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.
32 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * Implementation of cl_lock for OSC layer.
41 * Author: Nikita Danilov <nikita.danilov@sun.com>
44 #define DEBUG_SUBSYSTEM S_OSC
47 # include <libcfs/libcfs.h>
49 # include <liblustre.h>
51 /* fid_build_reg_res_name() */
52 #include <lustre_fid.h>
54 #include "osc_cl_internal.h"
60 #define _PAGEREF_MAGIC (-10000000)
62 /*****************************************************************************
68 static const struct cl_lock_operations osc_lock_ops;
69 static const struct cl_lock_operations osc_lock_lockless_ops;
70 static void osc_lock_to_lockless(const struct lu_env *env,
71 struct osc_lock *ols, int force);
72 static int osc_lock_has_pages(struct osc_lock *olck);
74 int osc_lock_is_lockless(const struct osc_lock *olck)
76 return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
80 * Returns a weak pointer to the ldlm lock identified by a handle. Returned
81 * pointer cannot be dereferenced, as lock is not protected from concurrent
82 * reclaim. This function is a helper for osc_lock_invariant().
84 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
86 struct ldlm_lock *lock;
88 lock = ldlm_handle2lock(handle);
95 * Invariant that has to be true all of the time.
97 static int osc_lock_invariant(struct osc_lock *ols)
99 struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle);
100 struct ldlm_lock *olock = ols->ols_lock;
101 int handle_used = lustre_handle_is_used(&ols->ols_handle);
104 ergo(osc_lock_is_lockless(ols),
105 ols->ols_locklessable && ols->ols_lock == NULL) ||
106 (ergo(olock != NULL, handle_used) &&
108 olock->l_handle.h_cookie == ols->ols_handle.cookie) &&
110 * Check that ->ols_handle and ->ols_lock are consistent, but
111 * take into account that they are set at the different time.
114 ergo(lock != NULL && olock != NULL, lock == olock) &&
115 ergo(lock == NULL, olock == NULL)) &&
116 ergo(ols->ols_state == OLS_CANCELLED,
117 olock == NULL && !handle_used) &&
119 * DLM lock is destroyed only after we have seen cancellation
122 ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
123 !olock->l_destroyed) &&
124 ergo(ols->ols_state == OLS_GRANTED,
126 olock->l_req_mode == olock->l_granted_mode &&
130 /*****************************************************************************
137 * Breaks a link between osc_lock and dlm_lock.
139 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
141 struct ldlm_lock *dlmlock;
143 cfs_spin_lock(&osc_ast_guard);
144 dlmlock = olck->ols_lock;
145 if (dlmlock == NULL) {
146 cfs_spin_unlock(&osc_ast_guard);
150 olck->ols_lock = NULL;
151 /* wb(); --- for all who checks (ols->ols_lock != NULL) before
152 * call to osc_lock_detach() */
153 dlmlock->l_ast_data = NULL;
154 olck->ols_handle.cookie = 0ULL;
155 cfs_spin_unlock(&osc_ast_guard);
157 lock_res_and_lock(dlmlock);
158 if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
159 struct cl_object *obj = olck->ols_cl.cls_obj;
160 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
163 cl_object_attr_lock(obj);
164 /* Must get the value under the lock to avoid possible races. */
165 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
166 /* Update the kms. Need to loop all granted locks.
167 * Not a problem for the client */
168 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
170 cl_object_attr_set(env, obj, attr, CAT_KMS);
171 cl_object_attr_unlock(obj);
173 unlock_res_and_lock(dlmlock);
175 /* release a reference taken in osc_lock_upcall0(). */
176 LASSERT(olck->ols_has_ref);
177 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
178 LDLM_LOCK_RELEASE(dlmlock);
179 olck->ols_has_ref = 0;
182 static int osc_lock_unhold(struct osc_lock *ols)
188 result = osc_cancel_base(&ols->ols_handle,
189 ols->ols_einfo.ei_mode);
194 static int osc_lock_unuse(const struct lu_env *env,
195 const struct cl_lock_slice *slice)
197 struct osc_lock *ols = cl2osc_lock(slice);
199 LINVRNT(osc_lock_invariant(ols));
201 switch (ols->ols_state) {
203 LASSERT(!ols->ols_hold);
204 LASSERT(ols->ols_agl);
206 case OLS_UPCALL_RECEIVED:
207 LASSERT(!ols->ols_hold);
208 ols->ols_state = OLS_NEW;
211 LASSERT(!ols->ols_glimpse);
212 LASSERT(ols->ols_hold);
214 * Move lock into OLS_RELEASED state before calling
215 * osc_cancel_base() so that possible synchronous cancellation
216 * (that always happens e.g., for liblustre) sees that lock is
219 ols->ols_state = OLS_RELEASED;
220 return osc_lock_unhold(ols);
222 CERROR("Impossible state: %d\n", ols->ols_state);
227 static void osc_lock_fini(const struct lu_env *env,
228 struct cl_lock_slice *slice)
230 struct osc_lock *ols = cl2osc_lock(slice);
232 LINVRNT(osc_lock_invariant(ols));
234 * ->ols_hold can still be true at this point if, for example, a
235 * thread that requested a lock was killed (and released a reference
236 * to the lock), before reply from a server was received. In this case
237 * lock is destroyed immediately after upcall.
239 osc_lock_unhold(ols);
240 LASSERT(ols->ols_lock == NULL);
241 LASSERT(cfs_atomic_read(&ols->ols_pageref) == 0 ||
242 cfs_atomic_read(&ols->ols_pageref) == _PAGEREF_MAGIC);
244 OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
247 void osc_lock_build_res(const struct lu_env *env, const struct osc_object *obj,
248 struct ldlm_res_id *resname)
250 const struct lu_fid *fid = lu_object_fid(&obj->oo_cl.co_lu);
253 * In the perfect world of the future, where ost servers talk
256 fid_build_reg_res_name(fid, resname);
259 * In reality, where ost server expects ->lsm_object_id and
260 * ->lsm_object_seq in rename.
