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) || PageDirty(page->cp_vmpage) ||
615 PageWriteback(page->cp_vmpage))
616 return CLP_GANG_ABORT;
618 *(pgoff_t *)cbdata = osc_index(ops) + 1;
619 return CLP_GANG_OKAY;
622 static unsigned long osc_lock_weight(const struct lu_env *env,
623 struct osc_object *oscobj,
624 loff_t start, loff_t end)
626 struct cl_io *io = osc_env_thread_io(env);
627 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, start);
641 result = osc_page_gang_lookup(env, io, oscobj,
642 page_index, cl_index(obj, end),
643 weigh_cb, (void *)&page_index);
644 if (result == CLP_GANG_ABORT)
646 if (result == CLP_GANG_RESCHED)
648 } while (result != CLP_GANG_OKAY);
651 return result == CLP_GANG_ABORT ? 1 : 0;
655 * Get the weight of dlm lock for early cancellation.
657 unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
660 struct osc_object *obj;
661 struct osc_lock *oscl;
662 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 ldlm_has_dom(dlmlock));
683 lock_res_and_lock(dlmlock);
684 obj = dlmlock->l_ast_data;
686 cl_object_get(osc2cl(obj));
687 unlock_res_and_lock(dlmlock);
690 GOTO(out, weight = 1);
692 spin_lock(&obj->oo_ol_spin);
693 list_for_each_entry(oscl, &obj->oo_ol_list, ols_nextlock_oscobj) {
694 if (oscl->ols_dlmlock != NULL && oscl->ols_dlmlock != dlmlock)
698 spin_unlock(&obj->oo_ol_spin);
701 * If the lock is being used by an IO, definitely not cancel it.
703 GOTO(out, weight = 1);
706 if (ldlm_has_dom(dlmlock))
707 weight = osc_lock_weight(env, obj, 0, OBD_OBJECT_EOF);
709 weight = osc_lock_weight(env, obj,
710 dlmlock->l_policy_data.l_extent.start,
711 dlmlock->l_policy_data.l_extent.end);
717 cl_object_put(env, osc2cl(obj));
719 cl_env_put(env, &refcheck);
722 EXPORT_SYMBOL(osc_ldlm_weigh_ast);
724 static void osc_lock_build_einfo(const struct lu_env *env,
725 const struct cl_lock *lock,
726 struct osc_object *osc,
727 struct ldlm_enqueue_info *einfo)
729 einfo->ei_type = LDLM_EXTENT;
730 einfo->ei_mode = osc_cl_lock2ldlm(lock->cll_descr.cld_mode);
731 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
732 einfo->ei_cb_cp = ldlm_completion_ast;
733 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
734 einfo->ei_cbdata = osc; /* value to be put into ->l_ast_data */
738 * Determine if the lock should be converted into a lockless lock.
741 * - if the lock has an explicite requirment for a non-lockless lock;
742 * - if the io lock request type ci_lockreq;
743 * - send the enqueue rpc to ost to make the further decision;
744 * - special treat to truncate lockless lock
746 * Additional policy can be implemented here, e.g., never do lockless-io
749 void osc_lock_to_lockless(const struct lu_env *env,
750 struct osc_lock *ols, int force)
752 struct cl_lock_slice *slice = &ols->ols_cl;
753 struct osc_io *oio = osc_env_io(env);
754 struct cl_io *io = oio->oi_cl.cis_io;
755 struct cl_object *obj = slice->cls_obj;
756 struct osc_object *oob = cl2osc(obj);
757 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
758 struct obd_connect_data *ocd;
760 LASSERT(ols->ols_state == OLS_NEW ||
761 ols->ols_state == OLS_UPCALL_RECEIVED);
764 ols->ols_locklessable = 1;
765 slice->cls_ops = ols->ols_lockless_ops;
767 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
768 io->ci_lockreq == CILR_MAYBE ||
769 io->ci_lockreq == CILR_NEVER);
