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, 2015, 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 <linux/libcfs/libcfs.h>
41 /* fid_build_reg_res_name() */
42 #include <lustre_fid.h>
44 #include "osc_cl_internal.h"
50 /*****************************************************************************
56 static const struct cl_lock_operations osc_lock_ops;
57 static const struct cl_lock_operations osc_lock_lockless_ops;
58 static void osc_lock_to_lockless(const struct lu_env *env,
59 struct osc_lock *ols, int force);
61 int osc_lock_is_lockless(const struct osc_lock *olck)
63 return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
67 * Returns a weak pointer to the ldlm lock identified by a handle. Returned
68 * pointer cannot be dereferenced, as lock is not protected from concurrent
69 * reclaim. This function is a helper for osc_lock_invariant().
71 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
73 struct ldlm_lock *lock;
75 lock = ldlm_handle2lock(handle);
82 * Invariant that has to be true all of the time.
84 static int osc_lock_invariant(struct osc_lock *ols)
86 struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle);
87 struct ldlm_lock *olock = ols->ols_dlmlock;
88 int handle_used = lustre_handle_is_used(&ols->ols_handle);
90 if (ergo(osc_lock_is_lockless(ols),
91 ols->ols_locklessable && !ols->ols_dlmlock))
95 * If all the following "ergo"s are true, return 1, otherwise 0
97 if (!ergo(olock, handle_used))
100 if (!ergo(olock, olock->l_handle.h_cookie == ols->ols_handle.cookie))
103 if (!ergo(handle_used,
104 ergo(lock && olock, lock == olock) &&
105 ergo(!lock, !olock)))
108 * Check that ->ols_handle and ->ols_dlmlock are consistent, but
109 * take into account that they are set at the different time.
111 if (!ergo(ols->ols_state == OLS_CANCELLED,
112 !olock && !handle_used))
115 * DLM lock is destroyed only after we have seen cancellation
118 if (!ergo(olock && ols->ols_state < OLS_CANCELLED,
119 !ldlm_is_destroyed(olock)))
122 if (!ergo(ols->ols_state == OLS_GRANTED,
123 olock && olock->l_req_mode == olock->l_granted_mode &&
129 /*****************************************************************************
135 static void osc_lock_fini(const struct lu_env *env,
136 struct cl_lock_slice *slice)
138 struct osc_lock *ols = cl2osc_lock(slice);
140 LINVRNT(osc_lock_invariant(ols));
141 LASSERT(!ols->ols_dlmlock);
143 kmem_cache_free(osc_lock_kmem, ols);
146 static void osc_lock_build_policy(const struct lu_env *env,
147 const struct cl_lock *lock,
148 union ldlm_policy_data *policy)
150 const struct cl_lock_descr *d = &lock->cll_descr;
152 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
153 policy->l_extent.gid = d->cld_gid;
156 static __u64 osc_enq2ldlm_flags(__u32 enqflags)
160 LASSERT((enqflags & ~CEF_MASK) == 0);
162 if (enqflags & CEF_NONBLOCK)
163 result |= LDLM_FL_BLOCK_NOWAIT;
164 if (enqflags & CEF_ASYNC)
165 result |= LDLM_FL_HAS_INTENT;
166 if (enqflags & CEF_DISCARD_DATA)
167 result |= LDLM_FL_AST_DISCARD_DATA;
168 if (enqflags & CEF_PEEK)
169 result |= LDLM_FL_TEST_LOCK;
170 if (enqflags & CEF_LOCK_MATCH)
171 result |= LDLM_FL_MATCH_LOCK;
176 * Updates object attributes from a lock value block (lvb) received together
177 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
180 * This can be optimized to not update attributes when lock is a result of a
183 * Called under lock and resource spin-locks.
185 static void osc_lock_lvb_update(const struct lu_env *env,
186 struct osc_object *osc,
187 struct ldlm_lock *dlmlock,
190 struct cl_object *obj = osc2cl(osc);
191 struct lov_oinfo *oinfo = osc->oo_oinfo;
192 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
195 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
197 lvb = dlmlock->l_lvb_data;
199 cl_lvb2attr(attr, lvb);
201 cl_object_attr_lock(obj);
205 check_res_locked(dlmlock->l_resource);
206 LASSERT(lvb == dlmlock->l_lvb_data);
207 size = lvb->lvb_size;
209 /* Extend KMS up to the end of this lock and no further
210 * A lock on [x,y] means a KMS of up to y + 1 bytes!
