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. */
309 } else if (rc < 0 && oscl->ols_flags & LDLM_FL_NDELAY) {
313 if (oscl->ols_owner != NULL)
314 cl_sync_io_note(env, oscl->ols_owner, rc);
315 cl_env_percpu_put(env);
320 static int osc_lock_upcall_speculative(void *cookie,
321 struct lustre_handle *lockh,
324 struct osc_object *osc = cookie;
325 struct ldlm_lock *dlmlock;
330 env = cl_env_get(&refcheck);
331 LASSERT(!IS_ERR(env));
333 if (errcode == ELDLM_LOCK_MATCHED)
334 GOTO(out, errcode = ELDLM_OK);
336 if (errcode != ELDLM_OK)
339 dlmlock = ldlm_handle2lock(lockh);
340 LASSERT(dlmlock != NULL);
342 lock_res_and_lock(dlmlock);
343 LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
345 /* there is no osc_lock associated with speculative locks */
346 osc_lock_lvb_update(env, osc, dlmlock, NULL);
348 unlock_res_and_lock(dlmlock);
349 LDLM_LOCK_PUT(dlmlock);
352 cl_object_put(env, osc2cl(osc));
353 cl_env_put(env, &refcheck);
354 RETURN(ldlm_error2errno(errcode));
357 static int osc_lock_flush(struct osc_object *obj, pgoff_t start, pgoff_t end,
358 enum cl_lock_mode mode, bool discard)
367 env = cl_env_get(&refcheck);
369 RETURN(PTR_ERR(env));
371 if (mode == CLM_WRITE) {
372 rc = osc_cache_writeback_range(env, obj, start, end, 1,
374 CDEBUG(D_CACHE, "object %p: [%lu -> %lu] %d pages were %s.\n",
376 discard ? "discarded" : "written back");
381 rc2 = osc_lock_discard_pages(env, obj, start, end, discard);
382 if (rc == 0 && rc2 < 0)
385 cl_env_put(env, &refcheck);
390 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
391 * and ldlm_lock caches.
393 static int osc_dlm_blocking_ast0(const struct lu_env *env,
394 struct ldlm_lock *dlmlock,
395 void *data, int flag)
397 struct cl_object *obj = NULL;
400 enum cl_lock_mode mode = CLM_READ;
403 LASSERT(flag == LDLM_CB_CANCELING);
405 lock_res_and_lock(dlmlock);
406 if (dlmlock->l_granted_mode != dlmlock->l_req_mode) {
407 dlmlock->l_ast_data = NULL;
408 unlock_res_and_lock(dlmlock);
412 discard = ldlm_is_discard_data(dlmlock);
413 if (dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))
416 if (dlmlock->l_ast_data != NULL) {
417 obj = osc2cl(dlmlock->l_ast_data);
418 dlmlock->l_ast_data = NULL;
423 unlock_res_and_lock(dlmlock);
425 /* if l_ast_data is NULL, the dlmlock was enqueued by AGL or
426 * the object has been destroyed. */
428 struct ldlm_extent *extent = &dlmlock->l_policy_data.l_extent;
429 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
432 /* Destroy pages covered by the extent of the DLM lock */
433 result = osc_lock_flush(cl2osc(obj),
434 cl_index(obj, extent->start),
435 cl_index(obj, extent->end),
438 /* losing a lock, update kms */
439 lock_res_and_lock(dlmlock);
440 cl_object_attr_lock(obj);
441 /* Must get the value under the lock to avoid race. */
442 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
443 /* Update the kms. Need to loop all granted locks.
444 * Not a problem for the client */
445 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
447 cl_object_attr_update(env, obj, attr, CAT_KMS);
448 cl_object_attr_unlock(obj);
449 unlock_res_and_lock(dlmlock);
451 cl_object_put(env, obj);
457 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
458 * some other lock, or is canceled. This function is installed as a
459 * ldlm_lock::l_blocking_ast() for client extent locks.
