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 2008 Sun Microsystems, Inc. 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 struct lock_class_key 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)
82 ergo(lock->cll_state == CLS_FREEING, lock->cll_holds == 0) &&
83 atomic_read(&lock->cll_ref) >= lock->cll_holds &&
84 lock->cll_holds >= lock->cll_users &&
85 lock->cll_holds >= 0 &&
86 lock->cll_users >= 0 &&
91 * Stronger lock invariant, checking that caller has a reference on a lock.
93 * \see cl_lock_invariant_trusted()
95 static int cl_lock_invariant(const struct lu_env *env,
96 const struct cl_lock *lock)
100 result = atomic_read(&lock->cll_ref) > 0 &&
101 cl_lock_invariant_trusted(env, lock);
102 if (!result && env != NULL)
103 CL_LOCK_DEBUG(D_ERROR, env, lock, "invariant broken");
108 * Returns lock "nesting": 0 for a top-lock and 1 for a sub-lock.
110 static enum clt_nesting_level cl_lock_nesting(const struct cl_lock *lock)
112 return cl_object_header(lock->cll_descr.cld_obj)->coh_nesting;
116 * Returns a set of counters for this lock, depending on a lock nesting.
118 static struct cl_thread_counters *cl_lock_counters(const struct lu_env *env,
119 const struct cl_lock *lock)
121 struct cl_thread_info *info;
122 enum clt_nesting_level nesting;
124 info = cl_env_info(env);
125 nesting = cl_lock_nesting(lock);
126 LASSERT(nesting < ARRAY_SIZE(info->clt_counters));
127 return &info->clt_counters[nesting];
130 static void cl_lock_trace0(int level, const struct lu_env *env,
131 const char *prefix, const struct cl_lock *lock,
132 const char *func, const int line)
134 struct cl_object_header *h = cl_object_header(lock->cll_descr.cld_obj);
135 CDEBUG(level, "%s: %p@(%i %p %i %d %d %d %d %lx)"
136 "(%p/%d/%i) at %s():%d\n",
138 atomic_read(&lock->cll_ref), lock->cll_guarder, lock->cll_depth,
139 lock->cll_state, lock->cll_error, lock->cll_holds,
140 lock->cll_users, lock->cll_flags,
141 env, h->coh_nesting, cl_lock_nr_mutexed(env),
144 #define cl_lock_trace(level, env, prefix, lock) \
145 cl_lock_trace0(level, env, prefix, lock, __FUNCTION__, __LINE__)
147 #define RETIP ((unsigned long)__builtin_return_address(0))
149 #ifdef CONFIG_LOCKDEP
150 static struct lock_class_key cl_lock_key;
152 static void cl_lock_lockdep_init(struct cl_lock *lock)
154 lockdep_set_class_and_name(lock, &cl_lock_key, "EXT");
157 static void cl_lock_lockdep_acquire(const struct lu_env *env,
158 struct cl_lock *lock, __u32 enqflags)
160 cl_lock_counters(env, lock)->ctc_nr_locks_acquired++;
161 lock_acquire(&lock->dep_map, !!(enqflags & CEF_ASYNC),
162 /* try: */ 0, lock->cll_descr.cld_mode <= CLM_READ,
163 /* check: */ 2, RETIP);
166 static void cl_lock_lockdep_release(const struct lu_env *env,
167 struct cl_lock *lock)
169 cl_lock_counters(env, lock)->ctc_nr_locks_acquired--;
170 lock_release(&lock->dep_map, 0, RETIP);
173 #else /* !CONFIG_LOCKDEP */
175 static void cl_lock_lockdep_init(struct cl_lock *lock)
177 static void cl_lock_lockdep_acquire(const struct lu_env *env,
178 struct cl_lock *lock, __u32 enqflags)
180 static void cl_lock_lockdep_release(const struct lu_env *env,
181 struct cl_lock *lock)
184 #endif /* !CONFIG_LOCKDEP */
187 * Adds lock slice to the compound lock.
189 * This is called by cl_object_operations::coo_lock_init() methods to add a
190 * per-layer state to the lock. New state is added at the end of
191 * cl_lock::cll_layers list, that is, it is at the bottom of the stack.
193 * \see cl_req_slice_add(), cl_page_slice_add(), cl_io_slice_add()
195 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
196 struct cl_object *obj,
197 const struct cl_lock_operations *ops)
200 slice->cls_lock = lock;
201 list_add_tail(&slice->cls_linkage, &lock->cll_layers);
202 slice->cls_obj = obj;
203 slice->cls_ops = ops;
206 EXPORT_SYMBOL(cl_lock_slice_add);
209 * Returns true iff a lock with the mode \a has provides at least the same
210 * guarantees as a lock with the mode \a need.
212 int cl_lock_mode_match(enum cl_lock_mode has, enum cl_lock_mode need)
214 LINVRNT(need == CLM_READ || need == CLM_WRITE ||
215 need == CLM_PHANTOM || need == CLM_GROUP);
216 LINVRNT(has == CLM_READ || has == CLM_WRITE ||
217 has == CLM_PHANTOM || has == CLM_GROUP);
218 CLASSERT(CLM_PHANTOM < CLM_READ);
219 CLASSERT(CLM_READ < CLM_WRITE);
220 CLASSERT(CLM_WRITE < CLM_GROUP);
222 if (has != CLM_GROUP)
227 EXPORT_SYMBOL(cl_lock_mode_match);
230 * Returns true iff extent portions of lock descriptions match.
232 int cl_lock_ext_match(const struct cl_lock_descr *has,
233 const struct cl_lock_descr *need)
236 has->cld_start <= need->cld_start &&
237 has->cld_end >= need->cld_end &&
238 cl_lock_mode_match(has->cld_mode, need->cld_mode) &&
239 (has->cld_mode != CLM_GROUP || has->cld_gid == need->cld_gid);
241 EXPORT_SYMBOL(cl_lock_ext_match);
244 * Returns true iff a lock with the description \a has provides at least the
245 * same guarantees as a lock with the description \a need.
247 int cl_lock_descr_match(const struct cl_lock_descr *has,
248 const struct cl_lock_descr *need)
251 cl_object_same(has->cld_obj, need->cld_obj) &&
252 cl_lock_ext_match(has, need);
254 EXPORT_SYMBOL(cl_lock_descr_match);
256 static void cl_lock_free(const struct lu_env *env, struct cl_lock *lock)
258 struct cl_object *obj = lock->cll_descr.cld_obj;
260 LASSERT(cl_is_lock(lock));
261 LINVRNT(!cl_lock_is_mutexed(lock));
264 cl_lock_trace(D_DLMTRACE, env, "free lock", lock);
266 while (!list_empty(&lock->cll_layers)) {
267 struct cl_lock_slice *slice;
269 slice = list_entry(lock->cll_layers.next, struct cl_lock_slice,
271 list_del_init(lock->cll_layers.next);
272 slice->cls_ops->clo_fini(env, slice);
274 atomic_dec(&cl_object_site(obj)->cs_locks.cs_total);
275 atomic_dec(&cl_object_site(obj)->cs_locks_state[lock->cll_state]);
276 lu_object_ref_del_at(&obj->co_lu, lock->cll_obj_ref, "cl_lock", lock);
277 cl_object_put(env, obj);
278 lu_ref_fini(&lock->cll_reference);
279 lu_ref_fini(&lock->cll_holders);
280 mutex_destroy(&lock->cll_guard);
281 OBD_SLAB_FREE_PTR(lock, cl_lock_kmem);
286 * Releases a reference on a lock.
288 * When last reference is released, lock is returned to the cache, unless it
289 * is in cl_lock_state::CLS_FREEING state, in which case it is destroyed
292 * \see cl_object_put(), cl_page_put()
294 void cl_lock_put(const struct lu_env *env, struct cl_lock *lock)
296 struct cl_object *obj;
297 struct cl_object_header *head;
298 struct cl_site *site;
300 LINVRNT(cl_lock_invariant(env, lock));
302 obj = lock->cll_descr.cld_obj;
303 LINVRNT(obj != NULL);
304 head = cl_object_header(obj);
305 site = cl_object_site(obj);
307 CDEBUG(D_TRACE, "releasing reference: %d %p %lu\n",
308 atomic_read(&lock->cll_ref), lock, RETIP);
310 if (atomic_dec_and_test(&lock->cll_ref)) {
311 if (lock->cll_state == CLS_FREEING) {
312 LASSERT(list_empty(&lock->cll_linkage));
313 cl_lock_free(env, lock);
315 atomic_dec(&site->cs_locks.cs_busy);
319 EXPORT_SYMBOL(cl_lock_put);
322 * Acquires an additional reference to a lock.
