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>
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 == NULL))
95 * If all the following "ergo"s are true, return 1, otherwise 0
97 if (! ergo(olock != NULL, handle_used))
100 if (! ergo(olock != NULL,
101 olock->l_handle.h_cookie == ols->ols_handle.cookie))
104 if (! ergo(handle_used,
105 ergo(lock != NULL && olock != NULL, lock == olock) &&
106 ergo(lock == NULL, olock == NULL)))
109 * Check that ->ols_handle and ->ols_dlmlock are consistent, but
110 * take into account that they are set at the different time.
112 if (! ergo(ols->ols_state == OLS_CANCELLED,
113 olock == NULL && !handle_used))
116 * DLM lock is destroyed only after we have seen cancellation
119 if (! ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
120 !ldlm_is_destroyed(olock)))
123 if (! ergo(ols->ols_state == OLS_GRANTED,
125 olock->l_req_mode == olock->l_granted_mode &&
131 /*****************************************************************************
137 static void osc_lock_fini(const struct lu_env *env,
138 struct cl_lock_slice *slice)
140 struct osc_lock *ols = cl2osc_lock(slice);
142 LINVRNT(osc_lock_invariant(ols));
143 LASSERT(ols->ols_dlmlock == NULL);
145 OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
148 static void osc_lock_build_policy(const struct lu_env *env,
149 const struct cl_lock *lock,
150 union ldlm_policy_data *policy)
152 const struct cl_lock_descr *d = &lock->cll_descr;
154 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
155 policy->l_extent.gid = d->cld_gid;
158 static __u64 osc_enq2ldlm_flags(__u32 enqflags)
162 LASSERT((enqflags & ~CEF_MASK) == 0);
164 if (enqflags & CEF_NONBLOCK)
165 result |= LDLM_FL_BLOCK_NOWAIT;
166 if (enqflags & CEF_ASYNC)
167 result |= LDLM_FL_HAS_INTENT;
168 if (enqflags & CEF_DISCARD_DATA)
169 result |= LDLM_FL_AST_DISCARD_DATA;
170 if (enqflags & CEF_PEEK)
171 result |= LDLM_FL_TEST_LOCK;
172 if (enqflags & CEF_LOCK_MATCH)
173 result |= LDLM_FL_MATCH_LOCK;
178 * Updates object attributes from a lock value block (lvb) received together
179 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
182 * This can be optimized to not update attributes when lock is a result of a
185 * Called under lock and resource spin-locks.
187 static void osc_lock_lvb_update(const struct lu_env *env,
188 struct osc_object *osc,
189 struct ldlm_lock *dlmlock,
192 struct cl_object *obj = osc2cl(osc);
193 struct lov_oinfo *oinfo = osc->oo_oinfo;
194 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
199 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
201 LASSERT(dlmlock != NULL);
202 lvb = dlmlock->l_lvb_data;
204 cl_lvb2attr(attr, lvb);
206 cl_object_attr_lock(obj);
207 if (dlmlock != NULL) {
210 check_res_locked(dlmlock->l_resource);
212 LASSERT(lvb == dlmlock->l_lvb_data);
213 size = lvb->lvb_size;
215 /* Extend KMS up to the end of this lock and no further
216 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
217 if (size > dlmlock->l_policy_data.l_extent.end)
218 size = dlmlock->l_policy_data.l_extent.end + 1;
219 if (size >= oinfo->loi_kms) {
220 LDLM_DEBUG(dlmlock, "lock acquired, setting rss=%llu"
221 ", kms=%llu", lvb->lvb_size, size);
223 attr->cat_kms = size;
225 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
226 "%llu; leaving kms=%llu, end=%llu",
227 lvb->lvb_size, oinfo->loi_kms,
228 dlmlock->l_policy_data.l_extent.end);
230 ldlm_lock_allow_match_locked(dlmlock);
233 cl_object_attr_update(env, obj, attr, valid);
234 cl_object_attr_unlock(obj);
239 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *oscl,
240 struct lustre_handle *lockh, bool lvb_update)
242 struct ldlm_lock *dlmlock;
244 dlmlock = ldlm_handle2lock_long(lockh, 0);
245 LASSERT(dlmlock != NULL);
247 /* lock reference taken by ldlm_handle2lock_long() is
248 * owned by osc_lock and released in osc_lock_detach()
250 lu_ref_add(&dlmlock->l_reference, "osc_lock", oscl);
251 oscl->ols_has_ref = 1;
253 LASSERT(oscl->ols_dlmlock == NULL);
254 oscl->ols_dlmlock = dlmlock;
