1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_lock for OSC layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
41 #define DEBUG_SUBSYSTEM S_OSC
44 # include <libcfs/libcfs.h>
46 # include <liblustre.h>
48 /* fid_build_reg_res_name() */
49 #include <lustre_fid.h>
51 #include "osc_cl_internal.h"
57 /*****************************************************************************
63 static const struct cl_lock_operations osc_lock_ops;
64 static const struct cl_lock_operations osc_lock_lockless_ops;
65 static void osc_lock_to_lockless(const struct lu_env *env,
66 struct osc_lock *ols, int force);
67 static int osc_lock_has_pages(struct osc_lock *olck);
69 int osc_lock_is_lockless(const struct osc_lock *olck)
71 return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
75 * Returns a weak pointer to the ldlm lock identified by a handle. Returned
76 * pointer cannot be dereferenced, as lock is not protected from concurrent
77 * reclaim. This function is a helper for osc_lock_invariant().
79 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
81 struct ldlm_lock *lock;
83 lock = ldlm_handle2lock(handle);
90 * Invariant that has to be true all of the time.
92 static int osc_lock_invariant(struct osc_lock *ols)
94 struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle);
95 struct ldlm_lock *olock = ols->ols_lock;
96 int handle_used = lustre_handle_is_used(&ols->ols_handle);
99 ergo(osc_lock_is_lockless(ols),
100 ols->ols_locklessable && ols->ols_lock == NULL) ||
101 (ergo(olock != NULL, handle_used) &&
103 olock->l_handle.h_cookie == ols->ols_handle.cookie) &&
105 * Check that ->ols_handle and ->ols_lock are consistent, but
106 * take into account that they are set at the different time.
109 ergo(lock != NULL && olock != NULL, lock == olock) &&
110 ergo(lock == NULL, olock == NULL)) &&
111 ergo(ols->ols_state == OLS_CANCELLED,
112 olock == NULL && !handle_used) &&
114 * DLM lock is destroyed only after we have seen cancellation
117 ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
118 !olock->l_destroyed) &&
119 ergo(ols->ols_state == OLS_GRANTED,
121 olock->l_req_mode == olock->l_granted_mode &&
125 /*****************************************************************************
132 * Breaks a link between osc_lock and dlm_lock.
134 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
136 struct ldlm_lock *dlmlock;
138 cfs_spin_lock(&osc_ast_guard);
139 dlmlock = olck->ols_lock;
140 if (dlmlock == NULL) {
141 cfs_spin_unlock(&osc_ast_guard);
145 olck->ols_lock = NULL;
146 /* wb(); --- for all who checks (ols->ols_lock != NULL) before
147 * call to osc_lock_detach() */
148 dlmlock->l_ast_data = NULL;
149 olck->ols_handle.cookie = 0ULL;
150 cfs_spin_unlock(&osc_ast_guard);
152 lock_res_and_lock(dlmlock);
153 if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
154 struct cl_object *obj = olck->ols_cl.cls_obj;
155 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
156 __u64 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
158 /* Update the kms. Need to loop all granted locks.
159 * Not a problem for the client */
160 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
161 unlock_res_and_lock(dlmlock);
163 cl_object_attr_lock(obj);
164 cl_object_attr_set(env, obj, attr, CAT_KMS);
165 cl_object_attr_unlock(obj);
167 unlock_res_and_lock(dlmlock);
169 /* release a reference taken in osc_lock_upcall0(). */
170 LASSERT(olck->ols_has_ref);
171 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
172 LDLM_LOCK_RELEASE(dlmlock);
173 olck->ols_has_ref = 0;
176 static int osc_lock_unhold(struct osc_lock *ols)
182 result = osc_cancel_base(&ols->ols_handle,
183 ols->ols_einfo.ei_mode);
188 static int osc_lock_unuse(const struct lu_env *env,
189 const struct cl_lock_slice *slice)
191 struct osc_lock *ols = cl2osc_lock(slice);
193 LASSERT(ols->ols_state == OLS_GRANTED ||
194 ols->ols_state == OLS_UPCALL_RECEIVED);
195 LINVRNT(osc_lock_invariant(ols));
197 if (ols->ols_glimpse) {
198 LASSERT(ols->ols_hold == 0);
201 LASSERT(ols->ols_hold);
204 * Move lock into OLS_RELEASED state before calling osc_cancel_base()
205 * so that possible synchronous cancellation (that always happens
206 * e.g., for liblustre) sees that lock is released.
208 ols->ols_state = OLS_RELEASED;
209 return osc_lock_unhold(ols);
212 static void osc_lock_fini(const struct lu_env *env,
213 struct cl_lock_slice *slice)
215 struct osc_lock *ols = cl2osc_lock(slice);
217 LINVRNT(osc_lock_invariant(ols));
219 * ->ols_hold can still be true at this point if, for example, a
220 * thread that requested a lock was killed (and released a reference
221 * to the lock), before reply from a server was received. In this case
222 * lock is destroyed immediately after upcall.
224 osc_lock_unhold(ols);
225 LASSERT(ols->ols_lock == NULL);
227 OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
230 void osc_lock_build_res(const struct lu_env *env, const struct osc_object *obj,
231 struct ldlm_res_id *resname)
233 const struct lu_fid *fid = lu_object_fid(&obj->oo_cl.co_lu);
236 * In the perfect world of the future, where ost servers talk
239 fid_build_reg_res_name(fid, resname);
242 * In reality, where ost server expects ->lsm_object_id and
243 * ->lsm_object_gr in rename.
245 osc_build_res_name(obj->oo_oinfo->loi_id, obj->oo_oinfo->loi_gr,
250 static void osc_lock_build_policy(const struct lu_env *env,
251 const struct cl_lock *lock,
252 ldlm_policy_data_t *policy)
254 const struct cl_lock_descr *d = &lock->cll_descr;
256 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
257 policy->l_extent.gid = d->cld_gid;
260 static int osc_enq2ldlm_flags(__u32 enqflags)
264 LASSERT((enqflags & ~CEF_MASK) == 0);
266 if (enqflags & CEF_NONBLOCK)
267 result |= LDLM_FL_BLOCK_NOWAIT;
268 if (enqflags & CEF_ASYNC)
269 result |= LDLM_FL_HAS_INTENT;
270 if (enqflags & CEF_DISCARD_DATA)
271 result |= LDLM_AST_DISCARD_DATA;
276 * Global spin-lock protecting consistency of ldlm_lock::l_ast_data
277 * pointers. Initialized in osc_init().
