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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Implementation of cl_lock for OSC layer.
34 * Author: Nikita Danilov <nikita.danilov@sun.com>
35 * Author: Jinshan Xiong <jinshan.xiong@intel.com>
38 #define DEBUG_SUBSYSTEM S_OSC
40 /* fid_build_reg_res_name() */
41 #include <lustre_fid.h>
42 #include <lustre_osc.h>
44 #include "osc_internal.h"
51 * Returns a weak pointer to the ldlm lock identified by a handle. Returned
52 * pointer cannot be dereferenced, as lock is not protected from concurrent
53 * reclaim. This function is a helper for osc_lock_invariant().
55 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
57 struct ldlm_lock *lock;
59 lock = ldlm_handle2lock(handle);
66 * Invariant that has to be true all of the time.
68 static int osc_lock_invariant(struct osc_lock *ols)
70 struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle);
71 struct ldlm_lock *olock = ols->ols_dlmlock;
72 int handle_used = lustre_handle_is_used(&ols->ols_handle);
74 if (ergo(osc_lock_is_lockless(ols),
75 ols->ols_locklessable && ols->ols_dlmlock == NULL))
79 * If all the following "ergo"s are true, return 1, otherwise 0
81 if (! ergo(olock != NULL, handle_used))
84 if (! ergo(olock != NULL,
85 olock->l_handle.h_cookie == ols->ols_handle.cookie))
88 if (! ergo(handle_used,
89 ergo(lock != NULL && olock != NULL, lock == olock) &&
90 ergo(lock == NULL, olock == NULL)))
93 * Check that ->ols_handle and ->ols_dlmlock are consistent, but
94 * take into account that they are set at the different time.
96 if (! ergo(ols->ols_state == OLS_CANCELLED,
97 olock == NULL && !handle_used))
100 * DLM lock is destroyed only after we have seen cancellation
103 if (! ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
104 !ldlm_is_destroyed(olock)))
107 if (! ergo(ols->ols_state == OLS_GRANTED,
109 olock->l_req_mode == olock->l_granted_mode &&
115 /*****************************************************************************
121 void osc_lock_fini(const struct lu_env *env, struct cl_lock_slice *slice)
123 struct osc_lock *ols = cl2osc_lock(slice);
125 LINVRNT(osc_lock_invariant(ols));
126 LASSERT(ols->ols_dlmlock == NULL);
128 OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
130 EXPORT_SYMBOL(osc_lock_fini);
132 static void osc_lock_build_policy(const struct lu_env *env,
133 const struct cl_lock *lock,
134 union ldlm_policy_data *policy)
136 const struct cl_lock_descr *d = &lock->cll_descr;
138 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
139 policy->l_extent.gid = d->cld_gid;
143 * Updates object attributes from a lock value block (lvb) received together
144 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
147 * This can be optimized to not update attributes when lock is a result of a
150 * Called under lock and resource spin-locks.
152 static void osc_lock_lvb_update(const struct lu_env *env,
153 struct osc_object *osc,
154 struct ldlm_lock *dlmlock,
157 struct cl_object *obj = osc2cl(osc);
158 struct lov_oinfo *oinfo = osc->oo_oinfo;
159 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
164 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
166 LASSERT(dlmlock != NULL);
167 lvb = dlmlock->l_lvb_data;
169 cl_lvb2attr(attr, lvb);
171 cl_object_attr_lock(obj);
172 if (dlmlock != NULL) {
175 check_res_locked(dlmlock->l_resource);
177 LASSERT(lvb == dlmlock->l_lvb_data);
178 size = lvb->lvb_size;
180 /* Extend KMS up to the end of this lock and no further
181 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
182 if (size > dlmlock->l_policy_data.l_extent.end)
183 size = dlmlock->l_policy_data.l_extent.end + 1;
184 if (size >= oinfo->loi_kms) {
185 LDLM_DEBUG(dlmlock, "lock acquired, setting rss=%llu"
186 ", kms=%llu", lvb->lvb_size, size);
188 attr->cat_kms = size;
190 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
191 "%llu; leaving kms=%llu, end=%llu",
192 lvb->lvb_size, oinfo->loi_kms,
193 dlmlock->l_policy_data.l_extent.end);
195 ldlm_lock_allow_match_locked(dlmlock);
198 cl_object_attr_update(env, obj, attr, valid);
199 cl_object_attr_unlock(obj);
204 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *oscl,
205 struct lustre_handle *lockh, bool lvb_update)
207 struct ldlm_lock *dlmlock;
209 dlmlock = ldlm_handle2lock_long(lockh, 0);
210 LASSERT(dlmlock != NULL);