256 /* This may be a matched lock for glimpse request, do not hold
257 * lock reference in that case. */
258 if (!oscl->ols_glimpse) {
259 /* hold a refc for non glimpse lock which will
260 * be released in osc_lock_cancel() */
261 lustre_handle_copy(&oscl->ols_handle, lockh);
262 ldlm_lock_addref(lockh, oscl->ols_einfo.ei_mode);
266 /* Lock must have been granted. */
267 lock_res_and_lock(dlmlock);
268 if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
269 struct ldlm_extent *ext = &dlmlock->l_policy_data.l_extent;
270 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
272 /* extend the lock extent, otherwise it will have problem when
273 * we decide whether to grant a lockless lock. */
274 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
275 descr->cld_start = cl_index(descr->cld_obj, ext->start);
276 descr->cld_end = cl_index(descr->cld_obj, ext->end);
277 descr->cld_gid = ext->gid;
279 /* no lvb update for matched lock */
281 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
282 osc_lock_lvb_update(env, cl2osc(oscl->ols_cl.cls_obj),
285 LINVRNT(osc_lock_invariant(oscl));
287 unlock_res_and_lock(dlmlock);
289 LASSERT(oscl->ols_state != OLS_GRANTED);
290 oscl->ols_state = OLS_GRANTED;
294 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
295 * received from a server, or after osc_enqueue_base() matched a local DLM
298 static int osc_lock_upcall(void *cookie, struct lustre_handle *lockh,
301 struct osc_lock *oscl = cookie;
302 struct cl_lock_slice *slice = &oscl->ols_cl;
308 env = cl_env_percpu_get();
309 /* should never happen, similar to osc_ldlm_blocking_ast(). */
310 LASSERT(!IS_ERR(env));
312 rc = ldlm_error2errno(errcode);
313 if (oscl->ols_state == OLS_ENQUEUED) {
314 oscl->ols_state = OLS_UPCALL_RECEIVED;
315 } else if (oscl->ols_state == OLS_CANCELLED) {
318 CERROR("Impossible state: %d\n", oscl->ols_state);
323 osc_lock_granted(env, oscl, lockh, errcode == ELDLM_OK);
325 /* Error handling, some errors are tolerable. */
326 if (oscl->ols_locklessable && rc == -EUSERS) {
327 /* This is a tolerable error, turn this lock into
330 osc_object_set_contended(cl2osc(slice->cls_obj));
331 LASSERT(slice->cls_ops == &osc_lock_ops);
333 /* Change this lock to ldlmlock-less lock. */
334 osc_lock_to_lockless(env, oscl, 1);
335 oscl->ols_state = OLS_GRANTED;
337 } else if (oscl->ols_glimpse && rc == -ENAVAIL) {
338 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
339 osc_lock_lvb_update(env, cl2osc(slice->cls_obj),
340 NULL, &oscl->ols_lvb);
341 /* Hide the error. */
345 if (oscl->ols_owner != NULL)
346 cl_sync_io_note(env, oscl->ols_owner, rc);
347 cl_env_percpu_put(env);
352 static int osc_lock_upcall_agl(void *cookie, struct lustre_handle *lockh,
355 struct osc_object *osc = cookie;
356 struct ldlm_lock *dlmlock;
361 env = cl_env_get(&refcheck);
362 LASSERT(!IS_ERR(env));
364 if (errcode == ELDLM_LOCK_MATCHED)
365 GOTO(out, errcode = ELDLM_OK);
367 if (errcode != ELDLM_OK)
370 dlmlock = ldlm_handle2lock(lockh);
371 LASSERT(dlmlock != NULL);
373 lock_res_and_lock(dlmlock);
374 LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
376 /* there is no osc_lock associated with AGL lock */
377 osc_lock_lvb_update(env, osc, dlmlock, NULL);
379 unlock_res_and_lock(dlmlock);
380 LDLM_LOCK_PUT(dlmlock);
383 cl_object_put(env, osc2cl(osc));
384 cl_env_put(env, &refcheck);
385 RETURN(ldlm_error2errno(errcode));
388 static int osc_lock_flush(struct osc_object *obj, pgoff_t start, pgoff_t end,
389 enum cl_lock_mode mode, int discard)
398 env = cl_env_get(&refcheck);
400 RETURN(PTR_ERR(env));
402 if (mode == CLM_WRITE) {
403 rc = osc_cache_writeback_range(env, obj, start, end, 1,
405 CDEBUG(D_CACHE, "object %p: [%lu -> %lu] %d pages were %s.\n",
407 discard ? "discarded" : "written back");
412 rc2 = osc_lock_discard_pages(env, obj, start, end, mode);
413 if (rc == 0 && rc2 < 0)
416 cl_env_put(env, &refcheck);
421 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
422 * and ldlm_lock caches.
