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, 2012, Whamcloud, Inc.
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)
272 LASSERT(result <= 0 || result == CLO_REPEAT || result == CLO_WAIT);
273 LASSERT(rc <= 0 || rc == CLO_REPEAT || rc == CLO_WAIT);
274 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_LARGE(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_LARGE(lck->lls_sub,
489 lck->lls_nr * sizeof lck->lls_sub[0]);
491 OBD_SLAB_FREE_PTR(lck, lov_lock_kmem);
495 static int lov_lock_enqueue_wait(const struct lu_env *env,
496 struct lov_lock *lck,
497 struct cl_lock *sublock)
499 struct cl_lock *lock = lck->lls_cl.cls_lock;
503 LASSERT(cl_lock_is_mutexed(lock));
505 cl_lock_mutex_put(env, lock);
506 result = cl_lock_enqueue_wait(env, sublock, 0);
507 cl_lock_mutex_get(env, lock);
508 RETURN(result ?: CLO_REPEAT);
512 * Tries to advance a state machine of a given sub-lock toward enqueuing of
515 * \retval 0 if state-transition can proceed
516 * \retval -ve otherwise.
518 static int lov_lock_enqueue_one(const struct lu_env *env, struct lov_lock *lck,
519 struct cl_lock *sublock,
520 struct cl_io *io, __u32 enqflags, int last)
525 /* first, try to enqueue a sub-lock ... */
526 result = cl_enqueue_try(env, sublock, io, enqflags);
527 if ((sublock->cll_state == CLS_ENQUEUED) && !(enqflags & CEF_AGL))
528 /* if it is enqueued, try to `wait' on it---maybe it's already
530 result = cl_wait_try(env, sublock);
532 * If CEF_ASYNC flag is set, then all sub-locks can be enqueued in
533 * parallel, otherwise---enqueue has to wait until sub-lock is granted
534 * before proceeding to the next one.
536 if ((result == CLO_WAIT) && (sublock->cll_state <= CLS_HELD) &&
537 (enqflags & CEF_ASYNC) && (!last || (enqflags & CEF_AGL)))
543 * Helper function for lov_lock_enqueue() that creates missing sub-lock.
545 static int lov_sublock_fill(const struct lu_env *env, struct cl_lock *parent,
546 struct cl_io *io, struct lov_lock *lck, int idx)
548 struct lov_lock_link *link;
549 struct cl_lock *sublock;
552 LASSERT(parent->cll_depth == 1);
553 cl_lock_mutex_put(env, parent);
554 sublock = lov_sublock_alloc(env, io, lck, idx, &link);
555 if (!IS_ERR(sublock))
556 cl_lock_mutex_get(env, sublock);
557 cl_lock_mutex_get(env, parent);
559 if (!IS_ERR(sublock)) {
560 cl_lock_get_trust(sublock);
561 if (parent->cll_state == CLS_QUEUING &&
562 lck->lls_sub[idx].sub_lock == NULL) {
563 lov_sublock_adopt(env, lck, sublock, idx, link);
565 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
566 /* other thread allocated sub-lock, or enqueue is no
568 cl_lock_mutex_put(env, parent);
569 cl_lock_unhold(env, sublock, "lov-parent", parent);
570 cl_lock_mutex_get(env, parent);
572 cl_lock_mutex_put(env, sublock);
573 cl_lock_put(env, sublock);
576 result = PTR_ERR(sublock);
581 * Implementation of cl_lock_operations::clo_enqueue() for lov layer. This
582 * function is rather subtle, as it enqueues top-lock (i.e., advances top-lock
583 * state machine from CLS_QUEUING to CLS_ENQUEUED states) by juggling sub-lock
584 * state machines in the face of sub-locks sharing (by multiple top-locks),
585 * and concurrent sub-lock cancellations.
587 static int lov_lock_enqueue(const struct lu_env *env,
588 const struct cl_lock_slice *slice,
589 struct cl_io *io, __u32 enqflags)
591 struct cl_lock *lock = slice->cls_lock;
592 struct lov_lock *lck = cl2lov_lock(slice);
593 struct cl_lock_closure *closure = lov_closure_get(env, lock);
596 enum cl_lock_state minstate;
600 for (result = 0, minstate = CLS_FREEING, i = 0; i < lck->lls_nr; ++i) {
602 struct lovsub_lock *sub;
603 struct lov_lock_sub *lls;
604 struct cl_lock *sublock;
605 struct lov_sublock_env *subenv;
607 if (lock->cll_state != CLS_QUEUING) {
609 * Lock might have left QUEUING state if previous
610 * iteration released its mutex. Stop enqueing in this
611 * case and let the upper layer to decide what to do.
