1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
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, CFS_ALLOC_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_descr;
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)
270 LASSERT(result <= 0 || result == CLO_REPEAT || result == CLO_WAIT);
271 LASSERT(rc <= 0 || rc == CLO_REPEAT || rc == CLO_WAIT);
272 CLASSERT(CLO_WAIT < CLO_REPEAT);
276 /* calculate ranks in the ordering above */
277 result_rank = result < 0 ? 1 + CLO_REPEAT : result;
278 rc_rank = rc < 0 ? 1 + CLO_REPEAT : rc;
280 if (result_rank < rc_rank)
286 * Creates sub-locks for a given lov_lock for the first time.
288 * Goes through all sub-objects of top-object, and creates sub-locks on every
289 * sub-object intersecting with top-lock extent. This is complicated by the
290 * fact that top-lock (that is being created) can be accessed concurrently
291 * through already created sub-locks (possibly shared with other top-locks).
293 static int lov_lock_sub_init(const struct lu_env *env,
294 struct lov_lock *lck, const struct cl_io *io)
304 struct lov_object *loo = cl2lov(lck->lls_cl.cls_obj);
305 struct lov_layout_raid0 *r0 = lov_r0(loo);
306 struct cl_lock *parent = lck->lls_cl.cls_lock;
310 lck->lls_orig = parent->cll_descr;
311 file_start = cl_offset(lov2cl(loo), parent->cll_descr.cld_start);
312 file_end = cl_offset(lov2cl(loo), parent->cll_descr.cld_end + 1) - 1;
314 for (i = 0, nr = 0; i < r0->lo_nr; i++) {
316 * XXX for wide striping smarter algorithm is desirable,
317 * breaking out of the loop, early.
319 if (lov_stripe_intersects(r0->lo_lsm, i,
320 file_start, file_end, &start, &end))
324 OBD_ALLOC(lck->lls_sub, nr * sizeof lck->lls_sub[0]);
325 if (lck->lls_sub == NULL)
330 * First, fill in sub-lock descriptions in
331 * lck->lls_sub[].sub_descr. They are used by lov_sublock_alloc()
332 * (called below in this function, and by lov_lock_enqueue()) to
333 * create sub-locks. At this moment, no other thread can access
336 for (i = 0, nr = 0; i < r0->lo_nr; ++i) {
337 if (lov_stripe_intersects(r0->lo_lsm, i,
338 file_start, file_end, &start, &end)) {
339 struct cl_lock_descr *descr;
341 descr = &lck->lls_sub[nr].sub_descr;
343 LASSERT(descr->cld_obj == NULL);
344 descr->cld_obj = lovsub2cl(r0->lo_sub[i]);
345 descr->cld_start = cl_index(descr->cld_obj, start);
346 descr->cld_end = cl_index(descr->cld_obj, end);
347 descr->cld_mode = parent->cll_descr.cld_mode;
348 descr->cld_gid = parent->cll_descr.cld_gid;
349 descr->cld_enq_flags = parent->cll_descr.cld_enq_flags;
350 /* XXX has no effect */
351 lck->lls_sub[nr].sub_got = *descr;
352 lck->lls_sub[nr].sub_stripe = i;
356 LASSERT(nr == lck->lls_nr);
358 * Then, create sub-locks. Once at least one sub-lock was created,
359 * top-lock can be reached by other threads.
361 for (i = 0; i < lck->lls_nr; ++i) {
362 struct cl_lock *sublock;
363 struct lov_lock_link *link;
365 if (lck->lls_sub[i].sub_lock == NULL) {
366 sublock = lov_sublock_alloc(env, io, lck, i, &link);
367 if (IS_ERR(sublock)) {
368 result = PTR_ERR(sublock);
371 cl_lock_get_trust(sublock);
372 cl_lock_mutex_get(env, sublock);
373 cl_lock_mutex_get(env, parent);
375 * recheck under mutex that sub-lock wasn't created
376 * concurrently, and that top-lock is still alive.
