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) 2017 Intel Corporation.
26 * This file is part of Lustre, http://www.lustre.org/
28 * Implementation of cl_device, cl_req for MDC layer.
30 * Author: Mikhail Pershin <mike.pershin@intel.com>
33 #define DEBUG_SUBSYSTEM S_MDC
35 #include <obd_class.h>
36 #include <lustre_osc.h>
38 #include "mdc_internal.h"
40 static void mdc_lock_build_policy(const struct lu_env *env,
41 union ldlm_policy_data *policy)
43 memset(policy, 0, sizeof *policy);
44 policy->l_inodebits.bits = MDS_INODELOCK_DOM;
47 int mdc_ldlm_glimpse_ast(struct ldlm_lock *dlmlock, void *data)
49 return osc_ldlm_glimpse_ast(dlmlock, data);
52 static void mdc_lock_build_einfo(const struct lu_env *env,
53 const struct cl_lock *lock,
54 struct osc_object *osc,
55 struct ldlm_enqueue_info *einfo)
57 einfo->ei_type = LDLM_IBITS;
58 einfo->ei_mode = osc_cl_lock2ldlm(lock->cll_descr.cld_mode);
59 einfo->ei_cb_bl = mdc_ldlm_blocking_ast;
60 einfo->ei_cb_cp = ldlm_completion_ast;
61 einfo->ei_cb_gl = mdc_ldlm_glimpse_ast;
62 einfo->ei_cbdata = osc; /* value to be put into ->l_ast_data */
65 static void mdc_lock_lvb_update(const struct lu_env *env,
66 struct osc_object *osc,
67 struct ldlm_lock *dlmlock,
70 static int mdc_set_dom_lock_data(const struct lu_env *env,
71 struct ldlm_lock *lock, void *data)
73 struct osc_object *obj = data;
76 LASSERT(lock != NULL);
77 LASSERT(lock->l_glimpse_ast == mdc_ldlm_glimpse_ast);
79 lock_res_and_lock(lock);
80 if (lock->l_ast_data == NULL) {
81 lock->l_ast_data = data;
82 mdc_lock_lvb_update(env, obj, lock, NULL);
85 if (lock->l_ast_data == data)
88 unlock_res_and_lock(lock);
93 int mdc_dom_lock_match(const struct lu_env *env, struct obd_export *exp,
94 struct ldlm_res_id *res_id,
95 enum ldlm_type type, union ldlm_policy_data *policy,
96 enum ldlm_mode mode, __u64 *flags, void *data,
97 struct lustre_handle *lockh, int unref)
99 struct obd_device *obd = exp->exp_obd;
100 __u64 lflags = *flags;
105 rc = ldlm_lock_match(obd->obd_namespace, lflags,
106 res_id, type, policy, mode, lockh, unref);
107 if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
111 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
113 LASSERT(lock != NULL);
114 if (!mdc_set_dom_lock_data(env, lock, data)) {
115 ldlm_lock_decref(lockh, rc);
124 * Finds an existing lock covering a page with given index.
125 * Copy of osc_obj_dlmlock_at_pgoff() but for DoM IBITS lock.
127 struct ldlm_lock *mdc_dlmlock_at_pgoff(const struct lu_env *env,
128 struct osc_object *obj, pgoff_t index,
129 enum osc_dap_flags dap_flags)
131 struct osc_thread_info *info = osc_env_info(env);
132 struct ldlm_res_id *resname = &info->oti_resname;
133 union ldlm_policy_data *policy = &info->oti_policy;
134 struct lustre_handle lockh;
135 struct ldlm_lock *lock = NULL;
141 fid_build_reg_res_name(lu_object_fid(osc2lu(obj)), resname);
142 mdc_lock_build_policy(env, policy);
144 flags = LDLM_FL_BLOCK_GRANTED | LDLM_FL_CBPENDING;
145 if (dap_flags & OSC_DAP_FL_TEST_LOCK)
146 flags |= LDLM_FL_TEST_LOCK;
149 /* Next, search for already existing extent locks that will cover us */
150 /* If we're trying to read, we also search for an existing PW lock. The
151 * VFS and page cache already protect us locally, so lots of readers/
152 * writers can share a single PW lock. */
153 mode = mdc_dom_lock_match(env, osc_export(obj), resname, LDLM_IBITS,
154 policy, LCK_PR | LCK_PW, &flags, obj, &lockh,
155 dap_flags & OSC_DAP_FL_CANCELING);
157 lock = ldlm_handle2lock(&lockh);
158 /* RACE: the lock is cancelled so let's try again */
159 if (unlikely(lock == NULL))
167 * Check if page @page is covered by an extra lock or discard it.
169 static int mdc_check_and_discard_cb(const struct lu_env *env, struct cl_io *io,
170 struct osc_page *ops, void *cbdata)
172 struct osc_thread_info *info = osc_env_info(env);
173 struct osc_object *osc = cbdata;
176 index = osc_index(ops);
177 if (index >= info->oti_fn_index) {
178 struct ldlm_lock *tmp;
179 struct cl_page *page = ops->ops_cl.cpl_page;
181 /* refresh non-overlapped index */
182 tmp = mdc_dlmlock_at_pgoff(env, osc, index,
183 OSC_DAP_FL_TEST_LOCK);
185 info->oti_fn_index = CL_PAGE_EOF;
187 } else if (cl_page_own(env, io, page) == 0) {
188 /* discard the page */
189 cl_page_discard(env, io, page);
190 cl_page_disown(env, io, page);
192 LASSERT(page->cp_state == CPS_FREEING);
196 info->oti_next_index = index + 1;
197 return CLP_GANG_OKAY;
201 * Discard pages protected by the given lock. This function traverses radix
202 * tree to find all covering pages and discard them. If a page is being covered
203 * by other locks, it should remain in cache.