424 static int osc_dlm_blocking_ast0(const struct lu_env *env,
425 struct ldlm_lock *dlmlock,
426 void *data, int flag)
428 struct cl_object *obj = NULL;
431 enum cl_lock_mode mode = CLM_READ;
434 LASSERT(flag == LDLM_CB_CANCELING);
436 lock_res_and_lock(dlmlock);
437 if (dlmlock->l_granted_mode != dlmlock->l_req_mode) {
438 dlmlock->l_ast_data = NULL;
439 unlock_res_and_lock(dlmlock);
443 discard = ldlm_is_discard_data(dlmlock);
444 if (dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))
447 if (dlmlock->l_ast_data != NULL) {
448 obj = osc2cl(dlmlock->l_ast_data);
449 dlmlock->l_ast_data = NULL;
454 unlock_res_and_lock(dlmlock);
456 /* if l_ast_data is NULL, the dlmlock was enqueued by AGL or
457 * the object has been destroyed. */
459 struct ldlm_extent *extent = &dlmlock->l_policy_data.l_extent;
460 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
463 /* Destroy pages covered by the extent of the DLM lock */
464 result = osc_lock_flush(cl2osc(obj),
465 cl_index(obj, extent->start),
466 cl_index(obj, extent->end),
469 /* losing a lock, update kms */
470 lock_res_and_lock(dlmlock);
471 cl_object_attr_lock(obj);
472 /* Must get the value under the lock to avoid race. */
473 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
474 /* Update the kms. Need to loop all granted locks.
475 * Not a problem for the client */
476 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
478 cl_object_attr_update(env, obj, attr, CAT_KMS);
479 cl_object_attr_unlock(obj);
480 unlock_res_and_lock(dlmlock);
482 cl_object_put(env, obj);
488 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
489 * some other lock, or is canceled. This function is installed as a
490 * ldlm_lock::l_blocking_ast() for client extent locks.
492 * Control flow is tricky, because ldlm uses the same call-back
493 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
495 * \param dlmlock lock for which ast occurred.
497 * \param new description of a conflicting lock in case of blocking ast.
499 * \param data value of dlmlock->l_ast_data
501 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
502 * cancellation and blocking ast's.
504 * Possible use cases:
506 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
507 * lock due to lock lru pressure, or explicit user request to purge
510 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
511 * us that dlmlock conflicts with another lock that some client is
512 * enqueuing. Lock is canceled.
514 * - cl_lock_cancel() is called. osc_lock_cancel() calls
515 * ldlm_cli_cancel() that calls
517 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
519 * recursively entering osc_ldlm_blocking_ast().
521 * - client cancels lock voluntary (e.g., as a part of early cancellation):
524 * osc_lock_cancel()->
525 * ldlm_cli_cancel()->
526 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
529 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
530 struct ldlm_lock_desc *new, void *data,
537 case LDLM_CB_BLOCKING: {
538 struct lustre_handle lockh;
540 ldlm_lock2handle(dlmlock, &lockh);
541 result = ldlm_cli_cancel(&lockh, LCF_ASYNC);
542 if (result == -ENODATA)
546 case LDLM_CB_CANCELING: {
551 * This can be called in the context of outer IO, e.g.,
553 * osc_enqueue_base()->...