378 if (lck->lls_sub[i].sub_lock == NULL &&
379 parent->cll_state < CLS_FREEING) {
380 lov_sublock_adopt(env, lck, sublock, i, link);
381 cl_lock_mutex_put(env, parent);
383 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
384 cl_lock_mutex_put(env, parent);
385 cl_lock_unhold(env, sublock,
386 "lov-parent", parent);
388 cl_lock_mutex_put(env, sublock);
389 cl_lock_put(env, sublock);
393 * Some sub-locks can be missing at this point. This is not a problem,
394 * because enqueue will create them anyway. Main duty of this function
395 * is to fill in sub-lock descriptions in a race free manner.
400 static int lov_sublock_release(const struct lu_env *env, struct lov_lock *lck,
401 int i, int deluser, int rc)
403 struct cl_lock *parent = lck->lls_cl.cls_lock;
405 LASSERT(cl_lock_is_mutexed(parent));
408 if (lck->lls_sub[i].sub_flags & LSF_HELD) {
409 struct cl_lock *sublock;
412 LASSERT(lck->lls_sub[i].sub_lock != NULL);
413 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
414 LASSERT(cl_lock_is_mutexed(sublock));
416 lck->lls_sub[i].sub_flags &= ~LSF_HELD;
418 cl_lock_user_del(env, sublock);
420 * If the last hold is released, and cancellation is pending
421 * for a sub-lock, release parent mutex, to avoid keeping it
422 * while sub-lock is being paged out.
424 dying = (sublock->cll_descr.cld_mode == CLM_PHANTOM ||
425 sublock->cll_descr.cld_mode == CLM_GROUP ||
426 (sublock->cll_flags & (CLF_CANCELPEND|CLF_DOOMED))) &&
427 sublock->cll_holds == 1;
429 cl_lock_mutex_put(env, parent);
430 cl_lock_unhold(env, sublock, "lov-parent", parent);
432 cl_lock_mutex_get(env, parent);
433 rc = lov_subresult(rc, CLO_REPEAT);
436 * From now on lck->lls_sub[i].sub_lock is a "weak" pointer,
437 * not backed by a reference on a
438 * sub-lock. lovsub_lock_delete() will clear
439 * lck->lls_sub[i].sub_lock under semaphores, just before
440 * sub-lock is destroyed.
446 static void lov_sublock_hold(const struct lu_env *env, struct lov_lock *lck,
449 struct cl_lock *parent = lck->lls_cl.cls_lock;
451 LASSERT(cl_lock_is_mutexed(parent));
454 if (!(lck->lls_sub[i].sub_flags & LSF_HELD)) {
455 struct cl_lock *sublock;
457 LASSERT(lck->lls_sub[i].sub_lock != NULL);
458 sublock = lck->lls_sub[i].sub_lock->lss_cl.cls_lock;
459 LASSERT(cl_lock_is_mutexed(sublock));
460 LASSERT(sublock->cll_state != CLS_FREEING);
462 lck->lls_sub[i].sub_flags |= LSF_HELD;
464 cl_lock_get_trust(sublock);
465 cl_lock_hold_add(env, sublock, "lov-parent", parent);
466 cl_lock_user_add(env, sublock);
467 cl_lock_put(env, sublock);
472 static void lov_lock_fini(const struct lu_env *env,
473 struct cl_lock_slice *slice)
475 struct lov_lock *lck;
479 lck = cl2lov_lock(slice);
480 LASSERT(lck->lls_nr_filled == 0);
481 if (lck->lls_sub != NULL) {
482 for (i = 0; i < lck->lls_nr; ++i)
484 * No sub-locks exists at this point, as sub-lock has
485 * a reference on its parent.
487 LASSERT(lck->lls_sub[i].sub_lock == NULL);
488 OBD_FREE(lck->lls_sub, lck->lls_nr * sizeof lck->lls_sub[0]);
490 OBD_SLAB_FREE_PTR(lck, lov_lock_kmem);
494 static int lov_lock_enqueue_wait(const struct lu_env *env,
495 struct lov_lock *lck,
496 struct cl_lock *sublock)
498 struct cl_lock *lock = lck->lls_cl.cls_lock;
502 LASSERT(cl_lock_is_mutexed(lock));
504 cl_lock_mutex_put(env, lock);
505 result = cl_lock_enqueue_wait(env, sublock, 0);
506 cl_lock_mutex_get(env, lock);
507 RETURN(result ?: CLO_REPEAT);
511 * Tries to advance a state machine of a given sub-lock toward enqueuing of
514 * \retval 0 if state-transition can proceed
515 * \retval -ve otherwise.