279 cfs_spinlock_t osc_ast_guard;
281 static struct osc_lock *osc_ast_data_get(struct ldlm_lock *dlm_lock)
283 struct osc_lock *olck;
285 lock_res_and_lock(dlm_lock);
286 cfs_spin_lock(&osc_ast_guard);
287 olck = dlm_lock->l_ast_data;
289 struct cl_lock *lock = olck->ols_cl.cls_lock;
291 * If osc_lock holds a reference on ldlm lock, return it even
292 * when cl_lock is in CLS_FREEING state. This way
294 * osc_ast_data_get(dlmlock) == NULL
296 * guarantees that all osc references on dlmlock were
297 * released. osc_dlm_blocking_ast0() relies on that.
299 if (lock->cll_state < CLS_FREEING || olck->ols_has_ref) {
300 cl_lock_get_trust(lock);
301 lu_ref_add_atomic(&lock->cll_reference,
302 "ast", cfs_current());
306 cfs_spin_unlock(&osc_ast_guard);
307 unlock_res_and_lock(dlm_lock);
311 static void osc_ast_data_put(const struct lu_env *env, struct osc_lock *olck)
313 struct cl_lock *lock;
315 lock = olck->ols_cl.cls_lock;
316 lu_ref_del(&lock->cll_reference, "ast", cfs_current());
317 cl_lock_put(env, lock);
321 * Updates object attributes from a lock value block (lvb) received together
322 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
325 * This can be optimized to not update attributes when lock is a result of a
328 * Called under lock and resource spin-locks.
330 static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
334 struct cl_object *obj;
335 struct lov_oinfo *oinfo;
336 struct cl_attr *attr;
341 if (!(olck->ols_flags & LDLM_FL_LVB_READY)) {
346 lvb = &olck->ols_lvb;
347 obj = olck->ols_cl.cls_obj;
348 oinfo = cl2osc(obj)->oo_oinfo;
349 attr = &osc_env_info(env)->oti_attr;
350 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
351 cl_lvb2attr(attr, lvb);
353 cl_object_attr_lock(obj);
355 struct ldlm_lock *dlmlock;
358 dlmlock = olck->ols_lock;
359 LASSERT(dlmlock != NULL);
361 /* re-grab LVB from a dlm lock under DLM spin-locks. */
362 *lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
363 size = lvb->lvb_size;
364 /* Extend KMS up to the end of this lock and no further
365 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
366 if (size > dlmlock->l_policy_data.l_extent.end)
367 size = dlmlock->l_policy_data.l_extent.end + 1;
368 if (size >= oinfo->loi_kms) {
369 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="LPU64
370 ", kms="LPU64, lvb->lvb_size, size);
372 attr->cat_kms = size;
374 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
375 LPU64"; leaving kms="LPU64", end="LPU64,
376 lvb->lvb_size, oinfo->loi_kms,
377 dlmlock->l_policy_data.l_extent.end);
379 ldlm_lock_allow_match_locked(dlmlock);
380 } else if (rc == -ENAVAIL && olck->ols_glimpse) {
381 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
382 " kms="LPU64"\n", lvb->lvb_size, oinfo->loi_kms);
387 cl_object_attr_set(env, obj, attr, valid);
389 cl_object_attr_unlock(obj);
395 * Called when a lock is granted, from an upcall (when server returned a
396 * granted lock), or from completion AST, when server returned a blocked lock.
398 * Called under lock and resource spin-locks, that are released temporarily
401 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
402 struct ldlm_lock *dlmlock, int rc)
404 struct ldlm_extent *ext;
405 struct cl_lock *lock;
406 struct cl_lock_descr *descr;
408 LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
411 if (olck->ols_state != OLS_GRANTED) {
412 lock = olck->ols_cl.cls_lock;
413 ext = &dlmlock->l_policy_data.l_extent;
414 descr = &osc_env_info(env)->oti_descr;
415 descr->cld_obj = lock->cll_descr.cld_obj;
417 /* XXX check that ->l_granted_mode is valid. */
418 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
419 descr->cld_start = cl_index(descr->cld_obj, ext->start);
420 descr->cld_end = cl_index(descr->cld_obj, ext->end);
421 descr->cld_gid = ext->gid;
423 * tell upper layers the extent of the lock that was actually
426 olck->ols_state = OLS_GRANTED;
427 osc_lock_lvb_update(env, olck, rc);
429 /* release DLM spin-locks to allow cl_lock_{modify,signal}()
430 * to take a semaphore on a parent lock. This is safe, because
431 * spin-locks are needed to protect consistency of
432 * dlmlock->l_*_mode and LVB, and we have finished processing
434 unlock_res_and_lock(dlmlock);
435 cl_lock_modify(env, lock, descr);
436 cl_lock_signal(env, lock);
437 LINVRNT(osc_lock_invariant(olck));
438 lock_res_and_lock(dlmlock);
443 static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
446 struct ldlm_lock *dlmlock;
450 dlmlock = ldlm_handle2lock_long(&olck->ols_handle, 0);
451 LASSERT(dlmlock != NULL);
453 lock_res_and_lock(dlmlock);
454 cfs_spin_lock(&osc_ast_guard);
455 LASSERT(dlmlock->l_ast_data == olck);
456 LASSERT(olck->ols_lock == NULL);
457 olck->ols_lock = dlmlock;
458 cfs_spin_unlock(&osc_ast_guard);
461 * Lock might be not yet granted. In this case, completion ast
462 * (osc_ldlm_completion_ast()) comes later and finishes lock
465 if (dlmlock->l_granted_mode == dlmlock->l_req_mode)
466 osc_lock_granted(env, olck, dlmlock, 0);
467 unlock_res_and_lock(dlmlock);
470 * osc_enqueue_interpret() decrefs asynchronous locks, counter
473 ldlm_lock_addref(&olck->ols_handle, olck->ols_einfo.ei_mode);
476 /* lock reference taken by ldlm_handle2lock_long() is owned by