324 * This can be called only by caller already possessing a reference to \a
327 * \see cl_object_get(), cl_page_get()
329 void cl_lock_get(struct cl_lock *lock)
331 LINVRNT(cl_lock_invariant(NULL, lock));
332 CDEBUG(D_TRACE, "acquiring reference: %d %p %lu\n",
333 atomic_read(&lock->cll_ref), lock, RETIP);
334 atomic_inc(&lock->cll_ref);
336 EXPORT_SYMBOL(cl_lock_get);
339 * Acquires a reference to a lock.
341 * This is much like cl_lock_get(), except that this function can be used to
342 * acquire initial reference to the cached lock. Caller has to deal with all
343 * possible races. Use with care!
345 * \see cl_page_get_trust()
347 void cl_lock_get_trust(struct cl_lock *lock)
349 struct cl_site *site = cl_object_site(lock->cll_descr.cld_obj);
351 LASSERT(cl_is_lock(lock));
352 CDEBUG(D_TRACE, "acquiring trusted reference: %d %p %lu\n",
353 atomic_read(&lock->cll_ref), lock, RETIP);
354 if (atomic_inc_return(&lock->cll_ref) == 1)
355 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 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 mutex_init(&lock->cll_guard);
398 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 atomic_inc(&site->cs_locks_state[CLS_NEW]);
402 atomic_inc(&site->cs_locks.cs_total);
403 atomic_inc(&site->cs_locks.cs_created);
404 cl_lock_lockdep_init(lock);
405 list_for_each_entry(obj, &head->loh_layers, co_lu.lo_linkage) {
408 err = obj->co_ops->coo_lock_init(env, obj, lock, io);
410 cl_lock_finish(env, lock);
416 lock = ERR_PTR(-ENOMEM);
421 * Transfer the lock into INTRANSIT state and return the original state.
423 * \pre state: CLS_CACHED, CLS_HELD or CLS_ENQUEUED
424 * \post state: CLS_INTRANSIT
427 enum cl_lock_state cl_lock_intransit(const struct lu_env *env,
428 struct cl_lock *lock)
430 enum cl_lock_state state = lock->cll_state;
432 LASSERT(cl_lock_is_mutexed(lock));
433 LASSERT(state != CLS_INTRANSIT);
434 LASSERTF(state >= CLS_ENQUEUED && state <= CLS_CACHED,
435 "Malformed lock state %d.\n", state);
437 cl_lock_state_set(env, lock, CLS_INTRANSIT);
438 lock->cll_intransit_owner = cfs_current();
439 cl_lock_hold_add(env, lock, "intransit", cfs_current());
442 EXPORT_SYMBOL(cl_lock_intransit);
445 * Exit the intransit state and restore the lock state to the original state
447 void cl_lock_extransit(const struct lu_env *env, struct cl_lock *lock,
448 enum cl_lock_state state)
450 LASSERT(cl_lock_is_mutexed(lock));
451 LASSERT(lock->cll_state == CLS_INTRANSIT);
452 LASSERT(state != CLS_INTRANSIT);
453 LASSERT(lock->cll_intransit_owner == cfs_current());
455 lock->cll_intransit_owner = NULL;
456 cl_lock_state_set(env, lock, state);
457 cl_lock_unhold(env, lock, "intransit", cfs_current());
459 EXPORT_SYMBOL(cl_lock_extransit);
462 * Checking whether the lock is intransit state
464 int cl_lock_is_intransit(struct cl_lock *lock)
466 LASSERT(cl_lock_is_mutexed(lock));
467 return lock->cll_state == CLS_INTRANSIT &&
468 lock->cll_intransit_owner != cfs_current();
470 EXPORT_SYMBOL(cl_lock_is_intransit);
472 * Returns true iff lock is "suitable" for given io. E.g., locks acquired by
473 * truncate and O_APPEND cannot be reused for read/non-append-write, as they
474 * cover multiple stripes and can trigger cascading timeouts.
476 static int cl_lock_fits_into(const struct lu_env *env,
477 const struct cl_lock *lock,
478 const struct cl_lock_descr *need,
479 const struct cl_io *io)
481 const struct cl_lock_slice *slice;
483 LINVRNT(cl_lock_invariant_trusted(env, lock));
485 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
486 if (slice->cls_ops->clo_fits_into != NULL &&
487 !slice->cls_ops->clo_fits_into(env, slice, need, io))
493 static struct cl_lock *cl_lock_lookup(const struct lu_env *env,
494 struct cl_object *obj,
495 const struct cl_io *io,
496 const struct cl_lock_descr *need)
498 struct cl_lock *lock;
499 struct cl_object_header *head;
500 struct cl_site *site;
504 head = cl_object_header(obj);
505 site = cl_object_site(obj);
506 LINVRNT_SPIN_LOCKED(&head->coh_lock_guard);
507 atomic_inc(&site->cs_locks.cs_lookup);
508 list_for_each_entry(lock, &head->coh_locks, cll_linkage) {
511 LASSERT(cl_is_lock(lock));
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"(%i) need: "DDESCR": %d\n",
518 PDESCR(&lock->cll_descr), lock->cll_state, PDESCR(need),
521 cl_lock_get_trust(lock);
522 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 spin_lock(&head->coh_lock_guard);
555 lock = cl_lock_lookup(env, obj, io, need);
556 spin_unlock(&head->coh_lock_guard);
559 lock = cl_lock_alloc(env, obj, io, need);
561 struct cl_lock *ghost;
563 spin_lock(&head->coh_lock_guard);
564 ghost = cl_lock_lookup(env, obj, io, need);
566 list_add_tail(&lock->cll_linkage, &head->coh_locks);
567 spin_unlock(&head->coh_lock_guard);
568 atomic_inc(&site->cs_locks.cs_busy);
570 spin_unlock(&head->coh_lock_guard);
572 * Other threads can acquire references to the
573 * top-lock through its sub-locks. Hence, it
574 * cannot be cl_lock_free()-ed immediately.
576 cl_lock_finish(env, lock);
585 * Returns existing lock matching given description. This is similar to
586 * cl_lock_find() except that no new lock is created, and returned lock is
587 * guaranteed to be in enum cl_lock_state::CLS_HELD state.
589 struct cl_lock *cl_lock_peek(const struct lu_env *env, const struct cl_io *io,
590 const struct cl_lock_descr *need,
591 const char *scope, const void *source)
593 struct cl_object_header *head;
594 struct cl_object *obj;
595 struct cl_lock *lock;
599 head = cl_object_header(obj);
601 spin_lock(&head->coh_lock_guard);
602 lock = cl_lock_lookup(env, obj, io, need);
603 spin_unlock(&head->coh_lock_guard);
608 cl_lock_mutex_get(env, lock);
609 if (lock->cll_state == CLS_INTRANSIT)
610 cl_lock_state_wait(env, lock); /* Don't care return value. */
611 if (lock->cll_state == CLS_CACHED) {
613 result = cl_use_try(env, lock, 1);
615 cl_lock_error(env, lock, result);
617 ok = lock->cll_state == CLS_HELD;
619 cl_lock_hold_add(env, lock, scope, source);
620 cl_lock_user_add(env, lock);
621 cl_lock_put(env, lock);
623 cl_lock_mutex_put(env, lock);
625 cl_lock_put(env, lock);
631 EXPORT_SYMBOL(cl_lock_peek);
634 * Returns a slice within a lock, corresponding to the given layer in the
639 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
640 const struct lu_device_type *dtype)
642 const struct cl_lock_slice *slice;
644 LINVRNT(cl_lock_invariant_trusted(NULL, lock));
647 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
648 if (slice->cls_obj->co_lu.lo_dev->ld_type == dtype)
653 EXPORT_SYMBOL(cl_lock_at);
655 static void cl_lock_mutex_tail(const struct lu_env *env, struct cl_lock *lock)
657 struct cl_thread_counters *counters;
659 counters = cl_lock_counters(env, lock);
661 counters->ctc_nr_locks_locked++;
662 lu_ref_add(&counters->ctc_locks_locked, "cll_guard", lock);
663 cl_lock_trace(D_TRACE, env, "got mutex", lock);
667 * Locks cl_lock object.