461 * Control flow is tricky, because ldlm uses the same call-back
462 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
464 * \param dlmlock lock for which ast occurred.
466 * \param new description of a conflicting lock in case of blocking ast.
468 * \param data value of dlmlock->l_ast_data
470 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
471 * cancellation and blocking ast's.
473 * Possible use cases:
475 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
476 * lock due to lock lru pressure, or explicit user request to purge
479 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
480 * us that dlmlock conflicts with another lock that some client is
481 * enqueuing. Lock is canceled.
483 * - cl_lock_cancel() is called. osc_lock_cancel() calls
484 * ldlm_cli_cancel() that calls
486 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
488 * recursively entering osc_ldlm_blocking_ast().
490 * - client cancels lock voluntary (e.g., as a part of early cancellation):
493 * osc_lock_cancel()->
494 * ldlm_cli_cancel()->
495 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
498 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
499 struct ldlm_lock_desc *new, void *data,
506 case LDLM_CB_BLOCKING: {
507 struct lustre_handle lockh;
509 ldlm_lock2handle(dlmlock, &lockh);
510 result = ldlm_cli_cancel(&lockh, LCF_ASYNC);
511 if (result == -ENODATA)
515 case LDLM_CB_CANCELING: {
520 * This can be called in the context of outer IO, e.g.,
522 * osc_enqueue_base()->...
523 * ->ldlm_prep_elc_req()->...
524 * ->ldlm_cancel_callback()->...
525 * ->osc_ldlm_blocking_ast()
527 * new environment has to be created to not corrupt outer
530 env = cl_env_get(&refcheck);
532 result = PTR_ERR(env);
536 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
537 cl_env_put(env, &refcheck);
546 int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
548 struct ptlrpc_request *req = data;
551 struct req_capsule *cap;
552 struct cl_object *obj = NULL;
558 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
560 env = cl_env_get(&refcheck);
562 GOTO(out, result = PTR_ERR(env));
565 lock_res_and_lock(dlmlock);
566 if (dlmlock->l_ast_data != NULL) {
567 obj = osc2cl(dlmlock->l_ast_data);
570 unlock_res_and_lock(dlmlock);
573 /* Do not grab the mutex of cl_lock for glimpse.
574 * See LU-1274 for details.
575 * BTW, it's okay for cl_lock to be cancelled during
576 * this period because server can handle this race.
577 * See ldlm_server_glimpse_ast() for details.
578 * cl_lock_mutex_get(env, lock); */
580 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
581 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
583 result = req_capsule_server_pack(cap);
585 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
586 result = cl_object_glimpse(env, obj, lvb);
588 if (!exp_connect_lvb_type(req->rq_export))
589 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
590 sizeof(struct ost_lvb_v1), RCL_SERVER);
591 cl_object_put(env, obj);
594 * These errors are normal races, so we don't want to
595 * fill the console with messages by calling
598 lustre_pack_reply(req, 1, NULL, NULL);
599 result = -ELDLM_NO_LOCK_DATA;
601 cl_env_put(env, &refcheck);
605 req->rq_status = result;
608 EXPORT_SYMBOL(osc_ldlm_glimpse_ast);
610 static int weigh_cb(const struct lu_env *env, struct cl_io *io,
611 struct osc_page *ops, void *cbdata)
613 struct cl_page *page = ops->ops_cl.cpl_page;
615 if (cl_page_is_vmlocked(env, page)
616 || PageDirty(page->cp_vmpage) || PageWriteback(page->cp_vmpage)
618 return CLP_GANG_ABORT;
620 *(pgoff_t *)cbdata = osc_index(ops) + 1;
621 return CLP_GANG_OKAY;
624 static unsigned long osc_lock_weight(const struct lu_env *env,
625 struct osc_object *oscobj,
626 struct ldlm_extent *extent)
628 struct cl_io *io = osc_env_thread_io(env);
629 struct cl_object *obj = cl_object_top(&oscobj->oo_cl);
635 io->ci_ignore_layout = 1;
636 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
640 page_index = cl_index(obj, extent->start);
642 result = osc_page_gang_lookup(env, io, oscobj,
644 cl_index(obj, extent->end),
645 weigh_cb, (void *)&page_index);
646 if (result == CLP_GANG_ABORT)
648 if (result == CLP_GANG_RESCHED)
650 } while (result != CLP_GANG_OKAY);
653 return result == CLP_GANG_ABORT ? 1 : 0;
657 * Get the weight of dlm lock for early cancellation.