477 * osc_lock and released in osc_lock_detach() */
478 lu_ref_add(&dlmlock->l_reference, "osc_lock", olck);
479 olck->ols_has_ref = 1;
483 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
484 * received from a server, or after osc_enqueue_base() matched a local DLM
487 static int osc_lock_upcall(void *cookie, int errcode)
489 struct osc_lock *olck = cookie;
490 struct cl_lock_slice *slice = &olck->ols_cl;
491 struct cl_lock *lock = slice->cls_lock;
493 struct cl_env_nest nest;
496 env = cl_env_nested_get(&nest);
500 cl_lock_mutex_get(env, lock);
502 LASSERT(lock->cll_state >= CLS_QUEUING);
503 if (olck->ols_state == OLS_ENQUEUED) {
504 olck->ols_state = OLS_UPCALL_RECEIVED;
505 rc = ldlm_error2errno(errcode);
506 } else if (olck->ols_state == OLS_CANCELLED) {
509 CERROR("Impossible state: %i\n", olck->ols_state);
513 struct ldlm_lock *dlmlock;
515 dlmlock = ldlm_handle2lock(&olck->ols_handle);
516 if (dlmlock != NULL) {
517 lock_res_and_lock(dlmlock);
518 cfs_spin_lock(&osc_ast_guard);
519 LASSERT(olck->ols_lock == NULL);
520 dlmlock->l_ast_data = NULL;
521 olck->ols_handle.cookie = 0ULL;
522 cfs_spin_unlock(&osc_ast_guard);
523 unlock_res_and_lock(dlmlock);
524 LDLM_LOCK_PUT(dlmlock);
527 if (olck->ols_glimpse)
528 olck->ols_glimpse = 0;
529 osc_lock_upcall0(env, olck);
532 /* Error handling, some errors are tolerable. */
533 if (olck->ols_locklessable && rc == -EUSERS) {
534 /* This is a tolerable error, turn this lock into
537 osc_object_set_contended(cl2osc(slice->cls_obj));
538 LASSERT(slice->cls_ops == &osc_lock_ops);
540 /* Change this lock to ldlmlock-less lock. */
541 osc_lock_to_lockless(env, olck, 1);
542 olck->ols_state = OLS_GRANTED;
544 } else if (olck->ols_glimpse && rc == -ENAVAIL) {
545 osc_lock_lvb_update(env, olck, rc);
546 cl_lock_delete(env, lock);
547 /* Hide the error. */
552 /* on error, lock was signaled by cl_lock_error() */
553 cl_lock_signal(env, lock);
555 cl_lock_error(env, lock, rc);
557 cl_lock_mutex_put(env, lock);
559 /* release cookie reference, acquired by osc_lock_enqueue() */
560 lu_ref_del(&lock->cll_reference, "upcall", lock);
561 cl_lock_put(env, lock);
562 cl_env_nested_put(&nest, env);
564 /* should never happen, similar to osc_ldlm_blocking_ast(). */
570 * Core of osc_dlm_blocking_ast() logic.
572 static void osc_lock_blocking(const struct lu_env *env,
573 struct ldlm_lock *dlmlock,
574 struct osc_lock *olck, int blocking)
576 struct cl_lock *lock = olck->ols_cl.cls_lock;
578 LASSERT(olck->ols_lock == dlmlock);
579 CLASSERT(OLS_BLOCKED < OLS_CANCELLED);
580 LASSERT(!osc_lock_is_lockless(olck));
583 * Lock might be still addref-ed here, if e.g., blocking ast
584 * is sent for a failed lock.
586 osc_lock_unhold(olck);
588 if (blocking && olck->ols_state < OLS_BLOCKED)
590 * Move osc_lock into OLS_BLOCKED before canceling the lock,
591 * because it recursively re-enters osc_lock_blocking(), with
592 * the state set to OLS_CANCELLED.
594 olck->ols_state = OLS_BLOCKED;
596 * cancel and destroy lock at least once no matter how blocking ast is
597 * entered (see comment above osc_ldlm_blocking_ast() for use
598 * cases). cl_lock_cancel() and cl_lock_delete() are idempotent.
600 cl_lock_cancel(env, lock);
601 cl_lock_delete(env, lock);
605 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
606 * and ldlm_lock caches.
608 static int osc_dlm_blocking_ast0(const struct lu_env *env,
609 struct ldlm_lock *dlmlock,
610 void *data, int flag)
612 struct osc_lock *olck;
613 struct cl_lock *lock;
617 LASSERT(flag == LDLM_CB_BLOCKING || flag == LDLM_CB_CANCELING);
620 olck = osc_ast_data_get(dlmlock);
622 lock = olck->ols_cl.cls_lock;
623 cl_lock_mutex_get(env, lock);
624 LINVRNT(osc_lock_invariant(olck));
625 if (olck->ols_ast_wait) {
626 /* wake up osc_lock_use() */
627 cl_lock_signal(env, lock);
628 olck->ols_ast_wait = 0;
631 * Lock might have been canceled while this thread was
632 * sleeping for lock mutex, but olck is pinned in memory.
634 if (olck == dlmlock->l_ast_data) {
636 * NOTE: DLM sends blocking AST's for failed locks
637 * (that are still in pre-OLS_GRANTED state)
638 * too, and they have to be canceled otherwise
639 * DLM lock is never destroyed and stuck in
642 * Alternatively, ldlm_cli_cancel() can be
643 * called here directly for osc_locks with
644 * ols_state < OLS_GRANTED to maintain an
645 * invariant that ->clo_cancel() is only called
646 * for locks that were granted.
648 LASSERT(data == olck);
649 osc_lock_blocking(env, dlmlock,
650 olck, flag == LDLM_CB_BLOCKING);
653 cl_lock_mutex_put(env, lock);
654 osc_ast_data_put(env, olck);
657 * DLM lock exists, but there is no cl_lock attached to it.
658 * This is a `normal' race. cl_object and its cl_lock's can be
659 * removed by memory pressure, together with all pages.
661 cancel = (flag == LDLM_CB_BLOCKING);
664 struct lustre_handle *lockh;
666 lockh = &osc_env_info(env)->oti_handle;
667 ldlm_lock2handle(dlmlock, lockh);
668 result = ldlm_cli_cancel(lockh);
675 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
676 * some other lock, or is canceled. This function is installed as a
677 * ldlm_lock::l_blocking_ast() for client extent locks.
