4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Implementation of cl_lock for OSC layer.
34 * Author: Nikita Danilov <nikita.danilov@sun.com>
35 * Author: Jinshan Xiong <jinshan.xiong@intel.com>
38 #define DEBUG_SUBSYSTEM S_OSC
40 #include <libcfs/libcfs.h>
41 /* fid_build_reg_res_name() */
42 #include <lustre_fid.h>
43 #include <lustre_osc.h>
45 #include "osc_internal.h"
51 /*****************************************************************************
57 static const struct cl_lock_operations osc_lock_ops;
58 static const struct cl_lock_operations osc_lock_lockless_ops;
59 static void osc_lock_to_lockless(const struct lu_env *env,
60 struct osc_lock *ols, int force);
62 int osc_lock_is_lockless(const struct osc_lock *olck)
64 return (olck->ols_cl.cls_ops == &osc_lock_lockless_ops);
68 * Returns a weak pointer to the ldlm lock identified by a handle. Returned
69 * pointer cannot be dereferenced, as lock is not protected from concurrent
70 * reclaim. This function is a helper for osc_lock_invariant().
72 static struct ldlm_lock *osc_handle_ptr(struct lustre_handle *handle)
74 struct ldlm_lock *lock;
76 lock = ldlm_handle2lock(handle);
83 * Invariant that has to be true all of the time.
85 static int osc_lock_invariant(struct osc_lock *ols)
87 struct ldlm_lock *lock = osc_handle_ptr(&ols->ols_handle);
88 struct ldlm_lock *olock = ols->ols_dlmlock;
89 int handle_used = lustre_handle_is_used(&ols->ols_handle);
91 if (ergo(osc_lock_is_lockless(ols),
92 ols->ols_locklessable && ols->ols_dlmlock == NULL))
96 * If all the following "ergo"s are true, return 1, otherwise 0
98 if (! ergo(olock != NULL, handle_used))
101 if (! ergo(olock != NULL,
102 olock->l_handle.h_cookie == ols->ols_handle.cookie))
105 if (! ergo(handle_used,
106 ergo(lock != NULL && olock != NULL, lock == olock) &&
107 ergo(lock == NULL, olock == NULL)))
110 * Check that ->ols_handle and ->ols_dlmlock are consistent, but
111 * take into account that they are set at the different time.
113 if (! ergo(ols->ols_state == OLS_CANCELLED,
114 olock == NULL && !handle_used))
117 * DLM lock is destroyed only after we have seen cancellation
120 if (! ergo(olock != NULL && ols->ols_state < OLS_CANCELLED,
121 !ldlm_is_destroyed(olock)))
124 if (! ergo(ols->ols_state == OLS_GRANTED,
126 olock->l_req_mode == olock->l_granted_mode &&
132 /*****************************************************************************
138 static void osc_lock_fini(const struct lu_env *env,
139 struct cl_lock_slice *slice)
141 struct osc_lock *ols = cl2osc_lock(slice);
143 LINVRNT(osc_lock_invariant(ols));
144 LASSERT(ols->ols_dlmlock == NULL);
146 OBD_SLAB_FREE_PTR(ols, osc_lock_kmem);
149 static void osc_lock_build_policy(const struct lu_env *env,
150 const struct cl_lock *lock,
151 union ldlm_policy_data *policy)
153 const struct cl_lock_descr *d = &lock->cll_descr;
155 osc_index2policy(policy, d->cld_obj, d->cld_start, d->cld_end);
156 policy->l_extent.gid = d->cld_gid;
159 static __u64 osc_enq2ldlm_flags(__u32 enqflags)
163 LASSERT((enqflags & ~CEF_MASK) == 0);
165 if (enqflags & CEF_NONBLOCK)
166 result |= LDLM_FL_BLOCK_NOWAIT;
167 if (enqflags & CEF_ASYNC)
168 result |= LDLM_FL_HAS_INTENT;
169 if (enqflags & CEF_DISCARD_DATA)
170 result |= LDLM_FL_AST_DISCARD_DATA;
171 if (enqflags & CEF_PEEK)
172 result |= LDLM_FL_TEST_LOCK;
173 if (enqflags & CEF_LOCK_MATCH)
174 result |= LDLM_FL_MATCH_LOCK;
179 * Updates object attributes from a lock value block (lvb) received together
180 * with the DLM lock reply from the server. Copy of osc_update_enqueue()
183 * This can be optimized to not update attributes when lock is a result of a
186 * Called under lock and resource spin-locks.
