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 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;
157 unsigned valid, setkms = 0;
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) {
183 attr->cat_kms = size;
186 ldlm_lock_allow_match_locked(dlmlock);
189 /* The size should not be less than the kms */
190 if (attr->cat_size < oinfo->loi_kms)
191 attr->cat_size = oinfo->loi_kms;
193 LDLM_DEBUG(dlmlock, "acquired size %llu, setting rss=%llu;%s "
194 "kms=%llu, end=%llu", lvb->lvb_size, attr->cat_size,
195 setkms ? "" : " leaving",
196 setkms ? attr->cat_kms : oinfo->loi_kms,
197 dlmlock ? dlmlock->l_policy_data.l_extent.end : -1ull);
199 cl_object_attr_update(env, obj, attr, valid);
200 cl_object_attr_unlock(obj);
205 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *oscl,
206 struct lustre_handle *lockh)
208 struct osc_object *osc = cl2osc(oscl->ols_cl.cls_obj);
209 struct ldlm_lock *dlmlock;
211 dlmlock = ldlm_handle2lock_long(lockh, 0);
212 LASSERT(dlmlock != NULL);
214 /* lock reference taken by ldlm_handle2lock_long() is
215 * owned by osc_lock and released in osc_lock_detach()
217 lu_ref_add_atomic(&dlmlock->l_reference, "osc_lock", oscl);
218 oscl->ols_has_ref = 1;
220 LASSERT(oscl->ols_dlmlock == NULL);
221 oscl->ols_dlmlock = dlmlock;
223 /* This may be a matched lock for glimpse request, do not hold
224 * lock reference in that case. */
225 if (!oscl->ols_glimpse) {
226 /* hold a refc for non glimpse lock which will
227 * be released in osc_lock_cancel() */
228 lustre_handle_copy(&oscl->ols_handle, lockh);
229 ldlm_lock_addref(lockh, oscl->ols_einfo.ei_mode);
233 /* Lock must have been granted. */
234 lock_res_and_lock(dlmlock);
235 if (ldlm_is_granted(dlmlock)) {
236 struct ldlm_extent *ext = &dlmlock->l_policy_data.l_extent;
237 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
239 /* extend the lock extent, otherwise it will have problem when
240 * we decide whether to grant a lockless lock. */
241 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
242 descr->cld_start = cl_index(descr->cld_obj, ext->start);
243 descr->cld_end = cl_index(descr->cld_obj, ext->end);
244 descr->cld_gid = ext->gid;
246 /* no lvb update for matched lock */
247 if (!ldlm_is_lvb_cached(dlmlock)) {
248 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
249 LASSERT(osc == dlmlock->l_ast_data);
250 osc_lock_lvb_update(env, osc, dlmlock, NULL);
251 ldlm_set_lvb_cached(dlmlock);
253 LINVRNT(osc_lock_invariant(oscl));
255 unlock_res_and_lock(dlmlock);
257 LASSERT(oscl->ols_state != OLS_GRANTED);
258 oscl->ols_state = OLS_GRANTED;
262 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
263 * received from a server, or after osc_enqueue_base() matched a local DLM
266 static int osc_lock_upcall(void *cookie, struct lustre_handle *lockh,
269 struct osc_lock *oscl = cookie;
270 struct cl_lock_slice *slice = &oscl->ols_cl;
276 env = cl_env_percpu_get();
277 /* should never happen, similar to osc_ldlm_blocking_ast(). */
278 LASSERT(!IS_ERR(env));
280 rc = ldlm_error2errno(errcode);
281 if (oscl->ols_state == OLS_ENQUEUED) {
282 oscl->ols_state = OLS_UPCALL_RECEIVED;
283 } else if (oscl->ols_state == OLS_CANCELLED) {
286 CERROR("Impossible state: %d\n", oscl->ols_state);
291 osc_lock_granted(env, oscl, lockh);
293 /* Error handling, some errors are tolerable. */
294 if (oscl->ols_locklessable && rc == -EUSERS) {
295 /* This is a tolerable error, turn this lock into
298 osc_object_set_contended(cl2osc(slice->cls_obj));
299 LASSERT(slice->cls_ops != oscl->ols_lockless_ops);
301 /* Change this lock to ldlmlock-less lock. */
302 osc_lock_to_lockless(env, oscl, 1);
303 oscl->ols_state = OLS_GRANTED;
305 } else if (oscl->ols_glimpse && rc == -ENAVAIL) {
