4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_lock for LOV layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
41 #define DEBUG_SUBSYSTEM S_LOV
43 #include "lov_cl_internal.h"
49 static struct cl_lock_closure *lov_closure_get(const struct lu_env *env,
50 struct cl_lock *parent);
52 static int lov_lock_unuse(const struct lu_env *env,
53 const struct cl_lock_slice *slice);
54 /*****************************************************************************
56 * Lov lock operations.
60 static struct lov_sublock_env *lov_sublock_env_get(const struct lu_env *env,
61 struct cl_lock *parent,
62 struct lov_lock_sub *lls)
64 struct lov_sublock_env *subenv;
65 struct lov_io *lio = lov_env_io(env);
66 struct cl_io *io = lio->lis_cl.cis_io;
67 struct lov_io_sub *sub;
69 subenv = &lov_env_session(env)->ls_subenv;
72 * FIXME: We tend to use the subio's env & io to call the sublock
73 * lock operations because osc lock sometimes stores some control
74 * variables in thread's IO infomation(Now only lockless information).
75 * However, if the lock's host(object) is different from the object
76 * for current IO, we have no way to get the subenv and subio because
77 * they are not initialized at all. As a temp fix, in this case,
78 * we still borrow the parent's env to call sublock operations.
80 if (!io || !cl_object_same(io->ci_obj, parent->cll_descr.cld_obj)) {
81 subenv->lse_env = env;
83 subenv->lse_sub = NULL;
85 sub = lov_sub_get(env, lio, lls->sub_stripe);
87 subenv->lse_env = sub->sub_env;
88 subenv->lse_io = sub->sub_io;
89 subenv->lse_sub = sub;
97 static void lov_sublock_env_put(struct lov_sublock_env *subenv)
99 if (subenv && subenv->lse_sub)
100 lov_sub_put(subenv->lse_sub);
103 static void lov_sublock_adopt(const struct lu_env *env, struct lov_lock *lck,
104 struct cl_lock *sublock, int idx,
105 struct lov_lock_link *link)
107 struct lovsub_lock *lsl;
108 struct cl_lock *parent = lck->lls_cl.cls_lock;
111 LASSERT(cl_lock_is_mutexed(parent));
112 LASSERT(cl_lock_is_mutexed(sublock));
115 lsl = cl2sub_lock(sublock);
117 * check that sub-lock doesn't have lock link to this top-lock.
119 LASSERT(lov_lock_link_find(env, lck, lsl) == NULL);
120 LASSERT(idx < lck->lls_nr);
122 lck->lls_sub[idx].sub_lock = lsl;
123 lck->lls_nr_filled++;
124 LASSERT(lck->lls_nr_filled <= lck->lls_nr);
125 cfs_list_add_tail(&link->lll_list, &lsl->lss_parents);
127 link->lll_super = lck;
129 lu_ref_add(&parent->cll_reference, "lov-child", sublock);
130 lck->lls_sub[idx].sub_flags |= LSF_HELD;
131 cl_lock_user_add(env, sublock);
133 rc = lov_sublock_modify(env, lck, lsl, &sublock->cll_descr, idx);
134 LASSERT(rc == 0); /* there is no way this can fail, currently */
138 static struct cl_lock *lov_sublock_alloc(const struct lu_env *env,
139 const struct cl_io *io,
140 struct lov_lock *lck,
141 int idx, struct lov_lock_link **out)
143 struct cl_lock *sublock;
144 struct cl_lock *parent;
145 struct lov_lock_link *link;
147 LASSERT(idx < lck->lls_nr);
150 OBD_SLAB_ALLOC_PTR_GFP(link, lov_lock_link_kmem, __GFP_IO);
152 struct lov_sublock_env *subenv;
153 struct lov_lock_sub *lls;
154 struct cl_lock_descr *descr;
156 parent = lck->lls_cl.cls_lock;
157 lls = &lck->lls_sub[idx];
158 descr = &lls->sub_got;
160 subenv = lov_sublock_env_get(env, parent, lls);
161 if (!IS_ERR(subenv)) {
162 /* CAVEAT: Don't try to add a field in lov_lock_sub
163 * to remember the subio. This is because lock is able
164 * to be cached, but this is not true for IO. This
165 * further means a sublock might be referenced in
166 * different io context. -jay */
168 sublock = cl_lock_hold(subenv->lse_env, subenv->lse_io,
169 descr, "lov-parent", parent);
170 lov_sublock_env_put(subenv);
173 sublock = (void*)subenv;
176 if (!IS_ERR(sublock))
179 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
181 sublock = ERR_PTR(-ENOMEM);
185 static void lov_sublock_unlock(const struct lu_env *env,
186 struct lovsub_lock *lsl,
187 struct cl_lock_closure *closure,
188 struct lov_sublock_env *subenv)
191 lov_sublock_env_put(subenv);
192 lsl->lss_active = NULL;
193 cl_lock_disclosure(env, closure);
197 static int lov_sublock_lock(const struct lu_env *env,
198 struct lov_lock *lck,
199 struct lov_lock_sub *lls,
200 struct cl_lock_closure *closure,
201 struct lov_sublock_env **lsep)
203 struct lovsub_lock *sublock;
204 struct cl_lock *child;
208 LASSERT(cfs_list_empty(&closure->clc_list));
210 sublock = lls->sub_lock;
211 child = sublock->lss_cl.cls_lock;
212 result = cl_lock_closure_build(env, child, closure);
214 struct cl_lock *parent = closure->clc_origin;
216 LASSERT(cl_lock_is_mutexed(child));
217 sublock->lss_active = parent;
219 if (unlikely((child->cll_state == CLS_FREEING) ||
220 (child->cll_flags & CLF_CANCELLED))) {
221 struct lov_lock_link *link;
223 * we could race with lock deletion which temporarily
224 * put the lock in freeing state, bug 19080.