679 * Control flow is tricky, because ldlm uses the same call-back
680 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
682 * \param dlmlock lock for which ast occurred.
684 * \param new description of a conflicting lock in case of blocking ast.
686 * \param data value of dlmlock->l_ast_data
688 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
689 * cancellation and blocking ast's.
691 * Possible use cases:
693 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
694 * lock due to lock lru pressure, or explicit user request to purge
697 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
698 * us that dlmlock conflicts with another lock that some client is
699 * enqueing. Lock is canceled.
701 * - cl_lock_cancel() is called. osc_lock_cancel() calls
702 * ldlm_cli_cancel() that calls
704 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
706 * recursively entering osc_ldlm_blocking_ast().
708 * - client cancels lock voluntary (e.g., as a part of early cancellation):
711 * osc_lock_cancel()->
712 * ldlm_cli_cancel()->
713 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
716 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
717 struct ldlm_lock_desc *new, void *data,
721 struct cl_env_nest nest;
725 * This can be called in the context of outer IO, e.g.,
728 * ->osc_enqueue_base()->...
729 * ->ldlm_prep_elc_req()->...
730 * ->ldlm_cancel_callback()->...
731 * ->osc_ldlm_blocking_ast()
733 * new environment has to be created to not corrupt outer context.
735 env = cl_env_nested_get(&nest);
737 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
738 cl_env_nested_put(&nest, env);
740 result = PTR_ERR(env);
742 * XXX This should never happen, as cl_lock is
743 * stuck. Pre-allocated environment a la vvp_inode_fini_env
749 if (result == -ENODATA)
752 CERROR("BAST failed: %d\n", result);
757 static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
758 int flags, void *data)
760 struct cl_env_nest nest;
762 struct osc_lock *olck;
763 struct cl_lock *lock;
767 /* first, do dlm part of the work */
768 dlmrc = ldlm_completion_ast_async(dlmlock, flags, data);
769 /* then, notify cl_lock */
770 env = cl_env_nested_get(&nest);
772 olck = osc_ast_data_get(dlmlock);
774 lock = olck->ols_cl.cls_lock;
775 cl_lock_mutex_get(env, lock);
777 * ldlm_handle_cp_callback() copied LVB from request
778 * to lock->l_lvb_data, store it in osc_lock.
780 LASSERT(dlmlock->l_lvb_data != NULL);
781 lock_res_and_lock(dlmlock);
782 olck->ols_lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
783 if (olck->ols_lock == NULL) {
785 * upcall (osc_lock_upcall()) hasn't yet been
786 * called. Do nothing now, upcall will bind
787 * olck to dlmlock and signal the waiters.
789 * This maintains an invariant that osc_lock
790 * and ldlm_lock are always bound when
791 * osc_lock is in OLS_GRANTED state.
793 } else if (dlmlock->l_granted_mode ==
794 dlmlock->l_req_mode) {
795 osc_lock_granted(env, olck, dlmlock, dlmrc);
797 unlock_res_and_lock(dlmlock);
800 CL_LOCK_DEBUG(D_ERROR, env, lock,
801 "dlmlock returned %d\n", dlmrc);
802 cl_lock_error(env, lock, dlmrc);
804 cl_lock_mutex_put(env, lock);
805 osc_ast_data_put(env, olck);
808 result = -ELDLM_NO_LOCK_DATA;
809 cl_env_nested_put(&nest, env);
811 result = PTR_ERR(env);
812 return dlmrc ?: result;
815 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
817 struct ptlrpc_request *req = data;
818 struct osc_lock *olck;
819 struct cl_lock *lock;
820 struct cl_object *obj;
821 struct cl_env_nest nest;
824 struct req_capsule *cap;
827 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
829 env = cl_env_nested_get(&nest);
832 * osc_ast_data_get() has to go after environment is
833 * allocated, because osc_ast_data() acquires a
834 * reference to a lock, and it can only be released in
837 olck = osc_ast_data_get(dlmlock);
839 lock = olck->ols_cl.cls_lock;
840 cl_lock_mutex_get(env, lock);
842 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
843 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
845 result = req_capsule_server_pack(cap);
847 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
848 obj = lock->cll_descr.cld_obj;
849 result = cl_object_glimpse(env, obj, lvb);
851 cl_lock_mutex_put(env, lock);
852 osc_ast_data_put(env, olck);
855 * These errors are normal races, so we don't want to
856 * fill the console with messages by calling
859 lustre_pack_reply(req, 1, NULL, NULL);
860 result = -ELDLM_NO_LOCK_DATA;
862 cl_env_nested_put(&nest, env);
864 result = PTR_ERR(env);
865 req->rq_status = result;
869 static unsigned long osc_lock_weigh(const struct lu_env *env,
870 const struct cl_lock_slice *slice)
873 * don't need to grab coh_page_guard since we don't care the exact #
876 return cl_object_header(slice->cls_obj)->coh_pages;
880 * Get the weight of dlm lock for early cancellation.
882 * XXX: it should return the pages covered by this \a dlmlock.
884 static unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
886 struct cl_env_nest nest;
888 struct osc_lock *lock;
890 unsigned long weight;
895 * osc_ldlm_weigh_ast has a complex context since it might be called
896 * because of lock canceling, or from user's input. We have to make
897 * a new environment for it. Probably it is implementation safe to use
898 * the upper context because cl_lock_put don't modify environment
899 * variables. But in case of ..
901 env = cl_env_nested_get(&nest);
903 /* Mostly because lack of memory, tend to eliminate this lock*/
906 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
907 lock = osc_ast_data_get(dlmlock);
909 /* cl_lock was destroyed because of memory pressure.
910 * It is much reasonable to assign this type of lock
913 GOTO(out, weight = 0);
916 cll = lock->ols_cl.cls_lock;
917 cl_lock_mutex_get(env, cll);
918 weight = cl_lock_weigh(env, cll);
919 cl_lock_mutex_put(env, cll);
920 osc_ast_data_put(env, lock);
924 cl_env_nested_put(&nest, env);
928 static void osc_lock_build_einfo(const struct lu_env *env,
929 const struct cl_lock *clock,
930 struct osc_lock *lock,
931 struct ldlm_enqueue_info *einfo)
933 enum cl_lock_mode mode;
935 mode = clock->cll_descr.cld_mode;
936 if (mode == CLM_PHANTOM)
938 * For now, enqueue all glimpse locks in read mode. In the
939 * future, client might choose to enqueue LCK_PW lock for
940 * glimpse on a file opened for write.