205 * If error happens on any step, the process continues anyway (the reasoning
206 * behind this being that lock cancellation cannot be delayed indefinitely).
208 static int mdc_lock_discard_pages(const struct lu_env *env,
209 struct osc_object *osc,
210 pgoff_t start, pgoff_t end,
213 struct osc_thread_info *info = osc_env_info(env);
214 struct cl_io *io = &info->oti_io;
215 osc_page_gang_cbt cb;
221 io->ci_obj = cl_object_top(osc2cl(osc));
222 io->ci_ignore_layout = 1;
223 result = cl_io_init(env, io, CIT_MISC, io->ci_obj);
227 cb = discard ? osc_discard_cb : mdc_check_and_discard_cb;
228 info->oti_fn_index = info->oti_next_index = start;
230 res = osc_page_gang_lookup(env, io, osc, info->oti_next_index,
231 end, cb, (void *)osc);
232 if (info->oti_next_index > end)
235 if (res == CLP_GANG_RESCHED)
237 } while (res != CLP_GANG_OKAY);
243 static int mdc_lock_flush(const struct lu_env *env, struct osc_object *obj,
244 pgoff_t start, pgoff_t end, enum cl_lock_mode mode,
252 if (mode == CLM_WRITE) {
253 result = osc_cache_writeback_range(env, obj, start, end, 1,
255 CDEBUG(D_CACHE, "object %p: [%lu -> %lu] %d pages were %s.\n",
256 obj, start, end, result,
257 discard ? "discarded" : "written back");
262 rc = mdc_lock_discard_pages(env, obj, start, end, discard);
263 if (result == 0 && rc < 0)
269 void mdc_lock_lockless_cancel(const struct lu_env *env,
270 const struct cl_lock_slice *slice)
272 struct osc_lock *ols = cl2osc_lock(slice);
273 struct osc_object *osc = cl2osc(slice->cls_obj);
274 struct cl_lock_descr *descr = &slice->cls_lock->cll_descr;
277 LASSERT(ols->ols_dlmlock == NULL);
278 rc = mdc_lock_flush(env, osc, descr->cld_start, descr->cld_end,
281 CERROR("Pages for lockless lock %p were not purged(%d)\n",
284 osc_lock_wake_waiters(env, osc, ols);
288 * Helper for osc_dlm_blocking_ast() handling discrepancies between cl_lock
289 * and ldlm_lock caches.
291 static int mdc_dlm_blocking_ast0(const struct lu_env *env,
292 struct ldlm_lock *dlmlock,
293 void *data, int flag)
295 struct cl_object *obj = NULL;
298 enum cl_lock_mode mode = CLM_READ;
302 LASSERT(flag == LDLM_CB_CANCELING);
303 LASSERT(dlmlock != NULL);
305 lock_res_and_lock(dlmlock);
306 if (dlmlock->l_granted_mode != dlmlock->l_req_mode) {
307 dlmlock->l_ast_data = NULL;
308 unlock_res_and_lock(dlmlock);
312 discard = ldlm_is_discard_data(dlmlock);
313 if (dlmlock->l_granted_mode & (LCK_PW | LCK_GROUP))
316 if (dlmlock->l_ast_data != NULL) {
317 obj = osc2cl(dlmlock->l_ast_data);
318 dlmlock->l_ast_data = NULL;
321 ldlm_set_kms_ignore(dlmlock);
322 unlock_res_and_lock(dlmlock);
324 /* if l_ast_data is NULL, the dlmlock was enqueued by AGL or
325 * the object has been destroyed. */
327 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
329 /* Destroy pages covered by the extent of the DLM lock */
330 result = mdc_lock_flush(env, cl2osc(obj), cl_index(obj, 0),
331 CL_PAGE_EOF, mode, discard);
332 /* Losing a lock, set KMS to 0.
333 * NB: assumed that DOM lock covers whole data on MDT.
335 /* losing a lock, update kms */
336 lock_res_and_lock(dlmlock);
337 cl_object_attr_lock(obj);
339 cl_object_attr_update(env, obj, attr, CAT_KMS);
340 cl_object_attr_unlock(obj);
341 unlock_res_and_lock(dlmlock);
342 cl_object_put(env, obj);
347 int mdc_ldlm_blocking_ast(struct ldlm_lock *dlmlock,
348 struct ldlm_lock_desc *new, void *data, int flag)
355 case LDLM_CB_BLOCKING: {
356 struct lustre_handle lockh;
358 ldlm_lock2handle(dlmlock, &lockh);
359 rc = ldlm_cli_cancel(&lockh, LCF_ASYNC);
364 case LDLM_CB_CANCELING: {
369 * This can be called in the context of outer IO, e.g.,
371 * osc_enqueue_base()->...
372 * ->ldlm_prep_elc_req()->...
373 * ->ldlm_cancel_callback()->...
374 * ->osc_ldlm_blocking_ast()
376 * new environment has to be created to not corrupt outer
379 env = cl_env_get(&refcheck);
385 rc = mdc_dlm_blocking_ast0(env, dlmlock, data, flag);
386 cl_env_put(env, &refcheck);
396 * Updates object attributes from a lock value block (lvb) received together
397 * with the DLM lock reply from the server.
