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 /** \addtogroup osc osc @{ */
43 #define DEBUG_SUBSYSTEM S_OSC
46 # include <libcfs/libcfs.h>
48 # include <liblustre.h>
50 /* fid_build_reg_res_name() */
51 #include <lustre_fid.h>
53 #include "osc_cl_internal.h"
55 /*****************************************************************************
61 static const struct cl_lock_operations osc_lock_ops;
62 static const struct cl_lock_operations osc_lock_lockless_ops;
63 static void osc_lock_to_lockless(const struct lu_env *env,
64 struct osc_lock *ols, int force);
66 int osc_lock_is_lockless(const struct osc_lock *olck)
68 return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
72 * Returns a weak pointer to the ldlm lock identified by a handle. Returned
73 * pointer cannot be dereferenced, as lock is not protected from concurrent
74 * reclaim. This function is a helper for osc_lock_invariant().
76 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
78 struct ldlm_lock *lock;
80 lock = ldlm_handle2lock(handle);
87 * Invariant that has to be true all of the time.
89 static int osc_lock_invariant(struct osc_lock *ols)
91 struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle);
92 struct ldlm_lock *olock = ols->ols_lock;
93 int handle_used = lustre_handle_is_used(&ols->ols_handle);
96 ergo(osc_lock_is_lockless(ols),
97 ols->ols_locklessable && ols->ols_lock == NULL) ||
98 (ergo(olock != NULL, handle_used) &&
100 olock->l_handle.h_cookie == ols->ols_handle.cookie) &&
102 * Check that ->ols_handle and ->ols_lock are consistent, but
103 * take into account that they are set at the different time.
106 ergo(lock != NULL && olock != NULL, lock == olock) &&
107 ergo(lock == NULL, olock == NULL)) &&
108 ergo(ols->ols_state == OLS_CANCELLED,
109 olock == NULL && !handle_used) &&
111 * DLM lock is destroyed only after we have seen cancellation
114 ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
115 !olock->l_destroyed) &&
116 ergo(ols->ols_state == OLS_GRANTED,
118 olock->l_req_mode == olock->l_granted_mode &&
122 /*****************************************************************************
129 * Breaks a link between osc_lock and dlm_lock.
131 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
133 struct ldlm_lock *dlmlock;
135 spin_lock(&osc_ast_guard);
136 dlmlock = olck->ols_lock;
137 if (dlmlock == NULL) {
138 spin_unlock(&osc_ast_guard);
142 olck->ols_lock = NULL;
143 /* wb(); --- for all who checks (ols->ols_lock != NULL) before
144 * call to osc_lock_detach() */
145 dlmlock->l_ast_data = NULL;
146 olck->ols_handle.cookie = 0ULL;
147 spin_unlock(&osc_ast_guard);
149 lock_res_and_lock(dlmlock);
150 if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
151 struct cl_object *obj = olck->ols_cl.cls_obj;
152 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
153 __u64 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
155 /* Update the kms. Need to loop all granted locks.
156 * Not a problem for the client */
157 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
158 unlock_res_and_lock(dlmlock);
160 cl_object_attr_lock(obj);
161 cl_object_attr_set(env, obj, attr, CAT_KMS);
162 cl_object_attr_unlock(obj);
164 unlock_res_and_lock(dlmlock);
166 /* release a reference taken in osc_lock_upcall0(). */
167 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
168 LDLM_LOCK_RELEASE(dlmlock);
171 static int osc_lock_unuse(const struct lu_env *env,
172 const struct cl_lock_slice *slice)
174 struct osc_lock *ols = cl2osc_lock(slice);
177 LASSERT(ols->ols_state == OLS_GRANTED ||
178 ols->ols_state == OLS_UPCALL_RECEIVED);
179 LINVRNT(osc_lock_invariant(ols));
181 if (ols->ols_glimpse) {
182 LASSERT(ols->ols_hold == 0);
185 LASSERT(ols->ols_hold);
188 * Move lock into OLS_RELEASED state before calling osc_cancel_base()
189 * so that possible synchronous cancellation (that always happens
190 * e.g., for liblustre) sees that lock is released.
192 ols->ols_state = OLS_RELEASED;
194 result = osc_cancel_base(&ols->ols_handle, ols->ols_einfo.ei_mode);
195 ols->ols_has_ref = 0;
199 static void osc_lock_fini(const struct lu_env *env,
200 struct cl_lock_slice *slice)
202 struct osc_lock *ols = cl2osc_lock(slice);
204 LINVRNT(osc_lock_invariant(ols));
206 * ->ols_hold can still be true at this point if, for example, a
207 * thread that requested a lock was killed (and released a reference
208 * to the lock), before reply from a server was received. In this case
209 * lock is destroyed immediately after upcall.
212 osc_lock_unuse(env, slice);
213 if (ols->ols_lock != NULL)
214 osc_lock_detach(env, ols);
216 OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
219 void osc_lock_build_res(const struct lu_env *env, const struct osc_object *obj,
220 struct ldlm_res_id *resname)
222 const struct lu_fid *fid = lu_object_fid(&obj->oo_cl.co_lu);
225 * In the perfect world of the future, where ost servers talk
228 fid_build_reg_res_name(fid, resname);
231 * In reality, where ost server expects ->lsm_object_id and
232 * ->lsm_object_gr in rename.
234 osc_build_res_name(obj->oo_oinfo->loi_id, obj->oo_oinfo->loi_gr,
239 static void osc_lock_build_policy(const struct lu_env *env,
240 const struct cl_lock *lock,
241 ldlm_policy_data_t *policy)
243 const struct cl_lock_descr *d = &lock->cll_descr;
245 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
248 static int osc_enq2ldlm_flags(__u32 enqflags)
252 LASSERT((enqflags & ~CEF_MASK) == 0);
254 if (enqflags & CEF_NONBLOCK)
255 result |= LDLM_FL_BLOCK_NOWAIT;
256 if (enqflags & CEF_ASYNC)
257 result |= LDLM_FL_HAS_INTENT;
258 if (enqflags & CEF_DISCARD_DATA)
259 result |= LDLM_AST_DISCARD_DATA;
264 * Global spin-lock protecting consistency of ldlm_lock::l_ast_data
265 * pointers. Initialized in osc_init().