613 LASSERT(i > 0 && result != 0);
617 lls = &lck->lls_sub[i];
620 * Sub-lock might have been canceled, while top-lock was
624 result = lov_sublock_fill(env, lock, io, lck, i);
625 /* lov_sublock_fill() released @lock mutex,
629 sublock = sub->lss_cl.cls_lock;
630 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
632 lov_sublock_hold(env, lck, i);
633 rc = lov_lock_enqueue_one(subenv->lse_env, lck, sublock,
634 subenv->lse_io, enqflags,
635 i == lck->lls_nr - 1);
636 minstate = min(minstate, sublock->cll_state);
637 if (rc == CLO_WAIT) {
638 switch (sublock->cll_state) {
640 /* take recursive mutex, the lock is
641 * released in lov_lock_enqueue_wait.
643 cl_lock_mutex_get(env, sublock);
644 lov_sublock_unlock(env, sub, closure,
646 rc = lov_lock_enqueue_wait(env, lck,
650 rc = lov_sublock_release(env, lck, i,
653 lov_sublock_unlock(env, sub, closure,
658 LASSERT(sublock->cll_conflict == NULL);
659 lov_sublock_unlock(env, sub, closure, subenv);
662 result = lov_subresult(result, rc);
666 cl_lock_closure_fini(closure);
667 RETURN(result ?: minstate >= CLS_ENQUEUED ? 0 : CLO_WAIT);
670 static int lov_lock_unuse(const struct lu_env *env,
671 const struct cl_lock_slice *slice)
673 struct lov_lock *lck = cl2lov_lock(slice);
674 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
680 for (result = 0, i = 0; i < lck->lls_nr; ++i) {
682 struct lovsub_lock *sub;
683 struct cl_lock *sublock;
684 struct lov_lock_sub *lls;
685 struct lov_sublock_env *subenv;
687 /* top-lock state cannot change concurrently, because single
688 * thread (one that released the last hold) carries unlocking
689 * to the completion. */
690 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
691 lls = &lck->lls_sub[i];
696 sublock = sub->lss_cl.cls_lock;
697 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
699 if (lls->sub_flags & LSF_HELD) {
700 LASSERT(sublock->cll_state == CLS_HELD ||
701 sublock->cll_state == CLS_ENQUEUED);
702 rc = cl_unuse_try(subenv->lse_env, sublock);
703 rc = lov_sublock_release(env, lck, i, 0, rc);
705 lov_sublock_unlock(env, sub, closure, subenv);
707 result = lov_subresult(result, rc);
710 if (result == 0 && lck->lls_cancel_race) {
711 lck->lls_cancel_race = 0;
714 cl_lock_closure_fini(closure);
719 static void lov_lock_cancel(const struct lu_env *env,
720 const struct cl_lock_slice *slice)
722 struct lov_lock *lck = cl2lov_lock(slice);
723 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
729 for (result = 0, i = 0; i < lck->lls_nr; ++i) {
731 struct lovsub_lock *sub;
732 struct cl_lock *sublock;
733 struct lov_lock_sub *lls;
734 struct lov_sublock_env *subenv;
736 /* top-lock state cannot change concurrently, because single
737 * thread (one that released the last hold) carries unlocking
738 * to the completion. */
739 lls = &lck->lls_sub[i];
744 sublock = sub->lss_cl.cls_lock;
745 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
747 if (!(lls->sub_flags & LSF_HELD)) {
748 lov_sublock_unlock(env, sub, closure, subenv);
752 switch(sublock->cll_state) {
754 rc = cl_unuse_try(subenv->lse_env, sublock);
755 lov_sublock_release(env, lck, i, 0, 0);
758 lov_sublock_release(env, lck, i, 1, 0);
761 lov_sublock_unlock(env, sub, closure, subenv);
764 if (rc == CLO_REPEAT) {
769 result = lov_subresult(result, rc);
773 CL_LOCK_DEBUG(D_ERROR, env, slice->cls_lock,
774 "lov_lock_cancel fails with %d.\n", result);
776 cl_lock_closure_fini(closure);
779 static int lov_lock_wait(const struct lu_env *env,
780 const struct cl_lock_slice *slice)
782 struct lov_lock *lck = cl2lov_lock(slice);
