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, 2017, 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 /* fid_build_reg_res_name() */
41 #include <lustre_fid.h>
42 #include <lustre_osc.h>
44 #include "osc_internal.h"
51 * Returns a weak pointer to the ldlm lock identified by a handle. Returned
52 * pointer cannot be dereferenced, as lock is not protected from concurrent
53 * reclaim. This function is a helper for osc_lock_invariant().
55 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
57 struct ldlm_lock *lock;
59 lock = ldlm_handle2lock(handle);
66 * Invariant that has to be true all of the time.
68 static int osc_lock_invariant(struct osc_lock *ols)
70 struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle);
71 struct ldlm_lock *olock = ols->ols_dlmlock;
72 int handle_used = lustre_handle_is_used(&ols->ols_handle);
74 if (ergo(osc_lock_is_lockless(ols),
75 ols->ols_locklessable && ols->ols_dlmlock == NULL))
79 * If all the following "ergo"s are true, return 1, otherwise 0
81 if (! ergo(olock != NULL, handle_used))
84 if (! ergo(olock != NULL,
85 olock->l_handle.h_cookie == ols->ols_handle.cookie))
88 if (! ergo(handle_used,
89 ergo(lock != NULL && olock != NULL, lock == olock) &&
90 ergo(lock == NULL, olock == NULL)))
93 * Check that ->ols_handle and ->ols_dlmlock are consistent, but
94 * take into account that they are set at the different time.
96 if (! ergo(ols->ols_state == OLS_CANCELLED,
97 olock == NULL && !handle_used))
100 * DLM lock is destroyed only after we have seen cancellation
103 if (! ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
104 !ldlm_is_destroyed(olock)))
107 if (! ergo(ols->ols_state == OLS_GRANTED,
109 ldlm_is_granted(olock) &&
115 /*****************************************************************************
121 void osc_lock_fini(const struct lu_env *env, struct cl_lock_slice *slice)
123 struct osc_lock *ols = cl2osc_lock(slice);
125 LINVRNT(osc_lock_invariant(ols));
126 LASSERT(ols->ols_dlmlock == NULL);
128 OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
130 EXPORT_SYMBOL(osc_lock_fini);
132 static void osc_lock_build_policy(const struct lu_env *env,
133 const struct cl_lock *lock,
134 union ldlm_policy_data *policy)
136 const struct cl_lock_descr *d = &lock->cll_descr;
138 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
139 policy->l_extent.gid = d->cld_gid;
143 * Updates object attributes from a lock value block (lvb) received together
144 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
147 * Called under lock and resource spin-locks.
149 static void osc_lock_lvb_update(const struct lu_env *env,
150 struct osc_object *osc,
151 struct ldlm_lock *dlmlock,
154 struct cl_object *obj = osc2cl(osc);
155 struct lov_oinfo *oinfo = osc->oo_oinfo;
156 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
161 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
163 LASSERT(dlmlock != NULL);
164 lvb = dlmlock->l_lvb_data;
166 cl_lvb2attr(attr, lvb);
168 cl_object_attr_lock(obj);
169 if (dlmlock != NULL) {
172 check_res_locked(dlmlock->l_resource);
174 LASSERT(lvb == dlmlock->l_lvb_data);
175 size = lvb->lvb_size;
177 /* Extend KMS up to the end of this lock and no further
178 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
179 if (size > dlmlock->l_policy_data.l_extent.end)
180 size = dlmlock->l_policy_data.l_extent.end + 1;
181 if (size >= oinfo->loi_kms) {
182 LDLM_DEBUG(dlmlock, "lock acquired, setting rss=%llu"
183 ", kms=%llu", lvb->lvb_size, size);
185 attr->cat_kms = size;
187 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
188 "%llu; leaving kms=%llu, end=%llu",
189 lvb->lvb_size, oinfo->loi_kms,
190 dlmlock->l_policy_data.l_extent.end);
192 ldlm_lock_allow_match_locked(dlmlock);
195 cl_object_attr_update(env, obj, attr, valid);
196 cl_object_attr_unlock(obj);
201 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *oscl,
202 struct lustre_handle *lockh)
204 struct ldlm_lock *dlmlock;
206 dlmlock = ldlm_handle2lock_long(lockh, 0);
207 LASSERT(dlmlock != NULL);
209 /* lock reference taken by ldlm_handle2lock_long() is
210 * owned by osc_lock and released in osc_lock_detach()
212 lu_ref_add(&dlmlock->l_reference, "osc_lock", oscl);
213 oscl->ols_has_ref = 1;
215 LASSERT(oscl->ols_dlmlock == NULL);
