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.
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>
41 #define DEBUG_SUBSYSTEM S_CLASS
43 # define EXPORT_SYMTAB
46 #include <obd_class.h>
47 #include <obd_support.h>
48 #include <lustre_fid.h>
49 #include <libcfs/list.h>
50 /* lu_time_global_{init,fini}() */
53 #include <cl_object.h>
54 #include "cl_internal.h"
56 /** Lock class of cl_lock::cll_guard */
57 static cfs_lock_class_key_t cl_lock_guard_class;
58 static cfs_mem_cache_t *cl_lock_kmem;
60 static struct lu_kmem_descr cl_lock_caches[] = {
62 .ckd_cache = &cl_lock_kmem,
63 .ckd_name = "cl_lock_kmem",
64 .ckd_size = sizeof (struct cl_lock)
72 * Basic lock invariant that is maintained at all times. Caller either has a
73 * reference to \a lock, or somehow assures that \a lock cannot be freed.
75 * \see cl_lock_invariant()
77 static int cl_lock_invariant_trusted(const struct lu_env *env,
78 const struct cl_lock *lock)
80 return ergo(lock->cll_state == CLS_FREEING, lock->cll_holds == 0) &&
81 cfs_atomic_read(&lock->cll_ref) >= lock->cll_holds &&
82 lock->cll_holds >= lock->cll_users &&
83 lock->cll_holds >= 0 &&
84 lock->cll_users >= 0 &&
89 * Stronger lock invariant, checking that caller has a reference on a lock.
91 * \see cl_lock_invariant_trusted()
93 static int cl_lock_invariant(const struct lu_env *env,
94 const struct cl_lock *lock)
98 result = cfs_atomic_read(&lock->cll_ref) > 0 &&
99 cl_lock_invariant_trusted(env, lock);
100 if (!result && env != NULL)
101 CL_LOCK_DEBUG(D_ERROR, env, lock, "invariant broken");
106 * Returns lock "nesting": 0 for a top-lock and 1 for a sub-lock.
108 static enum clt_nesting_level cl_lock_nesting(const struct cl_lock *lock)
110 return cl_object_header(lock->cll_descr.cld_obj)->coh_nesting;
114 * Returns a set of counters for this lock, depending on a lock nesting.
116 static struct cl_thread_counters *cl_lock_counters(const struct lu_env *env,
117 const struct cl_lock *lock)
119 struct cl_thread_info *info;
120 enum clt_nesting_level nesting;
122 info = cl_env_info(env);
123 nesting = cl_lock_nesting(lock);
124 LASSERT(nesting < ARRAY_SIZE(info->clt_counters));
125 return &info->clt_counters[nesting];
128 static void cl_lock_trace0(int level, const struct lu_env *env,
129 const char *prefix, const struct cl_lock *lock,
130 const char *func, const int line)
132 struct cl_object_header *h = cl_object_header(lock->cll_descr.cld_obj);
133 CDEBUG(level, "%s: %p@(%d %p %d %d %d %d %d %lx)"
134 "(%p/%d/%d) at %s():%d\n",
135 prefix, lock, cfs_atomic_read(&lock->cll_ref),
136 lock->cll_guarder, lock->cll_depth,
137 lock->cll_state, lock->cll_error, lock->cll_holds,
138 lock->cll_users, lock->cll_flags,
139 env, h->coh_nesting, cl_lock_nr_mutexed(env),
142 #define cl_lock_trace(level, env, prefix, lock) \
143 cl_lock_trace0(level, env, prefix, lock, __FUNCTION__, __LINE__)
145 #define RETIP ((unsigned long)__builtin_return_address(0))
147 #ifdef CONFIG_LOCKDEP
148 static cfs_lock_class_key_t cl_lock_key;
150 static void cl_lock_lockdep_init(struct cl_lock *lock)
152 lockdep_set_class_and_name(lock, &cl_lock_key, "EXT");
155 static void cl_lock_lockdep_acquire(const struct lu_env *env,
156 struct cl_lock *lock, __u32 enqflags)
158 cl_lock_counters(env, lock)->ctc_nr_locks_acquired++;
159 #ifdef HAVE_LOCK_MAP_ACQUIRE
160 lock_map_acquire(&lock->dep_map);
161 #else /* HAVE_LOCK_MAP_ACQUIRE */
162 lock_acquire(&lock->dep_map, !!(enqflags & CEF_ASYNC),
163 /* try: */ 0, lock->cll_descr.cld_mode <= CLM_READ,
164 /* check: */ 2, RETIP);
165 #endif /* HAVE_LOCK_MAP_ACQUIRE */
168 static void cl_lock_lockdep_release(const struct lu_env *env,
169 struct cl_lock *lock)
171 cl_lock_counters(env, lock)->ctc_nr_locks_acquired--;
172 lock_release(&lock->dep_map, 0, RETIP);
175 #else /* !CONFIG_LOCKDEP */
177 static void cl_lock_lockdep_init(struct cl_lock *lock)
179 static void cl_lock_lockdep_acquire(const struct lu_env *env,
180 struct cl_lock *lock, __u32 enqflags)
182 static void cl_lock_lockdep_release(const struct lu_env *env,
183 struct cl_lock *lock)
186 #endif /* !CONFIG_LOCKDEP */
189 * Adds lock slice to the compound lock.
191 * This is called by cl_object_operations::coo_lock_init() methods to add a
192 * per-layer state to the lock. New state is added at the end of
193 * cl_lock::cll_layers list, that is, it is at the bottom of the stack.
195 * \see cl_req_slice_add(), cl_page_slice_add(), cl_io_slice_add()
197 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
198 struct cl_object *obj,
199 const struct cl_lock_operations *ops)
202 slice->cls_lock = lock;
203 cfs_list_add_tail(&slice->cls_linkage, &lock->cll_layers);
204 slice->cls_obj = obj;
205 slice->cls_ops = ops;
208 EXPORT_SYMBOL(cl_lock_slice_add);
211 * Returns true iff a lock with the mode \a has provides at least the same
212 * guarantees as a lock with the mode \a need.
214 int cl_lock_mode_match(enum cl_lock_mode has, enum cl_lock_mode need)
216 LINVRNT(need == CLM_READ || need == CLM_WRITE ||
217 need == CLM_PHANTOM || need == CLM_GROUP);
218 LINVRNT(has == CLM_READ || has == CLM_WRITE ||
219 has == CLM_PHANTOM || has == CLM_GROUP);
220 CLASSERT(CLM_PHANTOM < CLM_READ);
221 CLASSERT(CLM_READ < CLM_WRITE);
222 CLASSERT(CLM_WRITE < CLM_GROUP);
224 if (has != CLM_GROUP)
229 EXPORT_SYMBOL(cl_lock_mode_match);
232 * Returns true iff extent portions of lock descriptions match.
234 int cl_lock_ext_match(const struct cl_lock_descr *has,
235 const struct cl_lock_descr *need)
238 has->cld_start <= need->cld_start &&
239 has->cld_end >= need->cld_end &&
240 cl_lock_mode_match(has->cld_mode, need->cld_mode) &&
241 (has->cld_mode != CLM_GROUP || has->cld_gid == need->cld_gid);
243 EXPORT_SYMBOL(cl_lock_ext_match);
246 * Returns true iff a lock with the description \a has provides at least the
247 * same guarantees as a lock with the description \a need.
249 int cl_lock_descr_match(const struct cl_lock_descr *has,
250 const struct cl_lock_descr *need)
253 cl_object_same(has->cld_obj, need->cld_obj) &&
254 cl_lock_ext_match(has, need);
256 EXPORT_SYMBOL(cl_lock_descr_match);
258 static void cl_lock_free(const struct lu_env *env, struct cl_lock *lock)
260 struct cl_object *obj = lock->cll_descr.cld_obj;
262 LINVRNT(!cl_lock_is_mutexed(lock));
265 cl_lock_trace(D_DLMTRACE, env, "free lock", lock);
267 while (!cfs_list_empty(&lock->cll_layers)) {
268 struct cl_lock_slice *slice;
270 slice = cfs_list_entry(lock->cll_layers.next,
271 struct cl_lock_slice, cls_linkage);
272 cfs_list_del_init(lock->cll_layers.next);
273 slice->cls_ops->clo_fini(env, slice);
275 cfs_atomic_dec(&cl_object_site(obj)->cs_locks.cs_total);
276 cfs_atomic_dec(&cl_object_site(obj)->cs_locks_state[lock->cll_state]);
277 lu_object_ref_del_at(&obj->co_lu, lock->cll_obj_ref, "cl_lock", lock);
278 cl_object_put(env, obj);
279 lu_ref_fini(&lock->cll_reference);
280 lu_ref_fini(&lock->cll_holders);
281 cfs_mutex_destroy(&lock->cll_guard);
282 OBD_SLAB_FREE_PTR(lock, cl_lock_kmem);
287 * Releases a reference on a lock.
