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/
31 * Implementation of cl_lock for OSC layer.
33 * Author: Nikita Danilov <nikita.danilov@sun.com>
34 * Author: Jinshan Xiong <jinshan.xiong@intel.com>
37 #define DEBUG_SUBSYSTEM S_OSC
39 /* fid_build_reg_res_name() */
40 #include <lustre_fid.h>
41 #include <lustre_osc.h>
43 #include "osc_internal.h"
50 * Returns a weak pointer to the ldlm lock identified by a handle. Returned
51 * pointer cannot be dereferenced, as lock is not protected from concurrent
52 * reclaim. This function is a helper for osc_lock_invariant().
54 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
56 struct ldlm_lock *lock;
58 lock = ldlm_handle2lock(handle);
65 * Invariant that has to be true all of the time.
67 static int osc_lock_invariant(struct osc_lock *ols)
69 struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle);
70 struct ldlm_lock *olock = ols->ols_dlmlock;
71 int handle_used = lustre_handle_is_used(&ols->ols_handle);
73 if (ergo(osc_lock_is_lockless(ols),
74 ols->ols_locklessable && ols->ols_dlmlock == NULL))
78 * If all the following "ergo"s are true, return 1, otherwise 0
80 if (! ergo(olock != NULL, handle_used))
83 if (! ergo(olock != NULL,
84 olock->l_handle.h_cookie == ols->ols_handle.cookie))
87 if (! ergo(handle_used,
88 ergo(lock != NULL && olock != NULL, lock == olock) &&
89 ergo(lock == NULL, olock == NULL)))
92 * Check that ->ols_handle and ->ols_dlmlock are consistent, but
93 * take into account that they are set at the different time.
95 if (! ergo(ols->ols_state == OLS_CANCELLED,
96 olock == NULL && !handle_used))
99 * DLM lock is destroyed only after we have seen cancellation
102 if (! ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
103 !ldlm_is_destroyed(olock)))
106 if (! ergo(ols->ols_state == OLS_GRANTED,
108 ldlm_is_granted(olock) &&
114 /*****************************************************************************
120 void osc_lock_fini(const struct lu_env *env, struct cl_lock_slice *slice)
122 struct osc_lock *ols = cl2osc_lock(slice);
124 LINVRNT(osc_lock_invariant(ols));
125 LASSERT(ols->ols_dlmlock == NULL);
127 OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
129 EXPORT_SYMBOL(osc_lock_fini);
131 static void osc_lock_build_policy(const struct lu_env *env,
132 const struct cl_lock *lock,
133 union ldlm_policy_data *policy)
135 const struct cl_lock_descr *d = &lock->cll_descr;
137 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
138 policy->l_extent.gid = d->cld_gid;
142 * Updates object attributes from a lock value block (lvb) received together
143 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
146 * Called under lock and resource spin-locks.
148 void osc_lock_lvb_update(const struct lu_env *env,
149 struct osc_object *osc,
150 struct ldlm_lock *dlmlock,
153 struct cl_object *obj = osc2cl(osc);
154 struct lov_oinfo *oinfo = osc->oo_oinfo;
155 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
156 unsigned valid, setkms = 0;
160 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
162 LASSERT(dlmlock != NULL);
163 lvb = dlmlock->l_lvb_data;
165 cl_lvb2attr(attr, lvb);
167 cl_object_attr_lock(obj);
168 if (dlmlock != NULL) {
171 check_res_locked(dlmlock->l_resource);
173 LASSERT(lvb == dlmlock->l_lvb_data);
174 size = lvb->lvb_size;
176 /* Extend KMS up to the end of this lock and no further
177 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
178 if (size > dlmlock->l_policy_data.l_extent.end)
179 size = dlmlock->l_policy_data.l_extent.end + 1;
180 if (size >= oinfo->loi_kms) {
182 attr->cat_kms = size;
185 ldlm_lock_allow_match_locked(dlmlock);
188 /* The size should not be less than the kms */
189 if (attr->cat_size < oinfo->loi_kms)
190 attr->cat_size = oinfo->loi_kms;
192 LDLM_DEBUG(dlmlock, "acquired size %llu, setting rss=%llu;%s "
193 "kms=%llu, end=%llu", lvb->lvb_size, attr->cat_size,
194 setkms ? "" : " leaving",
195 setkms ? attr->cat_kms : oinfo->loi_kms,
196 dlmlock ? dlmlock->l_policy_data.l_extent.end : -1ull);
198 cl_object_attr_update(env, obj, attr, valid);
199 cl_object_attr_unlock(obj);
204 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *oscl,
205 struct lustre_handle *lockh)
207 struct osc_object *osc = cl2osc(oscl->ols_cl.cls_obj);
208 struct ldlm_lock *dlmlock;