306 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
307 osc_lock_lvb_update(env, cl2osc(slice->cls_obj),
308 NULL, &oscl->ols_lvb);
309 /* Hide the error. */
311 } else if (rc < 0 && oscl->ols_flags & LDLM_FL_NDELAY) {
315 if (oscl->ols_owner != NULL)
316 cl_sync_io_note(env, oscl->ols_owner, rc);
317 cl_env_percpu_put(env);
322 static int osc_lock_upcall_speculative(void *cookie,
323 struct lustre_handle *lockh,
326 struct osc_object *osc = cookie;
327 struct ldlm_lock *dlmlock;
332 env = cl_env_get(&refcheck);
333 LASSERT(!IS_ERR(env));
335 if (errcode == ELDLM_LOCK_MATCHED)
336 GOTO(out, errcode = ELDLM_OK);
338 if (errcode != ELDLM_OK)
341 dlmlock = ldlm_handle2lock(lockh);
342 LASSERT(dlmlock != NULL);
344 lock_res_and_lock(dlmlock);
345 LASSERT(ldlm_is_granted(dlmlock));
347 /* there is no osc_lock associated with speculative locks
348 * thus no need to set LDLM_FL_LVB_CACHED */
349 osc_lock_lvb_update(env, osc, dlmlock, NULL);
351 unlock_res_and_lock(dlmlock);
352 LDLM_LOCK_PUT(dlmlock);
355 cl_object_put(env, osc2cl(osc));
356 cl_env_put(env, &refcheck);
357 RETURN(ldlm_error2errno(errcode));
360 static int osc_lock_flush(struct osc_object *obj, pgoff_t start, pgoff_t end,
361 enum cl_lock_mode mode, bool discard)
370 env = cl_env_get(&refcheck);
372 RETURN(PTR_ERR(env));
374 if (mode == CLM_WRITE) {
375 rc = osc_cache_writeback_range(env, obj, start, end, 1,
377 CDEBUG(D_CACHE, "object %p: [%lu -> %lu] %d pages were %s.\n",
379 discard ? "discarded" : "written back");
385 * Do not try to match other locks with CLM_WRITE since we already
388 rc2 = osc_lock_discard_pages(env, obj, start, end,
389 mode == CLM_WRITE || discard);
390 if (rc == 0 && rc2 < 0)
393 cl_env_put(env, &refcheck);
398 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
399 * and ldlm_lock caches.
401 static int osc_dlm_blocking_ast0(const struct lu_env *env,
402 struct ldlm_lock *dlmlock,
403 void *data, int flag)
405 struct cl_object *obj = NULL;
408 enum cl_lock_mode mode = CLM_READ;
411 LASSERT(flag == LDLM_CB_CANCELING);
413 lock_res_and_lock(dlmlock);
414 if (!ldlm_is_granted(dlmlock)) {
415 dlmlock->l_ast_data = NULL;
416 unlock_res_and_lock(dlmlock);
420 discard = ldlm_is_discard_data(dlmlock);
421 if (dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))
424 if (dlmlock->l_ast_data != NULL) {
425 obj = osc2cl(dlmlock->l_ast_data);
429 unlock_res_and_lock(dlmlock);
431 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_CANCEL, 5);
433 /* if l_ast_data is NULL, the dlmlock was enqueued by AGL or
434 * the object has been destroyed. */
436 struct ldlm_extent *extent = &dlmlock->l_policy_data.l_extent;
437 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
440 /* Destroy pages covered by the extent of the DLM lock */
441 result = osc_lock_flush(cl2osc(obj),
442 cl_index(obj, extent->start),
443 cl_index(obj, extent->end),
446 /* losing a lock, update kms */
447 lock_res_and_lock(dlmlock);
448 /* clearing l_ast_data after flushing data,
449 * to let glimpse ast find the lock and the object */
450 dlmlock->l_ast_data = NULL;
451 cl_object_attr_lock(obj);
452 /* Must get the value under the lock to avoid race. */
453 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
454 /* Update the kms. Need to loop all granted locks.
455 * Not a problem for the client */
456 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
458 cl_object_attr_update(env, obj, attr, CAT_KMS);
459 cl_object_attr_unlock(obj);
460 unlock_res_and_lock(dlmlock);
462 cl_object_put(env, obj);
468 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
469 * some other lock, or is canceled. This function is installed as a
470 * ldlm_lock::l_blocking_ast() for client extent locks.