216 oscl->ols_dlmlock = dlmlock;
218 /* This may be a matched lock for glimpse request, do not hold
219 * lock reference in that case. */
220 if (!oscl->ols_glimpse) {
221 /* hold a refc for non glimpse lock which will
222 * be released in osc_lock_cancel() */
223 lustre_handle_copy(&oscl->ols_handle, lockh);
224 ldlm_lock_addref(lockh, oscl->ols_einfo.ei_mode);
228 /* Lock must have been granted. */
229 lock_res_and_lock(dlmlock);
230 if (ldlm_is_granted(dlmlock)) {
231 struct ldlm_extent *ext = &dlmlock->l_policy_data.l_extent;
232 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
234 /* extend the lock extent, otherwise it will have problem when
235 * we decide whether to grant a lockless lock. */
236 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
237 descr->cld_start = cl_index(descr->cld_obj, ext->start);
238 descr->cld_end = cl_index(descr->cld_obj, ext->end);
239 descr->cld_gid = ext->gid;
241 /* no lvb update for matched lock */
242 if (!ldlm_is_lvb_cached(dlmlock)) {
243 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
244 osc_lock_lvb_update(env, cl2osc(oscl->ols_cl.cls_obj),
246 ldlm_set_lvb_cached(dlmlock);
248 LINVRNT(osc_lock_invariant(oscl));
250 unlock_res_and_lock(dlmlock);
252 LASSERT(oscl->ols_state != OLS_GRANTED);
253 oscl->ols_state = OLS_GRANTED;
257 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
258 * received from a server, or after osc_enqueue_base() matched a local DLM
261 static int osc_lock_upcall(void *cookie, struct lustre_handle *lockh,
264 struct osc_lock *oscl = cookie;
265 struct cl_lock_slice *slice = &oscl->ols_cl;
271 env = cl_env_percpu_get();
272 /* should never happen, similar to osc_ldlm_blocking_ast(). */
273 LASSERT(!IS_ERR(env));
275 rc = ldlm_error2errno(errcode);
276 if (oscl->ols_state == OLS_ENQUEUED) {
277 oscl->ols_state = OLS_UPCALL_RECEIVED;
278 } else if (oscl->ols_state == OLS_CANCELLED) {
281 CERROR("Impossible state: %d\n", oscl->ols_state);
286 osc_lock_granted(env, oscl, lockh);
288 /* Error handling, some errors are tolerable. */
289 if (oscl->ols_locklessable && rc == -EUSERS) {
290 /* This is a tolerable error, turn this lock into
293 osc_object_set_contended(cl2osc(slice->cls_obj));
294 LASSERT(slice->cls_ops != oscl->ols_lockless_ops);
296 /* Change this lock to ldlmlock-less lock. */
297 osc_lock_to_lockless(env, oscl, 1);
298 oscl->ols_state = OLS_GRANTED;
300 } else if (oscl->ols_glimpse && rc == -ENAVAIL) {
301 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
302 osc_lock_lvb_update(env, cl2osc(slice->cls_obj),
303 NULL, &oscl->ols_lvb);
304 /* Hide the error. */
306 } else if (rc < 0 && oscl->ols_flags & LDLM_FL_NDELAY) {
310 if (oscl->ols_owner != NULL)
311 cl_sync_io_note(env, oscl->ols_owner, rc);
312 cl_env_percpu_put(env);
317 static int osc_lock_upcall_speculative(void *cookie,
318 struct lustre_handle *lockh,
321 struct osc_object *osc = cookie;
322 struct ldlm_lock *dlmlock;
327 env = cl_env_get(&refcheck);
328 LASSERT(!IS_ERR(env));
330 if (errcode == ELDLM_LOCK_MATCHED)
331 GOTO(out, errcode = ELDLM_OK);
333 if (errcode != ELDLM_OK)
336 dlmlock = ldlm_handle2lock(lockh);
337 LASSERT(dlmlock != NULL);
339 lock_res_and_lock(dlmlock);
340 LASSERT(ldlm_is_granted(dlmlock));
342 /* there is no osc_lock associated with speculative locks
343 * thus no need to set LDLM_FL_LVB_CACHED */
344 osc_lock_lvb_update(env, osc, dlmlock, NULL);
346 unlock_res_and_lock(dlmlock);
347 LDLM_LOCK_PUT(dlmlock);
350 cl_object_put(env, osc2cl(osc));
351 cl_env_put(env, &refcheck);
352 RETURN(ldlm_error2errno(errcode));
355 static int osc_lock_flush(struct osc_object *obj, pgoff_t start, pgoff_t end,
356 enum cl_lock_mode mode, bool discard)
365 env = cl_env_get(&refcheck);
367 RETURN(PTR_ERR(env));
369 if (mode == CLM_WRITE) {
370 rc = osc_cache_writeback_range(env, obj, start, end, 1,
372 CDEBUG(D_CACHE, "object %p: [%lu -> %lu] %d pages were %s.\n",
374 discard ? "discarded" : "written back");
380 * Do not try to match other locks with CLM_WRITE since we already
383 rc2 = osc_lock_discard_pages(env, obj, start, end,
384 mode == CLM_WRITE || discard);
385 if (rc == 0 && rc2 < 0)
388 cl_env_put(env, &refcheck);
393 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
394 * and ldlm_lock caches.