262 osc_build_res_name(obj->oo_oinfo->loi_id, obj->oo_oinfo->loi_seq,
267 static void osc_lock_build_policy(const struct lu_env *env,
268 const struct cl_lock *lock,
269 ldlm_policy_data_t *policy)
271 const struct cl_lock_descr *d = &lock->cll_descr;
273 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
274 policy->l_extent.gid = d->cld_gid;
277 static int osc_enq2ldlm_flags(__u32 enqflags)
281 LASSERT((enqflags & ~CEF_MASK) == 0);
283 if (enqflags & CEF_NONBLOCK)
284 result |= LDLM_FL_BLOCK_NOWAIT;
285 if (enqflags & CEF_ASYNC)
286 result |= LDLM_FL_HAS_INTENT;
287 if (enqflags & CEF_DISCARD_DATA)
288 result |= LDLM_AST_DISCARD_DATA;
293 * Global spin-lock protecting consistency of ldlm_lock::l_ast_data
294 * pointers. Initialized in osc_init().
296 cfs_spinlock_t osc_ast_guard;
298 static struct osc_lock *osc_ast_data_get(struct ldlm_lock *dlm_lock)
300 struct osc_lock *olck;
302 lock_res_and_lock(dlm_lock);
303 cfs_spin_lock(&osc_ast_guard);
304 olck = dlm_lock->l_ast_data;
306 struct cl_lock *lock = olck->ols_cl.cls_lock;
308 * If osc_lock holds a reference on ldlm lock, return it even
309 * when cl_lock is in CLS_FREEING state. This way
311 * osc_ast_data_get(dlmlock) == NULL
313 * guarantees that all osc references on dlmlock were
314 * released. osc_dlm_blocking_ast0() relies on that.
316 if (lock->cll_state < CLS_FREEING || olck->ols_has_ref) {
317 cl_lock_get_trust(lock);
318 lu_ref_add_atomic(&lock->cll_reference,
319 "ast", cfs_current());
323 cfs_spin_unlock(&osc_ast_guard);
324 unlock_res_and_lock(dlm_lock);
328 static void osc_ast_data_put(const struct lu_env *env, struct osc_lock *olck)
330 struct cl_lock *lock;
332 lock = olck->ols_cl.cls_lock;
333 lu_ref_del(&lock->cll_reference, "ast", cfs_current());
334 cl_lock_put(env, lock);
338 * Updates object attributes from a lock value block (lvb) received together
339 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
342 * This can be optimized to not update attributes when lock is a result of a
345 * Called under lock and resource spin-locks.
347 static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
351 struct cl_object *obj;
352 struct lov_oinfo *oinfo;
353 struct cl_attr *attr;
358 if (!(olck->ols_flags & LDLM_FL_LVB_READY))
361 lvb = &olck->ols_lvb;
362 obj = olck->ols_cl.cls_obj;
363 oinfo = cl2osc(obj)->oo_oinfo;
364 attr = &osc_env_info(env)->oti_attr;
365 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
366 cl_lvb2attr(attr, lvb);
368 cl_object_attr_lock(obj);
370 struct ldlm_lock *dlmlock;
373 dlmlock = olck->ols_lock;
374 LASSERT(dlmlock != NULL);
376 /* re-grab LVB from a dlm lock under DLM spin-locks. */
377 *lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
378 size = lvb->lvb_size;
379 /* Extend KMS up to the end of this lock and no further
380 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
381 if (size > dlmlock->l_policy_data.l_extent.end)
382 size = dlmlock->l_policy_data.l_extent.end + 1;
383 if (size >= oinfo->loi_kms) {
384 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="LPU64
385 ", kms="LPU64, lvb->lvb_size, size);
387 attr->cat_kms = size;
389 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
390 LPU64"; leaving kms="LPU64", end="LPU64,
391 lvb->lvb_size, oinfo->loi_kms,
392 dlmlock->l_policy_data.l_extent.end);
394 ldlm_lock_allow_match_locked(dlmlock);
395 } else if (rc == -ENAVAIL && olck->ols_glimpse) {
396 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
397 " kms="LPU64"\n", lvb->lvb_size, oinfo->loi_kms);
402 cl_object_attr_set(env, obj, attr, valid);
404 cl_object_attr_unlock(obj);
410 * Called when a lock is granted, from an upcall (when server returned a
411 * granted lock), or from completion AST, when server returned a blocked lock.
413 * Called under lock and resource spin-locks, that are released temporarily
416 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
417 struct ldlm_lock *dlmlock, int rc)
419 struct ldlm_extent *ext;
420 struct cl_lock *lock;
421 struct cl_lock_descr *descr;
423 LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
426 if (olck->ols_state < OLS_GRANTED) {
427 lock = olck->ols_cl.cls_lock;
428 ext = &dlmlock->l_policy_data.l_extent;
429 descr = &osc_env_info(env)->oti_descr;
430 descr->cld_obj = lock->cll_descr.cld_obj;
432 /* XXX check that ->l_granted_mode is valid. */
433 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
434 descr->cld_start = cl_index(descr->cld_obj, ext->start);
435 descr->cld_end = cl_index(descr->cld_obj, ext->end);
436 descr->cld_gid = ext->gid;
438 * tell upper layers the extent of the lock that was actually
441 olck->ols_state = OLS_GRANTED;
442 osc_lock_lvb_update(env, olck, rc);
444 /* release DLM spin-locks to allow cl_lock_{modify,signal}()
445 * to take a semaphore on a parent lock. This is safe, because
446 * spin-locks are needed to protect consistency of
447 * dlmlock->l_*_mode and LVB, and we have finished processing
449 unlock_res_and_lock(dlmlock);
450 cl_lock_modify(env, lock, descr);
451 cl_lock_signal(env, lock);
452 LINVRNT(osc_lock_invariant(olck));
453 lock_res_and_lock(dlmlock);
458 static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
461 struct ldlm_lock *dlmlock;
465 dlmlock = ldlm_handle2lock_long(&olck->ols_handle, 0);
466 LASSERT(dlmlock != NULL);
468 lock_res_and_lock(dlmlock);
469 cfs_spin_lock(&osc_ast_guard);
470 LASSERT(dlmlock->l_ast_data == olck);
471 LASSERT(olck->ols_lock == NULL);