212 /* lock reference taken by ldlm_handle2lock_long() is
213 * owned by osc_lock and released in osc_lock_detach()
215 lu_ref_add(&dlmlock->l_reference, "osc_lock", oscl);
216 oscl->ols_has_ref = 1;
218 LASSERT(oscl->ols_dlmlock == NULL);
219 oscl->ols_dlmlock = dlmlock;
221 /* This may be a matched lock for glimpse request, do not hold
222 * lock reference in that case. */
223 if (!oscl->ols_glimpse) {
224 /* hold a refc for non glimpse lock which will
225 * be released in osc_lock_cancel() */
226 lustre_handle_copy(&oscl->ols_handle, lockh);
227 ldlm_lock_addref(lockh, oscl->ols_einfo.ei_mode);
231 /* Lock must have been granted. */
232 lock_res_and_lock(dlmlock);
233 if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
234 struct ldlm_extent *ext = &dlmlock->l_policy_data.l_extent;
235 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
237 /* extend the lock extent, otherwise it will have problem when
238 * we decide whether to grant a lockless lock. */
239 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
240 descr->cld_start = cl_index(descr->cld_obj, ext->start);
241 descr->cld_end = cl_index(descr->cld_obj, ext->end);
242 descr->cld_gid = ext->gid;
244 /* no lvb update for matched lock */
246 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
247 osc_lock_lvb_update(env, cl2osc(oscl->ols_cl.cls_obj),
250 LINVRNT(osc_lock_invariant(oscl));
252 unlock_res_and_lock(dlmlock);
254 LASSERT(oscl->ols_state != OLS_GRANTED);
255 oscl->ols_state = OLS_GRANTED;
259 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
260 * received from a server, or after osc_enqueue_base() matched a local DLM
263 static int osc_lock_upcall(void *cookie, struct lustre_handle *lockh,
266 struct osc_lock *oscl = cookie;
267 struct cl_lock_slice *slice = &oscl->ols_cl;
273 env = cl_env_percpu_get();
274 /* should never happen, similar to osc_ldlm_blocking_ast(). */
275 LASSERT(!IS_ERR(env));
277 rc = ldlm_error2errno(errcode);
278 if (oscl->ols_state == OLS_ENQUEUED) {
279 oscl->ols_state = OLS_UPCALL_RECEIVED;
280 } else if (oscl->ols_state == OLS_CANCELLED) {
283 CERROR("Impossible state: %d\n", oscl->ols_state);
288 osc_lock_granted(env, oscl, lockh, errcode == ELDLM_OK);
290 /* Error handling, some errors are tolerable. */
291 if (oscl->ols_locklessable && rc == -EUSERS) {
292 /* This is a tolerable error, turn this lock into
295 osc_object_set_contended(cl2osc(slice->cls_obj));
296 LASSERT(slice->cls_ops != oscl->ols_lockless_ops);
298 /* Change this lock to ldlmlock-less lock. */
299 osc_lock_to_lockless(env, oscl, 1);
300 oscl->ols_state = OLS_GRANTED;
302 } else if (oscl->ols_glimpse && rc == -ENAVAIL) {
303 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
304 osc_lock_lvb_update(env, cl2osc(slice->cls_obj),
305 NULL, &oscl->ols_lvb);
306 /* Hide the error. */
308 } else if (rc < 0 && oscl->ols_flags & LDLM_FL_NDELAY) {
312 if (oscl->ols_owner != NULL)
313 cl_sync_io_note(env, oscl->ols_owner, rc);
314 cl_env_percpu_put(env);
319 static int osc_lock_upcall_speculative(void *cookie,
320 struct lustre_handle *lockh,
323 struct osc_object *osc = cookie;
324 struct ldlm_lock *dlmlock;
329 env = cl_env_get(&refcheck);
330 LASSERT(!IS_ERR(env));
332 if (errcode == ELDLM_LOCK_MATCHED)
333 GOTO(out, errcode = ELDLM_OK);
335 if (errcode != ELDLM_OK)
338 dlmlock = ldlm_handle2lock(lockh);
339 LASSERT(dlmlock != NULL);
341 lock_res_and_lock(dlmlock);
342 LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
344 /* there is no osc_lock associated with speculative locks */
345 osc_lock_lvb_update(env, osc, dlmlock, NULL);
347 unlock_res_and_lock(dlmlock);
348 LDLM_LOCK_PUT(dlmlock);
351 cl_object_put(env, osc2cl(osc));
352 cl_env_put(env, &refcheck);
353 RETURN(ldlm_error2errno(errcode));
356 static int osc_lock_flush(struct osc_object *obj, pgoff_t start, pgoff_t end,
357 enum cl_lock_mode mode, bool discard)
366 env = cl_env_get(&refcheck);
368 RETURN(PTR_ERR(env));
370 if (mode == CLM_WRITE) {
371 rc = osc_cache_writeback_range(env, obj, start, end, 1,
373 CDEBUG(D_CACHE, "object %p: [%lu -> %lu] %d pages were %s.\n",
375 discard ? "discarded" : "written back");
380 rc2 = osc_lock_discard_pages(env, obj, start, end, discard);
381 if (rc == 0 && rc2 < 0)
384 cl_env_put(env, &refcheck);
389 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
390 * and ldlm_lock caches.