472 * Control flow is tricky, because ldlm uses the same call-back
473 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
475 * \param dlmlock lock for which ast occurred.
477 * \param new description of a conflicting lock in case of blocking ast.
479 * \param data value of dlmlock->l_ast_data
481 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
482 * cancellation and blocking ast's.
484 * Possible use cases:
486 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
487 * lock due to lock lru pressure, or explicit user request to purge
490 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
491 * us that dlmlock conflicts with another lock that some client is
492 * enqueuing. Lock is canceled.
494 * - cl_lock_cancel() is called. osc_lock_cancel() calls
495 * ldlm_cli_cancel() that calls
497 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
499 * recursively entering osc_ldlm_blocking_ast().
501 * - client cancels lock voluntary (e.g., as a part of early cancellation):
504 * osc_lock_cancel()->
505 * ldlm_cli_cancel()->
506 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
509 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
510 struct ldlm_lock_desc *new, void *data,
517 case LDLM_CB_BLOCKING: {
518 struct lustre_handle lockh;
520 ldlm_lock2handle(dlmlock, &lockh);
521 result = ldlm_cli_cancel(&lockh, LCF_ASYNC);
522 if (result == -ENODATA)
526 case LDLM_CB_CANCELING: {
531 * This can be called in the context of outer IO, e.g.,
533 * osc_enqueue_base()->...
534 * ->ldlm_prep_elc_req()->...
535 * ->ldlm_cancel_callback()->...
536 * ->osc_ldlm_blocking_ast()
538 * new environment has to be created to not corrupt outer
541 env = cl_env_get(&refcheck);
543 result = PTR_ERR(env);
547 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
548 cl_env_put(env, &refcheck);
557 int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
559 struct ptlrpc_request *req = data;
562 struct req_capsule *cap;
563 struct cl_object *obj = NULL;
564 struct ldlm_resource *res = dlmlock->l_resource;
565 struct ldlm_match_data matchdata = { 0 };
566 union ldlm_policy_data policy;
567 enum ldlm_mode mode = LCK_PW | LCK_GROUP | LCK_PR;
573 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
575 env = cl_env_get(&refcheck);
577 GOTO(out, result = PTR_ERR(env));
579 policy.l_extent.start = 0;
580 policy.l_extent.end = LUSTRE_EOF;
582 matchdata.lmd_mode = &mode;
583 matchdata.lmd_policy = &policy;
584 matchdata.lmd_flags = LDLM_FL_TEST_LOCK | LDLM_FL_CBPENDING;
585 matchdata.lmd_unref = 1;
586 matchdata.lmd_has_ast_data = true;
588 LDLM_LOCK_GET(dlmlock);
590 /* If any dlmlock has l_ast_data set, we must find it or we risk
591 * missing a size update done under a different lock.
594 lock_res_and_lock(dlmlock);
595 if (dlmlock->l_ast_data) {
596 obj = osc2cl(dlmlock->l_ast_data);
599 unlock_res_and_lock(dlmlock);
600 LDLM_LOCK_RELEASE(dlmlock);
604 if (obj == NULL && res->lr_type == LDLM_EXTENT) {
605 if (OBD_FAIL_CHECK(OBD_FAIL_OSC_NO_SIZE_DATA))
609 dlmlock = search_itree(res, &matchdata);
615 /* Do not grab the mutex of cl_lock for glimpse.
616 * See LU-1274 for details.
617 * BTW, it's okay for cl_lock to be cancelled during
618 * this period because server can handle this race.
619 * See ldlm_server_glimpse_ast() for details.