944 einfo->ei_type = LDLM_EXTENT;
945 einfo->ei_mode = osc_cl_lock2ldlm(mode);
946 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
947 einfo->ei_cb_cp = osc_ldlm_completion_ast;
948 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
949 einfo->ei_cb_wg = osc_ldlm_weigh_ast;
950 einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */
954 * Determine if the lock should be converted into a lockless lock.
957 * - if the lock has an explicite requirment for a non-lockless lock;
958 * - if the io lock request type ci_lockreq;
959 * - send the enqueue rpc to ost to make the further decision;
960 * - special treat to truncate lockless lock
962 * Additional policy can be implemented here, e.g., never do lockless-io
965 static void osc_lock_to_lockless(const struct lu_env *env,
966 struct osc_lock *ols, int force)
968 struct cl_lock_slice *slice = &ols->ols_cl;
969 struct cl_lock *lock = slice->cls_lock;
971 LASSERT(ols->ols_state == OLS_NEW ||
972 ols->ols_state == OLS_UPCALL_RECEIVED);
975 ols->ols_locklessable = 1;
976 LASSERT(cl_lock_is_mutexed(lock));
977 slice->cls_ops = &osc_lock_lockless_ops;
979 struct osc_io *oio = osc_env_io(env);
980 struct cl_io *io = oio->oi_cl.cis_io;
981 struct cl_object *obj = slice->cls_obj;
982 struct osc_object *oob = cl2osc(obj);
983 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
984 struct obd_connect_data *ocd;
986 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
987 io->ci_lockreq == CILR_MAYBE ||
988 io->ci_lockreq == CILR_NEVER);
990 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
991 ols->ols_locklessable = (io->ci_type != CIT_TRUNC) &&
992 (io->ci_lockreq == CILR_MAYBE) &&
993 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
994 if (io->ci_lockreq == CILR_NEVER ||
996 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
997 /* lockless truncate */
998 (io->ci_type == CIT_TRUNC &&
999 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
1000 osd->od_lockless_truncate)) {
1001 ols->ols_locklessable = 1;
1002 slice->cls_ops = &osc_lock_lockless_ops;
1005 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1008 static int osc_lock_compatible(const struct osc_lock *qing,
1009 const struct osc_lock *qed)
1011 enum cl_lock_mode qing_mode;
1012 enum cl_lock_mode qed_mode;
1014 qing_mode = qing->ols_cl.cls_lock->cll_descr.cld_mode;
1015 if (qed->ols_glimpse &&
1016 (qed->ols_state >= OLS_UPCALL_RECEIVED || qing_mode == CLM_READ))
1019 qed_mode = qed->ols_cl.cls_lock->cll_descr.cld_mode;
1020 return ((qing_mode == CLM_READ) && (qed_mode == CLM_READ));
1024 * Cancel all conflicting locks and wait for them to be destroyed.
1026 * This function is used for two purposes:
1028 * - early cancel all conflicting locks before starting IO, and
1030 * - guarantee that pages added to the page cache by lockless IO are never
1031 * covered by locks other than lockless IO lock, and, hence, are not
1032 * visible to other threads.
1034 static int osc_lock_enqueue_wait(const struct lu_env *env,
1035 const struct osc_lock *olck)
1037 struct cl_lock *lock = olck->ols_cl.cls_lock;
1038 struct cl_lock_descr *descr = &lock->cll_descr;
1039 struct cl_object_header *hdr = cl_object_header(descr->cld_obj);
1040 struct cl_lock *scan = lock;
1041 struct cl_lock *conflict= NULL;
1042 int lockless = osc_lock_is_lockless(olck);
1046 LASSERT(cl_lock_is_mutexed(lock));
1047 LASSERT(lock->cll_state == CLS_QUEUING);
1049 /* make it enqueue anyway for glimpse lock, because we actually
1050 * don't need to cancel any conflicting locks. */
1051 if (olck->ols_glimpse)
1054 cfs_spin_lock(&hdr->coh_lock_guard);
1055 cfs_list_for_each_entry_continue(scan, &hdr->coh_locks, cll_linkage) {
1056 struct cl_lock_descr *cld = &scan->cll_descr;
1057 const struct osc_lock *scan_ols;
1059 if (scan->cll_state < CLS_QUEUING ||
1060 scan->cll_state == CLS_FREEING ||
1061 cld->cld_start > descr->cld_end ||
1062 cld->cld_end < descr->cld_start)
1065 /* overlapped and living locks. */
1067 /* We're not supposed to give up group lock. */
1068 if (scan->cll_descr.cld_mode == CLM_GROUP) {
1069 LASSERT(descr->cld_mode != CLM_GROUP ||
1070 descr->cld_gid != scan->cll_descr.cld_gid);
1074 scan_ols = osc_lock_at(scan);
1076 /* We need to cancel the compatible locks if we're enqueuing
1077 * a lockless lock, for example:
1078 * imagine that client has PR lock on [0, 1000], and thread T0
1079 * is doing lockless IO in [500, 1500] region. Concurrent
1080 * thread T1 can see lockless data in [500, 1000], which is
1081 * wrong, because these data are possibly stale. */
1082 if (!lockless && osc_lock_compatible(olck, scan_ols))
1085 /* Now @scan is conflicting with @lock, this means current
1086 * thread have to sleep for @scan being destroyed. */
1087 if (scan_ols->ols_owner == osc_env_io(env)) {
1088 CERROR("DEADLOCK POSSIBLE!\n");
1089 CL_LOCK_DEBUG(D_ERROR, env, scan, "queued.\n");
1090 CL_LOCK_DEBUG(D_ERROR, env, lock, "queuing.\n");
1091 libcfs_debug_dumpstack(NULL);
1093 cl_lock_get_trust(scan);
1097 cfs_spin_unlock(&hdr->coh_lock_guard);
1100 CDEBUG(D_DLMTRACE, "lock %p is confliced with %p, will wait\n",
1102 lu_ref_add(&conflict->cll_reference, "cancel-wait", lock);
1103 LASSERT(lock->cll_conflict == NULL);
1104 lock->cll_conflict = conflict;
1111 * Deadlock avoidance for osc_lock_enqueue(). Consider following scenario:
1113 * - Thread0: obtains PR:[0, 10]. Lock is busy.
