, "closure-w", closure);
1710 cl_lock_put(env, lock);
1712 result = CLO_REPEAT;
1716 EXPORT_SYMBOL(cl_lock_enclosure);
1718 /** Releases mutices of enclosed locks. */
1719 void cl_lock_disclosure(const struct lu_env *env,
1720 struct cl_lock_closure *closure)
1722 struct cl_lock *scan;
1723 struct cl_lock *temp;
1725 cl_lock_trace(D_DLMTRACE, env, "disclosure lock", closure->clc_origin);
1726 cfs_list_for_each_entry_safe(scan, temp, &closure->clc_list,
1728 cfs_list_del_init(&scan->cll_inclosure);
1729 cl_lock_mutex_put(env, scan);
1730 lu_ref_del(&scan->cll_reference, "closure", closure);
1731 cl_lock_put(env, scan);
1734 LASSERT(closure->clc_nr == 0);
1736 EXPORT_SYMBOL(cl_lock_disclosure);
1738 /** Finalizes a closure. */
1739 void cl_lock_closure_fini(struct cl_lock_closure *closure)
1741 LASSERT(closure->clc_nr == 0);
1742 LASSERT(cfs_list_empty(&closure->clc_list));
1744 EXPORT_SYMBOL(cl_lock_closure_fini);
1747 * Destroys this lock. Notifies layers (bottom-to-top) that lock is being
1748 * destroyed, then destroy the lock. If there are holds on the lock, postpone
1749 * destruction until all holds are released. This is called when a decision is
1750 * made to destroy the lock in the future. E.g., when a blocking AST is
1751 * received on it, or fatal communication error happens.
1753 * Caller must have a reference on this lock to prevent a situation, when
1754 * deleted lock lingers in memory for indefinite time, because nobody calls
1755 * cl_lock_put() to finish it.
1757 * \pre atomic_read(&lock->cll_ref) > 0
1758 * \pre ergo(cl_lock_nesting(lock) == CNL_TOP,
1759 * cl_lock_nr_mutexed(env) == 1)
1760 * [i.e., if a top-lock is deleted, mutices of no other locks can be
1761 * held, as deletion of sub-locks might require releasing a top-lock
1764 * \see cl_lock_operations::clo_delete()
1765 * \see cl_lock::cll_holds
1767 void cl_lock_delete(const struct lu_env *env, struct cl_lock *lock)
1769 LINVRNT(cl_lock_is_mutexed(lock));
1770 LINVRNT(cl_lock_invariant(env, lock));
1771 LASSERT(ergo(cl_lock_nesting(lock) == CNL_TOP,
1772 cl_lock_nr_mutexed(env) == 1));
1775 cl_lock_trace(D_DLMTRACE, env, "delete lock", lock);
1776 if (lock->cll_holds == 0)
1777 cl_lock_delete0(env, lock);
1779 lock->cll_flags |= CLF_DOOMED;
1782 EXPORT_SYMBOL(cl_lock_delete);
1785 * Mark lock as irrecoverably failed, and mark it for destruction. This
1786 * happens when, e.g., server fails to grant a lock to us, or networking
1789 * \pre atomic_read(&lock->cll_ref) > 0
1791 * \see clo_lock_delete()
1792 * \see cl_lock::cll_holds
1794 void cl_lock_error(const struct lu_env *env, struct cl_lock *lock, int error)
1796 LINVRNT(cl_lock_is_mutexed(lock));
1797 LINVRNT(cl_lock_invariant(env, lock));
1800 cl_lock_trace(D_DLMTRACE, env, "set lock error", lock);
1801 if (lock->cll_error == 0 && error != 0) {
1802 lock->cll_error = error;
1803 cl_lock_signal(env, lock);
1804 cl_lock_cancel(env, lock);
1805 cl_lock_delete(env, lock);
1809 EXPORT_SYMBOL(cl_lock_error);
1812 * Cancels this lock. Notifies layers
1813 * (bottom-to-top) that lock is being cancelled, then destroy the lock. If
1814 * there are holds on the lock, postpone cancellation until
1815 * all holds are released.