472 olck->ols_lock = dlmlock;
473 cfs_spin_unlock(&osc_ast_guard);
476 * Lock might be not yet granted. In this case, completion ast
477 * (osc_ldlm_completion_ast()) comes later and finishes lock
480 if (dlmlock->l_granted_mode == dlmlock->l_req_mode)
481 osc_lock_granted(env, olck, dlmlock, 0);
482 unlock_res_and_lock(dlmlock);
485 * osc_enqueue_interpret() decrefs asynchronous locks, counter
488 ldlm_lock_addref(&olck->ols_handle, olck->ols_einfo.ei_mode);
491 /* lock reference taken by ldlm_handle2lock_long() is owned by
492 * osc_lock and released in osc_lock_detach() */
493 lu_ref_add(&dlmlock->l_reference, "osc_lock", olck);
494 olck->ols_has_ref = 1;
498 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
499 * received from a server, or after osc_enqueue_base() matched a local DLM
502 static int osc_lock_upcall(void *cookie, int errcode)
504 struct osc_lock *olck = cookie;
505 struct cl_lock_slice *slice = &olck->ols_cl;
506 struct cl_lock *lock = slice->cls_lock;
508 struct cl_env_nest nest;
511 env = cl_env_nested_get(&nest);
515 cl_lock_mutex_get(env, lock);
517 LASSERT(lock->cll_state >= CLS_QUEUING);
518 if (olck->ols_state == OLS_ENQUEUED) {
519 olck->ols_state = OLS_UPCALL_RECEIVED;
520 rc = ldlm_error2errno(errcode);
521 } else if (olck->ols_state == OLS_CANCELLED) {
524 CERROR("Impossible state: %d\n", olck->ols_state);
528 struct ldlm_lock *dlmlock;
530 dlmlock = ldlm_handle2lock(&olck->ols_handle);
531 if (dlmlock != NULL) {
532 lock_res_and_lock(dlmlock);
533 cfs_spin_lock(&osc_ast_guard);
534 LASSERT(olck->ols_lock == NULL);
535 dlmlock->l_ast_data = NULL;
536 olck->ols_handle.cookie = 0ULL;
537 cfs_spin_unlock(&osc_ast_guard);
538 ldlm_lock_fail_match_locked(dlmlock, rc);
539 unlock_res_and_lock(dlmlock);
540 LDLM_LOCK_PUT(dlmlock);
543 if (olck->ols_glimpse)
544 olck->ols_glimpse = 0;
545 osc_lock_upcall0(env, olck);
548 /* Error handling, some errors are tolerable. */
549 if (olck->ols_locklessable && rc == -EUSERS) {
550 /* This is a tolerable error, turn this lock into
553 osc_object_set_contended(cl2osc(slice->cls_obj));
554 LASSERT(slice->cls_ops == &osc_lock_ops);
556 /* Change this lock to ldlmlock-less lock. */
557 osc_lock_to_lockless(env, olck, 1);
558 olck->ols_state = OLS_GRANTED;
560 } else if (olck->ols_glimpse && rc == -ENAVAIL) {
561 osc_lock_lvb_update(env, olck, rc);
562 cl_lock_delete(env, lock);
563 /* Hide the error. */
568 cl_lock_signal(env, lock);
569 /* del user for lock upcall cookie */
570 cl_unuse_try(env, lock);
572 /* del user for lock upcall cookie */
573 cl_lock_user_del(env, lock);
574 cl_lock_error(env, lock, rc);
577 cl_lock_mutex_put(env, lock);
579 /* release cookie reference, acquired by osc_lock_enqueue() */
580 lu_ref_del(&lock->cll_reference, "upcall", lock);
581 cl_lock_put(env, lock);
583 cl_env_nested_put(&nest, env);
585 /* should never happen, similar to osc_ldlm_blocking_ast(). */
591 * Core of osc_dlm_blocking_ast() logic.
593 static void osc_lock_blocking(const struct lu_env *env,
594 struct ldlm_lock *dlmlock,
595 struct osc_lock *olck, int blocking)
597 struct cl_lock *lock = olck->ols_cl.cls_lock;
599 LASSERT(olck->ols_lock == dlmlock);
600 CLASSERT(OLS_BLOCKED < OLS_CANCELLED);
601 LASSERT(!osc_lock_is_lockless(olck));
604 * Lock might be still addref-ed here, if e.g., blocking ast
605 * is sent for a failed lock.
607 osc_lock_unhold(olck);
609 if (blocking && olck->ols_state < OLS_BLOCKED)
611 * Move osc_lock into OLS_BLOCKED before canceling the lock,
612 * because it recursively re-enters osc_lock_blocking(), with
613 * the state set to OLS_CANCELLED.
615 olck->ols_state = OLS_BLOCKED;
617 * cancel and destroy lock at least once no matter how blocking ast is
618 * entered (see comment above osc_ldlm_blocking_ast() for use
619 * cases). cl_lock_cancel() and cl_lock_delete() are idempotent.
621 cl_lock_cancel(env, lock);
622 cl_lock_delete(env, lock);
626 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
627 * and ldlm_lock caches.
629 static int osc_dlm_blocking_ast0(const struct lu_env *env,
630 struct ldlm_lock *dlmlock,
631 void *data, int flag)
633 struct osc_lock *olck;
634 struct cl_lock *lock;
638 LASSERT(flag == LDLM_CB_BLOCKING || flag == LDLM_CB_CANCELING);
641 olck = osc_ast_data_get(dlmlock);
643 lock = olck->ols_cl.cls_lock;
644 cl_lock_mutex_get(env, lock);
645 LINVRNT(osc_lock_invariant(olck));
646 if (olck->ols_ast_wait) {
647 /* wake up osc_lock_use() */
648 cl_lock_signal(env, lock);
649 olck->ols_ast_wait = 0;
652 * Lock might have been canceled while this thread was
653 * sleeping for lock mutex, but olck is pinned in memory.
655 if (olck == dlmlock->l_ast_data) {
657 * NOTE: DLM sends blocking AST's for failed locks
658 * (that are still in pre-OLS_GRANTED state)
659 * too, and they have to be canceled otherwise
660 * DLM lock is never destroyed and stuck in
663 * Alternatively, ldlm_cli_cancel() can be
664 * called here directly for osc_locks with
665 * ols_state < OLS_GRANTED to maintain an
666 * invariant that ->clo_cancel() is only called
667 * for locks that were granted.
669 LASSERT(data == olck);
670 osc_lock_blocking(env, dlmlock,
671 olck, flag == LDLM_CB_BLOCKING);
674 cl_lock_mutex_put(env, lock);
675 osc_ast_data_put(env, olck);
678 * DLM lock exists, but there is no cl_lock attached to it.
679 * This is a `normal' race. cl_object and its cl_lock's can be
680 * removed by memory pressure, together with all pages.