210 dlmlock = ldlm_handle2lock_long(lockh, 0);
211 LASSERT(dlmlock != NULL);
213 /* lock reference taken by ldlm_handle2lock_long() is
214 * owned by osc_lock and released in osc_lock_detach()
216 lu_ref_add_atomic(&dlmlock->l_reference, "osc_lock", oscl);
217 oscl->ols_has_ref = 1;
219 LASSERT(oscl->ols_dlmlock == NULL);
220 oscl->ols_dlmlock = dlmlock;
222 /* This may be a matched lock for glimpse request, do not hold
223 * lock reference in that case. */
224 if (!oscl->ols_glimpse) {
225 /* hold a refc for non glimpse lock which will
226 * be released in osc_lock_cancel() */
227 lustre_handle_copy(&oscl->ols_handle, lockh);
228 ldlm_lock_addref(lockh, oscl->ols_einfo.ei_mode);
232 /* Lock must have been granted. */
233 lock_res_and_lock(dlmlock);
234 if (ldlm_is_granted(dlmlock)) {
235 struct ldlm_extent *ext = &dlmlock->l_policy_data.l_extent;
236 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
238 /* extend the lock extent, otherwise it will have problem when
239 * we decide whether to grant a lockless lock. */
240 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
241 descr->cld_start = cl_index(descr->cld_obj, ext->start);
242 descr->cld_end = cl_index(descr->cld_obj, ext->end);
243 descr->cld_gid = ext->gid;
245 /* no lvb update for matched lock */
246 if (!ldlm_is_lvb_cached(dlmlock)) {
247 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
248 LASSERT(osc == dlmlock->l_ast_data);
249 osc_lock_lvb_update(env, osc, dlmlock, NULL);
250 ldlm_set_lvb_cached(dlmlock);
252 LINVRNT(osc_lock_invariant(oscl));
254 unlock_res_and_lock(dlmlock);
256 LASSERT(oscl->ols_state != OLS_GRANTED);
257 oscl->ols_state = OLS_GRANTED;
261 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
262 * received from a server, or after osc_enqueue_base() matched a local DLM
265 static int osc_lock_upcall(void *cookie, struct lustre_handle *lockh,
268 struct osc_lock *oscl = cookie;
269 struct cl_lock_slice *slice = &oscl->ols_cl;
275 env = cl_env_percpu_get();
276 /* should never happen, similar to osc_ldlm_blocking_ast(). */
277 LASSERT(!IS_ERR(env));
279 rc = ldlm_error2errno(errcode);
280 if (oscl->ols_state == OLS_ENQUEUED) {
281 oscl->ols_state = OLS_UPCALL_RECEIVED;
282 } else if (oscl->ols_state == OLS_CANCELLED) {
285 CERROR("Impossible state: %d\n", oscl->ols_state);
290 osc_lock_granted(env, oscl, lockh);
292 /* Error handling, some errors are tolerable. */
293 if (oscl->ols_glimpse && rc == -ENAVAIL) {
294 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
295 osc_lock_lvb_update(env, cl2osc(slice->cls_obj),
296 NULL, &oscl->ols_lvb);
297 /* Hide the error. */
299 } else if (rc < 0 && oscl->ols_flags & LDLM_FL_NDELAY) {
303 if (oscl->ols_owner != NULL)
304 cl_sync_io_note(env, oscl->ols_owner, rc);
305 cl_env_percpu_put(env);
310 static int osc_lock_upcall_speculative(void *cookie,
311 struct lustre_handle *lockh,
314 struct osc_object *osc = cookie;
315 struct ldlm_lock *dlmlock;
320 env = cl_env_get(&refcheck);
321 LASSERT(!IS_ERR(env));
323 if (errcode == ELDLM_LOCK_MATCHED)
324 GOTO(out, errcode = ELDLM_OK);
326 if (errcode != ELDLM_OK)
329 dlmlock = ldlm_handle2lock(lockh);
330 LASSERT(dlmlock != NULL);
332 lock_res_and_lock(dlmlock);
333 LASSERT(ldlm_is_granted(dlmlock));
335 /* there is no osc_lock associated with speculative locks
336 * thus no need to set LDLM_FL_LVB_CACHED */
337 osc_lock_lvb_update(env, osc, dlmlock, NULL);
339 unlock_res_and_lock(dlmlock);
340 LDLM_LOCK_PUT(dlmlock);
343 cl_object_put(env, osc2cl(osc));
344 cl_env_put(env, &refcheck);
345 RETURN(ldlm_error2errno(errcode));
348 static int osc_lock_flush(struct osc_object *obj, pgoff_t start, pgoff_t end,
349 enum cl_lock_mode mode, bool discard)
358 env = cl_env_get(&refcheck);
360 RETURN(PTR_ERR(env));
362 if (mode == CLM_WRITE) {
363 rc = osc_cache_writeback_range(env, obj, start, end, 1,
365 CDEBUG(D_CACHE, "object %p: [%lu -> %lu] %d pages were %s.\n",
367 discard ? "discarded" : "written back");
373 * Do not try to match other locks with CLM_WRITE since we already
376 rc2 = osc_lock_discard_pages(env, obj, start, end,
377 mode == CLM_WRITE || discard);
378 if (rc == 0 && rc2 < 0)
381 cl_env_put(env, &refcheck);
386 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
387 * and ldlm_lock caches.