267 spinlock_t osc_ast_guard;
269 static struct osc_lock *osc_ast_data_get(struct ldlm_lock *dlm_lock)
271 struct osc_lock *olck;
273 lock_res_and_lock(dlm_lock);
274 spin_lock(&osc_ast_guard);
275 olck = dlm_lock->l_ast_data;
277 struct cl_lock *lock = olck->ols_cl.cls_lock;
279 * If osc_lock holds a reference on ldlm lock, return it even
280 * when cl_lock is in CLS_FREEING state. This way
282 * osc_ast_data_get(dlmlock) == NULL
284 * guarantees that all osc references on dlmlock were
285 * released. osc_dlm_blocking_ast0() relies on that.
287 if (lock->cll_state < CLS_FREEING || olck->ols_has_ref) {
288 cl_lock_get_trust(lock);
289 lu_ref_add_atomic(&lock->cll_reference,
290 "ast", cfs_current());
294 spin_unlock(&osc_ast_guard);
295 unlock_res_and_lock(dlm_lock);
299 static void osc_ast_data_put(const struct lu_env *env, struct osc_lock *olck)
301 struct cl_lock *lock;
303 lock = olck->ols_cl.cls_lock;
304 lu_ref_del(&lock->cll_reference, "ast", cfs_current());
305 cl_lock_put(env, lock);
309 * Updates object attributes from a lock value block (lvb) received together
310 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
313 * This can be optimized to not update attributes when lock is a result of a
316 * Called under lock and resource spin-locks.
318 static void osc_lock_lvb_update(const struct lu_env *env, struct osc_lock *olck,
322 struct cl_object *obj;
323 struct lov_oinfo *oinfo;
324 struct cl_attr *attr;
329 if (!(olck->ols_flags & LDLM_FL_LVB_READY)) {
334 lvb = &olck->ols_lvb;
335 obj = olck->ols_cl.cls_obj;
336 oinfo = cl2osc(obj)->oo_oinfo;
337 attr = &osc_env_info(env)->oti_attr;
338 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
339 cl_lvb2attr(attr, lvb);
341 cl_object_attr_lock(obj);
343 struct ldlm_lock *dlmlock;
346 dlmlock = olck->ols_lock;
347 LASSERT(dlmlock != NULL);
349 /* re-grab LVB from a dlm lock under DLM spin-locks. */
350 *lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
351 size = lvb->lvb_size;
352 /* Extend KMS up to the end of this lock and no further
353 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
354 if (size > dlmlock->l_policy_data.l_extent.end)
355 size = dlmlock->l_policy_data.l_extent.end + 1;
356 if (size >= oinfo->loi_kms) {
357 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="LPU64
358 ", kms="LPU64, lvb->lvb_size, size);
360 attr->cat_kms = size;
362 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
363 LPU64"; leaving kms="LPU64", end="LPU64,
364 lvb->lvb_size, oinfo->loi_kms,
365 dlmlock->l_policy_data.l_extent.end);
367 ldlm_lock_allow_match_locked(dlmlock);
368 } else if (rc == -ENAVAIL && olck->ols_glimpse) {
369 CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving"
370 " kms="LPU64"\n", lvb->lvb_size, oinfo->loi_kms);
375 cl_object_attr_set(env, obj, attr, valid);
377 cl_object_attr_unlock(obj);
383 * Called when a lock is granted, from an upcall (when server returned a
384 * granted lock), or from completion AST, when server returned a blocked lock.
386 * Called under lock and resource spin-locks, that are released temporarily
389 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *olck,
390 struct ldlm_lock *dlmlock, int rc)
392 struct ldlm_extent *ext;
393 struct cl_lock *lock;
394 struct cl_lock_descr *descr;
396 LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
399 if (olck->ols_state != OLS_GRANTED) {
400 lock = olck->ols_cl.cls_lock;
401 ext = &dlmlock->l_policy_data.l_extent;
402 descr = &osc_env_info(env)->oti_descr;
403 descr->cld_obj = lock->cll_descr.cld_obj;
405 /* XXX check that ->l_granted_mode is valid. */
406 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
407 descr->cld_start = cl_index(descr->cld_obj, ext->start);
408 descr->cld_end = cl_index(descr->cld_obj, ext->end);
410 * tell upper layers the extent of the lock that was actually
413 olck->ols_state = OLS_GRANTED;
414 osc_lock_lvb_update(env, olck, rc);
416 /* release DLM spin-locks to allow cl_lock_{modify,signal}()
417 * to take a semaphore on a parent lock. This is safe, because
418 * spin-locks are needed to protect consistency of
419 * dlmlock->l_*_mode and LVB, and we have finished processing
421 unlock_res_and_lock(dlmlock);
422 cl_lock_modify(env, lock, descr);
423 cl_lock_signal(env, lock);
424 LINVRNT(osc_lock_invariant(olck));
425 lock_res_and_lock(dlmlock);
430 static void osc_lock_upcall0(const struct lu_env *env, struct osc_lock *olck)
433 struct ldlm_lock *dlmlock;
437 dlmlock = ldlm_handle2lock_long(&olck->ols_handle, 0);
438 LASSERT(dlmlock != NULL);
440 lock_res_and_lock(dlmlock);
441 spin_lock(&osc_ast_guard);
442 LASSERT(dlmlock->l_ast_data == olck);
443 LASSERT(olck->ols_lock == NULL);
444 olck->ols_lock = dlmlock;
445 spin_unlock(&osc_ast_guard);
448 * Lock might be not yet granted. In this case, completion ast
449 * (osc_ldlm_completion_ast()) comes later and finishes lock
452 if (dlmlock->l_granted_mode == dlmlock->l_req_mode)
453 osc_lock_granted(env, olck, dlmlock, 0);
454 unlock_res_and_lock(dlmlock);
457 * osc_enqueue_interpret() decrefs asynchronous locks, counter
460 ldlm_lock_addref(&olck->ols_handle, olck->ols_einfo.ei_mode);
461 olck->ols_hold = olck->ols_has_ref = 1;
463 /* lock reference taken by ldlm_handle2lock_long() is owned by
464 * osc_lock and released in osc_lock_detach() */
465 lu_ref_add(&dlmlock->l_reference, "osc_lock", olck);
469 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
470 * received from a server, or after osc_enqueue_base() matched a local DLM
473 static int osc_lock_upcall(void *cookie, int errcode)
475 struct osc_lock *olck = cookie;
476 struct cl_lock_slice *slice = &olck->ols_cl;
477 struct cl_lock *lock = slice->cls_lock;
484 * XXX environment should be created in ptlrpcd.
