4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_lock for OSC layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
39 * Author: Jinshan Xiong <jinshan.xiong@intel.com>
42 #define DEBUG_SUBSYSTEM S_OSC
45 # include <libcfs/libcfs.h>
47 # include <liblustre.h>
49 /* fid_build_reg_res_name() */
50 #include <lustre_fid.h>
52 #include "osc_cl_internal.h"
58 /*****************************************************************************
64 static const struct cl_lock_operations osc_lock_ops;
65 static const struct cl_lock_operations osc_lock_lockless_ops;
66 static void osc_lock_to_lockless(const struct lu_env *env,
67 struct osc_lock *ols, int force);
68 static bool osc_lock_has_pages(struct osc_lock *olck);
70 int osc_lock_is_lockless(const struct osc_lock *olck)
72 return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
76 * Returns a weak pointer to the ldlm lock identified by a handle. Returned
77 * pointer cannot be dereferenced, as lock is not protected from concurrent
78 * reclaim. This function is a helper for osc_lock_invariant().
80 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
82 struct ldlm_lock *lock;
84 lock = ldlm_handle2lock(handle);
91 * Invariant that has to be true all of the time.
93 static int osc_lock_invariant(struct osc_lock *ols)
95 struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle);
96 struct ldlm_lock *olock = ols->ols_lock;
97 int handle_used = lustre_handle_is_used(&ols->ols_handle);
99 if (ergo(osc_lock_is_lockless(ols),
100 ols->ols_locklessable && ols->ols_lock == NULL))
104 * If all the following "ergo"s are true, return 1, otherwise 0
106 if (! ergo(olock != NULL, handle_used))
109 if (! ergo(olock != NULL,
110 olock->l_handle.h_cookie == ols->ols_handle.cookie))
113 if (! ergo(handle_used,
114 ergo(lock != NULL && olock != NULL, lock == olock) &&
115 ergo(lock == NULL, olock == NULL)))
118 * Check that ->ols_handle and ->ols_lock are consistent, but
119 * take into account that they are set at the different time.
121 if (! ergo(ols->ols_state == OLS_CANCELLED,
122 olock == NULL && !handle_used))
125 * DLM lock is destroyed only after we have seen cancellation
128 if (! ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
129 !ldlm_is_destroyed(olock)))
132 if (! ergo(ols->ols_state == OLS_GRANTED,
134 olock->l_req_mode == olock->l_granted_mode &&
140 /*****************************************************************************
147 * Breaks a link between osc_lock and dlm_lock.
149 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
151 struct ldlm_lock *dlmlock;
153 spin_lock(&osc_ast_guard);
154 dlmlock = olck->ols_lock;
155 if (dlmlock == NULL) {
156 spin_unlock(&osc_ast_guard);
160 olck->ols_lock = NULL;
161 /* wb(); --- for all who checks (ols->ols_lock != NULL) before
162 * call to osc_lock_detach() */
163 dlmlock->l_ast_data = NULL;
164 olck->ols_handle.cookie = 0ULL;
165 spin_unlock(&osc_ast_guard);
167 lock_res_and_lock(dlmlock);
168 if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
169 struct cl_object *obj = olck->ols_cl.cls_obj;
170 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
173 cl_object_attr_lock(obj);
174 /* Must get the value under the lock to avoid possible races. */
175 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
176 /* Update the kms. Need to loop all granted locks.
177 * Not a problem for the client */
178 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
180 cl_object_attr_set(env, obj, attr, CAT_KMS);
181 cl_object_attr_unlock(obj);
183 unlock_res_and_lock(dlmlock);
185 /* release a reference taken in osc_lock_upcall0(). */
186 LASSERT(olck->ols_has_ref);
187 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
188 LDLM_LOCK_RELEASE(dlmlock);
189 olck->ols_has_ref = 0;
192 static int osc_lock_unhold(struct osc_lock *ols)
198 result = osc_cancel_base(&ols->ols_handle,
199 ols->ols_einfo.ei_mode);
204 static int osc_lock_unuse(const struct lu_env *env,
205 const struct cl_lock_slice *slice)
207 struct osc_lock *ols = cl2osc_lock(slice);
209 LINVRNT(osc_lock_invariant(ols));
211 switch (ols->ols_state) {
213 LASSERT(!ols->ols_hold);
214 LASSERT(ols->ols_agl);
216 case OLS_UPCALL_RECEIVED:
217 osc_lock_unhold(ols);
219 LASSERT(!ols->ols_hold);
220 osc_lock_detach(env, ols);
221 ols->ols_state = OLS_NEW;
224 LASSERT(!ols->ols_glimpse);
225 LASSERT(ols->ols_hold);
227 * Move lock into OLS_RELEASED state before calling
228 * osc_cancel_base() so that possible synchronous cancellation
229 * (that always happens e.g., for liblustre) sees that lock is
232 ols->ols_state = OLS_RELEASED;
233 return osc_lock_unhold(ols);
235 CERROR("Impossible state: %d\n", ols->ols_state);
240 static void osc_lock_fini(const struct lu_env *env,
241 struct cl_lock_slice *slice)
243 struct osc_lock *ols = cl2osc_lock(slice);
245 LINVRNT(osc_lock_invariant(ols));
247 * ->ols_hold can still be true at this point if, for example, a
248 * thread that requested a lock was killed (and released a reference
249 * to the lock), before reply from a server was received. In this case
250 * lock is destroyed immediately after upcall.
252 osc_lock_unhold(ols);
253 LASSERT(ols->ols_lock == NULL);
254 OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
257 static void osc_lock_build_policy(const struct lu_env *env,
258 const struct cl_lock *lock,
259 ldlm_policy_data_t *policy)
261 const struct cl_lock_descr *d = &lock->cll_descr;
263 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
264 policy->l_extent.gid = d->cld_gid;
267 static __u64 osc_enq2ldlm_flags(__u32 enqflags)
271 LASSERT((enqflags & ~CEF_MASK) == 0);
273 if (enqflags & CEF_NONBLOCK)
274 result |= LDLM_FL_BLOCK_NOWAIT;
275 if (enqflags & CEF_ASYNC)
276 result |= LDLM_FL_HAS_INTENT;
277 if (enqflags & CEF_DISCARD_DATA)
278 result |= LDLM_FL_AST_DISCARD_DATA;
283 * Global spin-lock protecting consistency of ldlm_lock::l_ast_data
284 * pointers. Initialized in osc_init().