389 static int osc_dlm_blocking_ast0(const struct lu_env *env,
390 struct ldlm_lock *dlmlock,
391 void *data, int flag)
393 struct cl_object *obj = NULL;
396 enum cl_lock_mode mode = CLM_READ;
399 LASSERT(flag == LDLM_CB_CANCELING);
401 lock_res_and_lock(dlmlock);
402 if (!ldlm_is_granted(dlmlock)) {
403 dlmlock->l_ast_data = NULL;
404 unlock_res_and_lock(dlmlock);
408 discard = ldlm_is_discard_data(dlmlock);
409 if (dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))
412 if (dlmlock->l_ast_data != NULL) {
413 obj = osc2cl(dlmlock->l_ast_data);
417 unlock_res_and_lock(dlmlock);
419 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_CANCEL, 5);
421 /* if l_ast_data is NULL, the dlmlock was enqueued by AGL or
422 * the object has been destroyed. */
424 struct ldlm_extent *extent = &dlmlock->l_policy_data.l_extent;
425 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
428 /* Destroy pages covered by the extent of the DLM lock */
429 result = osc_lock_flush(cl2osc(obj),
430 cl_index(obj, extent->start),
431 cl_index(obj, extent->end),
434 /* losing a lock, update kms */
435 lock_res_and_lock(dlmlock);
436 /* clearing l_ast_data after flushing data,
437 * to let glimpse ast find the lock and the object */
438 dlmlock->l_ast_data = NULL;
439 cl_object_attr_lock(obj);
440 /* Must get the value under the lock to avoid race. */
441 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
442 /* Update the kms. Need to loop all granted locks.
443 * Not a problem for the client */
444 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
446 cl_object_attr_update(env, obj, attr, CAT_KMS);
447 cl_object_attr_unlock(obj);
448 unlock_res_and_lock(dlmlock);
450 cl_object_put(env, obj);
456 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
457 * some other lock, or is canceled. This function is installed as a
458 * ldlm_lock::l_blocking_ast() for client extent locks.
460 * Control flow is tricky, because ldlm uses the same call-back
461 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
463 * \param dlmlock lock for which ast occurred.
465 * \param new description of a conflicting lock in case of blocking ast.
467 * \param data value of dlmlock->l_ast_data
469 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
470 * cancellation and blocking ast's.
472 * Possible use cases:
474 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
475 * lock due to lock lru pressure, or explicit user request to purge
478 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
479 * us that dlmlock conflicts with another lock that some client is
480 * enqueuing. Lock is canceled.
482 * - cl_lock_cancel() is called. osc_lock_cancel() calls
483 * ldlm_cli_cancel() that calls
485 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
487 * recursively entering osc_ldlm_blocking_ast().
489 * - client cancels lock voluntary (e.g., as a part of early cancellation):
492 * osc_lock_cancel()->
493 * ldlm_cli_cancel()->
494 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
497 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
498 struct ldlm_lock_desc *new, void *data,
505 case LDLM_CB_BLOCKING: {
506 struct lustre_handle lockh;
508 ldlm_lock2handle(dlmlock, &lockh);
509 result = ldlm_cli_cancel(&lockh, LCF_ASYNC);
510 if (result == -ENODATA)
514 case LDLM_CB_CANCELING: {
519 * This can be called in the context of outer IO, e.g.,
521 * osc_enqueue_base()->...
522 * ->ldlm_prep_elc_req()->...
523 * ->ldlm_cancel_callback()->...
