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 #define RETIP ((unsigned long)__builtin_return_address(0))
132 #ifdef CONFIG_LOCKDEP
133 static struct lock_class_key cl_lock_key;
135 static void cl_lock_lockdep_init(struct cl_lock *lock)
137 lockdep_set_class_and_name(lock, &cl_lock_key, "EXT");
140 static void cl_lock_lockdep_acquire(const struct lu_env *env,
141 struct cl_lock *lock, __u32 enqflags)
143 cl_lock_counters(env, lock)->ctc_nr_locks_acquired++;
144 lock_acquire(&lock->dep_map, !!(enqflags & CEF_ASYNC),
145 /* try: */ 0, lock->cll_descr.cld_mode <= CLM_READ,
146 /* check: */ 2, RETIP);
149 static void cl_lock_lockdep_release(const struct lu_env *env,
150 struct cl_lock *lock)
152 cl_lock_counters(env, lock)->ctc_nr_locks_acquired--;
153 lock_release(&lock->dep_map, 0, RETIP);
156 #else /* !CONFIG_LOCKDEP */
158 static void cl_lock_lockdep_init(struct cl_lock *lock)
160 static void cl_lock_lockdep_acquire(const struct lu_env *env,
161 struct cl_lock *lock, __u32 enqflags)
163 static void cl_lock_lockdep_release(const struct lu_env *env,
164 struct cl_lock *lock)
167 #endif /* !CONFIG_LOCKDEP */
170 * Adds lock slice to the compound lock.
172 * This is called by cl_object_operations::coo_lock_init() methods to add a
173 * per-layer state to the lock. New state is added at the end of
174 * cl_lock::cll_layers list, that is, it is at the bottom of the stack.
176 * \see cl_req_slice_add(), cl_page_slice_add(), cl_io_slice_add()
178 void cl_lock_slice_add(struct cl_lock *lock, struct cl_lock_slice *slice,
179 struct cl_object *obj,
180 const struct cl_lock_operations *ops)
183 slice->cls_lock = lock;
184 list_add_tail(&slice->cls_linkage, &lock->cll_layers);
185 slice->cls_obj = obj;
186 slice->cls_ops = ops;
189 EXPORT_SYMBOL(cl_lock_slice_add);
192 * Returns true iff a lock with the mode \a has provides at least the same
193 * guarantees as a lock with the mode \a need.
195 int cl_lock_mode_match(enum cl_lock_mode has, enum cl_lock_mode need)
197 LINVRNT(need == CLM_READ || need == CLM_WRITE ||
198 need == CLM_PHANTOM || need == CLM_GROUP);
199 LINVRNT(has == CLM_READ || has == CLM_WRITE ||
200 has == CLM_PHANTOM || has == CLM_GROUP);
201 CLASSERT(CLM_PHANTOM < CLM_READ);
202 CLASSERT(CLM_READ < CLM_WRITE);
203 CLASSERT(CLM_WRITE < CLM_GROUP);
205 if (has != CLM_GROUP)
210 EXPORT_SYMBOL(cl_lock_mode_match);
213 * Returns true iff extent portions of lock descriptions match.
215 int cl_lock_ext_match(const struct cl_lock_descr *has,
216 const struct cl_lock_descr *need)
219 has->cld_start <= need->cld_start &&
220 has->cld_end >= need->cld_end &&
221 cl_lock_mode_match(has->cld_mode, need->cld_mode) &&
222 (has->cld_mode != CLM_GROUP || has->cld_gid == need->cld_gid);
224 EXPORT_SYMBOL(cl_lock_ext_match);
227 * Returns true iff a lock with the description \a has provides at least the
228 * same guarantees as a lock with the description \a need.
230 int cl_lock_descr_match(const struct cl_lock_descr *has,
231 const struct cl_lock_descr *need)
234 cl_object_same(has->cld_obj, need->cld_obj) &&
235 cl_lock_ext_match(has, need);
237 EXPORT_SYMBOL(cl_lock_descr_match);
239 static void cl_lock_free(const struct lu_env *env, struct cl_lock *lock)
241 struct cl_object *obj = lock->cll_descr.cld_obj;
243 LASSERT(cl_is_lock(lock));
244 LINVRNT(!cl_lock_is_mutexed(lock));
248 while (!list_empty(&lock->cll_layers)) {
249 struct cl_lock_slice *slice;
251 slice = list_entry(lock->cll_layers.next, struct cl_lock_slice,
253 list_del_init(lock->cll_layers.next);
254 slice->cls_ops->clo_fini(env, slice);
256 atomic_dec(&cl_object_site(obj)->cs_locks.cs_total);
257 atomic_dec(&cl_object_site(obj)->cs_locks_state[lock->cll_state]);
258 lu_object_ref_del_at(&obj->co_lu, lock->cll_obj_ref, "cl_lock", lock);
259 cl_object_put(env, obj);
260 lu_ref_fini(&lock->cll_reference);
261 lu_ref_fini(&lock->cll_holders);
262 mutex_destroy(&lock->cll_guard);
263 OBD_SLAB_FREE_PTR(lock, cl_lock_kmem);
268 * Releases a reference on a lock.
270 * When last reference is released, lock is returned to the cache, unless it
271 * is in cl_lock_state::CLS_FREEING state, in which case it is destroyed
274 * \see cl_object_put(), cl_page_put()
276 void cl_lock_put(const struct lu_env *env, struct cl_lock *lock)
278 struct cl_object *obj;
279 struct cl_object_header *head;
280 struct cl_site *site;
282 LINVRNT(cl_lock_invariant(env, lock));
284 obj = lock->cll_descr.cld_obj;
285 LINVRNT(obj != NULL);
286 head = cl_object_header(obj);
287 site = cl_object_site(obj);
289 CDEBUG(D_DLMTRACE, "releasing reference: %d %p %lu\n",
290 atomic_read(&lock->cll_ref), lock, RETIP);
292 if (atomic_dec_and_test(&lock->cll_ref)) {
293 if (lock->cll_state == CLS_FREEING) {
294 LASSERT(list_empty(&lock->cll_linkage));
295 cl_lock_free(env, lock);
297 atomic_dec(&site->cs_locks.cs_busy);
301 EXPORT_SYMBOL(cl_lock_put);
304 * Acquires an additional reference to a lock.
306 * This can be called only by caller already possessing a reference to \a
309 * \see cl_object_get(), cl_page_get()
311 void cl_lock_get(struct cl_lock *lock)
313 LINVRNT(cl_lock_invariant(NULL, lock));
314 CDEBUG(D_DLMTRACE|D_TRACE, "acquiring reference: %d %p %lu\n",
315 atomic_read(&lock->cll_ref), lock, RETIP);
316 atomic_inc(&lock->cll_ref);
318 EXPORT_SYMBOL(cl_lock_get);
321 * Acquires a reference to a lock.
323 * This is much like cl_lock_get(), except that this function can be used to
324 * acquire initial reference to the cached lock. Caller has to deal with all
325 * possible races. Use with care!
327 * \see cl_page_get_trust()
329 void cl_lock_get_trust(struct cl_lock *lock)
331 struct cl_site *site = cl_object_site(lock->cll_descr.cld_obj);
333 LASSERT(cl_is_lock(lock));
334 CDEBUG(D_DLMTRACE|D_TRACE, "acquiring trusted reference: %d %p %lu\n",
335 atomic_read(&lock->cll_ref), lock, RETIP);
336 if (atomic_inc_return(&lock->cll_ref) == 1)
337 atomic_inc(&site->cs_locks.cs_busy);
339 EXPORT_SYMBOL(cl_lock_get_trust);
342 * Helper function destroying the lock that wasn't completely initialized.
344 * Other threads can acquire references to the top-lock through its
345 * sub-locks. Hence, it cannot be cl_lock_free()-ed immediately.