212 if (size > dlmlock->l_policy_data.l_extent.end)
213 size = dlmlock->l_policy_data.l_extent.end + 1;
214 if (size >= oinfo->loi_kms) {
215 LDLM_DEBUG(dlmlock, "lock acquired, setting rss=%llu, kms=%llu",
216 lvb->lvb_size, size);
218 attr->cat_kms = size;
220 LDLM_DEBUG(dlmlock, "lock acquired, setting rss=%llu; leaving kms=%llu, end=%llu",
221 lvb->lvb_size, oinfo->loi_kms,
222 dlmlock->l_policy_data.l_extent.end);
224 ldlm_lock_allow_match_locked(dlmlock);
227 cl_object_attr_update(env, obj, attr, valid);
228 cl_object_attr_unlock(obj);
231 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *oscl,
232 struct lustre_handle *lockh, bool lvb_update)
234 struct ldlm_lock *dlmlock;
236 dlmlock = ldlm_handle2lock_long(lockh, 0);
239 /* lock reference taken by ldlm_handle2lock_long() is
240 * owned by osc_lock and released in osc_lock_detach()
242 lu_ref_add(&dlmlock->l_reference, "osc_lock", oscl);
243 oscl->ols_has_ref = 1;
245 LASSERT(!oscl->ols_dlmlock);
246 oscl->ols_dlmlock = dlmlock;
248 /* This may be a matched lock for glimpse request, do not hold
249 * lock reference in that case.
251 if (!oscl->ols_glimpse) {
252 /* hold a refc for non glimpse lock which will
253 * be released in osc_lock_cancel()
255 lustre_handle_copy(&oscl->ols_handle, lockh);
256 ldlm_lock_addref(lockh, oscl->ols_einfo.ei_mode);
260 /* Lock must have been granted. */
261 lock_res_and_lock(dlmlock);
262 if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
263 struct ldlm_extent *ext = &dlmlock->l_policy_data.l_extent;
264 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
266 /* extend the lock extent, otherwise it will have problem when
267 * we decide whether to grant a lockless lock.
269 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
270 descr->cld_start = cl_index(descr->cld_obj, ext->start);
271 descr->cld_end = cl_index(descr->cld_obj, ext->end);
272 descr->cld_gid = ext->gid;
274 /* no lvb update for matched lock */
276 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
277 osc_lock_lvb_update(env, cl2osc(oscl->ols_cl.cls_obj),
280 LINVRNT(osc_lock_invariant(oscl));
282 unlock_res_and_lock(dlmlock);
284 LASSERT(oscl->ols_state != OLS_GRANTED);
285 oscl->ols_state = OLS_GRANTED;
289 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
290 * received from a server, or after osc_enqueue_base() matched a local DLM
293 static int osc_lock_upcall(void *cookie, struct lustre_handle *lockh,
296 struct osc_lock *oscl = cookie;
297 struct cl_lock_slice *slice = &oscl->ols_cl;
302 env = cl_env_get(&refcheck);
303 /* should never happen, similar to osc_ldlm_blocking_ast(). */
304 LASSERT(!IS_ERR(env));
306 rc = ldlm_error2errno(errcode);
307 if (oscl->ols_state == OLS_ENQUEUED) {
308 oscl->ols_state = OLS_UPCALL_RECEIVED;
309 } else if (oscl->ols_state == OLS_CANCELLED) {
312 CERROR("Impossible state: %d\n", oscl->ols_state);
317 osc_lock_granted(env, oscl, lockh, errcode == ELDLM_OK);
319 /* Error handling, some errors are tolerable. */
320 if (oscl->ols_locklessable && rc == -EUSERS) {
321 /* This is a tolerable error, turn this lock into
324 osc_object_set_contended(cl2osc(slice->cls_obj));
325 LASSERT(slice->cls_ops == &osc_lock_ops);
327 /* Change this lock to ldlmlock-less lock. */
328 osc_lock_to_lockless(env, oscl, 1);
329 oscl->ols_state = OLS_GRANTED;
331 } else if (oscl->ols_glimpse && rc == -ENAVAIL) {
332 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
333 osc_lock_lvb_update(env, cl2osc(slice->cls_obj),