226 LASSERT(!(lls->sub_flags & LSF_HELD));
228 link = lov_lock_link_find(env, lck, sublock);
229 LASSERT(link != NULL);
230 lov_lock_unlink(env, link, sublock);
231 lov_sublock_unlock(env, sublock, closure, NULL);
232 lck->lls_cancel_race = 1;
235 struct lov_sublock_env *subenv;
236 subenv = lov_sublock_env_get(env, parent, lls);
237 if (IS_ERR(subenv)) {
238 lov_sublock_unlock(env, sublock,
240 result = PTR_ERR(subenv);
250 * Updates the result of a top-lock operation from a result of sub-lock
251 * sub-operations. Top-operations like lov_lock_{enqueue,use,unuse}() iterate
252 * over sub-locks and lov_subresult() is used to calculate return value of a
253 * top-operation. To this end, possible return values of sub-operations are
257 * - CLO_WAIT wait for event
258 * - CLO_REPEAT repeat top-operation
259 * - -ne fundamental error
261 * Top-level return code can only go down through this list. CLO_REPEAT
262 * overwrites CLO_WAIT, because lock mutex was released and sleeping condition
263 * has to be rechecked by the upper layer.
265 static int lov_subresult(int result, int rc)
272 LASSERTF(result <= 0 || result == CLO_REPEAT || result == CLO_WAIT,
273 "result = %d", result);
274 LASSERTF(rc <= 0 || rc == CLO_REPEAT || rc == CLO_WAIT,
276 CLASSERT(CLO_WAIT < CLO_REPEAT);
278 /* calculate ranks in the ordering above */
279 result_rank = result < 0 ? 1 + CLO_REPEAT : result;
280 rc_rank = rc < 0 ? 1 + CLO_REPEAT : rc;
282 if (result_rank < rc_rank)
288 * Creates sub-locks for a given lov_lock for the first time.
290 * Goes through all sub-objects of top-object, and creates sub-locks on every
291 * sub-object intersecting with top-lock extent. This is complicated by the
292 * fact that top-lock (that is being created) can be accessed concurrently
293 * through already created sub-locks (possibly shared with other top-locks).
295 static int lov_lock_sub_init(const struct lu_env *env,
296 struct lov_lock *lck, const struct cl_io *io)
306 struct lov_object *loo = cl2lov(lck->lls_cl.cls_obj);
307 struct lov_layout_raid0 *r0 = lov_r0(loo);
308 struct cl_lock *parent = lck->lls_cl.cls_lock;
312 lck->lls_orig = parent->cll_descr;
313 file_start = cl_offset(lov2cl(loo), parent->cll_descr.cld_start);
314 file_end = cl_offset(lov2cl(loo), parent->cll_descr.cld_end + 1) - 1;
316 for (i = 0, nr = 0; i < r0->lo_nr; i++) {
318 * XXX for wide striping smarter algorithm is desirable,
319 * breaking out of the loop, early.