396 static int osc_dlm_blocking_ast0(const struct lu_env *env,
397 struct ldlm_lock *dlmlock,
398 void *data, int flag)
400 struct cl_object *obj = NULL;
403 enum cl_lock_mode mode = CLM_READ;
406 LASSERT(flag == LDLM_CB_CANCELING);
408 lock_res_and_lock(dlmlock);
409 if (!ldlm_is_granted(dlmlock)) {
410 dlmlock->l_ast_data = NULL;
411 unlock_res_and_lock(dlmlock);
415 discard = ldlm_is_discard_data(dlmlock);
416 if (dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))
419 if (dlmlock->l_ast_data != NULL) {
420 obj = osc2cl(dlmlock->l_ast_data);
421 dlmlock->l_ast_data = NULL;
426 unlock_res_and_lock(dlmlock);
428 /* if l_ast_data is NULL, the dlmlock was enqueued by AGL or
429 * the object has been destroyed. */
431 struct ldlm_extent *extent = &dlmlock->l_policy_data.l_extent;
432 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
435 /* Destroy pages covered by the extent of the DLM lock */
436 result = osc_lock_flush(cl2osc(obj),
437 cl_index(obj, extent->start),
438 cl_index(obj, extent->end),
441 /* losing a lock, update kms */
442 lock_res_and_lock(dlmlock);
443 cl_object_attr_lock(obj);
444 /* Must get the value under the lock to avoid race. */
445 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
446 /* Update the kms. Need to loop all granted locks.
447 * Not a problem for the client */
448 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
450 cl_object_attr_update(env, obj, attr, CAT_KMS);
451 cl_object_attr_unlock(obj);
452 unlock_res_and_lock(dlmlock);
454 cl_object_put(env, obj);
460 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
461 * some other lock, or is canceled. This function is installed as a
462 * ldlm_lock::l_blocking_ast() for client extent locks.
464 * Control flow is tricky, because ldlm uses the same call-back
465 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
467 * \param dlmlock lock for which ast occurred.
469 * \param new description of a conflicting lock in case of blocking ast.
471 * \param data value of dlmlock->l_ast_data
473 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
474 * cancellation and blocking ast's.
476 * Possible use cases:
478 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
479 * lock due to lock lru pressure, or explicit user request to purge
482 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
483 * us that dlmlock conflicts with another lock that some client is
484 * enqueuing. Lock is canceled.
486 * - cl_lock_cancel() is called. osc_lock_cancel() calls
487 * ldlm_cli_cancel() that calls
489 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
491 * recursively entering osc_ldlm_blocking_ast().
493 * - client cancels lock voluntary (e.g., as a part of early cancellation):
496 * osc_lock_cancel()->
497 * ldlm_cli_cancel()->
498 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
501 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
502 struct ldlm_lock_desc *new, void *data,
509 case LDLM_CB_BLOCKING: {
510 struct lustre_handle lockh;
512 ldlm_lock2handle(dlmlock, &lockh);
513 result = ldlm_cli_cancel(&lockh, LCF_ASYNC);
514 if (result == -ENODATA)
518 case LDLM_CB_CANCELING: {
523 * This can be called in the context of outer IO, e.g.,
525 * osc_enqueue_base()->...
526 * ->ldlm_prep_elc_req()->...
527 * ->ldlm_cancel_callback()->...
