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, 2016, 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 #include <libcfs/libcfs.h>
41 /* fid_build_reg_res_name() */
42 #include <lustre_fid.h>
43 #include <lustre_osc.h>
45 #include "osc_internal.h"
52 * Returns a weak pointer to the ldlm lock identified by a handle. Returned
53 * pointer cannot be dereferenced, as lock is not protected from concurrent
54 * reclaim. This function is a helper for osc_lock_invariant().
56 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
58 struct ldlm_lock *lock;
60 lock = ldlm_handle2lock(handle);
67 * Invariant that has to be true all of the time.
69 static int osc_lock_invariant(struct osc_lock *ols)
71 struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle);
72 struct ldlm_lock *olock = ols->ols_dlmlock;
73 int handle_used = lustre_handle_is_used(&ols->ols_handle);
75 if (ergo(osc_lock_is_lockless(ols),
76 ols->ols_locklessable && ols->ols_dlmlock == NULL))
80 * If all the following "ergo"s are true, return 1, otherwise 0
82 if (! ergo(olock != NULL, handle_used))
85 if (! ergo(olock != NULL,
86 olock->l_handle.h_cookie == ols->ols_handle.cookie))
89 if (! ergo(handle_used,
90 ergo(lock != NULL && olock != NULL, lock == olock) &&
91 ergo(lock == NULL, olock == NULL)))
94 * Check that ->ols_handle and ->ols_dlmlock are consistent, but
95 * take into account that they are set at the different time.
97 if (! ergo(ols->ols_state == OLS_CANCELLED,
98 olock == NULL && !handle_used))
101 * DLM lock is destroyed only after we have seen cancellation
104 if (! ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
105 !ldlm_is_destroyed(olock)))
108 if (! ergo(ols->ols_state == OLS_GRANTED,
110 olock->l_req_mode == olock->l_granted_mode &&
116 /*****************************************************************************
122 void osc_lock_fini(const struct lu_env *env, struct cl_lock_slice *slice)
124 struct osc_lock *ols = cl2osc_lock(slice);
126 LINVRNT(osc_lock_invariant(ols));
127 LASSERT(ols->ols_dlmlock == NULL);
129 OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
131 EXPORT_SYMBOL(osc_lock_fini);
133 static void osc_lock_build_policy(const struct lu_env *env,
134 const struct cl_lock *lock,
135 union ldlm_policy_data *policy)
137 const struct cl_lock_descr *d = &lock->cll_descr;
139 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
140 policy->l_extent.gid = d->cld_gid;
144 * Updates object attributes from a lock value block (lvb) received together
145 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
148 * This can be optimized to not update attributes when lock is a result of a
151 * Called under lock and resource spin-locks.
153 static void osc_lock_lvb_update(const struct lu_env *env,
154 struct osc_object *osc,
155 struct ldlm_lock *dlmlock,
158 struct cl_object *obj = osc2cl(osc);
159 struct lov_oinfo *oinfo = osc->oo_oinfo;
160 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
165 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
167 LASSERT(dlmlock != NULL);
168 lvb = dlmlock->l_lvb_data;
170 cl_lvb2attr(attr, lvb);
172 cl_object_attr_lock(obj);
173 if (dlmlock != NULL) {
176 check_res_locked(dlmlock->l_resource);
178 LASSERT(lvb == dlmlock->l_lvb_data);
179 size = lvb->lvb_size;
181 /* Extend KMS up to the end of this lock and no further
182 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
183 if (size > dlmlock->l_policy_data.l_extent.end)
184 size = dlmlock->l_policy_data.l_extent.end + 1;
185 if (size >= oinfo->loi_kms) {
186 LDLM_DEBUG(dlmlock, "lock acquired, setting rss=%llu"
187 ", kms=%llu", lvb->lvb_size, size);
189 attr->cat_kms = size;
191 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
192 "%llu; leaving kms=%llu, end=%llu",
193 lvb->lvb_size, oinfo->loi_kms,
194 dlmlock->l_policy_data.l_extent.end);
196 ldlm_lock_allow_match_locked(dlmlock);
199 cl_object_attr_update(env, obj, attr, valid);
200 cl_object_attr_unlock(obj);
205 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *oscl,
206 struct lustre_handle *lockh, bool lvb_update)
208 struct ldlm_lock *dlmlock;
210 dlmlock = ldlm_handle2lock_long(lockh, 0);
211 LASSERT(dlmlock != NULL);
213 /* lock reference taken by ldlm_handle2lock_long() is