334 NULL, &oscl->ols_lvb);
335 /* Hide the error. */
340 cl_sync_io_note(env, oscl->ols_owner, rc);
341 cl_env_put(env, &refcheck);
346 static int osc_lock_upcall_agl(void *cookie, struct lustre_handle *lockh,
349 struct osc_object *osc = cookie;
350 struct ldlm_lock *dlmlock;
354 env = cl_env_get(&refcheck);
355 LASSERT(!IS_ERR(env));
357 if (errcode == ELDLM_LOCK_MATCHED) {
362 if (errcode != ELDLM_OK)
365 dlmlock = ldlm_handle2lock(lockh);
368 lock_res_and_lock(dlmlock);
369 LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
371 /* there is no osc_lock associated with AGL lock */
372 osc_lock_lvb_update(env, osc, dlmlock, NULL);
374 unlock_res_and_lock(dlmlock);
375 LDLM_LOCK_PUT(dlmlock);
378 cl_object_put(env, osc2cl(osc));
379 cl_env_put(env, &refcheck);
380 return ldlm_error2errno(errcode);
383 static int osc_lock_flush(struct osc_object *obj, pgoff_t start, pgoff_t end,
384 enum cl_lock_mode mode, int discard)
391 env = cl_env_get(&refcheck);
395 if (mode == CLM_WRITE) {
396 rc = osc_cache_writeback_range(env, obj, start, end, 1,
398 CDEBUG(D_CACHE, "object %p: [%lu -> %lu] %d pages were %s.\n",
400 discard ? "discarded" : "written back");
405 rc2 = osc_lock_discard_pages(env, obj, start, end, mode);
406 if (rc == 0 && rc2 < 0)
409 cl_env_put(env, &refcheck);
414 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
415 * and ldlm_lock caches.
417 static int osc_dlm_blocking_ast0(const struct lu_env *env,
418 struct ldlm_lock *dlmlock,
419 void *data, int flag)
421 struct cl_object *obj = NULL;
424 enum cl_lock_mode mode = CLM_READ;
426 LASSERT(flag == LDLM_CB_CANCELING);
428 lock_res_and_lock(dlmlock);
429 if (dlmlock->l_granted_mode != dlmlock->l_req_mode) {
430 dlmlock->l_ast_data = NULL;
431 unlock_res_and_lock(dlmlock);
435 discard = ldlm_is_discard_data(dlmlock);
436 if (dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))
439 if (dlmlock->l_ast_data) {
440 obj = osc2cl(dlmlock->l_ast_data);
441 dlmlock->l_ast_data = NULL;
446 unlock_res_and_lock(dlmlock);
448 /* if l_ast_data is NULL, the dlmlock was enqueued by AGL or
449 * the object has been destroyed.
452 struct ldlm_extent *extent = &dlmlock->l_policy_data.l_extent;
453 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
456 /* Destroy pages covered by the extent of the DLM lock */
457 result = osc_lock_flush(cl2osc(obj),
458 cl_index(obj, extent->start),
459 cl_index(obj, extent->end),
462 /* losing a lock, update kms */
463 lock_res_and_lock(dlmlock);
464 cl_object_attr_lock(obj);
465 /* Must get the value under the lock to avoid race. */
466 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
467 /* Update the kms. Need to loop all granted locks.
468 * Not a problem for the client
470 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
472 cl_object_attr_update(env, obj, attr, CAT_KMS);
473 cl_object_attr_unlock(obj);
474 unlock_res_and_lock(dlmlock);
476 cl_object_put(env, obj);
482 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
483 * some other lock, or is canceled. This function is installed as a
484 * ldlm_lock::l_blocking_ast() for client extent locks.
486 * Control flow is tricky, because ldlm uses the same call-back
487 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
489 * \param dlmlock lock for which ast occurred.
491 * \param new description of a conflicting lock in case of blocking ast.
493 * \param data value of dlmlock->l_ast_data
495 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
496 * cancellation and blocking ast's.