771 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
772 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
773 (io->ci_lockreq == CILR_MAYBE) &&
774 (ocd->ocd_connect_flags &
775 OBD_CONNECT_SRVLOCK);
776 if (io->ci_lockreq == CILR_NEVER ||
778 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
779 /* lockless truncate */
780 (cl_io_is_trunc(io) && osd->od_lockless_truncate &&
781 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK))) {
782 ols->ols_locklessable = 1;
783 slice->cls_ops = ols->ols_lockless_ops;
786 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
788 EXPORT_SYMBOL(osc_lock_to_lockless);
790 static bool osc_lock_compatible(const struct osc_lock *qing,
791 const struct osc_lock *qed)
793 struct cl_lock_descr *qed_descr = &qed->ols_cl.cls_lock->cll_descr;
794 struct cl_lock_descr *qing_descr = &qing->ols_cl.cls_lock->cll_descr;
796 if (qed->ols_glimpse || qed->ols_speculative)
799 if (qing_descr->cld_mode == CLM_READ && qed_descr->cld_mode == CLM_READ)
802 if (qed->ols_state < OLS_GRANTED)
805 if (qed_descr->cld_mode >= qing_descr->cld_mode &&
806 qed_descr->cld_start <= qing_descr->cld_start &&
807 qed_descr->cld_end >= qing_descr->cld_end)
813 void osc_lock_wake_waiters(const struct lu_env *env, struct osc_object *osc,
814 struct osc_lock *oscl)
816 spin_lock(&osc->oo_ol_spin);
817 list_del_init(&oscl->ols_nextlock_oscobj);
818 spin_unlock(&osc->oo_ol_spin);
820 spin_lock(&oscl->ols_lock);
821 while (!list_empty(&oscl->ols_waiting_list)) {
822 struct osc_lock *scan;
824 scan = list_entry(oscl->ols_waiting_list.next, struct osc_lock,
826 list_del_init(&scan->ols_wait_entry);
828 cl_sync_io_note(env, scan->ols_owner, 0);
830 spin_unlock(&oscl->ols_lock);
832 EXPORT_SYMBOL(osc_lock_wake_waiters);
834 int osc_lock_enqueue_wait(const struct lu_env *env, struct osc_object *obj,
835 struct osc_lock *oscl)
837 struct osc_lock *tmp_oscl;
838 struct cl_lock_descr *need = &oscl->ols_cl.cls_lock->cll_descr;
839 struct cl_sync_io *waiter = &osc_env_info(env)->oti_anchor;
844 spin_lock(&obj->oo_ol_spin);
845 list_add_tail(&oscl->ols_nextlock_oscobj, &obj->oo_ol_list);
848 list_for_each_entry(tmp_oscl, &obj->oo_ol_list,
849 ols_nextlock_oscobj) {
850 struct cl_lock_descr *descr;
852 if (tmp_oscl == oscl)
855 descr = &tmp_oscl->ols_cl.cls_lock->cll_descr;
856 if (descr->cld_start > need->cld_end ||
857 descr->cld_end < need->cld_start)
860 /* We're not supposed to give up group lock */
861 if (descr->cld_mode == CLM_GROUP)
864 if (!osc_lock_is_lockless(oscl) &&
865 osc_lock_compatible(oscl, tmp_oscl))
868 /* wait for conflicting lock to be canceled */
869 cl_sync_io_init(waiter, 1, cl_sync_io_end);
870 oscl->ols_owner = waiter;
872 spin_lock(&tmp_oscl->ols_lock);
873 /* add oscl into tmp's ols_waiting list */
874 list_add_tail(&oscl->ols_wait_entry,
875 &tmp_oscl->ols_waiting_list);
876 spin_unlock(&tmp_oscl->ols_lock);
878 spin_unlock(&obj->oo_ol_spin);
879 rc = cl_sync_io_wait(env, waiter, 0);
880 spin_lock(&obj->oo_ol_spin);
885 oscl->ols_owner = NULL;
888 spin_unlock(&obj->oo_ol_spin);
892 EXPORT_SYMBOL(osc_lock_enqueue_wait);
895 * Implementation of cl_lock_operations::clo_enqueue() method for osc
896 * layer. This initiates ldlm enqueue:
898 * - cancels conflicting locks early (osc_lock_enqueue_wait());
900 * - calls osc_enqueue_base() to do actual enqueue.