392 static int osc_dlm_blocking_ast0(const struct lu_env *env,
393 struct ldlm_lock *dlmlock,
394 void *data, int flag)
396 struct cl_object *obj = NULL;
399 enum cl_lock_mode mode = CLM_READ;
402 LASSERT(flag == LDLM_CB_CANCELING);
404 lock_res_and_lock(dlmlock);
405 if (dlmlock->l_granted_mode != dlmlock->l_req_mode) {
406 dlmlock->l_ast_data = NULL;
407 unlock_res_and_lock(dlmlock);
411 discard = ldlm_is_discard_data(dlmlock);
412 if (dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))
415 if (dlmlock->l_ast_data != NULL) {
416 obj = osc2cl(dlmlock->l_ast_data);
417 dlmlock->l_ast_data = NULL;
422 unlock_res_and_lock(dlmlock);
424 /* if l_ast_data is NULL, the dlmlock was enqueued by AGL or
425 * the object has been destroyed. */
427 struct ldlm_extent *extent = &dlmlock->l_policy_data.l_extent;
428 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
431 /* Destroy pages covered by the extent of the DLM lock */
432 result = osc_lock_flush(cl2osc(obj),
433 cl_index(obj, extent->start),
434 cl_index(obj, extent->end),
437 /* losing a lock, update kms */
438 lock_res_and_lock(dlmlock);
439 cl_object_attr_lock(obj);
440 /* Must get the value under the lock to avoid race. */
441 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
442 /* Update the kms. Need to loop all granted locks.
443 * Not a problem for the client */
444 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
446 cl_object_attr_update(env, obj, attr, CAT_KMS);
447 cl_object_attr_unlock(obj);
448 unlock_res_and_lock(dlmlock);
450 cl_object_put(env, obj);
456 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
457 * some other lock, or is canceled. This function is installed as a
458 * ldlm_lock::l_blocking_ast() for client extent locks.
460 * Control flow is tricky, because ldlm uses the same call-back
461 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
463 * \param dlmlock lock for which ast occurred.
465 * \param new description of a conflicting lock in case of blocking ast.
467 * \param data value of dlmlock->l_ast_data
469 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
470 * cancellation and blocking ast's.
472 * Possible use cases:
474 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
475 * lock due to lock lru pressure, or explicit user request to purge
478 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
479 * us that dlmlock conflicts with another lock that some client is
480 * enqueuing. Lock is canceled.
482 * - cl_lock_cancel() is called. osc_lock_cancel() calls
483 * ldlm_cli_cancel() that calls
485 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
487 * recursively entering osc_ldlm_blocking_ast().
489 * - client cancels lock voluntary (e.g., as a part of early cancellation):
492 * osc_lock_cancel()->
493 * ldlm_cli_cancel()->
494 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
497 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
498 struct ldlm_lock_desc *new, void *data,
505 case LDLM_CB_BLOCKING: {
506 struct lustre_handle lockh;
508 ldlm_lock2handle(dlmlock, &lockh);
509 result = ldlm_cli_cancel(&lockh, LCF_ASYNC);
510 if (result == -ENODATA)
514 case LDLM_CB_CANCELING: {
519 * This can be called in the context of outer IO, e.g.,
521 * osc_enqueue_base()->...
522 * ->ldlm_prep_elc_req()->...
523 * ->ldlm_cancel_callback()->...