554 * ->ldlm_prep_elc_req()->...
555 * ->ldlm_cancel_callback()->...
556 * ->osc_ldlm_blocking_ast()
558 * new environment has to be created to not corrupt outer
561 env = cl_env_get(&refcheck);
563 result = PTR_ERR(env);
567 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
568 cl_env_put(env, &refcheck);
577 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
579 struct ptlrpc_request *req = data;
582 struct req_capsule *cap;
583 struct cl_object *obj = NULL;
589 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
591 env = cl_env_get(&refcheck);
593 GOTO(out, result = PTR_ERR(env));
596 lock_res_and_lock(dlmlock);
597 if (dlmlock->l_ast_data != NULL) {
598 obj = osc2cl(dlmlock->l_ast_data);
601 unlock_res_and_lock(dlmlock);
604 /* Do not grab the mutex of cl_lock for glimpse.
605 * See LU-1274 for details.
606 * BTW, it's okay for cl_lock to be cancelled during
607 * this period because server can handle this race.
608 * See ldlm_server_glimpse_ast() for details.
609 * cl_lock_mutex_get(env, lock); */
611 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
612 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
614 result = req_capsule_server_pack(cap);
616 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
617 result = cl_object_glimpse(env, obj, lvb);
619 if (!exp_connect_lvb_type(req->rq_export))
620 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
621 sizeof(struct ost_lvb_v1), RCL_SERVER);
622 cl_object_put(env, obj);
625 * These errors are normal races, so we don't want to
626 * fill the console with messages by calling
629 lustre_pack_reply(req, 1, NULL, NULL);
630 result = -ELDLM_NO_LOCK_DATA;
632 cl_env_put(env, &refcheck);
636 req->rq_status = result;
640 static int weigh_cb(const struct lu_env *env, struct cl_io *io,
641 struct osc_page *ops, void *cbdata)
643 struct cl_page *page = ops->ops_cl.cpl_page;
645 if (cl_page_is_vmlocked(env, page)
646 || PageDirty(page->cp_vmpage) || PageWriteback(page->cp_vmpage)
648 return CLP_GANG_ABORT;
650 *(pgoff_t *)cbdata = osc_index(ops) + 1;
651 return CLP_GANG_OKAY;
654 static unsigned long osc_lock_weight(const struct lu_env *env,
655 struct osc_object *oscobj,
656 struct ldlm_extent *extent)
658 struct cl_io *io = &osc_env_info(env)->oti_io;
659 struct cl_object *obj = cl_object_top(&oscobj->oo_cl);
665 io->ci_ignore_layout = 1;
666 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
670 page_index = cl_index(obj, extent->start);
672 result = osc_page_gang_lookup(env, io, oscobj,
674 cl_index(obj, extent->end),
675 weigh_cb, (void *)&page_index);
676 if (result == CLP_GANG_ABORT)
678 if (result == CLP_GANG_RESCHED)
680 } while (result != CLP_GANG_OKAY);
683 return result == CLP_GANG_ABORT ? 1 : 0;
687 * Get the weight of dlm lock for early cancellation.
689 unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
692 struct osc_object *obj;
693 struct osc_lock *oscl;
694 unsigned long weight;
701 * osc_ldlm_weigh_ast has a complex context since it might be called
702 * because of lock canceling, or from user's input. We have to make
703 * a new environment for it. Probably it is implementation safe to use
704 * the upper context because cl_lock_put don't modify environment
705 * variables. But just in case ..
707 env = cl_env_get(&refcheck);
709 /* Mostly because lack of memory, do not eliminate this lock */
712 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT);
713 obj = dlmlock->l_ast_data;
715 GOTO(out, weight = 1);
717 spin_lock(&obj->oo_ol_spin);
718 list_for_each_entry(oscl, &obj->oo_ol_list, ols_nextlock_oscobj) {
719 if (oscl->ols_dlmlock != NULL && oscl->ols_dlmlock != dlmlock)
723 spin_unlock(&obj->oo_ol_spin);
726 * If the lock is being used by an IO, definitely not cancel it.