188 static void osc_lock_lvb_update(const struct lu_env *env,
189 struct osc_object *osc,
190 struct ldlm_lock *dlmlock,
193 struct cl_object *obj = osc2cl(osc);
194 struct lov_oinfo *oinfo = osc->oo_oinfo;
195 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
200 valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME | CAT_SIZE;
202 LASSERT(dlmlock != NULL);
203 lvb = dlmlock->l_lvb_data;
205 cl_lvb2attr(attr, lvb);
207 cl_object_attr_lock(obj);
208 if (dlmlock != NULL) {
211 check_res_locked(dlmlock->l_resource);
213 LASSERT(lvb == dlmlock->l_lvb_data);
214 size = lvb->lvb_size;
216 /* Extend KMS up to the end of this lock and no further
217 * A lock on [x,y] means a KMS of up to y + 1 bytes! */
218 if (size > dlmlock->l_policy_data.l_extent.end)
219 size = dlmlock->l_policy_data.l_extent.end + 1;
220 if (size >= oinfo->loi_kms) {
221 LDLM_DEBUG(dlmlock, "lock acquired, setting rss=%llu"
222 ", kms=%llu", lvb->lvb_size, size);
224 attr->cat_kms = size;
226 LDLM_DEBUG(dlmlock, "lock acquired, setting rss="
227 "%llu; leaving kms=%llu, end=%llu",
228 lvb->lvb_size, oinfo->loi_kms,
229 dlmlock->l_policy_data.l_extent.end);
231 ldlm_lock_allow_match_locked(dlmlock);
234 cl_object_attr_update(env, obj, attr, valid);
235 cl_object_attr_unlock(obj);
240 static void osc_lock_granted(const struct lu_env *env, struct osc_lock *oscl,
241 struct lustre_handle *lockh, bool lvb_update)
243 struct ldlm_lock *dlmlock;
245 dlmlock = ldlm_handle2lock_long(lockh, 0);
246 LASSERT(dlmlock != NULL);
248 /* lock reference taken by ldlm_handle2lock_long() is
249 * owned by osc_lock and released in osc_lock_detach()
251 lu_ref_add(&dlmlock->l_reference, "osc_lock", oscl);
252 oscl->ols_has_ref = 1;
254 LASSERT(oscl->ols_dlmlock == NULL);
255 oscl->ols_dlmlock = dlmlock;
257 /* This may be a matched lock for glimpse request, do not hold
258 * lock reference in that case. */
259 if (!oscl->ols_glimpse) {
260 /* hold a refc for non glimpse lock which will
261 * be released in osc_lock_cancel() */
262 lustre_handle_copy(&oscl->ols_handle, lockh);
263 ldlm_lock_addref(lockh, oscl->ols_einfo.ei_mode);
267 /* Lock must have been granted. */
268 lock_res_and_lock(dlmlock);
269 if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
270 struct ldlm_extent *ext = &dlmlock->l_policy_data.l_extent;
271 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
273 /* extend the lock extent, otherwise it will have problem when
274 * we decide whether to grant a lockless lock. */
275 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
276 descr->cld_start = cl_index(descr->cld_obj, ext->start);
277 descr->cld_end = cl_index(descr->cld_obj, ext->end);
278 descr->cld_gid = ext->gid;
280 /* no lvb update for matched lock */
282 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
283 osc_lock_lvb_update(env, cl2osc(oscl->ols_cl.cls_obj),
286 LINVRNT(osc_lock_invariant(oscl));
288 unlock_res_and_lock(dlmlock);
290 LASSERT(oscl->ols_state != OLS_GRANTED);
291 oscl->ols_state = OLS_GRANTED;
295 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
296 * received from a server, or after osc_enqueue_base() matched a local DLM
299 static int osc_lock_upcall(void *cookie, struct lustre_handle *lockh,
302 struct osc_lock *oscl = cookie;
303 struct cl_lock_slice *slice = &oscl->ols_cl;
309 env = cl_env_percpu_get();
310 /* should never happen, similar to osc_ldlm_blocking_ast(). */
311 LASSERT(!IS_ERR(env));
313 rc = ldlm_error2errno(errcode);
314 if (oscl->ols_state == OLS_ENQUEUED) {
315 oscl->ols_state = OLS_UPCALL_RECEIVED;
316 } else if (oscl->ols_state == OLS_CANCELLED) {
319 CERROR("Impossible state: %d\n", oscl->ols_state);
324 osc_lock_granted(env, oscl, lockh, errcode == ELDLM_OK);
326 /* Error handling, some errors are tolerable. */
327 if (oscl->ols_locklessable && rc == -EUSERS) {
328 /* This is a tolerable error, turn this lock into
331 osc_object_set_contended(cl2osc(slice->cls_obj));
332 LASSERT(slice->cls_ops == &osc_lock_ops);