517 static int lov_lock_enqueue_one(const struct lu_env *env, struct lov_lock *lck,
518 struct cl_lock *sublock,
519 struct cl_io *io, __u32 enqflags, int last)
524 /* first, try to enqueue a sub-lock ... */
525 result = cl_enqueue_try(env, sublock, io, enqflags);
526 if (sublock->cll_state == CLS_ENQUEUED)
527 /* if it is enqueued, try to `wait' on it---maybe it's already
529 result = cl_wait_try(env, sublock);
531 * If CEF_ASYNC flag is set, then all sub-locks can be enqueued in
532 * parallel, otherwise---enqueue has to wait until sub-lock is granted
533 * before proceeding to the next one.
535 if (result == CLO_WAIT && sublock->cll_state <= CLS_HELD &&
536 enqflags & CEF_ASYNC && !last)
542 * Helper function for lov_lock_enqueue() that creates missing sub-lock.
544 static int lov_sublock_fill(const struct lu_env *env, struct cl_lock *parent,
545 struct cl_io *io, struct lov_lock *lck, int idx)
547 struct lov_lock_link *link;
548 struct cl_lock *sublock;
551 LASSERT(parent->cll_depth == 1);
552 cl_lock_mutex_put(env, parent);
553 sublock = lov_sublock_alloc(env, io, lck, idx, &link);
554 if (!IS_ERR(sublock))
555 cl_lock_mutex_get(env, sublock);
556 cl_lock_mutex_get(env, parent);
558 if (!IS_ERR(sublock)) {
559 cl_lock_get_trust(sublock);
560 if (parent->cll_state == CLS_QUEUING &&
561 lck->lls_sub[idx].sub_lock == NULL) {
562 lov_sublock_adopt(env, lck, sublock, idx, link);
564 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
565 /* other thread allocated sub-lock, or enqueue is no
567 cl_lock_mutex_put(env, parent);
568 cl_lock_unhold(env, sublock, "lov-parent", parent);
569 cl_lock_mutex_get(env, parent);
571 cl_lock_mutex_put(env, sublock);
572 cl_lock_put(env, sublock);
575 result = PTR_ERR(sublock);
580 * Implementation of cl_lock_operations::clo_enqueue() for lov layer. This
581 * function is rather subtle, as it enqueues top-lock (i.e., advances top-lock
582 * state machine from CLS_QUEUING to CLS_ENQUEUED states) by juggling sub-lock
583 * state machines in the face of sub-locks sharing (by multiple top-locks),
584 * and concurrent sub-lock cancellations.
586 static int lov_lock_enqueue(const struct lu_env *env,
587 const struct cl_lock_slice *slice,
588 struct cl_io *io, __u32 enqflags)
590 struct cl_lock *lock = slice->cls_lock;
591 struct lov_lock *lck = cl2lov_lock(slice);
592 struct cl_lock_closure *closure = lov_closure_get(env, lock);
595 enum cl_lock_state minstate;
599 for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
601 struct lovsub_lock *sub;
602 struct lov_lock_sub *lls;
603 struct cl_lock *sublock;
604 struct lov_sublock_env *subenv;
606 if (lock->cll_state != CLS_QUEUING) {
608 * Lock might have left QUEUING state if previous
609 * iteration released its mutex. Stop enqueing in this
610 * case and let the upper layer to decide what to do.
612 LASSERT(i > 0 && result != 0);
616 lls = &lck->lls_sub[i];
619 * Sub-lock might have been canceled, while top-lock was
623 result = lov_sublock_fill(env, lock, io, lck, i);
624 /* lov_sublock_fill() released @lock mutex,
628 sublock = sub->lss_cl.cls_lock;
629 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
631 lov_sublock_hold(env, lck, i);
632 rc = lov_lock_enqueue_one(subenv->lse_env, lck, sublock,
633 subenv->lse_io, enqflags,
634 i == lck->lls_nr - 1);
635 minstate = min(minstate, sublock->cll_state);
636 if (rc == CLO_WAIT) {
637 switch (sublock->cll_state) {
639 /* take recursive mutex, the lock is
640 * released in lov_lock_enqueue_wait.