524 * ->osc_ldlm_blocking_ast()
526 * new environment has to be created to not corrupt outer
529 env = cl_env_get(&refcheck);
531 result = PTR_ERR(env);
535 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
536 cl_env_put(env, &refcheck);
545 int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
547 struct ptlrpc_request *req = data;
550 struct req_capsule *cap;
551 struct cl_object *obj = NULL;
552 struct ldlm_resource *res = dlmlock->l_resource;
553 struct ldlm_match_data matchdata = { 0 };
554 union ldlm_policy_data policy;
555 enum ldlm_mode mode = LCK_PW | LCK_GROUP | LCK_PR;
561 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
563 env = cl_env_get(&refcheck);
565 GOTO(out, result = PTR_ERR(env));
567 policy.l_extent.start = 0;
568 policy.l_extent.end = LUSTRE_EOF;
570 matchdata.lmd_mode = &mode;
571 matchdata.lmd_policy = &policy;
572 matchdata.lmd_flags = LDLM_FL_TEST_LOCK | LDLM_FL_CBPENDING;
573 matchdata.lmd_match = LDLM_MATCH_UNREF | LDLM_MATCH_AST_ANY;
575 LDLM_LOCK_GET(dlmlock);
577 /* If any dlmlock has l_ast_data set, we must find it or we risk
578 * missing a size update done under a different lock.
581 lock_res_and_lock(dlmlock);
582 if (dlmlock->l_ast_data) {
583 obj = osc2cl(dlmlock->l_ast_data);
586 unlock_res_and_lock(dlmlock);
587 LDLM_LOCK_RELEASE(dlmlock);
591 if (obj == NULL && res->lr_type == LDLM_EXTENT) {
592 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_SIZE_DATA))
596 dlmlock = search_itree(res, &matchdata);
602 /* Do not grab the mutex of cl_lock for glimpse.
603 * See LU-1274 for details.
604 * BTW, it's okay for cl_lock to be cancelled during
605 * this period because server can handle this race.
606 * See ldlm_server_glimpse_ast() for details.
607 * cl_lock_mutex_get(env, lock); */
609 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
610 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
612 result = req_capsule_server_pack(cap);
614 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
615 result = cl_object_glimpse(env, obj, lvb);
617 if (!exp_connect_lvb_type(req->rq_export))
618 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
619 sizeof(struct ost_lvb_v1), RCL_SERVER);
620 cl_object_put(env, obj);
623 * These errors are normal races, so we don't want to
624 * fill the console with messages by calling
627 lustre_pack_reply(req, 1, NULL, NULL);
628 result = -ELDLM_NO_LOCK_DATA;
630 cl_env_put(env, &refcheck);
634 req->rq_status = result;
637 EXPORT_SYMBOL(osc_ldlm_glimpse_ast);
639 static bool weigh_cb(const struct lu_env *env, struct cl_io *io,
640 void **pvec, int count, void *cbdata)
644 for (i = 0; i < count; i++) {
645 struct osc_page *ops = pvec[i];
646 struct cl_page *page = ops->ops_cl.cpl_page;
648 if (cl_page_is_vmlocked(env, page) ||
649 PageDirty(page->cp_vmpage) ||
650 PageWriteback(page->cp_vmpage))
653 *(pgoff_t *)cbdata = osc_index(ops) + 1;
658 static unsigned long osc_lock_weight(const struct lu_env *env,
659 struct osc_object *oscobj,
660 loff_t start, loff_t end)
662 struct cl_io *io = osc_env_thread_io(env);
663 struct cl_object *obj = cl_object_top(&oscobj->oo_cl);
670 io->ci_ignore_layout = 1;
671 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
675 page_index = cl_index(obj, start);
677 if (!osc_page_gang_lookup(env, io, oscobj,
678 page_index, cl_index(obj, end),
679 weigh_cb, (void *)&page_index))
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;
702 * osc_ldlm_weigh_ast has a complex context since it might be called
703 * because of lock canceling, or from user's input. We have to make
704 * a new environment for it. Probably it is implementation safe to use
705 * the upper context because cl_lock_put don't modify environment
706 * variables. But just in case ..
708 env = cl_env_get(&refcheck);
710 /* Mostly because lack of memory, do not eliminate this lock */
713 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT ||
714 dlmlock->l_resource->lr_type == LDLM_IBITS);
716 lock_res_and_lock(dlmlock);
717 obj = dlmlock->l_ast_data;
719 cl_object_get(osc2cl(obj));
720 unlock_res_and_lock(dlmlock);
723 GOTO(out, weight = 0);
725 spin_lock(&obj->oo_ol_spin);
726 list_for_each_entry(oscl, &obj->oo_ol_list, ols_nextlock_oscobj) {
727 if (oscl->ols_dlmlock == dlmlock) {
732 spin_unlock(&obj->oo_ol_spin);
735 * If the lock is being used by an IO, definitely not cancel it.