1817 * Cancellation notification is delivered to layers at most once.
1819 * \see cl_lock_operations::clo_cancel()
1820 * \see cl_lock::cll_holds
1822 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock)
1824 LINVRNT(cl_lock_is_mutexed(lock));
1825 LINVRNT(cl_lock_invariant(env, lock));
1828 cl_lock_trace(D_DLMTRACE, env, "cancel lock", lock);
1829 if (lock->cll_holds == 0)
1830 cl_lock_cancel0(env, lock);
1832 lock->cll_flags |= CLF_CANCELPEND;
1835 EXPORT_SYMBOL(cl_lock_cancel);
1838 * Finds an existing lock covering given page and optionally different from a
1839 * given \a except lock.
1841 struct cl_lock *cl_lock_at_page(const struct lu_env *env, struct cl_object *obj,
1842 struct cl_page *page, struct cl_lock *except,
1843 int pending, int canceld)
1845 struct cl_object_header *head;
1846 struct cl_lock *scan;
1847 struct cl_lock *lock;
1848 struct cl_lock_descr *need;
1852 head = cl_object_header(obj);
1853 need = &cl_env_info(env)->clt_descr;
1856 need->cld_mode = CLM_READ; /* CLM_READ matches both READ & WRITE, but
1858 need->cld_start = need->cld_end = page->cp_index;
1859 need->cld_enq_flags = 0;
1861 cfs_spin_lock(&head->coh_lock_guard);
1862 /* It is fine to match any group lock since there could be only one
1863 * with a uniq gid and it conflicts with all other lock modes too */
1864 cfs_list_for_each_entry(scan, &head->coh_locks, cll_linkage) {
1865 if (scan != except &&
1866 (scan->cll_descr.cld_mode == CLM_GROUP ||
1867 cl_lock_ext_match(&scan->cll_descr, need)) &&
1868 scan->cll_state >= CLS_HELD &&
1869 scan->cll_state < CLS_FREEING &&
1871 * This check is racy as the lock can be canceled right
1872 * after it is done, but this is fine, because page exists
1875 (canceld || !(scan->cll_flags & CLF_CANCELLED)) &&
1876 (pending || !(scan->cll_flags & CLF_CANCELPEND))) {
1877 /* Don't increase cs_hit here since this
1878 * is just a helper function. */
1879 cl_lock_get_trust(scan);
1884 cfs_spin_unlock(&head->coh_lock_guard);
1887 EXPORT_SYMBOL(cl_lock_at_page);
1890 * Returns a list of pages protected (only) by a given lock.
1892 * Scans an extent of page radix tree, corresponding to the \a lock and queues
1893 * all pages that are not protected by locks other than \a lock into \a queue.
1895 void cl_lock_page_list_fixup(const struct lu_env *env,
1896 struct cl_io *io, struct cl_lock *lock,
1897 struct cl_page_list *queue)
1899 struct cl_page *page;
1900 struct cl_page *temp;
1901 struct cl_page_list *plist = &cl_env_info(env)->clt_list;
1903 LINVRNT(cl_lock_invariant(env, lock));
1906 /* No need to fix for WRITE lock because it is exclusive. */
1907 if (lock->cll_descr.cld_mode >= CLM_WRITE)
1910 /* For those pages who are still covered by other PR locks, we should
1911 * not discard them otherwise a [0, EOF) PR lock will discard all
1914 cl_page_list_init(plist);
1915 cl_page_list_for_each_safe(page, temp, queue) {
1916 pgoff_t idx = page->cp_index;
1917 struct cl_lock *found;
1918 struct cl_lock_descr *descr;
1920 /* The algorithm counts on the index-ascending page index. */
1921 LASSERT(ergo(&temp->cp_batch != &queue->pl_pages,
1922 page->cp_index < temp->cp_index));
1924 found = cl_lock_at_page(env, lock->cll_descr.cld_obj,
1929 descr = &found->cll_descr;
1930 cfs_list_for_each_entry_safe_from(page, temp, &queue->pl_pages,
1932 idx = page->cp_index;
1933 if (descr->cld_start > idx || descr->cld_end < idx)
1935 cl_page_list_move(plist, queue, page);
1937 cl_lock_put(env, found);
1940 /* The pages in plist are covered by other locks, don't handle them
1944 cl_page_list_disown(env, io, plist);
1945 cl_page_list_fini(env, plist);
1948 EXPORT_SYMBOL(cl_lock_page_list_fixup);
1951 * Invalidate pages protected by the given lock, sending them out to the
1952 * server first, if necessary.