289 * When last reference is released, lock is returned to the cache, unless it
290 * is in cl_lock_state::CLS_FREEING state, in which case it is destroyed
293 * \see cl_object_put(), cl_page_put()
295 void cl_lock_put(const struct lu_env *env, struct cl_lock *lock)
297 struct cl_object *obj;
298 struct cl_object_header *head;
299 struct cl_site *site;
301 LINVRNT(cl_lock_invariant(env, lock));
303 obj = lock->cll_descr.cld_obj;
304 LINVRNT(obj != NULL);
305 head = cl_object_header(obj);
306 site = cl_object_site(obj);
308 CDEBUG(D_TRACE, "releasing reference: %d %p %lu\n",
309 cfs_atomic_read(&lock->cll_ref), lock, RETIP);
311 if (cfs_atomic_dec_and_test(&lock->cll_ref)) {
312 if (lock->cll_state == CLS_FREEING) {
313 LASSERT(cfs_list_empty(&lock->cll_linkage));
314 cl_lock_free(env, lock);
316 cfs_atomic_dec(&site->cs_locks.cs_busy);
320 EXPORT_SYMBOL(cl_lock_put);
323 * Acquires an additional reference to a lock.
325 * This can be called only by caller already possessing a reference to \a
328 * \see cl_object_get(), cl_page_get()
330 void cl_lock_get(struct cl_lock *lock)
332 LINVRNT(cl_lock_invariant(NULL, lock));
333 CDEBUG(D_TRACE, "acquiring reference: %d %p %lu\n",
334 cfs_atomic_read(&lock->cll_ref), lock, RETIP);
335 cfs_atomic_inc(&lock->cll_ref);
337 EXPORT_SYMBOL(cl_lock_get);
340 * Acquires a reference to a lock.
342 * This is much like cl_lock_get(), except that this function can be used to
343 * acquire initial reference to the cached lock. Caller has to deal with all
344 * possible races. Use with care!
346 * \see cl_page_get_trust()
348 void cl_lock_get_trust(struct cl_lock *lock)
350 struct cl_site *site = cl_object_site(lock->cll_descr.cld_obj);
352 CDEBUG(D_TRACE, "acquiring trusted reference: %d %p %lu\n",
353 cfs_atomic_read(&lock->cll_ref), lock, RETIP);
354 if (cfs_atomic_inc_return(&lock->cll_ref) == 1)
355 cfs_atomic_inc(&site->cs_locks.cs_busy);
357 EXPORT_SYMBOL(cl_lock_get_trust);
360 * Helper function destroying the lock that wasn't completely initialized.
362 * Other threads can acquire references to the top-lock through its
363 * sub-locks. Hence, it cannot be cl_lock_free()-ed immediately.
365 static void cl_lock_finish(const struct lu_env *env, struct cl_lock *lock)
367 cl_lock_mutex_get(env, lock);
368 cl_lock_cancel(env, lock);
369 cl_lock_delete(env, lock);
370 cl_lock_mutex_put(env, lock);
371 cl_lock_put(env, lock);
374 static struct cl_lock *cl_lock_alloc(const struct lu_env *env,
375 struct cl_object *obj,
376 const struct cl_io *io,
377 const struct cl_lock_descr *descr)
379 struct cl_lock *lock;
380 struct lu_object_header *head;
381 struct cl_site *site = cl_object_site(obj);
384 OBD_SLAB_ALLOC_PTR_GFP(lock, cl_lock_kmem, CFS_ALLOC_IO);
386 cfs_atomic_set(&lock->cll_ref, 1);
387 lock->cll_descr = *descr;
388 lock->cll_state = CLS_NEW;
390 lock->cll_obj_ref = lu_object_ref_add(&obj->co_lu,
392 CFS_INIT_LIST_HEAD(&lock->cll_layers);
393 CFS_INIT_LIST_HEAD(&lock->cll_linkage);
394 CFS_INIT_LIST_HEAD(&lock->cll_inclosure);
395 lu_ref_init(&lock->cll_reference);
396 lu_ref_init(&lock->cll_holders);
397 cfs_mutex_init(&lock->cll_guard);
398 cfs_lockdep_set_class(&lock->cll_guard, &cl_lock_guard_class);
399 cfs_waitq_init(&lock->cll_wq);
400 head = obj->co_lu.lo_header;
401 cfs_atomic_inc(&site->cs_locks_state[CLS_NEW]);
402 cfs_atomic_inc(&site->cs_locks.cs_total);
403 cfs_atomic_inc(&site->cs_locks.cs_created);
404 cl_lock_lockdep_init(lock);
405 cfs_list_for_each_entry(obj, &head->loh_layers,
409 err = obj->co_ops->coo_lock_init(env, obj, lock, io);
411 cl_lock_finish(env, lock);
417 lock = ERR_PTR(-ENOMEM);
422 * Transfer the lock into INTRANSIT state and return the original state.
424 * \pre state: CLS_CACHED, CLS_HELD or CLS_ENQUEUED
425 * \post state: CLS_INTRANSIT
428 enum cl_lock_state cl_lock_intransit(const struct lu_env *env,
429 struct cl_lock *lock)
431 enum cl_lock_state state = lock->cll_state;
433 LASSERT(cl_lock_is_mutexed(lock));
434 LASSERT(state != CLS_INTRANSIT);
435 LASSERTF(state >= CLS_ENQUEUED && state <= CLS_CACHED,
436 "Malformed lock state %d.\n", state);
438 cl_lock_state_set(env, lock, CLS_INTRANSIT);
439 lock->cll_intransit_owner = cfs_current();
440 cl_lock_hold_add(env, lock, "intransit", cfs_current());
443 EXPORT_SYMBOL(cl_lock_intransit);
446 * Exit the intransit state and restore the lock state to the original state
448 void cl_lock_extransit(const struct lu_env *env, struct cl_lock *lock,
449 enum cl_lock_state state)
451 LASSERT(cl_lock_is_mutexed(lock));
452 LASSERT(lock->cll_state == CLS_INTRANSIT);
453 LASSERT(state != CLS_INTRANSIT);
454 LASSERT(lock->cll_intransit_owner == cfs_current());
456 lock->cll_intransit_owner = NULL;
457 cl_lock_state_set(env, lock, state);
458 cl_lock_unhold(env, lock, "intransit", cfs_current());
460 EXPORT_SYMBOL(cl_lock_extransit);
463 * Checking whether the lock is intransit state
465 int cl_lock_is_intransit(struct cl_lock *lock)
467 LASSERT(cl_lock_is_mutexed(lock));
468 return lock->cll_state == CLS_INTRANSIT &&
469 lock->cll_intransit_owner != cfs_current();
471 EXPORT_SYMBOL(cl_lock_is_intransit);
473 * Returns true iff lock is "suitable" for given io. E.g., locks acquired by
474 * truncate and O_APPEND cannot be reused for read/non-append-write, as they
475 * cover multiple stripes and can trigger cascading timeouts.
477 static int cl_lock_fits_into(const struct lu_env *env,
478 const struct cl_lock *lock,
479 const struct cl_lock_descr *need,
480 const struct cl_io *io)
482 const struct cl_lock_slice *slice;
484 LINVRNT(cl_lock_invariant_trusted(env, lock));
486 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
487 if (slice->cls_ops->clo_fits_into != NULL &&
488 !slice->cls_ops->clo_fits_into(env, slice, need, io))
494 static struct cl_lock *cl_lock_lookup(const struct lu_env *env,
495 struct cl_object *obj,
496 const struct cl_io *io,
497 const struct cl_lock_descr *need)
499 struct cl_lock *lock;
500 struct cl_object_header *head;
501 struct cl_site *site;
505 head = cl_object_header(obj);
506 site = cl_object_site(obj);
507 LINVRNT_SPIN_LOCKED(&head->coh_lock_guard);
508 cfs_atomic_inc(&site->cs_locks.cs_lookup);
509 cfs_list_for_each_entry(lock, &head->coh_locks, cll_linkage) {
512 matched = cl_lock_ext_match(&lock->cll_descr, need) &&
513 lock->cll_state < CLS_FREEING &&
514 lock->cll_error == 0 &&
515 !(lock->cll_flags & CLF_CANCELLED) &&
516 cl_lock_fits_into(env, lock, need, io);
517 CDEBUG(D_DLMTRACE, "has: "DDESCR"(%d) need: "DDESCR": %d\n",
518 PDESCR(&lock->cll_descr), lock->cll_state, PDESCR(need),
521 cl_lock_get_trust(lock);
522 cfs_atomic_inc(&cl_object_site(obj)->cs_locks.cs_hit);
530 * Returns a lock matching description \a need.
532 * This is the main entry point into the cl_lock caching interface. First, a
533 * cache (implemented as a per-object linked list) is consulted. If lock is
534 * found there, it is returned immediately. Otherwise new lock is allocated
535 * and returned. In any case, additional reference to lock is acquired.