347 static void cl_lock_finish(const struct lu_env *env, struct cl_lock *lock)
349 cl_lock_mutex_get(env, lock);
350 cl_lock_delete(env, lock);
351 cl_lock_mutex_put(env, lock);
352 cl_lock_put(env, lock);
355 static struct cl_lock *cl_lock_alloc(const struct lu_env *env,
356 struct cl_object *obj,
357 const struct cl_io *io,
358 const struct cl_lock_descr *descr)
360 struct cl_lock *lock;
361 struct lu_object_header *head;
362 struct cl_site *site = cl_object_site(obj);
365 OBD_SLAB_ALLOC_PTR_GFP(lock, cl_lock_kmem, CFS_ALLOC_IO);
367 atomic_set(&lock->cll_ref, 1);
368 lock->cll_descr = *descr;
369 lock->cll_state = CLS_NEW;
371 lock->cll_obj_ref = lu_object_ref_add(&obj->co_lu,
373 CFS_INIT_LIST_HEAD(&lock->cll_layers);
374 CFS_INIT_LIST_HEAD(&lock->cll_linkage);
375 CFS_INIT_LIST_HEAD(&lock->cll_inclosure);
376 lu_ref_init(&lock->cll_reference);
377 lu_ref_init(&lock->cll_holders);
378 mutex_init(&lock->cll_guard);
379 lockdep_set_class(&lock->cll_guard, &cl_lock_guard_class);
380 cfs_waitq_init(&lock->cll_wq);
381 head = obj->co_lu.lo_header;
382 atomic_inc(&site->cs_locks_state[CLS_NEW]);
383 atomic_inc(&site->cs_locks.cs_total);
384 atomic_inc(&site->cs_locks.cs_created);
385 cl_lock_lockdep_init(lock);
386 list_for_each_entry(obj, &head->loh_layers, co_lu.lo_linkage) {
389 err = obj->co_ops->coo_lock_init(env, obj, lock, io);
391 cl_lock_finish(env, lock);
397 lock = ERR_PTR(-ENOMEM);
402 * Returns true iff lock is "suitable" for given io. E.g., locks acquired by
403 * truncate and O_APPEND cannot be reused for read/non-append-write, as they
404 * cover multiple stripes and can trigger cascading timeouts.
406 static int cl_lock_fits_into(const struct lu_env *env,
407 const struct cl_lock *lock,
408 const struct cl_lock_descr *need,
409 const struct cl_io *io)
411 const struct cl_lock_slice *slice;
413 LINVRNT(cl_lock_invariant_trusted(env, lock));
415 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
416 if (slice->cls_ops->clo_fits_into != NULL &&
417 !slice->cls_ops->clo_fits_into(env, slice, need, io))
423 static struct cl_lock *cl_lock_lookup(const struct lu_env *env,
424 struct cl_object *obj,
425 const struct cl_io *io,
426 const struct cl_lock_descr *need)
428 struct cl_lock *lock;
429 struct cl_object_header *head;
430 struct cl_site *site;
434 head = cl_object_header(obj);
435 site = cl_object_site(obj);
436 LINVRNT_SPIN_LOCKED(&head->coh_lock_guard);
437 atomic_inc(&site->cs_locks.cs_lookup);
438 list_for_each_entry(lock, &head->coh_locks, cll_linkage) {
441 LASSERT(cl_is_lock(lock));
442 matched = cl_lock_ext_match(&lock->cll_descr, need) &&
443 lock->cll_state < CLS_FREEING &&
444 !(lock->cll_flags & CLF_CANCELLED) &&
445 cl_lock_fits_into(env, lock, need, io);
446 CDEBUG(D_DLMTRACE, "has: "DDESCR"(%i) need: "DDESCR": %d\n",
447 PDESCR(&lock->cll_descr), lock->cll_state, PDESCR(need),
450 cl_lock_get_trust(lock);
451 /* move the lock to the LRU head */
452 list_move(&lock->cll_linkage, &head->coh_locks);
453 atomic_inc(&cl_object_site(obj)->cs_locks.cs_hit);
461 * Returns a lock matching description \a need.
463 * This is the main entry point into the cl_lock caching interface. First, a
464 * cache (implemented as a per-object linked list) is consulted. If lock is
465 * found there, it is returned immediately. Otherwise new lock is allocated
466 * and returned. In any case, additional reference to lock is acquired.
468 * \see cl_object_find(), cl_page_find()
470 static struct cl_lock *cl_lock_find(const struct lu_env *env,
471 const struct cl_io *io,
472 const struct cl_lock_descr *need)
474 struct cl_object_header *head;
475 struct cl_object *obj;
476 struct cl_lock *lock;
477 struct cl_site *site;
482 head = cl_object_header(obj);
483 site = cl_object_site(obj);
485 spin_lock(&head->coh_lock_guard);
486 lock = cl_lock_lookup(env, obj, io, need);
487 spin_unlock(&head->coh_lock_guard);
490 lock = cl_lock_alloc(env, obj, io, need);
492 struct cl_lock *ghost;
494 spin_lock(&head->coh_lock_guard);
495 ghost = cl_lock_lookup(env, obj, io, need);
497 list_add(&lock->cll_linkage, &head->coh_locks);
498 spin_unlock(&head->coh_lock_guard);
499 atomic_inc(&site->cs_locks.cs_busy);
501 spin_unlock(&head->coh_lock_guard);
503 * Other threads can acquire references to the
504 * top-lock through its sub-locks. Hence, it
505 * cannot be cl_lock_free()-ed immediately.
507 cl_lock_finish(env, lock);
516 * Returns existing lock matching given description. This is similar to
517 * cl_lock_find() except that no new lock is created, and returned lock is
518 * guaranteed to be in enum cl_lock_state::CLS_HELD state.
520 struct cl_lock *cl_lock_peek(const struct lu_env *env, const struct cl_io *io,
521 const struct cl_lock_descr *need,
522 const char *scope, const void *source)
524 struct cl_object_header *head;
525 struct cl_object *obj;
526 struct cl_lock *lock;
529 head = cl_object_header(obj);
531 spin_lock(&head->coh_lock_guard);
532 lock = cl_lock_lookup(env, obj, io, need);
533 spin_unlock(&head->coh_lock_guard);
538 cl_lock_mutex_get(env, lock);
539 if (lock->cll_state == CLS_CACHED)
540 cl_use_try(env, lock);
541 ok = lock->cll_state == CLS_HELD;
543 cl_lock_hold_add(env, lock, scope, source);
544 cl_lock_user_add(env, lock);
545 cl_lock_put(env, lock);
547 cl_lock_mutex_put(env, lock);
549 cl_lock_put(env, lock);
555 EXPORT_SYMBOL(cl_lock_peek);
558 * Returns a slice within a lock, corresponding to the given layer in the
563 const struct cl_lock_slice *cl_lock_at(const struct cl_lock *lock,
564 const struct lu_device_type *dtype)
566 const struct cl_lock_slice *slice;
568 LINVRNT(cl_lock_invariant_trusted(NULL, lock));
571 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
572 if (slice->cls_obj->co_lu.lo_dev->ld_type == dtype)
577 EXPORT_SYMBOL(cl_lock_at);
579 static void cl_lock_trace(struct cl_thread_counters *counters,
580 const char *prefix, const struct cl_lock *lock)
582 CDEBUG(D_DLMTRACE|D_TRACE, "%s: %i@%p %p %i %i\n", prefix,
583 atomic_read(&lock->cll_ref), lock, lock->cll_guarder,
584 lock->cll_depth, counters->ctc_nr_locks_locked);
587 static void cl_lock_mutex_tail(const struct lu_env *env, struct cl_lock *lock)
589 struct cl_thread_counters *counters;
591 counters = cl_lock_counters(env, lock);
593 counters->ctc_nr_locks_locked++;
594 lu_ref_add(&counters->ctc_locks_locked, "cll_guard", lock);
595 cl_lock_trace(counters, "got mutex", lock);
599 * Locks cl_lock object.
601 * This is used to manipulate cl_lock fields, and to serialize state
602 * transitions in the lock state machine.
604 * \post cl_lock_is_mutexed(lock)
606 * \see cl_lock_mutex_put()
608 void cl_lock_mutex_get(const struct lu_env *env, struct cl_lock *lock)
610 LINVRNT(cl_lock_invariant(env, lock));
612 if (lock->cll_guarder == cfs_current()) {
613 LINVRNT(cl_lock_is_mutexed(lock));
614 LINVRNT(lock->cll_depth > 0);
616 struct cl_object_header *hdr;
617 struct cl_thread_info *info;
620 LINVRNT(lock->cll_guarder != cfs_current());
621 hdr = cl_object_header(lock->cll_descr.cld_obj);
623 * Check that mutices are taken in the bottom-to-top order.