620 * cl_lock_mutex_get(env, lock); */
622 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
623 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
625 result = req_capsule_server_pack(cap);
627 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
628 result = cl_object_glimpse(env, obj, lvb);
630 if (!exp_connect_lvb_type(req->rq_export))
631 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
632 sizeof(struct ost_lvb_v1), RCL_SERVER);
633 cl_object_put(env, obj);
636 * These errors are normal races, so we don't want to
637 * fill the console with messages by calling
640 lustre_pack_reply(req, 1, NULL, NULL);
641 result = -ELDLM_NO_LOCK_DATA;
643 cl_env_put(env, &refcheck);
647 req->rq_status = result;
650 EXPORT_SYMBOL(osc_ldlm_glimpse_ast);
652 static bool weigh_cb(const struct lu_env *env, struct cl_io *io,
653 struct osc_page *ops, void *cbdata)
655 struct cl_page *page = ops->ops_cl.cpl_page;
657 if (cl_page_is_vmlocked(env, page) || PageDirty(page->cp_vmpage) ||
658 PageWriteback(page->cp_vmpage))
661 *(pgoff_t *)cbdata = osc_index(ops) + 1;
665 static unsigned long osc_lock_weight(const struct lu_env *env,
666 struct osc_object *oscobj,
667 loff_t start, loff_t end)
669 struct cl_io *io = osc_env_thread_io(env);
670 struct cl_object *obj = cl_object_top(&oscobj->oo_cl);
677 io->ci_ignore_layout = 1;
678 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
682 page_index = cl_index(obj, start);
684 if (!osc_page_gang_lookup(env, io, oscobj,
685 page_index, cl_index(obj, end),
686 weigh_cb, (void *)&page_index))
694 * Get the weight of dlm lock for early cancellation.
696 unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
699 struct osc_object *obj;
700 struct osc_lock *oscl;
701 unsigned long weight;
709 * osc_ldlm_weigh_ast has a complex context since it might be called
710 * because of lock canceling, or from user's input. We have to make
711 * a new environment for it. Probably it is implementation safe to use
712 * the upper context because cl_lock_put don't modify environment
713 * variables. But just in case ..
715 env = cl_env_get(&refcheck);
717 /* Mostly because lack of memory, do not eliminate this lock */
720 LASSERT(dlmlock->l_resource->lr_type == LDLM_EXTENT ||
721 dlmlock->l_resource->lr_type == LDLM_IBITS);
723 lock_res_and_lock(dlmlock);
724 obj = dlmlock->l_ast_data;
726 cl_object_get(osc2cl(obj));
727 unlock_res_and_lock(dlmlock);
730 GOTO(out, weight = 1);
732 spin_lock(&obj->oo_ol_spin);
733 list_for_each_entry(oscl, &obj->oo_ol_list, ols_nextlock_oscobj) {
734 if (oscl->ols_dlmlock == dlmlock) {
739 spin_unlock(&obj->oo_ol_spin);
742 * If the lock is being used by an IO, definitely not cancel it.
744 GOTO(out, weight = 1);
747 if (dlmlock->l_resource->lr_type == LDLM_EXTENT)
748 weight = osc_lock_weight(env, obj,
749 dlmlock->l_policy_data.l_extent.start,
750 dlmlock->l_policy_data.l_extent.end);
751 else if (ldlm_has_dom(dlmlock))
752 weight = osc_lock_weight(env, obj, 0, OBD_OBJECT_EOF);
753 /* The DOM bit can be cancelled at any time; in that case, we know
754 * there are no pages, so just return weight of 0
763 cl_object_put(env, osc2cl(obj));
765 cl_env_put(env, &refcheck);
768 EXPORT_SYMBOL(osc_ldlm_weigh_ast);
770 static void osc_lock_build_einfo(const struct lu_env *env,
771 const struct cl_lock *lock,
772 struct osc_object *osc,
773 struct ldlm_enqueue_info *einfo)
775 einfo->ei_type = LDLM_EXTENT;
776 einfo->ei_mode = osc_cl_lock2ldlm(lock->cll_descr.cld_mode);
777 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
778 einfo->ei_cb_cp = ldlm_completion_ast;
779 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
780 einfo->ei_cbdata = osc; /* value to be put into ->l_ast_data */
784 * Determine if the lock should be converted into a lockless lock.