1954 * This function does the following:
1956 * - collects a list of pages to be invalidated,
1958 * - unmaps them from the user virtual memory,
1960 * - sends dirty pages to the server,
1962 * - waits for transfer completion,
1964 * - discards pages, and throws them out of memory.
1966 * If \a discard is set, pages are discarded without sending them to the
1969 * If error happens on any step, the process continues anyway (the reasoning
1970 * behind this being that lock cancellation cannot be delayed indefinitely).
1972 int cl_lock_page_out(const struct lu_env *env, struct cl_lock *lock,
1975 struct cl_thread_info *info = cl_env_info(env);
1976 struct cl_io *io = &info->clt_io;
1977 struct cl_2queue *queue = &info->clt_queue;
1978 struct cl_lock_descr *descr = &lock->cll_descr;
1979 struct lu_device_type *dtype;
1985 LINVRNT(cl_lock_invariant(env, lock));
1988 io->ci_obj = cl_object_top(descr->cld_obj);
1989 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1993 dtype = descr->cld_obj->co_lu.lo_dev->ld_type;
1994 next_index = descr->cld_start;
1996 const struct cl_page_slice *slice;
1998 cl_2queue_init(queue);
1999 res = cl_page_gang_lookup(env, descr->cld_obj, io,
2000 next_index, descr->cld_end,
2002 page_count = queue->c2_qin.pl_nr;
2003 if (page_count == 0)
2006 /* cl_page_gang_lookup() uses subobj and sublock to look for
2007 * covered pages, but @queue->c2_qin contains the list of top
2008 * pages. We have to turn the page back to subpage so as to
2009 * get `correct' next index. -jay */
2010 slice = cl_page_at(cl_page_list_last(&queue->c2_qin), dtype);
2011 next_index = slice->cpl_page->cp_index + 1;
2013 result = cl_page_list_unmap(env, io, &queue->c2_qin);
2015 long timeout = 600; /* 10 minutes. */
2016 /* for debug purpose, if this request can't be
2017 * finished in 10 minutes, we hope it can notify us.
2019 result = cl_io_submit_sync(env, io, CRT_WRITE, queue,
2020 CRP_CANCEL, timeout);
2022 CWARN("Writing %lu pages error: %d\n",
2023 page_count, result);
2025 cl_lock_page_list_fixup(env, io, lock, &queue->c2_qout);
2026 cl_2queue_discard(env, io, queue);
2027 cl_2queue_disown(env, io, queue);
2028 cl_2queue_fini(env, queue);
2030 if (next_index > descr->cld_end)
2033 if (res == CLP_GANG_RESCHED)
2035 } while (res != CLP_GANG_OKAY);
2037 cl_io_fini(env, io);
2040 EXPORT_SYMBOL(cl_lock_page_out);
2043 * Eliminate all locks for a given object.
2045 * Caller has to guarantee that no lock is in active use.
2047 * \param cancel when this is set, cl_locks_prune() cancels locks before
2050 void cl_locks_prune(const struct lu_env *env, struct cl_object *obj, int cancel)
2052 struct cl_object_header *head;
2053 struct cl_lock *lock;
2056 head = cl_object_header(obj);
2058 * If locks are destroyed without cancellation, all pages must be
2059 * already destroyed (as otherwise they will be left unprotected).