321 if (lov_stripe_intersects(loo->lo_lsm, i,
322 file_start, file_end, &start, &end))
326 OBD_ALLOC_LARGE(lck->lls_sub, nr * sizeof lck->lls_sub[0]);
327 if (lck->lls_sub == NULL)
332 * First, fill in sub-lock descriptions in
333 * lck->lls_sub[].sub_descr. They are used by lov_sublock_alloc()
334 * (called below in this function, and by lov_lock_enqueue()) to
335 * create sub-locks. At this moment, no other thread can access
338 for (i = 0, nr = 0; i < r0->lo_nr; ++i) {
339 if (lov_stripe_intersects(loo->lo_lsm, i,
340 file_start, file_end, &start, &end)) {
341 struct cl_lock_descr *descr;
343 descr = &lck->lls_sub[nr].sub_descr;
345 LASSERT(descr->cld_obj == NULL);
346 descr->cld_obj = lovsub2cl(r0->lo_sub[i]);
347 descr->cld_start = cl_index(descr->cld_obj, start);
348 descr->cld_end = cl_index(descr->cld_obj, end);
349 descr->cld_mode = parent->cll_descr.cld_mode;
350 descr->cld_gid = parent->cll_descr.cld_gid;
351 descr->cld_enq_flags = parent->cll_descr.cld_enq_flags;
352 /* XXX has no effect */
353 lck->lls_sub[nr].sub_got = *descr;
354 lck->lls_sub[nr].sub_stripe = i;
358 LASSERT(nr == lck->lls_nr);
360 * Then, create sub-locks. Once at least one sub-lock was created,
361 * top-lock can be reached by other threads.
363 for (i = 0; i < lck->lls_nr; ++i) {
364 struct cl_lock *sublock;
365 struct lov_lock_link *link;
367 if (lck->lls_sub[i].sub_lock == NULL) {
368 sublock = lov_sublock_alloc(env, io, lck, i, &link);
369 if (IS_ERR(sublock)) {
370 result = PTR_ERR(sublock);
373 cl_lock_get_trust(sublock);
374 cl_lock_mutex_get(env, sublock);
375 cl_lock_mutex_get(env, parent);
377 * recheck under mutex that sub-lock wasn't created
378 * concurrently, and that top-lock is still alive.
380 if (lck->lls_sub[i].sub_lock == NULL &&
381 parent->cll_state < CLS_FREEING) {
382 lov_sublock_adopt(env, lck, sublock, i, link);
383 cl_lock_mutex_put(env, parent);
385 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
386 cl_lock_mutex_put(env, parent);
387 cl_lock_unhold(env, sublock,
388 "lov-parent", parent);
390 cl_lock_mutex_put(env, sublock);
391 cl_lock_put(env, sublock);
395 * Some sub-locks can be missing at this point. This is not a problem,
396 * because enqueue will create them anyway. Main duty of this function
397 * is to fill in sub-lock descriptions in a race free manner.
402 static int lov_sublock_release(const struct lu_env *env, struct lov_lock *lck,
403 int i, int deluser, int rc)
405 struct cl_lock *parent = lck->lls_cl.cls_lock;
407 LASSERT(cl_lock_is_mutexed(parent));
410 if (lck->lls_sub[i].sub_flags & LSF_HELD) {
411 struct cl_lock *sublock;
414 LASSERT(lck->lls_sub[i].sub_lock != NULL);
415 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
416 LASSERT(cl_lock_is_mutexed(sublock));
418 lck->lls_sub[i].sub_flags &= ~LSF_HELD;
420 cl_lock_user_del(env, sublock);
422 * If the last hold is released, and cancellation is pending
423 * for a sub-lock, release parent mutex, to avoid keeping it
424 * while sub-lock is being paged out.
426 dying = (sublock->cll_descr.cld_mode == CLM_PHANTOM ||
427 sublock->cll_descr.cld_mode == CLM_GROUP ||
428 (sublock->cll_flags & (CLF_CANCELPEND|CLF_DOOMED))) &&
429 sublock->cll_holds == 1;
431 cl_lock_mutex_put(env, parent);
432 cl_lock_unhold(env, sublock, "lov-parent", parent);
434 cl_lock_mutex_get(env, parent);
435 rc = lov_subresult(rc, CLO_REPEAT);
438 * From now on lck->lls_sub[i].sub_lock is a "weak" pointer,
439 * not backed by a reference on a
440 * sub-lock. lovsub_lock_delete() will clear
441 * lck->lls_sub[i].sub_lock under semaphores, just before
442 * sub-lock is destroyed.
448 static void lov_sublock_hold(const struct lu_env *env, struct lov_lock *lck,
451 struct cl_lock *parent = lck->lls_cl.cls_lock;
453 LASSERT(cl_lock_is_mutexed(parent));
456 if (!(lck->lls_sub[i].sub_flags & LSF_HELD)) {
457 struct cl_lock *sublock;
459 LASSERT(lck->lls_sub[i].sub_lock != NULL);
460 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
461 LASSERT(cl_lock_is_mutexed(sublock));
462 LASSERT(sublock->cll_state != CLS_FREEING);
464 lck->lls_sub[i].sub_flags |= LSF_HELD;
466 cl_lock_get_trust(sublock);
467 cl_lock_hold_add(env, sublock, "lov-parent", parent);
468 cl_lock_user_add(env, sublock);
469 cl_lock_put(env, sublock);
474 static void lov_lock_fini(const struct lu_env *env,
475 struct cl_lock_slice *slice)
477 struct lov_lock *lck;
481 lck = cl2lov_lock(slice);
482 LASSERT(lck->lls_nr_filled == 0);
483 if (lck->lls_sub != NULL) {
484 for (i = 0; i < lck->lls_nr; ++i)
486 * No sub-locks exists at this point, as sub-lock has
487 * a reference on its parent.