498 * Possible use cases:
500 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
501 * lock due to lock lru pressure, or explicit user request to purge
504 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
505 * us that dlmlock conflicts with another lock that some client is
506 * enqueing. Lock is canceled.
508 * - cl_lock_cancel() is called. osc_lock_cancel() calls
509 * ldlm_cli_cancel() that calls
511 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
513 * recursively entering osc_ldlm_blocking_ast().
515 * - client cancels lock voluntary (e.g., as a part of early cancellation):
518 * osc_lock_cancel()->
519 * ldlm_cli_cancel()->
520 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
523 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
524 struct ldlm_lock_desc *new, void *data,
530 case LDLM_CB_BLOCKING: {
531 struct lustre_handle lockh;
533 ldlm_lock2handle(dlmlock, &lockh);
534 result = ldlm_cli_cancel(&lockh, LCF_ASYNC);
535 if (result == -ENODATA)
539 case LDLM_CB_CANCELING: {
544 * This can be called in the context of outer IO, e.g.,
546 * osc_enqueue_base()->...
547 * ->ldlm_prep_elc_req()->...
548 * ->ldlm_cancel_callback()->...
549 * ->osc_ldlm_blocking_ast()
551 * new environment has to be created to not corrupt outer
554 env = cl_env_get(&refcheck);
556 result = PTR_ERR(env);
560 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
561 cl_env_put(env, &refcheck);
570 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
572 struct ptlrpc_request *req = data;
575 struct req_capsule *cap;
576 struct cl_object *obj = NULL;
580 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
582 env = cl_env_get(&refcheck);
584 result = PTR_ERR(env);
588 lock_res_and_lock(dlmlock);
589 if (dlmlock->l_ast_data) {
590 obj = osc2cl(dlmlock->l_ast_data);
593 unlock_res_and_lock(dlmlock);
596 /* Do not grab the mutex of cl_lock for glimpse.
597 * See LU-1274 for details.
598 * BTW, it's okay for cl_lock to be cancelled during
599 * this period because server can handle this race.
600 * See ldlm_server_glimpse_ast() for details.
601 * cl_lock_mutex_get(env, lock);
604 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
605 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
607 result = req_capsule_server_pack(cap);
609 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
610 result = cl_object_glimpse(env, obj, lvb);
612 if (!exp_connect_lvb_type(req->rq_export)) {
613 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
614 sizeof(struct ost_lvb_v1),
617 cl_object_put(env, obj);
620 * These errors are normal races, so we don't want to
621 * fill the console with messages by calling
624 lustre_pack_reply(req, 1, NULL, NULL);
625 result = -ELDLM_NO_LOCK_DATA;
627 cl_env_put(env, &refcheck);
630 req->rq_status = result;
634 static int weigh_cb(const struct lu_env *env, struct cl_io *io,
635 struct osc_page *ops, void *cbdata)
637 struct cl_page *page = ops->ops_cl.cpl_page;
639 if (cl_page_is_vmlocked(env, page) ||
640 PageDirty(page->cp_vmpage) || PageWriteback(page->cp_vmpage)
642 return CLP_GANG_ABORT;
644 *(pgoff_t *)cbdata = osc_index(ops) + 1;
645 return CLP_GANG_OKAY;
648 static unsigned long osc_lock_weight(const struct lu_env *env,
649 struct osc_object *oscobj,
650 struct ldlm_extent *extent)
652 struct cl_io *io = &osc_env_info(env)->oti_io;
653 struct cl_object *obj = cl_object_top(&oscobj->oo_cl);
658 io->ci_ignore_layout = 1;
659 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
663 page_index = cl_index(obj, extent->start);
665 result = osc_page_gang_lookup(env, io, oscobj,
667 cl_index(obj, extent->end),
668 weigh_cb, (void *)&page_index);
669 if (result == CLP_GANG_ABORT)
671 if (result == CLP_GANG_RESCHED)
673 } while (result != CLP_GANG_OKAY);
676 return result == CLP_GANG_ABORT ? 1 : 0;
680 * Get the weight of dlm lock for early cancellation.