214 * owned by osc_lock and released in osc_lock_detach()
216 lu_ref_add(&dlmlock->l_reference, "osc_lock", oscl);
217 oscl->ols_has_ref = 1;
219 LASSERT(oscl->ols_dlmlock == NULL);
220 oscl->ols_dlmlock = dlmlock;
222 /* This may be a matched lock for glimpse request, do not hold
223 * lock reference in that case. */
224 if (!oscl->ols_glimpse) {
225 /* hold a refc for non glimpse lock which will
226 * be released in osc_lock_cancel() */
227 lustre_handle_copy(&oscl->ols_handle, lockh);
228 ldlm_lock_addref(lockh, oscl->ols_einfo.ei_mode);
232 /* Lock must have been granted. */
233 lock_res_and_lock(dlmlock);
234 if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
235 struct ldlm_extent *ext = &dlmlock->l_policy_data.l_extent;
236 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
238 /* extend the lock extent, otherwise it will have problem when
239 * we decide whether to grant a lockless lock. */
240 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
241 descr->cld_start = cl_index(descr->cld_obj, ext->start);
242 descr->cld_end = cl_index(descr->cld_obj, ext->end);
243 descr->cld_gid = ext->gid;
245 /* no lvb update for matched lock */
247 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
248 osc_lock_lvb_update(env, cl2osc(oscl->ols_cl.cls_obj),
251 LINVRNT(osc_lock_invariant(oscl));
253 unlock_res_and_lock(dlmlock);
255 LASSERT(oscl->ols_state != OLS_GRANTED);
256 oscl->ols_state = OLS_GRANTED;
260 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
261 * received from a server, or after osc_enqueue_base() matched a local DLM
264 static int osc_lock_upcall(void *cookie, struct lustre_handle *lockh,
267 struct osc_lock *oscl = cookie;
268 struct cl_lock_slice *slice = &oscl->ols_cl;
274 env = cl_env_percpu_get();
275 /* should never happen, similar to osc_ldlm_blocking_ast(). */
276 LASSERT(!IS_ERR(env));
278 rc = ldlm_error2errno(errcode);
279 if (oscl->ols_state == OLS_ENQUEUED) {
280 oscl->ols_state = OLS_UPCALL_RECEIVED;
281 } else if (oscl->ols_state == OLS_CANCELLED) {
284 CERROR("Impossible state: %d\n", oscl->ols_state);
289 osc_lock_granted(env, oscl, lockh, errcode == ELDLM_OK);
291 /* Error handling, some errors are tolerable. */
292 if (oscl->ols_locklessable && rc == -EUSERS) {
293 /* This is a tolerable error, turn this lock into
296 osc_object_set_contended(cl2osc(slice->cls_obj));
297 LASSERT(slice->cls_ops != oscl->ols_lockless_ops);
299 /* Change this lock to ldlmlock-less lock. */
300 osc_lock_to_lockless(env, oscl, 1);
301 oscl->ols_state = OLS_GRANTED;
303 } else if (oscl->ols_glimpse && rc == -ENAVAIL) {
304 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
305 osc_lock_lvb_update(env, cl2osc(slice->cls_obj),
306 NULL, &oscl->ols_lvb);
307 /* Hide the error. */
311 if (oscl->ols_owner != NULL)
312 cl_sync_io_note(env, oscl->ols_owner, rc);
313 cl_env_percpu_put(env);
318 static int osc_lock_upcall_speculative(void *cookie,
319 struct lustre_handle *lockh,
322 struct osc_object *osc = cookie;
323 struct ldlm_lock *dlmlock;
328 env = cl_env_get(&refcheck);
329 LASSERT(!IS_ERR(env));
331 if (errcode == ELDLM_LOCK_MATCHED)
332 GOTO(out, errcode = ELDLM_OK);
334 if (errcode != ELDLM_OK)
337 dlmlock = ldlm_handle2lock(lockh);
338 LASSERT(dlmlock != NULL);
340 lock_res_and_lock(dlmlock);
341 LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
343 /* there is no osc_lock associated with speculative locks */
344 osc_lock_lvb_update(env, osc, dlmlock, NULL);
346 unlock_res_and_lock(dlmlock);
347 LDLM_LOCK_PUT(dlmlock);
350 cl_object_put(env, osc2cl(osc));
351 cl_env_put(env, &refcheck);
352 RETURN(ldlm_error2errno(errcode));
355 static int osc_lock_flush(struct osc_object *obj, pgoff_t start, pgoff_t end,
356 enum cl_lock_mode mode, bool discard)
365 env = cl_env_get(&refcheck);
367 RETURN(PTR_ERR(env));
369 if (mode == CLM_WRITE) {
370 rc = osc_cache_writeback_range(env, obj, start, end, 1,
372 CDEBUG(D_CACHE, "object %p: [%lu -> %lu] %d pages were %s.\n",
374 discard ? "discarded" : "written back");
379 rc2 = osc_lock_discard_pages(env, obj, start, end, discard);
380 if (rc == 0 && rc2 < 0)
383 cl_env_put(env, &refcheck);
388 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
389 * and ldlm_lock caches.