398 * This can be optimized to not update attributes when lock is a result of a
401 * Called under lock and resource spin-locks.
403 void mdc_lock_lvb_update(const struct lu_env *env, struct osc_object *osc,
404 struct ldlm_lock *dlmlock, struct ost_lvb *lvb)
406 struct cl_object *obj = osc2cl(osc);
407 struct lov_oinfo *oinfo = osc->oo_oinfo;
408 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
409 unsigned valid = CAT_BLOCKS | CAT_ATIME | CAT_CTIME | CAT_MTIME |
415 LASSERT(dlmlock != NULL);
416 lvb = &dlmlock->l_ost_lvb;
418 cl_lvb2attr(attr, lvb);
420 cl_object_attr_lock(obj);
421 if (dlmlock != NULL) {
424 check_res_locked(dlmlock->l_resource);
425 size = lvb->lvb_size;
427 if (size >= oinfo->loi_kms) {
428 LDLM_DEBUG(dlmlock, "lock acquired, setting rss=%llu,"
429 " kms=%llu", lvb->lvb_size, size);
431 attr->cat_kms = size;
433 LDLM_DEBUG(dlmlock, "lock acquired, setting rss=%llu,"
435 lvb->lvb_size, oinfo->loi_kms);
438 cl_object_attr_update(env, obj, attr, valid);
439 cl_object_attr_unlock(obj);
443 static void mdc_lock_granted(const struct lu_env *env, struct osc_lock *oscl,
444 struct lustre_handle *lockh, bool lvb_update)
446 struct ldlm_lock *dlmlock;
450 dlmlock = ldlm_handle2lock_long(lockh, 0);
451 LASSERT(dlmlock != NULL);
453 /* lock reference taken by ldlm_handle2lock_long() is
454 * owned by osc_lock and released in osc_lock_detach()
456 lu_ref_add(&dlmlock->l_reference, "osc_lock", oscl);
457 oscl->ols_has_ref = 1;
459 LASSERT(oscl->ols_dlmlock == NULL);
460 oscl->ols_dlmlock = dlmlock;
462 /* This may be a matched lock for glimpse request, do not hold
463 * lock reference in that case. */
464 if (!oscl->ols_glimpse) {
465 /* hold a refc for non glimpse lock which will
466 * be released in osc_lock_cancel() */
467 lustre_handle_copy(&oscl->ols_handle, lockh);
468 ldlm_lock_addref(lockh, oscl->ols_einfo.ei_mode);
472 /* Lock must have been granted. */
473 lock_res_and_lock(dlmlock);
474 if (dlmlock->l_granted_mode == dlmlock->l_req_mode) {
475 struct cl_lock_descr *descr = &oscl->ols_cl.cls_lock->cll_descr;
477 /* extend the lock extent, otherwise it will have problem when
478 * we decide whether to grant a lockless lock. */
479 descr->cld_mode = osc_ldlm2cl_lock(dlmlock->l_granted_mode);
480 descr->cld_start = cl_index(descr->cld_obj, 0);
481 descr->cld_end = CL_PAGE_EOF;
483 /* no lvb update for matched lock */
485 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
486 mdc_lock_lvb_update(env, cl2osc(oscl->ols_cl.cls_obj),
490 unlock_res_and_lock(dlmlock);
492 LASSERT(oscl->ols_state != OLS_GRANTED);
493 oscl->ols_state = OLS_GRANTED;
498 * Lock upcall function that is executed either when a reply to ENQUEUE rpc is
499 * received from a server, or after osc_enqueue_base() matched a local DLM
502 static int mdc_lock_upcall(void *cookie, struct lustre_handle *lockh,
505 struct osc_lock *oscl = cookie;
506 struct cl_lock_slice *slice = &oscl->ols_cl;
512 env = cl_env_percpu_get();
513 /* should never happen, similar to osc_ldlm_blocking_ast(). */
514 LASSERT(!IS_ERR(env));
516 rc = ldlm_error2errno(errcode);
517 if (oscl->ols_state == OLS_ENQUEUED) {
518 oscl->ols_state = OLS_UPCALL_RECEIVED;
519 } else if (oscl->ols_state == OLS_CANCELLED) {
522 CERROR("Impossible state: %d\n", oscl->ols_state);
526 CDEBUG(D_INODE, "rc %d, err %d\n", rc, errcode);
528 mdc_lock_granted(env, oscl, lockh, errcode == ELDLM_OK);
530 /* Error handling, some errors are tolerable. */
531 if (oscl->ols_locklessable && rc == -EUSERS) {
532 /* This is a tolerable error, turn this lock into
535 osc_object_set_contended(cl2osc(slice->cls_obj));
536 LASSERT(slice->cls_ops != oscl->ols_lockless_ops);
538 /* Change this lock to ldlmlock-less lock. */
539 osc_lock_to_lockless(env, oscl, 1);
540 oscl->ols_state = OLS_GRANTED;
542 } else if (oscl->ols_glimpse && rc == -ENAVAIL) {
543 LASSERT(oscl->ols_flags & LDLM_FL_LVB_READY);
544 mdc_lock_lvb_update(env, cl2osc(slice->cls_obj),
545 NULL, &oscl->ols_lvb);
546 /* Hide the error. */
550 if (oscl->ols_owner != NULL)
551 cl_sync_io_note(env, oscl->ols_owner, rc);
552 cl_env_percpu_put(env);
557 int mdc_fill_lvb(struct ptlrpc_request *req, struct ost_lvb *lvb)
559 struct mdt_body *body;
561 body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
565 lvb->lvb_mtime = body->mbo_mtime;
566 lvb->lvb_atime = body->mbo_atime;
567 lvb->lvb_ctime = body->mbo_ctime;
568 lvb->lvb_blocks = body->mbo_dom_blocks;
569 lvb->lvb_size = body->mbo_dom_size;
574 int mdc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
575 void *cookie, struct lustre_handle *lockh,
576 enum ldlm_mode mode, __u64 *flags, int errcode)
578 struct osc_lock *ols = cookie;
579 struct ldlm_lock *lock;
584 /* The request was created before ldlm_cli_enqueue call. */
585 if (errcode == ELDLM_LOCK_ABORTED) {
586 struct ldlm_reply *rep;
588 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
589 LASSERT(rep != NULL);
591 rep->lock_policy_res2 =
592 ptlrpc_status_ntoh(rep->lock_policy_res2);
593 if (rep->lock_policy_res2)
594 errcode = rep->lock_policy_res2;
596 rc = mdc_fill_lvb(req, &ols->ols_lvb);
597 *flags |= LDLM_FL_LVB_READY;