669 * This is used to manipulate cl_lock fields, and to serialize state
670 * transitions in the lock state machine.
672 * \post cl_lock_is_mutexed(lock)
674 * \see cl_lock_mutex_put()
676 void cl_lock_mutex_get(const struct lu_env *env, struct cl_lock *lock)
678 LINVRNT(cl_lock_invariant(env, lock));
680 if (lock->cll_guarder == cfs_current()) {
681 LINVRNT(cl_lock_is_mutexed(lock));
682 LINVRNT(lock->cll_depth > 0);
684 struct cl_object_header *hdr;
685 struct cl_thread_info *info;
688 LINVRNT(lock->cll_guarder != cfs_current());
689 hdr = cl_object_header(lock->cll_descr.cld_obj);
691 * Check that mutices are taken in the bottom-to-top order.
693 info = cl_env_info(env);
694 for (i = 0; i < hdr->coh_nesting; ++i)
695 LASSERT(info->clt_counters[i].ctc_nr_locks_locked == 0);
696 mutex_lock_nested(&lock->cll_guard, hdr->coh_nesting);
697 lock->cll_guarder = cfs_current();
698 LINVRNT(lock->cll_depth == 0);
700 cl_lock_mutex_tail(env, lock);
702 EXPORT_SYMBOL(cl_lock_mutex_get);
705 * Try-locks cl_lock object.
707 * \retval 0 \a lock was successfully locked
709 * \retval -EBUSY \a lock cannot be locked right now
711 * \post ergo(result == 0, cl_lock_is_mutexed(lock))
713 * \see cl_lock_mutex_get()
715 int cl_lock_mutex_try(const struct lu_env *env, struct cl_lock *lock)
719 LINVRNT(cl_lock_invariant_trusted(env, lock));
723 if (lock->cll_guarder == cfs_current()) {
724 LINVRNT(lock->cll_depth > 0);
725 cl_lock_mutex_tail(env, lock);
726 } else if (mutex_trylock(&lock->cll_guard)) {
727 LINVRNT(lock->cll_depth == 0);
728 lock->cll_guarder = cfs_current();
729 cl_lock_mutex_tail(env, lock);
734 EXPORT_SYMBOL(cl_lock_mutex_try);
737 {* Unlocks cl_lock object.
739 * \pre cl_lock_is_mutexed(lock)
741 * \see cl_lock_mutex_get()
743 void cl_lock_mutex_put(const struct lu_env *env, struct cl_lock *lock)
745 struct cl_thread_counters *counters;
747 LINVRNT(cl_lock_invariant(env, lock));
748 LINVRNT(cl_lock_is_mutexed(lock));
749 LINVRNT(lock->cll_guarder == cfs_current());
750 LINVRNT(lock->cll_depth > 0);
752 counters = cl_lock_counters(env, lock);
753 LINVRNT(counters->ctc_nr_locks_locked > 0);
755 cl_lock_trace(D_TRACE, env, "put mutex", lock);
756 lu_ref_del(&counters->ctc_locks_locked, "cll_guard", lock);
757 counters->ctc_nr_locks_locked--;
758 if (--lock->cll_depth == 0) {
759 lock->cll_guarder = NULL;
760 mutex_unlock(&lock->cll_guard);
763 EXPORT_SYMBOL(cl_lock_mutex_put);
766 * Returns true iff lock's mutex is owned by the current thread.
768 int cl_lock_is_mutexed(struct cl_lock *lock)
770 return lock->cll_guarder == cfs_current();
772 EXPORT_SYMBOL(cl_lock_is_mutexed);
775 * Returns number of cl_lock mutices held by the current thread (environment).
777 int cl_lock_nr_mutexed(const struct lu_env *env)
779 struct cl_thread_info *info;
784 * NOTE: if summation across all nesting levels (currently 2) proves
785 * too expensive, a summary counter can be added to
786 * struct cl_thread_info.
788 info = cl_env_info(env);
789 for (i = 0, locked = 0; i < ARRAY_SIZE(info->clt_counters); ++i)
790 locked += info->clt_counters[i].ctc_nr_locks_locked;
793 EXPORT_SYMBOL(cl_lock_nr_mutexed);
795 static void cl_lock_cancel0(const struct lu_env *env, struct cl_lock *lock)
797 LINVRNT(cl_lock_is_mutexed(lock));
798 LINVRNT(cl_lock_invariant(env, lock));
800 if (!(lock->cll_flags & CLF_CANCELLED)) {
801 const struct cl_lock_slice *slice;
803 lock->cll_flags |= CLF_CANCELLED;
804 list_for_each_entry_reverse(slice, &lock->cll_layers,
806 if (slice->cls_ops->clo_cancel != NULL)
807 slice->cls_ops->clo_cancel(env, slice);
813 static void cl_lock_delete0(const struct lu_env *env, struct cl_lock *lock)
815 struct cl_object_header *head;
816 const struct cl_lock_slice *slice;
818 LINVRNT(cl_lock_is_mutexed(lock));
819 LINVRNT(cl_lock_invariant(env, lock));
822 if (lock->cll_state < CLS_FREEING) {
823 LASSERT(lock->cll_state != CLS_INTRANSIT);
824 cl_lock_state_set(env, lock, CLS_FREEING);
826 head = cl_object_header(lock->cll_descr.cld_obj);
828 spin_lock(&head->coh_lock_guard);
829 list_del_init(&lock->cll_linkage);
831 spin_unlock(&head->coh_lock_guard);
833 * From now on, no new references to this lock can be acquired
834 * by cl_lock_lookup().
836 list_for_each_entry_reverse(slice, &lock->cll_layers,
838 if (slice->cls_ops->clo_delete != NULL)
839 slice->cls_ops->clo_delete(env, slice);
842 * From now on, no new references to this lock can be acquired
843 * by layer-specific means (like a pointer from struct
844 * ldlm_lock in osc, or a pointer from top-lock to sub-lock in
847 * Lock will be finally freed in cl_lock_put() when last of
848 * existing references goes away.
855 * Mod(ifie)s cl_lock::cll_holds counter for a given lock. Also, for a
856 * top-lock (nesting == 0) accounts for this modification in the per-thread
857 * debugging counters. Sub-lock holds can be released by a thread different
858 * from one that acquired it.
860 static void cl_lock_hold_mod(const struct lu_env *env, struct cl_lock *lock,
863 struct cl_thread_counters *counters;
864 enum clt_nesting_level nesting;
866 lock->cll_holds += delta;
867 nesting = cl_lock_nesting(lock);
868 if (nesting == CNL_TOP) {
869 counters = &cl_env_info(env)->clt_counters[CNL_TOP];
870 counters->ctc_nr_held += delta;
871 LASSERT(counters->ctc_nr_held >= 0);
876 * Mod(ifie)s cl_lock::cll_users counter for a given lock. See
877 * cl_lock_hold_mod() for the explanation of the debugging code.
879 static void cl_lock_used_mod(const struct lu_env *env, struct cl_lock *lock,
882 struct cl_thread_counters *counters;
883 enum clt_nesting_level nesting;
885 lock->cll_users += delta;
886 nesting = cl_lock_nesting(lock);
887 if (nesting == CNL_TOP) {
888 counters = &cl_env_info(env)->clt_counters[CNL_TOP];
889 counters->ctc_nr_used += delta;
890 LASSERT(counters->ctc_nr_used >= 0);
894 static void cl_lock_hold_release(const struct lu_env *env, struct cl_lock *lock,
895 const char *scope, const void *source)
897 LINVRNT(cl_lock_is_mutexed(lock));
898 LINVRNT(cl_lock_invariant(env, lock));
899 LASSERT(lock->cll_holds > 0);
902 cl_lock_trace(D_DLMTRACE, env, "hold release lock", lock);
903 lu_ref_del(&lock->cll_holders, scope, source);
904 cl_lock_hold_mod(env, lock, -1);
905 if (lock->cll_holds == 0) {
906 if (lock->cll_descr.cld_mode == CLM_PHANTOM ||
907 lock->cll_descr.cld_mode == CLM_GROUP)
909 * If lock is still phantom or grouplock when user is
910 * done with it---destroy the lock.
