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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_lock for OSC layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
39 * Author: Jinshan Xiong <jinshan.xiong@intel.com>
42 #define DEBUG_SUBSYSTEM S_OSC
44 #include <libcfs/libcfs.h>
45 /* fid_build_reg_res_name() */
46 #include <lustre_fid.h>
48 #include "osc_cl_internal.h"
54 /*****************************************************************************
60 static const struct cl_lock_operations osc_lock_ops;
61 static const struct cl_lock_operations osc_lock_lockless_ops;
62 static void osc_lock_to_lockless(const struct lu_env *env,
63 struct osc_lock *ols, int force);
65 int osc_lock_is_lockless(const struct osc_lock *olck)
67 return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
71 * Returns a weak pointer to the ldlm lock identified by a handle. Returned
72 * pointer cannot be dereferenced, as lock is not protected from concurrent
73 * reclaim. This function is a helper for osc_lock_invariant().
75 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
77 struct ldlm_lock *lock;
79 lock = ldlm_handle2lock(handle);
86 * Invariant that has to be true all of the time.
88 static int osc_lock_invariant(struct osc_lock *ols)
90 struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle);
91 struct ldlm_lock *olock = ols->ols_dlmlock;
92 int handle_used = lustre_handle_is_used(&ols->ols_handle);
94 if (ergo(osc_lock_is_lockless(ols),
95 ols->ols_locklessable && ols->ols_dlmlock == NULL))
99 * If all the following "ergo"s are true, return 1, otherwise 0
101 if (! ergo(olock != NULL, handle_used))
104 if (! ergo(olock != NULL,
105 olock->l_handle.h_cookie == ols->ols_handle.cookie))
108 if (! ergo(handle_used,
109 ergo(lock != NULL && olock != NULL, lock == olock) &&
110 ergo(lock == NULL, olock == NULL)))
113 * Check that ->ols_handle and ->ols_dlmlock are consistent, but
114 * take into account that they are set at the different time.
116 if (! ergo(ols->ols_state == OLS_CANCELLED,
117 olock == NULL && !handle_used))
120 * DLM lock is destroyed only after we have seen cancellation
123 if (! ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
124 !ldlm_is_destroyed(olock)))
127 if (! ergo(ols->ols_state == OLS_GRANTED,
129 olock->l_req_mode == olock->l_granted_mode &&
135 /*****************************************************************************
141 static void osc_lock_fini(const struct lu_env *env,
142 struct cl_lock_slice *slice)
144 struct osc_lock *ols = cl2osc_lock(slice);
146 LINVRNT(osc_lock_invariant(ols));
147 LASSERT(ols->ols_dlmlock == NULL);
149 OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
152 static void osc_lock_build_policy(const struct lu_env *env,
153 const struct cl_lock *lock,
154 union ldlm_policy_data *policy)
156 const struct cl_lock_descr *d = &lock->cll_descr;
158 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
159 policy->l_extent.gid = d->cld_gid;
162 static __u64 osc_enq2ldlm_flags(__u32 enqflags)
166 LASSERT((enqflags & ~CEF_MASK) == 0);
168 if (enqflags & CEF_NONBLOCK)
169 result |= LDLM_FL_BLOCK_NOWAIT;
170 if (enqflags & CEF_ASYNC)
171 result |= LDLM_FL_HAS_INTENT;
172 if (enqflags & CEF_DISCARD_DATA)
173 result |= LDLM_FL_AST_DISCARD_DATA;
174 if (enqflags & CEF_PEEK)
175 result |= LDLM_FL_TEST_LOCK;
176 if (enqflags & CEF_LOCK_MATCH)
177 result |= LDLM_FL_MATCH_LOCK;
182 * Updates object attributes from a lock value block (lvb) received together
183 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
186 * This can be optimized to not update attributes when lock is a result of a
189 * Called under lock and resource spin-locks.