737 GOTO(out, weight = 1);
740 if (dlmlock->l_resource->lr_type == LDLM_EXTENT)
741 weight = osc_lock_weight(env, obj,
742 dlmlock->l_policy_data.l_extent.start,
743 dlmlock->l_policy_data.l_extent.end);
744 else if (ldlm_has_dom(dlmlock))
745 weight = osc_lock_weight(env, obj, 0, OBD_OBJECT_EOF);
746 /* The DOM bit can be cancelled at any time; in that case, we know
747 * there are no pages, so just return weight of 0
756 cl_object_put(env, osc2cl(obj));
758 cl_env_put(env, &refcheck);
761 EXPORT_SYMBOL(osc_ldlm_weigh_ast);
763 static void osc_lock_build_einfo(const struct lu_env *env,
764 const struct cl_lock *lock,
765 struct osc_object *osc,
766 struct ldlm_enqueue_info *einfo)
768 einfo->ei_type = LDLM_EXTENT;
769 einfo->ei_mode = osc_cl_lock2ldlm(lock->cll_descr.cld_mode);
770 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
771 einfo->ei_cb_cp = ldlm_completion_ast;
772 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
773 einfo->ei_cbdata = osc; /* value to be put into ->l_ast_data */
777 * Determine if the lock should be converted into a lockless lock.
780 * - if the lock has an explicite requirment for a non-lockless lock;
781 * - if the io lock request type ci_lockreq;
782 * - send the enqueue rpc to ost to make the further decision;
783 * - special treat to truncate lockless lock
785 * Additional policy can be implemented here, e.g., never do lockless-io
788 void osc_lock_to_lockless(const struct lu_env *env,
789 struct osc_lock *ols, int force)
791 struct cl_lock_slice *slice = &ols->ols_cl;
792 struct osc_io *oio = osc_env_io(env);
793 struct cl_io *io = oio->oi_cl.cis_io;
794 struct cl_object *obj = slice->cls_obj;
795 struct osc_object *oob = cl2osc(obj);
796 struct obd_connect_data *ocd;
798 LASSERT(ols->ols_state == OLS_NEW ||
799 ols->ols_state == OLS_UPCALL_RECEIVED);
802 ols->ols_locklessable = 1;
803 slice->cls_ops = ols->ols_lockless_ops;
805 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
806 io->ci_lockreq == CILR_MAYBE ||
807 io->ci_lockreq == CILR_NEVER);
809 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
810 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
811 (io->ci_lockreq == CILR_MAYBE) &&
812 (ocd->ocd_connect_flags &
813 OBD_CONNECT_SRVLOCK);
814 if (io->ci_lockreq == CILR_NEVER) {
815 ols->ols_locklessable = 1;
816 slice->cls_ops = ols->ols_lockless_ops;
819 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
821 EXPORT_SYMBOL(osc_lock_to_lockless);
823 static bool osc_lock_compatible(const struct osc_lock *qing,
824 const struct osc_lock *qed)
826 struct cl_lock_descr *qed_descr = &qed->ols_cl.cls_lock->cll_descr;
827 struct cl_lock_descr *qing_descr = &qing->ols_cl.cls_lock->cll_descr;
829 if (qed->ols_glimpse || qed->ols_speculative)
832 if (qing_descr->cld_mode == CLM_READ && qed_descr->cld_mode == CLM_READ)
835 if (qed->ols_state < OLS_GRANTED)
838 if (qed_descr->cld_mode >= qing_descr->cld_mode &&
839 qed_descr->cld_start <= qing_descr->cld_start &&
840 qed_descr->cld_end >= qing_descr->cld_end)
846 void osc_lock_wake_waiters(const struct lu_env *env, struct osc_object *osc,
847 struct osc_lock *oscl)
849 struct osc_lock *scan;
851 spin_lock(&osc->oo_ol_spin);
852 list_del_init(&oscl->ols_nextlock_oscobj);
853 spin_unlock(&osc->oo_ol_spin);
855 spin_lock(&oscl->ols_lock);
856 while ((scan = list_first_entry_or_null(&oscl->ols_waiting_list,
858 ols_wait_entry)) != NULL) {
859 list_del_init(&scan->ols_wait_entry);
861 cl_sync_io_note(env, scan->ols_owner, 0);
863 spin_unlock(&oscl->ols_lock);
865 EXPORT_SYMBOL(osc_lock_wake_waiters);
867 int osc_lock_enqueue_wait(const struct lu_env *env, struct osc_object *obj,
868 struct osc_lock *oscl)
870 struct osc_lock *tmp_oscl;
871 struct cl_lock_descr *need = &oscl->ols_cl.cls_lock->cll_descr;
872 struct cl_sync_io *waiter = &osc_env_info(env)->oti_anchor;
877 spin_lock(&obj->oo_ol_spin);
878 list_add_tail(&oscl->ols_nextlock_oscobj, &obj->oo_ol_list);
881 list_for_each_entry(tmp_oscl, &obj->oo_ol_list,
882 ols_nextlock_oscobj) {
883 struct cl_lock_descr *descr;
885 if (tmp_oscl == oscl)
888 descr = &tmp_oscl->ols_cl.cls_lock->cll_descr;
889 if (descr->cld_start > need->cld_end ||
890 descr->cld_end < need->cld_start)
893 /* We're not supposed to give up group lock */
894 if (descr->cld_mode == CLM_GROUP)
897 if (!osc_lock_is_lockless(oscl) &&