334 /* Change this lock to ldlmlock-less lock. */
335 osc_lock_to_lockless(env, oscl, 1);
336 oscl->ols_state = OLS_GRANTED;
338 } else if (oscl->ols_glimpse && rc == -ENAVAIL) {
339 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
340 osc_lock_lvb_update(env, cl2osc(slice->cls_obj),
341 NULL, &oscl->ols_lvb);
342 /* Hide the error. */
346 if (oscl->ols_owner != NULL)
347 cl_sync_io_note(env, oscl->ols_owner, rc);
348 cl_env_percpu_put(env);
353 static int osc_lock_upcall_agl(void *cookie, struct lustre_handle *lockh,
356 struct osc_object *osc = cookie;
357 struct ldlm_lock *dlmlock;
362 env = cl_env_get(&refcheck);
363 LASSERT(!IS_ERR(env));
365 if (errcode == ELDLM_LOCK_MATCHED)
366 GOTO(out, errcode = ELDLM_OK);
368 if (errcode != ELDLM_OK)
371 dlmlock = ldlm_handle2lock(lockh);
372 LASSERT(dlmlock != NULL);
374 lock_res_and_lock(dlmlock);
375 LASSERT(dlmlock->l_granted_mode == dlmlock->l_req_mode);
377 /* there is no osc_lock associated with AGL lock */
378 osc_lock_lvb_update(env, osc, dlmlock, NULL);
380 unlock_res_and_lock(dlmlock);
381 LDLM_LOCK_PUT(dlmlock);
384 cl_object_put(env, osc2cl(osc));
385 cl_env_put(env, &refcheck);
386 RETURN(ldlm_error2errno(errcode));
389 static int osc_lock_flush(struct osc_object *obj, pgoff_t start, pgoff_t end,
390 enum cl_lock_mode mode, int discard)
399 env = cl_env_get(&refcheck);
401 RETURN(PTR_ERR(env));
403 if (mode == CLM_WRITE) {
404 rc = osc_cache_writeback_range(env, obj, start, end, 1,
406 CDEBUG(D_CACHE, "object %p: [%lu -> %lu] %d pages were %s.\n",
408 discard ? "discarded" : "written back");
413 rc2 = osc_lock_discard_pages(env, obj, start, end, mode);
414 if (rc == 0 && rc2 < 0)
417 cl_env_put(env, &refcheck);
422 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
423 * and ldlm_lock caches.
425 static int osc_dlm_blocking_ast0(const struct lu_env *env,
426 struct ldlm_lock *dlmlock,
427 void *data, int flag)
429 struct cl_object *obj = NULL;
432 enum cl_lock_mode mode = CLM_READ;
435 LASSERT(flag == LDLM_CB_CANCELING);
437 lock_res_and_lock(dlmlock);
438 if (dlmlock->l_granted_mode != dlmlock->l_req_mode) {
439 dlmlock->l_ast_data = NULL;
440 unlock_res_and_lock(dlmlock);
444 discard = ldlm_is_discard_data(dlmlock);
445 if (dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))
448 if (dlmlock->l_ast_data != NULL) {
449 obj = osc2cl(dlmlock->l_ast_data);
450 dlmlock->l_ast_data = NULL;
455 unlock_res_and_lock(dlmlock);
457 /* if l_ast_data is NULL, the dlmlock was enqueued by AGL or
458 * the object has been destroyed. */
460 struct ldlm_extent *extent = &dlmlock->l_policy_data.l_extent;
461 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
464 /* Destroy pages covered by the extent of the DLM lock */
465 result = osc_lock_flush(cl2osc(obj),
466 cl_index(obj, extent->start),
467 cl_index(obj, extent->end),
470 /* losing a lock, update kms */
471 lock_res_and_lock(dlmlock);
472 cl_object_attr_lock(obj);
473 /* Must get the value under the lock to avoid race. */
474 old_kms = cl2osc(obj)->oo_oinfo->loi_kms;
475 /* Update the kms. Need to loop all granted locks.
476 * Not a problem for the client */
477 attr->cat_kms = ldlm_extent_shift_kms(dlmlock, old_kms);
479 cl_object_attr_update(env, obj, attr, CAT_KMS);
480 cl_object_attr_unlock(obj);
481 unlock_res_and_lock(dlmlock);
483 cl_object_put(env, obj);
489 * Blocking ast invoked by ldlm when dlm lock is either blocking progress of
490 * some other lock, or is canceled. This function is installed as a
491 * ldlm_lock::l_blocking_ast() for client extent locks.
493 * Control flow is tricky, because ldlm uses the same call-back
494 * (ldlm_lock::l_blocking_ast()) for both blocking and cancellation ast's.
496 * \param dlmlock lock for which ast occurred.
498 * \param new description of a conflicting lock in case of blocking ast.
500 * \param data value of dlmlock->l_ast_data
502 * \param flag LDLM_CB_BLOCKING or LDLM_CB_CANCELING. Used to distinguish
503 * cancellation and blocking ast's.