783 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
784 enum cl_lock_state minstate;
792 for (result = 0, minstate = CLS_FREEING, i = 0, reenqueued = 0;
793 i < lck->lls_nr; ++i) {
795 struct lovsub_lock *sub;
796 struct cl_lock *sublock;
797 struct lov_lock_sub *lls;
798 struct lov_sublock_env *subenv;
800 lls = &lck->lls_sub[i];
802 LASSERT(sub != NULL);
803 sublock = sub->lss_cl.cls_lock;
804 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
806 LASSERT(sublock->cll_state >= CLS_ENQUEUED);
807 if (sublock->cll_state < CLS_HELD)
808 rc = cl_wait_try(env, sublock);
810 minstate = min(minstate, sublock->cll_state);
811 lov_sublock_unlock(env, sub, closure, subenv);
813 if (rc == CLO_REENQUEUED) {
817 result = lov_subresult(result, rc);
821 /* Each sublock only can be reenqueued once, so will not loop for
823 if (result == 0 && reenqueued != 0)
825 cl_lock_closure_fini(closure);
826 RETURN(result ?: minstate >= CLS_HELD ? 0 : CLO_WAIT);
829 static int lov_lock_use(const struct lu_env *env,
830 const struct cl_lock_slice *slice)
832 struct lov_lock *lck = cl2lov_lock(slice);
833 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
837 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
840 for (result = 0, i = 0; i < lck->lls_nr; ++i) {
842 struct lovsub_lock *sub;
843 struct cl_lock *sublock;
844 struct lov_lock_sub *lls;
845 struct lov_sublock_env *subenv;
847 LASSERT(slice->cls_lock->cll_state == CLS_INTRANSIT);
849 lls = &lck->lls_sub[i];
853 * Sub-lock might have been canceled, while top-lock was
860 sublock = sub->lss_cl.cls_lock;
861 rc = lov_sublock_lock(env, lck, lls, closure, &subenv);
863 LASSERT(sublock->cll_state != CLS_FREEING);
864 lov_sublock_hold(env, lck, i);
865 if (sublock->cll_state == CLS_CACHED) {
866 rc = cl_use_try(subenv->lse_env, sublock, 0);
868 rc = lov_sublock_release(env, lck,
870 } else if (sublock->cll_state == CLS_NEW) {
871 /* Sub-lock might have been canceled, while
872 * top-lock was cached. */
874 lov_sublock_release(env, lck, i, 1, result);
876 lov_sublock_unlock(env, sub, closure, subenv);
878 result = lov_subresult(result, rc);
883 if (lck->lls_cancel_race) {
885 * If there is unlocking happened at the same time, then
886 * sublock_lock state should be FREEING, and lov_sublock_lock
887 * should return CLO_REPEAT. In this case, it should return
888 * ESTALE, and up layer should reset the lock state to be NEW.
890 lck->lls_cancel_race = 0;
891 LASSERT(result != 0);
894 cl_lock_closure_fini(closure);
899 static int lock_lock_multi_match()
901 struct cl_lock *lock = slice->cls_lock;
902 struct cl_lock_descr *subneed = &lov_env_info(env)->lti_ldescr;
903 struct lov_object *loo = cl2lov(lov->lls_cl.cls_obj);
904 struct lov_layout_raid0 *r0 = lov_r0(loo);
905 struct lov_lock_sub *sub;
906 struct cl_object *subobj;
913 fstart = cl_offset(need->cld_obj, need->cld_start);
914 fend = cl_offset(need->cld_obj, need->cld_end + 1) - 1;
915 subneed->cld_mode = need->cld_mode;
916 cl_lock_mutex_get(env, lock);
917 for (i = 0; i < lov->lls_nr; ++i) {
918 sub = &lov->lls_sub[i];
919 if (sub->sub_lock == NULL)
921 subobj = sub->sub_descr.cld_obj;
922 if (!lov_stripe_intersects(r0->lo_lsm, sub->sub_stripe,
923 fstart, fend, &start, &end))
925 subneed->cld_start = cl_index(subobj, start);
926 subneed->cld_end = cl_index(subobj, end);
927 subneed->cld_obj = subobj;
928 if (!cl_lock_ext_match(&sub->sub_got, subneed)) {
933 cl_lock_mutex_put(env, lock);
938 * Check if the extent region \a descr is covered by \a child against the
939 * specific \a stripe.
