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) 2014, Intel Corporation.
27 * lustre/target/update_recovery.c
29 * This file implement the methods to handle the update recovery.
31 * During DNE recovery, the recovery thread will redo the operation according
32 * to the transaction no, and these replay are either from client replay req
33 * or update replay records(for distribute transaction) in the update log.
34 * For distribute transaction replay, the replay thread will call
35 * distribute_txn_replay_handle() to handle the updates.
37 * After the Master MDT restarts, it will retrieve the update records from all
38 * of MDTs, for each distributed operation, it will check updates on all MDTs,
39 * if some updates records are missing on some MDTs, the replay thread will redo
40 * updates on these MDTs.
42 * Author: Di Wang <di.wang@intel.com>
44 #define DEBUG_SUBSYSTEM S_CLASS
46 #include <lu_target.h>
47 #include <md_object.h>
48 #include <lustre_update.h>
50 #include <obd_class.h>
51 #include "tgt_internal.h"
54 * Lookup distribute_txn_replay req
56 * Lookup distribute_txn_replay in the replay list by batchid.
57 * It is assumed the list has been locked before calling this function.
59 * \param[in] tdtd distribute_txn_data, which holds the replay
61 * \param[in] batchid batchid used by lookup.
63 * \retval pointer of the replay if succeeds.
64 * \retval NULL if can not find it.
66 static struct distribute_txn_replay_req *
67 dtrq_lookup(struct target_distribute_txn_data *tdtd, __u64 batchid)
69 struct distribute_txn_replay_req *tmp;
70 struct distribute_txn_replay_req *dtrq = NULL;
72 list_for_each_entry(tmp, &tdtd->tdtd_replay_list, dtrq_list) {
73 if (tmp->dtrq_batchid == batchid) {
82 * insert distribute txn replay req
84 * Insert distribute txn replay to the replay list, and it assumes the
85 * list has been looked. Note: the replay list is a sorted list, which
86 * is sorted by master transno. It is assumed the replay list has been
87 * locked before calling this function.
89 * \param[in] tdtd target distribute txn data where replay list is
90 * \param[in] new distribute txn replay to be inserted
92 * \retval 0 if insertion succeeds
93 * \retval EEXIST if the dtrq already exists
95 static int dtrq_insert(struct target_distribute_txn_data *tdtd,
96 struct distribute_txn_replay_req *new)
98 struct distribute_txn_replay_req *iter;
100 /* Check if the dtrq has been added to the list */
101 iter = dtrq_lookup(tdtd, new->dtrq_batchid);
105 list_for_each_entry_reverse(iter, &tdtd->tdtd_replay_list, dtrq_list) {
106 if (iter->dtrq_master_transno > new->dtrq_master_transno)
109 /* If there are mulitple replay req with same transno, then
110 * sort them with batchid */
111 if (iter->dtrq_master_transno == new->dtrq_master_transno &&
112 iter->dtrq_batchid > new->dtrq_batchid)
115 list_add(&new->dtrq_list, &iter->dtrq_list);
119 if (list_empty(&new->dtrq_list))
120 list_add(&new->dtrq_list, &tdtd->tdtd_replay_list);
126 * create distribute txn replay req
128 * Allocate distribute txn replay req according to the update records.
130 * \param[in] tdtd target distribute txn data where replay list is.
131 * \param[in] record update records from the update log.
133 * \retval the pointer of distribute txn replay req if
134 * the creation succeeds.
135 * \retval NULL if the creation fails.
137 static struct distribute_txn_replay_req *
138 dtrq_create(struct target_distribute_txn_data *tdtd,
139 struct llog_update_record *lur)
141 struct distribute_txn_replay_req *new;
145 RETURN(ERR_PTR(-ENOMEM));
147 new->dtrq_lur_size = llog_update_record_size(lur);
148 OBD_ALLOC_LARGE(new->dtrq_lur, new->dtrq_lur_size);
149 if (new->dtrq_lur == NULL) {
151 RETURN(ERR_PTR(-ENOMEM));
154 memcpy(new->dtrq_lur, lur, new->dtrq_lur_size);
156 /* If the transno in the update record is 0, it means the
157 * update are from master MDT, and it will use the master
158 * last committed transno as its master transno. Later, if
159 * the update records are gotten from slave MDTs, then these
160 * transno will be replaced.