682 unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
685 struct osc_object *obj;
686 struct osc_lock *oscl;
687 unsigned long weight;
693 * osc_ldlm_weigh_ast has a complex context since it might be called
694 * because of lock canceling, or from user's input. We have to make
695 * a new environment for it. Probably it is implementation safe to use
696 * the upper context because cl_lock_put don't modify environment
697 * variables. But just in case ..
699 env = cl_env_get(&refcheck);
701 /* Mostly because lack of memory, do not eliminate this lock */
704 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
705 obj = dlmlock->l_ast_data;
711 spin_lock(&obj->oo_ol_spin);
712 list_for_each_entry(oscl, &obj->oo_ol_list, ols_nextlock_oscobj) {
713 if (oscl->ols_dlmlock && oscl->ols_dlmlock != dlmlock)
717 spin_unlock(&obj->oo_ol_spin);
720 * If the lock is being used by an IO, definitely not cancel it.
726 weight = osc_lock_weight(env, obj, &dlmlock->l_policy_data.l_extent);
729 cl_env_put(env, &refcheck);
733 static void osc_lock_build_einfo(const struct lu_env *env,
734 const struct cl_lock *lock,
735 struct osc_object *osc,
736 struct ldlm_enqueue_info *einfo)
738 einfo->ei_type = LDLM_EXTENT;
739 einfo->ei_mode = osc_cl_lock2ldlm(lock->cll_descr.cld_mode);
740 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
741 einfo->ei_cb_cp = ldlm_completion_ast;
742 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
743 einfo->ei_cbdata = osc; /* value to be put into ->l_ast_data */
747 * Determine if the lock should be converted into a lockless lock.
750 * - if the lock has an explicit requirement for a non-lockless lock;
751 * - if the io lock request type ci_lockreq;
752 * - send the enqueue rpc to ost to make the further decision;
753 * - special treat to truncate lockless lock
755 * Additional policy can be implemented here, e.g., never do lockless-io
758 static void osc_lock_to_lockless(const struct lu_env *env,
759 struct osc_lock *ols, int force)
761 struct cl_lock_slice *slice = &ols->ols_cl;
763 LASSERT(ols->ols_state == OLS_NEW ||
764 ols->ols_state == OLS_UPCALL_RECEIVED);
767 ols->ols_locklessable = 1;
768 slice->cls_ops = &osc_lock_lockless_ops;
770 struct osc_io *oio = osc_env_io(env);
771 struct cl_io *io = oio->oi_cl.cis_io;
772 struct cl_object *obj = slice->cls_obj;
773 struct osc_object *oob = cl2osc(obj);
774 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
775 struct obd_connect_data *ocd;
777 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
778 io->ci_lockreq == CILR_MAYBE ||
779 io->ci_lockreq == CILR_NEVER);
781 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
782 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
783 (io->ci_lockreq == CILR_MAYBE) &&
784 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
785 if (io->ci_lockreq == CILR_NEVER ||
787 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
788 /* lockless truncate */
789 (cl_io_is_trunc(io) &&
790 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
791 osd->od_lockless_truncate)) {
792 ols->ols_locklessable = 1;
793 slice->cls_ops = &osc_lock_lockless_ops;
796 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
799 static bool osc_lock_compatible(const struct osc_lock *qing,
800 const struct osc_lock *qed)
802 struct cl_lock_descr *qed_descr = &qed->ols_cl.cls_lock->cll_descr;
803 struct cl_lock_descr *qing_descr = &qing->ols_cl.cls_lock->cll_descr;
805 if (qed->ols_glimpse)
808 if (qing_descr->cld_mode == CLM_READ && qed_descr->cld_mode == CLM_READ)
811 if (qed->ols_state < OLS_GRANTED)
814 if (qed_descr->cld_mode >= qing_descr->cld_mode &&
815 qed_descr->cld_start <= qing_descr->cld_start &&
816 qed_descr->cld_end >= qing_descr->cld_end)
822 static void osc_lock_wake_waiters(const struct lu_env *env,
823 struct osc_object *osc,
824 struct osc_lock *oscl)
826 spin_lock(&osc->oo_ol_spin);
827 list_del_init(&oscl->ols_nextlock_oscobj);
828 spin_unlock(&osc->oo_ol_spin);
830 spin_lock(&oscl->ols_lock);
831 while (!list_empty(&oscl->ols_waiting_list)) {