598 } else if (errcode == ELDLM_OK) {
599 /* Callers have references, should be valid always */
600 lock = ldlm_handle2lock(lockh);
603 rc = mdc_fill_lvb(req, &lock->l_ost_lvb);
605 *flags |= LDLM_FL_LVB_READY;
608 /* Call the update callback. */
609 rc = (*upcall)(cookie, lockh, rc < 0 ? rc : errcode);
611 /* release the reference taken in ldlm_cli_enqueue() */
612 if (errcode == ELDLM_LOCK_MATCHED)
614 if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
615 ldlm_lock_decref(lockh, mode);
620 int mdc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
621 struct osc_enqueue_args *aa, int rc)
623 struct ldlm_lock *lock;
624 struct lustre_handle *lockh = &aa->oa_lockh;
625 enum ldlm_mode mode = aa->oa_mode;
629 LASSERT(!aa->oa_speculative);
631 /* ldlm_cli_enqueue is holding a reference on the lock, so it must
633 lock = ldlm_handle2lock(lockh);
634 LASSERTF(lock != NULL,
635 "lockh %#llx, req %p, aa %p - client evicted?\n",
636 lockh->cookie, req, aa);
638 /* Take an additional reference so that a blocking AST that
639 * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
640 * to arrive after an upcall has been executed by
641 * osc_enqueue_fini(). */
642 ldlm_lock_addref(lockh, mode);
644 /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
645 OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
647 /* Let CP AST to grant the lock first. */
648 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
650 /* Complete obtaining the lock procedure. */
651 rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, aa->oa_type, 1,
652 aa->oa_mode, aa->oa_flags, NULL, 0,
654 /* Complete mdc stuff. */
655 rc = mdc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
658 OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
660 ldlm_lock_decref(lockh, mode);
665 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
666 * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
667 * other synchronous requests, however keeping some locks and trying to obtain
668 * others may take a considerable amount of time in a case of ost failure; and
669 * when other sync requests do not get released lock from a client, the client
670 * is excluded from the cluster -- such scenarious make the life difficult, so
671 * release locks just after they are obtained. */
672 int mdc_enqueue_send(const struct lu_env *env, struct obd_export *exp,
673 struct ldlm_res_id *res_id, __u64 *flags,
674 union ldlm_policy_data *policy,
675 struct ost_lvb *lvb, int kms_valid,
676 osc_enqueue_upcall_f upcall, void *cookie,
677 struct ldlm_enqueue_info *einfo, int async)
679 struct obd_device *obd = exp->exp_obd;
680 struct lustre_handle lockh = { 0 };
681 struct ptlrpc_request *req = NULL;
682 struct ldlm_intent *lit;
684 bool glimpse = *flags & LDLM_FL_HAS_INTENT;
685 __u64 match_flags = *flags;
690 mode = einfo->ei_mode;
691 if (einfo->ei_mode == LCK_PR)
695 match_flags |= LDLM_FL_BLOCK_GRANTED;
696 mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
697 einfo->ei_type, policy, mode, &lockh, 0);
699 struct ldlm_lock *matched;
701 if (*flags & LDLM_FL_TEST_LOCK)
704 matched = ldlm_handle2lock(&lockh);
705 if (ldlm_is_kms_ignore(matched))
708 if (mdc_set_dom_lock_data(env, matched, einfo->ei_cbdata)) {
709 *flags |= LDLM_FL_LVB_READY;
711 /* We already have a lock, and it's referenced. */
712 (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
714 ldlm_lock_decref(&lockh, mode);
715 LDLM_LOCK_PUT(matched);
719 ldlm_lock_decref(&lockh, mode);
720 LDLM_LOCK_PUT(matched);
723 if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
726 req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LDLM_INTENT);
730 rc = ldlm_prep_enqueue_req(exp, req, NULL, 0);
732 ptlrpc_request_free(req);
736 /* pack the intent */
737 lit = req_capsule_client_get(&req->rq_pill, &RMF_LDLM_INTENT);
738 lit->opc = glimpse ? IT_GLIMPSE : IT_BRW;
740 req_capsule_set_size(&req->rq_pill, &RMF_MDT_MD, RCL_SERVER, 0);
741 req_capsule_set_size(&req->rq_pill, &RMF_ACL, RCL_SERVER, 0);
742 ptlrpc_request_set_replen(req);
744 /* users of mdc_enqueue() can pass this flag for ldlm_lock_match() */
745 *flags &= ~LDLM_FL_BLOCK_GRANTED;
746 /* All MDC IO locks are intents */
747 *flags |= LDLM_FL_HAS_INTENT;
748 rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, NULL,
749 0, LVB_T_NONE, &lockh, async);
752 struct osc_enqueue_args *aa;
754 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
755 aa = ptlrpc_req_async_args(req);
757 aa->oa_mode = einfo->ei_mode;
758 aa->oa_type = einfo->ei_type;
759 lustre_handle_copy(&aa->oa_lockh, &lockh);
760 aa->oa_upcall = upcall;
761 aa->oa_cookie = cookie;
762 aa->oa_speculative = false;
763 aa->oa_flags = flags;
766 req->rq_interpret_reply =
767 (ptlrpc_interpterer_t)mdc_enqueue_interpret;
768 ptlrpcd_add_req(req);
770 ptlrpc_req_finished(req);
775 rc = mdc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
777 ptlrpc_req_finished(req);
782 * Implementation of cl_lock_operations::clo_enqueue() method for osc
783 * layer. This initiates ldlm enqueue:
785 * - cancels conflicting locks early (osc_lock_enqueue_wait());
787 * - calls osc_enqueue_base() to do actual enqueue.