902 * osc_enqueue_base() is supplied with an upcall function that is executed
903 * when lock is received either after a local cached ldlm lock is matched, or
904 * when a reply from the server is received.
906 * This function does not wait for the network communication to complete.
908 static int osc_lock_enqueue(const struct lu_env *env,
909 const struct cl_lock_slice *slice,
910 struct cl_io *unused, struct cl_sync_io *anchor)
912 struct osc_thread_info *info = osc_env_info(env);
913 struct osc_io *oio = osc_env_io(env);
914 struct osc_object *osc = cl2osc(slice->cls_obj);
915 struct osc_lock *oscl = cl2osc_lock(slice);
916 struct obd_export *exp = osc_export(osc);
917 struct cl_lock *lock = slice->cls_lock;
918 struct ldlm_res_id *resname = &info->oti_resname;
919 union ldlm_policy_data *policy = &info->oti_policy;
920 osc_enqueue_upcall_f upcall = osc_lock_upcall;
927 LASSERTF(ergo(oscl->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
928 "lock = %p, ols = %p\n", lock, oscl);
930 if (oscl->ols_state == OLS_GRANTED)
933 if ((oscl->ols_flags & LDLM_FL_NO_EXPANSION) &&
934 !(exp_connect_lockahead_old(exp) || exp_connect_lockahead(exp))) {
935 result = -EOPNOTSUPP;
936 CERROR("%s: server does not support lockahead/locknoexpand:"
937 "rc = %d\n", exp->exp_obd->obd_name, result);
941 if (oscl->ols_flags & LDLM_FL_TEST_LOCK)
942 GOTO(enqueue_base, 0);
944 /* For glimpse and/or speculative locks, do not wait for reply from
945 * server on LDLM request */
946 if (oscl->ols_glimpse || oscl->ols_speculative) {
947 /* Speculative and glimpse locks do not have an anchor */
948 LASSERT(equi(oscl->ols_speculative, anchor == NULL));
950 GOTO(enqueue_base, 0);
953 result = osc_lock_enqueue_wait(env, osc, oscl);
957 /* we can grant lockless lock right after all conflicting locks
959 if (osc_lock_is_lockless(oscl)) {
960 oscl->ols_state = OLS_GRANTED;
961 oio->oi_lockless = 1;
966 oscl->ols_state = OLS_ENQUEUED;
967 if (anchor != NULL) {
968 atomic_inc(&anchor->csi_sync_nr);
969 oscl->ols_owner = anchor;
973 * DLM lock's ast data must be osc_object;
974 * if glimpse or speculative lock, async of osc_enqueue_base()
977 * For non-speculative locks:
978 * DLM's enqueue callback set to osc_lock_upcall() with cookie as
980 * For speculative locks:
981 * osc_lock_upcall_speculative & cookie is the osc object, since
982 * there is no osc_lock
984 ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
985 osc_lock_build_policy(env, lock, policy);
986 if (oscl->ols_speculative) {
987 oscl->ols_einfo.ei_cbdata = NULL;
988 /* hold a reference for callback */
989 cl_object_get(osc2cl(osc));
990 upcall = osc_lock_upcall_speculative;
993 result = osc_enqueue_base(exp, resname, &oscl->ols_flags,
994 policy, &oscl->ols_lvb,
995 osc->oo_oinfo->loi_kms_valid,
997 &oscl->ols_einfo, PTLRPCD_SET, async,
998 oscl->ols_speculative);
1000 if (osc_lock_is_lockless(oscl)) {
1001 oio->oi_lockless = 1;
1002 } else if (!async) {
1003 LASSERT(oscl->ols_state == OLS_GRANTED);
1004 LASSERT(oscl->ols_hold);
1005 LASSERT(oscl->ols_dlmlock != NULL);
1007 } else if (oscl->ols_speculative) {
1008 cl_object_put(env, osc2cl(osc));
1009 if (oscl->ols_glimpse) {
1010 /* hide error for AGL request */
1017 oscl->ols_state = OLS_CANCELLED;
1018 osc_lock_wake_waiters(env, osc, oscl);
1021 cl_sync_io_note(env, anchor, result);
1027 * Breaks a link between osc_lock and dlm_lock.