489 LASSERT(lck->lls_sub[i].sub_lock == NULL);
490 OBD_FREE_LARGE(lck->lls_sub,
491 lck->lls_nr * sizeof lck->lls_sub[0]);
493 OBD_SLAB_FREE_PTR(lck, lov_lock_kmem);
497 static int lov_lock_enqueue_wait(const struct lu_env *env,
498 struct lov_lock *lck,
499 struct cl_lock *sublock)
501 struct cl_lock *lock = lck->lls_cl.cls_lock;
505 LASSERT(cl_lock_is_mutexed(lock));
507 cl_lock_mutex_put(env, lock);
508 result = cl_lock_enqueue_wait(env, sublock, 0);
509 cl_lock_mutex_get(env, lock);
510 RETURN(result ?: CLO_REPEAT);
514 * Tries to advance a state machine of a given sub-lock toward enqueuing of
517 * \retval 0 if state-transition can proceed
518 * \retval -ve otherwise.
520 static int lov_lock_enqueue_one(const struct lu_env *env, struct lov_lock *lck,
521 struct cl_lock *sublock,
522 struct cl_io *io, __u32 enqflags, int last)
527 /* first, try to enqueue a sub-lock ... */
528 result = cl_enqueue_try(env, sublock, io, enqflags);
529 if ((sublock->cll_state == CLS_ENQUEUED) && !(enqflags & CEF_AGL)) {
530 /* if it is enqueued, try to `wait' on it---maybe it's already
532 result = cl_wait_try(env, sublock);
533 if (result == CLO_REENQUEUED)
537 * If CEF_ASYNC flag is set, then all sub-locks can be enqueued in
538 * parallel, otherwise---enqueue has to wait until sub-lock is granted
539 * before proceeding to the next one.
541 if ((result == CLO_WAIT) && (sublock->cll_state <= CLS_HELD) &&
542 (enqflags & CEF_ASYNC) && (!last || (enqflags & CEF_AGL)))
548 * Helper function for lov_lock_enqueue() that creates missing sub-lock.
550 static int lov_sublock_fill(const struct lu_env *env, struct cl_lock *parent,
551 struct cl_io *io, struct lov_lock *lck, int idx)
553 struct lov_lock_link *link;
554 struct cl_lock *sublock;
557 LASSERT(parent->cll_depth == 1);
558 cl_lock_mutex_put(env, parent);
559 sublock = lov_sublock_alloc(env, io, lck, idx, &link);
560 if (!IS_ERR(sublock))
561 cl_lock_mutex_get(env, sublock);
562 cl_lock_mutex_get(env, parent);
564 if (!IS_ERR(sublock)) {
565 cl_lock_get_trust(sublock);
566 if (parent->cll_state == CLS_QUEUING &&
567 lck->lls_sub[idx].sub_lock == NULL) {
568 lov_sublock_adopt(env, lck, sublock, idx, link);
570 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
571 /* other thread allocated sub-lock, or enqueue is no
573 cl_lock_mutex_put(env, parent);
574 cl_lock_unhold(env, sublock, "lov-parent", parent);
575 cl_lock_mutex_get(env, parent);
577 cl_lock_mutex_put(env, sublock);
578 cl_lock_put(env, sublock);
581 result = PTR_ERR(sublock);
586 * Implementation of cl_lock_operations::clo_enqueue() for lov layer. This
587 * function is rather subtle, as it enqueues top-lock (i.e., advances top-lock
588 * state machine from CLS_QUEUING to CLS_ENQUEUED states) by juggling sub-lock
589 * state machines in the face of sub-locks sharing (by multiple top-locks),
590 * and concurrent sub-lock cancellations.
592 static int lov_lock_enqueue(const struct lu_env *env,
593 const struct cl_lock_slice *slice,
594 struct cl_io *io, __u32 enqflags)
596 struct cl_lock *lock = slice->cls_lock;
597 struct lov_lock *lck = cl2lov_lock(slice);
598 struct cl_lock_closure *closure = lov_closure_get(env, lock);
601 enum cl_lock_state minstate;
605 for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
607 struct lovsub_lock *sub;
608 struct lov_lock_sub *lls;
609 struct cl_lock *sublock;
610 struct lov_sublock_env *subenv;
612 if (lock->cll_state != CLS_QUEUING) {
614 * Lock might have left QUEUING state if previous
615 * iteration released its mutex. Stop enqueing in this
616 * case and let the upper layer to decide what to do.