537 * \see cl_object_find(), cl_page_find()
539 static struct cl_lock *cl_lock_find(const struct lu_env *env,
540 const struct cl_io *io,
541 const struct cl_lock_descr *need)
543 struct cl_object_header *head;
544 struct cl_object *obj;
545 struct cl_lock *lock;
546 struct cl_site *site;
551 head = cl_object_header(obj);
552 site = cl_object_site(obj);
554 cfs_spin_lock(&head->coh_lock_guard);
555 lock = cl_lock_lookup(env, obj, io, need);
556 cfs_spin_unlock(&head->coh_lock_guard);
559 lock = cl_lock_alloc(env, obj, io, need);
561 struct cl_lock *ghost;
563 cfs_spin_lock(&head->coh_lock_guard);
564 ghost = cl_lock_lookup(env, obj, io, need);
566 cfs_list_add_tail(&lock->cll_linkage,
568 cfs_spin_unlock(&head->coh_lock_guard);
569 cfs_atomic_inc(&site->cs_locks.cs_busy);
571 cfs_spin_unlock(&head->coh_lock_guard);
573 * Other threads can acquire references to the
574 * top-lock through its sub-locks. Hence, it
575 * cannot be cl_lock_free()-ed immediately.
577 cl_lock_finish(env, lock);
586 * Returns existing lock matching given description. This is similar to
587 * cl_lock_find() except that no new lock is created, and returned lock is
588 * guaranteed to be in enum cl_lock_state::CLS_HELD state.
590 struct cl_lock *cl_lock_peek(const struct lu_env *env, const struct cl_io *io,
591 const struct cl_lock_descr *need,
592 const char *scope, const void *source)
594 struct cl_object_header *head;
595 struct cl_object *obj;
596 struct cl_lock *lock;
600 head = cl_object_header(obj);
602 cfs_spin_lock(&head->coh_lock_guard);
603 lock = cl_lock_lookup(env, obj, io, need);
604 cfs_spin_unlock(&head->coh_lock_guard);
609 cl_lock_mutex_get(env, lock);
610 if (lock->cll_state == CLS_INTRANSIT)
611 cl_lock_state_wait(env, lock); /* Don't care return value. */
612 if (lock->cll_state == CLS_CACHED) {
614 result = cl_use_try(env, lock, 1);
616 cl_lock_error(env, lock, result);
618 ok = lock->cll_state == CLS_HELD;
620 cl_lock_hold_add(env, lock, scope, source);
621 cl_lock_user_add(env, lock);
622 cl_lock_put(env, lock);
624 cl_lock_mutex_put(env, lock);
626 cl_lock_put(env, lock);
632 EXPORT_SYMBOL(cl_lock_peek);
635 * Returns a slice within a lock, corresponding to the given layer in the
640 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
641 const struct lu_device_type *dtype)
643 const struct cl_lock_slice *slice;
645 LINVRNT(cl_lock_invariant_trusted(NULL, lock));
648 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
649 if (slice->cls_obj->co_lu.lo_dev->ld_type == dtype)
654 EXPORT_SYMBOL(cl_lock_at);
656 static void cl_lock_mutex_tail(const struct lu_env *env, struct cl_lock *lock)
658 struct cl_thread_counters *counters;
660 counters = cl_lock_counters(env, lock);
662 counters->ctc_nr_locks_locked++;
663 lu_ref_add(&counters->ctc_locks_locked, "cll_guard", lock);
664 cl_lock_trace(D_TRACE, env, "got mutex", lock);
668 * Locks cl_lock object.
670 * This is used to manipulate cl_lock fields, and to serialize state
671 * transitions in the lock state machine.
673 * \post cl_lock_is_mutexed(lock)
675 * \see cl_lock_mutex_put()
677 void cl_lock_mutex_get(const struct lu_env *env, struct cl_lock *lock)
679 LINVRNT(cl_lock_invariant(env, lock));
681 if (lock->cll_guarder == cfs_current()) {
682 LINVRNT(cl_lock_is_mutexed(lock));
683 LINVRNT(lock->cll_depth > 0);
685 struct cl_object_header *hdr;
686 struct cl_thread_info *info;
689 LINVRNT(lock->cll_guarder != cfs_current());
690 hdr = cl_object_header(lock->cll_descr.cld_obj);
692 * Check that mutices are taken in the bottom-to-top order.
694 info = cl_env_info(env);
695 for (i = 0; i < hdr->coh_nesting; ++i)
696 LASSERT(info->clt_counters[i].ctc_nr_locks_locked == 0);
697 cfs_mutex_lock_nested(&lock->cll_guard, hdr->coh_nesting);
698 lock->cll_guarder = cfs_current();
699 LINVRNT(lock->cll_depth == 0);
701 cl_lock_mutex_tail(env, lock);
703 EXPORT_SYMBOL(cl_lock_mutex_get);
706 * Try-locks cl_lock object.
708 * \retval 0 \a lock was successfully locked
710 * \retval -EBUSY \a lock cannot be locked right now
712 * \post ergo(result == 0, cl_lock_is_mutexed(lock))
714 * \see cl_lock_mutex_get()
716 int cl_lock_mutex_try(const struct lu_env *env, struct cl_lock *lock)
720 LINVRNT(cl_lock_invariant_trusted(env, lock));
724 if (lock->cll_guarder == cfs_current()) {
725 LINVRNT(lock->cll_depth > 0);
726 cl_lock_mutex_tail(env, lock);
727 } else if (cfs_mutex_trylock(&lock->cll_guard)) {
728 LINVRNT(lock->cll_depth == 0);
729 lock->cll_guarder = cfs_current();
730 cl_lock_mutex_tail(env, lock);
735 EXPORT_SYMBOL(cl_lock_mutex_try);
738 {* Unlocks cl_lock object.
740 * \pre cl_lock_is_mutexed(lock)
742 * \see cl_lock_mutex_get()
744 void cl_lock_mutex_put(const struct lu_env *env, struct cl_lock *lock)
746 struct cl_thread_counters *counters;
748 LINVRNT(cl_lock_invariant(env, lock));
749 LINVRNT(cl_lock_is_mutexed(lock));
750 LINVRNT(lock->cll_guarder == cfs_current());
751 LINVRNT(lock->cll_depth > 0);
753 counters = cl_lock_counters(env, lock);
754 LINVRNT(counters->ctc_nr_locks_locked > 0);
756 cl_lock_trace(D_TRACE, env, "put mutex", lock);
757 lu_ref_del(&counters->ctc_locks_locked, "cll_guard", lock);
758 counters->ctc_nr_locks_locked--;
759 if (--lock->cll_depth == 0) {
760 lock->cll_guarder = NULL;
761 cfs_mutex_unlock(&lock->cll_guard);
764 EXPORT_SYMBOL(cl_lock_mutex_put);
767 * Returns true iff lock's mutex is owned by the current thread.
769 int cl_lock_is_mutexed(struct cl_lock *lock)
771 return lock->cll_guarder == cfs_current();
773 EXPORT_SYMBOL(cl_lock_is_mutexed);
776 * Returns number of cl_lock mutices held by the current thread (environment).
778 int cl_lock_nr_mutexed(const struct lu_env *env)
780 struct cl_thread_info *info;
785 * NOTE: if summation across all nesting levels (currently 2) proves
786 * too expensive, a summary counter can be added to
787 * struct cl_thread_info.
789 info = cl_env_info(env);
790 for (i = 0, locked = 0; i < ARRAY_SIZE(info->clt_counters); ++i)
791 locked += info->clt_counters[i].ctc_nr_locks_locked;
794 EXPORT_SYMBOL(cl_lock_nr_mutexed);
796 static void cl_lock_cancel0(const struct lu_env *env, struct cl_lock *lock)
798 LINVRNT(cl_lock_is_mutexed(lock));
799 LINVRNT(cl_lock_invariant(env, lock));
801 if (!(lock->cll_flags & CLF_CANCELLED)) {
802 const struct cl_lock_slice *slice;
804 lock->cll_flags |= CLF_CANCELLED;
805 cfs_list_for_each_entry_reverse(slice, &lock->cll_layers,
807 if (slice->cls_ops->clo_cancel != NULL)
808 slice->cls_ops->clo_cancel(env, slice);
814 static void cl_lock_delete0(const struct lu_env *env, struct cl_lock *lock)
816 struct cl_object_header *head;
817 const struct cl_lock_slice *slice;
819 LINVRNT(cl_lock_is_mutexed(lock));
820 LINVRNT(cl_lock_invariant(env, lock));
823 if (lock->cll_state < CLS_FREEING) {
824 LASSERT(lock->cll_state != CLS_INTRANSIT);
825 cl_lock_state_set(env, lock, CLS_FREEING);
827 head = cl_object_header(lock->cll_descr.cld_obj);
829 cfs_spin_lock(&head->coh_lock_guard);
830 cfs_list_del_init(&lock->cll_linkage);
832 cfs_spin_unlock(&head->coh_lock_guard);
834 * From now on, no new references to this lock can be acquired
835 * by cl_lock_lookup().
837 cfs_list_for_each_entry_reverse(slice, &lock->cll_layers,
839 if (slice->cls_ops->clo_delete != NULL)
840 slice->cls_ops->clo_delete(env, slice);
843 * From now on, no new references to this lock can be acquired
844 * by layer-specific means (like a pointer from struct
845 * ldlm_lock in osc, or a pointer from top-lock to sub-lock in
848 * Lock will be finally freed in cl_lock_put() when last of
849 * existing references goes away.