486 env = cl_env_get(&refcheck);
490 cl_lock_mutex_get(env, lock);
492 LASSERT(lock->cll_state >= CLS_QUEUING);
493 if (olck->ols_state == OLS_ENQUEUED) {
494 olck->ols_state = OLS_UPCALL_RECEIVED;
495 rc = ldlm_error2errno(errcode);
496 } else if (olck->ols_state == OLS_CANCELLED) {
499 CERROR("Impossible state: %i\n", olck->ols_state);
503 struct ldlm_lock *dlmlock;
505 dlmlock = ldlm_handle2lock(&olck->ols_handle);
506 if (dlmlock != NULL) {
507 lock_res_and_lock(dlmlock);
508 spin_lock(&osc_ast_guard);
509 LASSERT(olck->ols_lock == NULL);
510 dlmlock->l_ast_data = NULL;
511 olck->ols_handle.cookie = 0ULL;
512 spin_unlock(&osc_ast_guard);
513 unlock_res_and_lock(dlmlock);
514 LDLM_LOCK_PUT(dlmlock);
517 if (olck->ols_glimpse)
518 olck->ols_glimpse = 0;
519 osc_lock_upcall0(env, olck);
522 /* Error handling, some errors are tolerable. */
523 if (olck->ols_locklessable && rc == -EUSERS) {
524 /* This is a tolerable error, turn this lock into
527 osc_object_set_contended(cl2osc(slice->cls_obj));
528 LASSERT(slice->cls_ops == &osc_lock_ops);
530 /* Change this lock to ldlmlock-less lock. */
531 osc_lock_to_lockless(env, olck, 1);
532 olck->ols_state = OLS_GRANTED;
534 } else if (olck->ols_glimpse && rc == -ENAVAIL) {
535 osc_lock_lvb_update(env, olck, rc);
536 cl_lock_delete(env, lock);
537 /* Hide the error. */
542 /* on error, lock was signaled by cl_lock_error() */
543 cl_lock_signal(env, lock);
545 cl_lock_error(env, lock, rc);
547 cl_lock_mutex_put(env, lock);
549 /* release cookie reference, acquired by osc_lock_enqueue() */
550 lu_ref_del(&lock->cll_reference, "upcall", lock);
551 cl_lock_put(env, lock);
552 cl_env_put(env, &refcheck);
554 /* should never happen, similar to osc_ldlm_blocking_ast(). */
560 * Core of osc_dlm_blocking_ast() logic.
562 static void osc_lock_blocking(const struct lu_env *env,
563 struct ldlm_lock *dlmlock,
564 struct osc_lock *olck, int blocking)
566 struct cl_lock *lock = olck->ols_cl.cls_lock;
568 LASSERT(olck->ols_lock == dlmlock);
569 CLASSERT(OLS_BLOCKED < OLS_CANCELLED);
570 LASSERT(!osc_lock_is_lockless(olck));
574 * Lock might be still addref-ed here, if e.g., blocking ast
575 * is sent for a failed lock.
577 osc_lock_unuse(env, &olck->ols_cl);
579 if (blocking && olck->ols_state < OLS_BLOCKED)
581 * Move osc_lock into OLS_BLOCKED before canceling the lock,
582 * because it recursively re-enters osc_lock_blocking(), with
583 * the state set to OLS_CANCELLED.
585 olck->ols_state = OLS_BLOCKED;
587 * cancel and destroy lock at least once no matter how blocking ast is
588 * entered (see comment above osc_ldlm_blocking_ast() for use
589 * cases). cl_lock_cancel() and cl_lock_delete() are idempotent.
591 cl_lock_cancel(env, lock);
592 cl_lock_delete(env, lock);
596 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
597 * and ldlm_lock caches.
599 static int osc_dlm_blocking_ast0(const struct lu_env *env,
600 struct ldlm_lock *dlmlock,
601 void *data, int flag)
603 struct osc_lock *olck;
604 struct cl_lock *lock;
608 LASSERT(flag == LDLM_CB_BLOCKING || flag == LDLM_CB_CANCELING);
611 olck = osc_ast_data_get(dlmlock);
613 lock = olck->ols_cl.cls_lock;
614 cl_lock_mutex_get(env, lock);
615 LINVRNT(osc_lock_invariant(olck));
616 if (olck->ols_ast_wait) {
617 /* wake up osc_lock_use() */
618 cl_lock_signal(env, lock);
619 olck->ols_ast_wait = 0;
622 * Lock might have been canceled while this thread was
623 * sleeping for lock mutex, but olck is pinned in memory.
625 if (olck == dlmlock->l_ast_data) {
627 * NOTE: DLM sends blocking AST's for failed locks
628 * (that are still in pre-OLS_GRANTED state)
629 * too, and they have to be canceled otherwise
630 * DLM lock is never destroyed and stuck in
633 * Alternatively, ldlm_cli_cancel() can be
634 * called here directly for osc_locks with
635 * ols_state < OLS_GRANTED to maintain an
636 * invariant that ->clo_cancel() is only called
637 * for locks that were granted.
639 LASSERT(data == olck);
640 osc_lock_blocking(env, dlmlock,
641 olck, flag == LDLM_CB_BLOCKING);
644 cl_lock_mutex_put(env, lock);
645 osc_ast_data_put(env, olck);
648 * DLM lock exists, but there is no cl_lock attached to it.
649 * This is a `normal' race. cl_object and its cl_lock's can be
650 * removed by memory pressure, together with all pages.
652 cancel = (flag == LDLM_CB_BLOCKING);
655 struct lustre_handle *lockh;
657 lockh = &osc_env_info(env)->oti_handle;
658 ldlm_lock2handle(dlmlock, lockh);
659 result = ldlm_cli_cancel(lockh);
666 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
667 * some other lock, or is canceled. This function is installed as a
668 * ldlm_lock::l_blocking_ast() for client extent locks.
670 * Control flow is tricky, because ldlm uses the same call-back
671 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
673 * \param dlmlock lock for which ast occurred.
675 * \param new description of a conflicting lock in case of blocking ast.
677 * \param data value of dlmlock->l_ast_data
679 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
680 * cancellation and blocking ast's.