391 static int osc_dlm_blocking_ast0(const struct lu_env *env,
392 struct ldlm_lock *dlmlock,
393 void *data, int flag)
395 struct cl_object *obj = NULL;
398 enum cl_lock_mode mode = CLM_READ;
401 LASSERT(flag == LDLM_CB_CANCELING);
403 lock_res_and_lock(dlmlock);
404 if (dlmlock->l_granted_mode != dlmlock->l_req_mode) {
405 dlmlock->l_ast_data = NULL;
406 unlock_res_and_lock(dlmlock);
410 discard = ldlm_is_discard_data(dlmlock);
411 if (dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))
414 if (dlmlock->l_ast_data != NULL) {
415 obj = osc2cl(dlmlock->l_ast_data);
416 dlmlock->l_ast_data = NULL;
421 unlock_res_and_lock(dlmlock);
423 /* if l_ast_data is NULL, the dlmlock was enqueued by AGL or
424 * the object has been destroyed. */
426 struct ldlm_extent *extent = &dlmlock->l_policy_data.l_extent;
427 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
430 /* Destroy pages covered by the extent of the DLM lock */
431 result = osc_lock_flush(cl2osc(obj),
432 cl_index(obj, extent->start),
433 cl_index(obj, extent->end),
436 /* losing a lock, update kms */
437 lock_res_and_lock(dlmlock);
438 cl_object_attr_lock(obj);
439 /* Must get the value under the lock to avoid race. */
440 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
441 /* Update the kms. Need to loop all granted locks.
442 * Not a problem for the client */
443 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
445 cl_object_attr_update(env, obj, attr, CAT_KMS);
446 cl_object_attr_unlock(obj);
447 unlock_res_and_lock(dlmlock);
449 cl_object_put(env, obj);
455 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
456 * some other lock, or is canceled. This function is installed as a
457 * ldlm_lock::l_blocking_ast() for client extent locks.
459 * Control flow is tricky, because ldlm uses the same call-back
460 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
462 * \param dlmlock lock for which ast occurred.
464 * \param new description of a conflicting lock in case of blocking ast.
466 * \param data value of dlmlock->l_ast_data
468 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
469 * cancellation and blocking ast's.
471 * Possible use cases:
473 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
474 * lock due to lock lru pressure, or explicit user request to purge
477 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
478 * us that dlmlock conflicts with another lock that some client is
479 * enqueuing. Lock is canceled.
481 * - cl_lock_cancel() is called. osc_lock_cancel() calls
482 * ldlm_cli_cancel() that calls
484 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
486 * recursively entering osc_ldlm_blocking_ast().
488 * - client cancels lock voluntary (e.g., as a part of early cancellation):
491 * osc_lock_cancel()->
492 * ldlm_cli_cancel()->
493 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
496 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
497 struct ldlm_lock_desc *new, void *data,
504 case LDLM_CB_BLOCKING: {
505 struct lustre_handle lockh;
507 ldlm_lock2handle(dlmlock, &lockh);
508 result = ldlm_cli_cancel(&lockh, LCF_ASYNC);
509 if (result == -ENODATA)
513 case LDLM_CB_CANCELING: {
518 * This can be called in the context of outer IO, e.g.,
520 * osc_enqueue_base()->...
521 * ->ldlm_prep_elc_req()->...
522 * ->ldlm_cancel_callback()->...