898 osc_lock_compatible(oscl, tmp_oscl))
901 /* wait for conflicting lock to be canceled */
902 cl_sync_io_init(waiter, 1);
903 oscl->ols_owner = waiter;
905 spin_lock(&tmp_oscl->ols_lock);
906 /* add oscl into tmp's ols_waiting list */
907 list_add_tail(&oscl->ols_wait_entry,
908 &tmp_oscl->ols_waiting_list);
909 spin_unlock(&tmp_oscl->ols_lock);
911 spin_unlock(&obj->oo_ol_spin);
912 rc = cl_sync_io_wait(env, waiter, 0);
913 spin_lock(&obj->oo_ol_spin);
918 oscl->ols_owner = NULL;
921 spin_unlock(&obj->oo_ol_spin);
925 EXPORT_SYMBOL(osc_lock_enqueue_wait);
928 * Implementation of cl_lock_operations::clo_enqueue() method for osc
929 * layer. This initiates ldlm enqueue:
931 * - cancels conflicting locks early (osc_lock_enqueue_wait());
933 * - calls osc_enqueue_base() to do actual enqueue.
935 * osc_enqueue_base() is supplied with an upcall function that is executed
936 * when lock is received either after a local cached ldlm lock is matched, or
937 * when a reply from the server is received.
939 * This function does not wait for the network communication to complete.
941 static int osc_lock_enqueue(const struct lu_env *env,
942 const struct cl_lock_slice *slice,
943 struct cl_io *unused, struct cl_sync_io *anchor)
945 struct osc_thread_info *info = osc_env_info(env);
946 struct osc_io *oio = osc_env_io(env);
947 struct osc_object *osc = cl2osc(slice->cls_obj);
948 struct osc_lock *oscl = cl2osc_lock(slice);
949 struct obd_export *exp = osc_export(osc);
950 struct cl_lock *lock = slice->cls_lock;
951 struct ldlm_res_id *resname = &info->oti_resname;
952 union ldlm_policy_data *policy = &info->oti_policy;
953 osc_enqueue_upcall_f upcall = osc_lock_upcall;
960 LASSERTF(ergo(oscl->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
961 "lock = %p, ols = %p\n", lock, oscl);
963 if (oscl->ols_state == OLS_GRANTED)
966 if ((oscl->ols_flags & LDLM_FL_NO_EXPANSION) &&
967 !exp_connect_lockahead(exp)) {
968 result = -EOPNOTSUPP;
969 CERROR("%s: server does not support lockahead/locknoexpand: rc = %d\n",
970 exp->exp_obd->obd_name, result);
974 if (oscl->ols_flags & LDLM_FL_TEST_LOCK)
975 GOTO(enqueue_base, 0);
977 /* For glimpse and/or speculative locks, do not wait for reply from
978 * server on LDLM request */
979 if (oscl->ols_glimpse || oscl->ols_speculative) {
980 /* Speculative and glimpse locks do not have an anchor */
981 LASSERT(equi(oscl->ols_speculative, anchor == NULL));
983 GOTO(enqueue_base, 0);
986 result = osc_lock_enqueue_wait(env, osc, oscl);
990 /* we can grant lockless lock right after all conflicting locks
992 if (osc_lock_is_lockless(oscl)) {
993 oscl->ols_state = OLS_GRANTED;
994 oio->oi_lockless = 1;
999 oscl->ols_state = OLS_ENQUEUED;
1000 if (anchor != NULL) {
1001 atomic_inc(&anchor->csi_sync_nr);
1002 oscl->ols_owner = anchor;
1006 * DLM lock's ast data must be osc_object;
1007 * if glimpse or speculative lock, async of osc_enqueue_base()
1010 * For non-speculative locks:
1011 * DLM's enqueue callback set to osc_lock_upcall() with cookie as
1013 * For speculative locks:
1014 * osc_lock_upcall_speculative & cookie is the osc object, since
1015 * there is no osc_lock
1017 ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
1018 osc_lock_build_policy(env, lock, policy);
1019 if (oscl->ols_speculative) {
1020 oscl->ols_einfo.ei_cbdata = NULL;
1021 /* hold a reference for callback */
1022 cl_object_get(osc2cl(osc));
1023 upcall = osc_lock_upcall_speculative;
1026 result = osc_enqueue_base(exp, resname, &oscl->ols_flags,
1027 policy, &oscl->ols_lvb,
1029 &oscl->ols_einfo, PTLRPCD_SET, async,
1030 oscl->ols_speculative);
1032 if (osc_lock_is_lockless(oscl)) {
1033 oio->oi_lockless = 1;
1034 } else if (!async) {
1035 if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_IDLE_RACE)) {
1036 OBD_RACE(OBD_FAIL_PTLRPC_IDLE_RACE);
1037 set_current_state(TASK_UNINTERRUPTIBLE);
1038 schedule_timeout(cfs_time_seconds(1) / 2);
1040 LASSERT(oscl->ols_state == OLS_GRANTED);
1041 LASSERT(oscl->ols_hold);
1042 LASSERT(oscl->ols_dlmlock != NULL);
1044 } else if (oscl->ols_speculative) {
1045 cl_object_put(env, osc2cl(osc));
1046 if (oscl->ols_glimpse) {
1047 /* hide error for AGL request */
1054 oscl->ols_state = OLS_CANCELLED;
1055 osc_lock_wake_waiters(env, osc, oscl);
1058 cl_sync_io_note(env, anchor, result);
1064 * Breaks a link between osc_lock and dlm_lock.