787 * - if the lock has an explicite requirment for a non-lockless lock;
788 * - if the io lock request type ci_lockreq;
789 * - send the enqueue rpc to ost to make the further decision;
790 * - special treat to truncate lockless lock
792 * Additional policy can be implemented here, e.g., never do lockless-io
795 void osc_lock_to_lockless(const struct lu_env *env,
796 struct osc_lock *ols, int force)
798 struct cl_lock_slice *slice = &ols->ols_cl;
799 struct osc_io *oio = osc_env_io(env);
800 struct cl_io *io = oio->oi_cl.cis_io;
801 struct cl_object *obj = slice->cls_obj;
802 struct osc_object *oob = cl2osc(obj);
803 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
804 struct obd_connect_data *ocd;
806 LASSERT(ols->ols_state == OLS_NEW ||
807 ols->ols_state == OLS_UPCALL_RECEIVED);
810 ols->ols_locklessable = 1;
811 slice->cls_ops = ols->ols_lockless_ops;
813 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
814 io->ci_lockreq == CILR_MAYBE ||
815 io->ci_lockreq == CILR_NEVER);
817 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
818 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
819 (io->ci_lockreq == CILR_MAYBE) &&
820 (ocd->ocd_connect_flags &
821 OBD_CONNECT_SRVLOCK);
822 if (io->ci_lockreq == CILR_NEVER ||
824 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
825 /* lockless truncate */
826 (cl_io_is_trunc(io) && osd->od_lockless_truncate &&
827 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK))) {
828 ols->ols_locklessable = 1;
829 slice->cls_ops = ols->ols_lockless_ops;
832 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
834 EXPORT_SYMBOL(osc_lock_to_lockless);
836 static bool osc_lock_compatible(const struct osc_lock *qing,
837 const struct osc_lock *qed)
839 struct cl_lock_descr *qed_descr = &qed->ols_cl.cls_lock->cll_descr;
840 struct cl_lock_descr *qing_descr = &qing->ols_cl.cls_lock->cll_descr;
842 if (qed->ols_glimpse || qed->ols_speculative)
845 if (qing_descr->cld_mode == CLM_READ && qed_descr->cld_mode == CLM_READ)
848 if (qed->ols_state < OLS_GRANTED)
851 if (qed_descr->cld_mode >= qing_descr->cld_mode &&
852 qed_descr->cld_start <= qing_descr->cld_start &&
853 qed_descr->cld_end >= qing_descr->cld_end)
859 void osc_lock_wake_waiters(const struct lu_env *env, struct osc_object *osc,
860 struct osc_lock *oscl)
862 spin_lock(&osc->oo_ol_spin);
863 list_del_init(&oscl->ols_nextlock_oscobj);
864 spin_unlock(&osc->oo_ol_spin);
866 spin_lock(&oscl->ols_lock);
867 while (!list_empty(&oscl->ols_waiting_list)) {
868 struct osc_lock *scan;
870 scan = list_entry(oscl->ols_waiting_list.next, struct osc_lock,
872 list_del_init(&scan->ols_wait_entry);
874 cl_sync_io_note(env, scan->ols_owner, 0);
876 spin_unlock(&oscl->ols_lock);
878 EXPORT_SYMBOL(osc_lock_wake_waiters);
880 int osc_lock_enqueue_wait(const struct lu_env *env, struct osc_object *obj,
881 struct osc_lock *oscl)
883 struct osc_lock *tmp_oscl;
884 struct cl_lock_descr *need = &oscl->ols_cl.cls_lock->cll_descr;
885 struct cl_sync_io *waiter = &osc_env_info(env)->oti_anchor;
890 spin_lock(&obj->oo_ol_spin);
891 list_add_tail(&oscl->ols_nextlock_oscobj, &obj->oo_ol_list);
894 list_for_each_entry(tmp_oscl, &obj->oo_ol_list,
895 ols_nextlock_oscobj) {
896 struct cl_lock_descr *descr;
898 if (tmp_oscl == oscl)
901 descr = &tmp_oscl->ols_cl.cls_lock->cll_descr;
902 if (descr->cld_start > need->cld_end ||
903 descr->cld_end < need->cld_start)
906 /* We're not supposed to give up group lock */
907 if (descr->cld_mode == CLM_GROUP)
910 if (!osc_lock_is_lockless(oscl) &&
911 osc_lock_compatible(oscl, tmp_oscl))
914 /* wait for conflicting lock to be canceled */
915 cl_sync_io_init(waiter, 1);
916 oscl->ols_owner = waiter;
918 spin_lock(&tmp_oscl->ols_lock);
919 /* add oscl into tmp's ols_waiting list */
920 list_add_tail(&oscl->ols_wait_entry,
921 &tmp_oscl->ols_waiting_list);
922 spin_unlock(&tmp_oscl->ols_lock);
924 spin_unlock(&obj->oo_ol_spin);
925 rc = cl_sync_io_wait(env, waiter, 0);
926 spin_lock(&obj->oo_ol_spin);
931 oscl->ols_owner = NULL;
934 spin_unlock(&obj->oo_ol_spin);
938 EXPORT_SYMBOL(osc_lock_enqueue_wait);
941 * Implementation of cl_lock_operations::clo_enqueue() method for osc
942 * layer. This initiates ldlm enqueue:
944 * - cancels conflicting locks early (osc_lock_enqueue_wait());
946 * - calls osc_enqueue_base() to do actual enqueue.