625 info = cl_env_info(env);
626 for (i = 0; i < hdr->coh_nesting; ++i)
627 LASSERT(info->clt_counters[i].ctc_nr_locks_locked == 0);
628 mutex_lock_nested(&lock->cll_guard, hdr->coh_nesting);
629 lock->cll_guarder = cfs_current();
630 LINVRNT(lock->cll_depth == 0);
632 cl_lock_mutex_tail(env, lock);
634 EXPORT_SYMBOL(cl_lock_mutex_get);
637 * Try-locks cl_lock object.
639 * \retval 0 \a lock was successfully locked
641 * \retval -EBUSY \a lock cannot be locked right now
643 * \post ergo(result == 0, cl_lock_is_mutexed(lock))
645 * \see cl_lock_mutex_get()
647 int cl_lock_mutex_try(const struct lu_env *env, struct cl_lock *lock)
651 LINVRNT(cl_lock_invariant_trusted(env, lock));
655 if (lock->cll_guarder == cfs_current()) {
656 LINVRNT(lock->cll_depth > 0);
657 cl_lock_mutex_tail(env, lock);
658 } else if (mutex_trylock(&lock->cll_guard)) {
659 LINVRNT(lock->cll_depth == 0);
660 lock->cll_guarder = cfs_current();
661 cl_lock_mutex_tail(env, lock);
666 EXPORT_SYMBOL(cl_lock_mutex_try);
669 * Unlocks cl_lock object.
671 * \pre cl_lock_is_mutexed(lock)
673 * \see cl_lock_mutex_get()
675 void cl_lock_mutex_put(const struct lu_env *env, struct cl_lock *lock)
677 struct cl_thread_counters *counters;
679 LINVRNT(cl_lock_invariant(env, lock));
680 LINVRNT(cl_lock_is_mutexed(lock));
681 LINVRNT(lock->cll_guarder == cfs_current());
682 LINVRNT(lock->cll_depth > 0);
684 counters = cl_lock_counters(env, lock);
685 LINVRNT(counters->ctc_nr_locks_locked > 0);
687 cl_lock_trace(counters, "put mutex", lock);
688 lu_ref_del(&counters->ctc_locks_locked, "cll_guard", lock);
689 counters->ctc_nr_locks_locked--;
690 if (--lock->cll_depth == 0) {
691 lock->cll_guarder = NULL;
692 mutex_unlock(&lock->cll_guard);
695 EXPORT_SYMBOL(cl_lock_mutex_put);
698 * Returns true iff lock's mutex is owned by the current thread.
700 int cl_lock_is_mutexed(struct cl_lock *lock)
702 return lock->cll_guarder == cfs_current();
704 EXPORT_SYMBOL(cl_lock_is_mutexed);
707 * Returns number of cl_lock mutices held by the current thread (environment).
709 int cl_lock_nr_mutexed(const struct lu_env *env)
711 struct cl_thread_info *info;
716 * NOTE: if summation across all nesting levels (currently 2) proves
717 * too expensive, a summary counter can be added to
718 * struct cl_thread_info.
720 info = cl_env_info(env);
721 for (i = 0, locked = 0; i < ARRAY_SIZE(info->clt_counters); ++i)
722 locked += info->clt_counters[i].ctc_nr_locks_locked;
725 EXPORT_SYMBOL(cl_lock_nr_mutexed);
727 static void cl_lock_cancel0(const struct lu_env *env, struct cl_lock *lock)
729 LINVRNT(cl_lock_is_mutexed(lock));
730 LINVRNT(cl_lock_invariant(env, lock));
732 if (!(lock->cll_flags & CLF_CANCELLED)) {
733 const struct cl_lock_slice *slice;
735 lock->cll_flags |= CLF_CANCELLED;
736 list_for_each_entry_reverse(slice, &lock->cll_layers,
738 if (slice->cls_ops->clo_cancel != NULL)
739 slice->cls_ops->clo_cancel(env, slice);
745 static void cl_lock_delete0(const struct lu_env *env, struct cl_lock *lock)
747 struct cl_object_header *head;
748 const struct cl_lock_slice *slice;
750 LINVRNT(cl_lock_is_mutexed(lock));
751 LINVRNT(cl_lock_invariant(env, lock));
754 if (lock->cll_state < CLS_FREEING) {
755 cl_lock_state_set(env, lock, CLS_FREEING);
757 head = cl_object_header(lock->cll_descr.cld_obj);
759 spin_lock(&head->coh_lock_guard);
760 list_del_init(&lock->cll_linkage);
762 * No locks, no pages. This is only valid for bottom sub-locks
763 * and head->coh_nesting == 1 check assumes two level top-sub
767 * The count of pages of this object may NOT be zero because
768 * we don't cleanup the pages if they are in CPS_FREEING state.
769 * See cl_page_gang_lookup().
771 * It is safe to leave the CPS_FREEING pages in cache w/o
772 * a lock, because those page must not be uptodate.
773 * See cl_page_delete0 for details.
775 /* LASSERT(!ergo(head->coh_nesting == 1 &&
776 list_empty(&head->coh_locks), !head->coh_pages)); */
777 spin_unlock(&head->coh_lock_guard);
779 * From now on, no new references to this lock can be acquired
780 * by cl_lock_lookup().
782 list_for_each_entry_reverse(slice, &lock->cll_layers,
784 if (slice->cls_ops->clo_delete != NULL)
785 slice->cls_ops->clo_delete(env, slice);
788 * From now on, no new references to this lock can be acquired
789 * by layer-specific means (like a pointer from struct
790 * ldlm_lock in osc, or a pointer from top-lock to sub-lock in
793 * Lock will be finally freed in cl_lock_put() when last of
794 * existing references goes away.
801 * Mod(ifie)s cl_lock::cll_holds counter for a given lock. Also, for a
802 * top-lock (nesting == 0) accounts for this modification in the per-thread
803 * debugging counters. Sub-lock holds can be released by a thread different
804 * from one that acquired it.
806 static void cl_lock_hold_mod(const struct lu_env *env, struct cl_lock *lock,
809 struct cl_thread_counters *counters;
810 enum clt_nesting_level nesting;
812 lock->cll_holds += delta;
813 nesting = cl_lock_nesting(lock);
814 if (nesting == CNL_TOP) {
815 counters = &cl_env_info(env)->clt_counters[CNL_TOP];
816 counters->ctc_nr_held += delta;
817 LASSERT(counters->ctc_nr_held >= 0);
822 * Mod(ifie)s cl_lock::cll_users counter for a given lock. See
823 * cl_lock_hold_mod() for the explanation of the debugging code.
825 static void cl_lock_used_mod(const struct lu_env *env, struct cl_lock *lock,
828 struct cl_thread_counters *counters;
829 enum clt_nesting_level nesting;
831 lock->cll_users += delta;
832 nesting = cl_lock_nesting(lock);
833 if (nesting == CNL_TOP) {
834 counters = &cl_env_info(env)->clt_counters[CNL_TOP];
835 counters->ctc_nr_used += delta;
836 LASSERT(counters->ctc_nr_used >= 0);
840 static void cl_lock_hold_release(const struct lu_env *env, struct cl_lock *lock,
841 const char *scope, const void *source)
843 LINVRNT(cl_lock_is_mutexed(lock));
844 LINVRNT(cl_lock_invariant(env, lock));
845 LASSERT(lock->cll_holds > 0);
848 lu_ref_del(&lock->cll_holders, scope, source);
849 cl_lock_hold_mod(env, lock, -1);
850 if (lock->cll_holds == 0) {
851 if (lock->cll_descr.cld_mode == CLM_PHANTOM ||
852 lock->cll_descr.cld_mode == CLM_GROUP)
854 * If lock is still phantom or grouplock when user is
855 * done with it---destroy the lock.