789 * osc_enqueue_base() is supplied with an upcall function that is executed
790 * when lock is received either after a local cached ldlm lock is matched, or
791 * when a reply from the server is received.
793 * This function does not wait for the network communication to complete.
795 static int mdc_lock_enqueue(const struct lu_env *env,
796 const struct cl_lock_slice *slice,
797 struct cl_io *unused, struct cl_sync_io *anchor)
799 struct osc_thread_info *info = osc_env_info(env);
800 struct osc_io *oio = osc_env_io(env);
801 struct osc_object *osc = cl2osc(slice->cls_obj);
802 struct osc_lock *oscl = cl2osc_lock(slice);
803 struct cl_lock *lock = slice->cls_lock;
804 struct ldlm_res_id *resname = &info->oti_resname;
805 union ldlm_policy_data *policy = &info->oti_policy;
806 osc_enqueue_upcall_f upcall = mdc_lock_upcall;
807 void *cookie = (void *)oscl;
813 LASSERTF(ergo(oscl->ols_glimpse, lock->cll_descr.cld_mode <= CLM_READ),
814 "lock = %p, ols = %p\n", lock, oscl);
816 if (oscl->ols_state == OLS_GRANTED)
819 /* Lockahead is not supported on MDT yet */
820 if (oscl->ols_flags & LDLM_FL_NO_EXPANSION) {
821 result = -EOPNOTSUPP;
825 if (oscl->ols_flags & LDLM_FL_TEST_LOCK)
826 GOTO(enqueue_base, 0);
828 if (oscl->ols_glimpse) {
829 LASSERT(equi(oscl->ols_speculative, anchor == NULL));
831 GOTO(enqueue_base, 0);
834 result = osc_lock_enqueue_wait(env, osc, oscl);
838 /* we can grant lockless lock right after all conflicting locks
840 if (osc_lock_is_lockless(oscl)) {
841 oscl->ols_state = OLS_GRANTED;
842 oio->oi_lockless = 1;
847 oscl->ols_state = OLS_ENQUEUED;
848 if (anchor != NULL) {
849 atomic_inc(&anchor->csi_sync_nr);
850 oscl->ols_owner = anchor;
854 * DLM lock's ast data must be osc_object;
855 * DLM's enqueue callback set to osc_lock_upcall() with cookie as
858 fid_build_reg_res_name(lu_object_fid(osc2lu(osc)), resname);
859 mdc_lock_build_policy(env, policy);
860 LASSERT(!oscl->ols_speculative);
861 result = mdc_enqueue_send(env, osc_export(osc), resname,
862 &oscl->ols_flags, policy,
863 &oscl->ols_lvb, osc->oo_oinfo->loi_kms_valid,
864 upcall, cookie, &oscl->ols_einfo, async);
866 if (osc_lock_is_lockless(oscl)) {
867 oio->oi_lockless = 1;
869 LASSERT(oscl->ols_state == OLS_GRANTED);
870 LASSERT(oscl->ols_hold);
871 LASSERT(oscl->ols_dlmlock != NULL);
876 oscl->ols_state = OLS_CANCELLED;
877 osc_lock_wake_waiters(env, osc, oscl);
880 cl_sync_io_note(env, anchor, result);
885 static const struct cl_lock_operations mdc_lock_lockless_ops = {
886 .clo_fini = osc_lock_fini,
887 .clo_enqueue = mdc_lock_enqueue,
888 .clo_cancel = mdc_lock_lockless_cancel,
889 .clo_print = osc_lock_print
892 static const struct cl_lock_operations mdc_lock_ops = {
893 .clo_fini = osc_lock_fini,
894 .clo_enqueue = mdc_lock_enqueue,
895 .clo_cancel = osc_lock_cancel,
896 .clo_print = osc_lock_print,
899 int mdc_lock_init(const struct lu_env *env, struct cl_object *obj,
900 struct cl_lock *lock, const struct cl_io *io)
902 struct osc_lock *ols;
903 __u32 enqflags = lock->cll_descr.cld_enq_flags;
904 __u64 flags = osc_enq2ldlm_flags(enqflags);
908 /* Ignore AGL for Data-on-MDT, stat returns size data */
909 if ((enqflags & CEF_SPECULATIVE) != 0)
912 OBD_SLAB_ALLOC_PTR_GFP(ols, osc_lock_kmem, GFP_NOFS);
913 if (unlikely(ols == NULL))
916 ols->ols_state = OLS_NEW;
917 spin_lock_init(&ols->ols_lock);
918 INIT_LIST_HEAD(&ols->ols_waiting_list);
919 INIT_LIST_HEAD(&ols->ols_wait_entry);
920 INIT_LIST_HEAD(&ols->ols_nextlock_oscobj);
921 ols->ols_lockless_ops = &mdc_lock_lockless_ops;
923 ols->ols_flags = flags;
924 ols->ols_speculative = !!(enqflags & CEF_SPECULATIVE);
926 if (ols->ols_flags & LDLM_FL_HAS_INTENT) {
927 ols->ols_flags |= LDLM_FL_BLOCK_GRANTED;
928 ols->ols_glimpse = 1;
930 mdc_lock_build_einfo(env, lock, cl2osc(obj), &ols->ols_einfo);