505 * Possible use cases:
507 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING) to cancel
508 * lock due to lock lru pressure, or explicit user request to purge
511 * - ldlm calls dlmlock->l_blocking_ast(..., LDLM_CB_BLOCKING) to notify
512 * us that dlmlock conflicts with another lock that some client is
513 * enqueuing. Lock is canceled.
515 * - cl_lock_cancel() is called. osc_lock_cancel() calls
516 * ldlm_cli_cancel() that calls
518 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
520 * recursively entering osc_ldlm_blocking_ast().
522 * - client cancels lock voluntary (e.g., as a part of early cancellation):
525 * osc_lock_cancel()->
526 * ldlm_cli_cancel()->
527 * dlmlock->l_blocking_ast(..., LDLM_CB_CANCELING)
530 static int osc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
531 struct ldlm_lock_desc *new, void *data,
538 case LDLM_CB_BLOCKING: {
539 struct lustre_handle lockh;
541 ldlm_lock2handle(dlmlock, &lockh);
542 result = ldlm_cli_cancel(&lockh, LCF_ASYNC);
543 if (result == -ENODATA)
547 case LDLM_CB_CANCELING: {
552 * This can be called in the context of outer IO, e.g.,
554 * osc_enqueue_base()->...
555 * ->ldlm_prep_elc_req()->...
556 * ->ldlm_cancel_callback()->...
557 * ->osc_ldlm_blocking_ast()
559 * new environment has to be created to not corrupt outer
562 env = cl_env_get(&refcheck);
564 result = PTR_ERR(env);
568 result = osc_dlm_blocking_ast0(env, dlmlock, data, flag);
569 cl_env_put(env, &refcheck);
578 static int osc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
580 struct ptlrpc_request *req = data;
583 struct req_capsule *cap;
584 struct cl_object *obj = NULL;
590 LASSERT(lustre_msg_get_opc(req->rq_reqmsg) == LDLM_GL_CALLBACK);
592 env = cl_env_get(&refcheck);
594 GOTO(out, result = PTR_ERR(env));
597 lock_res_and_lock(dlmlock);
598 if (dlmlock->l_ast_data != NULL) {
599 obj = osc2cl(dlmlock->l_ast_data);
602 unlock_res_and_lock(dlmlock);
605 /* Do not grab the mutex of cl_lock for glimpse.
606 * See LU-1274 for details.
607 * BTW, it's okay for cl_lock to be cancelled during
608 * this period because server can handle this race.
609 * See ldlm_server_glimpse_ast() for details.
610 * cl_lock_mutex_get(env, lock); */
612 req_capsule_extend(cap, &RQF_LDLM_GL_CALLBACK);
613 req_capsule_set_size(cap, &RMF_DLM_LVB, RCL_SERVER,
615 result = req_capsule_server_pack(cap);
617 lvb = req_capsule_server_get(cap, &RMF_DLM_LVB);
618 result = cl_object_glimpse(env, obj, lvb);
620 if (!exp_connect_lvb_type(req->rq_export))
621 req_capsule_shrink(&req->rq_pill, &RMF_DLM_LVB,
622 sizeof(struct ost_lvb_v1), RCL_SERVER);
623 cl_object_put(env, obj);
626 * These errors are normal races, so we don't want to
627 * fill the console with messages by calling
630 lustre_pack_reply(req, 1, NULL, NULL);
631 result = -ELDLM_NO_LOCK_DATA;
633 cl_env_put(env, &refcheck);
637 req->rq_status = result;
641 static int weigh_cb(const struct lu_env *env, struct cl_io *io,
642 struct osc_page *ops, void *cbdata)
644 struct cl_page *page = ops->ops_cl.cpl_page;
646 if (cl_page_is_vmlocked(env, page)
647 || PageDirty(page->cp_vmpage) || PageWriteback(page->cp_vmpage)
649 return CLP_GANG_ABORT;
651 *(pgoff_t *)cbdata = osc_index(ops) + 1;
652 return CLP_GANG_OKAY;
655 static unsigned long osc_lock_weight(const struct lu_env *env,
656 struct osc_object *oscobj,
657 struct ldlm_extent *extent)
659 struct cl_io *io = &osc_env_info(env)->oti_io;
660 struct cl_object *obj = cl_object_top(&oscobj->oo_cl);
666 io->ci_ignore_layout = 1;
667 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
671 page_index = cl_index(obj, extent->start);
673 result = osc_page_gang_lookup(env, io, oscobj,
675 cl_index(obj, extent->end),
676 weigh_cb, (void *)&page_index);
677 if (result == CLP_GANG_ABORT)
679 if (result == CLP_GANG_RESCHED)
681 } while (result != CLP_GANG_OKAY);
684 return result == CLP_GANG_ABORT ? 1 : 0;
688 * Get the weight of dlm lock for early cancellation.