682 * Possible use cases:
684 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
685 * lock due to lock lru pressure, or explicit user request to purge
688 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
689 * us that dlmlock conflicts with another lock that some client is
690 * enqueing. Lock is canceled.
692 * - cl_lock_cancel() is called. osc_lock_cancel() calls
693 * ldlm_cli_cancel() that calls
695 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
697 * recursively entering osc_ldlm_blocking_ast().
699 * - client cancels lock voluntary (e.g., as a part of early cancellation):
702 * osc_lock_cancel()->
703 * ldlm_cli_cancel()->
704 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
707 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
708 struct ldlm_lock_desc *new, void *data,
712 struct cl_env_nest nest;
716 * This can be called in the context of outer IO, e.g.,
719 * ->osc_enqueue_base()->...
720 * ->ldlm_prep_elc_req()->...
721 * ->ldlm_cancel_callback()->...
722 * ->osc_ldlm_blocking_ast()
724 * new environment has to be created to not corrupt outer context.
726 env = cl_env_nested_get(&nest);
728 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
730 result = PTR_ERR(env);
732 * XXX This should never happen, as cl_lock is
733 * stuck. Pre-allocated environment a la vvp_inode_fini_env
739 if (result == -ENODATA)
742 CERROR("BAST failed: %d\n", result);
744 cl_env_nested_put(&nest, env);
748 static int osc_ldlm_completion_ast(struct ldlm_lock *dlmlock,
749 int flags, void *data)
753 struct osc_lock *olck;
754 struct cl_lock *lock;
759 /* first, do dlm part of the work */
760 dlmrc = ldlm_completion_ast_async(dlmlock, flags, data);
761 /* then, notify cl_lock */
762 env_cookie = cl_env_reenter();
763 env = cl_env_get(&refcheck);
765 olck = osc_ast_data_get(dlmlock);
767 lock = olck->ols_cl.cls_lock;
768 cl_lock_mutex_get(env, lock);
770 * ldlm_handle_cp_callback() copied LVB from request
771 * to lock->l_lvb_data, store it in osc_lock.
773 LASSERT(dlmlock->l_lvb_data != NULL);
774 lock_res_and_lock(dlmlock);
775 olck->ols_lvb = *(struct ost_lvb *)dlmlock->l_lvb_data;
776 if (olck->ols_lock == NULL)
778 * upcall (osc_lock_upcall()) hasn't yet been
779 * called. Do nothing now, upcall will bind
780 * olck to dlmlock and signal the waiters.
782 * This maintains an invariant that osc_lock
783 * and ldlm_lock are always bound when
784 * osc_lock is in OLS_GRANTED state.
787 else if (dlmlock->l_granted_mode != LCK_MINMODE)
788 osc_lock_granted(env, olck, dlmlock, dlmrc);
789 unlock_res_and_lock(dlmlock);
791 cl_lock_error(env, lock, dlmrc);
792 cl_lock_mutex_put(env, lock);
793 osc_ast_data_put(env, olck);
796 result = -ELDLM_NO_LOCK_DATA;
797 cl_env_put(env, &refcheck);
799 result = PTR_ERR(env);
800 cl_env_reexit(env_cookie);
801 return dlmrc ?: result;
804 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
806 struct ptlrpc_request *req = data;
807 struct osc_lock *olck;
808 struct cl_lock *lock;
809 struct cl_object *obj;
812 struct req_capsule *cap;
816 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
818 env = cl_env_get(&refcheck);
821 * osc_ast_data_get() has to go after environment is
822 * allocated, because osc_ast_data() acquires a
823 * reference to a lock, and it can only be released in
826 olck = osc_ast_data_get(dlmlock);
828 lock = olck->ols_cl.cls_lock;
829 cl_lock_mutex_get(env, lock);
831 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
832 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
834 result = req_capsule_server_pack(cap);
836 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
837 obj = lock->cll_descr.cld_obj;
838 result = cl_object_glimpse(env, obj, lvb);
840 cl_lock_mutex_put(env, lock);
841 osc_ast_data_put(env, olck);
844 * These errors are normal races, so we don't want to
845 * fill the console with messages by calling
848 lustre_pack_reply(req, 1, NULL, NULL);
849 result = -ELDLM_NO_LOCK_DATA;
851 cl_env_put(env, &refcheck);
853 result = PTR_ERR(env);
854 req->rq_status = result;
858 static unsigned long osc_lock_weigh(const struct lu_env *env,
859 const struct cl_lock_slice *slice)
862 * don't need to grab coh_page_guard since we don't care the exact #
865 return cl_object_header(slice->cls_obj)->coh_pages;
869 * Get the weight of dlm lock for early cancellation.
871 * XXX: it should return the pages covered by this \a dlmlock.
873 static unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
878 struct osc_lock *lock;
880 unsigned long weight;
884 cookie = cl_env_reenter();
886 * osc_ldlm_weigh_ast has a complex context since it might be called
887 * because of lock canceling, or from user's input. We have to make
888 * a new environment for it. Probably it is implementation safe to use
889 * the upper context because cl_lock_put don't modify environment
890 * variables. But in case of ..
892 env = cl_env_get(&refcheck);
894 /* Mostly because lack of memory, tend to eliminate this lock*/
895 cl_env_reexit(cookie);
899 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
900 lock = osc_ast_data_get(dlmlock);
902 /* cl_lock was destroyed because of memory pressure.
903 * It is much reasonable to assign this type of lock
906 GOTO(out, weight = 0);
909 cll = lock->ols_cl.cls_lock;
910 cl_lock_mutex_get(env, cll);
911 weight = cl_lock_weigh(env, cll);
912 cl_lock_mutex_put(env, cll);
913 osc_ast_data_put(env, lock);
917 cl_env_put(env, &refcheck);
918 cl_env_reexit(cookie);
922 static void osc_lock_build_einfo(const struct lu_env *env,
923 const struct cl_lock *clock,
924 struct osc_lock *lock,
925 struct ldlm_enqueue_info *einfo)
927 enum cl_lock_mode mode;
929 mode = clock->cll_descr.cld_mode;
930 if (mode == CLM_PHANTOM)
932 * For now, enqueue all glimpse locks in read mode. In the
933 * future, client might choose to enqueue LCK_PW lock for
934 * glimpse on a file opened for write.