528 * ->osc_ldlm_blocking_ast()
530 * new environment has to be created to not corrupt outer
533 env = cl_env_get(&refcheck);
535 result = PTR_ERR(env);
539 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
540 cl_env_put(env, &refcheck);
549 int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
551 struct ptlrpc_request *req = data;
554 struct req_capsule *cap;
555 struct cl_object *obj = NULL;
556 struct ldlm_resource *res = dlmlock->l_resource;
557 struct ldlm_match_data matchdata = { 0 };
558 union ldlm_policy_data policy;
559 enum ldlm_mode mode = LCK_PW | LCK_GROUP | LCK_PR;
565 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
567 env = cl_env_get(&refcheck);
569 GOTO(out, result = PTR_ERR(env));
571 policy.l_extent.start = 0;
572 policy.l_extent.end = LUSTRE_EOF;
574 matchdata.lmd_mode = &mode;
575 matchdata.lmd_policy = &policy;
576 matchdata.lmd_flags = LDLM_FL_TEST_LOCK | LDLM_FL_CBPENDING;
577 matchdata.lmd_unref = 1;
578 matchdata.lmd_has_ast_data = true;
580 LDLM_LOCK_GET(dlmlock);
582 /* If any dlmlock has l_ast_data set, we must find it or we risk
583 * missing a size update done under a different lock.
586 lock_res_and_lock(dlmlock);
587 if (dlmlock->l_ast_data) {
588 obj = osc2cl(dlmlock->l_ast_data);
591 unlock_res_and_lock(dlmlock);
592 LDLM_LOCK_RELEASE(dlmlock);
596 if (obj == NULL && res->lr_type == LDLM_EXTENT) {
597 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_SIZE_DATA))
601 dlmlock = search_itree(res, &matchdata);
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, &RMF_DLM_LVB,
624 sizeof(struct ost_lvb_v1), RCL_SERVER);
625 cl_object_put(env, obj);
628 * These errors are normal races, so we don't want to
629 * fill the console with messages by calling
632 lustre_pack_reply(req, 1, NULL, NULL);
633 result = -ELDLM_NO_LOCK_DATA;
635 cl_env_put(env, &refcheck);
639 req->rq_status = result;
642 EXPORT_SYMBOL(osc_ldlm_glimpse_ast);
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) || PageDirty(page->cp_vmpage) ||
650 PageWriteback(page->cp_vmpage))
651 return CLP_GANG_ABORT;
653 *(pgoff_t *)cbdata = osc_index(ops) + 1;
654 return CLP_GANG_OKAY;
657 static unsigned long osc_lock_weight(const struct lu_env *env,
658 struct osc_object *oscobj,
659 loff_t start, loff_t end)
661 struct cl_io *io = osc_env_thread_io(env);
662 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, start);
676 result = osc_page_gang_lookup(env, io, oscobj,
677 page_index, cl_index(obj, end),
678 weigh_cb, (void *)&page_index);
679 if (result == CLP_GANG_ABORT)
681 if (result == CLP_GANG_RESCHED)
683 } while (result != CLP_GANG_OKAY);
686 return result == CLP_GANG_ABORT ? 1 : 0;
690 * Get the weight of dlm lock for early cancellation.
692 unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
695 struct osc_object *obj;
696 struct osc_lock *oscl;
697 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_get(&refcheck);
713 /* Mostly because lack of memory, do not eliminate this lock */
716 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT ||
717 dlmlock->l_resource->lr_type == LDLM_IBITS);
719 lock_res_and_lock(dlmlock);
720 obj = dlmlock->l_ast_data;
722 cl_object_get(osc2cl(obj));
723 unlock_res_and_lock(dlmlock);
726 GOTO(out, weight = 1);
728 spin_lock(&obj->oo_ol_spin);
729 list_for_each_entry(oscl, &obj->oo_ol_list, ols_nextlock_oscobj) {
730 if (oscl->ols_dlmlock != NULL && oscl->ols_dlmlock != dlmlock)
734 spin_unlock(&obj->oo_ol_spin);
737 * If the lock is being used by an IO, definitely not cancel it.
739 GOTO(out, weight = 1);
742 if (dlmlock->l_resource->lr_type == LDLM_EXTENT)
743 weight = osc_lock_weight(env, obj,
744 dlmlock->l_policy_data.l_extent.start,
745 dlmlock->l_policy_data.l_extent.end);
746 else if (ldlm_has_dom(dlmlock))
747 weight = osc_lock_weight(env, obj, 0, OBD_OBJECT_EOF);
748 /* The DOM bit can be cancelled at any time; in that case, we know
749 * there are no pages, so just return weight of 0
758 cl_object_put(env, osc2cl(obj));
760 cl_env_put(env, &refcheck);
763 EXPORT_SYMBOL(osc_ldlm_weigh_ast);
765 static void osc_lock_build_einfo(const struct lu_env *env,
766 const struct cl_lock *lock,
767 struct osc_object *osc,
768 struct ldlm_enqueue_info *einfo)
770 einfo->ei_type = LDLM_EXTENT;
771 einfo->ei_mode = osc_cl_lock2ldlm(lock->cll_descr.cld_mode);
772 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
773 einfo->ei_cb_cp = ldlm_completion_ast;
774 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
775 einfo->ei_cbdata = osc; /* value to be put into ->l_ast_data */
779 * Determine if the lock should be converted into a lockless lock.