912 lock->cll_flags |= CLF_CANCELPEND|CLF_DOOMED;
913 if (lock->cll_flags & CLF_CANCELPEND) {
914 lock->cll_flags &= ~CLF_CANCELPEND;
915 cl_lock_cancel0(env, lock);
917 if (lock->cll_flags & CLF_DOOMED) {
918 /* no longer doomed: it's dead... Jim. */
919 lock->cll_flags &= ~CLF_DOOMED;
920 cl_lock_delete0(env, lock);
928 * Waits until lock state is changed.
930 * This function is called with cl_lock mutex locked, atomically releases
931 * mutex and goes to sleep, waiting for a lock state change (signaled by
932 * cl_lock_signal()), and re-acquires the mutex before return.
934 * This function is used to wait until lock state machine makes some progress
935 * and to emulate synchronous operations on top of asynchronous lock
938 * \retval -EINTR wait was interrupted
940 * \retval 0 wait wasn't interrupted
942 * \pre cl_lock_is_mutexed(lock)
944 * \see cl_lock_signal()
946 int cl_lock_state_wait(const struct lu_env *env, struct cl_lock *lock)
948 cfs_waitlink_t waiter;
952 LINVRNT(cl_lock_is_mutexed(lock));
953 LINVRNT(cl_lock_invariant(env, lock));
954 LASSERT(lock->cll_depth == 1);
955 LASSERT(lock->cll_state != CLS_FREEING); /* too late to wait */
957 cl_lock_trace(D_DLMTRACE, env, "state wait lock", lock);
958 result = lock->cll_error;
960 cfs_waitlink_init(&waiter);
961 cfs_waitq_add(&lock->cll_wq, &waiter);
962 set_current_state(CFS_TASK_INTERRUPTIBLE);
963 cl_lock_mutex_put(env, lock);
965 LASSERT(cl_lock_nr_mutexed(env) == 0);
966 cfs_waitq_wait(&waiter, CFS_TASK_INTERRUPTIBLE);
968 cl_lock_mutex_get(env, lock);
969 set_current_state(CFS_TASK_RUNNING);
970 cfs_waitq_del(&lock->cll_wq, &waiter);
971 result = cfs_signal_pending() ? -EINTR : 0;
975 EXPORT_SYMBOL(cl_lock_state_wait);
977 static void cl_lock_state_signal(const struct lu_env *env, struct cl_lock *lock,
978 enum cl_lock_state state)
980 const struct cl_lock_slice *slice;
983 LINVRNT(cl_lock_is_mutexed(lock));
984 LINVRNT(cl_lock_invariant(env, lock));
986 list_for_each_entry(slice, &lock->cll_layers, cls_linkage)
987 if (slice->cls_ops->clo_state != NULL)
988 slice->cls_ops->clo_state(env, slice, state);
989 cfs_waitq_broadcast(&lock->cll_wq);
994 * Notifies waiters that lock state changed.
996 * Wakes up all waiters sleeping in cl_lock_state_wait(), also notifies all
997 * layers about state change by calling cl_lock_operations::clo_state()
1000 void cl_lock_signal(const struct lu_env *env, struct cl_lock *lock)
1003 cl_lock_trace(D_DLMTRACE, env, "state signal lock", lock);
1004 cl_lock_state_signal(env, lock, lock->cll_state);
1007 EXPORT_SYMBOL(cl_lock_signal);
1010 * Changes lock state.
1012 * This function is invoked to notify layers that lock state changed, possible
1013 * as a result of an asynchronous event such as call-back reception.
1015 * \post lock->cll_state == state
1017 * \see cl_lock_operations::clo_state()
1019 void cl_lock_state_set(const struct lu_env *env, struct cl_lock *lock,
1020 enum cl_lock_state state)
1022 struct cl_site *site = cl_object_site(lock->cll_descr.cld_obj);
1025 LASSERT(lock->cll_state <= state ||
1026 (lock->cll_state == CLS_CACHED &&
1027 (state == CLS_HELD || /* lock found in cache */
1028 state == CLS_NEW || /* sub-lock canceled */
1029 state == CLS_INTRANSIT)) ||
1030 /* lock is in transit state */
1031 lock->cll_state == CLS_INTRANSIT);
1033 if (lock->cll_state != state) {
1034 atomic_dec(&site->cs_locks_state[lock->cll_state]);
1035 atomic_inc(&site->cs_locks_state[state]);
1037 cl_lock_state_signal(env, lock, state);
1038 lock->cll_state = state;
1042 EXPORT_SYMBOL(cl_lock_state_set);
1044 static int cl_unuse_try_internal(const struct lu_env *env, struct cl_lock *lock)
1046 const struct cl_lock_slice *slice;
1052 LINVRNT(cl_lock_is_mutexed(lock));
1053 LINVRNT(cl_lock_invariant(env, lock));
1054 LASSERT(lock->cll_state == CLS_INTRANSIT);
1057 list_for_each_entry_reverse(slice, &lock->cll_layers,
1059 if (slice->cls_ops->clo_unuse != NULL) {
1060 result = slice->cls_ops->clo_unuse(env, slice);
1065 LASSERT(result != -ENOSYS);
1066 } while (result == CLO_REPEAT);
1072 * Yanks lock from the cache (cl_lock_state::CLS_CACHED state) by calling
1073 * cl_lock_operations::clo_use() top-to-bottom to notify layers.
1074 * @atomic = 1, it must unuse the lock to recovery the lock to keep the
1075 * use process atomic
1077 int cl_use_try(const struct lu_env *env, struct cl_lock *lock, int atomic)
1079 const struct cl_lock_slice *slice;
1081 enum cl_lock_state state;
1084 cl_lock_trace(D_DLMTRACE, env, "use lock", lock);
1086 LASSERT(lock->cll_state == CLS_CACHED);
1087 if (lock->cll_error)
1088 RETURN(lock->cll_error);
1091 state = cl_lock_intransit(env, lock);
1092 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1093 if (slice->cls_ops->clo_use != NULL) {
1094 result = slice->cls_ops->clo_use(env, slice);
1099 LASSERT(result != -ENOSYS);
1101 LASSERTF(lock->cll_state == CLS_INTRANSIT, "Wrong state %d.\n",
1107 if (result == -ESTALE) {
1109 * ESTALE means sublock being cancelled
1110 * at this time, and set lock state to
1111 * be NEW here and ask the caller to repeat.
1114 result = CLO_REPEAT;
1117 /* @atomic means back-off-on-failure. */
1120 rc = cl_unuse_try_internal(env, lock);
1121 /* Vet the results. */
1122 if (rc < 0 && result > 0)
1127 cl_lock_extransit(env, lock, state);
1130 EXPORT_SYMBOL(cl_use_try);
1133 * Helper for cl_enqueue_try() that calls ->clo_enqueue() across all layers
1136 static int cl_enqueue_kick(const struct lu_env *env,
1137 struct cl_lock *lock,
1138 struct cl_io *io, __u32 flags)
1141 const struct cl_lock_slice *slice;
1145 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1146 if (slice->cls_ops->clo_enqueue != NULL) {
1147 result = slice->cls_ops->clo_enqueue(env,
1153 LASSERT(result != -ENOSYS);
1158 * Tries to enqueue a lock.
1160 * This function is called repeatedly by cl_enqueue() until either lock is
1161 * enqueued, or error occurs. This function does not block waiting for
1162 * networking communication to complete.
1164 * \post ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1165 * lock->cll_state == CLS_HELD)
1167 * \see cl_enqueue() cl_lock_operations::clo_enqueue()
1168 * \see cl_lock_state::CLS_ENQUEUED
1170 int cl_enqueue_try(const struct lu_env *env, struct cl_lock *lock,
1171 struct cl_io *io, __u32 flags)
1176 cl_lock_trace(D_DLMTRACE, env, "enqueue lock", lock);
1180 LINVRNT(cl_lock_is_mutexed(lock));
1182 if (lock->cll_error != 0)
1184 switch (lock->cll_state) {
1186 cl_lock_state_set(env, lock, CLS_QUEUING);
1190 result = cl_enqueue_kick(env, lock, io, flags);
1192 cl_lock_state_set(env, lock, CLS_ENQUEUED);
1195 LASSERT(cl_lock_is_intransit(lock));
1199 /* yank lock from the cache. */
1200 result = cl_use_try(env, lock, 0);
1209 * impossible, only held locks with increased
1210 * ->cll_holds can be enqueued, and they cannot be
1215 } while (result == CLO_REPEAT);
1217 cl_lock_error(env, lock, result);
1218 RETURN(result ?: lock->cll_error);
1220 EXPORT_SYMBOL(cl_enqueue_try);
1222 static int cl_enqueue_locked(const struct lu_env *env, struct cl_lock *lock,
1223 struct cl_io *io, __u32 enqflags)
1229 LINVRNT(cl_lock_is_mutexed(lock));
1230 LINVRNT(cl_lock_invariant(env, lock));
1231 LASSERT(lock->cll_holds > 0);
1233 cl_lock_user_add(env, lock);
1235 result = cl_enqueue_try(env, lock, io, enqflags);
1236 if (result == CLO_WAIT) {
1237 result = cl_lock_state_wait(env, lock);
1244 cl_lock_user_del(env, lock);
1245 cl_lock_error(env, lock, result);
1247 LASSERT(ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1248 lock->cll_state == CLS_HELD));
1255 * \pre current thread or io owns a hold on lock.