728 GOTO(out, weight = 1);
731 weight = osc_lock_weight(env, obj, &dlmlock->l_policy_data.l_extent);
735 cl_env_put(env, &refcheck);
739 static void osc_lock_build_einfo(const struct lu_env *env,
740 const struct cl_lock *lock,
741 struct osc_object *osc,
742 struct ldlm_enqueue_info *einfo)
744 einfo->ei_type = LDLM_EXTENT;
745 einfo->ei_mode = osc_cl_lock2ldlm(lock->cll_descr.cld_mode);
746 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
747 einfo->ei_cb_cp = ldlm_completion_ast;
748 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
749 einfo->ei_cbdata = osc; /* value to be put into ->l_ast_data */
753 * Determine if the lock should be converted into a lockless lock.
756 * - if the lock has an explicite requirment for a non-lockless lock;
757 * - if the io lock request type ci_lockreq;
758 * - send the enqueue rpc to ost to make the further decision;
759 * - special treat to truncate lockless lock
761 * Additional policy can be implemented here, e.g., never do lockless-io
764 static void osc_lock_to_lockless(const struct lu_env *env,
765 struct osc_lock *ols, int force)
767 struct cl_lock_slice *slice = &ols->ols_cl;
769 LASSERT(ols->ols_state == OLS_NEW ||
770 ols->ols_state == OLS_UPCALL_RECEIVED);
773 ols->ols_locklessable = 1;
774 slice->cls_ops = &osc_lock_lockless_ops;
776 struct osc_io *oio = osc_env_io(env);
777 struct cl_io *io = oio->oi_cl.cis_io;
778 struct cl_object *obj = slice->cls_obj;
779 struct osc_object *oob = cl2osc(obj);
780 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
781 struct obd_connect_data *ocd;
783 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
784 io->ci_lockreq == CILR_MAYBE ||
785 io->ci_lockreq == CILR_NEVER);
787 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
788 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
789 (io->ci_lockreq == CILR_MAYBE) &&
790 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
791 if (io->ci_lockreq == CILR_NEVER ||
793 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
794 /* lockless truncate */
795 (cl_io_is_trunc(io) &&
796 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
797 osd->od_lockless_truncate)) {
798 ols->ols_locklessable = 1;
799 slice->cls_ops = &osc_lock_lockless_ops;
802 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
805 static bool osc_lock_compatible(const struct osc_lock *qing,
806 const struct osc_lock *qed)
808 struct cl_lock_descr *qed_descr = &qed->ols_cl.cls_lock->cll_descr;
809 struct cl_lock_descr *qing_descr = &qing->ols_cl.cls_lock->cll_descr;
811 if (qed->ols_glimpse)
814 if (qing_descr->cld_mode == CLM_READ && qed_descr->cld_mode == CLM_READ)
817 if (qed->ols_state < OLS_GRANTED)
820 if (qed_descr->cld_mode >= qing_descr->cld_mode &&
821 qed_descr->cld_start <= qing_descr->cld_start &&
822 qed_descr->cld_end >= qing_descr->cld_end)
828 static void osc_lock_wake_waiters(const struct lu_env *env,
829 struct osc_object *osc,
830 struct osc_lock *oscl)
832 spin_lock(&osc->oo_ol_spin);
833 list_del_init(&oscl->ols_nextlock_oscobj);
834 spin_unlock(&osc->oo_ol_spin);
836 spin_lock(&oscl->ols_lock);
837 while (!list_empty(&oscl->ols_waiting_list)) {
838 struct osc_lock *scan;
840 scan = list_entry(oscl->ols_waiting_list.next, struct osc_lock,
842 list_del_init(&scan->ols_wait_entry);
844 cl_sync_io_note(env, scan->ols_owner, 0);
846 spin_unlock(&oscl->ols_lock);
849 static int osc_lock_enqueue_wait(const struct lu_env *env,
850 struct osc_object *obj, struct osc_lock *oscl)
852 struct osc_lock *tmp_oscl;
853 struct cl_lock_descr *need = &oscl->ols_cl.cls_lock->cll_descr;
854 struct cl_sync_io *waiter = &osc_env_info(env)->oti_anchor;
858 spin_lock(&obj->oo_ol_spin);
859 list_add_tail(&oscl->ols_nextlock_oscobj, &obj->oo_ol_list);