857 lock->cll_flags |= CLF_CANCELPEND|CLF_DOOMED;
858 if (lock->cll_flags & CLF_CANCELPEND) {
859 lock->cll_flags &= ~CLF_CANCELPEND;
860 cl_lock_cancel0(env, lock);
862 if (lock->cll_flags & CLF_DOOMED) {
863 /* no longer doomed: it's dead... Jim. */
864 lock->cll_flags &= ~CLF_DOOMED;
865 cl_lock_delete0(env, lock);
873 * Waits until lock state is changed.
875 * This function is called with cl_lock mutex locked, atomically releases
876 * mutex and goes to sleep, waiting for a lock state change (signaled by
877 * cl_lock_signal()), and re-acquires the mutex before return.
879 * This function is used to wait until lock state machine makes some progress
880 * and to emulate synchronous operations on top of asynchronous lock
883 * \retval -EINTR wait was interrupted
885 * \retval 0 wait wasn't interrupted
887 * \pre cl_lock_is_mutexed(lock)
889 * \see cl_lock_signal()
891 int cl_lock_state_wait(const struct lu_env *env, struct cl_lock *lock)
893 cfs_waitlink_t waiter;
897 LINVRNT(cl_lock_is_mutexed(lock));
898 LINVRNT(cl_lock_invariant(env, lock));
899 LASSERT(lock->cll_depth == 1);
900 LASSERT(lock->cll_state != CLS_FREEING); /* too late to wait */
902 result = lock->cll_error;
903 if (result == 0 && !(lock->cll_flags & CLF_STATE)) {
904 cfs_waitlink_init(&waiter);
905 cfs_waitq_add(&lock->cll_wq, &waiter);
906 set_current_state(CFS_TASK_INTERRUPTIBLE);
907 cl_lock_mutex_put(env, lock);
909 LASSERT(cl_lock_nr_mutexed(env) == 0);
910 cfs_waitq_wait(&waiter, CFS_TASK_INTERRUPTIBLE);
912 cl_lock_mutex_get(env, lock);
913 set_current_state(CFS_TASK_RUNNING);
914 cfs_waitq_del(&lock->cll_wq, &waiter);
915 result = cfs_signal_pending() ? -EINTR : 0;
917 lock->cll_flags &= ~CLF_STATE;
920 EXPORT_SYMBOL(cl_lock_state_wait);
922 static void cl_lock_state_signal(const struct lu_env *env, struct cl_lock *lock,
923 enum cl_lock_state state)
925 const struct cl_lock_slice *slice;
928 LINVRNT(cl_lock_is_mutexed(lock));
929 LINVRNT(cl_lock_invariant(env, lock));
931 list_for_each_entry(slice, &lock->cll_layers, cls_linkage)
932 if (slice->cls_ops->clo_state != NULL)
933 slice->cls_ops->clo_state(env, slice, state);
934 lock->cll_flags |= CLF_STATE;
935 cfs_waitq_broadcast(&lock->cll_wq);
940 * Notifies waiters that lock state changed.
942 * Wakes up all waiters sleeping in cl_lock_state_wait(), also notifies all
943 * layers about state change by calling cl_lock_operations::clo_state()
946 void cl_lock_signal(const struct lu_env *env, struct cl_lock *lock)
949 cl_lock_state_signal(env, lock, lock->cll_state);
952 EXPORT_SYMBOL(cl_lock_signal);
955 * Changes lock state.
957 * This function is invoked to notify layers that lock state changed, possible
958 * as a result of an asynchronous event such as call-back reception.
960 * \post lock->cll_state == state
962 * \see cl_lock_operations::clo_state()
964 void cl_lock_state_set(const struct lu_env *env, struct cl_lock *lock,
965 enum cl_lock_state state)
967 struct cl_site *site = cl_object_site(lock->cll_descr.cld_obj);
970 LASSERT(lock->cll_state <= state ||
971 (lock->cll_state == CLS_CACHED &&
972 (state == CLS_HELD || /* lock found in cache */
973 state == CLS_NEW /* sub-lock canceled */)) ||
974 /* sub-lock canceled during unlocking */
975 (lock->cll_state == CLS_UNLOCKING && state == CLS_NEW));
977 if (lock->cll_state != state) {
978 atomic_dec(&site->cs_locks_state[lock->cll_state]);
979 atomic_inc(&site->cs_locks_state[state]);
981 cl_lock_state_signal(env, lock, state);
982 lock->cll_state = state;
986 EXPORT_SYMBOL(cl_lock_state_set);
989 * Yanks lock from the cache (cl_lock_state::CLS_CACHED state) by calling
990 * cl_lock_operations::clo_use() top-to-bottom to notify layers.
992 int cl_use_try(const struct lu_env *env, struct cl_lock *lock)
995 const struct cl_lock_slice *slice;
999 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1000 if (slice->cls_ops->clo_use != NULL) {
1001 result = slice->cls_ops->clo_use(env, slice);
1006 LASSERT(result != -ENOSYS);
1008 cl_lock_state_set(env, lock, CLS_HELD);
1011 EXPORT_SYMBOL(cl_use_try);
1014 * Helper for cl_enqueue_try() that calls ->clo_enqueue() across all layers
1017 static int cl_enqueue_kick(const struct lu_env *env,
1018 struct cl_lock *lock,
1019 struct cl_io *io, __u32 flags)
1022 const struct cl_lock_slice *slice;
1026 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1027 if (slice->cls_ops->clo_enqueue != NULL) {
1028 result = slice->cls_ops->clo_enqueue(env,
1034 LASSERT(result != -ENOSYS);
1039 * Tries to enqueue a lock.
1041 * This function is called repeatedly by cl_enqueue() until either lock is
1042 * enqueued, or error occurs. This function does not block waiting for
1043 * networking communication to complete.
1045 * \post ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1046 * lock->cll_state == CLS_HELD)
1048 * \see cl_enqueue() cl_lock_operations::clo_enqueue()
1049 * \see cl_lock_state::CLS_ENQUEUED
1051 int cl_enqueue_try(const struct lu_env *env, struct cl_lock *lock,
1052 struct cl_io *io, __u32 flags)
1060 LINVRNT(cl_lock_is_mutexed(lock));
1062 if (lock->cll_error != 0)
1064 switch (lock->cll_state) {
1066 cl_lock_state_set(env, lock, CLS_QUEUING);
1070 result = cl_enqueue_kick(env, lock, io, flags);
1072 cl_lock_state_set(env, lock, CLS_ENQUEUED);
1075 /* wait until unlocking finishes, and enqueue lock
1080 /* yank lock from the cache. */
1081 result = cl_use_try(env, lock);
1090 * impossible, only held locks with increased
1091 * ->cll_holds can be enqueued, and they cannot be
1096 } while (result == CLO_REPEAT);
1098 cl_lock_error(env, lock, result);
1099 RETURN(result ?: lock->cll_error);
1101 EXPORT_SYMBOL(cl_enqueue_try);
1103 static int cl_enqueue_locked(const struct lu_env *env, struct cl_lock *lock,
1104 struct cl_io *io, __u32 enqflags)
1110 LINVRNT(cl_lock_is_mutexed(lock));
1111 LINVRNT(cl_lock_invariant(env, lock));
1112 LASSERT(lock->cll_holds > 0);
1114 cl_lock_user_add(env, lock);
1116 result = cl_enqueue_try(env, lock, io, enqflags);
1117 if (result == CLO_WAIT) {
1118 result = cl_lock_state_wait(env, lock);
1125 cl_lock_user_del(env, lock);
1126 if (result != -EINTR)
1127 cl_lock_error(env, lock, result);
1129 LASSERT(ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1130 lock->cll_state == CLS_HELD));
1137 * \pre current thread or io owns a hold on lock.
1139 * \post ergo(result == 0, lock->users increased)
1140 * \post ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1141 * lock->cll_state == CLS_HELD)
1143 int cl_enqueue(const struct lu_env *env, struct cl_lock *lock,
1144 struct cl_io *io, __u32 enqflags)
1150 cl_lock_lockdep_acquire(env, lock, enqflags);
1151 cl_lock_mutex_get(env, lock);
1152 result = cl_enqueue_locked(env, lock, io, enqflags);
1153 cl_lock_mutex_put(env, lock);
1155 cl_lock_lockdep_release(env, lock);
1156 LASSERT(ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1157 lock->cll_state == CLS_HELD));
1160 EXPORT_SYMBOL(cl_enqueue);
1163 * Tries to unlock a lock.