1257 * \post ergo(result == 0, lock->users increased)
1258 * \post ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1259 * lock->cll_state == CLS_HELD)
1261 int cl_enqueue(const struct lu_env *env, struct cl_lock *lock,
1262 struct cl_io *io, __u32 enqflags)
1268 cl_lock_lockdep_acquire(env, lock, enqflags);
1269 cl_lock_mutex_get(env, lock);
1270 result = cl_enqueue_locked(env, lock, io, enqflags);
1271 cl_lock_mutex_put(env, lock);
1273 cl_lock_lockdep_release(env, lock);
1274 LASSERT(ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1275 lock->cll_state == CLS_HELD));
1278 EXPORT_SYMBOL(cl_enqueue);
1281 * Tries to unlock a lock.
1283 * This function is called repeatedly by cl_unuse() until either lock is
1284 * unlocked, or error occurs.
1285 * cl_unuse_try is a one-shot operation, so it must NOT return CLO_WAIT.
1287 * \pre lock->cll_state == CLS_HELD
1289 * \post ergo(result == 0, lock->cll_state == CLS_CACHED)
1291 * \see cl_unuse() cl_lock_operations::clo_unuse()
1292 * \see cl_lock_state::CLS_CACHED
1294 int cl_unuse_try(const struct lu_env *env, struct cl_lock *lock)
1297 enum cl_lock_state state = CLS_NEW;
1300 cl_lock_trace(D_DLMTRACE, env, "unuse lock", lock);
1302 LASSERT(lock->cll_state == CLS_HELD || lock->cll_state == CLS_ENQUEUED);
1303 if (lock->cll_users > 1) {
1304 cl_lock_user_del(env, lock);
1309 * New lock users (->cll_users) are not protecting unlocking
1310 * from proceeding. From this point, lock eventually reaches
1311 * CLS_CACHED, is reinitialized to CLS_NEW or fails into
1314 state = cl_lock_intransit(env, lock);
1316 result = cl_unuse_try_internal(env, lock);
1317 LASSERT(lock->cll_state == CLS_INTRANSIT);
1318 LASSERT(result != CLO_WAIT);
1319 cl_lock_user_del(env, lock);
1320 if (result == 0 || result == -ESTALE) {
1322 * Return lock back to the cache. This is the only
1323 * place where lock is moved into CLS_CACHED state.
1325 * If one of ->clo_unuse() methods returned -ESTALE, lock
1326 * cannot be placed into cache and has to be
1327 * re-initialized. This happens e.g., when a sub-lock was
1328 * canceled while unlocking was in progress.
1330 if (state == CLS_HELD && result == 0)
1334 cl_lock_extransit(env, lock, state);
1337 * Hide -ESTALE error.
1338 * If the lock is a glimpse lock, and it has multiple
1339 * stripes. Assuming that one of its sublock returned -ENAVAIL,
1340 * and other sublocks are matched write locks. In this case,
1341 * we can't set this lock to error because otherwise some of
1342 * its sublocks may not be canceled. This causes some dirty
1343 * pages won't be written to OSTs. -jay
1347 CERROR("result = %d, this is unlikely!\n", result);
1348 cl_lock_extransit(env, lock, state);
1351 result = result ?: lock->cll_error;
1353 cl_lock_error(env, lock, result);
1356 EXPORT_SYMBOL(cl_unuse_try);
1358 static void cl_unuse_locked(const struct lu_env *env, struct cl_lock *lock)
1363 result = cl_unuse_try(env, lock);
1365 CL_LOCK_DEBUG(D_ERROR, env, lock, "unuse return %d\n", result);
1373 void cl_unuse(const struct lu_env *env, struct cl_lock *lock)
1376 cl_lock_mutex_get(env, lock);
1377 cl_unuse_locked(env, lock);
1378 cl_lock_mutex_put(env, lock);
1379 cl_lock_lockdep_release(env, lock);
1382 EXPORT_SYMBOL(cl_unuse);
1385 * Tries to wait for a lock.
1387 * This function is called repeatedly by cl_wait() until either lock is
1388 * granted, or error occurs. This function does not block waiting for network
1389 * communication to complete.
1391 * \see cl_wait() cl_lock_operations::clo_wait()
1392 * \see cl_lock_state::CLS_HELD
1394 int cl_wait_try(const struct lu_env *env, struct cl_lock *lock)
1396 const struct cl_lock_slice *slice;
1400 cl_lock_trace(D_DLMTRACE, env, "wait lock try", lock);
1402 LINVRNT(cl_lock_is_mutexed(lock));
1403 LINVRNT(cl_lock_invariant(env, lock));
1404 LASSERT(lock->cll_state == CLS_ENQUEUED ||
1405 lock->cll_state == CLS_HELD ||
1406 lock->cll_state == CLS_INTRANSIT);
1407 LASSERT(lock->cll_users > 0);
1408 LASSERT(lock->cll_holds > 0);
1411 if (lock->cll_error != 0)
1414 if (cl_lock_is_intransit(lock)) {
1419 if (lock->cll_state == CLS_HELD)
1424 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1425 if (slice->cls_ops->clo_wait != NULL) {
1426 result = slice->cls_ops->clo_wait(env, slice);
1431 LASSERT(result != -ENOSYS);
1433 LASSERT(lock->cll_state != CLS_INTRANSIT);
1434 cl_lock_state_set(env, lock, CLS_HELD);
1436 } while (result == CLO_REPEAT);
1437 RETURN(result ?: lock->cll_error);
1439 EXPORT_SYMBOL(cl_wait_try);
1442 * Waits until enqueued lock is granted.
1444 * \pre current thread or io owns a hold on the lock
1445 * \pre ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1446 * lock->cll_state == CLS_HELD)
1448 * \post ergo(result == 0, lock->cll_state == CLS_HELD)
1450 int cl_wait(const struct lu_env *env, struct cl_lock *lock)
1455 cl_lock_mutex_get(env, lock);
1457 LINVRNT(cl_lock_invariant(env, lock));
1458 LASSERTF(lock->cll_state == CLS_ENQUEUED || lock->cll_state == CLS_HELD,
1459 "Wrong state %d \n", lock->cll_state);
1460 LASSERT(lock->cll_holds > 0);
1461 cl_lock_trace(D_DLMTRACE, env, "wait lock", lock);
1464 result = cl_wait_try(env, lock);
1465 if (result == CLO_WAIT) {
1466 result = cl_lock_state_wait(env, lock);
1473 cl_lock_user_del(env, lock);
1474 cl_lock_error(env, lock, result);
1475 cl_lock_lockdep_release(env, lock);
1477 cl_lock_mutex_put(env, lock);
1478 LASSERT(ergo(result == 0, lock->cll_state == CLS_HELD));
1481 EXPORT_SYMBOL(cl_wait);
1484 * Executes cl_lock_operations::clo_weigh(), and sums results to estimate lock
1487 unsigned long cl_lock_weigh(const struct lu_env *env, struct cl_lock *lock)
1489 const struct cl_lock_slice *slice;
1490 unsigned long pound;
1491 unsigned long ounce;
1494 LINVRNT(cl_lock_is_mutexed(lock));
1495 LINVRNT(cl_lock_invariant(env, lock));
1498 list_for_each_entry_reverse(slice, &lock->cll_layers, cls_linkage) {
1499 if (slice->cls_ops->clo_weigh != NULL) {
1500 ounce = slice->cls_ops->clo_weigh(env, slice);
1502 if (pound < ounce) /* over-weight^Wflow */
1508 EXPORT_SYMBOL(cl_lock_weigh);
1511 * Notifies layers that lock description changed.