524 * ->osc_ldlm_blocking_ast()
526 * new environment has to be created to not corrupt outer
529 env = cl_env_get(&refcheck);
531 result = PTR_ERR(env);
535 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
536 cl_env_put(env, &refcheck);
545 int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
547 struct ptlrpc_request *req = data;
550 struct req_capsule *cap;
551 struct cl_object *obj = NULL;
557 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
559 env = cl_env_get(&refcheck);
561 GOTO(out, result = PTR_ERR(env));
564 lock_res_and_lock(dlmlock);
565 if (dlmlock->l_ast_data != NULL) {
566 obj = osc2cl(dlmlock->l_ast_data);
569 unlock_res_and_lock(dlmlock);
572 /* Do not grab the mutex of cl_lock for glimpse.
573 * See LU-1274 for details.
574 * BTW, it's okay for cl_lock to be cancelled during
575 * this period because server can handle this race.
576 * See ldlm_server_glimpse_ast() for details.
577 * cl_lock_mutex_get(env, lock); */
579 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
580 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
582 result = req_capsule_server_pack(cap);
584 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
585 result = cl_object_glimpse(env, obj, lvb);
587 if (!exp_connect_lvb_type(req->rq_export))
588 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
589 sizeof(struct ost_lvb_v1), RCL_SERVER);
590 cl_object_put(env, obj);
593 * These errors are normal races, so we don't want to
594 * fill the console with messages by calling
597 lustre_pack_reply(req, 1, NULL, NULL);
598 result = -ELDLM_NO_LOCK_DATA;
600 cl_env_put(env, &refcheck);
604 req->rq_status = result;
607 EXPORT_SYMBOL(osc_ldlm_glimpse_ast);
609 static int weigh_cb(const struct lu_env *env, struct cl_io *io,
610 struct osc_page *ops, void *cbdata)
612 struct cl_page *page = ops->ops_cl.cpl_page;
614 if (cl_page_is_vmlocked(env, page)
615 || PageDirty(page->cp_vmpage) || PageWriteback(page->cp_vmpage)
617 return CLP_GANG_ABORT;
619 *(pgoff_t *)cbdata = osc_index(ops) + 1;
620 return CLP_GANG_OKAY;
623 static unsigned long osc_lock_weight(const struct lu_env *env,
624 struct osc_object *oscobj,
625 struct ldlm_extent *extent)
627 struct cl_io *io = osc_env_thread_io(env);
628 struct cl_object *obj = cl_object_top(&oscobj->oo_cl);
634 io->ci_ignore_layout = 1;
635 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
639 page_index = cl_index(obj, extent->start);
641 result = osc_page_gang_lookup(env, io, oscobj,
643 cl_index(obj, extent->end),
644 weigh_cb, (void *)&page_index);
645 if (result == CLP_GANG_ABORT)
647 if (result == CLP_GANG_RESCHED)
649 } while (result != CLP_GANG_OKAY);
652 return result == CLP_GANG_ABORT ? 1 : 0;
656 * Get the weight of dlm lock for early cancellation.
658 unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
661 struct osc_object *obj;
662 struct osc_lock *oscl;
663 unsigned long weight;
670 * osc_ldlm_weigh_ast has a complex context since it might be called
671 * because of lock canceling, or from user's input. We have to make
672 * a new environment for it. Probably it is implementation safe to use
673 * the upper context because cl_lock_put don't modify environment
674 * variables. But just in case ..
676 env = cl_env_get(&refcheck);
678 /* Mostly because lack of memory, do not eliminate this lock */
681 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
682 lock_res_and_lock(dlmlock);
683 obj = dlmlock->l_ast_data;
685 cl_object_get(osc2cl(obj));
686 unlock_res_and_lock(dlmlock);
689 GOTO(out, weight = 1);
691 spin_lock(&obj->oo_ol_spin);
692 list_for_each_entry(oscl, &obj->oo_ol_list, ols_nextlock_oscobj) {
693 if (oscl->ols_dlmlock != NULL && oscl->ols_dlmlock != dlmlock)
697 spin_unlock(&obj->oo_ol_spin);
700 * If the lock is being used by an IO, definitely not cancel it.
702 GOTO(out, weight = 1);
705 weight = osc_lock_weight(env, obj, &dlmlock->l_policy_data.l_extent);
710 cl_object_put(env, osc2cl(obj));
712 cl_env_put(env, &refcheck);
716 static void osc_lock_build_einfo(const struct lu_env *env,
717 const struct cl_lock *lock,
718 struct osc_object *osc,
719 struct ldlm_enqueue_info *einfo)
721 einfo->ei_type = LDLM_EXTENT;
722 einfo->ei_mode = osc_cl_lock2ldlm(lock->cll_descr.cld_mode);
723 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
724 einfo->ei_cb_cp = ldlm_completion_ast;
725 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
726 einfo->ei_cbdata = osc; /* value to be put into ->l_ast_data */
730 * Determine if the lock should be converted into a lockless lock.