191 static void osc_lock_lvb_update(const struct lu_env *env,
192 struct osc_object *osc,
193 struct ldlm_lock *dlmlock,
196 struct cl_object *obj = osc2cl(osc);
197 struct lov_oinfo *oinfo = osc->oo_oinfo;
198 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
203 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
205 LASSERT(dlmlock != NULL);
206 lvb = dlmlock->l_lvb_data;
208 cl_lvb2attr(attr, lvb);
210 cl_object_attr_lock(obj);
211 if (dlmlock != NULL) {
214 check_res_locked(dlmlock->l_resource);
216 LASSERT(lvb == dlmlock->l_lvb_data);
217 size = lvb->lvb_size;
219 /* Extend KMS up to the end of this lock and no further
220 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
221 if (size > dlmlock->l_policy_data.l_extent.end)
222 size = dlmlock->l_policy_data.l_extent.end + 1;
223 if (size >= oinfo->loi_kms) {
224 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="LPU64
225 ", kms="LPU64, lvb->lvb_size, size);
227 attr->cat_kms = size;
229 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
230 LPU64"; leaving kms="LPU64", end="LPU64,
231 lvb->lvb_size, oinfo->loi_kms,
232 dlmlock->l_policy_data.l_extent.end);
234 ldlm_lock_allow_match_locked(dlmlock);
237 cl_object_attr_update(env, obj, attr, valid);
238 cl_object_attr_unlock(obj);
243 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *oscl,
244 struct lustre_handle *lockh, bool lvb_update)
246 struct ldlm_lock *dlmlock;
248 dlmlock = ldlm_handle2lock_long(lockh, 0);
249 LASSERT(dlmlock != NULL);
251 /* lock reference taken by ldlm_handle2lock_long() is
252 * owned by osc_lock and released in osc_lock_detach()
254 lu_ref_add(&dlmlock->l_reference, "osc_lock", oscl);
255 oscl->ols_has_ref = 1;
257 LASSERT(oscl->ols_dlmlock == NULL);
258 oscl->ols_dlmlock = dlmlock;
260 /* This may be a matched lock for glimpse request, do not hold
261 * lock reference in that case. */
262 if (!oscl->ols_glimpse) {
263 /* hold a refc for non glimpse lock which will
264 * be released in osc_lock_cancel() */
265 lustre_handle_copy(&oscl->ols_handle, lockh);
266 ldlm_lock_addref(lockh, oscl->ols_einfo.ei_mode);
270 /* Lock must have been granted. */
271 lock_res_and_lock(dlmlock);
272 if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
273 struct ldlm_extent *ext = &dlmlock->l_policy_data.l_extent;
274 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
276 /* extend the lock extent, otherwise it will have problem when
277 * we decide whether to grant a lockless lock. */
278 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
279 descr->cld_start = cl_index(descr->cld_obj, ext->start);
280 descr->cld_end = cl_index(descr->cld_obj, ext->end);
281 descr->cld_gid = ext->gid;
283 /* no lvb update for matched lock */
285 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
286 osc_lock_lvb_update(env, cl2osc(oscl->ols_cl.cls_obj),
289 LINVRNT(osc_lock_invariant(oscl));
291 unlock_res_and_lock(dlmlock);
293 LASSERT(oscl->ols_state != OLS_GRANTED);
294 oscl->ols_state = OLS_GRANTED;
298 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
299 * received from a server, or after osc_enqueue_base() matched a local DLM
302 static int osc_lock_upcall(void *cookie, struct lustre_handle *lockh,
305 struct osc_lock *oscl = cookie;
306 struct cl_lock_slice *slice = &oscl->ols_cl;
308 struct cl_env_nest nest;
313 env = cl_env_nested_get(&nest);
314 /* should never happen, similar to osc_ldlm_blocking_ast(). */
315 LASSERT(!IS_ERR(env));
317 rc = ldlm_error2errno(errcode);
318 if (oscl->ols_state == OLS_ENQUEUED) {
319 oscl->ols_state = OLS_UPCALL_RECEIVED;
320 } else if (oscl->ols_state == OLS_CANCELLED) {
323 CERROR("Impossible state: %d\n", oscl->ols_state);
328 osc_lock_granted(env, oscl, lockh, errcode == ELDLM_OK);
330 /* Error handling, some errors are tolerable. */