1513 * The server can grant client a lock different from one that was requested
1514 * (e.g., larger in extent). This method is called when actually granted lock
1515 * description becomes known to let layers to accommodate for changed lock
1518 * \see cl_lock_operations::clo_modify()
1520 int cl_lock_modify(const struct lu_env *env, struct cl_lock *lock,
1521 const struct cl_lock_descr *desc)
1523 const struct cl_lock_slice *slice;
1524 struct cl_object *obj = lock->cll_descr.cld_obj;
1525 struct cl_object_header *hdr = cl_object_header(obj);
1529 cl_lock_trace(D_DLMTRACE, env, "modify lock", lock);
1530 /* don't allow object to change */
1531 LASSERT(obj == desc->cld_obj);
1532 LINVRNT(cl_lock_is_mutexed(lock));
1533 LINVRNT(cl_lock_invariant(env, lock));
1535 list_for_each_entry_reverse(slice, &lock->cll_layers, cls_linkage) {
1536 if (slice->cls_ops->clo_modify != NULL) {
1537 result = slice->cls_ops->clo_modify(env, slice, desc);
1542 CL_LOCK_DEBUG(D_DLMTRACE, env, lock, " -> "DDESCR"@"DFID"\n",
1543 PDESCR(desc), PFID(lu_object_fid(&desc->cld_obj->co_lu)));
1545 * Just replace description in place. Nothing more is needed for
1546 * now. If locks were indexed according to their extent and/or mode,
1547 * that index would have to be updated here.
1549 spin_lock(&hdr->coh_lock_guard);
1550 lock->cll_descr = *desc;
1551 spin_unlock(&hdr->coh_lock_guard);
1554 EXPORT_SYMBOL(cl_lock_modify);
1557 * Initializes lock closure with a given origin.
1559 * \see cl_lock_closure
1561 void cl_lock_closure_init(const struct lu_env *env,
1562 struct cl_lock_closure *closure,
1563 struct cl_lock *origin, int wait)
1565 LINVRNT(cl_lock_is_mutexed(origin));
1566 LINVRNT(cl_lock_invariant(env, origin));
1568 CFS_INIT_LIST_HEAD(&closure->clc_list);
1569 closure->clc_origin = origin;
1570 closure->clc_wait = wait;
1571 closure->clc_nr = 0;
1573 EXPORT_SYMBOL(cl_lock_closure_init);
1576 * Builds a closure of \a lock.
1578 * Building of a closure consists of adding initial lock (\a lock) into it,
1579 * and calling cl_lock_operations::clo_closure() methods of \a lock. These
1580 * methods might call cl_lock_closure_build() recursively again, adding more
1581 * locks to the closure, etc.
1583 * \see cl_lock_closure
1585 int cl_lock_closure_build(const struct lu_env *env, struct cl_lock *lock,
1586 struct cl_lock_closure *closure)
1588 const struct cl_lock_slice *slice;
1592 LINVRNT(cl_lock_is_mutexed(closure->clc_origin));
1593 LINVRNT(cl_lock_invariant(env, closure->clc_origin));
1595 result = cl_lock_enclosure(env, lock, closure);
1597 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1598 if (slice->cls_ops->clo_closure != NULL) {
1599 result = slice->cls_ops->clo_closure(env, slice,
1607 cl_lock_disclosure(env, closure);
1610 EXPORT_SYMBOL(cl_lock_closure_build);
1613 * Adds new lock to a closure.
1615 * Try-locks \a lock and if succeeded, adds it to the closure (never more than
1616 * once). If try-lock failed, returns CLO_REPEAT, after optionally waiting
1617 * until next try-lock is likely to succeed.
1619 int cl_lock_enclosure(const struct lu_env *env, struct cl_lock *lock,
1620 struct cl_lock_closure *closure)
1624 cl_lock_trace(D_DLMTRACE, env, "enclosure lock", lock);
1625 if (!cl_lock_mutex_try(env, lock)) {
1627 * If lock->cll_inclosure is not empty, lock is already in
1630 if (list_empty(&lock->cll_inclosure)) {
1631 cl_lock_get_trust(lock);
1632 lu_ref_add(&lock->cll_reference, "closure", closure);
1633 list_add(&lock->cll_inclosure, &closure->clc_list);
1636 cl_lock_mutex_put(env, lock);
1639 cl_lock_disclosure(env, closure);
1640 if (closure->clc_wait) {
1641 cl_lock_get_trust(lock);
1642 lu_ref_add(&lock->cll_reference, "closure-w", closure);
1643 cl_lock_mutex_put(env, closure->clc_origin);
1645 LASSERT(cl_lock_nr_mutexed(env) == 0);
1646 cl_lock_mutex_get(env, lock);
1647 cl_lock_mutex_put(env, lock);
1649 cl_lock_mutex_get(env, closure->clc_origin);
1650 lu_ref_del(&lock->cll_reference, "closure-w", closure);
1651 cl_lock_put(env, lock);
1653 result = CLO_REPEAT;
1657 EXPORT_SYMBOL(cl_lock_enclosure);
1659 /** Releases mutices of enclosed locks. */
1660 void cl_lock_disclosure(const struct lu_env *env,
1661 struct cl_lock_closure *closure)
1663 struct cl_lock *scan;
1664 struct cl_lock *temp;
1666 cl_lock_trace(D_DLMTRACE, env, "disclosure lock", closure->clc_origin);
1667 list_for_each_entry_safe(scan, temp, &closure->clc_list, cll_inclosure){
1668 list_del_init(&scan->cll_inclosure);
1669 cl_lock_mutex_put(env, scan);
1670 lu_ref_del(&scan->cll_reference, "closure", closure);
1671 cl_lock_put(env, scan);
1674 LASSERT(closure->clc_nr == 0);
1676 EXPORT_SYMBOL(cl_lock_disclosure);
1678 /** Finalizes a closure. */
1679 void cl_lock_closure_fini(struct cl_lock_closure *closure)
1681 LASSERT(closure->clc_nr == 0);
1682 LASSERT(list_empty(&closure->clc_list));
1684 EXPORT_SYMBOL(cl_lock_closure_fini);
1687 * Destroys this lock. Notifies layers (bottom-to-top) that lock is being
1688 * destroyed, then destroy the lock. If there are holds on the lock, postpone
1689 * destruction until all holds are released. This is called when a decision is
1690 * made to destroy the lock in the future. E.g., when a blocking AST is
1691 * received on it, or fatal communication error happens.
1693 * Caller must have a reference on this lock to prevent a situation, when
1694 * deleted lock lingers in memory for indefinite time, because nobody calls
1695 * cl_lock_put() to finish it.
1697 * \pre atomic_read(&lock->cll_ref) > 0
1698 * \pre ergo(cl_lock_nesting(lock) == CNL_TOP,
1699 * cl_lock_nr_mutexed(env) == 1)
1700 * [i.e., if a top-lock is deleted, mutices of no other locks can be
1701 * held, as deletion of sub-locks might require releasing a top-lock
1704 * \see cl_lock_operations::clo_delete()
1705 * \see cl_lock::cll_holds
1707 void cl_lock_delete(const struct lu_env *env, struct cl_lock *lock)
1709 LINVRNT(cl_lock_is_mutexed(lock));
1710 LINVRNT(cl_lock_invariant(env, lock));
1711 LASSERT(ergo(cl_lock_nesting(lock) == CNL_TOP,
1712 cl_lock_nr_mutexed(env) == 1));
1715 cl_lock_trace(D_DLMTRACE, env, "delete lock", lock);
1716 if (lock->cll_holds == 0)
1717 cl_lock_delete0(env, lock);
1719 lock->cll_flags |= CLF_DOOMED;
1722 EXPORT_SYMBOL(cl_lock_delete);
1725 * Mark lock as irrecoverably failed, and mark it for destruction. This
1726 * happens when, e.g., server fails to grant a lock to us, or networking
1729 * \pre atomic_read(&lock->cll_ref) > 0
1731 * \see clo_lock_delete()
1732 * \see cl_lock::cll_holds
1734 void cl_lock_error(const struct lu_env *env, struct cl_lock *lock, int error)
1736 LINVRNT(cl_lock_is_mutexed(lock));
1737 LINVRNT(cl_lock_invariant(env, lock));
1740 cl_lock_trace(D_DLMTRACE, env, "set lock error", lock);
1741 if (lock->cll_error == 0 && error != 0) {
1742 lock->cll_error = error;
1743 cl_lock_signal(env, lock);
1744 cl_lock_cancel(env, lock);
1745 cl_lock_delete(env, lock);
1749 EXPORT_SYMBOL(cl_lock_error);
1752 * Cancels this lock. Notifies layers
1753 * (bottom-to-top) that lock is being cancelled, then destroy the lock. If
1754 * there are holds on the lock, postpone cancellation until
1755 * all holds are released.