941 static int lov_lock_stripe_is_matching(const struct lu_env *env,
942 struct lov_object *lov, int stripe,
943 const struct cl_lock_descr *child,
944 const struct cl_lock_descr *descr)
946 struct lov_stripe_md *lsm = lov_r0(lov)->lo_lsm;
951 if (lov_r0(lov)->lo_nr == 1)
952 return cl_lock_ext_match(child, descr);
955 * For a multi-stripes object:
956 * - make sure the descr only covers child's stripe, and
957 * - check if extent is matching.
959 start = cl_offset(&lov->lo_cl, descr->cld_start);
960 end = cl_offset(&lov->lo_cl, descr->cld_end + 1) - 1;
961 result = end - start <= lsm->lsm_stripe_size &&
962 stripe == lov_stripe_number(lsm, start) &&
963 stripe == lov_stripe_number(lsm, end);
965 struct cl_lock_descr *subd = &lov_env_info(env)->lti_ldescr;
969 subd->cld_obj = NULL; /* don't need sub object at all */
970 subd->cld_mode = descr->cld_mode;
971 subd->cld_gid = descr->cld_gid;
972 result = lov_stripe_intersects(lsm, stripe, start, end,
973 &sub_start, &sub_end);
975 subd->cld_start = cl_index(child->cld_obj, sub_start);
976 subd->cld_end = cl_index(child->cld_obj, sub_end);
977 result = cl_lock_ext_match(child, subd);
983 * An implementation of cl_lock_operations::clo_fits_into() method.
985 * Checks whether a lock (given by \a slice) is suitable for \a
986 * io. Multi-stripe locks can be used only for "quick" io, like truncate, or
989 * \see ccc_lock_fits_into().
991 static int lov_lock_fits_into(const struct lu_env *env,
992 const struct cl_lock_slice *slice,
993 const struct cl_lock_descr *need,
994 const struct cl_io *io)
996 struct lov_lock *lov = cl2lov_lock(slice);
997 struct lov_object *obj = cl2lov(slice->cls_obj);
1000 LASSERT(cl_object_same(need->cld_obj, slice->cls_obj));
1001 LASSERT(lov->lls_nr > 0);
1005 if (need->cld_mode == CLM_GROUP)
1007 * always allow to match group lock.
1009 result = cl_lock_ext_match(&lov->lls_orig, need);
1010 else if (lov->lls_nr == 1) {
1011 struct cl_lock_descr *got = &lov->lls_sub[0].sub_got;
1012 result = lov_lock_stripe_is_matching(env,
1013 cl2lov(slice->cls_obj),
1014 lov->lls_sub[0].sub_stripe,
1016 } else if (io->ci_type != CIT_SETATTR && io->ci_type != CIT_MISC &&
1017 !cl_io_is_append(io) && need->cld_mode != CLM_PHANTOM)
1019 * Multi-stripe locks are only suitable for `quick' IO and for
1025 * Most general case: multi-stripe existing lock, and
1026 * (potentially) multi-stripe @need lock. Check that @need is
1027 * covered by @lov's sub-locks.
1029 * For now, ignore lock expansions made by the server, and
1030 * match against original lock extent.
1032 result = cl_lock_ext_match(&lov->lls_orig, need);
1033 CDEBUG(D_DLMTRACE, DDESCR"/"DDESCR" %d %d/%d: %d\n",
1034 PDESCR(&lov->lls_orig), PDESCR(&lov->lls_sub[0].sub_got),
1035 lov->lls_sub[0].sub_stripe, lov->lls_nr, lov_r0(obj)->lo_nr,
1040 void lov_lock_unlink(const struct lu_env *env,
1041 struct lov_lock_link *link, struct lovsub_lock *sub)
1043 struct lov_lock *lck = link->lll_super;
1044 struct cl_lock *parent = lck->lls_cl.cls_lock;
1046 LASSERT(cl_lock_is_mutexed(parent));
1047 LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
1050 cfs_list_del_init(&link->lll_list);
1051 LASSERT(lck->lls_sub[link->lll_idx].sub_lock == sub);
1052 /* yank this sub-lock from parent's array */
1053 lck->lls_sub[link->lll_idx].sub_lock = NULL;
1054 LASSERT(lck->lls_nr_filled > 0);
1055 lck->lls_nr_filled--;
1056 lu_ref_del(&parent->cll_reference, "lov-child", sub->lss_cl.cls_lock);
1057 cl_lock_put(env, parent);
1058 OBD_SLAB_FREE_PTR(link, lov_lock_link_kmem);
1062 struct lov_lock_link *lov_lock_link_find(const struct lu_env *env,
1063 struct lov_lock *lck,
1064 struct lovsub_lock *sub)
1066 struct lov_lock_link *scan;
1068 LASSERT(cl_lock_is_mutexed(sub->lss_cl.cls_lock));
1071 cfs_list_for_each_entry(scan, &sub->lss_parents, lll_list) {
1072 if (scan->lll_super == lck)
1079 * An implementation of cl_lock_operations::clo_delete() method. This is
1080 * invoked for "top-to-bottom" delete, when lock destruction starts from the
1081 * top-lock, e.g., as a result of inode destruction.