856 * Mod(ifie)s cl_lock::cll_holds counter for a given lock. Also, for a
857 * top-lock (nesting == 0) accounts for this modification in the per-thread
858 * debugging counters. Sub-lock holds can be released by a thread different
859 * from one that acquired it.
861 static void cl_lock_hold_mod(const struct lu_env *env, struct cl_lock *lock,
864 struct cl_thread_counters *counters;
865 enum clt_nesting_level nesting;
867 lock->cll_holds += delta;
868 nesting = cl_lock_nesting(lock);
869 if (nesting == CNL_TOP) {
870 counters = &cl_env_info(env)->clt_counters[CNL_TOP];
871 counters->ctc_nr_held += delta;
872 LASSERT(counters->ctc_nr_held >= 0);
877 * Mod(ifie)s cl_lock::cll_users counter for a given lock. See
878 * cl_lock_hold_mod() for the explanation of the debugging code.
880 static void cl_lock_used_mod(const struct lu_env *env, struct cl_lock *lock,
883 struct cl_thread_counters *counters;
884 enum clt_nesting_level nesting;
886 lock->cll_users += delta;
887 nesting = cl_lock_nesting(lock);
888 if (nesting == CNL_TOP) {
889 counters = &cl_env_info(env)->clt_counters[CNL_TOP];
890 counters->ctc_nr_used += delta;
891 LASSERT(counters->ctc_nr_used >= 0);
895 static void cl_lock_hold_release(const struct lu_env *env, struct cl_lock *lock,
896 const char *scope, const void *source)
898 LINVRNT(cl_lock_is_mutexed(lock));
899 LINVRNT(cl_lock_invariant(env, lock));
900 LASSERT(lock->cll_holds > 0);
903 cl_lock_trace(D_DLMTRACE, env, "hold release lock", lock);
904 lu_ref_del(&lock->cll_holders, scope, source);
905 cl_lock_hold_mod(env, lock, -1);
906 if (lock->cll_holds == 0) {
907 if (lock->cll_descr.cld_mode == CLM_PHANTOM ||
908 lock->cll_descr.cld_mode == CLM_GROUP)
910 * If lock is still phantom or grouplock when user is
911 * done with it---destroy the lock.
913 lock->cll_flags |= CLF_CANCELPEND|CLF_DOOMED;
914 if (lock->cll_flags & CLF_CANCELPEND) {
915 lock->cll_flags &= ~CLF_CANCELPEND;
916 cl_lock_cancel0(env, lock);
918 if (lock->cll_flags & CLF_DOOMED) {
919 /* no longer doomed: it's dead... Jim. */
920 lock->cll_flags &= ~CLF_DOOMED;
921 cl_lock_delete0(env, lock);
929 * Waits until lock state is changed.
931 * This function is called with cl_lock mutex locked, atomically releases
932 * mutex and goes to sleep, waiting for a lock state change (signaled by
933 * cl_lock_signal()), and re-acquires the mutex before return.
935 * This function is used to wait until lock state machine makes some progress
936 * and to emulate synchronous operations on top of asynchronous lock
939 * \retval -EINTR wait was interrupted
941 * \retval 0 wait wasn't interrupted
943 * \pre cl_lock_is_mutexed(lock)
945 * \see cl_lock_signal()
947 int cl_lock_state_wait(const struct lu_env *env, struct cl_lock *lock)
949 cfs_waitlink_t waiter;
953 LINVRNT(cl_lock_is_mutexed(lock));
954 LINVRNT(cl_lock_invariant(env, lock));
955 LASSERT(lock->cll_depth == 1);
956 LASSERT(lock->cll_state != CLS_FREEING); /* too late to wait */
958 cl_lock_trace(D_DLMTRACE, env, "state wait lock", lock);
959 result = lock->cll_error;
961 cfs_waitlink_init(&waiter);
962 cfs_waitq_add(&lock->cll_wq, &waiter);
963 cfs_set_current_state(CFS_TASK_INTERRUPTIBLE);
964 cl_lock_mutex_put(env, lock);
966 LASSERT(cl_lock_nr_mutexed(env) == 0);
967 cfs_waitq_wait(&waiter, CFS_TASK_INTERRUPTIBLE);
969 cl_lock_mutex_get(env, lock);
970 cfs_set_current_state(CFS_TASK_RUNNING);
971 cfs_waitq_del(&lock->cll_wq, &waiter);
972 result = cfs_signal_pending() ? -EINTR : 0;
976 EXPORT_SYMBOL(cl_lock_state_wait);
978 static void cl_lock_state_signal(const struct lu_env *env, struct cl_lock *lock,
979 enum cl_lock_state state)
981 const struct cl_lock_slice *slice;
984 LINVRNT(cl_lock_is_mutexed(lock));
985 LINVRNT(cl_lock_invariant(env, lock));
987 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage)
988 if (slice->cls_ops->clo_state != NULL)
989 slice->cls_ops->clo_state(env, slice, state);
990 cfs_waitq_broadcast(&lock->cll_wq);
995 * Notifies waiters that lock state changed.
997 * Wakes up all waiters sleeping in cl_lock_state_wait(), also notifies all
998 * layers about state change by calling cl_lock_operations::clo_state()
1001 void cl_lock_signal(const struct lu_env *env, struct cl_lock *lock)
1004 cl_lock_trace(D_DLMTRACE, env, "state signal lock", lock);
1005 cl_lock_state_signal(env, lock, lock->cll_state);
1008 EXPORT_SYMBOL(cl_lock_signal);
1011 * Changes lock state.
1013 * This function is invoked to notify layers that lock state changed, possible
1014 * as a result of an asynchronous event such as call-back reception.
1016 * \post lock->cll_state == state
1018 * \see cl_lock_operations::clo_state()
1020 void cl_lock_state_set(const struct lu_env *env, struct cl_lock *lock,
1021 enum cl_lock_state state)
1023 struct cl_site *site = cl_object_site(lock->cll_descr.cld_obj);
1026 LASSERT(lock->cll_state <= state ||
1027 (lock->cll_state == CLS_CACHED &&
1028 (state == CLS_HELD || /* lock found in cache */
1029 state == CLS_NEW || /* sub-lock canceled */
1030 state == CLS_INTRANSIT)) ||
1031 /* lock is in transit state */
1032 lock->cll_state == CLS_INTRANSIT);
1034 if (lock->cll_state != state) {
1035 cfs_atomic_dec(&site->cs_locks_state[lock->cll_state]);
1036 cfs_atomic_inc(&site->cs_locks_state[state]);
1038 cl_lock_state_signal(env, lock, state);
1039 lock->cll_state = state;
1043 EXPORT_SYMBOL(cl_lock_state_set);
1045 static int cl_unuse_try_internal(const struct lu_env *env, struct cl_lock *lock)
1047 const struct cl_lock_slice *slice;
1053 LINVRNT(cl_lock_is_mutexed(lock));
1054 LINVRNT(cl_lock_invariant(env, lock));
1055 LASSERT(lock->cll_state == CLS_INTRANSIT);
1058 cfs_list_for_each_entry_reverse(slice, &lock->cll_layers,
1060 if (slice->cls_ops->clo_unuse != NULL) {
1061 result = slice->cls_ops->clo_unuse(env, slice);
1066 LASSERT(result != -ENOSYS);
1067 } while (result == CLO_REPEAT);
1073 * Yanks lock from the cache (cl_lock_state::CLS_CACHED state) by calling
1074 * cl_lock_operations::clo_use() top-to-bottom to notify layers.
1075 * @atomic = 1, it must unuse the lock to recovery the lock to keep the
1076 * use process atomic
1078 int cl_use_try(const struct lu_env *env, struct cl_lock *lock, int atomic)
1080 const struct cl_lock_slice *slice;
1082 enum cl_lock_state state;
1085 cl_lock_trace(D_DLMTRACE, env, "use lock", lock);
1087 LASSERT(lock->cll_state == CLS_CACHED);
1088 if (lock->cll_error)
1089 RETURN(lock->cll_error);
1092 state = cl_lock_intransit(env, lock);
1093 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1094 if (slice->cls_ops->clo_use != NULL) {
1095 result = slice->cls_ops->clo_use(env, slice);
1100 LASSERT(result != -ENOSYS);
1102 LASSERTF(lock->cll_state == CLS_INTRANSIT, "Wrong state %d.\n",
1108 if (result == -ESTALE) {
1110 * ESTALE means sublock being cancelled
1111 * at this time, and set lock state to
1112 * be NEW here and ask the caller to repeat.