690 unsigned long osc_ldlm_weigh_ast(struct ldlm_lock *dlmlock)
693 struct osc_object *obj;
694 struct osc_lock *oscl;
695 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 lock_res_and_lock(dlmlock);
715 obj = dlmlock->l_ast_data;
717 cl_object_get(osc2cl(obj));
718 unlock_res_and_lock(dlmlock);
721 GOTO(out, weight = 1);
723 spin_lock(&obj->oo_ol_spin);
724 list_for_each_entry(oscl, &obj->oo_ol_list, ols_nextlock_oscobj) {
725 if (oscl->ols_dlmlock != NULL && oscl->ols_dlmlock != dlmlock)
729 spin_unlock(&obj->oo_ol_spin);
732 * If the lock is being used by an IO, definitely not cancel it.
734 GOTO(out, weight = 1);
737 weight = osc_lock_weight(env, obj, &dlmlock->l_policy_data.l_extent);
742 cl_object_put(env, osc2cl(obj));
744 cl_env_put(env, &refcheck);
748 static void osc_lock_build_einfo(const struct lu_env *env,
749 const struct cl_lock *lock,
750 struct osc_object *osc,
751 struct ldlm_enqueue_info *einfo)
753 einfo->ei_type = LDLM_EXTENT;
754 einfo->ei_mode = osc_cl_lock2ldlm(lock->cll_descr.cld_mode);
755 einfo->ei_cb_bl = osc_ldlm_blocking_ast;
756 einfo->ei_cb_cp = ldlm_completion_ast;
757 einfo->ei_cb_gl = osc_ldlm_glimpse_ast;
758 einfo->ei_cbdata = osc; /* value to be put into ->l_ast_data */
762 * Determine if the lock should be converted into a lockless lock.
765 * - if the lock has an explicite requirment for a non-lockless lock;
766 * - if the io lock request type ci_lockreq;
767 * - send the enqueue rpc to ost to make the further decision;
768 * - special treat to truncate lockless lock
770 * Additional policy can be implemented here, e.g., never do lockless-io
773 static void osc_lock_to_lockless(const struct lu_env *env,
774 struct osc_lock *ols, int force)
776 struct cl_lock_slice *slice = &ols->ols_cl;
778 LASSERT(ols->ols_state == OLS_NEW ||
779 ols->ols_state == OLS_UPCALL_RECEIVED);
782 ols->ols_locklessable = 1;
783 slice->cls_ops = &osc_lock_lockless_ops;
785 struct osc_io *oio = osc_env_io(env);
786 struct cl_io *io = oio->oi_cl.cis_io;
787 struct cl_object *obj = slice->cls_obj;
788 struct osc_object *oob = cl2osc(obj);
789 const struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
790 struct obd_connect_data *ocd;
792 LASSERT(io->ci_lockreq == CILR_MANDATORY ||
793 io->ci_lockreq == CILR_MAYBE ||
794 io->ci_lockreq == CILR_NEVER);
796 ocd = &class_exp2cliimp(osc_export(oob))->imp_connect_data;
797 ols->ols_locklessable = (io->ci_type != CIT_SETATTR) &&
798 (io->ci_lockreq == CILR_MAYBE) &&
799 (ocd->ocd_connect_flags & OBD_CONNECT_SRVLOCK);
800 if (io->ci_lockreq == CILR_NEVER ||
802 (ols->ols_locklessable && osc_object_is_contended(oob)) ||
803 /* lockless truncate */
804 (cl_io_is_trunc(io) &&
805 (ocd->ocd_connect_flags & OBD_CONNECT_TRUNCLOCK) &&
806 osd->od_lockless_truncate)) {
807 ols->ols_locklessable = 1;
808 slice->cls_ops = &osc_lock_lockless_ops;
811 LASSERT(ergo(ols->ols_glimpse, !osc_lock_is_lockless(ols)));
814 static bool osc_lock_compatible(const struct osc_lock *qing,
815 const struct osc_lock *qed)
817 struct cl_lock_descr *qed_descr = &qed->ols_cl.cls_lock->cll_descr;
818 struct cl_lock_descr *qing_descr = &qing->ols_cl.cls_lock->cll_descr;
820 if (qed->ols_glimpse)
823 if (qing_descr->cld_mode == CLM_READ && qed_descr->cld_mode == CLM_READ)
826 if (qed->ols_state < OLS_GRANTED)
829 if (qed_descr->cld_mode >= qing_descr->cld_mode &&
830 qed_descr->cld_start <= qing_descr->cld_start &&
831 qed_descr->cld_end >= qing_descr->cld_end)
837 static void osc_lock_wake_waiters(const struct lu_env *env,
838 struct osc_object *osc,
839 struct osc_lock *oscl)
841 spin_lock(&osc->oo_ol_spin);
842 list_del_init(&oscl->ols_nextlock_oscobj);
843 spin_unlock(&osc->oo_ol_spin);
845 spin_lock(&oscl->ols_lock);
846 while (!list_empty(&oscl->ols_waiting_list)) {
847 struct osc_lock *scan;
849 scan = list_entry(oscl->ols_waiting_list.next, struct osc_lock,
851 list_del_init(&scan->ols_wait_entry);
853 cl_sync_io_note(env, scan->ols_owner, 0);
855 spin_unlock(&oscl->ols_lock);
858 static int osc_lock_enqueue_wait(const struct lu_env *env,