523 * ->osc_ldlm_blocking_ast()
525 * new environment has to be created to not corrupt outer
528 env = cl_env_get(&refcheck);
530 result = PTR_ERR(env);
534 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
535 cl_env_put(env, &refcheck);
544 int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
546 struct ptlrpc_request *req = data;
549 struct req_capsule *cap;
550 struct cl_object *obj = NULL;
556 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
558 env = cl_env_get(&refcheck);
560 GOTO(out, result = PTR_ERR(env));
563 lock_res_and_lock(dlmlock);
564 if (dlmlock->l_ast_data != NULL) {
565 obj = osc2cl(dlmlock->l_ast_data);
568 unlock_res_and_lock(dlmlock);
571 /* Do not grab the mutex of cl_lock for glimpse.
572 * See LU-1274 for details.
573 * BTW, it's okay for cl_lock to be cancelled during
574 * this period because server can handle this race.
575 * See ldlm_server_glimpse_ast() for details.
576 * cl_lock_mutex_get(env, lock); */
578 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
579 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
581 result = req_capsule_server_pack(cap);
583 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
584 result = cl_object_glimpse(env, obj, lvb);
586 if (!exp_connect_lvb_type(req->rq_export))
587 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
588 sizeof(struct ost_lvb_v1), RCL_SERVER);
589 cl_object_put(env, obj);
592 * These errors are normal races, so we don't want to
593 * fill the console with messages by calling
596 lustre_pack_reply(req, 1, NULL, NULL);
597 result = -ELDLM_NO_LOCK_DATA;
599 cl_env_put(env, &refcheck);
603 req->rq_status = result;
606 EXPORT_SYMBOL(osc_ldlm_glimpse_ast);
608 static int weigh_cb(const struct lu_env *env, struct cl_io *io,
609 struct osc_page *ops, void *cbdata)
611 struct cl_page *page = ops->ops_cl.cpl_page;
613 if (cl_page_is_vmlocked(env, page)
614 || PageDirty(page->cp_vmpage) || 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 struct ldlm_extent *extent)
626 struct cl_io *io = &osc_env_info(env)->oti_io;
627 struct cl_object *obj = cl_object_top(&oscobj->oo_cl);
633 io->ci_ignore_layout = 1;
634 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
638 page_index = cl_index(obj, extent->start);
640 result = osc_page_gang_lookup(env, io, oscobj,
642 cl_index(obj, extent->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;
669 * osc_ldlm_weigh_ast has a complex context since it might be called
670 * because of lock canceling, or from user's input. We have to make
671 * a new environment for it. Probably it is implementation safe to use
672 * the upper context because cl_lock_put don't modify environment
673 * variables. But just in case ..
675 env = cl_env_get(&refcheck);
677 /* Mostly because lack of memory, do not eliminate this lock */
680 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
681 lock_res_and_lock(dlmlock);
682 obj = dlmlock->l_ast_data;
684 cl_object_get(osc2cl(obj));
685 unlock_res_and_lock(dlmlock);
688 GOTO(out, weight = 1);
690 spin_lock(&obj->oo_ol_spin);
691 list_for_each_entry(oscl, &obj->oo_ol_list, ols_nextlock_oscobj) {
692 if (oscl->ols_dlmlock != NULL && oscl->ols_dlmlock != dlmlock)
696 spin_unlock(&obj->oo_ol_spin);
699 * If the lock is being used by an IO, definitely not cancel it.
701 GOTO(out, weight = 1);
704 weight = osc_lock_weight(env, obj, &dlmlock->l_policy_data.l_extent);
709 cl_object_put(env, osc2cl(obj));
711 cl_env_put(env, &refcheck);
715 static void osc_lock_build_einfo(const struct lu_env *env,
716 const struct cl_lock *lock,
717 struct osc_object *osc,
718 struct ldlm_enqueue_info *einfo)
720 einfo->ei_type = LDLM_EXTENT;
721 einfo->ei_mode = osc_cl_lock2ldlm(lock->cll_descr.cld_mode);
722 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
723 einfo->ei_cb_cp = ldlm_completion_ast;
724 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
725 einfo->ei_cbdata = osc; /* value to be put into ->l_ast_data */
729 * Determine if the lock should be converted into a lockless lock.