642 cl_lock_mutex_get(env, sublock);
643 lov_sublock_unlock(env, sub, closure,
645 rc = lov_lock_enqueue_wait(env, lck,
649 rc = lov_sublock_release(env, lck, i,
652 lov_sublock_unlock(env, sub, closure,
657 LASSERT(sublock->cll_conflict == NULL);
658 lov_sublock_unlock(env, sub, closure, subenv);
661 result = lov_subresult(result, rc);
665 cl_lock_closure_fini(closure);
666 RETURN(result ?: minstate >= CLS_ENQUEUED ? 0 : CLO_WAIT);
669 static int lov_lock_unuse(const struct lu_env *env,
670 const struct cl_lock_slice *slice)
672 struct lov_lock *lck = cl2lov_lock(slice);
673 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
679 for (result = 0, i = 0; i < lck->lls_nr; ++i) {
681 struct lovsub_lock *sub;
682 struct cl_lock *sublock;
683 struct lov_lock_sub *lls;
684 struct lov_sublock_env *subenv;
686 /* top-lock state cannot change concurrently, because single
687 * thread (one that released the last hold) carries unlocking
688 * to the completion. */
689 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
690 lls = &lck->lls_sub[i];
695 sublock = sub->lss_cl.cls_lock;
696 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
698 if (lls->sub_flags & LSF_HELD) {
699 LASSERT(sublock->cll_state == CLS_HELD);
700 rc = cl_unuse_try(subenv->lse_env, sublock);
701 rc = lov_sublock_release(env, lck, i, 0, rc);
703 lov_sublock_unlock(env, sub, closure, subenv);
705 result = lov_subresult(result, rc);
708 if (result == 0 && lck->lls_cancel_race) {
709 lck->lls_cancel_race = 0;
712 cl_lock_closure_fini(closure);
717 static void lov_lock_cancel(const struct lu_env *env,
718 const struct cl_lock_slice *slice)
720 struct lov_lock *lck = cl2lov_lock(slice);
721 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
727 for (result = 0, i = 0; i < lck->lls_nr; ++i) {
729 struct lovsub_lock *sub;
730 struct cl_lock *sublock;
731 struct lov_lock_sub *lls;
732 struct lov_sublock_env *subenv;
734 /* top-lock state cannot change concurrently, because single
735 * thread (one that released the last hold) carries unlocking
736 * to the completion. */
737 lls = &lck->lls_sub[i];
742 sublock = sub->lss_cl.cls_lock;
743 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
745 if (!(lls->sub_flags & LSF_HELD)) {
746 lov_sublock_unlock(env, sub, closure, subenv);
750 switch(sublock->cll_state) {
752 rc = cl_unuse_try(subenv->lse_env,
754 lov_sublock_release(env, lck, i, 0, 0);
757 /* TODO: it's not a good idea to cancel this
758 * lock because it's innocent. But it's
759 * acceptable. The better way would be to
760 * define a new lock method to unhold the
762 cl_lock_cancel(env, sublock);
764 lov_sublock_release(env, lck, i, 1, 0);
767 lov_sublock_unlock(env, sub, closure, subenv);
770 if (rc == CLO_REPEAT) {
775 result = lov_subresult(result, rc);
779 CL_LOCK_DEBUG(D_ERROR, env, slice->cls_lock,
780 "lov_lock_cancel fails with %d.\n", result);
782 cl_lock_closure_fini(closure);
785 static int lov_lock_wait(const struct lu_env *env,
786 const struct cl_lock_slice *slice)
788 struct lov_lock *lck = cl2lov_lock(slice);
789 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
790 enum cl_lock_state minstate;
796 for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
798 struct lovsub_lock *sub;
799 struct cl_lock *sublock;
800 struct lov_lock_sub *lls;
801 struct lov_sublock_env *subenv;
803 lls = &lck->lls_sub[i];