682 cancel = (flag == LDLM_CB_BLOCKING);
685 struct lustre_handle *lockh;
687 lockh = &osc_env_info(env)->oti_handle;
688 ldlm_lock2handle(dlmlock, lockh);
689 result = ldlm_cli_cancel(lockh);
696 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
697 * some other lock, or is canceled. This function is installed as a
698 * ldlm_lock::l_blocking_ast() for client extent locks.
700 * Control flow is tricky, because ldlm uses the same call-back
701 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
703 * \param dlmlock lock for which ast occurred.
705 * \param new description of a conflicting lock in case of blocking ast.
707 * \param data value of dlmlock->l_ast_data
709 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
710 * cancellation and blocking ast's.
712 * Possible use cases:
714 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
715 * lock due to lock lru pressure, or explicit user request to purge
718 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
719 * us that dlmlock conflicts with another lock that some client is
720 * enqueing. Lock is canceled.
722 * - cl_lock_cancel() is called. osc_lock_cancel() calls
723 * ldlm_cli_cancel() that calls
725 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
727 * recursively entering osc_ldlm_blocking_ast().
729 * - client cancels lock voluntary (e.g., as a part of early cancellation):
732 * osc_lock_cancel()->
733 * ldlm_cli_cancel()->
734 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
737 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
738 struct ldlm_lock_desc *new, void *data,
742 struct cl_env_nest nest;
746 * This can be called in the context of outer IO, e.g.,
749 * ->osc_enqueue_base()->...
750 * ->ldlm_prep_elc_req()->...
751 * ->ldlm_cancel_callback()->...
752 * ->osc_ldlm_blocking_ast()
754 * new environment has to be created to not corrupt outer context.
756 env = cl_env_nested_get(&nest);
758 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
759 cl_env_nested_put(&nest, env);
761 result = PTR_ERR(env);
763 * XXX This should never happen, as cl_lock is
764 * stuck. Pre-allocated environment a la vvp_inode_fini_env
770 if (result == -ENODATA)
773 CERROR("BAST failed: %d\n", result);
778 static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
779 int flags, void *data)
781 struct cl_env_nest nest;
783 struct osc_lock *olck;
784 struct cl_lock *lock;
788 /* first, do dlm part of the work */
789 dlmrc = ldlm_completion_ast_async(dlmlock, flags, data);
790 /* then, notify cl_lock */
791 env = cl_env_nested_get(&nest);
793 olck = osc_ast_data_get(dlmlock);
795 lock = olck->ols_cl.cls_lock;
796 cl_lock_mutex_get(env, lock);
798 * ldlm_handle_cp_callback() copied LVB from request
799 * to lock->l_lvb_data, store it in osc_lock.
801 LASSERT(dlmlock->l_lvb_data != NULL);
802 lock_res_and_lock(dlmlock);
803 olck->ols_lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
804 if (olck->ols_lock == NULL) {
806 * upcall (osc_lock_upcall()) hasn't yet been
807 * called. Do nothing now, upcall will bind
808 * olck to dlmlock and signal the waiters.
810 * This maintains an invariant that osc_lock
811 * and ldlm_lock are always bound when
812 * osc_lock is in OLS_GRANTED state.
814 } else if (dlmlock->l_granted_mode ==
815 dlmlock->l_req_mode) {
816 osc_lock_granted(env, olck, dlmlock, dlmrc);
818 unlock_res_and_lock(dlmlock);
821 CL_LOCK_DEBUG(D_ERROR, env, lock,
822 "dlmlock returned %d\n", dlmrc);
823 cl_lock_error(env, lock, dlmrc);
825 cl_lock_mutex_put(env, lock);
826 osc_ast_data_put(env, olck);
829 result = -ELDLM_NO_LOCK_DATA;
830 cl_env_nested_put(&nest, env);
832 result = PTR_ERR(env);
833 return dlmrc ?: result;
836 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
838 struct ptlrpc_request *req = data;
839 struct osc_lock *olck;
840 struct cl_lock *lock;
841 struct cl_object *obj;
842 struct cl_env_nest nest;
845 struct req_capsule *cap;
848 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
850 env = cl_env_nested_get(&nest);
853 * osc_ast_data_get() has to go after environment is
854 * allocated, because osc_ast_data() acquires a
855 * reference to a lock, and it can only be released in
858 olck = osc_ast_data_get(dlmlock);
860 lock = olck->ols_cl.cls_lock;
861 cl_lock_mutex_get(env, lock);
863 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
864 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
866 result = req_capsule_server_pack(cap);
868 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
869 obj = lock->cll_descr.cld_obj;
870 result = cl_object_glimpse(env, obj, lvb);
872 cl_lock_mutex_put(env, lock);
873 osc_ast_data_put(env, olck);
876 * These errors are normal races, so we don't want to
877 * fill the console with messages by calling
880 lustre_pack_reply(req, 1, NULL, NULL);
881 result = -ELDLM_NO_LOCK_DATA;
883 cl_env_nested_put(&nest, env);
885 result = PTR_ERR(env);
886 req->rq_status = result;
890 static unsigned long osc_lock_weigh(const struct lu_env *env,
891 const struct cl_lock_slice *slice)
894 * don't need to grab coh_page_guard since we don't care the exact #
897 return cl_object_header(slice->cls_obj)->coh_pages;
901 * Get the weight of dlm lock for early cancellation.
903 * XXX: it should return the pages covered by this \a dlmlock.
905 static unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
907 struct cl_env_nest nest;
909 struct osc_lock *lock;
911 unsigned long weight;
916 * osc_ldlm_weigh_ast has a complex context since it might be called
917 * because of lock canceling, or from user's input. We have to make
918 * a new environment for it. Probably it is implementation safe to use
919 * the upper context because cl_lock_put don't modify environment
920 * variables. But in case of ..
922 env = cl_env_nested_get(&nest);
924 /* Mostly because lack of memory, tend to eliminate this lock*/
927 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
928 lock = osc_ast_data_get(dlmlock);
930 /* cl_lock was destroyed because of memory pressure.