286 spinlock_t osc_ast_guard;
288 static struct osc_lock *osc_ast_data_get(struct ldlm_lock *dlm_lock)
290 struct osc_lock *olck;
292 lock_res_and_lock(dlm_lock);
293 spin_lock(&osc_ast_guard);
294 olck = dlm_lock->l_ast_data;
296 struct cl_lock *lock = olck->ols_cl.cls_lock;
298 * If osc_lock holds a reference on ldlm lock, return it even
299 * when cl_lock is in CLS_FREEING state. This way
301 * osc_ast_data_get(dlmlock) == NULL
303 * guarantees that all osc references on dlmlock were
304 * released. osc_dlm_blocking_ast0() relies on that.
306 if (lock->cll_state < CLS_FREEING || olck->ols_has_ref) {
307 cl_lock_get_trust(lock);
308 lu_ref_add_atomic(&lock->cll_reference,
313 spin_unlock(&osc_ast_guard);
314 unlock_res_and_lock(dlm_lock);
318 static void osc_ast_data_put(const struct lu_env *env, struct osc_lock *olck)
320 struct cl_lock *lock;
322 lock = olck->ols_cl.cls_lock;
323 lu_ref_del(&lock->cll_reference, "ast", current);
324 cl_lock_put(env, lock);
328 * Updates object attributes from a lock value block (lvb) received together
329 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
332 * This can be optimized to not update attributes when lock is a result of a
335 * Called under lock and resource spin-locks.
337 static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
341 struct cl_object *obj;
342 struct lov_oinfo *oinfo;
343 struct cl_attr *attr;
348 if (!(olck->ols_flags & LDLM_FL_LVB_READY))
351 lvb = &olck->ols_lvb;
352 obj = olck->ols_cl.cls_obj;
353 oinfo = cl2osc(obj)->oo_oinfo;
354 attr = &osc_env_info(env)->oti_attr;
355 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
356 cl_lvb2attr(attr, lvb);
358 cl_object_attr_lock(obj);
360 struct ldlm_lock *dlmlock;
363 dlmlock = olck->ols_lock;
364 LASSERT(dlmlock != NULL);
366 /* re-grab LVB from a dlm lock under DLM spin-locks. */
367 *lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
368 size = lvb->lvb_size;
369 /* Extend KMS up to the end of this lock and no further
370 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
371 if (size > dlmlock->l_policy_data.l_extent.end)
372 size = dlmlock->l_policy_data.l_extent.end + 1;
373 if (size >= oinfo->loi_kms) {
374 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="LPU64
375 ", kms="LPU64, lvb->lvb_size, size);
377 attr->cat_kms = size;
379 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
380 LPU64"; leaving kms="LPU64", end="LPU64,
381 lvb->lvb_size, oinfo->loi_kms,
382 dlmlock->l_policy_data.l_extent.end);
384 ldlm_lock_allow_match_locked(dlmlock);
385 } else if (rc == -ENAVAIL && olck->ols_glimpse) {
386 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
387 " kms="LPU64"\n", lvb->lvb_size, oinfo->loi_kms);
392 cl_object_attr_set(env, obj, attr, valid);
394 cl_object_attr_unlock(obj);
400 * Called when a lock is granted, from an upcall (when server returned a
401 * granted lock), or from completion AST, when server returned a blocked lock.
403 * Called under lock and resource spin-locks, that are released temporarily
406 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
407 struct ldlm_lock *dlmlock, int rc)
409 struct ldlm_extent *ext;
410 struct cl_lock *lock;
411 struct cl_lock_descr *descr;
413 LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
416 if (olck->ols_state < OLS_GRANTED) {
417 lock = olck->ols_cl.cls_lock;
418 ext = &dlmlock->l_policy_data.l_extent;
419 descr = &osc_env_info(env)->oti_descr;
420 descr->cld_obj = lock->cll_descr.cld_obj;
422 /* XXX check that ->l_granted_mode is valid. */
423 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
424 descr->cld_start = cl_index(descr->cld_obj, ext->start);
425 descr->cld_end = cl_index(descr->cld_obj, ext->end);
426 descr->cld_gid = ext->gid;
428 * tell upper layers the extent of the lock that was actually
431 olck->ols_state = OLS_GRANTED;
432 osc_lock_lvb_update(env, olck, rc);
434 /* release DLM spin-locks to allow cl_lock_{modify,signal}()
435 * to take a semaphore on a parent lock. This is safe, because
436 * spin-locks are needed to protect consistency of
437 * dlmlock->l_*_mode and LVB, and we have finished processing
439 unlock_res_and_lock(dlmlock);
440 cl_lock_modify(env, lock, descr);
441 cl_lock_signal(env, lock);
442 LINVRNT(osc_lock_invariant(olck));
443 lock_res_and_lock(dlmlock);
448 static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
451 struct ldlm_lock *dlmlock;
455 dlmlock = ldlm_handle2lock_long(&olck->ols_handle, 0);
456 LASSERT(dlmlock != NULL);
458 lock_res_and_lock(dlmlock);
459 spin_lock(&osc_ast_guard);
460 LASSERT(dlmlock->l_ast_data == olck);
461 LASSERT(olck->ols_lock == NULL);
462 olck->ols_lock = dlmlock;
463 spin_unlock(&osc_ast_guard);
466 * Lock might be not yet granted. In this case, completion ast
467 * (osc_ldlm_completion_ast()) comes later and finishes lock
470 if (dlmlock->l_granted_mode == dlmlock->l_req_mode)
471 osc_lock_granted(env, olck, dlmlock, 0);
472 unlock_res_and_lock(dlmlock);
475 * osc_enqueue_interpret() decrefs asynchronous locks, counter
478 ldlm_lock_addref(&olck->ols_handle, olck->ols_einfo.ei_mode);
481 /* lock reference taken by ldlm_handle2lock_long() is owned by
482 * osc_lock and released in osc_lock_detach() */
483 lu_ref_add(&dlmlock->l_reference, "osc_lock", olck);
484 olck->ols_has_ref = 1;
488 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
489 * received from a server, or after osc_enqueue_base() matched a local DLM
492 static int osc_lock_upcall(void *cookie, int errcode)