932 cl_lock_slice_add(lock, &ols->ols_cl, obj, &mdc_lock_ops);
934 if (!(enqflags & CEF_MUST))
935 osc_lock_to_lockless(env, ols, (enqflags & CEF_NEVER));
936 if (ols->ols_locklessable && !(enqflags & CEF_DISCARD_DATA))
937 ols->ols_flags |= LDLM_FL_DENY_ON_CONTENTION;
939 if (io->ci_type == CIT_WRITE || cl_io_is_mkwrite(io))
940 osc_lock_set_writer(env, io, obj, ols);
942 LDLM_DEBUG_NOLOCK("lock %p, mdc lock %p, flags %llx\n",
943 lock, ols, ols->ols_flags);
950 * An implementation of cl_io_operations specific methods for MDC layer.
953 static int mdc_async_upcall(void *a, int rc)
955 struct osc_async_cbargs *args = a;
958 complete(&args->opc_sync);
962 static int mdc_io_setattr_start(const struct lu_env *env,
963 const struct cl_io_slice *slice)
965 struct cl_io *io = slice->cis_io;
966 struct osc_io *oio = cl2osc_io(env, slice);
967 struct cl_object *obj = slice->cis_obj;
968 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
969 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
970 struct obdo *oa = &oio->oi_oa;
971 struct osc_async_cbargs *cbargs = &oio->oi_cbarg;
972 __u64 size = io->u.ci_setattr.sa_attr.lvb_size;
973 unsigned int ia_valid = io->u.ci_setattr.sa_valid;
976 /* silently ignore non-truncate setattr for Data-on-MDT object */
977 if (cl_io_is_trunc(io)) {
978 /* truncate cache dirty pages first */
979 rc = osc_cache_truncate_start(env, cl2osc(obj), size,
985 if (oio->oi_lockless == 0) {
986 cl_object_attr_lock(obj);
987 rc = cl_object_attr_get(env, obj, attr);
989 struct ost_lvb *lvb = &io->u.ci_setattr.sa_attr;
990 unsigned int cl_valid = 0;
992 if (ia_valid & ATTR_SIZE) {
993 attr->cat_size = attr->cat_kms = size;
994 cl_valid = (CAT_SIZE | CAT_KMS);
996 if (ia_valid & ATTR_MTIME_SET) {
997 attr->cat_mtime = lvb->lvb_mtime;
998 cl_valid |= CAT_MTIME;
1000 if (ia_valid & ATTR_ATIME_SET) {
1001 attr->cat_atime = lvb->lvb_atime;
1002 cl_valid |= CAT_ATIME;
1004 if (ia_valid & ATTR_CTIME_SET) {
1005 attr->cat_ctime = lvb->lvb_ctime;
1006 cl_valid |= CAT_CTIME;
1008 rc = cl_object_attr_update(env, obj, attr, cl_valid);
1010 cl_object_attr_unlock(obj);
1015 if (!(ia_valid & ATTR_SIZE))
1018 memset(oa, 0, sizeof(*oa));
1019 oa->o_oi = loi->loi_oi;
1020 oa->o_mtime = attr->cat_mtime;
1021 oa->o_atime = attr->cat_atime;
1022 oa->o_ctime = attr->cat_ctime;
1025 oa->o_blocks = OBD_OBJECT_EOF;
1026 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLATIME |
1027 OBD_MD_FLCTIME | OBD_MD_FLMTIME | OBD_MD_FLSIZE |
1029 if (oio->oi_lockless) {
1030 oa->o_flags = OBD_FL_SRVLOCK;
1031 oa->o_valid |= OBD_MD_FLFLAGS;
1034 init_completion(&cbargs->opc_sync);
1036 rc = osc_punch_send(osc_export(cl2osc(obj)), oa,
1037 mdc_async_upcall, cbargs);
1038 cbargs->opc_rpc_sent = rc == 0;
1042 static int mdc_io_read_ahead(const struct lu_env *env,
1043 const struct cl_io_slice *ios,
1044 pgoff_t start, struct cl_read_ahead *ra)
1046 struct osc_object *osc = cl2osc(ios->cis_obj);
1047 struct ldlm_lock *dlmlock;
1051 dlmlock = mdc_dlmlock_at_pgoff(env, osc, start, 0);
1052 if (dlmlock == NULL)
1055 if (dlmlock->l_req_mode != LCK_PR) {
1056 struct lustre_handle lockh;
1058 ldlm_lock2handle(dlmlock, &lockh);
1059 ldlm_lock_addref(&lockh, LCK_PR);
1060 ldlm_lock_decref(&lockh, dlmlock->l_req_mode);
1063 ra->cra_rpc_size = osc_cli(osc)->cl_max_pages_per_rpc;
1064 ra->cra_end = CL_PAGE_EOF;
1065 ra->cra_release = osc_read_ahead_release;
1066 ra->cra_cbdata = dlmlock;
1071 int mdc_io_fsync_start(const struct lu_env *env,
1072 const struct cl_io_slice *slice)
1074 struct cl_io *io = slice->cis_io;
1075 struct cl_fsync_io *fio = &io->u.ci_fsync;
1076 struct cl_object *obj = slice->cis_obj;
1077 struct osc_object *osc = cl2osc(obj);
1082 /* a MDC lock always covers whole object, do sync for whole
1083 * possible range despite of supplied start/end values.