659 unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
662 struct osc_object *obj;
663 struct osc_lock *oscl;
664 unsigned long weight;
671 * osc_ldlm_weigh_ast has a complex context since it might be called
672 * because of lock canceling, or from user's input. We have to make
673 * a new environment for it. Probably it is implementation safe to use
674 * the upper context because cl_lock_put don't modify environment
675 * variables. But just in case ..
677 env = cl_env_get(&refcheck);
679 /* Mostly because lack of memory, do not eliminate this lock */
682 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
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 weight = osc_lock_weight(env, obj, &dlmlock->l_policy_data.l_extent);
711 cl_object_put(env, osc2cl(obj));
713 cl_env_put(env, &refcheck);
717 static void osc_lock_build_einfo(const struct lu_env *env,
718 const struct cl_lock *lock,
719 struct osc_object *osc,
720 struct ldlm_enqueue_info *einfo)
722 einfo->ei_type = LDLM_EXTENT;
723 einfo->ei_mode = osc_cl_lock2ldlm(lock->cll_descr.cld_mode);
724 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
725 einfo->ei_cb_cp = ldlm_completion_ast;
726 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
727 einfo->ei_cbdata = osc; /* value to be put into ->l_ast_data */
731 * Determine if the lock should be converted into a lockless lock.
734 * - if the lock has an explicite requirment for a non-lockless lock;
735 * - if the io lock request type ci_lockreq;
736 * - send the enqueue rpc to ost to make the further decision;
737 * - special treat to truncate lockless lock
739 * Additional policy can be implemented here, e.g., never do lockless-io
742 void osc_lock_to_lockless(const struct lu_env *env,
743 struct osc_lock *ols, int force)
745 struct cl_lock_slice *slice = &ols->ols_cl;
746 struct osc_io *oio = osc_env_io(env);
747 struct cl_io *io = oio->oi_cl.cis_io;
748 struct cl_object *obj = slice->cls_obj;
749 struct osc_object *oob = cl2osc(obj);
750 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
751 struct obd_connect_data *ocd;
753 LASSERT(ols->ols_state == OLS_NEW ||
754 ols->ols_state == OLS_UPCALL_RECEIVED);
757 ols->ols_locklessable = 1;
758 slice->cls_ops = ols->ols_lockless_ops;
760 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
761 io->ci_lockreq == CILR_MAYBE ||
762 io->ci_lockreq == CILR_NEVER);
764 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
765 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
766 (io->ci_lockreq == CILR_MAYBE) &&
767 (ocd->ocd_connect_flags &
768 OBD_CONNECT_SRVLOCK);
769 if (io->ci_lockreq == CILR_NEVER ||
771 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
772 /* lockless truncate */
773 (cl_io_is_trunc(io) && osd->od_lockless_truncate &&
774 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK))) {
775 ols->ols_locklessable = 1;
776 slice->cls_ops = ols->ols_lockless_ops;
779 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
781 EXPORT_SYMBOL(osc_lock_to_lockless);