832 struct osc_lock *scan;
834 scan = list_entry(oscl->ols_waiting_list.next, struct osc_lock,
836 list_del_init(&scan->ols_wait_entry);
838 cl_sync_io_note(env, scan->ols_owner, 0);
840 spin_unlock(&oscl->ols_lock);
843 static int osc_lock_enqueue_wait(const struct lu_env *env,
844 struct osc_object *obj,
845 struct osc_lock *oscl)
847 struct osc_lock *tmp_oscl;
848 struct cl_lock_descr *need = &oscl->ols_cl.cls_lock->cll_descr;
849 struct cl_sync_io *waiter = &osc_env_info(env)->oti_anchor;
852 spin_lock(&obj->oo_ol_spin);
853 list_add_tail(&oscl->ols_nextlock_oscobj, &obj->oo_ol_list);
856 list_for_each_entry(tmp_oscl, &obj->oo_ol_list,
857 ols_nextlock_oscobj) {
858 struct cl_lock_descr *descr;
860 if (tmp_oscl == oscl)
863 descr = &tmp_oscl->ols_cl.cls_lock->cll_descr;
864 if (descr->cld_start > need->cld_end ||
865 descr->cld_end < need->cld_start)
868 /* We're not supposed to give up group lock */
869 if (descr->cld_mode == CLM_GROUP)
872 if (!osc_lock_is_lockless(oscl) &&
873 osc_lock_compatible(oscl, tmp_oscl))
876 /* wait for conflicting lock to be canceled */
877 cl_sync_io_init(waiter, 1, cl_sync_io_end);
878 oscl->ols_owner = waiter;
880 spin_lock(&tmp_oscl->ols_lock);
881 /* add oscl into tmp's ols_waiting list */
882 list_add_tail(&oscl->ols_wait_entry,
883 &tmp_oscl->ols_waiting_list);
884 spin_unlock(&tmp_oscl->ols_lock);
886 spin_unlock(&obj->oo_ol_spin);
887 rc = cl_sync_io_wait(env, waiter, 0);
888 spin_lock(&obj->oo_ol_spin);
892 oscl->ols_owner = NULL;
895 spin_unlock(&obj->oo_ol_spin);
901 * Implementation of cl_lock_operations::clo_enqueue() method for osc
902 * layer. This initiates ldlm enqueue:
904 * - cancels conflicting locks early (osc_lock_enqueue_wait());
906 * - calls osc_enqueue_base() to do actual enqueue.
908 * osc_enqueue_base() is supplied with an upcall function that is executed
909 * when lock is received either after a local cached ldlm lock is matched, or
910 * when a reply from the server is received.
912 * This function does not wait for the network communication to complete.
914 static int osc_lock_enqueue(const struct lu_env *env,
915 const struct cl_lock_slice *slice,
916 struct cl_io *unused, struct cl_sync_io *anchor)
918 struct osc_thread_info *info = osc_env_info(env);
919 struct osc_io *oio = osc_env_io(env);
920 struct osc_object *osc = cl2osc(slice->cls_obj);
921 struct osc_lock *oscl = cl2osc_lock(slice);
922 struct cl_lock *lock = slice->cls_lock;
923 struct ldlm_res_id *resname = &info->oti_resname;
924 union ldlm_policy_data *policy = &info->oti_policy;
925 osc_enqueue_upcall_f upcall = osc_lock_upcall;
930 LASSERTF(ergo(oscl->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
931 "lock = %p, ols = %p\n", lock, oscl);
933 if (oscl->ols_state == OLS_GRANTED)
936 if (oscl->ols_flags & LDLM_FL_TEST_LOCK)
939 if (oscl->ols_glimpse) {
940 LASSERT(equi(oscl->ols_agl, !anchor));
945 result = osc_lock_enqueue_wait(env, osc, oscl);
949 /* 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;
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 AGL lock, async of osc_enqueue_base() must be true,
968 * DLM's enqueue callback set to osc_lock_upcall() with cookie as
971 ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
972 osc_lock_build_policy(env, lock, policy);
974 oscl->ols_einfo.ei_cbdata = NULL;
975 /* hold a reference for callback */
976 cl_object_get(osc2cl(osc));
977 upcall = osc_lock_upcall_agl;
980 result = osc_enqueue_base(osc_export(osc), resname, &oscl->ols_flags,
981 policy, &oscl->ols_lvb,
982 osc->oo_oinfo->loi_kms_valid,
984 &oscl->ols_einfo, PTLRPCD_SET, async,
987 if (osc_lock_is_lockless(oscl)) {
988 oio->oi_lockless = 1;
990 LASSERT(oscl->ols_state == OLS_GRANTED);
991 LASSERT(oscl->ols_hold);
992 LASSERT(oscl->ols_dlmlock);
994 } else if (oscl->ols_agl) {
995 cl_object_put(env, osc2cl(osc));
1001 oscl->ols_state = OLS_CANCELLED;
1002 osc_lock_wake_waiters(env, osc, oscl);
1005 cl_sync_io_note(env, anchor, result);
1011 * Breaks a link between osc_lock and dlm_lock.