2061 LASSERT(ergo(!cancel,
2062 head->coh_tree.rnode == NULL && head->coh_pages == 0));
2064 cfs_spin_lock(&head->coh_lock_guard);
2065 while (!cfs_list_empty(&head->coh_locks)) {
2066 lock = container_of(head->coh_locks.next,
2067 struct cl_lock, cll_linkage);
2068 cl_lock_get_trust(lock);
2069 cfs_spin_unlock(&head->coh_lock_guard);
2070 lu_ref_add(&lock->cll_reference, "prune", cfs_current());
2071 cl_lock_mutex_get(env, lock);
2072 if (lock->cll_state < CLS_FREEING) {
2073 LASSERT(lock->cll_holds == 0);
2074 LASSERT(lock->cll_users == 0);
2076 cl_lock_cancel(env, lock);
2077 cl_lock_delete(env, lock);
2079 cl_lock_mutex_put(env, lock);
2080 lu_ref_del(&lock->cll_reference, "prune", cfs_current());
2081 cl_lock_put(env, lock);
2082 cfs_spin_lock(&head->coh_lock_guard);
2084 cfs_spin_unlock(&head->coh_lock_guard);
2087 EXPORT_SYMBOL(cl_locks_prune);
2089 static struct cl_lock *cl_lock_hold_mutex(const struct lu_env *env,
2090 const struct cl_io *io,
2091 const struct cl_lock_descr *need,
2092 const char *scope, const void *source)
2094 struct cl_lock *lock;
2099 lock = cl_lock_find(env, io, need);
2102 cl_lock_mutex_get(env, lock);
2103 if (lock->cll_state < CLS_FREEING &&
2104 !(lock->cll_flags & CLF_CANCELLED)) {
2105 cl_lock_hold_mod(env, lock, +1);
2106 lu_ref_add(&lock->cll_holders, scope, source);
2107 lu_ref_add(&lock->cll_reference, scope, source);
2110 cl_lock_mutex_put(env, lock);
2111 cl_lock_put(env, lock);
2117 * Returns a lock matching \a need description with a reference and a hold on
2120 * This is much like cl_lock_find(), except that cl_lock_hold() additionally
2121 * guarantees that lock is not in the CLS_FREEING state on return.
2123 struct cl_lock *cl_lock_hold(const struct lu_env *env, const struct cl_io *io,
2124 const struct cl_lock_descr *need,
2125 const char *scope, const void *source)
2127 struct cl_lock *lock;
2131 lock = cl_lock_hold_mutex(env, io, need, scope, source);
2133 cl_lock_mutex_put(env, lock);
2136 EXPORT_SYMBOL(cl_lock_hold);
2139 * Main high-level entry point of cl_lock interface that finds existing or
2140 * enqueues new lock matching given description.
2142 struct cl_lock *cl_lock_request(const struct lu_env *env, struct cl_io *io,
2143 const struct cl_lock_descr *need,
2144 const char *scope, const void *source)
2146 struct cl_lock *lock;
2148 __u32 enqflags = need->cld_enq_flags;
2152 lock = cl_lock_hold_mutex(env, io, need, scope, source);
2153 if (!IS_ERR(lock)) {
2154 rc = cl_enqueue_locked(env, lock, io, enqflags);
2156 if (cl_lock_fits_into(env, lock, need, io)) {
2157 cl_lock_mutex_put(env, lock);
2158 cl_lock_lockdep_acquire(env,
2162 cl_unuse_locked(env, lock);
2164 cl_lock_trace(D_DLMTRACE, env, "enqueue failed", lock);
2165 cl_lock_hold_release(env, lock, scope, source);
2166 cl_lock_mutex_put(env, lock);
2167 lu_ref_del(&lock->cll_reference, scope, source);
2168 cl_lock_put(env, lock);
2175 EXPORT_SYMBOL(cl_lock_request);
2178 * Adds a hold to a known lock.