782 * - if the lock has an explicite requirment for a non-lockless lock;
783 * - if the io lock request type ci_lockreq;
784 * - send the enqueue rpc to ost to make the further decision;
785 * - special treat to truncate lockless lock
787 * Additional policy can be implemented here, e.g., never do lockless-io
790 void osc_lock_to_lockless(const struct lu_env *env,
791 struct osc_lock *ols, int force)
793 struct cl_lock_slice *slice = &ols->ols_cl;
794 struct osc_io *oio = osc_env_io(env);
795 struct cl_io *io = oio->oi_cl.cis_io;
796 struct cl_object *obj = slice->cls_obj;
797 struct osc_object *oob = cl2osc(obj);
798 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
799 struct obd_connect_data *ocd;
801 LASSERT(ols->ols_state == OLS_NEW ||
802 ols->ols_state == OLS_UPCALL_RECEIVED);
805 ols->ols_locklessable = 1;
806 slice->cls_ops = ols->ols_lockless_ops;
808 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
809 io->ci_lockreq == CILR_MAYBE ||
810 io->ci_lockreq == CILR_NEVER);
812 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
813 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
814 (io->ci_lockreq == CILR_MAYBE) &&
815 (ocd->ocd_connect_flags &
816 OBD_CONNECT_SRVLOCK);
817 if (io->ci_lockreq == CILR_NEVER ||
819 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
820 /* lockless truncate */
821 (cl_io_is_trunc(io) && osd->od_lockless_truncate &&
822 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK))) {
823 ols->ols_locklessable = 1;
824 slice->cls_ops = ols->ols_lockless_ops;
827 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
829 EXPORT_SYMBOL(osc_lock_to_lockless);
831 static bool osc_lock_compatible(const struct osc_lock *qing,
832 const struct osc_lock *qed)
834 struct cl_lock_descr *qed_descr = &qed->ols_cl.cls_lock->cll_descr;
835 struct cl_lock_descr *qing_descr = &qing->ols_cl.cls_lock->cll_descr;
837 if (qed->ols_glimpse || qed->ols_speculative)
840 if (qing_descr->cld_mode == CLM_READ && qed_descr->cld_mode == CLM_READ)
843 if (qed->ols_state < OLS_GRANTED)
846 if (qed_descr->cld_mode >= qing_descr->cld_mode &&
847 qed_descr->cld_start <= qing_descr->cld_start &&
848 qed_descr->cld_end >= qing_descr->cld_end)
854 void osc_lock_wake_waiters(const struct lu_env *env, struct osc_object *osc,
855 struct osc_lock *oscl)
857 spin_lock(&osc->oo_ol_spin);
858 list_del_init(&oscl->ols_nextlock_oscobj);
859 spin_unlock(&osc->oo_ol_spin);
861 spin_lock(&oscl->ols_lock);
862 while (!list_empty(&oscl->ols_waiting_list)) {
863 struct osc_lock *scan;
865 scan = list_entry(oscl->ols_waiting_list.next, struct osc_lock,
867 list_del_init(&scan->ols_wait_entry);
869 cl_sync_io_note(env, scan->ols_owner, 0);
871 spin_unlock(&oscl->ols_lock);
873 EXPORT_SYMBOL(osc_lock_wake_waiters);
875 int osc_lock_enqueue_wait(const struct lu_env *env, struct osc_object *obj,
876 struct osc_lock *oscl)
878 struct osc_lock *tmp_oscl;
879 struct cl_lock_descr *need = &oscl->ols_cl.cls_lock->cll_descr;
880 struct cl_sync_io *waiter = &osc_env_info(env)->oti_anchor;
885 spin_lock(&obj->oo_ol_spin);
886 list_add_tail(&oscl->ols_nextlock_oscobj, &obj->oo_ol_list);
889 list_for_each_entry(tmp_oscl, &obj->oo_ol_list,
890 ols_nextlock_oscobj) {
891 struct cl_lock_descr *descr;
893 if (tmp_oscl == oscl)
896 descr = &tmp_oscl->ols_cl.cls_lock->cll_descr;
897 if (descr->cld_start > need->cld_end ||
898 descr->cld_end < need->cld_start)
901 /* We're not supposed to give up group lock */
902 if (descr->cld_mode == CLM_GROUP)
905 if (!osc_lock_is_lockless(oscl) &&
906 osc_lock_compatible(oscl, tmp_oscl))
909 /* wait for conflicting lock to be canceled */
910 cl_sync_io_init(waiter, 1, cl_sync_io_end);
911 oscl->ols_owner = waiter;
913 spin_lock(&tmp_oscl->ols_lock);
914 /* add oscl into tmp's ols_waiting list */
915 list_add_tail(&oscl->ols_wait_entry,
916 &tmp_oscl->ols_waiting_list);
917 spin_unlock(&tmp_oscl->ols_lock);
919 spin_unlock(&obj->oo_ol_spin);
920 rc = cl_sync_io_wait(env, waiter, 0);
921 spin_lock(&obj->oo_ol_spin);
926 oscl->ols_owner = NULL;
929 spin_unlock(&obj->oo_ol_spin);
933 EXPORT_SYMBOL(osc_lock_enqueue_wait);
936 * Implementation of cl_lock_operations::clo_enqueue() method for osc
937 * layer. This initiates ldlm enqueue:
939 * - cancels conflicting locks early (osc_lock_enqueue_wait());
941 * - calls osc_enqueue_base() to do actual enqueue.