805 LASSERT(sub != NULL);
806 sublock = sub->lss_cl.cls_lock;
807 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
809 LASSERT(sublock->cll_state >= CLS_ENQUEUED);
810 if (sublock->cll_state < CLS_HELD)
811 rc = cl_wait_try(env, sublock);
813 minstate = min(minstate, sublock->cll_state);
814 lov_sublock_unlock(env, sub, closure, subenv);
816 result = lov_subresult(result, rc);
820 cl_lock_closure_fini(closure);
821 RETURN(result ?: minstate >= CLS_HELD ? 0 : CLO_WAIT);
824 static int lov_lock_use(const struct lu_env *env,
825 const struct cl_lock_slice *slice)
827 struct lov_lock *lck = cl2lov_lock(slice);
828 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
832 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
835 for (result = 0, i = 0; i < lck->lls_nr; ++i) {
837 struct lovsub_lock *sub;
838 struct cl_lock *sublock;
839 struct lov_lock_sub *lls;
840 struct lov_sublock_env *subenv;
842 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
844 lls = &lck->lls_sub[i];
848 * Sub-lock might have been canceled, while top-lock was
855 sublock = sub->lss_cl.cls_lock;
856 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
858 LASSERT(sublock->cll_state != CLS_FREEING);
859 lov_sublock_hold(env, lck, i);
860 if (sublock->cll_state == CLS_CACHED) {
861 rc = cl_use_try(subenv->lse_env, sublock, 0);
863 rc = lov_sublock_release(env, lck,
866 lov_sublock_unlock(env, sub, closure, subenv);
868 result = lov_subresult(result, rc);
873 if (lck->lls_cancel_race) {
875 * If there is unlocking happened at the same time, then
876 * sublock_lock state should be FREEING, and lov_sublock_lock
877 * should return CLO_REPEAT. In this case, it should return
878 * ESTALE, and up layer should reset the lock state to be NEW.
880 lck->lls_cancel_race = 0;
881 LASSERT(result != 0);
884 cl_lock_closure_fini(closure);
889 static int lock_lock_multi_match()
891 struct cl_lock *lock = slice->cls_lock;
892 struct cl_lock_descr *subneed = &lov_env_info(env)->lti_ldescr;
893 struct lov_object *loo = cl2lov(lov->lls_cl.cls_obj);
894 struct lov_layout_raid0 *r0 = lov_r0(loo);
895 struct lov_lock_sub *sub;
896 struct cl_object *subobj;
903 fstart = cl_offset(need->cld_obj, need->cld_start);
904 fend = cl_offset(need->cld_obj, need->cld_end + 1) - 1;
905 subneed->cld_mode = need->cld_mode;
906 cl_lock_mutex_get(env, lock);
907 for (i = 0; i < lov->lls_nr; ++i) {
908 sub = &lov->lls_sub[i];
909 if (sub->sub_lock == NULL)
911 subobj = sub->sub_descr.cld_obj;
912 if (!lov_stripe_intersects(r0->lo_lsm, sub->sub_stripe,
913 fstart, fend, &start, &end))
915 subneed->cld_start = cl_index(subobj, start);
916 subneed->cld_end = cl_index(subobj, end);
917 subneed->cld_obj = subobj;
918 if (!cl_lock_ext_match(&sub->sub_got, subneed)) {
923 cl_lock_mutex_put(env, lock);
928 * Check if the extent region \a descr is covered by \a child against the
929 * specific \a stripe.
931 static int lov_lock_stripe_is_matching(const struct lu_env *env,
932 struct lov_object *lov, int stripe,
933 const struct cl_lock_descr *child,
934 const struct cl_lock_descr *descr)
936 struct lov_stripe_md *lsm = lov_r0(lov)->lo_lsm;
941 if (lov_r0(lov)->lo_nr == 1)
942 return cl_lock_ext_match(child, descr);
945 * For a multi-stripes object:
946 * - make sure the descr only covers child's stripe, and
947 * - check if extent is matching.