1115 result = CLO_REPEAT;
1118 /* @atomic means back-off-on-failure. */
1121 rc = cl_unuse_try_internal(env, lock);
1122 /* Vet the results. */
1123 if (rc < 0 && result > 0)
1128 cl_lock_extransit(env, lock, state);
1131 EXPORT_SYMBOL(cl_use_try);
1134 * Helper for cl_enqueue_try() that calls ->clo_enqueue() across all layers
1137 static int cl_enqueue_kick(const struct lu_env *env,
1138 struct cl_lock *lock,
1139 struct cl_io *io, __u32 flags)
1142 const struct cl_lock_slice *slice;
1146 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1147 if (slice->cls_ops->clo_enqueue != NULL) {
1148 result = slice->cls_ops->clo_enqueue(env,
1154 LASSERT(result != -ENOSYS);
1159 * Tries to enqueue a lock.
1161 * This function is called repeatedly by cl_enqueue() until either lock is
1162 * enqueued, or error occurs. This function does not block waiting for
1163 * networking communication to complete.
1165 * \post ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1166 * lock->cll_state == CLS_HELD)
1168 * \see cl_enqueue() cl_lock_operations::clo_enqueue()
1169 * \see cl_lock_state::CLS_ENQUEUED
1171 int cl_enqueue_try(const struct lu_env *env, struct cl_lock *lock,
1172 struct cl_io *io, __u32 flags)
1177 cl_lock_trace(D_DLMTRACE, env, "enqueue lock", lock);
1181 LINVRNT(cl_lock_is_mutexed(lock));
1183 if (lock->cll_error != 0)
1185 switch (lock->cll_state) {
1187 cl_lock_state_set(env, lock, CLS_QUEUING);
1191 result = cl_enqueue_kick(env, lock, io, flags);
1193 cl_lock_state_set(env, lock, CLS_ENQUEUED);
1196 LASSERT(cl_lock_is_intransit(lock));
1200 /* yank lock from the cache. */
1201 result = cl_use_try(env, lock, 0);
1210 * impossible, only held locks with increased
1211 * ->cll_holds can be enqueued, and they cannot be
1216 } while (result == CLO_REPEAT);
1218 cl_lock_error(env, lock, result);
1219 RETURN(result ?: lock->cll_error);
1221 EXPORT_SYMBOL(cl_enqueue_try);
1224 * Cancel the conflicting lock found during previous enqueue.
1226 * \retval 0 conflicting lock has been canceled.
1227 * \retval -ve error code.
1229 int cl_lock_enqueue_wait(const struct lu_env *env,
1230 struct cl_lock *lock,
1233 struct cl_lock *conflict;
1237 LASSERT(cl_lock_is_mutexed(lock));
1238 LASSERT(lock->cll_state == CLS_QUEUING);
1239 LASSERT(lock->cll_conflict != NULL);
1241 conflict = lock->cll_conflict;
1242 lock->cll_conflict = NULL;
1244 cl_lock_mutex_put(env, lock);
1245 LASSERT(cl_lock_nr_mutexed(env) == 0);
1247 cl_lock_mutex_get(env, conflict);
1248 cl_lock_cancel(env, conflict);
1249 cl_lock_delete(env, conflict);
1251 while (conflict->cll_state != CLS_FREEING) {
1252 rc = cl_lock_state_wait(env, conflict);
1256 cl_lock_mutex_put(env, conflict);
1257 lu_ref_del(&conflict->cll_reference, "cancel-wait", lock);
1258 cl_lock_put(env, conflict);
1261 cl_lock_mutex_get(env, lock);
1266 EXPORT_SYMBOL(cl_lock_enqueue_wait);
1268 static int cl_enqueue_locked(const struct lu_env *env, struct cl_lock *lock,
1269 struct cl_io *io, __u32 enqflags)
1275 LINVRNT(cl_lock_is_mutexed(lock));
1276 LINVRNT(cl_lock_invariant(env, lock));
1277 LASSERT(lock->cll_holds > 0);
1279 cl_lock_user_add(env, lock);
1281 result = cl_enqueue_try(env, lock, io, enqflags);
1282 if (result == CLO_WAIT) {
1283 if (lock->cll_conflict != NULL)
1284 result = cl_lock_enqueue_wait(env, lock, 1);
1286 result = cl_lock_state_wait(env, lock);
1293 cl_lock_user_del(env, lock);
1294 cl_lock_error(env, lock, result);
1296 LASSERT(ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1297 lock->cll_state == CLS_HELD));
1304 * \pre current thread or io owns a hold on lock.
1306 * \post ergo(result == 0, lock->users increased)
1307 * \post ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1308 * lock->cll_state == CLS_HELD)
1310 int cl_enqueue(const struct lu_env *env, struct cl_lock *lock,
1311 struct cl_io *io, __u32 enqflags)
1317 cl_lock_lockdep_acquire(env, lock, enqflags);
1318 cl_lock_mutex_get(env, lock);
1319 result = cl_enqueue_locked(env, lock, io, enqflags);
1320 cl_lock_mutex_put(env, lock);
1322 cl_lock_lockdep_release(env, lock);
1323 LASSERT(ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1324 lock->cll_state == CLS_HELD));
1327 EXPORT_SYMBOL(cl_enqueue);
1330 * Tries to unlock a lock.
1332 * This function is called repeatedly by cl_unuse() until either lock is
1333 * unlocked, or error occurs.
1334 * cl_unuse_try is a one-shot operation, so it must NOT return CLO_WAIT.
1336 * \pre lock->cll_state == CLS_HELD
1338 * \post ergo(result == 0, lock->cll_state == CLS_CACHED)
1340 * \see cl_unuse() cl_lock_operations::clo_unuse()
1341 * \see cl_lock_state::CLS_CACHED
1343 int cl_unuse_try(const struct lu_env *env, struct cl_lock *lock)
1346 enum cl_lock_state state = CLS_NEW;
1349 cl_lock_trace(D_DLMTRACE, env, "unuse lock", lock);
1351 LASSERT(lock->cll_state == CLS_HELD || lock->cll_state == CLS_ENQUEUED);
1352 if (lock->cll_users > 1) {
1353 cl_lock_user_del(env, lock);
1358 * New lock users (->cll_users) are not protecting unlocking
1359 * from proceeding. From this point, lock eventually reaches
1360 * CLS_CACHED, is reinitialized to CLS_NEW or fails into
1363 state = cl_lock_intransit(env, lock);
1365 result = cl_unuse_try_internal(env, lock);
1366 LASSERT(lock->cll_state == CLS_INTRANSIT);
1367 LASSERT(result != CLO_WAIT);
1368 cl_lock_user_del(env, lock);
1369 if (result == 0 || result == -ESTALE) {
1371 * Return lock back to the cache. This is the only
1372 * place where lock is moved into CLS_CACHED state.
1374 * If one of ->clo_unuse() methods returned -ESTALE, lock
1375 * cannot be placed into cache and has to be
1376 * re-initialized. This happens e.g., when a sub-lock was
1377 * canceled while unlocking was in progress.
1379 if (state == CLS_HELD && result == 0)
1383 cl_lock_extransit(env, lock, state);
1386 * Hide -ESTALE error.
1387 * If the lock is a glimpse lock, and it has multiple
1388 * stripes. Assuming that one of its sublock returned -ENAVAIL,
1389 * and other sublocks are matched write locks. In this case,
1390 * we can't set this lock to error because otherwise some of
1391 * its sublocks may not be canceled. This causes some dirty
1392 * pages won't be written to OSTs. -jay
1396 CERROR("result = %d, this is unlikely!\n", result);
1397 cl_lock_extransit(env, lock, state);
1400 result = result ?: lock->cll_error;
1402 cl_lock_error(env, lock, result);
1405 EXPORT_SYMBOL(cl_unuse_try);
1407 static void cl_unuse_locked(const struct lu_env *env, struct cl_lock *lock)
1412 result = cl_unuse_try(env, lock);
1414 CL_LOCK_DEBUG(D_ERROR, env, lock, "unuse return %d\n", result);
1422 void cl_unuse(const struct lu_env *env, struct cl_lock *lock)
1425 cl_lock_mutex_get(env, lock);
1426 cl_unuse_locked(env, lock);
1427 cl_lock_mutex_put(env, lock);
1428 cl_lock_lockdep_release(env, lock);
1431 EXPORT_SYMBOL(cl_unuse);
1434 * Tries to wait for a lock.
1436 * This function is called repeatedly by cl_wait() until either lock is
1437 * granted, or error occurs. This function does not block waiting for network
1438 * communication to complete.
1440 * \see cl_wait() cl_lock_operations::clo_wait()
1441 * \see cl_lock_state::CLS_HELD
1443 int cl_wait_try(const struct lu_env *env, struct cl_lock *lock)
1445 const struct cl_lock_slice *slice;
1449 cl_lock_trace(D_DLMTRACE, env, "wait lock try", lock);
1451 LINVRNT(cl_lock_is_mutexed(lock));
1452 LINVRNT(cl_lock_invariant(env, lock));
1453 LASSERT(lock->cll_state == CLS_ENQUEUED ||
1454 lock->cll_state == CLS_HELD ||
1455 lock->cll_state == CLS_INTRANSIT);
1456 LASSERT(lock->cll_users > 0);
1457 LASSERT(lock->cll_holds > 0);
1460 if (lock->cll_error != 0)
1463 if (cl_lock_is_intransit(lock)) {
1468 if (lock->cll_state == CLS_HELD)
1473 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1474 if (slice->cls_ops->clo_wait != NULL) {
1475 result = slice->cls_ops->clo_wait(env, slice);
1480 LASSERT(result != -ENOSYS);
1482 LASSERT(lock->cll_state != CLS_INTRANSIT);
1483 cl_lock_state_set(env, lock, CLS_HELD);
1485 } while (result == CLO_REPEAT);
1486 RETURN(result ?: lock->cll_error);
1488 EXPORT_SYMBOL(cl_wait_try);
1491 * Waits until enqueued lock is granted.