331 if (oscl->ols_locklessable && rc == -EUSERS) {
332 /* This is a tolerable error, turn this lock into
335 osc_object_set_contended(cl2osc(slice->cls_obj));
336 LASSERT(slice->cls_ops == &osc_lock_ops);
338 /* Change this lock to ldlmlock-less lock. */
339 osc_lock_to_lockless(env, oscl, 1);
340 oscl->ols_state = OLS_GRANTED;
342 } else if (oscl->ols_glimpse && rc == -ENAVAIL) {
343 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
344 osc_lock_lvb_update(env, cl2osc(slice->cls_obj),
345 NULL, &oscl->ols_lvb);
346 /* Hide the error. */
350 if (oscl->ols_owner != NULL)
351 cl_sync_io_note(env, oscl->ols_owner, rc);
352 cl_env_nested_put(&nest, env);
357 static int osc_lock_upcall_agl(void *cookie, struct lustre_handle *lockh,
360 struct osc_object *osc = cookie;
361 struct ldlm_lock *dlmlock;
363 struct cl_env_nest nest;
366 env = cl_env_nested_get(&nest);
367 LASSERT(!IS_ERR(env));
369 if (errcode == ELDLM_LOCK_MATCHED)
370 GOTO(out, errcode = ELDLM_OK);
372 if (errcode != ELDLM_OK)
375 dlmlock = ldlm_handle2lock(lockh);
376 LASSERT(dlmlock != NULL);
378 lock_res_and_lock(dlmlock);
379 LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
381 /* there is no osc_lock associated with AGL lock */
382 osc_lock_lvb_update(env, osc, dlmlock, NULL);
384 unlock_res_and_lock(dlmlock);
385 LDLM_LOCK_PUT(dlmlock);
388 cl_object_put(env, osc2cl(osc));
389 cl_env_nested_put(&nest, env);
390 RETURN(ldlm_error2errno(errcode));
393 static int osc_lock_flush(struct osc_object *obj, pgoff_t start, pgoff_t end,
394 enum cl_lock_mode mode, int discard)
397 struct cl_env_nest nest;
403 env = cl_env_nested_get(&nest);
405 RETURN(PTR_ERR(env));
407 if (mode == CLM_WRITE) {
408 rc = osc_cache_writeback_range(env, obj, start, end, 1,
410 CDEBUG(D_CACHE, "object %p: [%lu -> %lu] %d pages were %s.\n",
412 discard ? "discarded" : "written back");
417 rc2 = osc_lock_discard_pages(env, obj, start, end, mode);
418 if (rc == 0 && rc2 < 0)
421 cl_env_nested_put(&nest, env);
426 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
427 * and ldlm_lock caches.
429 static int osc_dlm_blocking_ast0(const struct lu_env *env,
430 struct ldlm_lock *dlmlock,
431 void *data, int flag)
433 struct cl_object *obj = NULL;
436 enum cl_lock_mode mode = CLM_READ;
439 LASSERT(flag == LDLM_CB_CANCELING);
441 lock_res_and_lock(dlmlock);
442 if (dlmlock->l_granted_mode != dlmlock->l_req_mode) {
443 dlmlock->l_ast_data = NULL;
444 unlock_res_and_lock(dlmlock);
448 discard = ldlm_is_discard_data(dlmlock);
449 if (dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))
452 if (dlmlock->l_ast_data != NULL) {
453 obj = osc2cl(dlmlock->l_ast_data);
454 dlmlock->l_ast_data = NULL;
459 unlock_res_and_lock(dlmlock);
461 /* if l_ast_data is NULL, the dlmlock was enqueued by AGL or
462 * the object has been destroyed. */
464 struct ldlm_extent *extent = &dlmlock->l_policy_data.l_extent;
465 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
468 /* Destroy pages covered by the extent of the DLM lock */
469 result = osc_lock_flush(cl2osc(obj),
470 cl_index(obj, extent->start),
471 cl_index(obj, extent->end),
474 /* losing a lock, update kms */
475 lock_res_and_lock(dlmlock);
476 cl_object_attr_lock(obj);
477 /* Must get the value under the lock to avoid race. */
478 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
479 /* Update the kms. Need to loop all granted locks.
480 * Not a problem for the client */
481 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
483 cl_object_attr_update(env, obj, attr, CAT_KMS);
484 cl_object_attr_unlock(obj);
485 unlock_res_and_lock(dlmlock);
487 cl_object_put(env, obj);
493 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
494 * some other lock, or is canceled. This function is installed as a
495 * ldlm_lock::l_blocking_ast() for client extent locks.