948 * osc_enqueue_base() is supplied with an upcall function that is executed
949 * when lock is received either after a local cached ldlm lock is matched, or
950 * when a reply from the server is received.
952 * This function does not wait for the network communication to complete.
954 static int osc_lock_enqueue(const struct lu_env *env,
955 const struct cl_lock_slice *slice,
956 struct cl_io *unused, struct cl_sync_io *anchor)
958 struct osc_thread_info *info = osc_env_info(env);
959 struct osc_io *oio = osc_env_io(env);
960 struct osc_object *osc = cl2osc(slice->cls_obj);
961 struct osc_lock *oscl = cl2osc_lock(slice);
962 struct obd_export *exp = osc_export(osc);
963 struct cl_lock *lock = slice->cls_lock;
964 struct ldlm_res_id *resname = &info->oti_resname;
965 union ldlm_policy_data *policy = &info->oti_policy;
966 osc_enqueue_upcall_f upcall = osc_lock_upcall;
973 LASSERTF(ergo(oscl->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
974 "lock = %p, ols = %p\n", lock, oscl);
976 if (oscl->ols_state == OLS_GRANTED)
979 if ((oscl->ols_flags & LDLM_FL_NO_EXPANSION) &&
980 !exp_connect_lockahead(exp)) {
981 result = -EOPNOTSUPP;
982 CERROR("%s: server does not support lockahead/locknoexpand: rc = %d\n",
983 exp->exp_obd->obd_name, result);
987 if (oscl->ols_flags & LDLM_FL_TEST_LOCK)
988 GOTO(enqueue_base, 0);
990 /* For glimpse and/or speculative locks, do not wait for reply from
991 * server on LDLM request */
992 if (oscl->ols_glimpse || oscl->ols_speculative) {
993 /* Speculative and glimpse locks do not have an anchor */
994 LASSERT(equi(oscl->ols_speculative, anchor == NULL));
996 GOTO(enqueue_base, 0);
999 result = osc_lock_enqueue_wait(env, osc, oscl);
1003 /* we can grant lockless lock right after all conflicting locks
1005 if (osc_lock_is_lockless(oscl)) {
1006 oscl->ols_state = OLS_GRANTED;
1007 oio->oi_lockless = 1;
1012 oscl->ols_state = OLS_ENQUEUED;
1013 if (anchor != NULL) {
1014 atomic_inc(&anchor->csi_sync_nr);
1015 oscl->ols_owner = anchor;
1019 * DLM lock's ast data must be osc_object;
1020 * if glimpse or speculative lock, async of osc_enqueue_base()
1023 * For non-speculative locks:
1024 * DLM's enqueue callback set to osc_lock_upcall() with cookie as
1026 * For speculative locks:
1027 * osc_lock_upcall_speculative & cookie is the osc object, since
1028 * there is no osc_lock
1030 ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
1031 osc_lock_build_policy(env, lock, policy);
1032 if (oscl->ols_speculative) {
1033 oscl->ols_einfo.ei_cbdata = NULL;
1034 /* hold a reference for callback */
1035 cl_object_get(osc2cl(osc));
1036 upcall = osc_lock_upcall_speculative;
1039 result = osc_enqueue_base(exp, resname, &oscl->ols_flags,
1040 policy, &oscl->ols_lvb,
1042 &oscl->ols_einfo, PTLRPCD_SET, async,
1043 oscl->ols_speculative);
1045 if (osc_lock_is_lockless(oscl)) {
1046 oio->oi_lockless = 1;
1047 } else if (!async) {
1048 LASSERT(oscl->ols_state == OLS_GRANTED);
1049 LASSERT(oscl->ols_hold);
1050 LASSERT(oscl->ols_dlmlock != NULL);
1052 } else if (oscl->ols_speculative) {
1053 cl_object_put(env, osc2cl(osc));
1054 if (oscl->ols_glimpse) {
1055 /* hide error for AGL request */
1062 oscl->ols_state = OLS_CANCELLED;
1063 osc_lock_wake_waiters(env, osc, oscl);
1066 cl_sync_io_note(env, anchor, result);
1072 * Breaks a link between osc_lock and dlm_lock.