1115 * - Thread1: enqueues PW:[5, 50]. Blocking ast is sent to
1116 * PR:[0, 10], but cancellation of busy lock is postponed.
1118 * - Thread0: enqueue PR:[30, 40]. Lock is locally matched to
1119 * PW:[5, 50], and thread0 waits for the lock completion never
1120 * releasing PR:[0, 10]---deadlock.
1122 * The second PR lock can be glimpse (it is to deal with that situation that
1123 * ll_glimpse_size() has second argument, preventing local match of
1124 * not-yet-granted locks, see bug 10295). Similar situation is possible in the
1125 * case of memory mapped user level buffer.
1127 * To prevent this we can detect a situation when current "thread" or "io"
1128 * already holds a lock on this object and either add LDLM_FL_BLOCK_GRANTED to
1129 * the ols->ols_flags, or prevent local match with PW locks.
1131 static int osc_deadlock_is_possible(const struct lu_env *env,
1132 struct cl_lock *lock)
1134 struct cl_object *obj;
1135 struct cl_object_header *head;
1136 struct cl_lock *scan;
1143 LASSERT(cl_lock_is_mutexed(lock));
1145 oio = osc_env_io(env);
1146 obj = lock->cll_descr.cld_obj;
1147 head = cl_object_header(obj);
1150 cfs_spin_lock(&head->coh_lock_guard);
1151 cfs_list_for_each_entry(scan, &head->coh_locks, cll_linkage) {
1153 struct osc_lock *oscan;
1155 oscan = osc_lock_at(scan);
1156 LASSERT(oscan != NULL);
1157 if (oscan->ols_owner == oio) {
1163 cfs_spin_unlock(&head->coh_lock_guard);
1168 * Implementation of cl_lock_operations::clo_enqueue() method for osc
1169 * layer. This initiates ldlm enqueue:
1171 * - checks for possible dead-lock conditions (osc_deadlock_is_possible());
1173 * - cancels conflicting locks early (osc_lock_enqueue_wait());
1175 * - calls osc_enqueue_base() to do actual enqueue.
1177 * osc_enqueue_base() is supplied with an upcall function that is executed
1178 * when lock is received either after a local cached ldlm lock is matched, or
1179 * when a reply from the server is received.
1181 * This function does not wait for the network communication to complete.
1183 static int osc_lock_enqueue(const struct lu_env *env,
1184 const struct cl_lock_slice *slice,
1185 struct cl_io *unused, __u32 enqflags)
1187 struct osc_lock *ols = cl2osc_lock(slice);
1188 struct cl_lock *lock = ols->ols_cl.cls_lock;
1189 struct osc_object *obj = cl2osc(slice->cls_obj);
1190 struct osc_thread_info *info = osc_env_info(env);
1191 struct ldlm_res_id *resname = &info->oti_resname;
1192 ldlm_policy_data_t *policy = &info->oti_policy;
1193 struct ldlm_enqueue_info *einfo = &ols->ols_einfo;
1197 LASSERT(cl_lock_is_mutexed(lock));
1198 LASSERT(lock->cll_state == CLS_QUEUING);
1199 LASSERT(ols->ols_state == OLS_NEW);
1201 osc_lock_build_res(env, obj, resname);
1202 osc_lock_build_policy(env, lock, policy);
1203 ols->ols_flags = osc_enq2ldlm_flags(enqflags);
1204 if (osc_deadlock_is_possible(env, lock))
1205 ols->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1206 if (ols->ols_flags & LDLM_FL_HAS_INTENT)
1207 ols->ols_glimpse = 1;
1208 if (!(enqflags & CEF_MUST))
1209 /* try to convert this lock to a lockless lock */
1210 osc_lock_to_lockless(env, ols, (enqflags & CEF_NEVER));
1212 result = osc_lock_enqueue_wait(env, ols);
1214 if (!osc_lock_is_lockless(ols)) {
1215 if (ols->ols_locklessable)
1216 ols->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1218 /* a reference for lock, passed as an upcall cookie */
1220 lu_ref_add(&lock->cll_reference, "upcall", lock);
1221 ols->ols_state = OLS_ENQUEUED;
1224 * XXX: this is possible blocking point as
1225 * ldlm_lock_match(LDLM_FL_LVB_READY) waits for
1228 result = osc_enqueue_base(osc_export(obj), resname,
1229 &ols->ols_flags, policy,
1231 obj->oo_oinfo->loi_kms_valid,
1233 ols, einfo, &ols->ols_handle,
1236 lu_ref_del(&lock->cll_reference,
1238 cl_lock_put(env, lock);
1241 ols->ols_state = OLS_GRANTED;
1244 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1248 static int osc_lock_wait(const struct lu_env *env,
1249 const struct cl_lock_slice *slice)
1251 struct osc_lock *olck = cl2osc_lock(slice);
1252 struct cl_lock *lock = olck->ols_cl.cls_lock;
1254 LINVRNT(osc_lock_invariant(olck));
1255 if (olck->ols_glimpse && olck->ols_state >= OLS_UPCALL_RECEIVED)
1258 LASSERT(equi(olck->ols_state >= OLS_UPCALL_RECEIVED &&
1259 lock->cll_error == 0, olck->ols_lock != NULL));
1261 return lock->cll_error ?: olck->ols_state >= OLS_GRANTED ? 0 : CLO_WAIT;
1265 * An implementation of cl_lock_operations::clo_use() method that pins cached
1268 static int osc_lock_use(const struct lu_env *env,
1269 const struct cl_lock_slice *slice)
1271 struct osc_lock *olck = cl2osc_lock(slice);
1274 LASSERT(!olck->ols_hold);
1277 * Atomically check for LDLM_FL_CBPENDING and addref a lock if this
1278 * flag is not set. This protects us from a concurrent blocking ast.