859 struct osc_object *obj, struct osc_lock *oscl)
861 struct osc_lock *tmp_oscl;
862 struct cl_lock_descr *need = &oscl->ols_cl.cls_lock->cll_descr;
863 struct cl_sync_io *waiter = &osc_env_info(env)->oti_anchor;
867 spin_lock(&obj->oo_ol_spin);
868 list_add_tail(&oscl->ols_nextlock_oscobj, &obj->oo_ol_list);
871 list_for_each_entry(tmp_oscl, &obj->oo_ol_list,
872 ols_nextlock_oscobj) {
873 struct cl_lock_descr *descr;
875 if (tmp_oscl == oscl)
878 descr = &tmp_oscl->ols_cl.cls_lock->cll_descr;
879 if (descr->cld_start > need->cld_end ||
880 descr->cld_end < need->cld_start)
883 /* We're not supposed to give up group lock */
884 if (descr->cld_mode == CLM_GROUP)
887 if (!osc_lock_is_lockless(oscl) &&
888 osc_lock_compatible(oscl, tmp_oscl))
891 /* wait for conflicting lock to be canceled */
892 cl_sync_io_init(waiter, 1, cl_sync_io_end);
893 oscl->ols_owner = waiter;
895 spin_lock(&tmp_oscl->ols_lock);
896 /* add oscl into tmp's ols_waiting list */
897 list_add_tail(&oscl->ols_wait_entry,
898 &tmp_oscl->ols_waiting_list);
899 spin_unlock(&tmp_oscl->ols_lock);
901 spin_unlock(&obj->oo_ol_spin);
902 rc = cl_sync_io_wait(env, waiter, 0);
903 spin_lock(&obj->oo_ol_spin);
908 oscl->ols_owner = NULL;
911 spin_unlock(&obj->oo_ol_spin);
917 * Implementation of cl_lock_operations::clo_enqueue() method for osc
918 * layer. This initiates ldlm enqueue:
920 * - cancels conflicting locks early (osc_lock_enqueue_wait());
922 * - calls osc_enqueue_base() to do actual enqueue.
924 * osc_enqueue_base() is supplied with an upcall function that is executed
925 * when lock is received either after a local cached ldlm lock is matched, or
926 * when a reply from the server is received.
928 * This function does not wait for the network communication to complete.
930 static int osc_lock_enqueue(const struct lu_env *env,
931 const struct cl_lock_slice *slice,
932 struct cl_io *unused, struct cl_sync_io *anchor)
934 struct osc_thread_info *info = osc_env_info(env);
935 struct osc_io *oio = osc_env_io(env);
936 struct osc_object *osc = cl2osc(slice->cls_obj);
937 struct osc_lock *oscl = cl2osc_lock(slice);
938 struct cl_lock *lock = slice->cls_lock;
939 struct ldlm_res_id *resname = &info->oti_resname;
940 union ldlm_policy_data *policy = &info->oti_policy;
941 osc_enqueue_upcall_f upcall = osc_lock_upcall;
948 LASSERTF(ergo(oscl->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
949 "lock = %p, ols = %p\n", lock, oscl);
951 if (oscl->ols_state == OLS_GRANTED)
954 if (oscl->ols_flags & LDLM_FL_TEST_LOCK)
955 GOTO(enqueue_base, 0);
957 if (oscl->ols_glimpse) {
958 LASSERT(equi(oscl->ols_agl, anchor == NULL));
960 GOTO(enqueue_base, 0);
963 result = osc_lock_enqueue_wait(env, osc, oscl);
967 /* we can grant lockless lock right after all conflicting locks
969 if (osc_lock_is_lockless(oscl)) {
970 oscl->ols_state = OLS_GRANTED;
971 oio->oi_lockless = 1;
976 oscl->ols_state = OLS_ENQUEUED;
977 if (anchor != NULL) {
978 atomic_inc(&anchor->csi_sync_nr);
979 oscl->ols_owner = anchor;
983 * DLM lock's ast data must be osc_object;
984 * if glimpse or AGL lock, async of osc_enqueue_base() must be true,
985 * DLM's enqueue callback set to osc_lock_upcall() with cookie as
988 ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
989 osc_lock_build_policy(env, lock, policy);
991 oscl->ols_einfo.ei_cbdata = NULL;
992 /* hold a reference for callback */
993 cl_object_get(osc2cl(osc));
994 upcall = osc_lock_upcall_agl;
997 result = osc_enqueue_base(osc_export(osc), resname, &oscl->ols_flags,
998 policy, &oscl->ols_lvb,
999 osc->oo_oinfo->loi_kms_valid,
1001 &oscl->ols_einfo, PTLRPCD_SET, async,
1004 if (osc_lock_is_lockless(oscl)) {
1005 oio->oi_lockless = 1;
1006 } else if (!async) {
1007 LASSERT(oscl->ols_state == OLS_GRANTED);
1008 LASSERT(oscl->ols_hold);
1009 LASSERT(oscl->ols_dlmlock != NULL);
1011 } else if (oscl->ols_agl) {
1012 cl_object_put(env, osc2cl(osc));
1018 oscl->ols_state = OLS_CANCELLED;
1019 osc_lock_wake_waiters(env, osc, oscl);
1022 cl_sync_io_note(env, anchor, result);
1028 * Breaks a link between osc_lock and dlm_lock.