1757 * Cancellation notification is delivered to layers at most once.
1759 * \see cl_lock_operations::clo_cancel()
1760 * \see cl_lock::cll_holds
1762 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock)
1764 LINVRNT(cl_lock_is_mutexed(lock));
1765 LINVRNT(cl_lock_invariant(env, lock));
1768 cl_lock_trace(D_DLMTRACE, env, "cancel lock", lock);
1769 if (lock->cll_holds == 0)
1770 cl_lock_cancel0(env, lock);
1772 lock->cll_flags |= CLF_CANCELPEND;
1775 EXPORT_SYMBOL(cl_lock_cancel);
1778 * Finds an existing lock covering given page and optionally different from a
1779 * given \a except lock.
1781 struct cl_lock *cl_lock_at_page(const struct lu_env *env, struct cl_object *obj,
1782 struct cl_page *page, struct cl_lock *except,
1783 int pending, int canceld)
1785 struct cl_object_header *head;
1786 struct cl_lock *scan;
1787 struct cl_lock *lock;
1788 struct cl_lock_descr *need;
1792 head = cl_object_header(obj);
1793 need = &cl_env_info(env)->clt_descr;
1796 need->cld_mode = CLM_READ; /* CLM_READ matches both READ & WRITE, but
1798 need->cld_start = need->cld_end = page->cp_index;
1799 need->cld_enq_flags = 0;
1801 spin_lock(&head->coh_lock_guard);
1802 /* It is fine to match any group lock since there could be only one
1803 * with a uniq gid and it conflicts with all other lock modes too */
1804 list_for_each_entry(scan, &head->coh_locks, cll_linkage) {
1805 if (scan != except &&
1806 (scan->cll_descr.cld_mode == CLM_GROUP ||
1807 cl_lock_ext_match(&scan->cll_descr, need)) &&
1808 scan->cll_state >= CLS_HELD &&
1809 scan->cll_state < CLS_FREEING &&
1811 * This check is racy as the lock can be canceled right
1812 * after it is done, but this is fine, because page exists
1815 (canceld || !(scan->cll_flags & CLF_CANCELLED)) &&
1816 (pending || !(scan->cll_flags & CLF_CANCELPEND))) {
1817 /* Don't increase cs_hit here since this
1818 * is just a helper function. */
1819 cl_lock_get_trust(scan);
1824 spin_unlock(&head->coh_lock_guard);
1827 EXPORT_SYMBOL(cl_lock_at_page);
1830 * Returns a list of pages protected (only) by a given lock.
1832 * Scans an extent of page radix tree, corresponding to the \a lock and queues
1833 * all pages that are not protected by locks other than \a lock into \a queue.
1835 void cl_lock_page_list_fixup(const struct lu_env *env,
1836 struct cl_io *io, struct cl_lock *lock,
1837 struct cl_page_list *queue)
1839 struct cl_page *page;
1840 struct cl_page *temp;
1841 struct cl_page_list *plist = &cl_env_info(env)->clt_list;
1843 LINVRNT(cl_lock_invariant(env, lock));
1846 /* Now, we have a list of cl_pages under the \a lock, we need
1847 * to check if some of pages are covered by other ldlm lock.
1848 * If this is the case, they aren't needed to be written out this time.
1850 * For example, we have A:[0,200] & B:[100,300] PW locks on client, now
1851 * the latter is to be canceled, this means other client is
1852 * reading/writing [200,300] since A won't canceled. Actually
1853 * we just need to write the pages covered by [200,300]. This is safe,
1854 * since [100,200] is also protected lock A.
1857 cl_page_list_init(plist);
1858 cl_page_list_for_each_safe(page, temp, queue) {
1859 pgoff_t idx = page->cp_index;
1860 struct cl_lock *found;
1861 struct cl_lock_descr *descr;
1863 /* The algorithm counts on the index-ascending page index. */
1864 LASSERT(ergo(&temp->cp_batch != &queue->pl_pages,
1865 page->cp_index < temp->cp_index));
1867 found = cl_lock_at_page(env, lock->cll_descr.cld_obj,
1872 descr = &found->cll_descr;
1873 list_for_each_entry_safe_from(page, temp, &queue->pl_pages,
1875 idx = page->cp_index;
1876 if (descr->cld_start > idx || descr->cld_end < idx)
1878 cl_page_list_move(plist, queue, page);
1880 cl_lock_put(env, found);
1883 /* The pages in plist are covered by other locks, don't handle them
1887 cl_page_list_disown(env, io, plist);
1888 cl_page_list_fini(env, plist);
1891 EXPORT_SYMBOL(cl_lock_page_list_fixup);
1894 * Invalidate pages protected by the given lock, sending them out to the
1895 * server first, if necessary.
1897 * This function does the following:
1899 * - collects a list of pages to be invalidated,
1901 * - unmaps them from the user virtual memory,
1903 * - sends dirty pages to the server,
1905 * - waits for transfer completion,
1907 * - discards pages, and throws them out of memory.
1909 * If \a discard is set, pages are discarded without sending them to the
1912 * If error happens on any step, the process continues anyway (the reasoning
1913 * behind this being that lock cancellation cannot be delayed indefinitely).
1915 int cl_lock_page_out(const struct lu_env *env, struct cl_lock *lock,
1918 struct cl_thread_info *info = cl_env_info(env);
1919 struct cl_io *io = &info->clt_io;
1920 struct cl_2queue *queue = &info->clt_queue;
1921 struct cl_lock_descr *descr = &lock->cll_descr;
1925 LINVRNT(cl_lock_invariant(env, lock));
1928 io->ci_obj = cl_object_top(descr->cld_obj);
1929 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1934 cl_2queue_init(queue);
1935 cl_page_gang_lookup(env, descr->cld_obj, io, descr->cld_start,
1936 descr->cld_end, &queue->c2_qin, nonblock);
1937 page_count = queue->c2_qin.pl_nr;
1938 if (page_count > 0) {
1939 result = cl_page_list_unmap(env, io, &queue->c2_qin);
1941 long timeout = 600; /* 10 minutes. */
1942 /* for debug purpose, if this request can't be
1943 * finished in 10 minutes, we hope it can
1946 result = cl_io_submit_sync(env, io, CRT_WRITE,
1950 CWARN("Writing %lu pages error: %d\n",
1951 page_count, result);
1953 cl_lock_page_list_fixup(env, io, lock, &queue->c2_qout);
1954 cl_2queue_discard(env, io, queue);
1955 cl_2queue_disown(env, io, queue);
1957 cl_2queue_fini(env, queue);
1964 cl_io_fini(env, io);
1967 EXPORT_SYMBOL(cl_lock_page_out);
1970 * Eliminate all locks for a given object.
1972 * Caller has to guarantee that no lock is in active use.
1974 * \param cancel when this is set, cl_locks_prune() cancels locks before
1977 void cl_locks_prune(const struct lu_env *env, struct cl_object *obj, int cancel)
1979 struct cl_object_header *head;
1980 struct cl_lock *lock;
1983 head = cl_object_header(obj);
1985 * If locks are destroyed without cancellation, all pages must be
1986 * already destroyed (as otherwise they will be left unprotected).