732 * - if the lock has an explicite requirment for a non-lockless lock;
733 * - if the io lock request type ci_lockreq;
734 * - send the enqueue rpc to ost to make the further decision;
735 * - special treat to truncate lockless lock
737 * Additional policy can be implemented here, e.g., never do lockless-io
740 void osc_lock_to_lockless(const struct lu_env *env,
741 struct osc_lock *ols, int force)
743 struct cl_lock_slice *slice = &ols->ols_cl;
744 struct osc_io *oio = osc_env_io(env);
745 struct cl_io *io = oio->oi_cl.cis_io;
746 struct cl_object *obj = slice->cls_obj;
747 struct osc_object *oob = cl2osc(obj);
748 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
749 struct obd_connect_data *ocd;
751 LASSERT(ols->ols_state == OLS_NEW ||
752 ols->ols_state == OLS_UPCALL_RECEIVED);
755 ols->ols_locklessable = 1;
756 slice->cls_ops = ols->ols_lockless_ops;
758 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
759 io->ci_lockreq == CILR_MAYBE ||
760 io->ci_lockreq == CILR_NEVER);
762 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
763 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
764 (io->ci_lockreq == CILR_MAYBE) &&
765 (ocd->ocd_connect_flags &
766 OBD_CONNECT_SRVLOCK);
767 if (io->ci_lockreq == CILR_NEVER ||
769 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
770 /* lockless truncate */
771 (cl_io_is_trunc(io) && osd->od_lockless_truncate &&
772 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK))) {
773 ols->ols_locklessable = 1;
774 slice->cls_ops = ols->ols_lockless_ops;
777 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
779 EXPORT_SYMBOL(osc_lock_to_lockless);
781 static bool osc_lock_compatible(const struct osc_lock *qing,
782 const struct osc_lock *qed)
784 struct cl_lock_descr *qed_descr = &qed->ols_cl.cls_lock->cll_descr;
785 struct cl_lock_descr *qing_descr = &qing->ols_cl.cls_lock->cll_descr;
787 if (qed->ols_glimpse || qed->ols_speculative)
790 if (qing_descr->cld_mode == CLM_READ && qed_descr->cld_mode == CLM_READ)
793 if (qed->ols_state < OLS_GRANTED)
796 if (qed_descr->cld_mode >= qing_descr->cld_mode &&
797 qed_descr->cld_start <= qing_descr->cld_start &&
798 qed_descr->cld_end >= qing_descr->cld_end)
804 void osc_lock_wake_waiters(const struct lu_env *env, struct osc_object *osc,
805 struct osc_lock *oscl)
807 spin_lock(&osc->oo_ol_spin);
808 list_del_init(&oscl->ols_nextlock_oscobj);
809 spin_unlock(&osc->oo_ol_spin);
811 spin_lock(&oscl->ols_lock);
812 while (!list_empty(&oscl->ols_waiting_list)) {
813 struct osc_lock *scan;
815 scan = list_entry(oscl->ols_waiting_list.next, struct osc_lock,
817 list_del_init(&scan->ols_wait_entry);
819 cl_sync_io_note(env, scan->ols_owner, 0);
821 spin_unlock(&oscl->ols_lock);
823 EXPORT_SYMBOL(osc_lock_wake_waiters);
825 int osc_lock_enqueue_wait(const struct lu_env *env, struct osc_object *obj,
826 struct osc_lock *oscl)
828 struct osc_lock *tmp_oscl;
829 struct cl_lock_descr *need = &oscl->ols_cl.cls_lock->cll_descr;
830 struct cl_sync_io *waiter = &osc_env_info(env)->oti_anchor;
835 spin_lock(&obj->oo_ol_spin);
836 list_add_tail(&oscl->ols_nextlock_oscobj, &obj->oo_ol_list);
839 list_for_each_entry(tmp_oscl, &obj->oo_ol_list,
840 ols_nextlock_oscobj) {
841 struct cl_lock_descr *descr;
843 if (tmp_oscl == oscl)
846 descr = &tmp_oscl->ols_cl.cls_lock->cll_descr;
847 if (descr->cld_start > need->cld_end ||
848 descr->cld_end < need->cld_start)
851 /* We're not supposed to give up group lock */
852 if (descr->cld_mode == CLM_GROUP)
855 if (!osc_lock_is_lockless(oscl) &&
856 osc_lock_compatible(oscl, tmp_oscl))
859 /* wait for conflicting lock to be canceled */
860 cl_sync_io_init(waiter, 1, cl_sync_io_end);
861 oscl->ols_owner = waiter;
863 spin_lock(&tmp_oscl->ols_lock);
864 /* add oscl into tmp's ols_waiting list */
865 list_add_tail(&oscl->ols_wait_entry,
866 &tmp_oscl->ols_waiting_list);
867 spin_unlock(&tmp_oscl->ols_lock);
869 spin_unlock(&obj->oo_ol_spin);
870 rc = cl_sync_io_wait(env, waiter, 0);
871 spin_lock(&obj->oo_ol_spin);
876 oscl->ols_owner = NULL;
879 spin_unlock(&obj->oo_ol_spin);
883 EXPORT_SYMBOL(osc_lock_enqueue_wait);
886 * Implementation of cl_lock_operations::clo_enqueue() method for osc
887 * layer. This initiates ldlm enqueue:
889 * - cancels conflicting locks early (osc_lock_enqueue_wait());
891 * - calls osc_enqueue_base() to do actual enqueue.