931 * It is much reasonable to assign this type of lock
934 GOTO(out, weight = 0);
937 cll = lock->ols_cl.cls_lock;
938 cl_lock_mutex_get(env, cll);
939 weight = cl_lock_weigh(env, cll);
940 cl_lock_mutex_put(env, cll);
941 osc_ast_data_put(env, lock);
945 cl_env_nested_put(&nest, env);
949 static void osc_lock_build_einfo(const struct lu_env *env,
950 const struct cl_lock *clock,
951 struct osc_lock *lock,
952 struct ldlm_enqueue_info *einfo)
954 enum cl_lock_mode mode;
956 mode = clock->cll_descr.cld_mode;
957 if (mode == CLM_PHANTOM)
959 * For now, enqueue all glimpse locks in read mode. In the
960 * future, client might choose to enqueue LCK_PW lock for
961 * glimpse on a file opened for write.
965 einfo->ei_type = LDLM_EXTENT;
966 einfo->ei_mode = osc_cl_lock2ldlm(mode);
967 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
968 einfo->ei_cb_cp = osc_ldlm_completion_ast;
969 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
970 einfo->ei_cb_wg = osc_ldlm_weigh_ast;
971 einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */
975 * Determine if the lock should be converted into a lockless lock.
978 * - if the lock has an explicite requirment for a non-lockless lock;
979 * - if the io lock request type ci_lockreq;
980 * - send the enqueue rpc to ost to make the further decision;
981 * - special treat to truncate lockless lock
983 * Additional policy can be implemented here, e.g., never do lockless-io
986 static void osc_lock_to_lockless(const struct lu_env *env,
987 struct osc_lock *ols, int force)
989 struct cl_lock_slice *slice = &ols->ols_cl;
990 struct cl_lock *lock = slice->cls_lock;
992 LASSERT(ols->ols_state == OLS_NEW ||
993 ols->ols_state == OLS_UPCALL_RECEIVED);
996 ols->ols_locklessable = 1;
997 LASSERT(cl_lock_is_mutexed(lock));
998 slice->cls_ops = &osc_lock_lockless_ops;
1000 struct osc_io *oio = osc_env_io(env);
1001 struct cl_io *io = oio->oi_cl.cis_io;
1002 struct cl_object *obj = slice->cls_obj;
1003 struct osc_object *oob = cl2osc(obj);
1004 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
1005 struct obd_connect_data *ocd;
1007 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
1008 io->ci_lockreq == CILR_MAYBE ||
1009 io->ci_lockreq == CILR_NEVER);
1011 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
1012 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
1013 (io->ci_lockreq == CILR_MAYBE) &&
1014 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
1015 if (io->ci_lockreq == CILR_NEVER ||
1017 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
1018 /* lockless truncate */
1019 (cl_io_is_trunc(io) &&
1020 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
1021 osd->od_lockless_truncate)) {
1022 ols->ols_locklessable = 1;
1023 slice->cls_ops = &osc_lock_lockless_ops;
1026 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1029 static int osc_lock_compatible(const struct osc_lock *qing,
1030 const struct osc_lock *qed)
1032 enum cl_lock_mode qing_mode;
1033 enum cl_lock_mode qed_mode;
1035 qing_mode = qing->ols_cl.cls_lock->cll_descr.cld_mode;
1036 if (qed->ols_glimpse &&
1037 (qed->ols_state >= OLS_UPCALL_RECEIVED || qing_mode == CLM_READ))
1040 qed_mode = qed->ols_cl.cls_lock->cll_descr.cld_mode;
1041 return ((qing_mode == CLM_READ) && (qed_mode == CLM_READ));
1045 * Cancel all conflicting locks and wait for them to be destroyed.
1047 * This function is used for two purposes:
1049 * - early cancel all conflicting locks before starting IO, and
1051 * - guarantee that pages added to the page cache by lockless IO are never
1052 * covered by locks other than lockless IO lock, and, hence, are not
1053 * visible to other threads.
1055 static int osc_lock_enqueue_wait(const struct lu_env *env,
1056 const struct osc_lock *olck)
1058 struct cl_lock *lock = olck->ols_cl.cls_lock;
1059 struct cl_lock_descr *descr = &lock->cll_descr;
1060 struct cl_object_header *hdr = cl_object_header(descr->cld_obj);
1061 struct cl_lock *scan;
1062 struct cl_lock *conflict= NULL;
1063 int lockless = osc_lock_is_lockless(olck);
1067 LASSERT(cl_lock_is_mutexed(lock));
1069 /* make it enqueue anyway for glimpse lock, because we actually
1070 * don't need to cancel any conflicting locks. */
1071 if (olck->ols_glimpse)
1074 cfs_spin_lock(&hdr->coh_lock_guard);
1075 cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
1076 struct cl_lock_descr *cld = &scan->cll_descr;
1077 const struct osc_lock *scan_ols;
1082 if (scan->cll_state < CLS_QUEUING ||
1083 scan->cll_state == CLS_FREEING ||
1084 cld->cld_start > descr->cld_end ||
1085 cld->cld_end < descr->cld_start)
1088 /* overlapped and living locks. */
1090 /* We're not supposed to give up group lock. */
1091 if (scan->cll_descr.cld_mode == CLM_GROUP) {
1092 LASSERT(descr->cld_mode != CLM_GROUP ||
1093 descr->cld_gid != scan->cll_descr.cld_gid);
1097 scan_ols = osc_lock_at(scan);
1099 /* We need to cancel the compatible locks if we're enqueuing
1100 * a lockless lock, for example:
1101 * imagine that client has PR lock on [0, 1000], and thread T0
1102 * is doing lockless IO in [500, 1500] region. Concurrent
1103 * thread T1 can see lockless data in [500, 1000], which is
1104 * wrong, because these data are possibly stale. */
1105 if (!lockless && osc_lock_compatible(olck, scan_ols))
1108 /* Now @scan is conflicting with @lock, this means current
1109 * thread have to sleep for @scan being destroyed. */
1110 if (scan_ols->ols_owner == osc_env_io(env)) {
1111 CERROR("DEADLOCK POSSIBLE!\n");
1112 CL_LOCK_DEBUG(D_ERROR, env, scan, "queued.\n");
1113 CL_LOCK_DEBUG(D_ERROR, env, lock, "queuing.\n");
1114 libcfs_debug_dumpstack(NULL);
1116 cl_lock_get_trust(scan);
1120 cfs_spin_unlock(&hdr->coh_lock_guard);
1123 if (lock->cll_descr.cld_mode == CLM_GROUP) {
1124 /* we want a group lock but a previous lock request
1125 * conflicts, we do not wait but return 0 so the
1126 * request is send to the server
1128 CDEBUG(D_DLMTRACE, "group lock %p is conflicted "
1129 "with %p, no wait, send to server\n",
1131 cl_lock_put(env, conflict);
1134 CDEBUG(D_DLMTRACE, "lock %p is conflicted with %p, "
1137 LASSERT(lock->cll_conflict == NULL);
1138 lu_ref_add(&conflict->cll_reference, "cancel-wait",
1140 lock->cll_conflict = conflict;
1148 * Implementation of cl_lock_operations::clo_enqueue() method for osc
1149 * layer. This initiates ldlm enqueue:
1151 * - cancels conflicting locks early (osc_lock_enqueue_wait());
1153 * - calls osc_enqueue_base() to do actual enqueue.