1493 * \pre current thread or io owns a hold on the lock
1494 * \pre ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1495 * lock->cll_state == CLS_HELD)
1497 * \post ergo(result == 0, lock->cll_state == CLS_HELD)
1499 int cl_wait(const struct lu_env *env, struct cl_lock *lock)
1504 cl_lock_mutex_get(env, lock);
1506 LINVRNT(cl_lock_invariant(env, lock));
1507 LASSERTF(lock->cll_state == CLS_ENQUEUED || lock->cll_state == CLS_HELD,
1508 "Wrong state %d \n", lock->cll_state);
1509 LASSERT(lock->cll_holds > 0);
1512 result = cl_wait_try(env, lock);
1513 if (result == CLO_WAIT) {
1514 result = cl_lock_state_wait(env, lock);
1521 cl_lock_user_del(env, lock);
1522 cl_lock_error(env, lock, result);
1523 cl_lock_lockdep_release(env, lock);
1525 cl_lock_trace(D_DLMTRACE, env, "wait lock", lock);
1526 cl_lock_mutex_put(env, lock);
1527 LASSERT(ergo(result == 0, lock->cll_state == CLS_HELD));
1530 EXPORT_SYMBOL(cl_wait);
1533 * Executes cl_lock_operations::clo_weigh(), and sums results to estimate lock
1536 unsigned long cl_lock_weigh(const struct lu_env *env, struct cl_lock *lock)
1538 const struct cl_lock_slice *slice;
1539 unsigned long pound;
1540 unsigned long ounce;
1543 LINVRNT(cl_lock_is_mutexed(lock));
1544 LINVRNT(cl_lock_invariant(env, lock));
1547 cfs_list_for_each_entry_reverse(slice, &lock->cll_layers, cls_linkage) {
1548 if (slice->cls_ops->clo_weigh != NULL) {
1549 ounce = slice->cls_ops->clo_weigh(env, slice);
1551 if (pound < ounce) /* over-weight^Wflow */
1557 EXPORT_SYMBOL(cl_lock_weigh);
1560 * Notifies layers that lock description changed.
1562 * The server can grant client a lock different from one that was requested
1563 * (e.g., larger in extent). This method is called when actually granted lock
1564 * description becomes known to let layers to accommodate for changed lock
1567 * \see cl_lock_operations::clo_modify()
1569 int cl_lock_modify(const struct lu_env *env, struct cl_lock *lock,
1570 const struct cl_lock_descr *desc)
1572 const struct cl_lock_slice *slice;
1573 struct cl_object *obj = lock->cll_descr.cld_obj;
1574 struct cl_object_header *hdr = cl_object_header(obj);
1578 cl_lock_trace(D_DLMTRACE, env, "modify lock", lock);
1579 /* don't allow object to change */
1580 LASSERT(obj == desc->cld_obj);
1581 LINVRNT(cl_lock_is_mutexed(lock));
1582 LINVRNT(cl_lock_invariant(env, lock));
1584 cfs_list_for_each_entry_reverse(slice, &lock->cll_layers, cls_linkage) {
1585 if (slice->cls_ops->clo_modify != NULL) {
1586 result = slice->cls_ops->clo_modify(env, slice, desc);
1591 CL_LOCK_DEBUG(D_DLMTRACE, env, lock, " -> "DDESCR"@"DFID"\n",
1592 PDESCR(desc), PFID(lu_object_fid(&desc->cld_obj->co_lu)));
1594 * Just replace description in place. Nothing more is needed for
1595 * now. If locks were indexed according to their extent and/or mode,
1596 * that index would have to be updated here.
1598 cfs_spin_lock(&hdr->coh_lock_guard);
1599 lock->cll_descr = *desc;
1600 cfs_spin_unlock(&hdr->coh_lock_guard);
1603 EXPORT_SYMBOL(cl_lock_modify);
1606 * Initializes lock closure with a given origin.
1608 * \see cl_lock_closure
1610 void cl_lock_closure_init(const struct lu_env *env,
1611 struct cl_lock_closure *closure,
1612 struct cl_lock *origin, int wait)
1614 LINVRNT(cl_lock_is_mutexed(origin));
1615 LINVRNT(cl_lock_invariant(env, origin));
1617 CFS_INIT_LIST_HEAD(&closure->clc_list);
1618 closure->clc_origin = origin;
1619 closure->clc_wait = wait;
1620 closure->clc_nr = 0;
1622 EXPORT_SYMBOL(cl_lock_closure_init);
1625 * Builds a closure of \a lock.
1627 * Building of a closure consists of adding initial lock (\a lock) into it,
1628 * and calling cl_lock_operations::clo_closure() methods of \a lock. These
1629 * methods might call cl_lock_closure_build() recursively again, adding more
1630 * locks to the closure, etc.
1632 * \see cl_lock_closure
1634 int cl_lock_closure_build(const struct lu_env *env, struct cl_lock *lock,
1635 struct cl_lock_closure *closure)
1637 const struct cl_lock_slice *slice;
1641 LINVRNT(cl_lock_is_mutexed(closure->clc_origin));
1642 LINVRNT(cl_lock_invariant(env, closure->clc_origin));
1644 result = cl_lock_enclosure(env, lock, closure);
1646 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1647 if (slice->cls_ops->clo_closure != NULL) {
1648 result = slice->cls_ops->clo_closure(env, slice,
1656 cl_lock_disclosure(env, closure);
1659 EXPORT_SYMBOL(cl_lock_closure_build);
1662 * Adds new lock to a closure.
1664 * Try-locks \a lock and if succeeded, adds it to the closure (never more than
1665 * once). If try-lock failed, returns CLO_REPEAT, after optionally waiting
1666 * until next try-lock is likely to succeed.
1668 int cl_lock_enclosure(const struct lu_env *env, struct cl_lock *lock,
1669 struct cl_lock_closure *closure)
1673 cl_lock_trace(D_DLMTRACE, env, "enclosure lock", lock);
1674 if (!cl_lock_mutex_try(env, lock)) {
1676 * If lock->cll_inclosure is not empty, lock is already in
1679 if (cfs_list_empty(&lock->cll_inclosure)) {
1680 cl_lock_get_trust(lock);
1681 lu_ref_add(&lock->cll_reference, "closure", closure);
1682 cfs_list_add(&lock->cll_inclosure, &closure->clc_list);
1685 cl_lock_mutex_put(env, lock);
1688 cl_lock_disclosure(env, closure);
1689 if (closure->clc_wait) {
1690 cl_lock_get_trust(lock);
1691 lu_ref_add(&lock->cll_reference, "closure-w", closure);
1692 cl_lock_mutex_put(env, closure->clc_origin);
1694 LASSERT(cl_lock_nr_mutexed(env) == 0);
1695 cl_lock_mutex_get(env, lock);
1696 cl_lock_mutex_put(env, lock);
1698 cl_lock_mutex_get(env, closure->clc_origin);
1699 lu_ref_del(&lock->cll_reference, "closure-w", closure);
1700 cl_lock_put(env, lock);
1702 result = CLO_REPEAT;
1706 EXPORT_SYMBOL(cl_lock_enclosure);
1708 /** Releases mutices of enclosed locks. */
1709 void cl_lock_disclosure(const struct lu_env *env,
1710 struct cl_lock_closure *closure)
1712 struct cl_lock *scan;
1713 struct cl_lock *temp;
1715 cl_lock_trace(D_DLMTRACE, env, "disclosure lock", closure->clc_origin);
1716 cfs_list_for_each_entry_safe(scan, temp, &closure->clc_list,
1718 cfs_list_del_init(&scan->cll_inclosure);
1719 cl_lock_mutex_put(env, scan);
1720 lu_ref_del(&scan->cll_reference, "closure", closure);
1721 cl_lock_put(env, scan);
1724 LASSERT(closure->clc_nr == 0);
1726 EXPORT_SYMBOL(cl_lock_disclosure);
1728 /** Finalizes a closure. */
1729 void cl_lock_closure_fini(struct cl_lock_closure *closure)
1731 LASSERT(closure->clc_nr == 0);
1732 LASSERT(cfs_list_empty(&closure->clc_list));
1734 EXPORT_SYMBOL(cl_lock_closure_fini);
1737 * Destroys this lock. Notifies layers (bottom-to-top) that lock is being
1738 * destroyed, then destroy the lock. If there are holds on the lock, postpone
1739 * destruction until all holds are released. This is called when a decision is
1740 * made to destroy the lock in the future. E.g., when a blocking AST is
1741 * received on it, or fatal communication error happens.