1066 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
1068 struct ldlm_lock *dlmlock;
1072 dlmlock = olck->ols_dlmlock;
1073 if (dlmlock == NULL)
1076 if (olck->ols_hold) {
1078 ldlm_lock_decref(&olck->ols_handle, olck->ols_einfo.ei_mode);
1079 olck->ols_handle.cookie = 0ULL;
1082 olck->ols_dlmlock = NULL;
1084 /* release a reference taken in osc_lock_upcall(). */
1085 LASSERT(olck->ols_has_ref);
1086 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
1087 LDLM_LOCK_RELEASE(dlmlock);
1088 olck->ols_has_ref = 0;
1094 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1095 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1096 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1097 * with some other lock some where in the cluster. This function does the
1100 * - invalidates all pages protected by this lock (after sending dirty
1101 * ones to the server, as necessary);
1103 * - decref's underlying ldlm lock;
1105 * - cancels ldlm lock (ldlm_cli_cancel()).
1107 void osc_lock_cancel(const struct lu_env *env,
1108 const struct cl_lock_slice *slice)
1110 struct osc_object *obj = cl2osc(slice->cls_obj);
1111 struct osc_lock *oscl = cl2osc_lock(slice);
1115 LINVRNT(osc_lock_invariant(oscl));
1117 osc_lock_detach(env, oscl);
1118 oscl->ols_state = OLS_CANCELLED;
1119 oscl->ols_flags &= ~LDLM_FL_LVB_READY;
1121 osc_lock_wake_waiters(env, obj, oscl);
1124 EXPORT_SYMBOL(osc_lock_cancel);
1126 int osc_lock_print(const struct lu_env *env, void *cookie,
1127 lu_printer_t p, const struct cl_lock_slice *slice)
1129 struct osc_lock *lock = cl2osc_lock(slice);
1131 (*p)(env, cookie, "%p %#llx %#llx %d %p ",
1132 lock->ols_dlmlock, lock->ols_flags, lock->ols_handle.cookie,
1133 lock->ols_state, lock->ols_owner);
1134 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1137 EXPORT_SYMBOL(osc_lock_print);
1139 static const struct cl_lock_operations osc_lock_ops = {
1140 .clo_fini = osc_lock_fini,
1141 .clo_enqueue = osc_lock_enqueue,
1142 .clo_cancel = osc_lock_cancel,
1143 .clo_print = osc_lock_print,
1146 static void osc_lock_lockless_cancel(const struct lu_env *env,
1147 const struct cl_lock_slice *slice)
1149 struct osc_lock *ols = cl2osc_lock(slice);
1150 struct osc_object *osc = cl2osc(slice->cls_obj);
1152 LASSERT(ols->ols_dlmlock == NULL);
1153 osc_lock_wake_waiters(env, osc, ols);
1156 static const struct cl_lock_operations osc_lock_lockless_ops = {
1157 .clo_fini = osc_lock_fini,
1158 .clo_enqueue = osc_lock_enqueue,
1159 .clo_cancel = osc_lock_lockless_cancel,
1160 .clo_print = osc_lock_print
1163 void osc_lock_set_writer(const struct lu_env *env, const struct cl_io *io,
1164 struct cl_object *obj, struct osc_lock *oscl)
1166 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
1170 if (!cl_object_same(io->ci_obj, obj))
1173 if (likely(io->ci_type == CIT_WRITE)) {
1174 io_start = cl_index(obj, io->u.ci_rw.crw_pos);
1175 io_end = cl_index(obj, io->u.ci_rw.crw_pos +
1176 io->u.ci_rw.crw_count - 1);
1178 LASSERT(cl_io_is_mkwrite(io));
1179 io_start = io_end = io->u.ci_fault.ft_index;
1182 if (descr->cld_mode >= CLM_WRITE &&
1183 (cl_io_is_append(io) ||
1184 (descr->cld_start <= io_start && descr->cld_end >= io_end))) {
1185 struct osc_io *oio = osc_env_io(env);
1187 /* There must be only one lock to match the write region */