1165 * This function is called repeatedly by cl_unuse() until either lock is
1166 * unlocked, or error occurs.
1168 * \ppre lock->cll_state <= CLS_HELD || lock->cll_state == CLS_UNLOCKING
1170 * \post ergo(result == 0, lock->cll_state == CLS_CACHED)
1172 * \see cl_unuse() cl_lock_operations::clo_unuse()
1173 * \see cl_lock_state::CLS_CACHED
1175 int cl_unuse_try(const struct lu_env *env, struct cl_lock *lock)
1177 const struct cl_lock_slice *slice;
1181 if (lock->cll_state != CLS_UNLOCKING) {
1182 if (lock->cll_users > 1) {
1183 cl_lock_user_del(env, lock);
1187 * New lock users (->cll_users) are not protecting unlocking
1188 * from proceeding. From this point, lock eventually reaches
1189 * CLS_CACHED, is reinitialized to CLS_NEW or fails into
1192 cl_lock_state_set(env, lock, CLS_UNLOCKING);
1197 if (lock->cll_error != 0)
1200 LINVRNT(cl_lock_is_mutexed(lock));
1201 LINVRNT(cl_lock_invariant(env, lock));
1202 LASSERT(lock->cll_state == CLS_UNLOCKING);
1203 LASSERT(lock->cll_users > 0);
1204 LASSERT(lock->cll_holds > 0);
1207 list_for_each_entry_reverse(slice, &lock->cll_layers,
1209 if (slice->cls_ops->clo_unuse != NULL) {
1210 result = slice->cls_ops->clo_unuse(env, slice);
1215 LASSERT(result != -ENOSYS);
1216 } while (result == CLO_REPEAT);
1217 if (result != CLO_WAIT)
1219 * Once there is no more need to iterate ->clo_unuse() calls,
1220 * remove lock user. This is done even if unrecoverable error
1221 * happened during unlocking, because nothing else can be
1224 cl_lock_user_del(env, lock);
1225 if (result == 0 || result == -ESTALE) {
1226 enum cl_lock_state state;
1229 * Return lock back to the cache. This is the only
1230 * place where lock is moved into CLS_CACHED state.
1232 * If one of ->clo_unuse() methods returned -ESTALE, lock
1233 * cannot be placed into cache and has to be
1234 * re-initialized. This happens e.g., when a sub-lock was
1235 * canceled while unlocking was in progress.
1237 state = result == 0 ? CLS_CACHED : CLS_NEW;
1238 cl_lock_state_set(env, lock, state);
1241 * Hide -ESTALE error.
1242 * If the lock is a glimpse lock, and it has multiple
1243 * stripes. Assuming that one of its sublock returned -ENAVAIL,
1244 * and other sublocks are matched write locks. In this case,
1245 * we can't set this lock to error because otherwise some of
1246 * its sublocks may not be canceled. This causes some dirty
1247 * pages won't be written to OSTs. -jay
1251 result = result ?: lock->cll_error;
1253 cl_lock_error(env, lock, result);
1256 EXPORT_SYMBOL(cl_unuse_try);
1258 static void cl_unuse_locked(const struct lu_env *env, struct cl_lock *lock)
1261 LASSERT(lock->cll_state <= CLS_HELD);
1265 result = cl_unuse_try(env, lock);
1266 if (result == CLO_WAIT) {
1267 result = cl_lock_state_wait(env, lock);
1279 void cl_unuse(const struct lu_env *env, struct cl_lock *lock)
1282 cl_lock_mutex_get(env, lock);
1283 cl_unuse_locked(env, lock);
1284 cl_lock_mutex_put(env, lock);
1285 cl_lock_lockdep_release(env, lock);
1288 EXPORT_SYMBOL(cl_unuse);
1291 * Tries to wait for a lock.
1293 * This function is called repeatedly by cl_wait() until either lock is
1294 * granted, or error occurs. This function does not block waiting for network
1295 * communication to complete.
1297 * \see cl_wait() cl_lock_operations::clo_wait()
1298 * \see cl_lock_state::CLS_HELD
1300 int cl_wait_try(const struct lu_env *env, struct cl_lock *lock)
1302 const struct cl_lock_slice *slice;
1307 LINVRNT(cl_lock_is_mutexed(lock));
1308 LINVRNT(cl_lock_invariant(env, lock));
1309 LASSERT(lock->cll_state == CLS_ENQUEUED ||
1310 lock->cll_state == CLS_HELD);
1311 LASSERT(lock->cll_users > 0);
1312 LASSERT(lock->cll_holds > 0);
1315 if (lock->cll_error != 0)
1317 if (lock->cll_state == CLS_HELD)
1322 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1323 if (slice->cls_ops->clo_wait != NULL) {
1324 result = slice->cls_ops->clo_wait(env, slice);
1329 LASSERT(result != -ENOSYS);
1331 cl_lock_state_set(env, lock, CLS_HELD);
1332 } while (result == CLO_REPEAT);
1333 RETURN(result ?: lock->cll_error);
1335 EXPORT_SYMBOL(cl_wait_try);
1338 * Waits until enqueued lock is granted.
1340 * \pre current thread or io owns a hold on the lock
1341 * \pre ergo(result == 0, lock->cll_state == CLS_ENQUEUED ||
1342 * lock->cll_state == CLS_HELD)
1344 * \post ergo(result == 0, lock->cll_state == CLS_HELD)
1346 int cl_wait(const struct lu_env *env, struct cl_lock *lock)
1351 cl_lock_mutex_get(env, lock);
1353 LINVRNT(cl_lock_invariant(env, lock));
1354 LASSERT(lock->cll_state == CLS_ENQUEUED || lock->cll_state == CLS_HELD);
1355 LASSERT(lock->cll_holds > 0);
1358 result = cl_wait_try(env, lock);
1359 if (result == CLO_WAIT) {
1360 result = cl_lock_state_wait(env, lock);
1367 cl_lock_user_del(env, lock);
1368 if (result != -EINTR)
1369 cl_lock_error(env, lock, result);
1370 cl_lock_lockdep_release(env, lock);
1372 cl_lock_mutex_put(env, lock);
1373 LASSERT(ergo(result == 0, lock->cll_state == CLS_HELD));
1376 EXPORT_SYMBOL(cl_wait);
1379 * Executes cl_lock_operations::clo_weigh(), and sums results to estimate lock
1382 unsigned long cl_lock_weigh(const struct lu_env *env, struct cl_lock *lock)
1384 const struct cl_lock_slice *slice;
1385 unsigned long pound;
1386 unsigned long ounce;
1389 LINVRNT(cl_lock_is_mutexed(lock));
1390 LINVRNT(cl_lock_invariant(env, lock));
1393 list_for_each_entry_reverse(slice, &lock->cll_layers, cls_linkage) {
1394 if (slice->cls_ops->clo_weigh != NULL) {
1395 ounce = slice->cls_ops->clo_weigh(env, slice);
1397 if (pound < ounce) /* over-weight^Wflow */
1403 EXPORT_SYMBOL(cl_lock_weigh);
1406 * Notifies layers that lock description changed.
1408 * The server can grant client a lock different from one that was requested
1409 * (e.g., larger in extent). This method is called when actually granted lock
1410 * description becomes known to let layers to accommodate for changed lock
1413 * \see cl_lock_operations::clo_modify()
1415 int cl_lock_modify(const struct lu_env *env, struct cl_lock *lock,
1416 const struct cl_lock_descr *desc)
1418 const struct cl_lock_slice *slice;
1419 struct cl_object *obj = lock->cll_descr.cld_obj;
1420 struct cl_object_header *hdr = cl_object_header(obj);
1424 /* don't allow object to change */
1425 LASSERT(obj == desc->cld_obj);
1426 LINVRNT(cl_lock_is_mutexed(lock));
1427 LINVRNT(cl_lock_invariant(env, lock));
1429 list_for_each_entry_reverse(slice, &lock->cll_layers, cls_linkage) {
1430 if (slice->cls_ops->clo_modify != NULL) {
1431 result = slice->cls_ops->clo_modify(env, slice, desc);
1436 CL_LOCK_DEBUG(D_DLMTRACE, env, lock, " -> "DDESCR"@"DFID"\n",
1437 PDESCR(desc), PFID(lu_object_fid(&desc->cld_obj->co_lu)));
1439 * Just replace description in place. Nothing more is needed for
1440 * now. If locks were indexed according to their extent and/or mode,
1441 * that index would have to be updated here.