1743 * Caller must have a reference on this lock to prevent a situation, when
1744 * deleted lock lingers in memory for indefinite time, because nobody calls
1745 * cl_lock_put() to finish it.
1747 * \pre atomic_read(&lock->cll_ref) > 0
1748 * \pre ergo(cl_lock_nesting(lock) == CNL_TOP,
1749 * cl_lock_nr_mutexed(env) == 1)
1750 * [i.e., if a top-lock is deleted, mutices of no other locks can be
1751 * held, as deletion of sub-locks might require releasing a top-lock
1754 * \see cl_lock_operations::clo_delete()
1755 * \see cl_lock::cll_holds
1757 void cl_lock_delete(const struct lu_env *env, struct cl_lock *lock)
1759 LINVRNT(cl_lock_is_mutexed(lock));
1760 LINVRNT(cl_lock_invariant(env, lock));
1761 LASSERT(ergo(cl_lock_nesting(lock) == CNL_TOP,
1762 cl_lock_nr_mutexed(env) == 1));
1765 cl_lock_trace(D_DLMTRACE, env, "delete lock", lock);
1766 if (lock->cll_holds == 0)
1767 cl_lock_delete0(env, lock);
1769 lock->cll_flags |= CLF_DOOMED;
1772 EXPORT_SYMBOL(cl_lock_delete);
1775 * Mark lock as irrecoverably failed, and mark it for destruction. This
1776 * happens when, e.g., server fails to grant a lock to us, or networking
1779 * \pre atomic_read(&lock->cll_ref) > 0
1781 * \see clo_lock_delete()
1782 * \see cl_lock::cll_holds
1784 void cl_lock_error(const struct lu_env *env, struct cl_lock *lock, int error)
1786 LINVRNT(cl_lock_is_mutexed(lock));
1787 LINVRNT(cl_lock_invariant(env, lock));
1790 cl_lock_trace(D_DLMTRACE, env, "set lock error", lock);
1791 if (lock->cll_error == 0 && error != 0) {
1792 lock->cll_error = error;
1793 cl_lock_signal(env, lock);
1794 cl_lock_cancel(env, lock);
1795 cl_lock_delete(env, lock);
1799 EXPORT_SYMBOL(cl_lock_error);
1802 * Cancels this lock. Notifies layers
1803 * (bottom-to-top) that lock is being cancelled, then destroy the lock. If
1804 * there are holds on the lock, postpone cancellation until
1805 * all holds are released.
1807 * Cancellation notification is delivered to layers at most once.
1809 * \see cl_lock_operations::clo_cancel()
1810 * \see cl_lock::cll_holds
1812 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock)
1814 LINVRNT(cl_lock_is_mutexed(lock));
1815 LINVRNT(cl_lock_invariant(env, lock));
1818 cl_lock_trace(D_DLMTRACE, env, "cancel lock", lock);
1819 if (lock->cll_holds == 0)
1820 cl_lock_cancel0(env, lock);
1822 lock->cll_flags |= CLF_CANCELPEND;
1825 EXPORT_SYMBOL(cl_lock_cancel);
1828 * Finds an existing lock covering given page and optionally different from a
1829 * given \a except lock.
1831 struct cl_lock *cl_lock_at_page(const struct lu_env *env, struct cl_object *obj,
1832 struct cl_page *page, struct cl_lock *except,
1833 int pending, int canceld)
1835 struct cl_object_header *head;
1836 struct cl_lock *scan;
1837 struct cl_lock *lock;
1838 struct cl_lock_descr *need;
1842 head = cl_object_header(obj);
1843 need = &cl_env_info(env)->clt_descr;
1846 need->cld_mode = CLM_READ; /* CLM_READ matches both READ & WRITE, but
1848 need->cld_start = need->cld_end = page->cp_index;
1849 need->cld_enq_flags = 0;
1851 cfs_spin_lock(&head->coh_lock_guard);
1852 /* It is fine to match any group lock since there could be only one
1853 * with a uniq gid and it conflicts with all other lock modes too */
1854 cfs_list_for_each_entry(scan, &head->coh_locks, cll_linkage) {
1855 if (scan != except &&
1856 (scan->cll_descr.cld_mode == CLM_GROUP ||
1857 cl_lock_ext_match(&scan->cll_descr, need)) &&
1858 scan->cll_state >= CLS_HELD &&
1859 scan->cll_state < CLS_FREEING &&
1861 * This check is racy as the lock can be canceled right
1862 * after it is done, but this is fine, because page exists
1865 (canceld || !(scan->cll_flags & CLF_CANCELLED)) &&
1866 (pending || !(scan->cll_flags & CLF_CANCELPEND))) {
1867 /* Don't increase cs_hit here since this
1868 * is just a helper function. */
1869 cl_lock_get_trust(scan);
1874 cfs_spin_unlock(&head->coh_lock_guard);
1877 EXPORT_SYMBOL(cl_lock_at_page);
1880 * Returns a list of pages protected (only) by a given lock.
1882 * Scans an extent of page radix tree, corresponding to the \a lock and queues
1883 * all pages that are not protected by locks other than \a lock into \a queue.
1885 void cl_lock_page_list_fixup(const struct lu_env *env,
1886 struct cl_io *io, struct cl_lock *lock,
1887 struct cl_page_list *queue)
1889 struct cl_page *page;
1890 struct cl_page *temp;
1891 struct cl_page_list *plist = &cl_env_info(env)->clt_list;
1893 LINVRNT(cl_lock_invariant(env, lock));
1896 /* Now, we have a list of cl_pages under the \a lock, we need
1897 * to check if some of pages are covered by other ldlm lock.
1898 * If this is the case, they aren't needed to be written out this time.
1900 * For example, we have A:[0,200] & B:[100,300] PW locks on client, now
1901 * the latter is to be canceled, this means other client is
1902 * reading/writing [200,300] since A won't canceled. Actually
1903 * we just need to write the pages covered by [200,300]. This is safe,
1904 * since [100,200] is also protected lock A.
1907 cl_page_list_init(plist);
1908 cl_page_list_for_each_safe(page, temp, queue) {
1909 pgoff_t idx = page->cp_index;
1910 struct cl_lock *found;
1911 struct cl_lock_descr *descr;
1913 /* The algorithm counts on the index-ascending page index. */
1914 LASSERT(ergo(&temp->cp_batch != &queue->pl_pages,
1915 page->cp_index < temp->cp_index));
1917 found = cl_lock_at_page(env, lock->cll_descr.cld_obj,
1922 descr = &found->cll_descr;
1923 cfs_list_for_each_entry_safe_from(page, temp, &queue->pl_pages,
1925 idx = page->cp_index;
1926 if (descr->cld_start > idx || descr->cld_end < idx)
1928 cl_page_list_move(plist, queue, page);
1930 cl_lock_put(env, found);
1933 /* The pages in plist are covered by other locks, don't handle them
1937 cl_page_list_disown(env, io, plist);
1938 cl_page_list_fini(env, plist);
1941 EXPORT_SYMBOL(cl_lock_page_list_fixup);
1944 * Invalidate pages protected by the given lock, sending them out to the
1945 * server first, if necessary.
1947 * This function does the following:
1949 * - collects a list of pages to be invalidated,
1951 * - unmaps them from the user virtual memory,
1953 * - sends dirty pages to the server,
1955 * - waits for transfer completion,
1957 * - discards pages, and throws them out of memory.
1959 * If \a discard is set, pages are discarded without sending them to the
1962 * If error happens on any step, the process continues anyway (the reasoning
1963 * behind this being that lock cancellation cannot be delayed indefinitely).
1965 int cl_lock_page_out(const struct lu_env *env, struct cl_lock *lock,
1968 struct cl_thread_info *info = cl_env_info(env);
1969 struct cl_io *io = &info->clt_io;
1970 struct cl_2queue *queue = &info->clt_queue;
1971 struct cl_lock_descr *descr = &lock->cll_descr;
1973 int nonblock = 1, resched;
1976 LINVRNT(cl_lock_invariant(env, lock));
1979 io->ci_obj = cl_object_top(descr->cld_obj);
1980 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1985 cl_2queue_init(queue);
1986 cl_page_gang_lookup(env, descr->cld_obj, io, descr->cld_start,
1987 descr->cld_end, &queue->c2_qin, nonblock,
1989 page_count = queue->c2_qin.pl_nr;
1990 if (page_count > 0) {
1991 result = cl_page_list_unmap(env, io, &queue->c2_qin);
1993 long timeout = 600; /* 10 minutes. */
1994 /* for debug purpose, if this request can't be
1995 * finished in 10 minutes, we hope it can
1998 result = cl_io_submit_sync(env, io, CRT_WRITE,
2002 CWARN("Writing %lu pages error: %d\n",
2003 page_count, result);
2005 cl_lock_page_list_fixup(env, io, lock, &queue->c2_qout);
2006 cl_2queue_discard(env, io, queue);
2007 cl_2queue_disown(env, io, queue);
2009 cl_2queue_fini(env, queue);
2013 } while (resched || nonblock--);
2015 cl_io_fini(env, io);
2018 EXPORT_SYMBOL(cl_lock_page_out);
2021 * Eliminate all locks for a given object.