1085 result = osc_cache_writeback_range(env, osc, 0, CL_PAGE_EOF, 0,
1086 fio->fi_mode == CL_FSYNC_DISCARD);
1088 fio->fi_nr_written += result;
1091 if (fio->fi_mode == CL_FSYNC_ALL) {
1094 rc = osc_cache_wait_range(env, osc, 0, CL_PAGE_EOF);
1097 /* Use OSC sync code because it is asynchronous.
1098 * It is to be added into MDC and avoid the using of
1099 * OST_SYNC at both MDC and MDT.
1101 rc = osc_fsync_ost(env, osc, fio);
1109 static struct cl_io_operations mdc_io_ops = {
1112 .cio_iter_init = osc_io_iter_init,
1113 .cio_iter_fini = osc_io_iter_fini,
1114 .cio_start = osc_io_read_start,
1117 .cio_iter_init = osc_io_write_iter_init,
1118 .cio_iter_fini = osc_io_write_iter_fini,
1119 .cio_start = osc_io_write_start,
1120 .cio_end = osc_io_end,
1123 .cio_iter_init = osc_io_iter_init,
1124 .cio_iter_fini = osc_io_iter_fini,
1125 .cio_start = mdc_io_setattr_start,
1126 .cio_end = osc_io_setattr_end,
1128 /* no support for data version so far */
1129 [CIT_DATA_VERSION] = {
1134 .cio_iter_init = osc_io_iter_init,
1135 .cio_iter_fini = osc_io_iter_fini,
1136 .cio_start = osc_io_fault_start,
1137 .cio_end = osc_io_end,
1140 .cio_start = mdc_io_fsync_start,
1141 .cio_end = osc_io_fsync_end,
1144 .cio_read_ahead = mdc_io_read_ahead,
1145 .cio_submit = osc_io_submit,
1146 .cio_commit_async = osc_io_commit_async,
1149 int mdc_io_init(const struct lu_env *env, struct cl_object *obj,
1152 struct osc_io *oio = osc_env_io(env);
1154 CL_IO_SLICE_CLEAN(oio, oi_cl);
1155 cl_io_slice_add(io, &oio->oi_cl, obj, &mdc_io_ops);
1159 static void mdc_build_res_name(struct osc_object *osc,
1160 struct ldlm_res_id *resname)
1162 fid_build_reg_res_name(lu_object_fid(osc2lu(osc)), resname);
1166 * Implementation of struct cl_req_operations::cro_attr_set() for MDC
1167 * layer. MDC is responsible for struct obdo::o_id and struct obdo::o_seq
1170 static void mdc_req_attr_set(const struct lu_env *env, struct cl_object *obj,
1171 struct cl_req_attr *attr)
1173 u64 flags = attr->cra_flags;
1175 /* Copy object FID to cl_attr */
1176 attr->cra_oa->o_oi.oi_fid = *lu_object_fid(&obj->co_lu);
1178 if (flags & OBD_MD_FLGROUP)
1179 attr->cra_oa->o_valid |= OBD_MD_FLGROUP;
1181 if (flags & OBD_MD_FLID)
1182 attr->cra_oa->o_valid |= OBD_MD_FLID;
1184 if (flags & OBD_MD_FLHANDLE) {
1185 struct ldlm_lock *lock; /* _some_ lock protecting @apage */
1186 struct osc_page *opg;
1188 opg = osc_cl_page_osc(attr->cra_page, cl2osc(obj));
1189 lock = mdc_dlmlock_at_pgoff(env, cl2osc(obj), osc_index(opg),
1190 OSC_DAP_FL_TEST_LOCK | OSC_DAP_FL_CANCELING);
1191 if (lock == NULL && !opg->ops_srvlock) {
1192 struct ldlm_resource *res;
1193 struct ldlm_res_id *resname;
1195 CL_PAGE_DEBUG(D_ERROR, env, attr->cra_page,
1196 "uncovered page!\n");
1198 resname = &osc_env_info(env)->oti_resname;
1199 mdc_build_res_name(cl2osc(obj), resname);
1200 res = ldlm_resource_get(
1201 osc_export(cl2osc(obj))->exp_obd->obd_namespace,
1202 NULL, resname, LDLM_IBITS, 0);
1203 ldlm_resource_dump(D_ERROR, res);
1205 libcfs_debug_dumpstack(NULL);
1209 /* check for lockless io. */
1211 attr->cra_oa->o_handle = lock->l_remote_handle;
1212 attr->cra_oa->o_valid |= OBD_MD_FLHANDLE;
1213 LDLM_LOCK_PUT(lock);
1218 static int mdc_attr_get(const struct lu_env *env, struct cl_object *obj,
1219 struct cl_attr *attr)
1221 struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
1223 if (OST_LVB_IS_ERR(oinfo->loi_lvb.lvb_blocks))
1224 return OST_LVB_GET_ERR(oinfo->loi_lvb.lvb_blocks);
1226 return osc_attr_get(env, obj, attr);
1229 static int mdc_object_ast_clear(struct ldlm_lock *lock, void *data)