938 einfo->ei_type = LDLM_EXTENT;
939 einfo->ei_mode = osc_cl_lock2ldlm(mode);
940 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
941 einfo->ei_cb_cp = osc_ldlm_completion_ast;
942 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
943 einfo->ei_cb_wg = osc_ldlm_weigh_ast;
944 einfo->ei_cbdata = lock; /* value to be put into ->l_ast_data */
947 static int osc_lock_delete0(struct cl_lock *conflict)
949 struct cl_env_nest nest;
953 env = cl_env_nested_get(&nest);
955 cl_lock_delete(env, conflict);
956 cl_env_nested_put(&nest, env);
962 * Cancels \a conflict lock and waits until it reached CLS_FREEING state. This
963 * is called as a part of enqueuing to cancel conflicting locks early.
965 * \retval 0: success, \a conflict was cancelled and destroyed.
967 * \retval CLO_REPEAT: \a conflict was cancelled, but \a lock mutex was
968 * released in the process. Repeat enqueing.
970 * \retval -EWOULDBLOCK: \a conflict cannot be cancelled immediately, and
971 * either \a lock is non-blocking, or current thread
972 * holds other locks, that prevent it from waiting
973 * for cancel to complete.
975 * \retval -ve: other error, including -EINTR.
978 static int osc_lock_cancel_wait(const struct lu_env *env, struct cl_lock *lock,
979 struct cl_lock *conflict, int canwait)
983 LASSERT(cl_lock_is_mutexed(lock));
984 LASSERT(cl_lock_is_mutexed(conflict));
987 if (conflict->cll_state != CLS_FREEING) {
988 cl_lock_cancel(env, conflict);
989 rc = osc_lock_delete0(conflict);
992 if (conflict->cll_flags & (CLF_CANCELPEND|CLF_DOOMED)) {
994 if (cl_lock_nr_mutexed(env) > 2)
996 * If mutices of locks other than @lock and
997 * @scan are held by the current thread, it
998 * cannot wait on @scan state change in a
999 * dead-lock safe matter, so simply skip early
1000 * cancellation in this case.
1002 * This means that early cancellation doesn't
1003 * work when there is even slight mutex
1004 * contention, as top-lock's mutex is usually
1005 * held at this time.
1009 /* Waiting for @scan to be destroyed */
1010 cl_lock_mutex_put(env, lock);
1012 rc = cl_lock_state_wait(env, conflict);
1014 conflict->cll_state < CLS_FREEING);
1015 /* mutex was released, repeat enqueue. */
1016 rc = rc ?: CLO_REPEAT;
1017 cl_lock_mutex_get(env, lock);
1020 LASSERT(ergo(!rc, conflict->cll_state == CLS_FREEING));
1021 CDEBUG(D_INFO, "lock %p was %s freed now, rc (%d)\n",
1022 conflict, rc ? "not":"", rc);
1028 * Determine if the lock should be converted into a lockless lock.
1031 * - if the lock has an explicite requirment for a non-lockless lock;
1032 * - if the io lock request type ci_lockreq;
1033 * - send the enqueue rpc to ost to make the further decision;
1034 * - special treat to truncate lockless lock
1036 * Additional policy can be implemented here, e.g., never do lockless-io
1037 * for large extents.
1039 static void osc_lock_to_lockless(const struct lu_env *env,
1040 struct osc_lock *ols, int force)
1042 struct cl_lock_slice *slice = &ols->ols_cl;
1043 struct cl_lock *lock = slice->cls_lock;
1045 LASSERT(ols->ols_state == OLS_NEW ||
1046 ols->ols_state == OLS_UPCALL_RECEIVED);
1049 ols->ols_locklessable = 1;
1050 LASSERT(cl_lock_is_mutexed(lock));
1051 slice->cls_ops = &osc_lock_lockless_ops;
1053 struct osc_io *oio = osc_env_io(env);
1054 struct cl_io *io = oio->oi_cl.cis_io;
1055 struct cl_object *obj = slice->cls_obj;
1056 struct osc_object *oob = cl2osc(obj);
1057 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
1058 struct obd_connect_data *ocd;
1060 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
1061 io->ci_lockreq == CILR_MAYBE ||
1062 io->ci_lockreq == CILR_NEVER);
1064 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
1065 ols->ols_locklessable = (io->ci_type != CIT_TRUNC) &&
1066 (io->ci_lockreq == CILR_MAYBE) &&
1067 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
1068 if (io->ci_lockreq == CILR_NEVER ||
1070 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
1071 /* lockless truncate */
1072 (io->ci_type == CIT_TRUNC &&
1073 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
1074 osd->od_lockless_truncate)) {
1075 ols->ols_locklessable = 1;
1076 slice->cls_ops = &osc_lock_lockless_ops;
1079 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1083 * Cancel all conflicting locks and wait for them to be destroyed.
1085 * This function is used for two purposes:
1087 * - early cancel all conflicting locks before starting IO, and
1089 * - guarantee that pages added to the page cache by lockless IO are never
1090 * covered by locks other than lockless IO lock, and, hence, are not
1091 * visible to other threads.
1093 static int osc_lock_enqueue_wait(const struct lu_env *env,
1094 const struct osc_lock *olck)
1096 struct cl_lock *lock = olck->ols_cl.cls_lock;
1097 struct cl_lock_descr *descr = &lock->cll_descr;
1098 struct cl_object_header *hdr = cl_object_header(descr->cld_obj);
1099 struct cl_lock_closure *closure = &osc_env_info(env)->oti_closure;
1100 struct cl_lock *scan;
1101 struct cl_lock *temp;
1102 int lockless = osc_lock_is_lockless(olck);
1108 LASSERT(cl_lock_is_mutexed(lock));
1109 LASSERT(lock->cll_state == CLS_QUEUING);
1112 * XXX This function could be sped up if we had asynchronous
1117 !(olck->ols_flags & LDLM_FL_BLOCK_NOWAIT) &&
1118 cl_lock_nr_mutexed(env) == 1;
1119 cl_lock_closure_init(env, closure, lock, canwait);
1120 spin_lock(&hdr->coh_lock_guard);
1121 list_for_each_entry_safe(scan, temp, &hdr->coh_locks, cll_linkage) {
1125 if (scan->cll_state < CLS_QUEUING ||
1126 scan->cll_state == CLS_FREEING ||
1127 scan->cll_descr.cld_start > descr->cld_end ||
1128 scan->cll_descr.cld_end < descr->cld_start)
1131 /* overlapped and living locks. */
1132 /* A tricky case for lockless pages:
1133 * We need to cancel the compatible locks if we're enqueuing
1134 * a lockless lock, for example:
1135 * imagine that client has PR lock on [0, 1000], and thread T0
1136 * is doing lockless IO in [500, 1500] region. Concurrent
1137 * thread T1 can see lockless data in [500, 1000], which is
1138 * wrong, because these data are possibly stale.