494 struct osc_lock *olck = cookie;
495 struct cl_lock_slice *slice = &olck->ols_cl;
496 struct cl_lock *lock = slice->cls_lock;
498 struct cl_env_nest nest;
501 env = cl_env_nested_get(&nest);
505 cl_lock_mutex_get(env, lock);
507 LASSERT(lock->cll_state >= CLS_QUEUING);
508 if (olck->ols_state == OLS_ENQUEUED) {
509 olck->ols_state = OLS_UPCALL_RECEIVED;
510 rc = ldlm_error2errno(errcode);
511 } else if (olck->ols_state == OLS_CANCELLED) {
514 CERROR("Impossible state: %d\n", olck->ols_state);
518 struct ldlm_lock *dlmlock;
520 dlmlock = ldlm_handle2lock(&olck->ols_handle);
521 if (dlmlock != NULL) {
522 lock_res_and_lock(dlmlock);
523 spin_lock(&osc_ast_guard);
524 LASSERT(olck->ols_lock == NULL);
525 dlmlock->l_ast_data = NULL;
526 olck->ols_handle.cookie = 0ULL;
527 spin_unlock(&osc_ast_guard);
528 ldlm_lock_fail_match_locked(dlmlock);
529 unlock_res_and_lock(dlmlock);
530 LDLM_LOCK_PUT(dlmlock);
533 if (olck->ols_glimpse)
534 olck->ols_glimpse = 0;
535 osc_lock_upcall0(env, olck);
538 /* Error handling, some errors are tolerable. */
539 if (olck->ols_locklessable && rc == -EUSERS) {
540 /* This is a tolerable error, turn this lock into
543 osc_object_set_contended(cl2osc(slice->cls_obj));
544 LASSERT(slice->cls_ops == &osc_lock_ops);
546 /* Change this lock to ldlmlock-less lock. */
547 osc_lock_to_lockless(env, olck, 1);
548 olck->ols_state = OLS_GRANTED;
550 } else if (olck->ols_glimpse && rc == -ENAVAIL) {
551 osc_lock_lvb_update(env, olck, rc);
552 cl_lock_delete(env, lock);
553 /* Hide the error. */
558 /* For AGL case, the RPC sponsor may exits the cl_lock
559 * processing without wait() called before related OSC
560 * lock upcall(). So update the lock status according
561 * to the enqueue result inside AGL upcall(). */
563 lock->cll_flags |= CLF_FROM_UPCALL;
564 cl_wait_try(env, lock);
565 lock->cll_flags &= ~CLF_FROM_UPCALL;
567 cl_lock_signal(env, lock);
568 /* del user for lock upcall cookie */
570 if (!olck->ols_glimpse)
572 cl_unuse_try(env, lock);
575 /* del user for lock upcall cookie */
577 cl_lock_user_del(env, lock);
578 cl_lock_error(env, lock, rc);
581 /* release cookie reference, acquired by osc_lock_enqueue() */
582 cl_lock_hold_release(env, lock, "upcall", lock);
583 cl_lock_mutex_put(env, lock);
585 lu_ref_del(&lock->cll_reference, "upcall", lock);
586 /* This maybe the last reference, so must be called after
587 * cl_lock_mutex_put(). */
588 cl_lock_put(env, lock);
590 cl_env_nested_put(&nest, env);
592 /* should never happen, similar to osc_ldlm_blocking_ast(). */
599 * Core of osc_dlm_blocking_ast() logic.
601 static void osc_lock_blocking(const struct lu_env *env,
602 struct ldlm_lock *dlmlock,
603 struct osc_lock *olck, int blocking)
605 struct cl_lock *lock = olck->ols_cl.cls_lock;
607 LASSERT(olck->ols_lock == dlmlock);
608 CLASSERT(OLS_BLOCKED < OLS_CANCELLED);
609 LASSERT(!osc_lock_is_lockless(olck));
612 * Lock might be still addref-ed here, if e.g., blocking ast
613 * is sent for a failed lock.
615 osc_lock_unhold(olck);
617 if (blocking && olck->ols_state < OLS_BLOCKED)
619 * Move osc_lock into OLS_BLOCKED before canceling the lock,
620 * because it recursively re-enters osc_lock_blocking(), with
621 * the state set to OLS_CANCELLED.
623 olck->ols_state = OLS_BLOCKED;
625 * cancel and destroy lock at least once no matter how blocking ast is
626 * entered (see comment above osc_ldlm_blocking_ast() for use
627 * cases). cl_lock_cancel() and cl_lock_delete() are idempotent.
629 cl_lock_cancel(env, lock);
630 cl_lock_delete(env, lock);
634 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
635 * and ldlm_lock caches.
637 static int osc_dlm_blocking_ast0(const struct lu_env *env,
638 struct ldlm_lock *dlmlock,
639 void *data, int flag)
641 struct osc_lock *olck;
642 struct cl_lock *lock;
646 LASSERT(flag == LDLM_CB_BLOCKING || flag == LDLM_CB_CANCELING);
649 olck = osc_ast_data_get(dlmlock);
651 lock = olck->ols_cl.cls_lock;
652 cl_lock_mutex_get(env, lock);
653 LINVRNT(osc_lock_invariant(olck));
654 if (olck->ols_ast_wait) {
655 /* wake up osc_lock_use() */
656 cl_lock_signal(env, lock);
657 olck->ols_ast_wait = 0;
660 * Lock might have been canceled while this thread was
661 * sleeping for lock mutex, but olck is pinned in memory.
663 if (olck == dlmlock->l_ast_data) {
665 * NOTE: DLM sends blocking AST's for failed locks
666 * (that are still in pre-OLS_GRANTED state)
667 * too, and they have to be canceled otherwise
668 * DLM lock is never destroyed and stuck in
671 * Alternatively, ldlm_cli_cancel() can be
672 * called here directly for osc_locks with
673 * ols_state < OLS_GRANTED to maintain an
674 * invariant that ->clo_cancel() is only called
675 * for locks that were granted.
677 LASSERT(data == olck);
678 osc_lock_blocking(env, dlmlock,
679 olck, flag == LDLM_CB_BLOCKING);
682 cl_lock_mutex_put(env, lock);
683 osc_ast_data_put(env, olck);
686 * DLM lock exists, but there is no cl_lock attached to it.
687 * This is a `normal' race. cl_object and its cl_lock's can be
688 * removed by memory pressure, together with all pages.
690 cancel = (flag == LDLM_CB_BLOCKING);
693 struct lustre_handle *lockh;
695 lockh = &osc_env_info(env)->oti_handle;
696 ldlm_lock2handle(dlmlock, lockh);
697 result = ldlm_cli_cancel(lockh, LCF_ASYNC);
704 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
705 * some other lock, or is canceled. This function is installed as a
706 * ldlm_lock::l_blocking_ast() for client extent locks.