893 * osc_enqueue_base() is supplied with an upcall function that is executed
894 * when lock is received either after a local cached ldlm lock is matched, or
895 * when a reply from the server is received.
897 * This function does not wait for the network communication to complete.
899 static int osc_lock_enqueue(const struct lu_env *env,
900 const struct cl_lock_slice *slice,
901 struct cl_io *unused, struct cl_sync_io *anchor)
903 struct osc_thread_info *info = osc_env_info(env);
904 struct osc_io *oio = osc_env_io(env);
905 struct osc_object *osc = cl2osc(slice->cls_obj);
906 struct osc_lock *oscl = cl2osc_lock(slice);
907 struct obd_export *exp = osc_export(osc);
908 struct cl_lock *lock = slice->cls_lock;
909 struct ldlm_res_id *resname = &info->oti_resname;
910 union ldlm_policy_data *policy = &info->oti_policy;
911 osc_enqueue_upcall_f upcall = osc_lock_upcall;
918 LASSERTF(ergo(oscl->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
919 "lock = %p, ols = %p\n", lock, oscl);
921 if (oscl->ols_state == OLS_GRANTED)
924 if ((oscl->ols_flags & LDLM_FL_NO_EXPANSION) &&
925 !(exp_connect_lockahead_old(exp) || exp_connect_lockahead(exp))) {
926 result = -EOPNOTSUPP;
927 CERROR("%s: server does not support lockahead/locknoexpand:"
928 "rc = %d\n", exp->exp_obd->obd_name, result);
932 if (oscl->ols_flags & LDLM_FL_TEST_LOCK)
933 GOTO(enqueue_base, 0);
935 /* For glimpse and/or speculative locks, do not wait for reply from
936 * server on LDLM request */
937 if (oscl->ols_glimpse || oscl->ols_speculative) {
938 /* Speculative and glimpse locks do not have an anchor */
939 LASSERT(equi(oscl->ols_speculative, anchor == NULL));
941 GOTO(enqueue_base, 0);
944 result = osc_lock_enqueue_wait(env, osc, oscl);
948 /* we can grant lockless lock right after all conflicting locks
950 if (osc_lock_is_lockless(oscl)) {
951 oscl->ols_state = OLS_GRANTED;
952 oio->oi_lockless = 1;
957 oscl->ols_state = OLS_ENQUEUED;
958 if (anchor != NULL) {
959 atomic_inc(&anchor->csi_sync_nr);
960 oscl->ols_owner = anchor;
964 * DLM lock's ast data must be osc_object;
965 * if glimpse or speculative lock, async of osc_enqueue_base()
968 * For non-speculative locks:
969 * DLM's enqueue callback set to osc_lock_upcall() with cookie as
971 * For speculative locks:
972 * osc_lock_upcall_speculative & cookie is the osc object, since
973 * there is no osc_lock
975 ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
976 osc_lock_build_policy(env, lock, policy);
977 if (oscl->ols_speculative) {
978 oscl->ols_einfo.ei_cbdata = NULL;
979 /* hold a reference for callback */
980 cl_object_get(osc2cl(osc));
981 upcall = osc_lock_upcall_speculative;
984 result = osc_enqueue_base(exp, resname, &oscl->ols_flags,
985 policy, &oscl->ols_lvb,
986 osc->oo_oinfo->loi_kms_valid,
988 &oscl->ols_einfo, PTLRPCD_SET, async,
989 oscl->ols_speculative);
991 if (osc_lock_is_lockless(oscl)) {
992 oio->oi_lockless = 1;
994 LASSERT(oscl->ols_state == OLS_GRANTED);
995 LASSERT(oscl->ols_hold);
996 LASSERT(oscl->ols_dlmlock != NULL);
998 } else if (oscl->ols_speculative) {
999 cl_object_put(env, osc2cl(osc));
1000 if (oscl->ols_glimpse) {
1001 /* hide error for AGL request */
1008 oscl->ols_state = OLS_CANCELLED;
1009 osc_lock_wake_waiters(env, osc, oscl);
1012 cl_sync_io_note(env, anchor, result);
1018 * Breaks a link between osc_lock and dlm_lock.