497 * Control flow is tricky, because ldlm uses the same call-back
498 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
500 * \param dlmlock lock for which ast occurred.
502 * \param new description of a conflicting lock in case of blocking ast.
504 * \param data value of dlmlock->l_ast_data
506 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
507 * cancellation and blocking ast's.
509 * Possible use cases:
511 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
512 * lock due to lock lru pressure, or explicit user request to purge
515 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
516 * us that dlmlock conflicts with another lock that some client is
517 * enqueuing. Lock is canceled.
519 * - cl_lock_cancel() is called. osc_lock_cancel() calls
520 * ldlm_cli_cancel() that calls
522 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
524 * recursively entering osc_ldlm_blocking_ast().
526 * - client cancels lock voluntary (e.g., as a part of early cancellation):
529 * osc_lock_cancel()->
530 * ldlm_cli_cancel()->
531 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
534 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
535 struct ldlm_lock_desc *new, void *data,
542 case LDLM_CB_BLOCKING: {
543 struct lustre_handle lockh;
545 ldlm_lock2handle(dlmlock, &lockh);
546 result = ldlm_cli_cancel(&lockh, LCF_ASYNC);
547 if (result == -ENODATA)
551 case LDLM_CB_CANCELING: {
553 struct cl_env_nest nest;
556 * This can be called in the context of outer IO, e.g.,
558 * osc_enqueue_base()->...
559 * ->ldlm_prep_elc_req()->...
560 * ->ldlm_cancel_callback()->...
561 * ->osc_ldlm_blocking_ast()
563 * new environment has to be created to not corrupt outer
566 env = cl_env_nested_get(&nest);
568 result = PTR_ERR(env);
572 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
573 cl_env_nested_put(&nest, env);
582 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
584 struct ptlrpc_request *req = data;
585 struct cl_env_nest nest;
588 struct req_capsule *cap;
593 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
595 env = cl_env_nested_get(&nest);
597 struct cl_object *obj = NULL;
599 lock_res_and_lock(dlmlock);
600 if (dlmlock->l_ast_data != NULL) {
601 obj = osc2cl(dlmlock->l_ast_data);
604 unlock_res_and_lock(dlmlock);
607 /* Do not grab the mutex of cl_lock for glimpse.
608 * See LU-1274 for details.
609 * BTW, it's okay for cl_lock to be cancelled during
610 * this period because server can handle this race.
611 * See ldlm_server_glimpse_ast() for details.
612 * cl_lock_mutex_get(env, lock); */
614 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
615 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
617 result = req_capsule_server_pack(cap);
619 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
620 result = cl_object_glimpse(env, obj, lvb);
622 if (!exp_connect_lvb_type(req->rq_export))
623 req_capsule_shrink(&req->rq_pill,
625 sizeof(struct ost_lvb_v1),
627 cl_object_put(env, obj);
630 * These errors are normal races, so we don't want to
631 * fill the console with messages by calling
634 lustre_pack_reply(req, 1, NULL, NULL);
635 result = -ELDLM_NO_LOCK_DATA;
637 cl_env_nested_put(&nest, env);
639 result = PTR_ERR(env);
640 req->rq_status = result;
644 static int weigh_cb(const struct lu_env *env, struct cl_io *io,
645 struct osc_page *ops, void *cbdata)
647 struct cl_page *page = ops->ops_cl.cpl_page;
649 if (cl_page_is_vmlocked(env, page)
650 || PageDirty(page->cp_vmpage) || PageWriteback(page->cp_vmpage)
652 return CLP_GANG_ABORT;
654 *(pgoff_t *)cbdata = osc_index(ops) + 1;
655 return CLP_GANG_OKAY;
658 static unsigned long osc_lock_weight(const struct lu_env *env,
659 struct osc_object *oscobj,
660 struct ldlm_extent *extent)
662 struct cl_io *io = &osc_env_info(env)->oti_io;
663 struct cl_object *obj = cl_object_top(&oscobj->oo_cl);
669 io->ci_ignore_layout = 1;
670 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
674 page_index = cl_index(obj, extent->start);
676 result = osc_page_gang_lookup(env, io, oscobj,
678 cl_index(obj, extent->end),
679 weigh_cb, (void *)&page_index);
680 if (result == CLP_GANG_ABORT)
682 if (result == CLP_GANG_RESCHED)
684 } while (result != CLP_GANG_OKAY);
687 return result == CLP_GANG_ABORT ? 1 : 0;
691 * Get the weight of dlm lock for early cancellation.