949 start = cl_offset(&lov->lo_cl, descr->cld_start);
950 end = cl_offset(&lov->lo_cl, descr->cld_end + 1) - 1;
951 result = end - start <= lsm->lsm_stripe_size &&
952 stripe == lov_stripe_number(lsm, start) &&
953 stripe == lov_stripe_number(lsm, end);
955 struct cl_lock_descr *subd = &lov_env_info(env)->lti_ldescr;
959 subd->cld_obj = NULL; /* don't need sub object at all */
960 subd->cld_mode = descr->cld_mode;
961 subd->cld_gid = descr->cld_gid;
962 result = lov_stripe_intersects(lsm, stripe, start, end,
963 &sub_start, &sub_end);
965 subd->cld_start = cl_index(child->cld_obj, sub_start);
966 subd->cld_end = cl_index(child->cld_obj, sub_end);
967 result = cl_lock_ext_match(child, subd);
973 * An implementation of cl_lock_operations::clo_fits_into() method.
975 * Checks whether a lock (given by \a slice) is suitable for \a
976 * io. Multi-stripe locks can be used only for "quick" io, like truncate, or
979 * \see ccc_lock_fits_into().
981 static int lov_lock_fits_into(const struct lu_env *env,
982 const struct cl_lock_slice *slice,
983 const struct cl_lock_descr *need,
984 const struct cl_io *io)
986 struct lov_lock *lov = cl2lov_lock(slice);
987 struct lov_object *obj = cl2lov(slice->cls_obj);
990 LASSERT(cl_object_same(need->cld_obj, slice->cls_obj));
991 LASSERT(lov->lls_nr > 0);
995 if (need->cld_mode == CLM_GROUP)
997 * always allow to match group lock.
999 result = cl_lock_ext_match(&lov->lls_orig, need);
1000 else if (lov->lls_nr == 1) {
1001 struct cl_lock_descr *got = &lov->lls_sub[0].sub_got;
1002 result = lov_lock_stripe_is_matching(env,
1003 cl2lov(slice->cls_obj),
1004 lov->lls_sub[0].sub_stripe,
1006 } else if (io->ci_type != CIT_SETATTR && io->ci_type != CIT_MISC &&
1007 !cl_io_is_append(io) && need->cld_mode != CLM_PHANTOM)
1009 * Multi-stripe locks are only suitable for `quick' IO and for
1015 * Most general case: multi-stripe existing lock, and
1016 * (potentially) multi-stripe @need lock. Check that @need is
1017 * covered by @lov's sub-locks.
1019 * For now, ignore lock expansions made by the server, and
1020 * match against original lock extent.
1022 result = cl_lock_ext_match(&lov->lls_orig, need);
1023 CDEBUG(D_DLMTRACE, DDESCR"/"DDESCR" %d %d/%d: %d\n",
1024 PDESCR(&lov->lls_orig), PDESCR(&lov->lls_sub[0].sub_got),
1025 lov->lls_sub[0].sub_stripe, lov->lls_nr, lov_r0(obj)->lo_nr,
1030 void lov_lock_unlink(const struct lu_env *env,
1031 struct lov_lock_link *link, struct lovsub_lock *sub)
1033 struct lov_lock *lck = link->lll_super;
1034 struct cl_lock *parent = lck->lls_cl.cls_lock;
1036 LASSERT(cl_lock_is_mutexed(parent));
1037 LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
1040 cfs_list_del_init(&link->lll_list);
1041 LASSERT(lck->lls_sub[link->lll_idx].sub_lock == sub);
1042 /* yank this sub-lock from parent's array */
1043 lck->lls_sub[link->lll_idx].sub_lock = NULL;
1044 LASSERT(lck->lls_nr_filled > 0);
1045 lck->lls_nr_filled--;
1046 lu_ref_del(&parent->cll_reference, "lov-child", sub->lss_cl.cls_lock);
1047 cl_lock_put(env, parent);
1048 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
1052 struct lov_lock_link *lov_lock_link_find(const struct lu_env *env,
1053 struct lov_lock *lck,
1054 struct lovsub_lock *sub)
1056 struct lov_lock_link *scan;
1058 LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
1061 cfs_list_for_each_entry(scan, &sub->lss_parents, lll_list) {
1062 if (scan->lll_super == lck)
1069 * An implementation of cl_lock_operations::clo_delete() method. This is
1070 * invoked for "top-to-bottom" delete, when lock destruction starts from the
1071 * top-lock, e.g., as a result of inode destruction.