1020 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
1022 struct ldlm_lock *dlmlock;
1026 dlmlock = olck->ols_dlmlock;
1027 if (dlmlock == NULL)
1030 if (olck->ols_hold) {
1032 ldlm_lock_decref(&olck->ols_handle, olck->ols_einfo.ei_mode);
1033 olck->ols_handle.cookie = 0ULL;
1036 olck->ols_dlmlock = NULL;
1038 /* release a reference taken in osc_lock_upcall(). */
1039 LASSERT(olck->ols_has_ref);
1040 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
1041 LDLM_LOCK_RELEASE(dlmlock);
1042 olck->ols_has_ref = 0;
1048 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1049 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1050 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1051 * with some other lock some where in the cluster. This function does the
1054 * - invalidates all pages protected by this lock (after sending dirty
1055 * ones to the server, as necessary);
1057 * - decref's underlying ldlm lock;
1059 * - cancels ldlm lock (ldlm_cli_cancel()).
1061 void osc_lock_cancel(const struct lu_env *env,
1062 const struct cl_lock_slice *slice)
1064 struct osc_object *obj = cl2osc(slice->cls_obj);
1065 struct osc_lock *oscl = cl2osc_lock(slice);
1069 LINVRNT(osc_lock_invariant(oscl));
1071 osc_lock_detach(env, oscl);
1072 oscl->ols_state = OLS_CANCELLED;
1073 oscl->ols_flags &= ~LDLM_FL_LVB_READY;
1075 osc_lock_wake_waiters(env, obj, oscl);
1078 EXPORT_SYMBOL(osc_lock_cancel);
1080 int osc_lock_print(const struct lu_env *env, void *cookie,
1081 lu_printer_t p, const struct cl_lock_slice *slice)
1083 struct osc_lock *lock = cl2osc_lock(slice);
1085 (*p)(env, cookie, "%p %#llx %#llx %d %p ",
1086 lock->ols_dlmlock, lock->ols_flags, lock->ols_handle.cookie,
1087 lock->ols_state, lock->ols_owner);
1088 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1091 EXPORT_SYMBOL(osc_lock_print);
1093 static const struct cl_lock_operations osc_lock_ops = {
1094 .clo_fini = osc_lock_fini,
1095 .clo_enqueue = osc_lock_enqueue,
1096 .clo_cancel = osc_lock_cancel,
1097 .clo_print = osc_lock_print,
1100 static void osc_lock_lockless_cancel(const struct lu_env *env,
1101 const struct cl_lock_slice *slice)
1103 struct osc_lock *ols = cl2osc_lock(slice);
1104 struct osc_object *osc = cl2osc(slice->cls_obj);
1105 struct cl_lock_descr *descr = &slice->cls_lock->cll_descr;
1108 LASSERT(ols->ols_dlmlock == NULL);
1109 result = osc_lock_flush(osc, descr->cld_start, descr->cld_end,
1110 descr->cld_mode, false);
1112 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1115 osc_lock_wake_waiters(env, osc, ols);
1118 static const struct cl_lock_operations osc_lock_lockless_ops = {
1119 .clo_fini = osc_lock_fini,
1120 .clo_enqueue = osc_lock_enqueue,
1121 .clo_cancel = osc_lock_lockless_cancel,
1122 .clo_print = osc_lock_print
1125 void osc_lock_set_writer(const struct lu_env *env, const struct cl_io *io,
1126 struct cl_object *obj, struct osc_lock *oscl)
1128 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
1132 if (!cl_object_same(io->ci_obj, obj))
1135 if (likely(io->ci_type == CIT_WRITE)) {
1136 io_start = cl_index(obj, io->u.ci_rw.rw_range.cir_pos);
1137 io_end = cl_index(obj, io->u.ci_rw.rw_range.cir_pos +
1138 io->u.ci_rw.rw_range.cir_count - 1);
1140 LASSERT(cl_io_is_mkwrite(io));
1141 io_start = io_end = io->u.ci_fault.ft_index;
1144 if (descr->cld_mode >= CLM_WRITE &&