1029 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
1031 struct ldlm_lock *dlmlock;
1035 dlmlock = olck->ols_dlmlock;
1036 if (dlmlock == NULL)
1039 if (olck->ols_hold) {
1041 ldlm_lock_decref(&olck->ols_handle, olck->ols_einfo.ei_mode);
1042 olck->ols_handle.cookie = 0ULL;
1045 olck->ols_dlmlock = NULL;
1047 /* release a reference taken in osc_lock_upcall(). */
1048 LASSERT(olck->ols_has_ref);
1049 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
1050 LDLM_LOCK_RELEASE(dlmlock);
1051 olck->ols_has_ref = 0;
1057 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1058 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1059 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1060 * with some other lock some where in the cluster. This function does the
1063 * - invalidates all pages protected by this lock (after sending dirty
1064 * ones to the server, as necessary);
1066 * - decref's underlying ldlm lock;
1068 * - cancels ldlm lock (ldlm_cli_cancel()).
1070 void osc_lock_cancel(const struct lu_env *env,
1071 const struct cl_lock_slice *slice)
1073 struct osc_object *obj = cl2osc(slice->cls_obj);
1074 struct osc_lock *oscl = cl2osc_lock(slice);
1078 LINVRNT(osc_lock_invariant(oscl));
1080 osc_lock_detach(env, oscl);
1081 oscl->ols_state = OLS_CANCELLED;
1082 oscl->ols_flags &= ~LDLM_FL_LVB_READY;
1084 osc_lock_wake_waiters(env, obj, oscl);
1087 EXPORT_SYMBOL(osc_lock_cancel);
1089 int osc_lock_print(const struct lu_env *env, void *cookie,
1090 lu_printer_t p, const struct cl_lock_slice *slice)
1092 struct osc_lock *lock = cl2osc_lock(slice);
1094 (*p)(env, cookie, "%p %#llx %#llx %d %p ",
1095 lock->ols_dlmlock, lock->ols_flags, lock->ols_handle.cookie,
1096 lock->ols_state, lock->ols_owner);
1097 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1100 EXPORT_SYMBOL(osc_lock_print);
1102 static const struct cl_lock_operations osc_lock_ops = {
1103 .clo_fini = osc_lock_fini,
1104 .clo_enqueue = osc_lock_enqueue,
1105 .clo_cancel = osc_lock_cancel,
1106 .clo_print = osc_lock_print,
1109 static void osc_lock_lockless_cancel(const struct lu_env *env,
1110 const struct cl_lock_slice *slice)
1112 struct osc_lock *ols = cl2osc_lock(slice);
1113 struct osc_object *osc = cl2osc(slice->cls_obj);
1114 struct cl_lock_descr *descr = &slice->cls_lock->cll_descr;
1117 LASSERT(ols->ols_dlmlock == NULL);
1118 result = osc_lock_flush(osc, descr->cld_start, descr->cld_end,
1119 descr->cld_mode, false);
1121 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1124 osc_lock_wake_waiters(env, osc, ols);
1127 static const struct cl_lock_operations osc_lock_lockless_ops = {
1128 .clo_fini = osc_lock_fini,
1129 .clo_enqueue = osc_lock_enqueue,
1130 .clo_cancel = osc_lock_lockless_cancel,
1131 .clo_print = osc_lock_print
1134 void osc_lock_set_writer(const struct lu_env *env, const struct cl_io *io,
1135 struct cl_object *obj, struct osc_lock *oscl)
1137 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
1141 if (!cl_object_same(io->ci_obj, obj))
1144 if (likely(io->ci_type == CIT_WRITE)) {
1145 io_start = cl_index(obj, io->u.ci_rw.rw_range.cir_pos);
1146 io_end = cl_index(obj, io->u.ci_rw.rw_range.cir_pos +
1147 io->u.ci_rw.rw_range.cir_count - 1);
1149 LASSERT(cl_io_is_mkwrite(io));
1150 io_start = io_end = io->u.ci_fault.ft_index;
1153 if (descr->cld_mode >= CLM_WRITE &&
1154 (cl_io_is_append(io) ||