783 static bool osc_lock_compatible(const struct osc_lock *qing,
784 const struct osc_lock *qed)
786 struct cl_lock_descr *qed_descr = &qed->ols_cl.cls_lock->cll_descr;
787 struct cl_lock_descr *qing_descr = &qing->ols_cl.cls_lock->cll_descr;
789 if (qed->ols_glimpse || qed->ols_speculative)
792 if (qing_descr->cld_mode == CLM_READ && qed_descr->cld_mode == CLM_READ)
795 if (qed->ols_state < OLS_GRANTED)
798 if (qed_descr->cld_mode >= qing_descr->cld_mode &&
799 qed_descr->cld_start <= qing_descr->cld_start &&
800 qed_descr->cld_end >= qing_descr->cld_end)
806 void osc_lock_wake_waiters(const struct lu_env *env, struct osc_object *osc,
807 struct osc_lock *oscl)
809 spin_lock(&osc->oo_ol_spin);
810 list_del_init(&oscl->ols_nextlock_oscobj);
811 spin_unlock(&osc->oo_ol_spin);
813 spin_lock(&oscl->ols_lock);
814 while (!list_empty(&oscl->ols_waiting_list)) {
815 struct osc_lock *scan;
817 scan = list_entry(oscl->ols_waiting_list.next, struct osc_lock,
819 list_del_init(&scan->ols_wait_entry);
821 cl_sync_io_note(env, scan->ols_owner, 0);
823 spin_unlock(&oscl->ols_lock);
825 EXPORT_SYMBOL(osc_lock_wake_waiters);
827 int osc_lock_enqueue_wait(const struct lu_env *env, struct osc_object *obj,
828 struct osc_lock *oscl)
830 struct osc_lock *tmp_oscl;
831 struct cl_lock_descr *need = &oscl->ols_cl.cls_lock->cll_descr;
832 struct cl_sync_io *waiter = &osc_env_info(env)->oti_anchor;
837 spin_lock(&obj->oo_ol_spin);
838 list_add_tail(&oscl->ols_nextlock_oscobj, &obj->oo_ol_list);
841 list_for_each_entry(tmp_oscl, &obj->oo_ol_list,
842 ols_nextlock_oscobj) {
843 struct cl_lock_descr *descr;
845 if (tmp_oscl == oscl)
848 descr = &tmp_oscl->ols_cl.cls_lock->cll_descr;
849 if (descr->cld_start > need->cld_end ||
850 descr->cld_end < need->cld_start)
853 /* We're not supposed to give up group lock */
854 if (descr->cld_mode == CLM_GROUP)
857 if (!osc_lock_is_lockless(oscl) &&
858 osc_lock_compatible(oscl, tmp_oscl))
861 /* wait for conflicting lock to be canceled */
862 cl_sync_io_init(waiter, 1, cl_sync_io_end);
863 oscl->ols_owner = waiter;
865 spin_lock(&tmp_oscl->ols_lock);
866 /* add oscl into tmp's ols_waiting list */
867 list_add_tail(&oscl->ols_wait_entry,
868 &tmp_oscl->ols_waiting_list);
869 spin_unlock(&tmp_oscl->ols_lock);
871 spin_unlock(&obj->oo_ol_spin);
872 rc = cl_sync_io_wait(env, waiter, 0);
873 spin_lock(&obj->oo_ol_spin);
878 oscl->ols_owner = NULL;
881 spin_unlock(&obj->oo_ol_spin);
885 EXPORT_SYMBOL(osc_lock_enqueue_wait);
888 * Implementation of cl_lock_operations::clo_enqueue() method for osc
889 * layer. This initiates ldlm enqueue:
891 * - cancels conflicting locks early (osc_lock_enqueue_wait());
893 * - calls osc_enqueue_base() to do actual enqueue.
895 * osc_enqueue_base() is supplied with an upcall function that is executed
896 * when lock is received either after a local cached ldlm lock is matched, or
897 * when a reply from the server is received.
899 * This function does not wait for the network communication to complete.