733 * - if the lock has an explicite requirment for a non-lockless lock;
734 * - if the io lock request type ci_lockreq;
735 * - send the enqueue rpc to ost to make the further decision;
736 * - special treat to truncate lockless lock
738 * Additional policy can be implemented here, e.g., never do lockless-io
741 void osc_lock_to_lockless(const struct lu_env *env,
742 struct osc_lock *ols, int force)
744 struct cl_lock_slice *slice = &ols->ols_cl;
745 struct osc_io *oio = osc_env_io(env);
746 struct cl_io *io = oio->oi_cl.cis_io;
747 struct cl_object *obj = slice->cls_obj;
748 struct osc_object *oob = cl2osc(obj);
749 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
750 struct obd_connect_data *ocd;
752 LASSERT(ols->ols_state == OLS_NEW ||
753 ols->ols_state == OLS_UPCALL_RECEIVED);
756 ols->ols_locklessable = 1;
757 slice->cls_ops = ols->ols_lockless_ops;
759 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
760 io->ci_lockreq == CILR_MAYBE ||
761 io->ci_lockreq == CILR_NEVER);
763 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
764 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
765 (io->ci_lockreq == CILR_MAYBE) &&
766 (ocd->ocd_connect_flags &
767 OBD_CONNECT_SRVLOCK);
768 if (io->ci_lockreq == CILR_NEVER ||
770 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
771 /* lockless truncate */
772 (cl_io_is_trunc(io) && osd->od_lockless_truncate &&
773 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK))) {
774 ols->ols_locklessable = 1;
775 slice->cls_ops = ols->ols_lockless_ops;
778 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
780 EXPORT_SYMBOL(osc_lock_to_lockless);
782 static bool osc_lock_compatible(const struct osc_lock *qing,
783 const struct osc_lock *qed)
785 struct cl_lock_descr *qed_descr = &qed->ols_cl.cls_lock->cll_descr;
786 struct cl_lock_descr *qing_descr = &qing->ols_cl.cls_lock->cll_descr;
788 if (qed->ols_glimpse || qed->ols_speculative)
791 if (qing_descr->cld_mode == CLM_READ && qed_descr->cld_mode == CLM_READ)
794 if (qed->ols_state < OLS_GRANTED)
797 if (qed_descr->cld_mode >= qing_descr->cld_mode &&
798 qed_descr->cld_start <= qing_descr->cld_start &&
799 qed_descr->cld_end >= qing_descr->cld_end)
805 void osc_lock_wake_waiters(const struct lu_env *env, struct osc_object *osc,
806 struct osc_lock *oscl)
808 spin_lock(&osc->oo_ol_spin);
809 list_del_init(&oscl->ols_nextlock_oscobj);
810 spin_unlock(&osc->oo_ol_spin);
812 spin_lock(&oscl->ols_lock);
813 while (!list_empty(&oscl->ols_waiting_list)) {
814 struct osc_lock *scan;
816 scan = list_entry(oscl->ols_waiting_list.next, struct osc_lock,
818 list_del_init(&scan->ols_wait_entry);
820 cl_sync_io_note(env, scan->ols_owner, 0);
822 spin_unlock(&oscl->ols_lock);
824 EXPORT_SYMBOL(osc_lock_wake_waiters);
826 int osc_lock_enqueue_wait(const struct lu_env *env, struct osc_object *obj,
827 struct osc_lock *oscl)
829 struct osc_lock *tmp_oscl;
830 struct cl_lock_descr *need = &oscl->ols_cl.cls_lock->cll_descr;
831 struct cl_sync_io *waiter = &osc_env_info(env)->oti_anchor;
836 spin_lock(&obj->oo_ol_spin);
837 list_add_tail(&oscl->ols_nextlock_oscobj, &obj->oo_ol_list);
840 list_for_each_entry(tmp_oscl, &obj->oo_ol_list,
841 ols_nextlock_oscobj) {
842 struct cl_lock_descr *descr;
844 if (tmp_oscl == oscl)
847 descr = &tmp_oscl->ols_cl.cls_lock->cll_descr;
848 if (descr->cld_start > need->cld_end ||
849 descr->cld_end < need->cld_start)
852 /* We're not supposed to give up group lock */
853 if (descr->cld_mode == CLM_GROUP)
856 if (!osc_lock_is_lockless(oscl) &&
857 osc_lock_compatible(oscl, tmp_oscl))
860 /* wait for conflicting lock to be canceled */
861 cl_sync_io_init(waiter, 1, cl_sync_io_end);
862 oscl->ols_owner = waiter;
864 spin_lock(&tmp_oscl->ols_lock);
865 /* add oscl into tmp's ols_waiting list */
866 list_add_tail(&oscl->ols_wait_entry,
867 &tmp_oscl->ols_waiting_list);
868 spin_unlock(&tmp_oscl->ols_lock);
870 spin_unlock(&obj->oo_ol_spin);
871 rc = cl_sync_io_wait(env, waiter, 0);
872 spin_lock(&obj->oo_ol_spin);
877 oscl->ols_owner = NULL;
880 spin_unlock(&obj->oo_ol_spin);
884 EXPORT_SYMBOL(osc_lock_enqueue_wait);
887 * Implementation of cl_lock_operations::clo_enqueue() method for osc
888 * layer. This initiates ldlm enqueue:
890 * - cancels conflicting locks early (osc_lock_enqueue_wait());
892 * - calls osc_enqueue_base() to do actual enqueue.