2023 * Caller has to guarantee that no lock is in active use.
2025 * \param cancel when this is set, cl_locks_prune() cancels locks before
2028 void cl_locks_prune(const struct lu_env *env, struct cl_object *obj, int cancel)
2030 struct cl_object_header *head;
2031 struct cl_lock *lock;
2034 head = cl_object_header(obj);
2036 * If locks are destroyed without cancellation, all pages must be
2037 * already destroyed (as otherwise they will be left unprotected).
2039 LASSERT(ergo(!cancel,
2040 head->coh_tree.rnode == NULL && head->coh_pages == 0));
2042 cfs_spin_lock(&head->coh_lock_guard);
2043 while (!cfs_list_empty(&head->coh_locks)) {
2044 lock = container_of(head->coh_locks.next,
2045 struct cl_lock, cll_linkage);
2046 cl_lock_get_trust(lock);
2047 cfs_spin_unlock(&head->coh_lock_guard);
2048 lu_ref_add(&lock->cll_reference, "prune", cfs_current());
2049 cl_lock_mutex_get(env, lock);
2050 if (lock->cll_state < CLS_FREEING) {
2051 LASSERT(lock->cll_holds == 0);
2052 LASSERT(lock->cll_users == 0);
2054 cl_lock_cancel(env, lock);
2055 cl_lock_delete(env, lock);
2057 cl_lock_mutex_put(env, lock);
2058 lu_ref_del(&lock->cll_reference, "prune", cfs_current());
2059 cl_lock_put(env, lock);
2060 cfs_spin_lock(&head->coh_lock_guard);
2062 cfs_spin_unlock(&head->coh_lock_guard);
2065 EXPORT_SYMBOL(cl_locks_prune);
2067 static struct cl_lock *cl_lock_hold_mutex(const struct lu_env *env,
2068 const struct cl_io *io,
2069 const struct cl_lock_descr *need,
2070 const char *scope, const void *source)
2072 struct cl_lock *lock;
2077 lock = cl_lock_find(env, io, need);
2080 cl_lock_mutex_get(env, lock);
2081 if (lock->cll_state < CLS_FREEING &&
2082 !(lock->cll_flags & CLF_CANCELLED)) {
2083 cl_lock_hold_mod(env, lock, +1);
2084 lu_ref_add(&lock->cll_holders, scope, source);
2085 lu_ref_add(&lock->cll_reference, scope, source);
2088 cl_lock_mutex_put(env, lock);
2089 cl_lock_put(env, lock);
2095 * Returns a lock matching \a need description with a reference and a hold on
2098 * This is much like cl_lock_find(), except that cl_lock_hold() additionally
2099 * guarantees that lock is not in the CLS_FREEING state on return.
2101 struct cl_lock *cl_lock_hold(const struct lu_env *env, const struct cl_io *io,
2102 const struct cl_lock_descr *need,
2103 const char *scope, const void *source)
2105 struct cl_lock *lock;
2109 lock = cl_lock_hold_mutex(env, io, need, scope, source);
2111 cl_lock_mutex_put(env, lock);
2114 EXPORT_SYMBOL(cl_lock_hold);
2117 * Main high-level entry point of cl_lock interface that finds existing or
2118 * enqueues new lock matching given description.
2120 struct cl_lock *cl_lock_request(const struct lu_env *env, struct cl_io *io,
2121 const struct cl_lock_descr *need,
2122 const char *scope, const void *source)
2124 struct cl_lock *lock;
2126 __u32 enqflags = need->cld_enq_flags;
2130 lock = cl_lock_hold_mutex(env, io, need, scope, source);
2131 if (!IS_ERR(lock)) {
2132 rc = cl_enqueue_locked(env, lock, io, enqflags);
2134 if (cl_lock_fits_into(env, lock, need, io)) {
2135 cl_lock_mutex_put(env, lock);
2136 cl_lock_lockdep_acquire(env,
2140 cl_unuse_locked(env, lock);
2142 cl_lock_trace(D_DLMTRACE, env, "enqueue failed", lock);
2143 cl_lock_hold_release(env, lock, scope, source);
2144 cl_lock_mutex_put(env, lock);
2145 lu_ref_del(&lock->cll_reference, scope, source);
2146 cl_lock_put(env, lock);
2153 EXPORT_SYMBOL(cl_lock_request);
2156 * Adds a hold to a known lock.
2158 void cl_lock_hold_add(const struct lu_env *env, struct cl_lock *lock,
2159 const char *scope, const void *source)
2161 LINVRNT(cl_lock_is_mutexed(lock));
2162 LINVRNT(cl_lock_invariant(env, lock));
2163 LASSERT(lock->cll_state != CLS_FREEING);
2166 cl_lock_hold_mod(env, lock, +1);
2168 lu_ref_add(&lock->cll_holders, scope, source);
2169 lu_ref_add(&lock->cll_reference, scope, source);
2172 EXPORT_SYMBOL(cl_lock_hold_add);
2175 * Releases a hold and a reference on a lock, on which caller acquired a
2178 void cl_lock_unhold(const struct lu_env *env, struct cl_lock *lock,
2179 const char *scope, const void *source)
2181 LINVRNT(cl_lock_invariant(env, lock));
2183 cl_lock_hold_release(env, lock, scope, source);
2184 lu_ref_del(&lock->cll_reference, scope, source);
2185 cl_lock_put(env, lock);
2188 EXPORT_SYMBOL(cl_lock_unhold);
2191 * Releases a hold and a reference on a lock, obtained by cl_lock_hold().
2193 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock,
2194 const char *scope, const void *source)
2196 LINVRNT(cl_lock_invariant(env, lock));
2198 cl_lock_trace(D_DLMTRACE, env, "release lock", lock);
2199 cl_lock_mutex_get(env, lock);
2200 cl_lock_hold_release(env, lock, scope, source);
2201 cl_lock_mutex_put(env, lock);
2202 lu_ref_del(&lock->cll_reference, scope, source);
2203 cl_lock_put(env, lock);
2206 EXPORT_SYMBOL(cl_lock_release);
2208 void cl_lock_user_add(const struct lu_env *env, struct cl_lock *lock)
2210 LINVRNT(cl_lock_is_mutexed(lock));
2211 LINVRNT(cl_lock_invariant(env, lock));
2214 cl_lock_used_mod(env, lock, +1);
2217 EXPORT_SYMBOL(cl_lock_user_add);
2219 int cl_lock_user_del(const struct lu_env *env, struct cl_lock *lock)
2221 LINVRNT(cl_lock_is_mutexed(lock));
2222 LINVRNT(cl_lock_invariant(env, lock));
2223 LASSERT(lock->cll_users > 0);
2226 cl_lock_used_mod(env, lock, -1);
2227 RETURN(lock->cll_users == 0);
2229 EXPORT_SYMBOL(cl_lock_user_del);
2231 const char *cl_lock_mode_name(const enum cl_lock_mode mode)
2233 static const char *names[] = {
2234 [CLM_PHANTOM] = "P",
2239 if (0 <= mode && mode < ARRAY_SIZE(names))
2244 EXPORT_SYMBOL(cl_lock_mode_name);
2247 * Prints human readable representation of a lock description.
2249 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2250 lu_printer_t printer,
2251 const struct cl_lock_descr *descr)
2253 const struct lu_fid *fid;
2255 fid = lu_object_fid(&descr->cld_obj->co_lu);
2256 (*printer)(env, cookie, DDESCR"@"DFID, PDESCR(descr), PFID(fid));
2258 EXPORT_SYMBOL(cl_lock_descr_print);
2261 * Prints human readable representation of \a lock to the \a f.
2263 void cl_lock_print(const struct lu_env *env, void *cookie,
2264 lu_printer_t printer, const struct cl_lock *lock)
2266 const struct cl_lock_slice *slice;
2267 (*printer)(env, cookie, "lock@%p[%d %d %d %d %d %08lx] ",
2268 lock, cfs_atomic_read(&lock->cll_ref),
2269 lock->cll_state, lock->cll_error, lock->cll_holds,
2270 lock->cll_users, lock->cll_flags);
2271 cl_lock_descr_print(env, cookie, printer, &lock->cll_descr);
2272 (*printer)(env, cookie, " {\n");
2274 cfs_list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
2275 (*printer)(env, cookie, " %s@%p: ",
2276 slice->cls_obj->co_lu.lo_dev->ld_type->ldt_name,
2278 if (slice->cls_ops->clo_print != NULL)
2279 slice->cls_ops->clo_print(env, cookie, printer, slice);
2280 (*printer)(env, cookie, "\n");
2282 (*printer)(env, cookie, "} lock@%p\n", lock);
2284 EXPORT_SYMBOL(cl_lock_print);
2286 int cl_lock_init(void)
2288 return lu_kmem_init(cl_lock_caches);
2291 void cl_lock_fini(void)
2293 lu_kmem_fini(cl_lock_caches);