943 * osc_enqueue_base() is supplied with an upcall function that is executed
944 * when lock is received either after a local cached ldlm lock is matched, or
945 * when a reply from the server is received.
947 * This function does not wait for the network communication to complete.
949 static int osc_lock_enqueue(const struct lu_env *env,
950 const struct cl_lock_slice *slice,
951 struct cl_io *unused, struct cl_sync_io *anchor)
953 struct osc_thread_info *info = osc_env_info(env);
954 struct osc_io *oio = osc_env_io(env);
955 struct osc_object *osc = cl2osc(slice->cls_obj);
956 struct osc_lock *oscl = cl2osc_lock(slice);
957 struct obd_export *exp = osc_export(osc);
958 struct cl_lock *lock = slice->cls_lock;
959 struct ldlm_res_id *resname = &info->oti_resname;
960 union ldlm_policy_data *policy = &info->oti_policy;
961 osc_enqueue_upcall_f upcall = osc_lock_upcall;
968 LASSERTF(ergo(oscl->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
969 "lock = %p, ols = %p\n", lock, oscl);
971 if (oscl->ols_state == OLS_GRANTED)
974 if ((oscl->ols_flags & LDLM_FL_NO_EXPANSION) &&
975 !(exp_connect_lockahead_old(exp) || exp_connect_lockahead(exp))) {
976 result = -EOPNOTSUPP;
977 CERROR("%s: server does not support lockahead/locknoexpand:"
978 "rc = %d\n", exp->exp_obd->obd_name, result);
982 if (oscl->ols_flags & LDLM_FL_TEST_LOCK)
983 GOTO(enqueue_base, 0);
985 /* For glimpse and/or speculative locks, do not wait for reply from
986 * server on LDLM request */
987 if (oscl->ols_glimpse || oscl->ols_speculative) {
988 /* Speculative and glimpse locks do not have an anchor */
989 LASSERT(equi(oscl->ols_speculative, anchor == NULL));
991 GOTO(enqueue_base, 0);
994 result = osc_lock_enqueue_wait(env, osc, oscl);
998 /* we can grant lockless lock right after all conflicting locks
1000 if (osc_lock_is_lockless(oscl)) {
1001 oscl->ols_state = OLS_GRANTED;
1002 oio->oi_lockless = 1;
1007 oscl->ols_state = OLS_ENQUEUED;
1008 if (anchor != NULL) {
1009 atomic_inc(&anchor->csi_sync_nr);
1010 oscl->ols_owner = anchor;
1014 * DLM lock's ast data must be osc_object;
1015 * if glimpse or speculative lock, async of osc_enqueue_base()
1018 * For non-speculative locks:
1019 * DLM's enqueue callback set to osc_lock_upcall() with cookie as
1021 * For speculative locks:
1022 * osc_lock_upcall_speculative & cookie is the osc object, since
1023 * there is no osc_lock
1025 ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
1026 osc_lock_build_policy(env, lock, policy);
1027 if (oscl->ols_speculative) {
1028 oscl->ols_einfo.ei_cbdata = NULL;
1029 /* hold a reference for callback */
1030 cl_object_get(osc2cl(osc));
1031 upcall = osc_lock_upcall_speculative;
1034 result = osc_enqueue_base(exp, resname, &oscl->ols_flags,
1035 policy, &oscl->ols_lvb,
1037 &oscl->ols_einfo, PTLRPCD_SET, async,
1038 oscl->ols_speculative);
1040 if (osc_lock_is_lockless(oscl)) {
1041 oio->oi_lockless = 1;
1042 } else if (!async) {
1043 LASSERT(oscl->ols_state == OLS_GRANTED);
1044 LASSERT(oscl->ols_hold);
1045 LASSERT(oscl->ols_dlmlock != NULL);
1047 } else if (oscl->ols_speculative) {
1048 cl_object_put(env, osc2cl(osc));
1049 if (oscl->ols_glimpse) {
1050 /* hide error for AGL request */
1057 oscl->ols_state = OLS_CANCELLED;
1058 osc_lock_wake_waiters(env, osc, oscl);
1061 cl_sync_io_note(env, anchor, result);
1067 * Breaks a link between osc_lock and dlm_lock.