1013 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
1015 struct ldlm_lock *dlmlock;
1017 dlmlock = olck->ols_dlmlock;
1021 if (olck->ols_hold) {
1023 ldlm_lock_decref(&olck->ols_handle, olck->ols_einfo.ei_mode);
1024 olck->ols_handle.cookie = 0ULL;
1027 olck->ols_dlmlock = NULL;
1029 /* release a reference taken in osc_lock_upcall(). */
1030 LASSERT(olck->ols_has_ref);
1031 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
1032 LDLM_LOCK_RELEASE(dlmlock);
1033 olck->ols_has_ref = 0;
1037 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1038 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1039 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1040 * with some other lock some where in the cluster. This function does the
1043 * - invalidates all pages protected by this lock (after sending dirty
1044 * ones to the server, as necessary);
1046 * - decref's underlying ldlm lock;
1048 * - cancels ldlm lock (ldlm_cli_cancel()).
1050 static void osc_lock_cancel(const struct lu_env *env,
1051 const struct cl_lock_slice *slice)
1053 struct osc_object *obj = cl2osc(slice->cls_obj);
1054 struct osc_lock *oscl = cl2osc_lock(slice);
1056 LINVRNT(osc_lock_invariant(oscl));
1058 osc_lock_detach(env, oscl);
1059 oscl->ols_state = OLS_CANCELLED;
1060 oscl->ols_flags &= ~LDLM_FL_LVB_READY;
1062 osc_lock_wake_waiters(env, obj, oscl);
1065 static int osc_lock_print(const struct lu_env *env, void *cookie,
1066 lu_printer_t p, const struct cl_lock_slice *slice)
1068 struct osc_lock *lock = cl2osc_lock(slice);
1070 (*p)(env, cookie, "%p %#16llx %#llx %d %p ",
1071 lock->ols_dlmlock, lock->ols_flags, lock->ols_handle.cookie,
1072 lock->ols_state, lock->ols_owner);
1073 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1077 static const struct cl_lock_operations osc_lock_ops = {
1078 .clo_fini = osc_lock_fini,
1079 .clo_enqueue = osc_lock_enqueue,
1080 .clo_cancel = osc_lock_cancel,
1081 .clo_print = osc_lock_print,
1084 static void osc_lock_lockless_cancel(const struct lu_env *env,
1085 const struct cl_lock_slice *slice)
1087 struct osc_lock *ols = cl2osc_lock(slice);
1088 struct osc_object *osc = cl2osc(slice->cls_obj);
1089 struct cl_lock_descr *descr = &slice->cls_lock->cll_descr;
1092 LASSERT(!ols->ols_dlmlock);
1093 result = osc_lock_flush(osc, descr->cld_start, descr->cld_end,
1094 descr->cld_mode, 0);
1096 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1099 osc_lock_wake_waiters(env, osc, ols);
1102 static const struct cl_lock_operations osc_lock_lockless_ops = {
1103 .clo_fini = osc_lock_fini,
1104 .clo_enqueue = osc_lock_enqueue,
1105 .clo_cancel = osc_lock_lockless_cancel,
1106 .clo_print = osc_lock_print
1109 static void osc_lock_set_writer(const struct lu_env *env,
1110 const struct cl_io *io,
1111 struct cl_object *obj, struct osc_lock *oscl)
1113 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
1117 if (!cl_object_same(io->ci_obj, obj))
1120 if (likely(io->ci_type == CIT_WRITE)) {
1121 io_start = cl_index(obj, io->u.ci_rw.crw_pos);
1122 io_end = cl_index(obj, io->u.ci_rw.crw_pos +
1123 io->u.ci_rw.crw_count - 1);