708 * Control flow is tricky, because ldlm uses the same call-back
709 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
711 * \param dlmlock lock for which ast occurred.
713 * \param new description of a conflicting lock in case of blocking ast.
715 * \param data value of dlmlock->l_ast_data
717 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
718 * cancellation and blocking ast's.
720 * Possible use cases:
722 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
723 * lock due to lock lru pressure, or explicit user request to purge
726 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
727 * us that dlmlock conflicts with another lock that some client is
728 * enqueing. Lock is canceled.
730 * - cl_lock_cancel() is called. osc_lock_cancel() calls
731 * ldlm_cli_cancel() that calls
733 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
735 * recursively entering osc_ldlm_blocking_ast().
737 * - client cancels lock voluntary (e.g., as a part of early cancellation):
740 * osc_lock_cancel()->
741 * ldlm_cli_cancel()->
742 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
745 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
746 struct ldlm_lock_desc *new, void *data,
750 struct cl_env_nest nest;
754 * This can be called in the context of outer IO, e.g.,
757 * ->osc_enqueue_base()->...
758 * ->ldlm_prep_elc_req()->...
759 * ->ldlm_cancel_callback()->...
760 * ->osc_ldlm_blocking_ast()
762 * new environment has to be created to not corrupt outer context.
764 env = cl_env_nested_get(&nest);
766 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
767 cl_env_nested_put(&nest, env);
769 result = PTR_ERR(env);
771 * XXX This should never happen, as cl_lock is
772 * stuck. Pre-allocated environment a la vvp_inode_fini_env
778 if (result == -ENODATA)
781 CERROR("BAST failed: %d\n", result);
786 static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
787 __u64 flags, void *data)
789 struct cl_env_nest nest;
791 struct osc_lock *olck;
792 struct cl_lock *lock;
796 /* first, do dlm part of the work */
797 dlmrc = ldlm_completion_ast_async(dlmlock, flags, data);
798 if (flags == LDLM_FL_WAIT_NOREPROC)
801 /* then, notify cl_lock */
802 env = cl_env_nested_get(&nest);
804 olck = osc_ast_data_get(dlmlock);
806 lock = olck->ols_cl.cls_lock;
807 cl_lock_mutex_get(env, lock);
809 * ldlm_handle_cp_callback() copied LVB from request
810 * to lock->l_lvb_data, store it in osc_lock.
812 LASSERT(dlmlock->l_lvb_data != NULL);
813 lock_res_and_lock(dlmlock);
814 olck->ols_lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
815 if (olck->ols_lock == NULL) {
817 * upcall (osc_lock_upcall()) hasn't yet been
818 * called. Do nothing now, upcall will bind
819 * olck to dlmlock and signal the waiters.
821 * This maintains an invariant that osc_lock
822 * and ldlm_lock are always bound when
823 * osc_lock is in OLS_GRANTED state.
825 } else if (dlmlock->l_granted_mode ==
826 dlmlock->l_req_mode) {
827 osc_lock_granted(env, olck, dlmlock, dlmrc);
829 unlock_res_and_lock(dlmlock);
832 CL_LOCK_DEBUG(D_ERROR, env, lock,
833 "dlmlock returned %d\n", dlmrc);
834 cl_lock_error(env, lock, dlmrc);
836 cl_lock_mutex_put(env, lock);
837 osc_ast_data_put(env, olck);
840 result = -ELDLM_NO_LOCK_DATA;
841 cl_env_nested_put(&nest, env);
843 result = PTR_ERR(env);
844 return dlmrc ?: result;
847 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
849 struct ptlrpc_request *req = data;
850 struct osc_lock *olck;
851 struct cl_lock *lock;
852 struct cl_object *obj;
853 struct cl_env_nest nest;
856 struct req_capsule *cap;
859 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
861 env = cl_env_nested_get(&nest);
863 /* osc_ast_data_get() has to go after environment is
864 * allocated, because osc_ast_data() acquires a
865 * reference to a lock, and it can only be released in
868 olck = osc_ast_data_get(dlmlock);
870 lock = olck->ols_cl.cls_lock;
871 /* Do not grab the mutex of cl_lock for glimpse.
872 * See LU-1274 for details.
873 * BTW, it's okay for cl_lock to be cancelled during
874 * this period because server can handle this race.
875 * See ldlm_server_glimpse_ast() for details.
876 * cl_lock_mutex_get(env, lock); */
878 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
879 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
881 result = req_capsule_server_pack(cap);
883 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
884 obj = lock->cll_descr.cld_obj;
885 result = cl_object_glimpse(env, obj, lvb);
887 if (!exp_connect_lvb_type(req->rq_export))
888 req_capsule_shrink(&req->rq_pill,
890 sizeof(struct ost_lvb_v1),
892 osc_ast_data_put(env, olck);
895 * These errors are normal races, so we don't want to
896 * fill the console with messages by calling
899 lustre_pack_reply(req, 1, NULL, NULL);
900 result = -ELDLM_NO_LOCK_DATA;
902 cl_env_nested_put(&nest, env);
904 result = PTR_ERR(env);
905 req->rq_status = result;
909 static int weigh_cb(const struct lu_env *env, struct cl_io *io,
910 struct osc_page *ops, void *cbdata)
912 struct cl_page *page = ops->ops_cl.cpl_page;
914 if (cl_page_is_vmlocked(env, page)
915 #if defined(__KERNEL__)
916 || PageDirty(page->cp_vmpage) || PageWriteback(page->cp_vmpage)
919 (*(unsigned long *)cbdata)++;
920 return CLP_GANG_ABORT;
923 return CLP_GANG_OKAY;
926 static unsigned long osc_lock_weight(const struct lu_env *env,
927 const struct osc_lock *ols)
929 struct cl_io *io = &osc_env_info(env)->oti_io;
930 struct cl_lock_descr *descr = &ols->ols_cl.cls_lock->cll_descr;
931 struct cl_object *obj = ols->ols_cl.cls_obj;
932 unsigned long npages = 0;
936 io->ci_obj = cl_object_top(obj);
937 io->ci_ignore_layout = 1;
938 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
943 result = osc_page_gang_lookup(env, io, cl2osc(obj),
944 descr->cld_start, descr->cld_end,
945 weigh_cb, (void *)&npages);
946 if (result == CLP_GANG_ABORT)
948 if (result == CLP_GANG_RESCHED)
950 } while (result != CLP_GANG_OKAY);
957 * Get the weight of dlm lock for early cancellation.