693 unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
695 struct cl_env_nest nest;
697 struct osc_object *obj;
698 struct osc_lock *oscl;
699 unsigned long weight;
705 * osc_ldlm_weigh_ast has a complex context since it might be called
706 * because of lock canceling, or from user's input. We have to make
707 * a new environment for it. Probably it is implementation safe to use
708 * the upper context because cl_lock_put don't modify environment
709 * variables. But just in case ..
711 env = cl_env_nested_get(&nest);
713 /* Mostly because lack of memory, do not eliminate this lock */
716 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
717 obj = dlmlock->l_ast_data;
719 GOTO(out, weight = 1);
721 spin_lock(&obj->oo_ol_spin);
722 list_for_each_entry(oscl, &obj->oo_ol_list, ols_nextlock_oscobj) {
723 if (oscl->ols_dlmlock != NULL && oscl->ols_dlmlock != dlmlock)
727 spin_unlock(&obj->oo_ol_spin);
730 * If the lock is being used by an IO, definitely not cancel it.
732 GOTO(out, weight = 1);
735 weight = osc_lock_weight(env, obj, &dlmlock->l_policy_data.l_extent);
739 cl_env_nested_put(&nest, env);
743 static void osc_lock_build_einfo(const struct lu_env *env,
744 const struct cl_lock *lock,
745 struct osc_object *osc,
746 struct ldlm_enqueue_info *einfo)
748 einfo->ei_type = LDLM_EXTENT;
749 einfo->ei_mode = osc_cl_lock2ldlm(lock->cll_descr.cld_mode);
750 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
751 einfo->ei_cb_cp = ldlm_completion_ast;
752 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
753 einfo->ei_cbdata = osc; /* value to be put into ->l_ast_data */
757 * Determine if the lock should be converted into a lockless lock.
760 * - if the lock has an explicite requirment for a non-lockless lock;
761 * - if the io lock request type ci_lockreq;
762 * - send the enqueue rpc to ost to make the further decision;
763 * - special treat to truncate lockless lock
765 * Additional policy can be implemented here, e.g., never do lockless-io
768 static void osc_lock_to_lockless(const struct lu_env *env,
769 struct osc_lock *ols, int force)
771 struct cl_lock_slice *slice = &ols->ols_cl;
773 LASSERT(ols->ols_state == OLS_NEW ||
774 ols->ols_state == OLS_UPCALL_RECEIVED);
777 ols->ols_locklessable = 1;
778 slice->cls_ops = &osc_lock_lockless_ops;
780 struct osc_io *oio = osc_env_io(env);
781 struct cl_io *io = oio->oi_cl.cis_io;
782 struct cl_object *obj = slice->cls_obj;
783 struct osc_object *oob = cl2osc(obj);
784 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
785 struct obd_connect_data *ocd;
787 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
788 io->ci_lockreq == CILR_MAYBE ||
789 io->ci_lockreq == CILR_NEVER);
791 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
792 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
793 (io->ci_lockreq == CILR_MAYBE) &&
794 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
795 if (io->ci_lockreq == CILR_NEVER ||
797 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
798 /* lockless truncate */
799 (cl_io_is_trunc(io) &&
800 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