1443 spin_lock(&hdr->coh_lock_guard);
1444 lock->cll_descr = *desc;
1445 spin_unlock(&hdr->coh_lock_guard);
1448 EXPORT_SYMBOL(cl_lock_modify);
1451 * Initializes lock closure with a given origin.
1453 * \see cl_lock_closure
1455 void cl_lock_closure_init(const struct lu_env *env,
1456 struct cl_lock_closure *closure,
1457 struct cl_lock *origin, int wait)
1459 LINVRNT(cl_lock_is_mutexed(origin));
1460 LINVRNT(cl_lock_invariant(env, origin));
1462 CFS_INIT_LIST_HEAD(&closure->clc_list);
1463 closure->clc_origin = origin;
1464 closure->clc_wait = wait;
1465 closure->clc_nr = 0;
1467 EXPORT_SYMBOL(cl_lock_closure_init);
1470 * Builds a closure of \a lock.
1472 * Building of a closure consists of adding initial lock (\a lock) into it,
1473 * and calling cl_lock_operations::clo_closure() methods of \a lock. These
1474 * methods might call cl_lock_closure_build() recursively again, adding more
1475 * locks to the closure, etc.
1477 * \see cl_lock_closure
1479 int cl_lock_closure_build(const struct lu_env *env, struct cl_lock *lock,
1480 struct cl_lock_closure *closure)
1482 const struct cl_lock_slice *slice;
1486 LINVRNT(cl_lock_is_mutexed(closure->clc_origin));
1487 LINVRNT(cl_lock_invariant(env, closure->clc_origin));
1489 result = cl_lock_enclosure(env, lock, closure);
1491 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
1492 if (slice->cls_ops->clo_closure != NULL) {
1493 result = slice->cls_ops->clo_closure(env, slice,
1501 cl_lock_disclosure(env, closure);
1504 EXPORT_SYMBOL(cl_lock_closure_build);
1507 * Adds new lock to a closure.
1509 * Try-locks \a lock and if succeeded, adds it to the closure (never more than
1510 * once). If try-lock failed, returns CLO_REPEAT, after optionally waiting
1511 * until next try-lock is likely to succeed.
1513 int cl_lock_enclosure(const struct lu_env *env, struct cl_lock *lock,
1514 struct cl_lock_closure *closure)
1518 if (!cl_lock_mutex_try(env, lock)) {
1520 * If lock->cll_inclosure is not empty, lock is already in
1523 if (list_empty(&lock->cll_inclosure)) {
1524 cl_lock_get_trust(lock);
1525 lu_ref_add(&lock->cll_reference, "closure", closure);
1526 list_add(&lock->cll_inclosure, &closure->clc_list);
1529 cl_lock_mutex_put(env, lock);
1532 cl_lock_disclosure(env, closure);
1533 if (closure->clc_wait) {
1534 cl_lock_get_trust(lock);
1535 lu_ref_add(&lock->cll_reference, "closure-w", closure);
1536 cl_lock_mutex_put(env, closure->clc_origin);
1538 LASSERT(cl_lock_nr_mutexed(env) == 0);
1539 cl_lock_mutex_get(env, lock);
1540 cl_lock_mutex_put(env, lock);
1542 cl_lock_mutex_get(env, closure->clc_origin);
1543 lu_ref_del(&lock->cll_reference, "closure-w", closure);
1544 cl_lock_put(env, lock);
1546 result = CLO_REPEAT;
1550 EXPORT_SYMBOL(cl_lock_enclosure);
1552 /** Releases mutices of enclosed locks. */
1553 void cl_lock_disclosure(const struct lu_env *env,
1554 struct cl_lock_closure *closure)
1556 struct cl_lock *scan;
1557 struct cl_lock *temp;
1559 list_for_each_entry_safe(scan, temp, &closure->clc_list, cll_inclosure){
1560 list_del_init(&scan->cll_inclosure);
1561 cl_lock_mutex_put(env, scan);
1562 lu_ref_del(&scan->cll_reference, "closure", closure);
1563 cl_lock_put(env, scan);
1566 LASSERT(closure->clc_nr == 0);
1568 EXPORT_SYMBOL(cl_lock_disclosure);
1570 /** Finalizes a closure. */
1571 void cl_lock_closure_fini(struct cl_lock_closure *closure)
1573 LASSERT(closure->clc_nr == 0);
1574 LASSERT(list_empty(&closure->clc_list));
1576 EXPORT_SYMBOL(cl_lock_closure_fini);
1579 * Destroys this lock. Notifies layers (bottom-to-top) that lock is being
1580 * destroyed, then destroy the lock. If there are holds on the lock, postpone
1581 * destruction until all holds are released. This is called when a decision is
1582 * made to destroy the lock in the future. E.g., when a blocking AST is
1583 * received on it, or fatal communication error happens.
1585 * Caller must have a reference on this lock to prevent a situation, when
1586 * deleted lock lingers in memory for indefinite time, because nobody calls
1587 * cl_lock_put() to finish it.
1589 * \pre atomic_read(&lock->cll_ref) > 0
1590 * \pre ergo(cl_lock_nesting(lock) == CNL_TOP,
1591 * cl_lock_nr_mutexed(env) == 1)
1592 * [i.e., if a top-lock is deleted, mutices of no other locks can be
1593 * held, as deletion of sub-locks might require releasing a top-lock
1596 * \see cl_lock_operations::clo_delete()
1597 * \see cl_lock::cll_holds
1599 void cl_lock_delete(const struct lu_env *env, struct cl_lock *lock)
1601 LINVRNT(cl_lock_is_mutexed(lock));
1602 LINVRNT(cl_lock_invariant(env, lock));
1603 LASSERT(ergo(cl_lock_nesting(lock) == CNL_TOP,
1604 cl_lock_nr_mutexed(env) == 1));
1607 if (lock->cll_holds == 0)
1608 cl_lock_delete0(env, lock);
1610 lock->cll_flags |= CLF_DOOMED;
1613 EXPORT_SYMBOL(cl_lock_delete);
1616 * Mark lock as irrecoverably failed, and mark it for destruction. This
1617 * happens when, e.g., server fails to grant a lock to us, or networking
1620 * \pre atomic_read(&lock->cll_ref) > 0
1622 * \see clo_lock_delete()
1623 * \see cl_lock::cll_holds
1625 void cl_lock_error(const struct lu_env *env, struct cl_lock *lock, int error)
1627 LINVRNT(cl_lock_is_mutexed(lock));
1628 LINVRNT(cl_lock_invariant(env, lock));
1631 if (lock->cll_error == 0 && error != 0) {
1632 lock->cll_error = error;
1633 cl_lock_signal(env, lock);
1634 cl_lock_cancel(env, lock);
1635 cl_lock_delete(env, lock);
1639 EXPORT_SYMBOL(cl_lock_error);
1642 * Cancels this lock. Notifies layers
1643 * (bottom-to-top) that lock is being cancelled, then destroy the lock. If
1644 * there are holds on the lock, postpone cancellation until
1645 * all holds are released.
1647 * Cancellation notification is delivered to layers at most once.
1649 * \see cl_lock_operations::clo_cancel()
1650 * \see cl_lock::cll_holds
1652 void cl_lock_cancel(const struct lu_env *env, struct cl_lock *lock)
1654 LINVRNT(cl_lock_is_mutexed(lock));
1655 LINVRNT(cl_lock_invariant(env, lock));
1658 if (lock->cll_holds == 0)
1659 cl_lock_cancel0(env, lock);
1661 lock->cll_flags |= CLF_CANCELPEND;
1664 EXPORT_SYMBOL(cl_lock_cancel);
1667 * Finds an existing lock covering given page and optionally different from a
1668 * given \a except lock.