1155 * osc_enqueue_base() is supplied with an upcall function that is executed
1156 * when lock is received either after a local cached ldlm lock is matched, or
1157 * when a reply from the server is received.
1159 * This function does not wait for the network communication to complete.
1161 static int osc_lock_enqueue(const struct lu_env *env,
1162 const struct cl_lock_slice *slice,
1163 struct cl_io *unused, __u32 enqflags)
1165 struct osc_lock *ols = cl2osc_lock(slice);
1166 struct cl_lock *lock = ols->ols_cl.cls_lock;
1170 LASSERT(cl_lock_is_mutexed(lock));
1171 LASSERTF(ols->ols_state == OLS_NEW,
1172 "Impossible state: %d\n", ols->ols_state);
1174 ols->ols_flags = osc_enq2ldlm_flags(enqflags);
1175 if (enqflags & CEF_AGL) {
1176 ols->ols_flags |= LDLM_FL_BLOCK_NOWAIT;
1179 if (ols->ols_flags & LDLM_FL_HAS_INTENT)
1180 ols->ols_glimpse = 1;
1181 if (!osc_lock_is_lockless(ols) && !(enqflags & CEF_MUST))
1182 /* try to convert this lock to a lockless lock */
1183 osc_lock_to_lockless(env, ols, (enqflags & CEF_NEVER));
1185 result = osc_lock_enqueue_wait(env, ols);
1187 if (!osc_lock_is_lockless(ols)) {
1188 struct osc_object *obj = cl2osc(slice->cls_obj);
1189 struct osc_thread_info *info = osc_env_info(env);
1190 struct ldlm_res_id *resname = &info->oti_resname;
1191 ldlm_policy_data_t *policy = &info->oti_policy;
1192 struct ldlm_enqueue_info *einfo = &ols->ols_einfo;
1194 if (ols->ols_locklessable)
1195 ols->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1197 /* a reference for lock, passed as an upcall cookie */
1199 lu_ref_add(&lock->cll_reference, "upcall", lock);
1200 /* a user for lock also */
1201 cl_lock_user_add(env, lock);
1202 ols->ols_state = OLS_ENQUEUED;
1205 * XXX: this is possible blocking point as
1206 * ldlm_lock_match(LDLM_FL_LVB_READY) waits for
1209 osc_lock_build_res(env, obj, resname);
1210 osc_lock_build_policy(env, lock, policy);
1211 result = osc_enqueue_base(osc_export(obj), resname,
1212 &ols->ols_flags, policy,
1214 obj->oo_oinfo->loi_kms_valid,
1216 ols, einfo, &ols->ols_handle,
1217 PTLRPCD_SET, 1, ols->ols_agl);
1219 cl_lock_user_del(env, lock);
1220 lu_ref_del(&lock->cll_reference,
1222 cl_lock_put(env, lock);
1223 if (unlikely(result == -ECANCELED)) {
1224 ols->ols_state = OLS_NEW;
1229 ols->ols_state = OLS_GRANTED;
1230 ols->ols_owner = osc_env_io(env);
1233 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1237 static int osc_lock_wait(const struct lu_env *env,
1238 const struct cl_lock_slice *slice)
1240 struct osc_lock *olck = cl2osc_lock(slice);
1241 struct cl_lock *lock = olck->ols_cl.cls_lock;
1243 LINVRNT(osc_lock_invariant(olck));
1245 if (olck->ols_glimpse && olck->ols_state >= OLS_UPCALL_RECEIVED) {
1246 if (olck->ols_flags & LDLM_FL_LVB_READY) {
1248 } else if (olck->ols_agl) {
1249 olck->ols_state = OLS_NEW;
1251 LASSERT(lock->cll_error);
1252 return lock->cll_error;
1256 if (olck->ols_state == OLS_NEW) {
1257 if (lock->cll_descr.cld_enq_flags & CEF_NO_REENQUEUE) {
1262 LASSERT(olck->ols_agl);
1264 rc = osc_lock_enqueue(env, slice, NULL, CEF_ASYNC |
1269 return CLO_REENQUEUED;
1273 LASSERT(equi(olck->ols_state >= OLS_UPCALL_RECEIVED &&
1274 lock->cll_error == 0, olck->ols_lock != NULL));
1276 return lock->cll_error ?: olck->ols_state >= OLS_GRANTED ? 0 : CLO_WAIT;
1280 * An implementation of cl_lock_operations::clo_use() method that pins cached
1283 static int osc_lock_use(const struct lu_env *env,
1284 const struct cl_lock_slice *slice)
1286 struct osc_lock *olck = cl2osc_lock(slice);
1289 LASSERT(!olck->ols_hold);
1292 * Atomically check for LDLM_FL_CBPENDING and addref a lock if this
1293 * flag is not set. This protects us from a concurrent blocking ast.
1295 rc = ldlm_lock_addref_try(&olck->ols_handle, olck->ols_einfo.ei_mode);
1298 olck->ols_state = OLS_GRANTED;
1300 struct cl_lock *lock;
1303 * Lock is being cancelled somewhere within
1304 * ldlm_handle_bl_callback(): LDLM_FL_CBPENDING is already
1305 * set, but osc_ldlm_blocking_ast() hasn't yet acquired
1308 lock = slice->cls_lock;
1309 LASSERT(lock->cll_state == CLS_INTRANSIT);
1310 LASSERT(lock->cll_users > 0);
1311 /* set a flag for osc_dlm_blocking_ast0() to signal the
1313 olck->ols_ast_wait = 1;
1319 static int osc_lock_flush(struct osc_lock *ols, int discard)
1321 struct cl_lock *lock = ols->ols_cl.cls_lock;
1322 struct cl_env_nest nest;
1326 env = cl_env_nested_get(&nest);
1328 result = cl_lock_page_out(env, lock, discard);
1329 cl_env_nested_put(&nest, env);
1331 result = PTR_ERR(env);
1334 LINVRNT(!osc_lock_has_pages(ols));
1340 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1341 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1342 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1343 * with some other lock some where in the cluster. This function does the
1346 * - invalidates all pages protected by this lock (after sending dirty
1347 * ones to the server, as necessary);
1349 * - decref's underlying ldlm lock;
1351 * - cancels ldlm lock (ldlm_cli_cancel()).