1155 (descr->cld_start <= io_start && descr->cld_end >= io_end))) {
1156 struct osc_io *oio = osc_env_io(env);
1158 /* There must be only one lock to match the write region */
1159 LASSERT(oio->oi_write_osclock == NULL);
1160 oio->oi_write_osclock = oscl;
1163 EXPORT_SYMBOL(osc_lock_set_writer);
1165 int osc_lock_init(const struct lu_env *env,
1166 struct cl_object *obj, struct cl_lock *lock,
1167 const struct cl_io *io)
1169 struct osc_lock *oscl;
1170 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1172 OBD_SLAB_ALLOC_PTR_GFP(oscl, osc_lock_kmem, GFP_NOFS);
1176 oscl->ols_state = OLS_NEW;
1177 spin_lock_init(&oscl->ols_lock);
1178 INIT_LIST_HEAD(&oscl->ols_waiting_list);
1179 INIT_LIST_HEAD(&oscl->ols_wait_entry);
1180 INIT_LIST_HEAD(&oscl->ols_nextlock_oscobj);
1181 oscl->ols_lockless_ops = &osc_lock_lockless_ops;
1183 /* Speculative lock requests must be either no_expand or glimpse
1184 * request (CEF_GLIMPSE). non-glimpse no_expand speculative extent
1185 * locks will break ofd_intent_cb. (see comment there)*/
1186 LASSERT(ergo((enqflags & CEF_SPECULATIVE) != 0,
1187 (enqflags & (CEF_LOCK_NO_EXPAND | CEF_GLIMPSE)) != 0));
1189 oscl->ols_flags = osc_enq2ldlm_flags(enqflags);
1190 oscl->ols_speculative = !!(enqflags & CEF_SPECULATIVE);
1192 if (oscl->ols_flags & LDLM_FL_HAS_INTENT) {
1193 oscl->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1194 oscl->ols_glimpse = 1;
1196 if (io->ci_ndelay && cl_object_same(io->ci_obj, obj))
1197 oscl->ols_flags |= LDLM_FL_NDELAY;
1198 osc_lock_build_einfo(env, lock, cl2osc(obj), &oscl->ols_einfo);
1200 cl_lock_slice_add(lock, &oscl->ols_cl, obj, &osc_lock_ops);
1202 if (!(enqflags & CEF_MUST))
1203 /* try to convert this lock to a lockless lock */
1204 osc_lock_to_lockless(env, oscl, (enqflags & CEF_NEVER));
1205 if (oscl->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
1206 oscl->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1208 if (io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io))
1209 osc_lock_set_writer(env, io, obj, oscl);
1211 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %#llx",
1212 lock, oscl, oscl->ols_flags);
1218 * Finds an existing lock covering given index and optionally different from a
1219 * given \a except lock.
1221 struct ldlm_lock *osc_obj_dlmlock_at_pgoff(const struct lu_env *env,
1222 struct osc_object *obj,
1224 enum osc_dap_flags dap_flags)
1226 struct osc_thread_info *info = osc_env_info(env);
1227 struct ldlm_res_id *resname = &info->oti_resname;
1228 union ldlm_policy_data *policy = &info->oti_policy;
1229 struct lustre_handle lockh;
1230 struct ldlm_lock *lock = NULL;
1231 enum ldlm_mode mode;
1236 ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1237 osc_index2policy(policy, osc2cl(obj), index, index);
1238 policy->l_extent.gid = LDLM_GID_ANY;
1240 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1241 if (dap_flags & OSC_DAP_FL_TEST_LOCK)
1242 flags |= LDLM_FL_TEST_LOCK;
1244 * It is fine to match any group lock since there could be only one
1245 * with a uniq gid and it conflicts with all other lock modes too
1248 mode = osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
1249 LCK_PR | LCK_PW | LCK_GROUP, &flags, obj, &lockh,
1250 dap_flags & OSC_DAP_FL_CANCELING);
1252 lock = ldlm_handle2lock(&lockh);
1253 /* RACE: the lock is cancelled so let's try again */
1254 if (unlikely(lock == NULL))