1670 struct cl_lock *cl_lock_at_page(const struct lu_env *env, struct cl_object *obj,
1671 struct cl_page *page, struct cl_lock *except,
1672 int pending, int canceld)
1674 struct cl_object_header *head;
1675 struct cl_lock *scan;
1676 struct cl_lock *lock;
1677 struct cl_lock_descr *need;
1681 head = cl_object_header(obj);
1682 need = &cl_env_info(env)->clt_descr;
1685 need->cld_mode = CLM_READ; /* CLM_READ matches both READ & WRITE, but
1687 need->cld_start = need->cld_end = page->cp_index;
1689 spin_lock(&head->coh_lock_guard);
1690 list_for_each_entry(scan, &head->coh_locks, cll_linkage) {
1691 if (scan != except &&
1692 cl_lock_ext_match(&scan->cll_descr, need) &&
1693 scan->cll_state >= CLS_HELD &&
1694 scan->cll_state < CLS_FREEING &&
1696 * This check is racy as the lock can be canceled right
1697 * after it is done, but this is fine, because page exists
1700 (canceld || !(scan->cll_flags & CLF_CANCELLED)) &&
1701 (pending || !(scan->cll_flags & CLF_CANCELPEND))) {
1702 /* Don't increase cs_hit here since this
1703 * is just a helper function. */
1704 cl_lock_get_trust(scan);
1709 spin_unlock(&head->coh_lock_guard);
1712 EXPORT_SYMBOL(cl_lock_at_page);
1715 * Returns a list of pages protected (only) by a given lock.
1717 * Scans an extent of page radix tree, corresponding to the \a lock and queues
1718 * all pages that are not protected by locks other than \a lock into \a queue.
1720 void cl_lock_page_list_fixup(const struct lu_env *env,
1721 struct cl_io *io, struct cl_lock *lock,
1722 struct cl_page_list *queue)
1724 struct cl_page *page;
1725 struct cl_page *temp;
1726 struct cl_page_list *plist = &cl_env_info(env)->clt_list;
1728 LINVRNT(cl_lock_invariant(env, lock));
1731 /* Now, we have a list of cl_pages under the \a lock, we need
1732 * to check if some of pages are covered by other ldlm lock.
1733 * If this is the case, they aren't needed to be written out this time.
1735 * For example, we have A:[0,200] & B:[100,300] PW locks on client, now
1736 * the latter is to be canceled, this means other client is
1737 * reading/writing [200,300] since A won't canceled. Actually
1738 * we just need to write the pages covered by [200,300]. This is safe,
1739 * since [100,200] is also protected lock A.
1742 cl_page_list_init(plist);
1743 cl_page_list_for_each_safe(page, temp, queue) {
1744 pgoff_t idx = page->cp_index;
1745 struct cl_lock *found;
1746 struct cl_lock_descr *descr;
1748 /* The algorithm counts on the index-ascending page index. */
1749 LASSERT(ergo(&temp->cp_batch != &queue->pl_pages,
1750 page->cp_index < temp->cp_index));
1752 found = cl_lock_at_page(env, lock->cll_descr.cld_obj,
1757 descr = &found->cll_descr;
1758 list_for_each_entry_safe_from(page, temp, &queue->pl_pages,
1760 idx = page->cp_index;
1761 if (descr->cld_start > idx || descr->cld_end < idx)
1763 cl_page_list_move(plist, queue, page);
1765 cl_lock_put(env, found);
1768 /* The pages in plist are covered by other locks, don't handle them
1772 cl_page_list_disown(env, io, plist);
1773 cl_page_list_fini(env, plist);
1776 EXPORT_SYMBOL(cl_lock_page_list_fixup);
1779 * Invalidate pages protected by the given lock, sending them out to the
1780 * server first, if necessary.
1782 * This function does the following:
1784 * - collects a list of pages to be invalidated,
1786 * - unmaps them from the user virtual memory,
1788 * - sends dirty pages to the server,
1790 * - waits for transfer completion,
1792 * - discards pages, and throws them out of memory.
1794 * If \a discard is set, pages are discarded without sending them to the
1797 * If error happens on any step, the process continues anyway (the reasoning
1798 * behind this being that lock cancellation cannot be delayed indefinitely).
1800 int cl_lock_page_out(const struct lu_env *env, struct cl_lock *lock,
1803 struct cl_thread_info *info = cl_env_info(env);
1804 struct cl_io *io = &info->clt_io;
1805 struct cl_2queue *queue = &info->clt_queue;
1806 struct cl_lock_descr *descr = &lock->cll_descr;
1810 LINVRNT(cl_lock_invariant(env, lock));
1813 io->ci_obj = cl_object_top(descr->cld_obj);
1814 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1816 cl_2queue_init(queue);
1817 cl_page_gang_lookup(env, descr->cld_obj, io, descr->cld_start,
1818 descr->cld_end, &queue->c2_qin);
1819 page_count = queue->c2_qin.pl_nr;
1820 if (page_count > 0) {
1821 result = cl_page_list_unmap(env, io, &queue->c2_qin);
1823 long timeout = 600; /* 10 minutes. */
1824 /* for debug purpose, if this request can't be
1825 * finished in 10 minutes, we hope it can
1828 result = cl_io_submit_sync(env, io, CRT_WRITE,
1832 CWARN("Writing %lu pages error: %d\n",
1833 page_count, result);
1835 cl_lock_page_list_fixup(env, io, lock, &queue->c2_qout);
1836 cl_2queue_discard(env, io, queue);
1837 cl_2queue_disown(env, io, queue);
1839 cl_2queue_fini(env, queue);
1841 cl_io_fini(env, io);
1844 EXPORT_SYMBOL(cl_lock_page_out);
1847 * Eliminate all locks for a given object.
1849 * Caller has to guarantee that no lock is in active use.
1851 * \param cancel when this is set, cl_locks_prune() cancels locks before
1854 void cl_locks_prune(const struct lu_env *env, struct cl_object *obj, int cancel)
1856 struct cl_object_header *head;
1857 struct cl_lock *lock;
1860 head = cl_object_header(obj);
1862 * If locks are destroyed without cancellation, all pages must be
1863 * already destroyed (as otherwise they will be left unprotected).
1865 LASSERT(ergo(!cancel,
1866 head->coh_tree.rnode == NULL && head->coh_pages == 0));
1868 spin_lock(&head->coh_lock_guard);
1869 while (!list_empty(&head->coh_locks)) {
1870 lock = container_of(head->coh_locks.next,
1871 struct cl_lock, cll_linkage);
1872 cl_lock_get_trust(lock);
1873 spin_unlock(&head->coh_lock_guard);
1874 lu_ref_add(&lock->cll_reference, "prune", cfs_current());
1875 cl_lock_mutex_get(env, lock);
1876 if (lock->cll_state < CLS_FREEING) {
1877 LASSERT(lock->cll_holds == 0);
1878 LASSERT(lock->cll_users == 0);
1880 cl_lock_cancel(env, lock);
1881 cl_lock_delete(env, lock);
1883 cl_lock_mutex_put(env, lock);
1884 lu_ref_del(&lock->cll_reference, "prune", cfs_current());
1885 cl_lock_put(env, lock);
1886 spin_lock(&head->coh_lock_guard);
1888 spin_unlock(&head->coh_lock_guard);
1891 EXPORT_SYMBOL(cl_locks_prune);
1894 * Returns true if \a addr is an address of an allocated cl_lock. Used in
1895 * assertions. This check is optimistically imprecise, i.e., it occasionally
1896 * returns true for the incorrect addresses, but if it returns false, then the
1897 * address is guaranteed to be incorrect. (Should be named cl_lockp().)