618 LASSERT(i > 0 && result != 0);
622 lls = &lck->lls_sub[i];
625 * Sub-lock might have been canceled, while top-lock was
629 result = lov_sublock_fill(env, lock, io, lck, i);
630 /* lov_sublock_fill() released @lock mutex,
634 sublock = sub->lss_cl.cls_lock;
635 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
637 lov_sublock_hold(env, lck, i);
638 rc = lov_lock_enqueue_one(subenv->lse_env, lck, sublock,
639 subenv->lse_io, enqflags,
640 i == lck->lls_nr - 1);
641 minstate = min(minstate, sublock->cll_state);
642 if (rc == CLO_WAIT) {
643 switch (sublock->cll_state) {
645 /* take recursive mutex, the lock is
646 * released in lov_lock_enqueue_wait.
648 cl_lock_mutex_get(env, sublock);
649 lov_sublock_unlock(env, sub, closure,
651 rc = lov_lock_enqueue_wait(env, lck,
655 cl_lock_get(sublock);
656 /* take recursive mutex of sublock */
657 cl_lock_mutex_get(env, sublock);
658 /* need to release all locks in closure
659 * otherwise it may deadlock. LU-2683.*/
660 lov_sublock_unlock(env, sub, closure,
662 /* sublock and parent are held. */
663 rc = lov_sublock_release(env, lck, i,
665 cl_lock_mutex_put(env, sublock);
666 cl_lock_put(env, sublock);
669 lov_sublock_unlock(env, sub, closure,
674 LASSERT(sublock->cll_conflict == NULL);
675 lov_sublock_unlock(env, sub, closure, subenv);
678 result = lov_subresult(result, rc);
682 cl_lock_closure_fini(closure);
683 RETURN(result ?: minstate >= CLS_ENQUEUED ? 0 : CLO_WAIT);
686 static int lov_lock_unuse(const struct lu_env *env,
687 const struct cl_lock_slice *slice)
689 struct lov_lock *lck = cl2lov_lock(slice);
690 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
696 for (result = 0, i = 0; i < lck->lls_nr; ++i) {
698 struct lovsub_lock *sub;
699 struct cl_lock *sublock;
700 struct lov_lock_sub *lls;
701 struct lov_sublock_env *subenv;
703 /* top-lock state cannot change concurrently, because single
704 * thread (one that released the last hold) carries unlocking
705 * to the completion. */
706 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
707 lls = &lck->lls_sub[i];
712 sublock = sub->lss_cl.cls_lock;
713 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
715 if (lls->sub_flags & LSF_HELD) {
716 LASSERT(sublock->cll_state == CLS_HELD ||
717 sublock->cll_state == CLS_ENQUEUED);
718 rc = cl_unuse_try(subenv->lse_env, sublock);
719 rc = lov_sublock_release(env, lck, i, 0, rc);
721 lov_sublock_unlock(env, sub, closure, subenv);
723 result = lov_subresult(result, rc);
726 if (result == 0 && lck->lls_cancel_race) {
727 lck->lls_cancel_race = 0;
730 cl_lock_closure_fini(closure);
735 static void lov_lock_cancel(const struct lu_env *env,
736 const struct cl_lock_slice *slice)
738 struct lov_lock *lck = cl2lov_lock(slice);
739 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
745 for (result = 0, i = 0; i < lck->lls_nr; ++i) {
747 struct lovsub_lock *sub;
748 struct cl_lock *sublock;
749 struct lov_lock_sub *lls;
750 struct lov_sublock_env *subenv;
752 /* top-lock state cannot change concurrently, because single
753 * thread (one that released the last hold) carries unlocking
754 * to the completion. */
755 lls = &lck->lls_sub[i];
760 sublock = sub->lss_cl.cls_lock;
761 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
763 if (!(lls->sub_flags & LSF_HELD)) {
764 lov_sublock_unlock(env, sub, closure, subenv);
768 switch(sublock->cll_state) {
770 rc = cl_unuse_try(subenv->lse_env, sublock);
771 lov_sublock_release(env, lck, i, 0, 0);
774 lov_sublock_release(env, lck, i, 1, 0);
777 lov_sublock_unlock(env, sub, closure, subenv);
780 if (rc == CLO_REPEAT) {
785 result = lov_subresult(result, rc);
789 CL_LOCK_DEBUG(D_ERROR, env, slice->cls_lock,
790 "lov_lock_cancel fails with %d.\n", result);
792 cl_lock_closure_fini(closure);
795 static int lov_lock_wait(const struct lu_env *env,
796 const struct cl_lock_slice *slice)
798 struct lov_lock *lck = cl2lov_lock(slice);
799 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