1145 (cl_io_is_append(io) ||
1146 (descr->cld_start <= io_start && descr->cld_end >= io_end))) {
1147 struct osc_io *oio = osc_env_io(env);
1149 /* There must be only one lock to match the write region */
1150 LASSERT(oio->oi_write_osclock == NULL);
1151 oio->oi_write_osclock = oscl;
1154 EXPORT_SYMBOL(osc_lock_set_writer);
1156 int osc_lock_init(const struct lu_env *env,
1157 struct cl_object *obj, struct cl_lock *lock,
1158 const struct cl_io *io)
1160 struct osc_lock *oscl;
1161 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1163 OBD_SLAB_ALLOC_PTR_GFP(oscl, osc_lock_kmem, GFP_NOFS);
1167 oscl->ols_state = OLS_NEW;
1168 spin_lock_init(&oscl->ols_lock);
1169 INIT_LIST_HEAD(&oscl->ols_waiting_list);
1170 INIT_LIST_HEAD(&oscl->ols_wait_entry);
1171 INIT_LIST_HEAD(&oscl->ols_nextlock_oscobj);
1172 oscl->ols_lockless_ops = &osc_lock_lockless_ops;
1174 /* Speculative lock requests must be either no_expand or glimpse
1175 * request (CEF_GLIMPSE). non-glimpse no_expand speculative extent
1176 * locks will break ofd_intent_cb. (see comment there)*/
1177 LASSERT(ergo((enqflags & CEF_SPECULATIVE) != 0,
1178 (enqflags & (CEF_LOCK_NO_EXPAND | CEF_GLIMPSE)) != 0));
1180 oscl->ols_flags = osc_enq2ldlm_flags(enqflags);
1181 oscl->ols_speculative = !!(enqflags & CEF_SPECULATIVE);
1183 if (oscl->ols_flags & LDLM_FL_HAS_INTENT) {
1184 oscl->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1185 oscl->ols_glimpse = 1;
1187 osc_lock_build_einfo(env, lock, cl2osc(obj), &oscl->ols_einfo);
1189 cl_lock_slice_add(lock, &oscl->ols_cl, obj, &osc_lock_ops);
1191 if (!(enqflags & CEF_MUST))
1192 /* try to convert this lock to a lockless lock */
1193 osc_lock_to_lockless(env, oscl, (enqflags & CEF_NEVER));
1194 if (oscl->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
1195 oscl->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1197 if (io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io))
1198 osc_lock_set_writer(env, io, obj, oscl);
1200 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %#llx",
1201 lock, oscl, oscl->ols_flags);
1207 * Finds an existing lock covering given index and optionally different from a
1208 * given \a except lock.
1210 struct ldlm_lock *osc_obj_dlmlock_at_pgoff(const struct lu_env *env,
1211 struct osc_object *obj,
1213 enum osc_dap_flags dap_flags)
1215 struct osc_thread_info *info = osc_env_info(env);
1216 struct ldlm_res_id *resname = &info->oti_resname;
1217 union ldlm_policy_data *policy = &info->oti_policy;
1218 struct lustre_handle lockh;
1219 struct ldlm_lock *lock = NULL;
1220 enum ldlm_mode mode;
1225 ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1226 osc_index2policy(policy, osc2cl(obj), index, index);
1227 policy->l_extent.gid = LDLM_GID_ANY;
1229 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1230 if (dap_flags & OSC_DAP_FL_TEST_LOCK)
1231 flags |= LDLM_FL_TEST_LOCK;
1233 * It is fine to match any group lock since there could be only one
1234 * with a uniq gid and it conflicts with all other lock modes too
1237 mode = osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
1238 LCK_PR | LCK_PW | LCK_GROUP, &flags, obj, &lockh,
1239 dap_flags & OSC_DAP_FL_CANCELING);
1241 lock = ldlm_handle2lock(&lockh);
1242 /* RACE: the lock is cancelled so let's try again */
1243 if (unlikely(lock == NULL))