1074 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
1076 struct ldlm_lock *dlmlock;
1080 dlmlock = olck->ols_dlmlock;
1081 if (dlmlock == NULL)
1084 if (olck->ols_hold) {
1086 ldlm_lock_decref(&olck->ols_handle, olck->ols_einfo.ei_mode);
1087 olck->ols_handle.cookie = 0ULL;
1090 olck->ols_dlmlock = NULL;
1092 /* release a reference taken in osc_lock_upcall(). */
1093 LASSERT(olck->ols_has_ref);
1094 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
1095 LDLM_LOCK_RELEASE(dlmlock);
1096 olck->ols_has_ref = 0;
1102 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1103 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1104 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1105 * with some other lock some where in the cluster. This function does the
1108 * - invalidates all pages protected by this lock (after sending dirty
1109 * ones to the server, as necessary);
1111 * - decref's underlying ldlm lock;
1113 * - cancels ldlm lock (ldlm_cli_cancel()).
1115 void osc_lock_cancel(const struct lu_env *env,
1116 const struct cl_lock_slice *slice)
1118 struct osc_object *obj = cl2osc(slice->cls_obj);
1119 struct osc_lock *oscl = cl2osc_lock(slice);
1123 LINVRNT(osc_lock_invariant(oscl));
1125 osc_lock_detach(env, oscl);
1126 oscl->ols_state = OLS_CANCELLED;
1127 oscl->ols_flags &= ~LDLM_FL_LVB_READY;
1129 osc_lock_wake_waiters(env, obj, oscl);
1132 EXPORT_SYMBOL(osc_lock_cancel);
1134 int osc_lock_print(const struct lu_env *env, void *cookie,
1135 lu_printer_t p, const struct cl_lock_slice *slice)
1137 struct osc_lock *lock = cl2osc_lock(slice);
1139 (*p)(env, cookie, "%p %#llx %#llx %d %p ",
1140 lock->ols_dlmlock, lock->ols_flags, lock->ols_handle.cookie,
1141 lock->ols_state, lock->ols_owner);
1142 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1145 EXPORT_SYMBOL(osc_lock_print);
1147 static const struct cl_lock_operations osc_lock_ops = {
1148 .clo_fini = osc_lock_fini,
1149 .clo_enqueue = osc_lock_enqueue,
1150 .clo_cancel = osc_lock_cancel,
1151 .clo_print = osc_lock_print,
1154 static void osc_lock_lockless_cancel(const struct lu_env *env,
1155 const struct cl_lock_slice *slice)
1157 struct osc_lock *ols = cl2osc_lock(slice);
1158 struct osc_object *osc = cl2osc(slice->cls_obj);
1159 struct cl_lock_descr *descr = &slice->cls_lock->cll_descr;
1162 LASSERT(ols->ols_dlmlock == NULL);
1163 result = osc_lock_flush(osc, descr->cld_start, descr->cld_end,
1164 descr->cld_mode, false);
1166 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1169 osc_lock_wake_waiters(env, osc, ols);
1172 static const struct cl_lock_operations osc_lock_lockless_ops = {
1173 .clo_fini = osc_lock_fini,
1174 .clo_enqueue = osc_lock_enqueue,
1175 .clo_cancel = osc_lock_lockless_cancel,
1176 .clo_print = osc_lock_print
1179 void osc_lock_set_writer(const struct lu_env *env, const struct cl_io *io,
1180 struct cl_object *obj, struct osc_lock *oscl)
1182 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
1186 if (!cl_object_same(io->ci_obj, obj))
1189 if (likely(io->ci_type == CIT_WRITE)) {
1190 io_start = cl_index(obj, io->u.ci_rw.crw_pos);
1191 io_end = cl_index(obj, io->u.ci_rw.crw_pos +
1192 io->u.ci_rw.crw_count - 1);
1194 LASSERT(cl_io_is_mkwrite(io));
1195 io_start = io_end = io->u.ci_fault.ft_index;
1198 if (descr->cld_mode >= CLM_WRITE &&
1199 (cl_io_is_append(io) ||
1200 (descr->cld_start <= io_start && descr->cld_end >= io_end))) {