1233 if ((lock->l_ast_data == NULL && !ldlm_is_kms_ignore(lock)) ||
1234 (lock->l_ast_data == data)) {
1235 lock->l_ast_data = NULL;
1236 ldlm_set_kms_ignore(lock);
1238 RETURN(LDLM_ITER_CONTINUE);
1241 int mdc_object_prune(const struct lu_env *env, struct cl_object *obj)
1243 struct osc_object *osc = cl2osc(obj);
1244 struct ldlm_res_id *resname = &osc_env_info(env)->oti_resname;
1246 /* DLM locks don't hold a reference of osc_object so we have to
1247 * clear it before the object is being destroyed. */
1248 osc_build_res_name(osc, resname);
1249 ldlm_resource_iterate(osc_export(osc)->exp_obd->obd_namespace, resname,
1250 mdc_object_ast_clear, osc);
1254 static const struct cl_object_operations mdc_ops = {
1255 .coo_page_init = osc_page_init,
1256 .coo_lock_init = mdc_lock_init,
1257 .coo_io_init = mdc_io_init,
1258 .coo_attr_get = mdc_attr_get,
1259 .coo_attr_update = osc_attr_update,
1260 .coo_glimpse = osc_object_glimpse,
1261 .coo_req_attr_set = mdc_req_attr_set,
1262 .coo_prune = mdc_object_prune,
1265 static const struct osc_object_operations mdc_object_ops = {
1266 .oto_build_res_name = mdc_build_res_name,
1267 .oto_dlmlock_at_pgoff = mdc_dlmlock_at_pgoff,
1270 static int mdc_object_init(const struct lu_env *env, struct lu_object *obj,
1271 const struct lu_object_conf *conf)
1273 struct osc_object *osc = lu2osc(obj);
1275 if (osc->oo_initialized)
1278 osc->oo_initialized = true;
1280 return osc_object_init(env, obj, conf);
1283 static void mdc_object_free(const struct lu_env *env, struct lu_object *obj)
1285 osc_object_free(env, obj);
1288 static const struct lu_object_operations mdc_lu_obj_ops = {
1289 .loo_object_init = mdc_object_init,
1290 .loo_object_delete = NULL,
1291 .loo_object_release = NULL,
1292 .loo_object_free = mdc_object_free,
1293 .loo_object_print = osc_object_print,
1294 .loo_object_invariant = NULL
1297 struct lu_object *mdc_object_alloc(const struct lu_env *env,
1298 const struct lu_object_header *unused,
1299 struct lu_device *dev)
1301 struct osc_object *osc;
1302 struct lu_object *obj;
1304 OBD_SLAB_ALLOC_PTR_GFP(osc, osc_object_kmem, GFP_NOFS);
1307 lu_object_init(obj, NULL, dev);
1308 osc->oo_cl.co_ops = &mdc_ops;
1309 obj->lo_ops = &mdc_lu_obj_ops;
1310 osc->oo_obj_ops = &mdc_object_ops;
1311 osc->oo_initialized = false;
1318 static int mdc_cl_process_config(const struct lu_env *env,
1319 struct lu_device *d, struct lustre_cfg *cfg)
1321 return mdc_process_config(d->ld_obd, 0, cfg);
1324 const struct lu_device_operations mdc_lu_ops = {
1325 .ldo_object_alloc = mdc_object_alloc,
1326 .ldo_process_config = mdc_cl_process_config,
1327 .ldo_recovery_complete = NULL,
1330 static struct lu_device *mdc_device_alloc(const struct lu_env *env,
1331 struct lu_device_type *t,
1332 struct lustre_cfg *cfg)
1334 struct lu_device *d;
1335 struct osc_device *od;
1336 struct obd_device *obd;
1341 RETURN(ERR_PTR(-ENOMEM));
1343 cl_device_init(&od->od_cl, t);
1345 d->ld_ops = &mdc_lu_ops;
1348 obd = class_name2obd(lustre_cfg_string(cfg, 0));
1350 RETURN(ERR_PTR(-ENODEV));
1352 rc = mdc_setup(obd, cfg);
1354 osc_device_free(env, d);
1355 RETURN(ERR_PTR(rc));
1357 od->od_exp = obd->obd_self_export;
1361 static const struct lu_device_type_operations mdc_device_type_ops = {
1362 .ldto_device_alloc = mdc_device_alloc,
1363 .ldto_device_free = osc_device_free,
1364 .ldto_device_init = osc_device_init,
1365 .ldto_device_fini = osc_device_fini
1368 struct lu_device_type mdc_device_type = {
1369 .ldt_tags = LU_DEVICE_CL,
1370 .ldt_name = LUSTRE_MDC_NAME,
1371 .ldt_ops = &mdc_device_type_ops,
1372 .ldt_ctx_tags = LCT_CL_THREAD