1069 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
1071 struct ldlm_lock *dlmlock;
1075 dlmlock = olck->ols_dlmlock;
1076 if (dlmlock == NULL)
1079 if (olck->ols_hold) {
1081 ldlm_lock_decref(&olck->ols_handle, olck->ols_einfo.ei_mode);
1082 olck->ols_handle.cookie = 0ULL;
1085 olck->ols_dlmlock = NULL;
1087 /* release a reference taken in osc_lock_upcall(). */
1088 LASSERT(olck->ols_has_ref);
1089 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
1090 LDLM_LOCK_RELEASE(dlmlock);
1091 olck->ols_has_ref = 0;
1097 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1098 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1099 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1100 * with some other lock some where in the cluster. This function does the
1103 * - invalidates all pages protected by this lock (after sending dirty
1104 * ones to the server, as necessary);
1106 * - decref's underlying ldlm lock;
1108 * - cancels ldlm lock (ldlm_cli_cancel()).
1110 void osc_lock_cancel(const struct lu_env *env,
1111 const struct cl_lock_slice *slice)
1113 struct osc_object *obj = cl2osc(slice->cls_obj);
1114 struct osc_lock *oscl = cl2osc_lock(slice);
1118 LINVRNT(osc_lock_invariant(oscl));
1120 osc_lock_detach(env, oscl);
1121 oscl->ols_state = OLS_CANCELLED;
1122 oscl->ols_flags &= ~LDLM_FL_LVB_READY;
1124 osc_lock_wake_waiters(env, obj, oscl);
1127 EXPORT_SYMBOL(osc_lock_cancel);
1129 int osc_lock_print(const struct lu_env *env, void *cookie,
1130 lu_printer_t p, const struct cl_lock_slice *slice)
1132 struct osc_lock *lock = cl2osc_lock(slice);
1134 (*p)(env, cookie, "%p %#llx %#llx %d %p ",
1135 lock->ols_dlmlock, lock->ols_flags, lock->ols_handle.cookie,
1136 lock->ols_state, lock->ols_owner);
1137 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1140 EXPORT_SYMBOL(osc_lock_print);
1142 static const struct cl_lock_operations osc_lock_ops = {
1143 .clo_fini = osc_lock_fini,
1144 .clo_enqueue = osc_lock_enqueue,
1145 .clo_cancel = osc_lock_cancel,
1146 .clo_print = osc_lock_print,
1149 static void osc_lock_lockless_cancel(const struct lu_env *env,
1150 const struct cl_lock_slice *slice)
1152 struct osc_lock *ols = cl2osc_lock(slice);
1153 struct osc_object *osc = cl2osc(slice->cls_obj);
1154 struct cl_lock_descr *descr = &slice->cls_lock->cll_descr;
1157 LASSERT(ols->ols_dlmlock == NULL);
1158 result = osc_lock_flush(osc, descr->cld_start, descr->cld_end,
1159 descr->cld_mode, false);
1161 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1164 osc_lock_wake_waiters(env, osc, ols);
1167 static const struct cl_lock_operations osc_lock_lockless_ops = {
1168 .clo_fini = osc_lock_fini,
1169 .clo_enqueue = osc_lock_enqueue,
1170 .clo_cancel = osc_lock_lockless_cancel,
1171 .clo_print = osc_lock_print
1174 void osc_lock_set_writer(const struct lu_env *env, const struct cl_io *io,
1175 struct cl_object *obj, struct osc_lock *oscl)
1177 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
1181 if (!cl_object_same(io->ci_obj, obj))
1184 if (likely(io->ci_type == CIT_WRITE)) {
1185 io_start = cl_index(obj, io->u.ci_rw.crw_pos);
1186 io_end = cl_index(obj, io->u.ci_rw.crw_pos +
1187 io->u.ci_rw.crw_count - 1);
1189 LASSERT(cl_io_is_mkwrite(io));
1190 io_start = io_end = io->u.ci_fault.ft_index;
1193 if (descr->cld_mode >= CLM_WRITE &&
1194 (cl_io_is_append(io) ||