894 * osc_enqueue_base() is supplied with an upcall function that is executed
895 * when lock is received either after a local cached ldlm lock is matched, or
896 * when a reply from the server is received.
898 * This function does not wait for the network communication to complete.
900 static int osc_lock_enqueue(const struct lu_env *env,
901 const struct cl_lock_slice *slice,
902 struct cl_io *unused, struct cl_sync_io *anchor)
904 struct osc_thread_info *info = osc_env_info(env);
905 struct osc_io *oio = osc_env_io(env);
906 struct osc_object *osc = cl2osc(slice->cls_obj);
907 struct osc_lock *oscl = cl2osc_lock(slice);
908 struct obd_export *exp = osc_export(osc);
909 struct cl_lock *lock = slice->cls_lock;
910 struct ldlm_res_id *resname = &info->oti_resname;
911 union ldlm_policy_data *policy = &info->oti_policy;
912 osc_enqueue_upcall_f upcall = osc_lock_upcall;
919 LASSERTF(ergo(oscl->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
920 "lock = %p, ols = %p\n", lock, oscl);
922 if (oscl->ols_state == OLS_GRANTED)
925 if ((oscl->ols_flags & LDLM_FL_NO_EXPANSION) &&
926 !(exp_connect_lockahead_old(exp) || exp_connect_lockahead(exp))) {
927 result = -EOPNOTSUPP;
928 CERROR("%s: server does not support lockahead/locknoexpand:"
929 "rc = %d\n", exp->exp_obd->obd_name, result);
933 if (oscl->ols_flags & LDLM_FL_TEST_LOCK)
934 GOTO(enqueue_base, 0);
936 /* For glimpse and/or speculative locks, do not wait for reply from
937 * server on LDLM request */
938 if (oscl->ols_glimpse || oscl->ols_speculative) {
939 /* Speculative and glimpse locks do not have an anchor */
940 LASSERT(equi(oscl->ols_speculative, anchor == NULL));
942 GOTO(enqueue_base, 0);
945 result = osc_lock_enqueue_wait(env, osc, oscl);
949 /* we can grant lockless lock right after all conflicting locks
951 if (osc_lock_is_lockless(oscl)) {
952 oscl->ols_state = OLS_GRANTED;
953 oio->oi_lockless = 1;
958 oscl->ols_state = OLS_ENQUEUED;
959 if (anchor != NULL) {
960 atomic_inc(&anchor->csi_sync_nr);
961 oscl->ols_owner = anchor;
965 * DLM lock's ast data must be osc_object;
966 * if glimpse or speculative lock, async of osc_enqueue_base()
969 * For non-speculative locks:
970 * DLM's enqueue callback set to osc_lock_upcall() with cookie as
972 * For speculative locks:
973 * osc_lock_upcall_speculative & cookie is the osc object, since
974 * there is no osc_lock
976 ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
977 osc_lock_build_policy(env, lock, policy);
978 if (oscl->ols_speculative) {
979 oscl->ols_einfo.ei_cbdata = NULL;
980 /* hold a reference for callback */
981 cl_object_get(osc2cl(osc));
982 upcall = osc_lock_upcall_speculative;
985 result = osc_enqueue_base(exp, resname, &oscl->ols_flags,
986 policy, &oscl->ols_lvb,
987 osc->oo_oinfo->loi_kms_valid,
989 &oscl->ols_einfo, PTLRPCD_SET, async,
990 oscl->ols_speculative);
992 if (osc_lock_is_lockless(oscl)) {
993 oio->oi_lockless = 1;
995 LASSERT(oscl->ols_state == OLS_GRANTED);
996 LASSERT(oscl->ols_hold);
997 LASSERT(oscl->ols_dlmlock != NULL);
999 } else if (oscl->ols_speculative) {
1000 cl_object_put(env, osc2cl(osc));
1001 if (oscl->ols_glimpse) {
1002 /* hide error for AGL request */
1009 oscl->ols_state = OLS_CANCELLED;
1010 osc_lock_wake_waiters(env, osc, oscl);
1013 cl_sync_io_note(env, anchor, result);
1019 * Breaks a link between osc_lock and dlm_lock.