800 enum cl_lock_state minstate;
808 for (result = 0, minstate = CLS_FREEING, i = 0, reenqueued = 0;
809 i < lck->lls_nr; ++i) {
811 struct lovsub_lock *sub;
812 struct cl_lock *sublock;
813 struct lov_lock_sub *lls;
814 struct lov_sublock_env *subenv;
816 lls = &lck->lls_sub[i];
818 LASSERT(sub != NULL);
819 sublock = sub->lss_cl.cls_lock;
820 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
822 LASSERT(sublock->cll_state >= CLS_ENQUEUED);
823 if (sublock->cll_state < CLS_HELD)
824 rc = cl_wait_try(env, sublock);
826 minstate = min(minstate, sublock->cll_state);
827 lov_sublock_unlock(env, sub, closure, subenv);
829 if (rc == CLO_REENQUEUED) {
833 result = lov_subresult(result, rc);
837 /* Each sublock only can be reenqueued once, so will not loop for
839 if (result == 0 && reenqueued != 0)
841 cl_lock_closure_fini(closure);
842 RETURN(result ?: minstate >= CLS_HELD ? 0 : CLO_WAIT);
845 static int lov_lock_use(const struct lu_env *env,
846 const struct cl_lock_slice *slice)
848 struct lov_lock *lck = cl2lov_lock(slice);
849 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
853 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
856 for (result = 0, i = 0; i < lck->lls_nr; ++i) {
858 struct lovsub_lock *sub;
859 struct cl_lock *sublock;
860 struct lov_lock_sub *lls;
861 struct lov_sublock_env *subenv;
863 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
865 lls = &lck->lls_sub[i];
869 * Sub-lock might have been canceled, while top-lock was
876 sublock = sub->lss_cl.cls_lock;
877 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
879 LASSERT(sublock->cll_state != CLS_FREEING);
880 lov_sublock_hold(env, lck, i);
881 if (sublock->cll_state == CLS_CACHED) {
882 rc = cl_use_try(subenv->lse_env, sublock, 0);
884 rc = lov_sublock_release(env, lck,
886 } else if (sublock->cll_state == CLS_NEW) {
887 /* Sub-lock might have been canceled, while
888 * top-lock was cached. */
890 lov_sublock_release(env, lck, i, 1, result);
892 lov_sublock_unlock(env, sub, closure, subenv);
894 result = lov_subresult(result, rc);
899 if (lck->lls_cancel_race) {
901 * If there is unlocking happened at the same time, then
902 * sublock_lock state should be FREEING, and lov_sublock_lock
903 * should return CLO_REPEAT. In this case, it should return
904 * ESTALE, and up layer should reset the lock state to be NEW.
906 lck->lls_cancel_race = 0;
907 LASSERT(result != 0);
910 cl_lock_closure_fini(closure);
915 static int lock_lock_multi_match()
917 struct cl_lock *lock = slice->cls_lock;
918 struct cl_lock_descr *subneed = &lov_env_info(env)->lti_ldescr;
919 struct lov_object *loo = cl2lov(lov->lls_cl.cls_obj);
920 struct lov_layout_raid0 *r0 = lov_r0(loo);
921 struct lov_lock_sub *sub;
922 struct cl_object *subobj;
929 fstart = cl_offset(need->cld_obj, need->cld_start);
930 fend = cl_offset(need->cld_obj, need->cld_end + 1) - 1;
931 subneed->cld_mode = need->cld_mode;
932 cl_lock_mutex_get(env, lock);
933 for (i = 0; i < lov->lls_nr; ++i) {
934 sub = &lov->lls_sub[i];
935 if (sub->sub_lock == NULL)
937 subobj = sub->sub_descr.cld_obj;
938 if (!lov_stripe_intersects(loo->lo_lsm, sub->sub_stripe,
939 fstart, fend, &start, &end))
941 subneed->cld_start = cl_index(subobj, start);
942 subneed->cld_end = cl_index(subobj, end);
943 subneed->cld_obj = subobj;
944 if (!cl_lock_ext_match(&sub->sub_got, subneed)) {
949 cl_lock_mutex_put(env, lock);
954 * Check if the extent region \a descr is covered by \a child against the
955 * specific \a stripe.
957 static int lov_lock_stripe_is_matching(const struct lu_env *env,
958 struct lov_object *lov, int stripe,
959 const struct cl_lock_descr *child,
960 const struct cl_lock_descr *descr)
962 struct lov_stripe_md *lsm = lov->lo_lsm;
967 if (lov_r0(lov)->lo_nr == 1)
968 return cl_lock_ext_match(child, descr);
971 * For a multi-stripes object:
972 * - make sure the descr only covers child's stripe, and
973 * - check if extent is matching.