1073 * Unlinks top-lock from all its sub-locks. Sub-locks are not deleted there:
1074 * this is done separately elsewhere:
1076 * - for inode destruction, lov_object_delete() calls cl_object_kill() for
1077 * each sub-object, purging its locks;
1079 * - in other cases (e.g., a fatal error with a top-lock) sub-locks are
1080 * left in the cache.
1082 static void lov_lock_delete(const struct lu_env *env,
1083 const struct cl_lock_slice *slice)
1085 struct lov_lock *lck = cl2lov_lock(slice);
1086 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
1087 struct lov_lock_link *link;
1091 LASSERT(slice->cls_lock->cll_state == CLS_FREEING);
1094 for (i = 0; i < lck->lls_nr; ++i) {
1095 struct lov_lock_sub *lls = &lck->lls_sub[i];
1096 struct lovsub_lock *lsl = lls->sub_lock;
1098 if (lsl == NULL) /* already removed */
1101 rc = lov_sublock_lock(env, lck, lls, closure, NULL);
1102 if (rc == CLO_REPEAT) {
1108 LASSERT(lsl->lss_cl.cls_lock->cll_state < CLS_FREEING);
1110 if (lls->sub_flags & LSF_HELD)
1111 lov_sublock_release(env, lck, i, 1, 0);
1113 link = lov_lock_link_find(env, lck, lsl);
1114 LASSERT(link != NULL);
1115 lov_lock_unlink(env, link, lsl);
1116 LASSERT(lck->lls_sub[i].sub_lock == NULL);
1118 lov_sublock_unlock(env, lsl, closure, NULL);
1121 cl_lock_closure_fini(closure);
1125 static int lov_lock_print(const struct lu_env *env, void *cookie,
1126 lu_printer_t p, const struct cl_lock_slice *slice)
1128 struct lov_lock *lck = cl2lov_lock(slice);
1131 (*p)(env, cookie, "%d\n", lck->lls_nr);
1132 for (i = 0; i < lck->lls_nr; ++i) {
1133 struct lov_lock_sub *sub;
1135 sub = &lck->lls_sub[i];
1136 (*p)(env, cookie, " %d %x: ", i, sub->sub_flags);
1137 if (sub->sub_lock != NULL)
1138 cl_lock_print(env, cookie, p,
1139 sub->sub_lock->lss_cl.cls_lock);
1141 (*p)(env, cookie, "---\n");
1146 static const struct cl_lock_operations lov_lock_ops = {
1147 .clo_fini = lov_lock_fini,
1148 .clo_enqueue = lov_lock_enqueue,
1149 .clo_wait = lov_lock_wait,
1150 .clo_use = lov_lock_use,
1151 .clo_unuse = lov_lock_unuse,
1152 .clo_cancel = lov_lock_cancel,
1153 .clo_fits_into = lov_lock_fits_into,
1154 .clo_delete = lov_lock_delete,
1155 .clo_print = lov_lock_print
1158 int lov_lock_init_raid0(const struct lu_env *env, struct cl_object *obj,
1159 struct cl_lock *lock, const struct cl_io *io)
1161 struct lov_lock *lck;
1165 OBD_SLAB_ALLOC_PTR_GFP(lck, lov_lock_kmem, CFS_ALLOC_IO);
1167 cl_lock_slice_add(lock, &lck->lls_cl, obj, &lov_lock_ops);
1168 result = lov_lock_sub_init(env, lck, io);
1174 static struct cl_lock_closure *lov_closure_get(const struct lu_env *env,
1175 struct cl_lock *parent)
1177 struct cl_lock_closure *closure;
1179 closure = &lov_env_info(env)->lti_closure;
1180 LASSERT(cfs_list_empty(&closure->clc_list));
1181 cl_lock_closure_init(env, closure, parent, 1);