959 unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
961 struct cl_env_nest nest;
963 struct osc_lock *lock;
964 unsigned long weight;
969 * osc_ldlm_weigh_ast has a complex context since it might be called
970 * because of lock canceling, or from user's input. We have to make
971 * a new environment for it. Probably it is implementation safe to use
972 * the upper context because cl_lock_put don't modify environment
973 * variables. But just in case ..
975 env = cl_env_nested_get(&nest);
977 /* Mostly because lack of memory, do not eliminate this lock */
980 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
981 lock = osc_ast_data_get(dlmlock);
983 /* cl_lock was destroyed because of memory pressure.
984 * It is much reasonable to assign this type of lock
987 GOTO(out, weight = 0);
990 weight = osc_lock_weight(env, lock);
991 osc_ast_data_put(env, lock);
995 cl_env_nested_put(&nest, env);
999 static void osc_lock_build_einfo(const struct lu_env *env,
1000 const struct cl_lock *clock,
1001 struct osc_lock *lock,
1002 struct ldlm_enqueue_info *einfo)
1004 enum cl_lock_mode mode;
1006 mode = clock->cll_descr.cld_mode;
1007 if (mode == CLM_PHANTOM)
1009 * For now, enqueue all glimpse locks in read mode. In the
1010 * future, client might choose to enqueue LCK_PW lock for
1011 * glimpse on a file opened for write.
1015 einfo->ei_type = LDLM_EXTENT;
1016 einfo->ei_mode = osc_cl_lock2ldlm(mode);
1017 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
1018 einfo->ei_cb_cp = osc_ldlm_completion_ast;
1019 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
1020 einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */
1024 * Determine if the lock should be converted into a lockless lock.
1027 * - if the lock has an explicite requirment for a non-lockless lock;
1028 * - if the io lock request type ci_lockreq;
1029 * - send the enqueue rpc to ost to make the further decision;
1030 * - special treat to truncate lockless lock
1032 * Additional policy can be implemented here, e.g., never do lockless-io
1033 * for large extents.
1035 static void osc_lock_to_lockless(const struct lu_env *env,
1036 struct osc_lock *ols, int force)
1038 struct cl_lock_slice *slice = &ols->ols_cl;
1040 LASSERT(ols->ols_state == OLS_NEW ||
1041 ols->ols_state == OLS_UPCALL_RECEIVED);
1044 ols->ols_locklessable = 1;
1045 slice->cls_ops = &osc_lock_lockless_ops;
1047 struct osc_io *oio = osc_env_io(env);
1048 struct cl_io *io = oio->oi_cl.cis_io;
1049 struct cl_object *obj = slice->cls_obj;
1050 struct osc_object *oob = cl2osc(obj);
1051 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
1052 struct obd_connect_data *ocd;
1054 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
1055 io->ci_lockreq == CILR_MAYBE ||
1056 io->ci_lockreq == CILR_NEVER);
1058 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
1059 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
1060 (io->ci_lockreq == CILR_MAYBE) &&
1061 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
1062 if (io->ci_lockreq == CILR_NEVER ||
1064 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
1065 /* lockless truncate */
1066 (cl_io_is_trunc(io) &&
1067 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
1068 osd->od_lockless_truncate)) {
1069 ols->ols_locklessable = 1;
1070 slice->cls_ops = &osc_lock_lockless_ops;
1073 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1076 static int osc_lock_compatible(const struct osc_lock *qing,
1077 const struct osc_lock *qed)
1079 enum cl_lock_mode qing_mode;
1080 enum cl_lock_mode qed_mode;
1082 qing_mode = qing->ols_cl.cls_lock->cll_descr.cld_mode;
1083 if (qed->ols_glimpse &&
1084 (qed->ols_state >= OLS_UPCALL_RECEIVED || qing_mode == CLM_READ))
1087 qed_mode = qed->ols_cl.cls_lock->cll_descr.cld_mode;
1088 return ((qing_mode == CLM_READ) && (qed_mode == CLM_READ));
1092 * Cancel all conflicting locks and wait for them to be destroyed.
1094 * This function is used for two purposes:
1096 * - early cancel all conflicting locks before starting IO, and
1098 * - guarantee that pages added to the page cache by lockless IO are never
1099 * covered by locks other than lockless IO lock, and, hence, are not
1100 * visible to other threads.
1102 static int osc_lock_enqueue_wait(const struct lu_env *env,
1103 const struct osc_lock *olck)
1105 struct cl_lock *lock = olck->ols_cl.cls_lock;
1106 struct cl_lock_descr *descr = &lock->cll_descr;
1107 struct cl_object_header *hdr = cl_object_header(descr->cld_obj);
1108 struct cl_lock *scan;
1109 struct cl_lock *conflict= NULL;
1110 int lockless = osc_lock_is_lockless(olck);
1114 LASSERT(cl_lock_is_mutexed(lock));
1116 /* make it enqueue anyway for glimpse lock, because we actually
1117 * don't need to cancel any conflicting locks. */
1118 if (olck->ols_glimpse)
1121 spin_lock(&hdr->coh_lock_guard);
1122 list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
1123 struct cl_lock_descr *cld = &scan->cll_descr;
1124 const struct osc_lock *scan_ols;
1129 if (scan->cll_state < CLS_QUEUING ||
1130 scan->cll_state == CLS_FREEING ||
1131 cld->cld_start > descr->cld_end ||
1132 cld->cld_end < descr->cld_start)
1135 /* overlapped and living locks. */
1137 /* We're not supposed to give up group lock. */
1138 if (scan->cll_descr.cld_mode == CLM_GROUP) {
1139 LASSERT(descr->cld_mode != CLM_GROUP ||
1140 descr->cld_gid != scan->cll_descr.cld_gid);
1144 scan_ols = osc_lock_at(scan);
1146 /* We need to cancel the compatible locks if we're enqueuing
1147 * a lockless lock, for example:
1148 * imagine that client has PR lock on [0, 1000], and thread T0
1149 * is doing lockless IO in [500, 1500] region. Concurrent
1150 * thread T1 can see lockless data in [500, 1000], which is
1151 * wrong, because these data are possibly stale. */
1152 if (!lockless && osc_lock_compatible(olck, scan_ols))
1155 cl_lock_get_trust(scan);
1159 spin_unlock(&hdr->coh_lock_guard);
1162 if (lock->cll_descr.cld_mode == CLM_GROUP) {
1163 /* we want a group lock but a previous lock request
1164 * conflicts, we do not wait but return 0 so the
1165 * request is send to the server
1167 CDEBUG(D_DLMTRACE, "group lock %p is conflicted "
1168 "with %p, no wait, send to server\n",
1170 cl_lock_put(env, conflict);
1173 CDEBUG(D_DLMTRACE, "lock %p is conflicted with %p, "
1176 LASSERT(lock->cll_conflict == NULL);
1177 lu_ref_add(&conflict->cll_reference, "cancel-wait",
1179 lock->cll_conflict = conflict;
1187 * Implementation of cl_lock_operations::clo_enqueue() method for osc
1188 * layer. This initiates ldlm enqueue:
1190 * - cancels conflicting locks early (osc_lock_enqueue_wait());
1192 * - calls osc_enqueue_base() to do actual enqueue.