1083 * Unlinks top-lock from all its sub-locks. Sub-locks are not deleted there:
1084 * this is done separately elsewhere:
1086 * - for inode destruction, lov_object_delete() calls cl_object_kill() for
1087 * each sub-object, purging its locks;
1089 * - in other cases (e.g., a fatal error with a top-lock) sub-locks are
1090 * left in the cache.
1092 static void lov_lock_delete(const struct lu_env *env,
1093 const struct cl_lock_slice *slice)
1095 struct lov_lock *lck = cl2lov_lock(slice);
1096 struct cl_lock_closure *closure = lov_closure_get(env, slice->cls_lock);
1097 struct lov_lock_link *link;
1101 LASSERT(slice->cls_lock->cll_state == CLS_FREEING);
1104 for (i = 0; i < lck->lls_nr; ++i) {
1105 struct lov_lock_sub *lls = &lck->lls_sub[i];
1106 struct lovsub_lock *lsl = lls->sub_lock;
1108 if (lsl == NULL) /* already removed */
1111 rc = lov_sublock_lock(env, lck, lls, closure, NULL);
1112 if (rc == CLO_REPEAT) {
1118 LASSERT(lsl->lss_cl.cls_lock->cll_state < CLS_FREEING);
1120 if (lls->sub_flags & LSF_HELD)
1121 lov_sublock_release(env, lck, i, 1, 0);
1123 link = lov_lock_link_find(env, lck, lsl);
1124 LASSERT(link != NULL);
1125 lov_lock_unlink(env, link, lsl);
1126 LASSERT(lck->lls_sub[i].sub_lock == NULL);
1128 lov_sublock_unlock(env, lsl, closure, NULL);
1131 cl_lock_closure_fini(closure);
1135 static int lov_lock_print(const struct lu_env *env, void *cookie,
1136 lu_printer_t p, const struct cl_lock_slice *slice)
1138 struct lov_lock *lck = cl2lov_lock(slice);
1141 (*p)(env, cookie, "%d\n", lck->lls_nr);
1142 for (i = 0; i < lck->lls_nr; ++i) {
1143 struct lov_lock_sub *sub;
1145 sub = &lck->lls_sub[i];
1146 (*p)(env, cookie, " %d %x: ", i, sub->sub_flags);
1147 if (sub->sub_lock != NULL)
1148 cl_lock_print(env, cookie, p,
1149 sub->sub_lock->lss_cl.cls_lock);
1151 (*p)(env, cookie, "---\n");
1156 static const struct cl_lock_operations lov_lock_ops = {
1157 .clo_fini = lov_lock_fini,
1158 .clo_enqueue = lov_lock_enqueue,
1159 .clo_wait = lov_lock_wait,
1160 .clo_use = lov_lock_use,
1161 .clo_unuse = lov_lock_unuse,
1162 .clo_cancel = lov_lock_cancel,
1163 .clo_fits_into = lov_lock_fits_into,
1164 .clo_delete = lov_lock_delete,
1165 .clo_print = lov_lock_print
1168 int lov_lock_init_raid0(const struct lu_env *env, struct cl_object *obj,
1169 struct cl_lock *lock, const struct cl_io *io)
1171 struct lov_lock *lck;
1175 OBD_SLAB_ALLOC_PTR_GFP(lck, lov_lock_kmem, CFS_ALLOC_IO);
1177 cl_lock_slice_add(lock, &lck->lls_cl, obj, &lov_lock_ops);
1178 result = lov_lock_sub_init(env, lck, io);
1184 static struct cl_lock_closure *lov_closure_get(const struct lu_env *env,
1185 struct cl_lock *parent)
1187 struct cl_lock_closure *closure;
1189 closure = &lov_env_info(env)->lti_closure;
1190 LASSERT(cfs_list_empty(&closure->clc_list));
1191 cl_lock_closure_init(env, closure, parent, 1);