975 start = cl_offset(&lov->lo_cl, descr->cld_start);
976 end = cl_offset(&lov->lo_cl, descr->cld_end + 1) - 1;
977 result = end - start <= lsm->lsm_stripe_size &&
978 stripe == lov_stripe_number(lsm, start) &&
979 stripe == lov_stripe_number(lsm, end);
981 struct cl_lock_descr *subd = &lov_env_info(env)->lti_ldescr;
985 subd->cld_obj = NULL; /* don't need sub object at all */
986 subd->cld_mode = descr->cld_mode;
987 subd->cld_gid = descr->cld_gid;
988 result = lov_stripe_intersects(lsm, stripe, start, end,
989 &sub_start, &sub_end);
991 subd->cld_start = cl_index(child->cld_obj, sub_start);
992 subd->cld_end = cl_index(child->cld_obj, sub_end);
993 result = cl_lock_ext_match(child, subd);
999 * An implementation of cl_lock_operations::clo_fits_into() method.
1001 * Checks whether a lock (given by \a slice) is suitable for \a
1002 * io. Multi-stripe locks can be used only for "quick" io, like truncate, or
1005 * \see ccc_lock_fits_into().
1007 static int lov_lock_fits_into(const struct lu_env *env,
1008 const struct cl_lock_slice *slice,
1009 const struct cl_lock_descr *need,
1010 const struct cl_io *io)
1012 struct lov_lock *lov = cl2lov_lock(slice);
1013 struct lov_object *obj = cl2lov(slice->cls_obj);
1016 LASSERT(cl_object_same(need->cld_obj, slice->cls_obj));
1017 LASSERT(lov->lls_nr > 0);
1021 /* for top lock, it's necessary to match enq flags otherwise it will
1022 * run into problem if a sublock is missing and reenqueue. */
1023 if (need->cld_enq_flags != lov->lls_orig.cld_enq_flags)
1026 if (need->cld_mode == CLM_GROUP)
1028 * always allow to match group lock.
1030 result = cl_lock_ext_match(&lov->lls_orig, need);
1031 else if (lov->lls_nr == 1) {
1032 struct cl_lock_descr *got = &lov->lls_sub[0].sub_got;
1033 result = lov_lock_stripe_is_matching(env,
1034 cl2lov(slice->cls_obj),
1035 lov->lls_sub[0].sub_stripe,
1037 } else if (io->ci_type != CIT_SETATTR && io->ci_type != CIT_MISC &&
1038 !cl_io_is_append(io) && need->cld_mode != CLM_PHANTOM)
1040 * Multi-stripe locks are only suitable for `quick' IO and for
1046 * Most general case: multi-stripe existing lock, and
1047 * (potentially) multi-stripe @need lock. Check that @need is
1048 * covered by @lov's sub-locks.
1050 * For now, ignore lock expansions made by the server, and
1051 * match against original lock extent.
1053 result = cl_lock_ext_match(&lov->lls_orig, need);
1054 CDEBUG(D_DLMTRACE, DDESCR"/"DDESCR" %d %d/%d: %d\n",
1055 PDESCR(&lov->lls_orig), PDESCR(&lov->lls_sub[0].sub_got),
1056 lov->lls_sub[0].sub_stripe, lov->lls_nr, lov_r0(obj)->lo_nr,
1061 void lov_lock_unlink(const struct lu_env *env,
1062 struct lov_lock_link *link, struct lovsub_lock *sub)
1064 struct lov_lock *lck = link->lll_super;
1065 struct cl_lock *parent = lck->lls_cl.cls_lock;
1067 LASSERT(cl_lock_is_mutexed(parent));
1068 LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
1071 cfs_list_del_init(&link->lll_list);
1072 LASSERT(lck->lls_sub[link->lll_idx].sub_lock == sub);
1073 /* yank this sub-lock from parent's array */
1074 lck->lls_sub[link->lll_idx].sub_lock = NULL;
1075 LASSERT(lck->lls_nr_filled > 0);
1076 lck->lls_nr_filled--;
1077 lu_ref_del(&parent->cll_reference, "lov-child", sub->lss_cl.cls_lock);
1078 cl_lock_put(env, parent);
1079 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
1083 struct lov_lock_link *lov_lock_link_find(const struct lu_env *env,
1084 struct lov_lock *lck,
1085 struct lovsub_lock *sub)
1087 struct lov_lock_link *scan;
1089 LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
1092 cfs_list_for_each_entry(scan, &sub->lss_parents, lll_list) {
1093 if (scan->lll_super == lck)
1100 * An implementation of cl_lock_operations::clo_delete() method. This is
1101 * invoked for "top-to-bottom" delete, when lock destruction starts from the
1102 * top-lock, e.g., as a result of inode destruction.
1104 * Unlinks top-lock from all its sub-locks. Sub-locks are not deleted there:
1105 * this is done separately elsewhere:
1107 * - for inode destruction, lov_object_delete() calls cl_object_kill() for
1108 * each sub-object, purging its locks;
1110 * - in other cases (e.g., a fatal error with a top-lock) sub-locks are
1111 * left in the cache.