862 list_for_each_entry(tmp_oscl, &obj->oo_ol_list,
863 ols_nextlock_oscobj) {
864 struct cl_lock_descr *descr;
866 if (tmp_oscl == oscl)
869 descr = &tmp_oscl->ols_cl.cls_lock->cll_descr;
870 if (descr->cld_start > need->cld_end ||
871 descr->cld_end < need->cld_start)
874 /* We're not supposed to give up group lock */
875 if (descr->cld_mode == CLM_GROUP)
878 if (!osc_lock_is_lockless(oscl) &&
879 osc_lock_compatible(oscl, tmp_oscl))
882 /* wait for conflicting lock to be canceled */
883 cl_sync_io_init(waiter, 1, cl_sync_io_end);
884 oscl->ols_owner = waiter;
886 spin_lock(&tmp_oscl->ols_lock);
887 /* add oscl into tmp's ols_waiting list */
888 list_add_tail(&oscl->ols_wait_entry,
889 &tmp_oscl->ols_waiting_list);
890 spin_unlock(&tmp_oscl->ols_lock);
892 spin_unlock(&obj->oo_ol_spin);
893 rc = cl_sync_io_wait(env, waiter, 0);
894 spin_lock(&obj->oo_ol_spin);
899 oscl->ols_owner = NULL;
902 spin_unlock(&obj->oo_ol_spin);
908 * Implementation of cl_lock_operations::clo_enqueue() method for osc
909 * layer. This initiates ldlm enqueue:
911 * - cancels conflicting locks early (osc_lock_enqueue_wait());
913 * - calls osc_enqueue_base() to do actual enqueue.
915 * osc_enqueue_base() is supplied with an upcall function that is executed
916 * when lock is received either after a local cached ldlm lock is matched, or
917 * when a reply from the server is received.
919 * This function does not wait for the network communication to complete.
921 static int osc_lock_enqueue(const struct lu_env *env,
922 const struct cl_lock_slice *slice,
923 struct cl_io *unused, struct cl_sync_io *anchor)
925 struct osc_thread_info *info = osc_env_info(env);
926 struct osc_io *oio = osc_env_io(env);
927 struct osc_object *osc = cl2osc(slice->cls_obj);
928 struct osc_lock *oscl = cl2osc_lock(slice);
929 struct cl_lock *lock = slice->cls_lock;
930 struct ldlm_res_id *resname = &info->oti_resname;
931 union ldlm_policy_data *policy = &info->oti_policy;
932 osc_enqueue_upcall_f upcall = osc_lock_upcall;
939 LASSERTF(ergo(oscl->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
940 "lock = %p, ols = %p\n", lock, oscl);
942 if (oscl->ols_state == OLS_GRANTED)
945 if (oscl->ols_flags & LDLM_FL_TEST_LOCK)
946 GOTO(enqueue_base, 0);
948 if (oscl->ols_glimpse) {
949 LASSERT(equi(oscl->ols_agl, anchor == NULL));
951 GOTO(enqueue_base, 0);
954 result = osc_lock_enqueue_wait(env, osc, oscl);
958 /* we can grant lockless lock right after all conflicting locks
960 if (osc_lock_is_lockless(oscl)) {
961 oscl->ols_state = OLS_GRANTED;
962 oio->oi_lockless = 1;
967 oscl->ols_state = OLS_ENQUEUED;
968 if (anchor != NULL) {
969 atomic_inc(&anchor->csi_sync_nr);
970 oscl->ols_owner = anchor;
974 * DLM lock's ast data must be osc_object;
975 * if glimpse or AGL lock, async of osc_enqueue_base() must be true,
976 * DLM's enqueue callback set to osc_lock_upcall() with cookie as
979 ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
980 osc_lock_build_policy(env, lock, policy);
982 oscl->ols_einfo.ei_cbdata = NULL;
983 /* hold a reference for callback */
984 cl_object_get(osc2cl(osc));
985 upcall = osc_lock_upcall_agl;
988 result = osc_enqueue_base(osc_export(osc), resname, &oscl->ols_flags,
989 policy, &oscl->ols_lvb,
990 osc->oo_oinfo->loi_kms_valid,
992 &oscl->ols_einfo, PTLRPCD_SET, async,
995 if (osc_lock_is_lockless(oscl)) {
996 oio->oi_lockless = 1;
998 LASSERT(oscl->ols_state == OLS_GRANTED);
999 LASSERT(oscl->ols_hold);
1000 LASSERT(oscl->ols_dlmlock != NULL);
1002 } else if (oscl->ols_agl) {
1003 cl_object_put(env, osc2cl(osc));
1009 oscl->ols_state = OLS_CANCELLED;
1010 osc_lock_wake_waiters(env, osc, oscl);
1013 cl_sync_io_note(env, anchor, result);
1019 * Breaks a link between osc_lock and dlm_lock.