161 * See insert_update_records_to_replay_list(). */
162 if (lur->lur_update_rec.ur_master_transno == 0) {
163 new->dtrq_lur->lur_update_rec.ur_master_transno =
164 tdtd->tdtd_lut->lut_obd->obd_last_committed;
165 new->dtrq_master_transno =
166 tdtd->tdtd_lut->lut_obd->obd_last_committed;
168 new->dtrq_master_transno =
169 lur->lur_update_rec.ur_master_transno;
172 new->dtrq_batchid = lur->lur_update_rec.ur_batchid;
174 spin_lock_init(&new->dtrq_sub_list_lock);
175 INIT_LIST_HEAD(&new->dtrq_sub_list);
176 INIT_LIST_HEAD(&new->dtrq_list);
182 * Lookup distribute sub replay
184 * Lookup distribute sub replay in the sub list of distribute_txn_replay by
187 * \param[in] distribute_txn_replay_req the distribute txn replay req to lookup
188 * \param[in] mdt_index the mdt_index as the key of lookup
190 * \retval the pointer of sub replay if it can be found.
191 * \retval NULL if it can not find.
193 struct distribute_txn_replay_req_sub *
194 dtrq_sub_lookup(struct distribute_txn_replay_req *dtrq, __u32 mdt_index)
196 struct distribute_txn_replay_req_sub *dtrqs = NULL;
197 struct distribute_txn_replay_req_sub *tmp;
199 list_for_each_entry(tmp, &dtrq->dtrq_sub_list, dtrqs_list) {
200 if (tmp->dtrqs_mdt_index == mdt_index) {
209 * Try to add cookie to sub distribute txn request
211 * Check if the update log cookie has been added to the request, if not,
212 * add it to the dtrqs_cookie_list.
214 * \param[in] dtrqs sub replay req where cookies to be added.
215 * \param[in] cookie cookie to be added.
217 * \retval 0 if the cookie is adding succeeds.
218 * \retval negative errno if adding fails.
220 static int dtrq_sub_add_cookie(struct distribute_txn_replay_req_sub *dtrqs,
221 struct llog_cookie *cookie)
223 struct sub_thandle_cookie *new;
229 INIT_LIST_HEAD(&new->stc_list);
230 new->stc_cookie = *cookie;
231 /* Note: only single thread will access one sub_request each time,
232 * so no need lock here */
233 list_add(&new->stc_list, &dtrqs->dtrqs_cookie_list);
239 * Insert distribute txn sub req replay
241 * Allocate sub replay req and insert distribute txn replay list.
243 * \param[in] dtrq d to be added
244 * \param[in] cookie the cookie of the update record
245 * \param[in] mdt_index the mdt_index of the update record
247 * \retval 0 if the adding succeeds.
248 * \retval negative errno if the adding fails.
251 dtrq_sub_create_and_insert(struct distribute_txn_replay_req *dtrq,
252 struct llog_cookie *cookie,
255 struct distribute_txn_replay_req_sub *dtrqs = NULL;
256 struct distribute_txn_replay_req_sub *new;
260 spin_lock(&dtrq->dtrq_sub_list_lock);
261 dtrqs = dtrq_sub_lookup(dtrq, mdt_index);
262 spin_unlock(&dtrq->dtrq_sub_list_lock);
264 rc = dtrq_sub_add_cookie(dtrqs, cookie);
272 INIT_LIST_HEAD(&new->dtrqs_list);
273 INIT_LIST_HEAD(&new->dtrqs_cookie_list);
274 new->dtrqs_mdt_index = mdt_index;
275 spin_lock(&dtrq->dtrq_sub_list_lock);
276 dtrqs = dtrq_sub_lookup(dtrq, mdt_index);
278 list_add(&new->dtrqs_list, &dtrq->dtrq_sub_list);
283 spin_unlock(&dtrq->dtrq_sub_list_lock);
285 rc = dtrq_sub_add_cookie(dtrqs, cookie);
291 * append updates to the current replay updates
293 * Append more updates to the existent replay update. And this is only
294 * used when combining mulitple updates into one large updates during
297 * \param[in] dtrq the update replay request where the new update
298 * records will be added.