801 osd->od_lockless_truncate)) {
802 ols->ols_locklessable = 1;
803 slice->cls_ops = &osc_lock_lockless_ops;
806 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
809 static bool osc_lock_compatible(const struct osc_lock *qing,
810 const struct osc_lock *qed)
812 struct cl_lock_descr *qed_descr = &qed->ols_cl.cls_lock->cll_descr;
813 struct cl_lock_descr *qing_descr = &qing->ols_cl.cls_lock->cll_descr;
815 if (qed->ols_glimpse)
818 if (qing_descr->cld_mode == CLM_READ && qed_descr->cld_mode == CLM_READ)
821 if (qed->ols_state < OLS_GRANTED)
824 if (qed_descr->cld_mode >= qing_descr->cld_mode &&
825 qed_descr->cld_start <= qing_descr->cld_start &&
826 qed_descr->cld_end >= qing_descr->cld_end)
832 static void osc_lock_wake_waiters(const struct lu_env *env,
833 struct osc_object *osc,
834 struct osc_lock *oscl)
836 spin_lock(&osc->oo_ol_spin);
837 list_del_init(&oscl->ols_nextlock_oscobj);
838 spin_unlock(&osc->oo_ol_spin);
840 spin_lock(&oscl->ols_lock);
841 while (!list_empty(&oscl->ols_waiting_list)) {
842 struct osc_lock *scan;
844 scan = list_entry(oscl->ols_waiting_list.next, struct osc_lock,
846 list_del_init(&scan->ols_wait_entry);
848 cl_sync_io_note(env, scan->ols_owner, 0);
850 spin_unlock(&oscl->ols_lock);
853 static int osc_lock_enqueue_wait(const struct lu_env *env,
854 struct osc_object *obj, struct osc_lock *oscl)
856 struct osc_lock *tmp_oscl;
857 struct cl_lock_descr *need = &oscl->ols_cl.cls_lock->cll_descr;
858 struct cl_sync_io *waiter = &osc_env_info(env)->oti_anchor;
862 spin_lock(&obj->oo_ol_spin);
863 list_add_tail(&oscl->ols_nextlock_oscobj, &obj->oo_ol_list);
866 list_for_each_entry(tmp_oscl, &obj->oo_ol_list,
867 ols_nextlock_oscobj) {
868 struct cl_lock_descr *descr;
870 if (tmp_oscl == oscl)
873 descr = &tmp_oscl->ols_cl.cls_lock->cll_descr;
874 if (descr->cld_start > need->cld_end ||
875 descr->cld_end < need->cld_start)
878 /* We're not supposed to give up group lock */
879 if (descr->cld_mode == CLM_GROUP)
882 if (!osc_lock_is_lockless(oscl) &&
883 osc_lock_compatible(oscl, tmp_oscl))
886 /* wait for conflicting lock to be canceled */
887 cl_sync_io_init(waiter, 1, cl_sync_io_end);
888 oscl->ols_owner = waiter;
890 spin_lock(&tmp_oscl->ols_lock);
891 /* add oscl into tmp's ols_waiting list */
892 list_add_tail(&oscl->ols_wait_entry,
893 &tmp_oscl->ols_waiting_list);
894 spin_unlock(&tmp_oscl->ols_lock);
896 spin_unlock(&obj->oo_ol_spin);
897 rc = cl_sync_io_wait(env, waiter, 0);
898 spin_lock(&obj->oo_ol_spin);
903 oscl->ols_owner = NULL;
906 spin_unlock(&obj->oo_ol_spin);
912 * Implementation of cl_lock_operations::clo_enqueue() method for osc
913 * layer. This initiates ldlm enqueue:
915 * - cancels conflicting locks early (osc_lock_enqueue_wait());
917 * - calls osc_enqueue_base() to do actual enqueue.
919 * osc_enqueue_base() is supplied with an upcall function that is executed
920 * when lock is received either after a local cached ldlm lock is matched, or
921 * when a reply from the server is received.
923 * This function does not wait for the network communication to complete.