1021 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
1023 struct ldlm_lock *dlmlock;
1027 dlmlock = olck->ols_dlmlock;
1028 if (dlmlock == NULL)
1031 if (olck->ols_hold) {
1033 ldlm_lock_decref(&olck->ols_handle, olck->ols_einfo.ei_mode);
1034 olck->ols_handle.cookie = 0ULL;
1037 olck->ols_dlmlock = NULL;
1039 /* release a reference taken in osc_lock_upcall(). */
1040 LASSERT(olck->ols_has_ref);
1041 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
1042 LDLM_LOCK_RELEASE(dlmlock);
1043 olck->ols_has_ref = 0;
1049 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1050 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1051 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1052 * with some other lock some where in the cluster. This function does the
1055 * - invalidates all pages protected by this lock (after sending dirty
1056 * ones to the server, as necessary);
1058 * - decref's underlying ldlm lock;
1060 * - cancels ldlm lock (ldlm_cli_cancel()).
1062 static void osc_lock_cancel(const struct lu_env *env,
1063 const struct cl_lock_slice *slice)
1065 struct osc_object *obj = cl2osc(slice->cls_obj);
1066 struct osc_lock *oscl = cl2osc_lock(slice);
1070 LINVRNT(osc_lock_invariant(oscl));
1072 osc_lock_detach(env, oscl);
1073 oscl->ols_state = OLS_CANCELLED;
1074 oscl->ols_flags &= ~LDLM_FL_LVB_READY;
1076 osc_lock_wake_waiters(env, obj, oscl);
1080 static int osc_lock_print(const struct lu_env *env, void *cookie,
1081 lu_printer_t p, const struct cl_lock_slice *slice)
1083 struct osc_lock *lock = cl2osc_lock(slice);
1085 (*p)(env, cookie, "%p %#llx %#llx %d %p ",
1086 lock->ols_dlmlock, lock->ols_flags, lock->ols_handle.cookie,
1087 lock->ols_state, lock->ols_owner);
1088 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1092 static const struct cl_lock_operations osc_lock_ops = {
1093 .clo_fini = osc_lock_fini,
1094 .clo_enqueue = osc_lock_enqueue,
1095 .clo_cancel = osc_lock_cancel,
1096 .clo_print = osc_lock_print,
1099 static void osc_lock_lockless_cancel(const struct lu_env *env,
1100 const struct cl_lock_slice *slice)
1102 struct osc_lock *ols = cl2osc_lock(slice);
1103 struct osc_object *osc = cl2osc(slice->cls_obj);
1104 struct cl_lock_descr *descr = &slice->cls_lock->cll_descr;
1107 LASSERT(ols->ols_dlmlock == NULL);
1108 result = osc_lock_flush(osc, descr->cld_start, descr->cld_end,
1109 descr->cld_mode, 0);
1111 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1114 osc_lock_wake_waiters(env, osc, ols);
1117 static const struct cl_lock_operations osc_lock_lockless_ops = {
1118 .clo_fini = osc_lock_fini,
1119 .clo_enqueue = osc_lock_enqueue,
1120 .clo_cancel = osc_lock_lockless_cancel,
1121 .clo_print = osc_lock_print
1124 static void osc_lock_set_writer(const struct lu_env *env,
1125 const struct cl_io *io,
1126 struct cl_object *obj, struct osc_lock *oscl)
1128 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
1132 if (!cl_object_same(io->ci_obj, obj))
1135 if (likely(io->ci_type == CIT_WRITE)) {
1136 io_start = cl_index(obj, io->u.ci_rw.crw_pos);
1137 io_end = cl_index(obj, io->u.ci_rw.crw_pos +