1140 if (!lockless && cl_lock_compatible(scan, lock))
1143 /* Now @scan is conflicting with @lock, this means current
1144 * thread have to sleep for @scan being destroyed. */
1145 cl_lock_get_trust(scan);
1146 if (&temp->cll_linkage != &hdr->coh_locks)
1147 cl_lock_get_trust(temp);
1148 spin_unlock(&hdr->coh_lock_guard);
1149 lu_ref_add(&scan->cll_reference, "cancel-wait", lock);
1151 LASSERT(list_empty(&closure->clc_list));
1152 rc = cl_lock_closure_build(env, scan, closure);
1154 rc = osc_lock_cancel_wait(env, lock, scan, canwait);
1155 cl_lock_disclosure(env, closure);
1156 if (rc == -EWOULDBLOCK)
1159 if (rc == CLO_REPEAT && !canwait)
1160 /* cannot wait... no early cancellation. */
1163 lu_ref_del(&scan->cll_reference, "cancel-wait", lock);
1164 cl_lock_put(env, scan);
1165 spin_lock(&hdr->coh_lock_guard);
1167 * Lock list could have been modified, while spin-lock was
1168 * released. Check that it is safe to continue.
1170 stop = list_empty(&temp->cll_linkage);
1171 if (&temp->cll_linkage != &hdr->coh_locks)
1172 cl_lock_put(env, temp);
1173 if (stop || rc != 0)
1176 spin_unlock(&hdr->coh_lock_guard);
1177 cl_lock_closure_fini(closure);
1182 * Deadlock avoidance for osc_lock_enqueue(). Consider following scenario:
1184 * - Thread0: obtains PR:[0, 10]. Lock is busy.
1186 * - Thread1: enqueues PW:[5, 50]. Blocking ast is sent to
1187 * PR:[0, 10], but cancellation of busy lock is postponed.
1189 * - Thread0: enqueue PR:[30, 40]. Lock is locally matched to
1190 * PW:[5, 50], and thread0 waits for the lock completion never
1191 * releasing PR:[0, 10]---deadlock.
1193 * The second PR lock can be glimpse (it is to deal with that situation that
1194 * ll_glimpse_size() has second argument, preventing local match of
1195 * not-yet-granted locks, see bug 10295). Similar situation is possible in the
1196 * case of memory mapped user level buffer.
1198 * To prevent this we can detect a situation when current "thread" or "io"
1199 * already holds a lock on this object and either add LDLM_FL_BLOCK_GRANTED to
1200 * the ols->ols_flags, or prevent local match with PW locks.
1202 static int osc_deadlock_is_possible(const struct lu_env *env,
1203 struct cl_lock *lock)
1205 struct cl_object *obj;
1206 struct cl_object_header *head;
1207 struct cl_lock *scan;
1214 LASSERT(cl_lock_is_mutexed(lock));
1216 oio = osc_env_io(env);
1217 obj = lock->cll_descr.cld_obj;
1218 head = cl_object_header(obj);
1221 spin_lock(&head->coh_lock_guard);
1222 list_for_each_entry(scan, &head->coh_locks, cll_linkage) {
1224 struct osc_lock *oscan;
1226 oscan = osc_lock_at(scan);
1227 LASSERT(oscan != NULL);
1228 if (oscan->ols_owner == oio) {
1234 spin_unlock(&head->coh_lock_guard);
1239 * Implementation of cl_lock_operations::clo_enqueue() method for osc
1240 * layer. This initiates ldlm enqueue:
1242 * - checks for possible dead-lock conditions (osc_deadlock_is_possible());
1244 * - cancels conflicting locks early (osc_lock_enqueue_wait());
1246 * - calls osc_enqueue_base() to do actual enqueue.
1248 * osc_enqueue_base() is supplied with an upcall function that is executed
1249 * when lock is received either after a local cached ldlm lock is matched, or
1250 * when a reply from the server is received.
1252 * This function does not wait for the network communication to complete.
1254 static int osc_lock_enqueue(const struct lu_env *env,
1255 const struct cl_lock_slice *slice,
1256 struct cl_io *_, __u32 enqflags)
1258 struct osc_lock *ols = cl2osc_lock(slice);
1259 struct cl_lock *lock = ols->ols_cl.cls_lock;
1260 struct osc_object *obj = cl2osc(slice->cls_obj);
1261 struct osc_thread_info *info = osc_env_info(env);
1262 struct ldlm_res_id *resname = &info->oti_resname;
1263 ldlm_policy_data_t *policy = &info->oti_policy;
1264 struct ldlm_enqueue_info *einfo = &ols->ols_einfo;
1268 LASSERT(cl_lock_is_mutexed(lock));
1269 LASSERT(lock->cll_state == CLS_QUEUING);
1270 LASSERT(ols->ols_state == OLS_NEW);
1272 osc_lock_build_res(env, obj, resname);
1273 osc_lock_build_policy(env, lock, policy);
1274 ols->ols_flags = osc_enq2ldlm_flags(enqflags);
1275 if (osc_deadlock_is_possible(env, lock))
1276 ols->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1277 if (ols->ols_flags & LDLM_FL_HAS_INTENT)
1278 ols->ols_glimpse = 1;
1280 result = osc_lock_enqueue_wait(env, ols);
1282 if (!(enqflags & CEF_MUST))
1283 /* try to convert this lock to a lockless lock */
1284 osc_lock_to_lockless(env, ols, (enqflags & CEF_NEVER));
1285 if (!osc_lock_is_lockless(ols)) {
1286 if (ols->ols_locklessable)
1287 ols->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1289 /* a reference for lock, passed as an upcall cookie */
1291 lu_ref_add(&lock->cll_reference, "upcall", lock);
1292 ols->ols_state = OLS_ENQUEUED;
1295 * XXX: this is possible blocking point as
1296 * ldlm_lock_match(LDLM_FL_LVB_READY) waits for
1299 result = osc_enqueue_base(osc_export(obj), resname,
1300 &ols->ols_flags, policy,
1302 obj->oo_oinfo->loi_kms_valid,
1304 ols, einfo, &ols->ols_handle,
1307 lu_ref_del(&lock->cll_reference,
1309 cl_lock_put(env, lock);
1312 ols->ols_state = OLS_GRANTED;
1315 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
1319 static int osc_lock_wait(const struct lu_env *env,
1320 const struct cl_lock_slice *slice)
1322 struct osc_lock *olck = cl2osc_lock(slice);
1323 struct cl_lock *lock = olck->ols_cl.cls_lock;
1325 LINVRNT(osc_lock_invariant(olck));
1326 if (olck->ols_glimpse && olck->ols_state >= OLS_UPCALL_RECEIVED)
1329 LASSERT(equi(olck->ols_state >= OLS_UPCALL_RECEIVED &&
1330 lock->cll_error == 0, olck->ols_lock != NULL));
1332 return lock->cll_error ?: olck->ols_state >= OLS_GRANTED ? 0 : CLO_WAIT;
1336 * An implementation of cl_lock_operations::clo_use() method that pins cached
1339 static int osc_lock_use(const struct lu_env *env,
1340 const struct cl_lock_slice *slice)
1342 struct osc_lock *olck = cl2osc_lock(slice);
1345 LASSERT(!olck->ols_hold);
1347 * Atomically check for LDLM_FL_CBPENDING and addref a lock if this
1348 * flag is not set. This protects us from a concurrent blocking ast.