925 static int osc_lock_enqueue(const struct lu_env *env,
926 const struct cl_lock_slice *slice,
927 struct cl_io *unused, struct cl_sync_io *anchor)
929 struct osc_thread_info *info = osc_env_info(env);
930 struct osc_io *oio = osc_env_io(env);
931 struct osc_object *osc = cl2osc(slice->cls_obj);
932 struct osc_lock *oscl = cl2osc_lock(slice);
933 struct cl_lock *lock = slice->cls_lock;
934 struct ldlm_res_id *resname = &info->oti_resname;
935 union ldlm_policy_data *policy = &info->oti_policy;
936 osc_enqueue_upcall_f upcall = osc_lock_upcall;
943 LASSERTF(ergo(oscl->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
944 "lock = %p, ols = %p\n", lock, oscl);
946 if (oscl->ols_state == OLS_GRANTED)
949 if (oscl->ols_flags & LDLM_FL_TEST_LOCK)
950 GOTO(enqueue_base, 0);
952 if (oscl->ols_glimpse) {
953 LASSERT(equi(oscl->ols_agl, anchor == NULL));
955 GOTO(enqueue_base, 0);
958 result = osc_lock_enqueue_wait(env, osc, oscl);
962 /* we can grant lockless lock right after all conflicting locks
964 if (osc_lock_is_lockless(oscl)) {
965 oscl->ols_state = OLS_GRANTED;
966 oio->oi_lockless = 1;
971 oscl->ols_state = OLS_ENQUEUED;
972 if (anchor != NULL) {
973 atomic_inc(&anchor->csi_sync_nr);
974 oscl->ols_owner = anchor;
978 * DLM lock's ast data must be osc_object;
979 * if glimpse or AGL lock, async of osc_enqueue_base() must be true,
980 * DLM's enqueue callback set to osc_lock_upcall() with cookie as
983 ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
984 osc_lock_build_policy(env, lock, policy);
986 oscl->ols_einfo.ei_cbdata = NULL;
987 /* hold a reference for callback */
988 cl_object_get(osc2cl(osc));
989 upcall = osc_lock_upcall_agl;
992 result = osc_enqueue_base(osc_export(osc), resname, &oscl->ols_flags,
993 policy, &oscl->ols_lvb,
994 osc->oo_oinfo->loi_kms_valid,
996 &oscl->ols_einfo, PTLRPCD_SET, async,
999 if (osc_lock_is_lockless(oscl)) {
1000 oio->oi_lockless = 1;
1001 } else if (!async) {
1002 LASSERT(oscl->ols_state == OLS_GRANTED);
1003 LASSERT(oscl->ols_hold);
1004 LASSERT(oscl->ols_dlmlock != NULL);
1006 } else if (oscl->ols_agl) {
1007 cl_object_put(env, osc2cl(osc));
1013 oscl->ols_state = OLS_CANCELLED;
1014 osc_lock_wake_waiters(env, osc, oscl);
1017 cl_sync_io_note(env, anchor, result);
1023 * Breaks a link between osc_lock and dlm_lock.
1025 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
1027 struct ldlm_lock *dlmlock;
1031 dlmlock = olck->ols_dlmlock;
1032 if (dlmlock == NULL)
1035 if (olck->ols_hold) {
1037 ldlm_lock_decref(&olck->ols_handle, olck->ols_einfo.ei_mode);
1038 olck->ols_handle.cookie = 0ULL;
1041 olck->ols_dlmlock = NULL;
1043 /* release a reference taken in osc_lock_upcall(). */
1044 LASSERT(olck->ols_has_ref);
1045 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
1046 LDLM_LOCK_RELEASE(dlmlock);
1047 olck->ols_has_ref = 0;
1053 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1054 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1055 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1056 * with some other lock some where in the cluster. This function does the
1059 * - invalidates all pages protected by this lock (after sending dirty
1060 * ones to the server, as necessary);
1062 * - decref's underlying ldlm lock;
1064 * - cancels ldlm lock (ldlm_cli_cancel()).
1066 static void osc_lock_cancel(const struct lu_env *env,
1067 const struct cl_lock_slice *slice)
1069 struct osc_object *obj = cl2osc(slice->cls_obj);
1070 struct osc_lock *oscl = cl2osc_lock(slice);
1074 LINVRNT(osc_lock_invariant(oscl));
1076 osc_lock_detach(env, oscl);
1077 oscl->ols_state = OLS_CANCELLED;
1078 oscl->ols_flags &= ~LDLM_FL_LVB_READY;
1080 osc_lock_wake_waiters(env, obj, oscl);
1084 static int osc_lock_print(const struct lu_env *env, void *cookie,
1085 lu_printer_t p, const struct cl_lock_slice *slice)
1087 struct osc_lock *lock = cl2osc_lock(slice);