1124 if (cl_io_is_append(io)) {
1126 io_end = CL_PAGE_EOF;
1129 LASSERT(cl_io_is_mkwrite(io));
1130 io_start = io->u.ci_fault.ft_index;
1131 io_end = io->u.ci_fault.ft_index;
1134 if (descr->cld_mode >= CLM_WRITE &&
1135 descr->cld_start <= io_start && descr->cld_end >= io_end) {
1136 struct osc_io *oio = osc_env_io(env);
1138 /* There must be only one lock to match the write region */
1139 LASSERT(!oio->oi_write_osclock);
1140 oio->oi_write_osclock = oscl;
1144 int osc_lock_init(const struct lu_env *env,
1145 struct cl_object *obj, struct cl_lock *lock,
1146 const struct cl_io *io)
1148 struct osc_lock *oscl;
1149 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1151 oscl = kmem_cache_zalloc(osc_lock_kmem, GFP_NOFS);
1155 oscl->ols_state = OLS_NEW;
1156 spin_lock_init(&oscl->ols_lock);
1157 INIT_LIST_HEAD(&oscl->ols_waiting_list);
1158 INIT_LIST_HEAD(&oscl->ols_wait_entry);
1159 INIT_LIST_HEAD(&oscl->ols_nextlock_oscobj);
1161 oscl->ols_flags = osc_enq2ldlm_flags(enqflags);
1162 oscl->ols_agl = !!(enqflags & CEF_AGL);
1164 oscl->ols_flags |= LDLM_FL_BLOCK_NOWAIT;
1165 if (oscl->ols_flags & LDLM_FL_HAS_INTENT) {
1166 oscl->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1167 oscl->ols_glimpse = 1;
1169 osc_lock_build_einfo(env, lock, cl2osc(obj), &oscl->ols_einfo);
1171 cl_lock_slice_add(lock, &oscl->ols_cl, obj, &osc_lock_ops);
1173 if (!(enqflags & CEF_MUST))
1174 /* try to convert this lock to a lockless lock */
1175 osc_lock_to_lockless(env, oscl, (enqflags & CEF_NEVER));
1176 if (oscl->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
1177 oscl->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1179 if (io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io))
1180 osc_lock_set_writer(env, io, obj, oscl);
1183 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %llx",
1184 lock, oscl, oscl->ols_flags);
1190 * Finds an existing lock covering given index and optionally different from a
1191 * given \a except lock.
1193 struct ldlm_lock *osc_dlmlock_at_pgoff(const struct lu_env *env,
1194 struct osc_object *obj, pgoff_t index,
1195 enum osc_dap_flags dap_flags)
1197 struct osc_thread_info *info = osc_env_info(env);
1198 struct ldlm_res_id *resname = &info->oti_resname;
1199 union ldlm_policy_data *policy = &info->oti_policy;
1200 struct lustre_handle lockh;
1201 struct ldlm_lock *lock = NULL;
1202 enum ldlm_mode mode;
1205 ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1206 osc_index2policy(policy, osc2cl(obj), index, index);
1207 policy->l_extent.gid = LDLM_GID_ANY;
1209 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1210 if (dap_flags & OSC_DAP_FL_TEST_LOCK)
1211 flags |= LDLM_FL_TEST_LOCK;
1214 * It is fine to match any group lock since there could be only one
1215 * with a uniq gid and it conflicts with all other lock modes too
1218 mode = osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
1219 LCK_PR | LCK_PW | LCK_GROUP, &flags, obj, &lockh,
1220 dap_flags & OSC_DAP_FL_CANCELING);
1222 lock = ldlm_handle2lock(&lockh);
1223 /* RACE: the lock is cancelled so let's try again */
1224 if (unlikely(!lock))