1350 rc = ldlm_lock_addref_try(&olck->ols_handle, olck->ols_einfo.ei_mode);
1352 olck->ols_hold = olck->ols_has_ref = 1;
1353 olck->ols_state = OLS_GRANTED;
1355 struct cl_lock *lock;
1358 * Lock is being cancelled somewhere within
1359 * ldlm_handle_bl_callback(): LDLM_FL_CBPENDING is already
1360 * set, but osc_ldlm_blocking_ast() hasn't yet acquired
1363 lock = slice->cls_lock;
1364 LASSERT(lock->cll_state == CLS_CACHED);
1365 LASSERT(lock->cll_users > 0);
1366 LASSERT(olck->ols_lock->l_flags & LDLM_FL_CBPENDING);
1367 /* set a flag for osc_dlm_blocking_ast0() to signal the
1369 olck->ols_ast_wait = 1;
1375 static int osc_lock_flush(struct osc_lock *ols, int discard)
1377 struct cl_lock *lock = ols->ols_cl.cls_lock;
1378 struct cl_env_nest nest;
1382 env = cl_env_nested_get(&nest);
1384 result = cl_lock_page_out(env, lock, discard);
1385 cl_env_nested_put(&nest, env);
1387 result = PTR_ERR(env);
1394 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1395 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1396 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1397 * with some other lock some where in the cluster. This function does the
1400 * - invalidates all pages protected by this lock (after sending dirty
1401 * ones to the server, as necessary);
1403 * - decref's underlying ldlm lock;
1405 * - cancels ldlm lock (ldlm_cli_cancel()).
1407 static void osc_lock_cancel(const struct lu_env *env,
1408 const struct cl_lock_slice *slice)
1410 struct cl_lock *lock = slice->cls_lock;
1411 struct osc_lock *olck = cl2osc_lock(slice);
1412 struct ldlm_lock *dlmlock = olck->ols_lock;
1416 LASSERT(cl_lock_is_mutexed(lock));
1417 LINVRNT(osc_lock_invariant(olck));
1419 if (dlmlock != NULL) {
1422 discard = dlmlock->l_flags & LDLM_FL_DISCARD_DATA;
1423 result = osc_lock_flush(olck, discard);
1425 osc_lock_unuse(env, slice);
1427 lock_res_and_lock(dlmlock);
1428 /* Now that we're the only user of dlm read/write reference,
1429 * mostly the ->l_readers + ->l_writers should be zero.
1430 * However, there is a corner case.
1431 * See bug 18829 for details.*/
1432 do_cancel = (dlmlock->l_readers == 0 &&
1433 dlmlock->l_writers == 0);
1434 dlmlock->l_flags |= LDLM_FL_CBPENDING;
1435 unlock_res_and_lock(dlmlock);
1437 result = ldlm_cli_cancel(&olck->ols_handle);
1439 CL_LOCK_DEBUG(D_ERROR, env, lock,
1440 "lock %p cancel failure with error(%d)\n",
1443 olck->ols_state = OLS_CANCELLED;
1444 osc_lock_detach(env, olck);
1447 void cl_lock_page_list_fixup(const struct lu_env *env,
1448 struct cl_io *io, struct cl_lock *lock,
1449 struct cl_page_list *queue);
1451 #ifdef INVARIANT_CHECK
1453 * Returns true iff there are pages under \a olck not protected by other
1456 static int osc_lock_has_pages(struct osc_lock *olck)
1458 struct cl_lock *lock;
1459 struct cl_lock_descr *descr;
1460 struct cl_object *obj;
1461 struct osc_object *oob;
1462 struct cl_page_list *plist;
1463 struct cl_page *page;
1464 struct cl_env_nest nest;
1469 env = cl_env_nested_get(&nest);
1471 obj = olck->ols_cl.cls_obj;
1473 io = &oob->oo_debug_io;
1474 lock = olck->ols_cl.cls_lock;
1475 descr = &lock->cll_descr;
1476 plist = &osc_env_info(env)->oti_plist;
1477 cl_page_list_init(plist);
1479 mutex_lock(&oob->oo_debug_mutex);
1481 io->ci_obj = cl_object_top(obj);
1482 cl_io_init(env, io, CIT_MISC, io->ci_obj);
1483 cl_page_gang_lookup(env, obj, io,
1484 descr->cld_start, descr->cld_end, plist);
1485 cl_lock_page_list_fixup(env, io, lock, plist);
1486 if (plist->pl_nr > 0) {
1487 CL_LOCK_DEBUG(D_ERROR, env, lock, "still has pages\n");
1488 cl_page_list_for_each(page, plist)
1489 CL_PAGE_DEBUG(D_ERROR, env, page, "\n");
1491 result = plist->pl_nr > 0;
1492 cl_page_list_disown(env, io, plist);
1493 cl_page_list_fini(env, plist);
1494 cl_io_fini(env, io);
1495 mutex_unlock(&oob->oo_debug_mutex);
1496 cl_env_nested_put(&nest, env);
1502 # define osc_lock_has_pages(olck) (0)
1503 #endif /* INVARIANT_CHECK */
1505 static void osc_lock_delete(const struct lu_env *env,
1506 const struct cl_lock_slice *slice)
1508 struct osc_lock *olck;
1510 olck = cl2osc_lock(slice);
1511 LINVRNT(osc_lock_invariant(olck));
1512 LINVRNT(!osc_lock_has_pages(olck));
1515 osc_lock_unuse(env, slice);
1516 osc_lock_detach(env, olck);
1520 * Implements cl_lock_operations::clo_state() method for osc layer.