1280 rc = ldlm_lock_addref_try(&olck->ols_handle, olck->ols_einfo.ei_mode);
1283 olck->ols_state = OLS_GRANTED;
1285 struct cl_lock *lock;
1288 * Lock is being cancelled somewhere within
1289 * ldlm_handle_bl_callback(): LDLM_FL_CBPENDING is already
1290 * set, but osc_ldlm_blocking_ast() hasn't yet acquired
1293 lock = slice->cls_lock;
1294 LASSERT(lock->cll_state == CLS_INTRANSIT);
1295 LASSERT(lock->cll_users > 0);
1296 /* set a flag for osc_dlm_blocking_ast0() to signal the
1298 olck->ols_ast_wait = 1;
1304 static int osc_lock_flush(struct osc_lock *ols, int discard)
1306 struct cl_lock *lock = ols->ols_cl.cls_lock;
1307 struct cl_env_nest nest;
1311 env = cl_env_nested_get(&nest);
1313 result = cl_lock_page_out(env, lock, discard);
1314 cl_env_nested_put(&nest, env);
1316 result = PTR_ERR(env);
1319 LINVRNT(!osc_lock_has_pages(ols));
1325 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1326 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1327 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1328 * with some other lock some where in the cluster. This function does the
1331 * - invalidates all pages protected by this lock (after sending dirty
1332 * ones to the server, as necessary);
1334 * - decref's underlying ldlm lock;
1336 * - cancels ldlm lock (ldlm_cli_cancel()).
1338 static void osc_lock_cancel(const struct lu_env *env,
1339 const struct cl_lock_slice *slice)
1341 struct cl_lock *lock = slice->cls_lock;
1342 struct osc_lock *olck = cl2osc_lock(slice);
1343 struct ldlm_lock *dlmlock = olck->ols_lock;
1347 LASSERT(cl_lock_is_mutexed(lock));
1348 LINVRNT(osc_lock_invariant(olck));
1350 if (dlmlock != NULL) {
1353 discard = dlmlock->l_flags & LDLM_FL_DISCARD_DATA;
1354 result = osc_lock_flush(olck, discard);
1355 osc_lock_unhold(olck);
1357 lock_res_and_lock(dlmlock);
1358 /* Now that we're the only user of dlm read/write reference,
1359 * mostly the ->l_readers + ->l_writers should be zero.
1360 * However, there is a corner case.
1361 * See bug 18829 for details.*/
1362 do_cancel = (dlmlock->l_readers == 0 &&
1363 dlmlock->l_writers == 0);
1364 dlmlock->l_flags |= LDLM_FL_CBPENDING;
1365 unlock_res_and_lock(dlmlock);
1367 result = ldlm_cli_cancel(&olck->ols_handle);
1369 CL_LOCK_DEBUG(D_ERROR, env, lock,
1370 "lock %p cancel failure with error(%d)\n",
1373 olck->ols_state = OLS_CANCELLED;
1374 osc_lock_detach(env, olck);
1377 void cl_lock_page_list_fixup(const struct lu_env *env,
1378 struct cl_io *io, struct cl_lock *lock,
1379 struct cl_page_list *queue);
1381 #ifdef INVARIANT_CHECK
1383 * Returns true iff there are pages under \a olck not protected by other
1386 static int osc_lock_has_pages(struct osc_lock *olck)
1388 struct cl_lock *lock;
1389 struct cl_lock_descr *descr;
1390 struct cl_object *obj;
1391 struct osc_object *oob;
1392 struct cl_page_list *plist;
1393 struct cl_page *page;
1394 struct cl_env_nest nest;
1399 env = cl_env_nested_get(&nest);
1401 obj = olck->ols_cl.cls_obj;
1403 io = &oob->oo_debug_io;
1404 lock = olck->ols_cl.cls_lock;
1405 descr = &lock->cll_descr;
1406 plist = &osc_env_info(env)->oti_plist;
1407 cl_page_list_init(plist);
1409 cfs_mutex_lock(&oob->oo_debug_mutex);
1411 io->ci_obj = cl_object_top(obj);
1412 cl_io_init(env, io, CIT_MISC, io->ci_obj);
1413 cl_page_gang_lookup(env, obj, io,
1414 descr->cld_start, descr->cld_end, plist, 0,
1416 cl_lock_page_list_fixup(env, io, lock, plist);
1417 if (plist->pl_nr > 0) {
1418 CL_LOCK_DEBUG(D_ERROR, env, lock, "still has pages\n");
1419 cl_page_list_for_each(page, plist)
1420 CL_PAGE_DEBUG(D_ERROR, env, page, "\n");
1422 result = plist->pl_nr > 0;
1423 cl_page_list_disown(env, io, plist);
1424 cl_page_list_fini(env, plist);
1425 cl_io_fini(env, io);
1426 cfs_mutex_unlock(&oob->oo_debug_mutex);
1427 cl_env_nested_put(&nest, env);
1433 static int osc_lock_has_pages(struct osc_lock *olck)
1437 #endif /* INVARIANT_CHECK */
1439 static void osc_lock_delete(const struct lu_env *env,
1440 const struct cl_lock_slice *slice)
1442 struct osc_lock *olck;
1444 olck = cl2osc_lock(slice);
1445 if (olck->ols_glimpse) {
1446 LASSERT(!olck->ols_hold);
1447 LASSERT(!olck->ols_lock);
1451 LINVRNT(osc_lock_invariant(olck));
1452 LINVRNT(!osc_lock_has_pages(olck));
1454 osc_lock_unhold(olck);
1455 osc_lock_detach(env, olck);
1459 * Implements cl_lock_operations::clo_state() method for osc layer.
1461 * Maintains osc_lock::ols_owner field.
1463 * This assumes that lock always enters CLS_HELD (from some other state) in
1464 * the same IO context as one that requested the lock. This should not be a
1465 * problem, because context is by definition shared by all activity pertaining
1466 * to the same high-level IO.
1468 static void osc_lock_state(const struct lu_env *env,
1469 const struct cl_lock_slice *slice,
1470 enum cl_lock_state state)
1472 struct osc_lock *lock = cl2osc_lock(slice);
1475 * XXX multiple io contexts can use the lock at the same time.