1194 * osc_enqueue_base() is supplied with an upcall function that is executed
1195 * when lock is received either after a local cached ldlm lock is matched, or
1196 * when a reply from the server is received.
1198 * This function does not wait for the network communication to complete.
1200 static int osc_lock_enqueue(const struct lu_env *env,
1201 const struct cl_lock_slice *slice,
1202 struct cl_io *unused, __u32 enqflags)
1204 struct osc_lock *ols = cl2osc_lock(slice);
1205 struct cl_lock *lock = ols->ols_cl.cls_lock;
1209 LASSERT(cl_lock_is_mutexed(lock));
1210 LASSERTF(ols->ols_state == OLS_NEW,
1211 "Impossible state: %d\n", ols->ols_state);
1213 LASSERTF(ergo(ols->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
1214 "lock = %p, ols = %p\n", lock, ols);
1216 result = osc_lock_enqueue_wait(env, ols);
1218 if (!osc_lock_is_lockless(ols)) {
1219 struct osc_object *obj = cl2osc(slice->cls_obj);
1220 struct osc_thread_info *info = osc_env_info(env);
1221 struct ldlm_res_id *resname = &info->oti_resname;
1222 ldlm_policy_data_t *policy = &info->oti_policy;
1223 struct ldlm_enqueue_info *einfo = &ols->ols_einfo;
1225 /* lock will be passed as upcall cookie,
1226 * hold ref to prevent to be released. */
1227 cl_lock_hold_add(env, lock, "upcall", lock);
1228 /* a user for agl lock also */
1230 cl_lock_user_add(env, lock);
1231 ols->ols_state = OLS_ENQUEUED;
1234 * XXX: this is possible blocking point as
1235 * ldlm_lock_match(LDLM_FL_LVB_READY) waits for
1238 ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1239 osc_lock_build_policy(env, lock, policy);
1240 result = osc_enqueue_base(osc_export(obj), resname,
1241 &ols->ols_flags, policy,
1243 obj->oo_oinfo->loi_kms_valid,
1245 ols, einfo, &ols->ols_handle,
1246 PTLRPCD_SET, 1, ols->ols_agl);
1249 cl_lock_user_del(env, lock);
1250 cl_lock_unhold(env, lock, "upcall", lock);
1251 if (unlikely(result == -ECANCELED)) {
1252 ols->ols_state = OLS_NEW;
1257 ols->ols_state = OLS_GRANTED;
1258 ols->ols_owner = osc_env_io(env);
1261 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1265 static int osc_lock_wait(const struct lu_env *env,
1266 const struct cl_lock_slice *slice)
1268 struct osc_lock *olck = cl2osc_lock(slice);
1269 struct cl_lock *lock = olck->ols_cl.cls_lock;
1271 LINVRNT(osc_lock_invariant(olck));
1273 if (olck->ols_glimpse && olck->ols_state >= OLS_UPCALL_RECEIVED) {
1274 if (olck->ols_flags & LDLM_FL_LVB_READY) {
1276 } else if (olck->ols_agl) {
1277 if (lock->cll_flags & CLF_FROM_UPCALL)
1278 /* It is from enqueue RPC reply upcall for
1279 * updating state. Do not re-enqueue. */
1282 olck->ols_state = OLS_NEW;
1284 LASSERT(lock->cll_error);
1285 return lock->cll_error;
1289 if (olck->ols_state == OLS_NEW) {
1292 LASSERT(olck->ols_agl);
1294 olck->ols_flags &= ~LDLM_FL_BLOCK_NOWAIT;
1295 rc = osc_lock_enqueue(env, slice, NULL, CEF_ASYNC | CEF_MUST);
1299 return CLO_REENQUEUED;
1302 LASSERT(equi(olck->ols_state >= OLS_UPCALL_RECEIVED &&
1303 lock->cll_error == 0, olck->ols_lock != NULL));
1305 return lock->cll_error ?: olck->ols_state >= OLS_GRANTED ? 0 : CLO_WAIT;
1309 * An implementation of cl_lock_operations::clo_use() method that pins cached
1312 static int osc_lock_use(const struct lu_env *env,
1313 const struct cl_lock_slice *slice)
1315 struct osc_lock *olck = cl2osc_lock(slice);
1318 LASSERT(!olck->ols_hold);
1321 * Atomically check for LDLM_FL_CBPENDING and addref a lock if this
1322 * flag is not set. This protects us from a concurrent blocking ast.
1324 rc = ldlm_lock_addref_try(&olck->ols_handle, olck->ols_einfo.ei_mode);
1327 olck->ols_state = OLS_GRANTED;
1329 struct cl_lock *lock;
1332 * Lock is being cancelled somewhere within
1333 * ldlm_handle_bl_callback(): LDLM_FL_CBPENDING is already
1334 * set, but osc_ldlm_blocking_ast() hasn't yet acquired
1337 lock = slice->cls_lock;
1338 LASSERT(lock->cll_state == CLS_INTRANSIT);
1339 LASSERT(lock->cll_users > 0);
1340 /* set a flag for osc_dlm_blocking_ast0() to signal the
1342 olck->ols_ast_wait = 1;
1348 static int osc_lock_flush(struct osc_lock *ols, int discard)
1350 struct cl_lock *lock = ols->ols_cl.cls_lock;
1351 struct cl_env_nest nest;
1356 env = cl_env_nested_get(&nest);
1358 struct osc_object *obj = cl2osc(ols->ols_cl.cls_obj);
1359 struct cl_lock_descr *descr = &lock->cll_descr;
1362 if (descr->cld_mode >= CLM_WRITE) {
1363 result = osc_cache_writeback_range(env, obj,
1364 descr->cld_start, descr->cld_end,
1366 LDLM_DEBUG(ols->ols_lock,
1367 "lock %p: %d pages were %s.\n", lock, result,
1368 discard ? "discarded" : "written");
1373 rc = osc_lock_discard_pages(env, ols);
1374 if (result == 0 && rc < 0)
1377 cl_env_nested_put(&nest, env);
1379 result = PTR_ERR(env);
1382 LINVRNT(!osc_lock_has_pages(ols));
1388 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1389 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1390 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1391 * with some other lock some where in the cluster. This function does the
1394 * - invalidates all pages protected by this lock (after sending dirty
1395 * ones to the server, as necessary);
1397 * - decref's underlying ldlm lock;
1399 * - cancels ldlm lock (ldlm_cli_cancel()).