1021 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
1023 struct ldlm_lock *dlmlock;
1027 dlmlock = olck->ols_dlmlock;
1028 if (dlmlock == NULL)
1031 if (olck->ols_hold) {
1033 ldlm_lock_decref(&olck->ols_handle, olck->ols_einfo.ei_mode);
1034 olck->ols_handle.cookie = 0ULL;
1037 olck->ols_dlmlock = NULL;
1039 /* release a reference taken in osc_lock_upcall(). */
1040 LASSERT(olck->ols_has_ref);
1041 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
1042 LDLM_LOCK_RELEASE(dlmlock);
1043 olck->ols_has_ref = 0;
1049 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1050 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1051 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1052 * with some other lock some where in the cluster. This function does the
1055 * - invalidates all pages protected by this lock (after sending dirty
1056 * ones to the server, as necessary);
1058 * - decref's underlying ldlm lock;
1060 * - cancels ldlm lock (ldlm_cli_cancel()).
1062 void osc_lock_cancel(const struct lu_env *env,
1063 const struct cl_lock_slice *slice)
1065 struct osc_object *obj = cl2osc(slice->cls_obj);
1066 struct osc_lock *oscl = cl2osc_lock(slice);
1070 LINVRNT(osc_lock_invariant(oscl));
1072 osc_lock_detach(env, oscl);
1073 oscl->ols_state = OLS_CANCELLED;
1074 oscl->ols_flags &= ~LDLM_FL_LVB_READY;
1076 osc_lock_wake_waiters(env, obj, oscl);
1079 EXPORT_SYMBOL(osc_lock_cancel);
1081 int osc_lock_print(const struct lu_env *env, void *cookie,
1082 lu_printer_t p, const struct cl_lock_slice *slice)
1084 struct osc_lock *lock = cl2osc_lock(slice);
1086 (*p)(env, cookie, "%p %#llx %#llx %d %p ",
1087 lock->ols_dlmlock, lock->ols_flags, lock->ols_handle.cookie,
1088 lock->ols_state, lock->ols_owner);
1089 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1092 EXPORT_SYMBOL(osc_lock_print);
1094 static const struct cl_lock_operations osc_lock_ops = {
1095 .clo_fini = osc_lock_fini,
1096 .clo_enqueue = osc_lock_enqueue,
1097 .clo_cancel = osc_lock_cancel,
1098 .clo_print = osc_lock_print,
1101 static void osc_lock_lockless_cancel(const struct lu_env *env,
1102 const struct cl_lock_slice *slice)
1104 struct osc_lock *ols = cl2osc_lock(slice);
1105 struct osc_object *osc = cl2osc(slice->cls_obj);
1106 struct cl_lock_descr *descr = &slice->cls_lock->cll_descr;
1109 LASSERT(ols->ols_dlmlock == NULL);
1110 result = osc_lock_flush(osc, descr->cld_start, descr->cld_end,
1111 descr->cld_mode, false);
1113 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1116 osc_lock_wake_waiters(env, osc, ols);
1119 static const struct cl_lock_operations osc_lock_lockless_ops = {
1120 .clo_fini = osc_lock_fini,
1121 .clo_enqueue = osc_lock_enqueue,
1122 .clo_cancel = osc_lock_lockless_cancel,
1123 .clo_print = osc_lock_print
1126 void osc_lock_set_writer(const struct lu_env *env, const struct cl_io *io,
1127 struct cl_object *obj, struct osc_lock *oscl)
1129 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
1133 if (!cl_object_same(io->ci_obj, obj))
1136 if (likely(io->ci_type == CIT_WRITE)) {
1137 io_start = cl_index(obj, io->u.ci_rw.rw_range.cir_pos);
1138 io_end = cl_index(obj, io->u.ci_rw.rw_range.cir_pos +
1139 io->u.ci_rw.rw_range.cir_count - 1);
1141 LASSERT(cl_io_is_mkwrite(io));
1142 io_start = io_end = io->u.ci_fault.ft_index;
1145 if (descr->cld_mode >= CLM_WRITE &&
1146 (cl_io_is_append(io) ||