1201 struct osc_io *oio = osc_env_io(env);
1203 /* There must be only one lock to match the write region */
1204 LASSERT(oio->oi_write_osclock == NULL);
1205 oio->oi_write_osclock = oscl;
1208 EXPORT_SYMBOL(osc_lock_set_writer);
1210 int osc_lock_init(const struct lu_env *env,
1211 struct cl_object *obj, struct cl_lock *lock,
1212 const struct cl_io *io)
1214 struct osc_lock *oscl;
1215 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1217 OBD_SLAB_ALLOC_PTR_GFP(oscl, osc_lock_kmem, GFP_NOFS);
1221 oscl->ols_state = OLS_NEW;
1222 spin_lock_init(&oscl->ols_lock);
1223 INIT_LIST_HEAD(&oscl->ols_waiting_list);
1224 INIT_LIST_HEAD(&oscl->ols_wait_entry);
1225 INIT_LIST_HEAD(&oscl->ols_nextlock_oscobj);
1226 oscl->ols_lockless_ops = &osc_lock_lockless_ops;
1228 /* Speculative lock requests must be either no_expand or glimpse
1229 * request (CEF_GLIMPSE). non-glimpse no_expand speculative extent
1230 * locks will break ofd_intent_cb. (see comment there)*/
1231 LASSERT(ergo((enqflags & CEF_SPECULATIVE) != 0,
1232 (enqflags & (CEF_LOCK_NO_EXPAND | CEF_GLIMPSE)) != 0));
1234 oscl->ols_flags = osc_enq2ldlm_flags(enqflags);
1235 oscl->ols_speculative = !!(enqflags & CEF_SPECULATIVE);
1236 if (lock->cll_descr.cld_mode == CLM_GROUP)
1237 oscl->ols_flags |= LDLM_FL_ATOMIC_CB;
1239 if (oscl->ols_flags & LDLM_FL_HAS_INTENT) {
1240 oscl->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1241 oscl->ols_glimpse = 1;
1243 if (io->ci_ndelay && cl_object_same(io->ci_obj, obj))
1244 oscl->ols_flags |= LDLM_FL_NDELAY;
1245 osc_lock_build_einfo(env, lock, cl2osc(obj), &oscl->ols_einfo);
1247 cl_lock_slice_add(lock, &oscl->ols_cl, obj, &osc_lock_ops);
1249 if (!(enqflags & CEF_MUST))
1250 /* try to convert this lock to a lockless lock */
1251 osc_lock_to_lockless(env, oscl, (enqflags & CEF_NEVER));
1252 if (oscl->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
1253 oscl->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1255 if (io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io))
1256 osc_lock_set_writer(env, io, obj, oscl);
1258 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %#llx",
1259 lock, oscl, oscl->ols_flags);
1265 * Finds an existing lock covering given index and optionally different from a
1266 * given \a except lock.
1268 struct ldlm_lock *osc_obj_dlmlock_at_pgoff(const struct lu_env *env,
1269 struct osc_object *obj,
1271 enum osc_dap_flags dap_flags)
1273 struct osc_thread_info *info = osc_env_info(env);
1274 struct ldlm_res_id *resname = &info->oti_resname;
1275 union ldlm_policy_data *policy = &info->oti_policy;
1276 struct lustre_handle lockh;
1277 struct ldlm_lock *lock = NULL;
1278 enum ldlm_mode mode;
1283 ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1284 osc_index2policy(policy, osc2cl(obj), index, index);
1285 policy->l_extent.gid = LDLM_GID_ANY;
1287 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1288 if (dap_flags & OSC_DAP_FL_TEST_LOCK)
1289 flags |= LDLM_FL_TEST_LOCK;
1291 * It is fine to match any group lock since there could be only one
1292 * with a uniq gid and it conflicts with all other lock modes too
1295 mode = osc_match_base(env, osc_export(obj), resname, LDLM_EXTENT,
1296 policy, LCK_PR | LCK_PW | LCK_GROUP, &flags,
1297 obj, &lockh, dap_flags & OSC_DAP_FL_CANCELING);
1299 lock = ldlm_handle2lock(&lockh);
1300 /* RACE: the lock is cancelled so let's try again */
1301 if (unlikely(lock == NULL))