1401 static void osc_lock_cancel(const struct lu_env *env,
1402 const struct cl_lock_slice *slice)
1404 struct cl_lock *lock = slice->cls_lock;
1405 struct osc_lock *olck = cl2osc_lock(slice);
1406 struct ldlm_lock *dlmlock = olck->ols_lock;
1408 LASSERT(cl_lock_is_mutexed(lock));
1409 LINVRNT(osc_lock_invariant(olck));
1411 if (dlmlock != NULL) {
1415 if (olck->ols_state >= OLS_GRANTED)
1416 result = osc_lock_flush(olck,
1417 ldlm_is_discard_data(dlmlock));
1418 osc_lock_unhold(olck);
1420 lock_res_and_lock(dlmlock);
1421 /* Now that we're the only user of dlm read/write reference,
1422 * mostly the ->l_readers + ->l_writers should be zero.
1423 * However, there is a corner case.
1424 * See b=18829 for details.*/
1425 do_cancel = (dlmlock->l_readers == 0 &&
1426 dlmlock->l_writers == 0);
1427 ldlm_set_cbpending(dlmlock);
1428 unlock_res_and_lock(dlmlock);
1430 result = ldlm_cli_cancel(&olck->ols_handle, LCF_ASYNC);
1432 CL_LOCK_DEBUG(D_ERROR, env, lock,
1433 "lock %p cancel failure with error(%d)\n",
1436 olck->ols_state = OLS_CANCELLED;
1437 olck->ols_flags &= ~LDLM_FL_LVB_READY;
1438 osc_lock_detach(env, olck);
1441 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
1442 static int check_cb(const struct lu_env *env, struct cl_io *io,
1443 struct osc_page *ops, void *cbdata)
1445 struct cl_lock *lock = cbdata;
1447 if (lock->cll_descr.cld_mode == CLM_READ) {
1448 struct cl_lock *tmp;
1449 tmp = cl_lock_at_pgoff(env, lock->cll_descr.cld_obj,
1450 osc_index(ops), lock, 1, 0);
1452 cl_lock_put(env, tmp);
1453 return CLP_GANG_OKAY;
1457 CL_LOCK_DEBUG(D_ERROR, env, lock, "still has pages\n");
1458 CL_PAGE_DEBUG(D_ERROR, env, ops->ops_cl.cpl_page, "\n");
1459 return CLP_GANG_ABORT;
1463 * Returns true iff there are pages under \a olck not protected by other
1466 static bool osc_lock_has_pages(struct osc_lock *olck)
1468 struct cl_lock *lock;
1469 struct cl_lock_descr *descr;
1470 struct cl_object *obj;
1471 struct osc_object *oob;
1472 struct cl_env_nest nest;
1478 env = cl_env_nested_get(&nest);
1482 obj = olck->ols_cl.cls_obj;
1484 io = &oob->oo_debug_io;
1485 lock = olck->ols_cl.cls_lock;
1486 descr = &lock->cll_descr;
1488 mutex_lock(&oob->oo_debug_mutex);
1489 io->ci_obj = cl_object_top(obj);
1490 io->ci_ignore_layout = 1;
1491 rc = cl_io_init(env, io, CIT_MISC, io->ci_obj);
1493 GOTO(out, has_pages = false);
1496 rc = osc_page_gang_lookup(env, io, oob,
1497 descr->cld_start, descr->cld_end,
1498 check_cb, (void *)lock);
1499 if (rc == CLP_GANG_ABORT)
1501 if (rc == CLP_GANG_RESCHED)
1503 } while (rc != CLP_GANG_OKAY);
1504 has_pages = (rc == CLP_GANG_ABORT);
1506 cl_io_fini(env, io);
1507 mutex_unlock(&oob->oo_debug_mutex);
1508 cl_env_nested_put(&nest, env);
1513 static bool osc_lock_has_pages(struct osc_lock *olck)
1517 #endif /* CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK */
1519 static void osc_lock_delete(const struct lu_env *env,
1520 const struct cl_lock_slice *slice)
1522 struct osc_lock *olck;
1524 olck = cl2osc_lock(slice);
1525 if (olck->ols_glimpse) {
1526 LASSERT(!olck->ols_hold);
1527 LASSERT(!olck->ols_lock);
1531 LINVRNT(osc_lock_invariant(olck));
1532 LINVRNT(!osc_lock_has_pages(olck));
1534 osc_lock_unhold(olck);
1535 osc_lock_detach(env, olck);
1539 * Implements cl_lock_operations::clo_state() method for osc layer.
1541 * Maintains osc_lock::ols_owner field.
1543 * This assumes that lock always enters CLS_HELD (from some other state) in
1544 * the same IO context as one that requested the lock. This should not be a
1545 * problem, because context is by definition shared by all activity pertaining
1546 * to the same high-level IO.
1548 static void osc_lock_state(const struct lu_env *env,
1549 const struct cl_lock_slice *slice,
1550 enum cl_lock_state state)
1552 struct osc_lock *lock = cl2osc_lock(slice);
1555 * XXX multiple io contexts can use the lock at the same time.
1557 LINVRNT(osc_lock_invariant(lock));
1558 if (state == CLS_HELD && slice->cls_lock->cll_state != CLS_HELD) {
1559 struct osc_io *oio = osc_env_io(env);
1561 LASSERT(lock->ols_owner == NULL);
1562 lock->ols_owner = oio;
1563 } else if (state != CLS_HELD)
1564 lock->ols_owner = NULL;
1567 static int osc_lock_print(const struct lu_env *env, void *cookie,
1568 lu_printer_t p, const struct cl_lock_slice *slice)
1570 struct osc_lock *lock = cl2osc_lock(slice);
1573 * XXX print ldlm lock and einfo properly.