1353 static void osc_lock_cancel(const struct lu_env *env,
1354 const struct cl_lock_slice *slice)
1356 struct cl_lock *lock = slice->cls_lock;
1357 struct osc_lock *olck = cl2osc_lock(slice);
1358 struct ldlm_lock *dlmlock = olck->ols_lock;
1362 LASSERT(cl_lock_is_mutexed(lock));
1363 LINVRNT(osc_lock_invariant(olck));
1365 if (dlmlock != NULL) {
1368 discard = !!(dlmlock->l_flags & LDLM_FL_DISCARD_DATA);
1369 result = osc_lock_flush(olck, discard);
1370 osc_lock_unhold(olck);
1372 lock_res_and_lock(dlmlock);
1373 /* Now that we're the only user of dlm read/write reference,
1374 * mostly the ->l_readers + ->l_writers should be zero.
1375 * However, there is a corner case.
1376 * See bug 18829 for details.*/
1377 do_cancel = (dlmlock->l_readers == 0 &&
1378 dlmlock->l_writers == 0);
1379 dlmlock->l_flags |= LDLM_FL_CBPENDING;
1380 unlock_res_and_lock(dlmlock);
1382 result = ldlm_cli_cancel(&olck->ols_handle);
1384 CL_LOCK_DEBUG(D_ERROR, env, lock,
1385 "lock %p cancel failure with error(%d)\n",
1388 olck->ols_state = OLS_CANCELLED;
1389 olck->ols_flags &= ~LDLM_FL_LVB_READY;
1390 osc_lock_detach(env, olck);
1393 #ifdef INVARIANT_CHECK
1394 static int check_cb(const struct lu_env *env, struct cl_io *io,
1395 struct cl_page *page, void *cbdata)
1397 struct cl_lock *lock = cbdata;
1399 if (lock->cll_descr.cld_mode == CLM_READ) {
1400 struct cl_lock *tmp;
1401 tmp = cl_lock_at_page(env, lock->cll_descr.cld_obj,
1404 cl_lock_put(env, tmp);
1405 return CLP_GANG_OKAY;
1409 CL_LOCK_DEBUG(D_ERROR, env, lock, "still has pages\n");
1410 CL_PAGE_DEBUG(D_ERROR, env, page, "\n");
1411 return CLP_GANG_ABORT;
1415 * Returns true iff there are pages under \a olck not protected by other
1418 static int osc_lock_has_pages(struct osc_lock *olck)
1420 struct cl_lock *lock;
1421 struct cl_lock_descr *descr;
1422 struct cl_object *obj;
1423 struct osc_object *oob;
1424 struct cl_env_nest nest;
1429 env = cl_env_nested_get(&nest);
1433 obj = olck->ols_cl.cls_obj;
1435 io = &oob->oo_debug_io;
1436 lock = olck->ols_cl.cls_lock;
1437 descr = &lock->cll_descr;
1439 cfs_mutex_lock(&oob->oo_debug_mutex);
1441 io->ci_obj = cl_object_top(obj);
1442 cl_io_init(env, io, CIT_MISC, io->ci_obj);
1444 result = cl_page_gang_lookup(env, obj, io,
1445 descr->cld_start, descr->cld_end,
1446 check_cb, (void *)lock);
1447 if (result == CLP_GANG_ABORT)
1449 if (result == CLP_GANG_RESCHED)
1451 } while (result != CLP_GANG_OKAY);
1452 cl_io_fini(env, io);
1453 cfs_mutex_unlock(&oob->oo_debug_mutex);
1454 cl_env_nested_put(&nest, env);
1456 return (result == CLP_GANG_ABORT);
1459 static int osc_lock_has_pages(struct osc_lock *olck)
1463 #endif /* INVARIANT_CHECK */
1465 static void osc_lock_delete(const struct lu_env *env,
1466 const struct cl_lock_slice *slice)
1468 struct osc_lock *olck;
1470 olck = cl2osc_lock(slice);
1471 if (olck->ols_glimpse) {
1472 LASSERT(!olck->ols_hold);
1473 LASSERT(!olck->ols_lock);
1477 LINVRNT(osc_lock_invariant(olck));
1478 LINVRNT(!osc_lock_has_pages(olck));
1480 osc_lock_unhold(olck);
1481 osc_lock_detach(env, olck);
1485 * Implements cl_lock_operations::clo_state() method for osc layer.
1487 * Maintains osc_lock::ols_owner field.
1489 * This assumes that lock always enters CLS_HELD (from some other state) in
1490 * the same IO context as one that requested the lock. This should not be a
1491 * problem, because context is by definition shared by all activity pertaining
1492 * to the same high-level IO.
1494 static void osc_lock_state(const struct lu_env *env,
1495 const struct cl_lock_slice *slice,
1496 enum cl_lock_state state)
1498 struct osc_lock *lock = cl2osc_lock(slice);
1501 * XXX multiple io contexts can use the lock at the same time.
1503 LINVRNT(osc_lock_invariant(lock));
1504 if (state == CLS_HELD && slice->cls_lock->cll_state != CLS_HELD) {
1505 struct osc_io *oio = osc_env_io(env);
1507 LASSERT(lock->ols_owner == NULL);
1508 lock->ols_owner = oio;
1509 } else if (state != CLS_HELD)
1510 lock->ols_owner = NULL;
1513 static int osc_lock_print(const struct lu_env *env, void *cookie,
1514 lu_printer_t p, const struct cl_lock_slice *slice)
1516 struct osc_lock *lock = cl2osc_lock(slice);
1519 * XXX print ldlm lock and einfo properly.