1477 LINVRNT(osc_lock_invariant(lock));
1478 if (state == CLS_HELD && slice->cls_lock->cll_state != CLS_HELD) {
1479 struct osc_io *oio = osc_env_io(env);
1481 LASSERT(lock->ols_owner == NULL);
1482 lock->ols_owner = oio;
1483 } else if (state != CLS_HELD)
1484 lock->ols_owner = NULL;
1487 static int osc_lock_print(const struct lu_env *env, void *cookie,
1488 lu_printer_t p, const struct cl_lock_slice *slice)
1490 struct osc_lock *lock = cl2osc_lock(slice);
1493 * XXX print ldlm lock and einfo properly.
1495 (*p)(env, cookie, "%p %08x "LPX64" %d %p ",
1496 lock->ols_lock, lock->ols_flags, lock->ols_handle.cookie,
1497 lock->ols_state, lock->ols_owner);
1498 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1502 static int osc_lock_fits_into(const struct lu_env *env,
1503 const struct cl_lock_slice *slice,
1504 const struct cl_lock_descr *need,
1505 const struct cl_io *io)
1507 struct osc_lock *ols = cl2osc_lock(slice);
1509 if (need->cld_enq_flags & CEF_NEVER)
1512 if (need->cld_mode == CLM_PHANTOM) {
1514 * Note: the QUEUED lock can't be matched here, otherwise
1515 * it might cause the deadlocks.
1517 * P1: enqueued read lock, create sublock1
1518 * P2: enqueued write lock, create sublock2(conflicted
1520 * P1: Grant read lock.
1521 * P1: enqueued glimpse lock(with holding sublock1_read),
1522 * matched with sublock2, waiting sublock2 to be granted.
1523 * But sublock2 can not be granted, because P1
1524 * will not release sublock1. Bang!
1526 if (ols->ols_state < OLS_GRANTED ||
1527 ols->ols_state > OLS_RELEASED)
1529 } else if (need->cld_enq_flags & CEF_MUST) {
1531 * If the lock hasn't ever enqueued, it can't be matched
1532 * because enqueue process brings in many information
1533 * which can be used to determine things such as lockless,
1536 if (ols->ols_state < OLS_GRANTED ||
1537 ols->ols_state > OLS_RELEASED)
1539 if (ols->ols_state < OLS_UPCALL_RECEIVED &&
1540 ols->ols_locklessable)
1546 static const struct cl_lock_operations osc_lock_ops = {
1547 .clo_fini = osc_lock_fini,
1548 .clo_enqueue = osc_lock_enqueue,
1549 .clo_wait = osc_lock_wait,
1550 .clo_unuse = osc_lock_unuse,
1551 .clo_use = osc_lock_use,
1552 .clo_delete = osc_lock_delete,
1553 .clo_state = osc_lock_state,
1554 .clo_cancel = osc_lock_cancel,
1555 .clo_weigh = osc_lock_weigh,
1556 .clo_print = osc_lock_print,
1557 .clo_fits_into = osc_lock_fits_into,
1560 static int osc_lock_lockless_enqueue(const struct lu_env *env,
1561 const struct cl_lock_slice *slice,
1562 struct cl_io *unused, __u32 enqflags)
1568 static int osc_lock_lockless_unuse(const struct lu_env *env,
1569 const struct cl_lock_slice *slice)
1571 struct osc_lock *ols = cl2osc_lock(slice);
1572 struct cl_lock *lock = slice->cls_lock;
1574 LASSERT(ols->ols_state == OLS_GRANTED);
1575 LINVRNT(osc_lock_invariant(ols));
1577 cl_lock_cancel(env, lock);
1578 cl_lock_delete(env, lock);
1582 static void osc_lock_lockless_cancel(const struct lu_env *env,
1583 const struct cl_lock_slice *slice)
1585 struct osc_lock *ols = cl2osc_lock(slice);
1588 result = osc_lock_flush(ols, 0);
1590 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1592 ols->ols_state = OLS_CANCELLED;
1595 static int osc_lock_lockless_wait(const struct lu_env *env,
1596 const struct cl_lock_slice *slice)
1598 struct osc_lock *olck = cl2osc_lock(slice);
1599 struct cl_lock *lock = olck->ols_cl.cls_lock;
1601 LINVRNT(osc_lock_invariant(olck));
1602 LASSERT(olck->ols_state >= OLS_UPCALL_RECEIVED);
1604 return lock->cll_error;
1607 static void osc_lock_lockless_state(const struct lu_env *env,
1608 const struct cl_lock_slice *slice,
1609 enum cl_lock_state state)
1611 struct osc_lock *lock = cl2osc_lock(slice);
1613 LINVRNT(osc_lock_invariant(lock));
1614 if (state == CLS_HELD) {
1615 struct osc_io *oio = osc_env_io(env);
1617 LASSERT(lock->ols_owner == NULL);
1618 lock->ols_owner = oio;
1620 /* set the io to be lockless if this lock is for io's
1622 if (cl_object_same(oio->oi_cl.cis_obj, slice->cls_obj))
1623 oio->oi_lockless = 1;
1625 lock->ols_owner = NULL;
1628 static int osc_lock_lockless_fits_into(const struct lu_env *env,
1629 const struct cl_lock_slice *slice,
1630 const struct cl_lock_descr *need,
1631 const struct cl_io *io)
1636 static const struct cl_lock_operations osc_lock_lockless_ops = {
1637 .clo_fini = osc_lock_fini,
1638 .clo_enqueue = osc_lock_lockless_enqueue,
1639 .clo_wait = osc_lock_lockless_wait,
1640 .clo_unuse = osc_lock_lockless_unuse,
1641 .clo_state = osc_lock_lockless_state,
1642 .clo_fits_into = osc_lock_lockless_fits_into,
1643 .clo_cancel = osc_lock_lockless_cancel,
1644 .clo_print = osc_lock_print
1647 int osc_lock_init(const struct lu_env *env,
1648 struct cl_object *obj, struct cl_lock *lock,
1649 const struct cl_io *unused)
1651 struct osc_lock *clk;
1654 OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, CFS_ALLOC_IO);
1656 osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo);
1657 clk->ols_state = OLS_NEW;
1658 cl_lock_slice_add(lock, &clk->ols_cl, obj, &osc_lock_ops);