1522 * Maintains osc_lock::ols_owner field.
1524 * This assumes that lock always enters CLS_HELD (from some other state) in
1525 * the same IO context as one that requested the lock. This should not be a
1526 * problem, because context is by definition shared by all activity pertaining
1527 * to the same high-level IO.
1529 static void osc_lock_state(const struct lu_env *env,
1530 const struct cl_lock_slice *slice,
1531 enum cl_lock_state state)
1533 struct osc_lock *lock = cl2osc_lock(slice);
1534 struct osc_io *oio = osc_env_io(env);
1537 * XXX multiple io contexts can use the lock at the same time.
1539 LINVRNT(osc_lock_invariant(lock));
1540 if (state == CLS_HELD && slice->cls_lock->cll_state != CLS_HELD) {
1541 LASSERT(lock->ols_owner == NULL);
1542 lock->ols_owner = oio;
1543 } else if (state != CLS_HELD)
1544 lock->ols_owner = NULL;
1547 static int osc_lock_print(const struct lu_env *env, void *cookie,
1548 lu_printer_t p, const struct cl_lock_slice *slice)
1550 struct osc_lock *lock = cl2osc_lock(slice);
1553 * XXX print ldlm lock and einfo properly.
1555 (*p)(env, cookie, "%p %08x "LPU64" %d %p ",
1556 lock->ols_lock, lock->ols_flags, lock->ols_handle.cookie,
1557 lock->ols_state, lock->ols_owner);
1558 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1562 static const struct cl_lock_operations osc_lock_ops = {
1563 .clo_fini = osc_lock_fini,
1564 .clo_enqueue = osc_lock_enqueue,
1565 .clo_wait = osc_lock_wait,
1566 .clo_unuse = osc_lock_unuse,
1567 .clo_use = osc_lock_use,
1568 .clo_delete = osc_lock_delete,
1569 .clo_state = osc_lock_state,
1570 .clo_cancel = osc_lock_cancel,
1571 .clo_weigh = osc_lock_weigh,
1572 .clo_print = osc_lock_print
1575 static int osc_lock_lockless_enqueue(const struct lu_env *env,
1576 const struct cl_lock_slice *slice,
1577 struct cl_io *_, __u32 enqflags)
1583 static int osc_lock_lockless_unuse(const struct lu_env *env,
1584 const struct cl_lock_slice *slice)
1586 struct osc_lock *ols = cl2osc_lock(slice);
1587 struct cl_lock *lock = slice->cls_lock;
1589 LASSERT(ols->ols_state == OLS_GRANTED);
1590 LINVRNT(osc_lock_invariant(ols));
1592 cl_lock_cancel(env, lock);
1593 cl_lock_delete(env, lock);
1597 static void osc_lock_lockless_cancel(const struct lu_env *env,
1598 const struct cl_lock_slice *slice)
1600 struct osc_lock *ols = cl2osc_lock(slice);
1603 result = osc_lock_flush(ols, 0);
1605 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1607 ols->ols_state = OLS_CANCELLED;
1610 static int osc_lock_lockless_wait(const struct lu_env *env,
1611 const struct cl_lock_slice *slice)
1613 struct osc_lock *olck = cl2osc_lock(slice);
1614 struct cl_lock *lock = olck->ols_cl.cls_lock;
1616 LINVRNT(osc_lock_invariant(olck));
1617 LASSERT(olck->ols_state >= OLS_UPCALL_RECEIVED);
1619 return lock->cll_error;
1622 static void osc_lock_lockless_state(const struct lu_env *env,
1623 const struct cl_lock_slice *slice,
1624 enum cl_lock_state state)
1626 struct osc_lock *lock = cl2osc_lock(slice);
1627 struct osc_io *oio = osc_env_io(env);
1629 LINVRNT(osc_lock_invariant(lock));
1630 if (state == CLS_HELD) {
1631 LASSERT(lock->ols_owner == NULL);
1632 lock->ols_owner = oio;
1634 /* set the io to be lockless if this lock is for io's
1636 if (cl_object_same(oio->oi_cl.cis_obj, slice->cls_obj))
1637 oio->oi_lockless = 1;
1639 lock->ols_owner = NULL;
1642 static int osc_lock_lockless_fits_into(const struct lu_env *env,
1643 const struct cl_lock_slice *slice,
1644 const struct cl_lock_descr *need,
1645 const struct cl_io *io)
1650 static const struct cl_lock_operations osc_lock_lockless_ops = {
1651 .clo_fini = osc_lock_fini,
1652 .clo_enqueue = osc_lock_lockless_enqueue,
1653 .clo_wait = osc_lock_lockless_wait,
1654 .clo_unuse = osc_lock_lockless_unuse,
1655 .clo_state = osc_lock_lockless_state,
1656 .clo_fits_into = osc_lock_lockless_fits_into,
1657 .clo_cancel = osc_lock_lockless_cancel,
1658 .clo_print = osc_lock_print
1661 int osc_lock_init(const struct lu_env *env,
1662 struct cl_object *obj, struct cl_lock *lock,
1663 const struct cl_io *_)
1665 struct osc_lock *clk;
1668 OBD_SLAB_ALLOC_PTR_GFP(clk, osc_lock_kmem, CFS_ALLOC_IO);
1670 osc_lock_build_einfo(env, lock, clk, &clk->ols_einfo);
1671 clk->ols_state = OLS_NEW;
1672 cl_lock_slice_add(lock, &clk->ols_cl, obj, &osc_lock_ops);