1089 (*p)(env, cookie, "%p "LPX64" "LPX64" %d %p ",
1090 lock->ols_dlmlock, lock->ols_flags, lock->ols_handle.cookie,
1091 lock->ols_state, lock->ols_owner);
1092 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1096 static const struct cl_lock_operations osc_lock_ops = {
1097 .clo_fini = osc_lock_fini,
1098 .clo_enqueue = osc_lock_enqueue,
1099 .clo_cancel = osc_lock_cancel,
1100 .clo_print = osc_lock_print,
1103 static void osc_lock_lockless_cancel(const struct lu_env *env,
1104 const struct cl_lock_slice *slice)
1106 struct osc_lock *ols = cl2osc_lock(slice);
1107 struct osc_object *osc = cl2osc(slice->cls_obj);
1108 struct cl_lock_descr *descr = &slice->cls_lock->cll_descr;
1111 LASSERT(ols->ols_dlmlock == NULL);
1112 result = osc_lock_flush(osc, descr->cld_start, descr->cld_end,
1113 descr->cld_mode, 0);
1115 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1118 osc_lock_wake_waiters(env, osc, ols);
1121 static const struct cl_lock_operations osc_lock_lockless_ops = {
1122 .clo_fini = osc_lock_fini,
1123 .clo_enqueue = osc_lock_enqueue,
1124 .clo_cancel = osc_lock_lockless_cancel,
1125 .clo_print = osc_lock_print
1128 static void osc_lock_set_writer(const struct lu_env *env,
1129 const struct cl_io *io,
1130 struct cl_object *obj, struct osc_lock *oscl)
1132 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
1136 if (!cl_object_same(io->ci_obj, obj))
1139 if (likely(io->ci_type == CIT_WRITE)) {
1140 io_start = cl_index(obj, io->u.ci_rw.crw_pos);
1141 io_end = cl_index(obj, io->u.ci_rw.crw_pos +
1142 io->u.ci_rw.crw_count - 1);
1143 if (cl_io_is_append(io)) {
1145 io_end = CL_PAGE_EOF;
1148 LASSERT(cl_io_is_mkwrite(io));
1149 io_start = io_end = io->u.ci_fault.ft_index;
1152 if (descr->cld_mode >= CLM_WRITE &&
1153 descr->cld_start <= io_start && descr->cld_end >= io_end) {
1154 struct osc_io *oio = osc_env_io(env);
1156 /* There must be only one lock to match the write region */
1157 LASSERT(oio->oi_write_osclock == NULL);
1158 oio->oi_write_osclock = oscl;
1162 int osc_lock_init(const struct lu_env *env,
1163 struct cl_object *obj, struct cl_lock *lock,
1164 const struct cl_io *io)
1166 struct osc_lock *oscl;
1167 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1169 OBD_SLAB_ALLOC_PTR_GFP(oscl, osc_lock_kmem, GFP_NOFS);
1173 oscl->ols_state = OLS_NEW;
1174 spin_lock_init(&oscl->ols_lock);
1175 INIT_LIST_HEAD(&oscl->ols_waiting_list);
1176 INIT_LIST_HEAD(&oscl->ols_wait_entry);
1177 INIT_LIST_HEAD(&oscl->ols_nextlock_oscobj);
1179 oscl->ols_flags = osc_enq2ldlm_flags(enqflags);
1180 oscl->ols_agl = !!(enqflags & CEF_AGL);
1182 oscl->ols_flags |= LDLM_FL_BLOCK_NOWAIT;
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 "LPX64,
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_dlmlock_at_pgoff(const struct lu_env *env,
1211 struct osc_object *obj, pgoff_t index,
1212 enum osc_dap_flags dap_flags)
1214 struct osc_thread_info *info = osc_env_info(env);
1215 struct ldlm_res_id *resname = &info->oti_resname;
1216 union ldlm_policy_data *policy = &info->oti_policy;
1217 struct lustre_handle lockh;
1218 struct ldlm_lock *lock = NULL;
1219 enum ldlm_mode mode;
1224 ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1225 osc_index2policy(policy, osc2cl(obj), index, index);
1226 policy->l_extent.gid = LDLM_GID_ANY;
1228 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1229 if (dap_flags & OSC_DAP_FL_TEST_LOCK)
1230 flags |= LDLM_FL_TEST_LOCK;
1232 * It is fine to match any group lock since there could be only one
1233 * with a uniq gid and it conflicts with all other lock modes too
1236 mode = osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
1237 LCK_PR | LCK_PW | LCK_GROUP, &flags, obj, &lockh,
1238 dap_flags & OSC_DAP_FL_CANCELING);
1240 lock = ldlm_handle2lock(&lockh);
1241 /* RACE: the lock is cancelled so let's try again */
1242 if (unlikely(lock == NULL))