1575 (*p)(env, cookie, "%p "LPX64" "LPX64" %d %p ",
1576 lock->ols_lock, lock->ols_flags, lock->ols_handle.cookie,
1577 lock->ols_state, lock->ols_owner);
1578 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1582 static int osc_lock_fits_into(const struct lu_env *env,
1583 const struct cl_lock_slice *slice,
1584 const struct cl_lock_descr *need,
1585 const struct cl_io *io)
1587 struct osc_lock *ols = cl2osc_lock(slice);
1589 if (need->cld_enq_flags & CEF_NEVER)
1592 if (ols->ols_state >= OLS_CANCELLED)
1595 if (need->cld_mode == CLM_PHANTOM) {
1597 return !(ols->ols_state > OLS_RELEASED);
1600 * Note: the QUEUED lock can't be matched here, otherwise
1601 * it might cause the deadlocks.
1603 * P1: enqueued read lock, create sublock1
1604 * P2: enqueued write lock, create sublock2(conflicted
1606 * P1: Grant read lock.
1607 * P1: enqueued glimpse lock(with holding sublock1_read),
1608 * matched with sublock2, waiting sublock2 to be granted.
1609 * But sublock2 can not be granted, because P1
1610 * will not release sublock1. Bang!
1612 if (ols->ols_state < OLS_GRANTED ||
1613 ols->ols_state > OLS_RELEASED)
1615 } else if (need->cld_enq_flags & CEF_MUST) {
1617 * If the lock hasn't ever enqueued, it can't be matched
1618 * because enqueue process brings in many information
1619 * which can be used to determine things such as lockless,
1622 if (ols->ols_state < OLS_UPCALL_RECEIVED &&
1623 ols->ols_locklessable)
1629 static const struct cl_lock_operations osc_lock_ops = {
1630 .clo_fini = osc_lock_fini,
1631 .clo_enqueue = osc_lock_enqueue,
1632 .clo_wait = osc_lock_wait,
1633 .clo_unuse = osc_lock_unuse,
1634 .clo_use = osc_lock_use,
1635 .clo_delete = osc_lock_delete,
1636 .clo_state = osc_lock_state,
1637 .clo_cancel = osc_lock_cancel,
1638 .clo_print = osc_lock_print,
1639 .clo_fits_into = osc_lock_fits_into,
1642 static int osc_lock_lockless_unuse(const struct lu_env *env,
1643 const struct cl_lock_slice *slice)
1645 struct osc_lock *ols = cl2osc_lock(slice);
1646 struct cl_lock *lock = slice->cls_lock;
1648 LASSERT(ols->ols_state == OLS_GRANTED);
1649 LINVRNT(osc_lock_invariant(ols));
1651 cl_lock_cancel(env, lock);
1652 cl_lock_delete(env, lock);
1656 static void osc_lock_lockless_cancel(const struct lu_env *env,
1657 const struct cl_lock_slice *slice)
1659 struct osc_lock *ols = cl2osc_lock(slice);
1662 result = osc_lock_flush(ols, 0);
1664 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1666 ols->ols_state = OLS_CANCELLED;
1669 static int osc_lock_lockless_wait(const struct lu_env *env,
1670 const struct cl_lock_slice *slice)
1672 struct osc_lock *olck = cl2osc_lock(slice);
1673 struct cl_lock *lock = olck->ols_cl.cls_lock;
1675 LINVRNT(osc_lock_invariant(olck));
1676 LASSERT(olck->ols_state >= OLS_UPCALL_RECEIVED);
1678 return lock->cll_error;
1681 static void osc_lock_lockless_state(const struct lu_env *env,
1682 const struct cl_lock_slice *slice,
1683 enum cl_lock_state state)
1685 struct osc_lock *lock = cl2osc_lock(slice);
1687 LINVRNT(osc_lock_invariant(lock));
1688 if (state == CLS_HELD) {
1689 struct osc_io *oio = osc_env_io(env);
1691 LASSERT(ergo(lock->ols_owner, lock->ols_owner == oio));
1692 lock->ols_owner = oio;
1694 /* set the io to be lockless if this lock is for io's
1696 if (cl_object_same(oio->oi_cl.cis_obj, slice->cls_obj))
1697 oio->oi_lockless = 1;
1701 static int osc_lock_lockless_fits_into(const struct lu_env *env,
1702 const struct cl_lock_slice *slice,
1703 const struct cl_lock_descr *need,
1704 const struct cl_io *io)
1706 struct osc_lock *lock = cl2osc_lock(slice);
1708 if (!(need->cld_enq_flags & CEF_NEVER))
1711 /* lockless lock should only be used by its owning io. b22147 */
1712 return (lock->ols_owner == osc_env_io(env));
1715 static const struct cl_lock_operations osc_lock_lockless_ops = {
1716 .clo_fini = osc_lock_fini,
1717 .clo_enqueue = osc_lock_enqueue,
1718 .clo_wait = osc_lock_lockless_wait,
1719 .clo_unuse = osc_lock_lockless_unuse,
1720 .clo_state = osc_lock_lockless_state,
1721 .clo_fits_into = osc_lock_lockless_fits_into,
1722 .clo_cancel = osc_lock_lockless_cancel,
1723 .clo_print = osc_lock_print
1726 int osc_lock_init(const struct lu_env *env,
1727 struct cl_object *obj, struct cl_lock *lock,
1728 const struct cl_io *unused)
1730 struct osc_lock *clk;
1733 OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, GFP_NOFS);
1735 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1737 osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo);
1738 clk->ols_state = OLS_NEW;
1740 clk->ols_flags = osc_enq2ldlm_flags(enqflags);
1741 clk->ols_agl = !!(enqflags & CEF_AGL);
1743 clk->ols_flags |= LDLM_FL_BLOCK_NOWAIT;
1744 if (clk->ols_flags & LDLM_FL_HAS_INTENT)
1745 clk->ols_glimpse = 1;
1747 cl_lock_slice_add(lock, &clk->ols_cl, obj, &osc_lock_ops);
1749 if (!(enqflags & CEF_MUST))
1750 /* try to convert this lock to a lockless lock */
1751 osc_lock_to_lockless(env, clk, (enqflags & CEF_NEVER));
1752 if (clk->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
1753 clk->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1755 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags "LPX64,
1756 lock, clk, clk->ols_flags);