1147 (descr->cld_start <= io_start && descr->cld_end >= io_end))) {
1148 struct osc_io *oio = osc_env_io(env);
1150 /* There must be only one lock to match the write region */
1151 LASSERT(oio->oi_write_osclock == NULL);
1152 oio->oi_write_osclock = oscl;
1155 EXPORT_SYMBOL(osc_lock_set_writer);
1157 int osc_lock_init(const struct lu_env *env,
1158 struct cl_object *obj, struct cl_lock *lock,
1159 const struct cl_io *io)
1161 struct osc_lock *oscl;
1162 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1164 OBD_SLAB_ALLOC_PTR_GFP(oscl, osc_lock_kmem, GFP_NOFS);
1168 oscl->ols_state = OLS_NEW;
1169 spin_lock_init(&oscl->ols_lock);
1170 INIT_LIST_HEAD(&oscl->ols_waiting_list);
1171 INIT_LIST_HEAD(&oscl->ols_wait_entry);
1172 INIT_LIST_HEAD(&oscl->ols_nextlock_oscobj);
1173 oscl->ols_lockless_ops = &osc_lock_lockless_ops;
1175 /* Speculative lock requests must be either no_expand or glimpse
1176 * request (CEF_GLIMPSE). non-glimpse no_expand speculative extent
1177 * locks will break ofd_intent_cb. (see comment there)*/
1178 LASSERT(ergo((enqflags & CEF_SPECULATIVE) != 0,
1179 (enqflags & (CEF_LOCK_NO_EXPAND | CEF_GLIMPSE)) != 0));
1181 oscl->ols_flags = osc_enq2ldlm_flags(enqflags);
1182 oscl->ols_speculative = !!(enqflags & CEF_SPECULATIVE);
1184 if (oscl->ols_flags & LDLM_FL_HAS_INTENT) {
1185 oscl->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1186 oscl->ols_glimpse = 1;
1188 if (io->ci_ndelay && cl_object_same(io->ci_obj, obj))
1189 oscl->ols_flags |= LDLM_FL_NDELAY;
1190 osc_lock_build_einfo(env, lock, cl2osc(obj), &oscl->ols_einfo);
1192 cl_lock_slice_add(lock, &oscl->ols_cl, obj, &osc_lock_ops);
1194 if (!(enqflags & CEF_MUST))
1195 /* try to convert this lock to a lockless lock */
1196 osc_lock_to_lockless(env, oscl, (enqflags & CEF_NEVER));
1197 if (oscl->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
1198 oscl->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1200 if (io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io))
1201 osc_lock_set_writer(env, io, obj, oscl);
1203 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %#llx",
1204 lock, oscl, oscl->ols_flags);
1210 * Finds an existing lock covering given index and optionally different from a
1211 * given \a except lock.
1213 struct ldlm_lock *osc_obj_dlmlock_at_pgoff(const struct lu_env *env,
1214 struct osc_object *obj,
1216 enum osc_dap_flags dap_flags)
1218 struct osc_thread_info *info = osc_env_info(env);
1219 struct ldlm_res_id *resname = &info->oti_resname;
1220 union ldlm_policy_data *policy = &info->oti_policy;
1221 struct lustre_handle lockh;
1222 struct ldlm_lock *lock = NULL;
1223 enum ldlm_mode mode;
1228 ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1229 osc_index2policy(policy, osc2cl(obj), index, index);
1230 policy->l_extent.gid = LDLM_GID_ANY;
1232 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1233 if (dap_flags & OSC_DAP_FL_TEST_LOCK)
1234 flags |= LDLM_FL_TEST_LOCK;
1236 * It is fine to match any group lock since there could be only one
1237 * with a uniq gid and it conflicts with all other lock modes too
1240 mode = osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
1241 LCK_PR | LCK_PW | LCK_GROUP, &flags, obj, &lockh,
1242 dap_flags & OSC_DAP_FL_CANCELING);
1244 lock = ldlm_handle2lock(&lockh);
1245 /* RACE: the lock is cancelled so let's try again */
1246 if (unlikely(lock == NULL))