1030 static void osc_lock_detach(const struct lu_env *env, struct osc_lock *olck)
1032 struct ldlm_lock *dlmlock;
1036 dlmlock = olck->ols_dlmlock;
1037 if (dlmlock == NULL)
1040 if (olck->ols_hold) {
1042 ldlm_lock_decref(&olck->ols_handle, olck->ols_einfo.ei_mode);
1043 olck->ols_handle.cookie = 0ULL;
1046 olck->ols_dlmlock = NULL;
1048 /* release a reference taken in osc_lock_upcall(). */
1049 LASSERT(olck->ols_has_ref);
1050 lu_ref_del(&dlmlock->l_reference, "osc_lock", olck);
1051 LDLM_LOCK_RELEASE(dlmlock);
1052 olck->ols_has_ref = 0;
1058 * Implements cl_lock_operations::clo_cancel() method for osc layer. This is
1059 * called (as part of cl_lock_cancel()) when lock is canceled either voluntary
1060 * (LRU pressure, early cancellation, umount, etc.) or due to the conflict
1061 * with some other lock some where in the cluster. This function does the
1064 * - invalidates all pages protected by this lock (after sending dirty
1065 * ones to the server, as necessary);
1067 * - decref's underlying ldlm lock;
1069 * - cancels ldlm lock (ldlm_cli_cancel()).
1071 static void osc_lock_cancel(const struct lu_env *env,
1072 const struct cl_lock_slice *slice)
1074 struct osc_object *obj = cl2osc(slice->cls_obj);
1075 struct osc_lock *oscl = cl2osc_lock(slice);
1079 LINVRNT(osc_lock_invariant(oscl));
1081 osc_lock_detach(env, oscl);
1082 oscl->ols_state = OLS_CANCELLED;
1083 oscl->ols_flags &= ~LDLM_FL_LVB_READY;
1085 osc_lock_wake_waiters(env, obj, oscl);
1089 static int osc_lock_print(const struct lu_env *env, void *cookie,
1090 lu_printer_t p, const struct cl_lock_slice *slice)
1092 struct osc_lock *lock = cl2osc_lock(slice);
1094 (*p)(env, cookie, "%p %#llx %#llx %d %p ",
1095 lock->ols_dlmlock, lock->ols_flags, lock->ols_handle.cookie,
1096 lock->ols_state, lock->ols_owner);
1097 osc_lvb_print(env, cookie, p, &lock->ols_lvb);
1101 static const struct cl_lock_operations osc_lock_ops = {
1102 .clo_fini = osc_lock_fini,
1103 .clo_enqueue = osc_lock_enqueue,
1104 .clo_cancel = osc_lock_cancel,
1105 .clo_print = osc_lock_print,
1108 static void osc_lock_lockless_cancel(const struct lu_env *env,
1109 const struct cl_lock_slice *slice)
1111 struct osc_lock *ols = cl2osc_lock(slice);
1112 struct osc_object *osc = cl2osc(slice->cls_obj);
1113 struct cl_lock_descr *descr = &slice->cls_lock->cll_descr;
1116 LASSERT(ols->ols_dlmlock == NULL);
1117 result = osc_lock_flush(osc, descr->cld_start, descr->cld_end,
1118 descr->cld_mode, 0);
1120 CERROR("Pages for lockless lock %p were not purged(%d)\n",
1123 osc_lock_wake_waiters(env, osc, ols);
1126 static const struct cl_lock_operations osc_lock_lockless_ops = {
1127 .clo_fini = osc_lock_fini,
1128 .clo_enqueue = osc_lock_enqueue,
1129 .clo_cancel = osc_lock_lockless_cancel,
1130 .clo_print = osc_lock_print
1133 static void osc_lock_set_writer(const struct lu_env *env,
1134 const struct cl_io *io,
1135 struct cl_object *obj, struct osc_lock *oscl)
1137 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
1141 if (!cl_object_same(io->ci_obj, obj))
1144 if (likely(io->ci_type == CIT_WRITE)) {
1145 io_start = cl_index(obj, io->u.ci_rw.rw_range.cir_pos);