1195 (descr->cld_start <= io_start && descr->cld_end >= io_end))) {
1196 struct osc_io *oio = osc_env_io(env);
1198 /* There must be only one lock to match the write region */
1199 LASSERT(oio->oi_write_osclock == NULL);
1200 oio->oi_write_osclock = oscl;
1203 EXPORT_SYMBOL(osc_lock_set_writer);
1205 int osc_lock_init(const struct lu_env *env,
1206 struct cl_object *obj, struct cl_lock *lock,
1207 const struct cl_io *io)
1209 struct osc_lock *oscl;
1210 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1212 OBD_SLAB_ALLOC_PTR_GFP(oscl, osc_lock_kmem, GFP_NOFS);
1216 oscl->ols_state = OLS_NEW;
1217 spin_lock_init(&oscl->ols_lock);
1218 INIT_LIST_HEAD(&oscl->ols_waiting_list);
1219 INIT_LIST_HEAD(&oscl->ols_wait_entry);
1220 INIT_LIST_HEAD(&oscl->ols_nextlock_oscobj);
1221 oscl->ols_lockless_ops = &osc_lock_lockless_ops;
1223 /* Speculative lock requests must be either no_expand or glimpse
1224 * request (CEF_GLIMPSE). non-glimpse no_expand speculative extent
1225 * locks will break ofd_intent_cb. (see comment there)*/
1226 LASSERT(ergo((enqflags & CEF_SPECULATIVE) != 0,
1227 (enqflags & (CEF_LOCK_NO_EXPAND | CEF_GLIMPSE)) != 0));
1229 oscl->ols_flags = osc_enq2ldlm_flags(enqflags);
1230 oscl->ols_speculative = !!(enqflags & CEF_SPECULATIVE);
1232 if (oscl->ols_flags & LDLM_FL_HAS_INTENT) {
1233 oscl->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1234 oscl->ols_glimpse = 1;
1236 if (io->ci_ndelay && cl_object_same(io->ci_obj, obj))
1237 oscl->ols_flags |= LDLM_FL_NDELAY;
1238 osc_lock_build_einfo(env, lock, cl2osc(obj), &oscl->ols_einfo);
1240 cl_lock_slice_add(lock, &oscl->ols_cl, obj, &osc_lock_ops);
1242 if (!(enqflags & CEF_MUST))
1243 /* try to convert this lock to a lockless lock */
1244 osc_lock_to_lockless(env, oscl, (enqflags & CEF_NEVER));
1245 if (oscl->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
1246 oscl->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1248 if (io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io))
1249 osc_lock_set_writer(env, io, obj, oscl);
1251 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %#llx",
1252 lock, oscl, oscl->ols_flags);
1258 * Finds an existing lock covering given index and optionally different from a
1259 * given \a except lock.
1261 struct ldlm_lock *osc_obj_dlmlock_at_pgoff(const struct lu_env *env,
1262 struct osc_object *obj,
1264 enum osc_dap_flags dap_flags)
1266 struct osc_thread_info *info = osc_env_info(env);
1267 struct ldlm_res_id *resname = &info->oti_resname;
1268 union ldlm_policy_data *policy = &info->oti_policy;
1269 struct lustre_handle lockh;
1270 struct ldlm_lock *lock = NULL;
1271 enum ldlm_mode mode;
1276 ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1277 osc_index2policy(policy, osc2cl(obj), index, index);
1278 policy->l_extent.gid = LDLM_GID_ANY;
1280 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1281 if (dap_flags & OSC_DAP_FL_TEST_LOCK)
1282 flags |= LDLM_FL_TEST_LOCK;
1284 * It is fine to match any group lock since there could be only one
1285 * with a uniq gid and it conflicts with all other lock modes too
1288 mode = osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
1289 LCK_PR | LCK_PW | LCK_GROUP, &flags, obj, &lockh,
1290 dap_flags & OSC_DAP_FL_CANCELING);
1292 lock = ldlm_handle2lock(&lockh);
1293 /* RACE: the lock is cancelled so let's try again */
1294 if (unlikely(lock == NULL))