1521 (*p)(env, cookie, "%p %08x "LPX64" %d %p ",
1522 lock->ols_lock, lock->ols_flags, lock->ols_handle.cookie,
1523 lock->ols_state, lock->ols_owner);
1524 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1528 static int osc_lock_fits_into(const struct lu_env *env,
1529 const struct cl_lock_slice *slice,
1530 const struct cl_lock_descr *need,
1531 const struct cl_io *io)
1533 struct osc_lock *ols = cl2osc_lock(slice);
1535 if (need->cld_enq_flags & CEF_NEVER)
1538 if (need->cld_mode == CLM_PHANTOM) {
1540 return !(ols->ols_state > OLS_RELEASED);
1543 * Note: the QUEUED lock can't be matched here, otherwise
1544 * it might cause the deadlocks.
1546 * P1: enqueued read lock, create sublock1
1547 * P2: enqueued write lock, create sublock2(conflicted
1549 * P1: Grant read lock.
1550 * P1: enqueued glimpse lock(with holding sublock1_read),
1551 * matched with sublock2, waiting sublock2 to be granted.
1552 * But sublock2 can not be granted, because P1
1553 * will not release sublock1. Bang!
1555 if (ols->ols_state < OLS_GRANTED ||
1556 ols->ols_state > OLS_RELEASED)
1558 } else if (need->cld_enq_flags & CEF_MUST) {
1560 * If the lock hasn't ever enqueued, it can't be matched
1561 * because enqueue process brings in many information
1562 * which can be used to determine things such as lockless,
1565 if (ols->ols_state < OLS_UPCALL_RECEIVED &&
1566 ols->ols_locklessable)
1572 static const struct cl_lock_operations osc_lock_ops = {
1573 .clo_fini = osc_lock_fini,
1574 .clo_enqueue = osc_lock_enqueue,
1575 .clo_wait = osc_lock_wait,
1576 .clo_unuse = osc_lock_unuse,
1577 .clo_use = osc_lock_use,
1578 .clo_delete = osc_lock_delete,
1579 .clo_state = osc_lock_state,
1580 .clo_cancel = osc_lock_cancel,
1581 .clo_weigh = osc_lock_weigh,
1582 .clo_print = osc_lock_print,
1583 .clo_fits_into = osc_lock_fits_into,
1586 static int osc_lock_lockless_unuse(const struct lu_env *env,
1587 const struct cl_lock_slice *slice)
1589 struct osc_lock *ols = cl2osc_lock(slice);
1590 struct cl_lock *lock = slice->cls_lock;
1592 LASSERT(ols->ols_state == OLS_GRANTED);
1593 LINVRNT(osc_lock_invariant(ols));
1595 cl_lock_cancel(env, lock);
1596 cl_lock_delete(env, lock);
1600 static void osc_lock_lockless_cancel(const struct lu_env *env,
1601 const struct cl_lock_slice *slice)
1603 struct osc_lock *ols = cl2osc_lock(slice);
1606 result = osc_lock_flush(ols, 0);
1608 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1610 ols->ols_state = OLS_CANCELLED;
1613 static int osc_lock_lockless_wait(const struct lu_env *env,
1614 const struct cl_lock_slice *slice)
1616 struct osc_lock *olck = cl2osc_lock(slice);
1617 struct cl_lock *lock = olck->ols_cl.cls_lock;
1619 LINVRNT(osc_lock_invariant(olck));
1620 LASSERT(olck->ols_state >= OLS_UPCALL_RECEIVED);
1622 return lock->cll_error;
1625 static void osc_lock_lockless_state(const struct lu_env *env,
1626 const struct cl_lock_slice *slice,
1627 enum cl_lock_state state)
1629 struct osc_lock *lock = cl2osc_lock(slice);
1631 LINVRNT(osc_lock_invariant(lock));
1632 if (state == CLS_HELD) {
1633 struct osc_io *oio = osc_env_io(env);
1635 LASSERT(ergo(lock->ols_owner, lock->ols_owner == oio));
1636 lock->ols_owner = oio;
1638 /* set the io to be lockless if this lock is for io's
1640 if (cl_object_same(oio->oi_cl.cis_obj, slice->cls_obj))
1641 oio->oi_lockless = 1;
1645 static int osc_lock_lockless_fits_into(const struct lu_env *env,
1646 const struct cl_lock_slice *slice,
1647 const struct cl_lock_descr *need,
1648 const struct cl_io *io)
1650 struct osc_lock *lock = cl2osc_lock(slice);
1652 if (!(need->cld_enq_flags & CEF_NEVER))
1655 /* lockless lock should only be used by its owning io. b22147 */
1656 return (lock->ols_owner == osc_env_io(env));
1659 static const struct cl_lock_operations osc_lock_lockless_ops = {
1660 .clo_fini = osc_lock_fini,
1661 .clo_enqueue = osc_lock_enqueue,
1662 .clo_wait = osc_lock_lockless_wait,
1663 .clo_unuse = osc_lock_lockless_unuse,
1664 .clo_state = osc_lock_lockless_state,
1665 .clo_fits_into = osc_lock_lockless_fits_into,
1666 .clo_cancel = osc_lock_lockless_cancel,
1667 .clo_print = osc_lock_print
1670 int osc_lock_init(const struct lu_env *env,
1671 struct cl_object *obj, struct cl_lock *lock,
1672 const struct cl_io *unused)
1674 struct osc_lock *clk;
1677 OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, CFS_ALLOC_IO);
1679 osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo);
1680 cfs_atomic_set(&clk->ols_pageref, 0);
1681 clk->ols_state = OLS_NEW;
1682 cl_lock_slice_add(lock, &clk->ols_cl, obj, &osc_lock_ops);
1689 int osc_dlm_lock_pageref(struct ldlm_lock *dlm)
1691 struct osc_lock *olock;
1694 cfs_spin_lock(&osc_ast_guard);
1695 olock = dlm->l_ast_data;
1697 * there's a very rare race with osc_page_addref_lock(), but that
1698 * doesn't matter because in the worst case we don't cancel a lock
1699 * which we actually can, that's no harm.
1701 if (olock != NULL &&
1702 cfs_atomic_add_return(_PAGEREF_MAGIC,
1703 &olock->ols_pageref) != _PAGEREF_MAGIC) {
1704 cfs_atomic_sub(_PAGEREF_MAGIC, &olock->ols_pageref);
1707 cfs_spin_unlock(&osc_ast_guard);