901 static int osc_lock_enqueue(const struct lu_env *env,
902 const struct cl_lock_slice *slice,
903 struct cl_io *unused, struct cl_sync_io *anchor)
905 struct osc_thread_info *info = osc_env_info(env);
906 struct osc_io *oio = osc_env_io(env);
907 struct osc_object *osc = cl2osc(slice->cls_obj);
908 struct osc_lock *oscl = cl2osc_lock(slice);
909 struct obd_export *exp = osc_export(osc);
910 struct cl_lock *lock = slice->cls_lock;
911 struct ldlm_res_id *resname = &info->oti_resname;
912 union ldlm_policy_data *policy = &info->oti_policy;
913 osc_enqueue_upcall_f upcall = osc_lock_upcall;
920 LASSERTF(ergo(oscl->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
921 "lock = %p, ols = %p\n", lock, oscl);
923 if (oscl->ols_state == OLS_GRANTED)
926 if ((oscl->ols_flags & LDLM_FL_NO_EXPANSION) &&
927 !(exp_connect_lockahead_old(exp) || exp_connect_lockahead(exp))) {
928 result = -EOPNOTSUPP;
929 CERROR("%s: server does not support lockahead/locknoexpand:"
930 "rc = %d\n", exp->exp_obd->obd_name, result);
934 if (oscl->ols_flags & LDLM_FL_TEST_LOCK)
935 GOTO(enqueue_base, 0);
937 /* For glimpse and/or speculative locks, do not wait for reply from
938 * server on LDLM request */
939 if (oscl->ols_glimpse || oscl->ols_speculative) {
940 /* Speculative and glimpse locks do not have an anchor */
941 LASSERT(equi(oscl->ols_speculative, anchor == NULL));
943 GOTO(enqueue_base, 0);
946 result = osc_lock_enqueue_wait(env, osc, oscl);
950 /* we can grant lockless lock right after all conflicting locks
952 if (osc_lock_is_lockless(oscl)) {
953 oscl->ols_state = OLS_GRANTED;
954 oio->oi_lockless = 1;
959 oscl->ols_state = OLS_ENQUEUED;
960 if (anchor != NULL) {
961 atomic_inc(&anchor->csi_sync_nr);
962 oscl->ols_owner = anchor;
966 * DLM lock's ast data must be osc_object;
967 * if glimpse or speculative lock, async of osc_enqueue_base()
970 * For non-speculative locks:
971 * DLM's enqueue callback set to osc_lock_upcall() with cookie as
973 * For speculative locks:
974 * osc_lock_upcall_speculative & cookie is the osc object, since
975 * there is no osc_lock
977 ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
978 osc_lock_build_policy(env, lock, policy);
979 if (oscl->ols_speculative) {
980 oscl->ols_einfo.ei_cbdata = NULL;
981 /* hold a reference for callback */
982 cl_object_get(osc2cl(osc));
983 upcall = osc_lock_upcall_speculative;
986 result = osc_enqueue_base(exp, resname, &oscl->ols_flags,
987 policy, &oscl->ols_lvb,
988 osc->oo_oinfo->loi_kms_valid,
990 &oscl->ols_einfo, PTLRPCD_SET, async,
991 oscl->ols_speculative);
993 if (osc_lock_is_lockless(oscl)) {
994 oio->oi_lockless = 1;
996 LASSERT(oscl->ols_state == OLS_GRANTED);
997 LASSERT(oscl->ols_hold);
998 LASSERT(oscl->ols_dlmlock != NULL);
1000 } else if (oscl->ols_speculative) {
1001 cl_object_put(env, osc2cl(osc));
1002 if (oscl->ols_glimpse) {
1003 /* hide error for AGL request */
1010 oscl->ols_state = OLS_CANCELLED;
1011 osc_lock_wake_waiters(env, osc, oscl);
1014 cl_sync_io_note(env, anchor, result);
1020 * Breaks a link between osc_lock and dlm_lock.