1188 LASSERT(oio->oi_write_osclock == NULL);
1189 oio->oi_write_osclock = oscl;
1192 EXPORT_SYMBOL(osc_lock_set_writer);
1194 int osc_lock_init(const struct lu_env *env,
1195 struct cl_object *obj, struct cl_lock *lock,
1196 const struct cl_io *io)
1198 struct osc_lock *oscl;
1199 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1201 OBD_SLAB_ALLOC_PTR_GFP(oscl, osc_lock_kmem, GFP_NOFS);
1205 oscl->ols_state = OLS_NEW;
1206 spin_lock_init(&oscl->ols_lock);
1207 INIT_LIST_HEAD(&oscl->ols_waiting_list);
1208 INIT_LIST_HEAD(&oscl->ols_wait_entry);
1209 INIT_LIST_HEAD(&oscl->ols_nextlock_oscobj);
1210 oscl->ols_lockless_ops = &osc_lock_lockless_ops;
1212 /* Speculative lock requests must be either no_expand or glimpse
1213 * request (CEF_GLIMPSE). non-glimpse no_expand speculative extent
1214 * locks will break ofd_intent_cb. (see comment there)*/
1215 LASSERT(ergo((enqflags & CEF_SPECULATIVE) != 0,
1216 (enqflags & (CEF_LOCK_NO_EXPAND | CEF_GLIMPSE)) != 0));
1218 oscl->ols_flags = osc_enq2ldlm_flags(enqflags);
1219 oscl->ols_speculative = !!(enqflags & CEF_SPECULATIVE);
1220 if (lock->cll_descr.cld_mode == CLM_GROUP)
1221 oscl->ols_flags |= LDLM_FL_ATOMIC_CB;
1223 if (oscl->ols_flags & LDLM_FL_HAS_INTENT) {
1224 oscl->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1225 oscl->ols_glimpse = 1;
1227 if (io->ci_ndelay && cl_object_same(io->ci_obj, obj))
1228 oscl->ols_flags |= LDLM_FL_NDELAY;
1229 osc_lock_build_einfo(env, lock, cl2osc(obj), &oscl->ols_einfo);
1231 cl_lock_slice_add(lock, &oscl->ols_cl, obj, &osc_lock_ops);
1233 if (!(enqflags & CEF_MUST))
1234 /* try to convert this lock to a lockless lock */
1235 osc_lock_to_lockless(env, oscl, (enqflags & CEF_NEVER));
1237 if (io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io))
1238 osc_lock_set_writer(env, io, obj, oscl);
1240 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %#llx",
1241 lock, oscl, oscl->ols_flags);
1247 * Finds an existing lock covering given index and optionally different from a
1248 * given \a except lock.
1250 struct ldlm_lock *osc_obj_dlmlock_at_pgoff(const struct lu_env *env,
1251 struct osc_object *obj,
1253 enum osc_dap_flags dap_flags)
1255 struct osc_thread_info *info = osc_env_info(env);
1256 struct ldlm_res_id *resname = &info->oti_resname;
1257 union ldlm_policy_data *policy = &info->oti_policy;
1258 struct lustre_handle lockh;
1259 struct ldlm_lock *lock = NULL;
1260 enum ldlm_mode mode;
1262 enum ldlm_match_flags match_flags = 0;
1266 ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1267 osc_index2policy(policy, osc2cl(obj), index, index);
1268 policy->l_extent.gid = LDLM_GID_ANY;
1270 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1271 if (dap_flags & OSC_DAP_FL_TEST_LOCK)
1272 flags |= LDLM_FL_TEST_LOCK;
1274 if (dap_flags & OSC_DAP_FL_AST)
1275 match_flags |= LDLM_MATCH_AST;
1277 if (dap_flags & OSC_DAP_FL_CANCELING)
1278 match_flags |= LDLM_MATCH_UNREF;
1280 if (dap_flags & OSC_DAP_FL_RIGHT)
1281 match_flags |= LDLM_MATCH_RIGHT;
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(env, osc_export(obj), resname, LDLM_EXTENT,
1289 policy, LCK_PR | LCK_PW | LCK_GROUP, &flags,
1290 obj, &lockh, match_flags);
1292 lock = ldlm_handle2lock(&lockh);
1293 /* RACE: the lock is cancelled so let's try again */
1294 if (unlikely(lock == NULL))