1113 static void lov_lock_delete(const struct lu_env *env,
1114 const struct cl_lock_slice *slice)
1116 struct lov_lock *lck = cl2lov_lock(slice);
1117 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
1118 struct lov_lock_link *link;
1122 LASSERT(slice->cls_lock->cll_state == CLS_FREEING);
1125 for (i = 0; i < lck->lls_nr; ++i) {
1126 struct lov_lock_sub *lls = &lck->lls_sub[i];
1127 struct lovsub_lock *lsl = lls->sub_lock;
1129 if (lsl == NULL) /* already removed */
1132 rc = lov_sublock_lock(env, lck, lls, closure, NULL);
1133 if (rc == CLO_REPEAT) {
1139 LASSERT(lsl->lss_cl.cls_lock->cll_state < CLS_FREEING);
1141 if (lls->sub_flags & LSF_HELD)
1142 lov_sublock_release(env, lck, i, 1, 0);
1144 link = lov_lock_link_find(env, lck, lsl);
1145 LASSERT(link != NULL);
1146 lov_lock_unlink(env, link, lsl);
1147 LASSERT(lck->lls_sub[i].sub_lock == NULL);
1149 lov_sublock_unlock(env, lsl, closure, NULL);
1152 cl_lock_closure_fini(closure);
1156 static int lov_lock_print(const struct lu_env *env, void *cookie,
1157 lu_printer_t p, const struct cl_lock_slice *slice)
1159 struct lov_lock *lck = cl2lov_lock(slice);
1162 (*p)(env, cookie, "%d\n", lck->lls_nr);
1163 for (i = 0; i < lck->lls_nr; ++i) {
1164 struct lov_lock_sub *sub;
1166 sub = &lck->lls_sub[i];
1167 (*p)(env, cookie, " %d %x: ", i, sub->sub_flags);
1168 if (sub->sub_lock != NULL)
1169 cl_lock_print(env, cookie, p,
1170 sub->sub_lock->lss_cl.cls_lock);
1172 (*p)(env, cookie, "---\n");
1177 static const struct cl_lock_operations lov_lock_ops = {
1178 .clo_fini = lov_lock_fini,
1179 .clo_enqueue = lov_lock_enqueue,
1180 .clo_wait = lov_lock_wait,
1181 .clo_use = lov_lock_use,
1182 .clo_unuse = lov_lock_unuse,
1183 .clo_cancel = lov_lock_cancel,
1184 .clo_fits_into = lov_lock_fits_into,
1185 .clo_delete = lov_lock_delete,
1186 .clo_print = lov_lock_print
1189 int lov_lock_init_raid0(const struct lu_env *env, struct cl_object *obj,
1190 struct cl_lock *lock, const struct cl_io *io)
1192 struct lov_lock *lck;
1196 OBD_SLAB_ALLOC_PTR_GFP(lck, lov_lock_kmem, __GFP_IO);
1198 cl_lock_slice_add(lock, &lck->lls_cl, obj, &lov_lock_ops);
1199 result = lov_lock_sub_init(env, lck, io);
1205 static void lov_empty_lock_fini(const struct lu_env *env,
1206 struct cl_lock_slice *slice)
1208 struct lov_lock *lck = cl2lov_lock(slice);
1209 OBD_SLAB_FREE_PTR(lck, lov_lock_kmem);
1212 static int lov_empty_lock_print(const struct lu_env *env, void *cookie,
1213 lu_printer_t p, const struct cl_lock_slice *slice)
1215 (*p)(env, cookie, "empty\n");
1219 /* XXX: more methods will be added later. */
1220 static const struct cl_lock_operations lov_empty_lock_ops = {
1221 .clo_fini = lov_empty_lock_fini,
1222 .clo_print = lov_empty_lock_print
1225 int lov_lock_init_empty(const struct lu_env *env, struct cl_object *obj,
1226 struct cl_lock *lock, const struct cl_io *io)
1228 struct lov_lock *lck;
1229 int result = -ENOMEM;
1232 OBD_SLAB_ALLOC_PTR_GFP(lck, lov_lock_kmem, __GFP_IO);
1234 cl_lock_slice_add(lock, &lck->lls_cl, obj, &lov_empty_lock_ops);
1235 lck->lls_orig = lock->cll_descr;
1241 static struct cl_lock_closure *lov_closure_get(const struct lu_env *env,
1242 struct cl_lock *parent)
1244 struct cl_lock_closure *closure;
1246 closure = &lov_env_info(env)->lti_closure;
1247 LASSERT(cfs_list_empty(&closure->clc_list));
1248 cl_lock_closure_init(env, closure, parent, 1);