1022 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
1024 struct ldlm_lock *dlmlock;
1028 dlmlock = olck->ols_dlmlock;
1029 if (dlmlock == NULL)
1032 if (olck->ols_hold) {
1034 ldlm_lock_decref(&olck->ols_handle, olck->ols_einfo.ei_mode);
1035 olck->ols_handle.cookie = 0ULL;
1038 olck->ols_dlmlock = NULL;
1040 /* release a reference taken in osc_lock_upcall(). */
1041 LASSERT(olck->ols_has_ref);
1042 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
1043 LDLM_LOCK_RELEASE(dlmlock);
1044 olck->ols_has_ref = 0;
1050 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1051 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1052 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1053 * with some other lock some where in the cluster. This function does the
1056 * - invalidates all pages protected by this lock (after sending dirty
1057 * ones to the server, as necessary);
1059 * - decref's underlying ldlm lock;
1061 * - cancels ldlm lock (ldlm_cli_cancel()).
1063 void osc_lock_cancel(const struct lu_env *env,
1064 const struct cl_lock_slice *slice)
1066 struct osc_object *obj = cl2osc(slice->cls_obj);
1067 struct osc_lock *oscl = cl2osc_lock(slice);
1071 LINVRNT(osc_lock_invariant(oscl));
1073 osc_lock_detach(env, oscl);
1074 oscl->ols_state = OLS_CANCELLED;
1075 oscl->ols_flags &= ~LDLM_FL_LVB_READY;
1077 osc_lock_wake_waiters(env, obj, oscl);
1080 EXPORT_SYMBOL(osc_lock_cancel);
1082 int osc_lock_print(const struct lu_env *env, void *cookie,
1083 lu_printer_t p, const struct cl_lock_slice *slice)
1085 struct osc_lock *lock = cl2osc_lock(slice);
1087 (*p)(env, cookie, "%p %#llx %#llx %d %p ",
1088 lock->ols_dlmlock, lock->ols_flags, lock->ols_handle.cookie,
1089 lock->ols_state, lock->ols_owner);
1090 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1093 EXPORT_SYMBOL(osc_lock_print);
1095 static const struct cl_lock_operations osc_lock_ops = {
1096 .clo_fini = osc_lock_fini,
1097 .clo_enqueue = osc_lock_enqueue,
1098 .clo_cancel = osc_lock_cancel,
1099 .clo_print = osc_lock_print,
1102 static void osc_lock_lockless_cancel(const struct lu_env *env,
1103 const struct cl_lock_slice *slice)
1105 struct osc_lock *ols = cl2osc_lock(slice);
1106 struct osc_object *osc = cl2osc(slice->cls_obj);
1107 struct cl_lock_descr *descr = &slice->cls_lock->cll_descr;
1110 LASSERT(ols->ols_dlmlock == NULL);
1111 result = osc_lock_flush(osc, descr->cld_start, descr->cld_end,
1112 descr->cld_mode, false);
1114 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1117 osc_lock_wake_waiters(env, osc, ols);
1120 static const struct cl_lock_operations osc_lock_lockless_ops = {
1121 .clo_fini = osc_lock_fini,
1122 .clo_enqueue = osc_lock_enqueue,
1123 .clo_cancel = osc_lock_lockless_cancel,
1124 .clo_print = osc_lock_print
1127 void osc_lock_set_writer(const struct lu_env *env, const struct cl_io *io,
1128 struct cl_object *obj, struct osc_lock *oscl)
1130 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
1134 if (!cl_object_same(io->ci_obj, obj))
1137 if (likely(io->ci_type == CIT_WRITE)) {
1138 io_start = cl_index(obj, io->u.ci_rw.rw_range.cir_pos);
1139 io_end = cl_index(obj, io->u.ci_rw.rw_range.cir_pos +
1140 io->u.ci_rw.rw_range.cir_count - 1);
1142 LASSERT(cl_io_is_mkwrite(io));
1143 io_start = io_end = io->u.ci_fault.ft_index;
1146 if (descr->cld_mode >= CLM_WRITE &&
1147 (cl_io_is_append(io) ||