1901 int cl_is_lock(const void *addr)
1903 return cfs_mem_is_in_cache(addr, cl_lock_kmem);
1905 EXPORT_SYMBOL(cl_is_lock);
1907 static struct cl_lock *cl_lock_hold_mutex(const struct lu_env *env,
1908 const struct cl_io *io,
1909 const struct cl_lock_descr *need,
1910 const char *scope, const void *source)
1912 struct cl_lock *lock;
1917 lock = cl_lock_find(env, io, need);
1920 cl_lock_mutex_get(env, lock);
1921 if (lock->cll_state < CLS_FREEING) {
1922 cl_lock_hold_mod(env, lock, +1);
1923 lu_ref_add(&lock->cll_holders, scope, source);
1924 lu_ref_add(&lock->cll_reference, scope, source);
1927 cl_lock_mutex_put(env, lock);
1928 cl_lock_put(env, lock);
1934 * Returns a lock matching \a need description with a reference and a hold on
1937 * This is much like cl_lock_find(), except that cl_lock_hold() additionally
1938 * guarantees that lock is not in the CLS_FREEING state on return.
1940 struct cl_lock *cl_lock_hold(const struct lu_env *env, const struct cl_io *io,
1941 const struct cl_lock_descr *need,
1942 const char *scope, const void *source)
1944 struct cl_lock *lock;
1948 lock = cl_lock_hold_mutex(env, io, need, scope, source);
1950 cl_lock_mutex_put(env, lock);
1953 EXPORT_SYMBOL(cl_lock_hold);
1956 * Main high-level entry point of cl_lock interface that finds existing or
1957 * enqueues new lock matching given description.
1959 struct cl_lock *cl_lock_request(const struct lu_env *env, struct cl_io *io,
1960 const struct cl_lock_descr *need,
1962 const char *scope, const void *source)
1964 struct cl_lock *lock;
1965 const struct lu_fid *fid;
1971 fid = lu_object_fid(&io->ci_obj->co_lu);
1974 warn = iter >= 16 && IS_PO2(iter);
1975 CDEBUG(warn ? D_WARNING : D_DLMTRACE,
1976 DDESCR"@"DFID" %i %08x `%s'\n",
1977 PDESCR(need), PFID(fid), iter, enqflags, scope);
1978 lock = cl_lock_hold_mutex(env, io, need, scope, source);
1979 if (!IS_ERR(lock)) {
1980 rc = cl_enqueue_locked(env, lock, io, enqflags);
1982 if (cl_lock_fits_into(env, lock, need, io)) {
1983 cl_lock_mutex_put(env, lock);
1984 cl_lock_lockdep_acquire(env,
1988 CL_LOCK_DEBUG(D_WARNING, env, lock,
1989 "got (see bug 17665)\n");
1990 cl_unuse_locked(env, lock);
1992 cl_lock_hold_release(env, lock, scope, source);
1993 cl_lock_mutex_put(env, lock);
1994 lu_ref_del(&lock->cll_reference, scope, source);
1995 cl_lock_put(env, lock);
2003 EXPORT_SYMBOL(cl_lock_request);
2006 * Adds a hold to a known lock.
2008 void cl_lock_hold_add(const struct lu_env *env, struct cl_lock *lock,
2009 const char *scope, const void *source)
2011 LINVRNT(cl_lock_is_mutexed(lock));
2012 LINVRNT(cl_lock_invariant(env, lock));
2013 LASSERT(lock->cll_state != CLS_FREEING);
2016 cl_lock_hold_mod(env, lock, +1);
2018 lu_ref_add(&lock->cll_holders, scope, source);
2019 lu_ref_add(&lock->cll_reference, scope, source);
2022 EXPORT_SYMBOL(cl_lock_hold_add);
2025 * Releases a hold and a reference on a lock, on which caller acquired a
2028 void cl_lock_unhold(const struct lu_env *env, struct cl_lock *lock,
2029 const char *scope, const void *source)
2031 LINVRNT(cl_lock_invariant(env, lock));
2033 cl_lock_hold_release(env, lock, scope, source);
2034 lu_ref_del(&lock->cll_reference, scope, source);
2035 cl_lock_put(env, lock);
2038 EXPORT_SYMBOL(cl_lock_unhold);
2041 * Releases a hold and a reference on a lock, obtained by cl_lock_hold().
2043 void cl_lock_release(const struct lu_env *env, struct cl_lock *lock,
2044 const char *scope, const void *source)
2046 LINVRNT(cl_lock_invariant(env, lock));
2048 cl_lock_mutex_get(env, lock);
2049 cl_lock_hold_release(env, lock, scope, source);
2050 cl_lock_mutex_put(env, lock);
2051 lu_ref_del(&lock->cll_reference, scope, source);
2052 cl_lock_put(env, lock);
2055 EXPORT_SYMBOL(cl_lock_release);
2057 void cl_lock_user_add(const struct lu_env *env, struct cl_lock *lock)
2059 LINVRNT(cl_lock_is_mutexed(lock));
2060 LINVRNT(cl_lock_invariant(env, lock));
2063 cl_lock_used_mod(env, lock, +1);
2066 EXPORT_SYMBOL(cl_lock_user_add);
2068 int cl_lock_user_del(const struct lu_env *env, struct cl_lock *lock)
2070 LINVRNT(cl_lock_is_mutexed(lock));
2071 LINVRNT(cl_lock_invariant(env, lock));
2072 LASSERT(lock->cll_users > 0);
2075 cl_lock_used_mod(env, lock, -1);
2076 RETURN(lock->cll_users == 0);
2078 EXPORT_SYMBOL(cl_lock_user_del);
2081 * Check if two lock's mode are compatible.
2083 * This returns true iff en-queuing \a lock2 won't cause cancellation of \a
2084 * lock1 even when these locks overlap.
2086 int cl_lock_compatible(const struct cl_lock *lock1, const struct cl_lock *lock2)
2088 enum cl_lock_mode mode1;
2089 enum cl_lock_mode mode2;
2092 mode1 = lock1->cll_descr.cld_mode;
2093 mode2 = lock2->cll_descr.cld_mode;
2094 RETURN(mode2 == CLM_PHANTOM ||
2095 (mode1 == CLM_READ && mode2 == CLM_READ));
2097 EXPORT_SYMBOL(cl_lock_compatible);
2099 const char *cl_lock_mode_name(const enum cl_lock_mode mode)
2101 static const char *names[] = {
2102 [CLM_PHANTOM] = "PHANTOM",
2103 [CLM_READ] = "READ",
2104 [CLM_WRITE] = "WRITE",
2105 [CLM_GROUP] = "GROUP"
2107 if (0 <= mode && mode < ARRAY_SIZE(names))
2112 EXPORT_SYMBOL(cl_lock_mode_name);
2115 * Prints human readable representation of a lock description.
2117 void cl_lock_descr_print(const struct lu_env *env, void *cookie,
2118 lu_printer_t printer,
2119 const struct cl_lock_descr *descr)
2121 const struct lu_fid *fid;
2123 fid = lu_object_fid(&descr->cld_obj->co_lu);
2124 (*printer)(env, cookie, DDESCR"@"DFID, PDESCR(descr), PFID(fid));
2126 EXPORT_SYMBOL(cl_lock_descr_print);
2129 * Prints human readable representation of \a lock to the \a f.
2131 void cl_lock_print(const struct lu_env *env, void *cookie,
2132 lu_printer_t printer, const struct cl_lock *lock)
2134 const struct cl_lock_slice *slice;
2135 (*printer)(env, cookie, "lock@%p[%d %d %d %d %d %08lx] ",
2136 lock, atomic_read(&lock->cll_ref),
2137 lock->cll_state, lock->cll_error, lock->cll_holds,
2138 lock->cll_users, lock->cll_flags);
2139 cl_lock_descr_print(env, cookie, printer, &lock->cll_descr);
2140 (*printer)(env, cookie, " {\n");
2142 list_for_each_entry(slice, &lock->cll_layers, cls_linkage) {
2143 (*printer)(env, cookie, " %s@%p: ",
2144 slice->cls_obj->co_lu.lo_dev->ld_type->ldt_name,
2146 if (slice->cls_ops->clo_print != NULL)
2147 slice->cls_ops->clo_print(env, cookie, printer, slice);
2148 (*printer)(env, cookie, "\n");
2150 (*printer)(env, cookie, "} lock@%p\n", lock);
2152 EXPORT_SYMBOL(cl_lock_print);
2154 int cl_lock_init(void)
2156 return lu_kmem_init(cl_lock_caches);
2159 void cl_lock_fini(void)
2161 lu_kmem_fini(cl_lock_caches);