1138 io->u.ci_rw.crw_count - 1);
1140 LASSERT(cl_io_is_mkwrite(io));
1141 io_start = io_end = io->u.ci_fault.ft_index;
1144 if (descr->cld_mode >= CLM_WRITE &&
1145 (cl_io_is_append(io) ||
1146 (descr->cld_start <= io_start && descr->cld_end >= io_end))) {
1147 struct osc_io *oio = osc_env_io(env);
1149 /* There must be only one lock to match the write region */
1150 LASSERT(oio->oi_write_osclock == NULL);
1151 oio->oi_write_osclock = oscl;
1155 int osc_lock_init(const struct lu_env *env,
1156 struct cl_object *obj, struct cl_lock *lock,
1157 const struct cl_io *io)
1159 struct osc_lock *oscl;
1160 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1162 OBD_SLAB_ALLOC_PTR_GFP(oscl, osc_lock_kmem, GFP_NOFS);
1166 oscl->ols_state = OLS_NEW;
1167 spin_lock_init(&oscl->ols_lock);
1168 INIT_LIST_HEAD(&oscl->ols_waiting_list);
1169 INIT_LIST_HEAD(&oscl->ols_wait_entry);
1170 INIT_LIST_HEAD(&oscl->ols_nextlock_oscobj);
1172 oscl->ols_flags = osc_enq2ldlm_flags(enqflags);
1173 oscl->ols_agl = !!(enqflags & CEF_AGL);
1175 oscl->ols_flags |= LDLM_FL_BLOCK_NOWAIT;
1176 if (oscl->ols_flags & LDLM_FL_HAS_INTENT) {
1177 oscl->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1178 oscl->ols_glimpse = 1;
1180 osc_lock_build_einfo(env, lock, cl2osc(obj), &oscl->ols_einfo);
1182 cl_lock_slice_add(lock, &oscl->ols_cl, obj, &osc_lock_ops);
1184 if (!(enqflags & CEF_MUST))
1185 /* try to convert this lock to a lockless lock */
1186 osc_lock_to_lockless(env, oscl, (enqflags & CEF_NEVER));
1187 if (oscl->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
1188 oscl->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1190 if (io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io))
1191 osc_lock_set_writer(env, io, obj, oscl);
1193 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %#llx",
1194 lock, oscl, oscl->ols_flags);
1200 * Finds an existing lock covering given index and optionally different from a
1201 * given \a except lock.
1203 struct ldlm_lock *osc_dlmlock_at_pgoff(const struct lu_env *env,
1204 struct osc_object *obj, pgoff_t index,
1205 enum osc_dap_flags dap_flags)
1207 struct osc_thread_info *info = osc_env_info(env);
1208 struct ldlm_res_id *resname = &info->oti_resname;
1209 union ldlm_policy_data *policy = &info->oti_policy;
1210 struct lustre_handle lockh;
1211 struct ldlm_lock *lock = NULL;
1212 enum ldlm_mode mode;
1217 ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1218 osc_index2policy(policy, osc2cl(obj), index, index);
1219 policy->l_extent.gid = LDLM_GID_ANY;
1221 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1222 if (dap_flags & OSC_DAP_FL_TEST_LOCK)
1223 flags |= LDLM_FL_TEST_LOCK;
1225 * It is fine to match any group lock since there could be only one
1226 * with a uniq gid and it conflicts with all other lock modes too
1229 mode = osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
1230 LCK_PR | LCK_PW | LCK_GROUP, &flags, obj, &lockh,
1231 dap_flags & OSC_DAP_FL_CANCELING);
1233 lock = ldlm_handle2lock(&lockh);
1234 /* RACE: the lock is cancelled so let's try again */
1235 if (unlikely(lock == NULL))