1146 io_end = cl_index(obj, io->u.ci_rw.rw_range.cir_pos +
1147 io->u.ci_rw.rw_range.cir_count - 1);
1149 LASSERT(cl_io_is_mkwrite(io));
1150 io_start = io_end = io->u.ci_fault.ft_index;
1153 if (descr->cld_mode >= CLM_WRITE &&
1154 (cl_io_is_append(io) ||
1155 (descr->cld_start <= io_start && descr->cld_end >= io_end))) {
1156 struct osc_io *oio = osc_env_io(env);
1158 /* There must be only one lock to match the write region */
1159 LASSERT(oio->oi_write_osclock == NULL);
1160 oio->oi_write_osclock = oscl;
1164 int osc_lock_init(const struct lu_env *env,
1165 struct cl_object *obj, struct cl_lock *lock,
1166 const struct cl_io *io)
1168 struct osc_lock *oscl;
1169 __u32 enqflags = lock->cll_descr.cld_enq_flags;
1171 OBD_SLAB_ALLOC_PTR_GFP(oscl, osc_lock_kmem, GFP_NOFS);
1175 oscl->ols_state = OLS_NEW;
1176 spin_lock_init(&oscl->ols_lock);
1177 INIT_LIST_HEAD(&oscl->ols_waiting_list);
1178 INIT_LIST_HEAD(&oscl->ols_wait_entry);
1179 INIT_LIST_HEAD(&oscl->ols_nextlock_oscobj);
1181 oscl->ols_flags = osc_enq2ldlm_flags(enqflags);
1182 oscl->ols_agl = !!(enqflags & CEF_AGL);
1184 oscl->ols_flags |= LDLM_FL_BLOCK_NOWAIT;
1185 if (oscl->ols_flags & LDLM_FL_HAS_INTENT) {
1186 oscl->ols_flags |= LDLM_FL_BLOCK_GRANTED;
1187 oscl->ols_glimpse = 1;
1189 osc_lock_build_einfo(env, lock, cl2osc(obj), &oscl->ols_einfo);
1191 cl_lock_slice_add(lock, &oscl->ols_cl, obj, &osc_lock_ops);
1193 if (!(enqflags & CEF_MUST))
1194 /* try to convert this lock to a lockless lock */
1195 osc_lock_to_lockless(env, oscl, (enqflags & CEF_NEVER));
1196 if (oscl->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
1197 oscl->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
1199 if (io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io))
1200 osc_lock_set_writer(env, io, obj, oscl);
1202 LDLM_DEBUG_NOLOCK("lock %p, osc lock %p, flags %#llx",
1203 lock, oscl, oscl->ols_flags);
1209 * Finds an existing lock covering given index and optionally different from a
1210 * given \a except lock.
1212 struct ldlm_lock *osc_dlmlock_at_pgoff(const struct lu_env *env,
1213 struct osc_object *obj, pgoff_t index,
1214 enum osc_dap_flags dap_flags)
1216 struct osc_thread_info *info = osc_env_info(env);
1217 struct ldlm_res_id *resname = &info->oti_resname;
1218 union ldlm_policy_data *policy = &info->oti_policy;
1219 struct lustre_handle lockh;
1220 struct ldlm_lock *lock = NULL;
1221 enum ldlm_mode mode;
1226 ostid_build_res_name(&obj->oo_oinfo->loi_oi, resname);
1227 osc_index2policy(policy, osc2cl(obj), index, index);
1228 policy->l_extent.gid = LDLM_GID_ANY;
1230 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
1231 if (dap_flags & OSC_DAP_FL_TEST_LOCK)
1232 flags |= LDLM_FL_TEST_LOCK;
1234 * It is fine to match any group lock since there could be only one
1235 * with a uniq gid and it conflicts with all other lock modes too
1238 mode = osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
1239 LCK_PR | LCK_PW | LCK_GROUP, &flags, obj, &lockh,
1240 dap_flags & OSC_DAP_FL_CANCELING);
1242 lock = ldlm_handle2lock(&lockh);
1243 /* RACE: the lock is cancelled so let's try again */
1244 if (unlikely(lock == NULL))