1988 LASSERT(ergo(!cancel,
1989 head->coh_tree.rnode == NULL && head->coh_pages == 0));
1991 spin_lock(&head->coh_lock_guard);
1992 while (!list_empty(&head->coh_locks)) {
1993 lock = container_of(head->coh_locks.next,
1994 struct cl_lock, cll_linkage);
1995 cl_lock_get_trust(lock);
1996 spin_unlock(&head->coh_lock_guard);
1997 lu_ref_add(&lock->cll_reference, "prune", cfs_current());
1998 cl_lock_mutex_get(env, lock);
1999 if (lock->cll_state < CLS_FREEING) {
2000 LASSERT(lock->cll_holds == 0);
2001 LASSERT(lock->cll_users == 0);
2003 cl_lock_cancel(env, lock);
2004 cl_lock_delete(env, lock);
2006 cl_lock_mutex_put(env, lock);
2007 lu_ref_del(&lock->cll_reference, "prune", cfs_current());
2008 cl_lock_put(env, lock);
2009 spin_lock(&head->coh_lock_guard);
2011 spin_unlock(&head->coh_lock_guard);
2014 EXPORT_SYMBOL(cl_locks_prune);
2017 * Returns true if \a addr is an address of an allocated cl_lock. Used in
2018 * assertions. This check is optimistically imprecise, i.e., it occasionally
2019 * returns true for the incorrect addresses, but if it returns false, then the
2020 * address is guaranteed to be incorrect. (Should be named cl_lockp().)
2024 int cl_is_lock(const void *addr)
2026 return cfs_mem_is_in_cache(addr, cl_lock_kmem);
2028 EXPORT_SYMBOL(cl_is_lock);
2030 static struct cl_lock *cl_lock_hold_mutex(const struct lu_env *env,
2031 const struct cl_io *io,
2032 const struct cl_lock_descr *need,
2033 const char *scope, const void *source)
2035 struct cl_lock *lock;
2040 lock = cl_lock_find(env, io, need);
2043 cl_lock_mutex_get(env, lock);
2044 if (lock->cll_state < CLS_FREEING &&
2045 !(lock->cll_flags & CLF_CANCELLED)) {
2046 cl_lock_hold_mod(env, lock, +1);
2047 lu_ref_add(&lock->cll_holders, scope, source);
2048 lu_ref_add(&lock->cll_reference, scope, source);
2051 cl_lock_mutex_put(env, lock);
2052 cl_lock_put(env, lock);
2058 * Returns a lock matching \a need description with a reference and a hold on
2061 * This is much like cl_lock_find(), except that cl_lock_hold() additionally
2062 * guarantees that lock is not in the CLS_FREEING state on return.
2064 struct cl_lock *cl_lock_hold(const struct lu_env *env, const struct cl_io *io,
2065 const struct cl_lock_descr *need,
2066 const char *scope, const void *source)
2068 struct cl_lock *lock;
2072 lock = cl_lock_hold_mutex(env, io, need, scope, source);
2074 cl_lock_mutex_put(env, lock);
2077 EXPORT_SYMBOL(cl_lock_hold);
2080 * Main high-level entry point of cl_lock interface that finds existing or
2081 * enqueues new lock matching given description.
2083 struct cl_lock *cl_lock_request(const struct lu_env *env, struct cl_io *io,
2084 const struct cl_lock_descr *need,
2085 const char *scope, const void *source)
2087 struct cl_lock *lock;
2088 const struct lu_fid *fid;
2091 __u32 enqflags = need->cld_enq_flags;
2094 fid = lu_object_fid(&io->ci_obj->co_lu);
2097 lock = cl_lock_hold_mutex(env, io, need, scope, source);
2098 if (!IS_ERR(lock)) {
2099 rc = cl_enqueue_locked(env, lock, io, enqflags);
2101 if (cl_lock_fits_into(env, lock, need, io)) {
2102 cl_lock_mutex_put(env, lock);
2103 cl_lock_lockdep_acquire(env,
2107 cl_unuse_locked(env, lock);
2109 cl_lock_trace(D_DLMTRACE, env, "enqueue failed", lock);
2110 cl_lock_hold_release(env, lock, scope, source);
2111 cl_lock_mutex_put(env, lock);
2112 lu_ref_del(&lock->cll_reference, scope, source);
2113 cl_lock_put(env, lock);
2121 EXPORT_SYMBOL(cl_lock_request);
2124 * Adds a hold to a known lock.
2126 void cl_lock_hold_add(const struct lu_env *env, struct cl_lock *lock,
2127 const char *scope, const void *source)
2129 LINVRNT(cl_lock_is_mutexed(lock));
2130 LINVRNT(cl_lock_invariant(env, lock));
2131 LASSERT(lock->cll_state != CLS_FREEING);
2134 cl_lock_hold_mod(env, lock, +1);
2136 lu_ref_add(&lock->cll_holders, scope, source);
2137 lu_ref_add(&lock->cll_reference, scope, source);
2140 EXPORT_SYMBOL(cl_lock_hold_add);
2143 * Releases a hold and a reference on a lock, on which caller acquired a
2146 void cl_lock_unhold(const struct lu_env *env, struct cl_lock *lock,
2147 const char *scope, const void *source)
2149 LINVRNT(cl_lock_invariant(env, lock));
2151 cl_lock_hold_release(env, lock, scope, source);
2152 lu_ref_del(&lock->cll_reference, scope, source);
2153 cl_lock_put(env, lock);
2156 EXPORT_SYMBOL(cl_lock_unhold);
2159 * Releases a hold and a reference on a lock, obtained by cl_lock_hold().
2161 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock,
2162 const char *scope, const void *source)
2164 LINVRNT(cl_lock_invariant(env, lock));
2166 cl_lock_trace(D_DLMTRACE, env, "release lock", lock);
2167 cl_lock_mutex_get(env, lock);
2168 cl_lock_hold_release(env, lock, scope, source);
2169 cl_lock_mutex_put(env, lock);
2170 lu_ref_del(&lock->cll_reference, scope, source);
2171 cl_lock_put(env, lock);
2174 EXPORT_SYMBOL(cl_lock_release);
2176 void cl_lock_user_add(const struct lu_env *env, struct cl_lock *lock)
2178 LINVRNT(cl_lock_is_mutexed(lock));
2179 LINVRNT(cl_lock_invariant(env, lock));
2182 cl_lock_used_mod(env, lock, +1);
2185 EXPORT_SYMBOL(cl_lock_user_add);
2187 int cl_lock_user_del(const struct lu_env *env, struct cl_lock *lock)
2189 LINVRNT(cl_lock_is_mutexed(lock));
2190 LINVRNT(cl_lock_invariant(env, lock));
2191 LASSERT(lock->cll_users > 0);
2194 cl_lock_used_mod(env, lock, -1);
2195 RETURN(lock->cll_users == 0);
2197 EXPORT_SYMBOL(cl_lock_user_del);
2199 const char *cl_lock_mode_name(const enum cl_lock_mode mode)
2201 static const char *names[] = {
2202 [CLM_PHANTOM] = "P",
2207 if (0 <= mode && mode < ARRAY_SIZE(names))
2212 EXPORT_SYMBOL(cl_lock_mode_name);
2215 * Prints human readable representation of a lock description.
2217 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2218 lu_printer_t printer,
2219 const struct cl_lock_descr *descr)
2221 const struct lu_fid *fid;
2223 fid = lu_object_fid(&descr->cld_obj->co_lu);
2224 (*printer)(env, cookie, DDESCR"@"DFID, PDESCR(descr), PFID(fid));
2226 EXPORT_SYMBOL(cl_lock_descr_print);
2229 * Prints human readable representation of \a lock to the \a f.
2231 void cl_lock_print(const struct lu_env *env, void *cookie,
2232 lu_printer_t printer, const struct cl_lock *lock)
2234 const struct cl_lock_slice *slice;
2235 (*printer)(env, cookie, "lock@%p[%d %d %d %d %d %08lx] ",
2236 lock, atomic_read(&lock->cll_ref),
2237 lock->cll_state, lock->cll_error, lock->cll_holds,
2238 lock->cll_users, lock->cll_flags);
2239 cl_lock_descr_print(env, cookie, printer, &lock->cll_descr);
2240 (*printer)(env, cookie, " {\n");
2242 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
2243 (*printer)(env, cookie, " %s@%p: ",
2244 slice->cls_obj->co_lu.lo_dev->ld_type->ldt_name,
2246 if (slice->cls_ops->clo_print != NULL)
2247 slice->cls_ops->clo_print(env, cookie, printer, slice);
2248 (*printer)(env, cookie, "\n");
2250 (*printer)(env, cookie, "} lock@%p\n", lock);
2252 EXPORT_SYMBOL(cl_lock_print);
2254 int cl_lock_init(void)
2256 return lu_kmem_init(cl_lock_caches);
2259 void cl_lock_fini(void)
2261 lu_kmem_fini(cl_lock_caches);