2180 void cl_lock_hold_add(const struct lu_env *env, struct cl_lock *lock,
2181 const char *scope, const void *source)
2183 LINVRNT(cl_lock_is_mutexed(lock));
2184 LINVRNT(cl_lock_invariant(env, lock));
2185 LASSERT(lock->cll_state != CLS_FREEING);
2188 cl_lock_hold_mod(env, lock, +1);
2190 lu_ref_add(&lock->cll_holders, scope, source);
2191 lu_ref_add(&lock->cll_reference, scope, source);
2194 EXPORT_SYMBOL(cl_lock_hold_add);
2197 * Releases a hold and a reference on a lock, on which caller acquired a
2200 void cl_lock_unhold(const struct lu_env *env, struct cl_lock *lock,
2201 const char *scope, const void *source)
2203 LINVRNT(cl_lock_invariant(env, lock));
2205 cl_lock_hold_release(env, lock, scope, source);
2206 lu_ref_del(&lock->cll_reference, scope, source);
2207 cl_lock_put(env, lock);
2210 EXPORT_SYMBOL(cl_lock_unhold);
2213 * Releases a hold and a reference on a lock, obtained by cl_lock_hold().
2215 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock,
2216 const char *scope, const void *source)
2218 LINVRNT(cl_lock_invariant(env, lock));
2220 cl_lock_trace(D_DLMTRACE, env, "release lock", lock);
2221 cl_lock_mutex_get(env, lock);
2222 cl_lock_hold_release(env, lock, scope, source);
2223 cl_lock_mutex_put(env, lock);
2224 lu_ref_del(&lock->cll_reference, scope, source);
2225 cl_lock_put(env, lock);
2228 EXPORT_SYMBOL(cl_lock_release);
2230 void cl_lock_user_add(const struct lu_env *env, struct cl_lock *lock)
2232 LINVRNT(cl_lock_is_mutexed(lock));
2233 LINVRNT(cl_lock_invariant(env, lock));
2236 cl_lock_used_mod(env, lock, +1);
2239 EXPORT_SYMBOL(cl_lock_user_add);
2241 int cl_lock_user_del(const struct lu_env *env, struct cl_lock *lock)
2243 LINVRNT(cl_lock_is_mutexed(lock));
2244 LINVRNT(cl_lock_invariant(env, lock));
2245 LASSERT(lock->cll_users > 0);
2248 cl_lock_used_mod(env, lock, -1);
2249 RETURN(lock->cll_users == 0);
2251 EXPORT_SYMBOL(cl_lock_user_del);
2253 const char *cl_lock_mode_name(const enum cl_lock_mode mode)
2255 static const char *names[] = {
2256 [CLM_PHANTOM] = "P",
2261 if (0 <= mode && mode < ARRAY_SIZE(names))
2266 EXPORT_SYMBOL(cl_lock_mode_name);
2269 * Prints human readable representation of a lock description.
2271 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2272 lu_printer_t printer,
2273 const struct cl_lock_descr *descr)
2275 const struct lu_fid *fid;
2277 fid = lu_object_fid(&descr->cld_obj->co_lu);
2278 (*printer)(env, cookie, DDESCR"@"DFID, PDESCR(descr), PFID(fid));
2280 EXPORT_SYMBOL(cl_lock_descr_print);
2283 * Prints human readable representation of \a lock to the \a f.
2285 void cl_lock_print(const struct lu_env *env, void *cookie,
2286 lu_printer_t printer, const struct cl_lock *lock)
2288 const struct cl_lock_slice *slice;
2289 (*printer)(env, cookie, "lock@%p[%d %d %d %d %d %08lx] ",
2290 lock, cfs_atomic_read(&lock->cll_ref),
2291 lock->cll_state, lock->cll_error, lock->cll_holds,
2292 lock->cll_users, lock->cll_flags);
2293 cl_lock_descr_print(env, cookie, printer, &lock->cll_descr);
2294 (*printer)(env, cookie, " {\n");
2296 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
2297 (*printer)(env, cookie, " %s@%p: ",
2298 slice->cls_obj->co_lu.lo_dev->ld_type->ldt_name,
2300 if (slice->cls_ops->clo_print != NULL)
2301 slice->cls_ops->clo_print(env, cookie, printer, slice);
2302 (*printer)(env, cookie, "\n");
2304 (*printer)(env, cookie, "} lock@%p\n", lock);
2306 EXPORT_SYMBOL(cl_lock_print);
2308 int cl_lock_init(void)
2310 return lu_kmem_init(cl_lock_caches);
2313 void cl_lock_fini(void)
2315 lu_kmem_fini(cl_lock_caches);