1148 (descr->cld_start <= io_start && descr->cld_end >= io_end))) {
1149 struct osc_io *oio = osc_env_io(env);
1151 /* There must be only one lock to match the write region */
1152 LASSERT(oio->oi_write_osclock == NULL);
1153 oio->oi_write_osclock = oscl;
1156 EXPORT_SYMBOL(osc_lock_set_writer);
1158 int osc_lock_init(const struct lu_env *env,
1159 struct cl_object *obj, struct cl_lock *lock,
1160 const struct cl_io *io)
1162 struct osc_lock *oscl;
1163 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1165 OBD_SLAB_ALLOC_PTR_GFP(oscl, osc_lock_kmem, GFP_NOFS);
1169 oscl->ols_state = OLS_NEW;
1170 spin_lock_init(&oscl->ols_lock);
1171 INIT_LIST_HEAD(&oscl->ols_waiting_list);
1172 INIT_LIST_HEAD(&oscl->ols_wait_entry);
1173 INIT_LIST_HEAD(&oscl->ols_nextlock_oscobj);
1174 oscl->ols_lockless_ops = &osc_lock_lockless_ops;
1176 /* Speculative lock requests must be either no_expand or glimpse
1177 * request (CEF_GLIMPSE). non-glimpse no_expand speculative extent
1178 * locks will break ofd_intent_cb. (see comment there)*/
1179 LASSERT(ergo((enqflags & CEF_SPECULATIVE) != 0,
1180 (enqflags & (CEF_LOCK_NO_EXPAND | CEF_GLIMPSE)) != 0));
1182 oscl->ols_flags = osc_enq2ldlm_flags(enqflags);
1183 oscl->ols_speculative = !!(enqflags & CEF_SPECULATIVE);
1185 if (oscl->ols_flags & LDLM_FL_HAS_INTENT) {
1186 oscl->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1187 oscl->ols_glimpse = 1;
1189 if (io->ci_ndelay && cl_object_same(io->ci_obj, obj))
1190 oscl->ols_flags |= LDLM_FL_NDELAY;
1191 osc_lock_build_einfo(env, lock, cl2osc(obj), &oscl->ols_einfo);
1193 cl_lock_slice_add(lock, &oscl->ols_cl, obj, &osc_lock_ops);
1195 if (!(enqflags & CEF_MUST))
1196 /* try to convert this lock to a lockless lock */
1197 osc_lock_to_lockless(env, oscl, (enqflags & CEF_NEVER));
1198 if (oscl->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
1199 oscl->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1201 if (io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io))
1202 osc_lock_set_writer(env, io, obj, oscl);
1204 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %#llx",
1205 lock, oscl, oscl->ols_flags);
1211 * Finds an existing lock covering given index and optionally different from a
1212 * given \a except lock.
1214 struct ldlm_lock *osc_obj_dlmlock_at_pgoff(const struct lu_env *env,
1215 struct osc_object *obj,
1217 enum osc_dap_flags dap_flags)
1219 struct osc_thread_info *info = osc_env_info(env);
1220 struct ldlm_res_id *resname = &info->oti_resname;
1221 union ldlm_policy_data *policy = &info->oti_policy;
1222 struct lustre_handle lockh;
1223 struct ldlm_lock *lock = NULL;
1224 enum ldlm_mode mode;
1229 ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1230 osc_index2policy(policy, osc2cl(obj), index, index);
1231 policy->l_extent.gid = LDLM_GID_ANY;
1233 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1234 if (dap_flags & OSC_DAP_FL_TEST_LOCK)
1235 flags |= LDLM_FL_TEST_LOCK;
1237 * It is fine to match any group lock since there could be only one
1238 * with a uniq gid and it conflicts with all other lock modes too
1241 mode = osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
1242 LCK_PR | LCK_PW | LCK_GROUP, &flags, obj, &lockh,
1243 dap_flags & OSC_DAP_FL_CANCELING);
1245 lock = ldlm_handle2lock(&lockh);
1246 /* RACE: the lock is cancelled so let's try again */
1247 if (unlikely(lock == NULL))