299 * \param[in] lur the new update record.
301 * \retval 0 if appending succeeds.
302 * \retval negative errno if appending fails.
304 static int dtrq_append_updates(struct distribute_txn_replay_req *dtrq,
305 struct update_records *record)
307 struct llog_update_record *new_lur;
308 size_t lur_size = dtrq->dtrq_lur_size;
312 /* Because several threads might retrieve the same records from
313 * different targets, and we only need one copy of records. So
314 * we will check if the records is in the next one, if not, just
316 spin_lock(&dtrq->dtrq_sub_list_lock);
317 if (dtrq->dtrq_lur->lur_update_rec.ur_index + 1 != record->ur_index) {
318 spin_unlock(&dtrq->dtrq_sub_list_lock);
321 dtrq->dtrq_lur->lur_update_rec.ur_index++;
322 spin_unlock(&dtrq->dtrq_sub_list_lock);
324 lur_size += update_records_size(record);
325 OBD_ALLOC_LARGE(new_lur, lur_size);
326 if (new_lur == NULL) {
327 spin_lock(&dtrq->dtrq_sub_list_lock);
328 dtrq->dtrq_lur->lur_update_rec.ur_index--;
329 spin_unlock(&dtrq->dtrq_sub_list_lock);
333 /* Copy the old and new records to the new allocated buffer */
334 memcpy(new_lur, dtrq->dtrq_lur, dtrq->dtrq_lur_size);
335 ptr = (char *)&new_lur->lur_update_rec +
336 update_records_size(&new_lur->lur_update_rec);
337 memcpy(ptr, &record->ur_ops,
338 update_records_size(record) -
339 offsetof(struct update_records, ur_ops));
341 new_lur->lur_update_rec.ur_update_count += record->ur_update_count;
342 new_lur->lur_update_rec.ur_param_count += record->ur_param_count;
343 new_lur->lur_hdr.lrh_len = llog_update_record_size(new_lur);
345 /* Replace the records */
346 OBD_FREE_LARGE(dtrq->dtrq_lur, dtrq->dtrq_lur_size);
347 dtrq->dtrq_lur = new_lur;
348 dtrq->dtrq_lur_size = lur_size;
349 dtrq->dtrq_lur->lur_update_rec.ur_flags = record->ur_flags;
350 update_records_dump(&new_lur->lur_update_rec, D_INFO, true);
355 * Insert update records to the replay list.
357 * Allocate distribute txn replay req and insert it into the replay
358 * list, then insert the update records into the replay req.
360 * \param[in] tdtd distribute txn replay data where the replay list
362 * \param[in] record the update record
363 * \param[in] cookie cookie of the record
364 * \param[in] index mdt index of the record
366 * \retval 0 if the adding succeeds.
367 * \retval negative errno if the adding fails.
370 insert_update_records_to_replay_list(struct target_distribute_txn_data *tdtd,
371 struct llog_update_record *lur,
372 struct llog_cookie *cookie,
375 struct distribute_txn_replay_req *dtrq;
376 struct update_records *record = &lur->lur_update_rec;
377 bool replace_record = false;
381 CDEBUG(D_HA, "%s: insert record batchid = "LPU64" transno = "LPU64
382 " mdt_index %u\n", tdtd->tdtd_lut->lut_obd->obd_name,
383 record->ur_batchid, record->ur_master_transno, mdt_index);
385 /* Update batchid if necessary */
386 spin_lock(&tdtd->tdtd_batchid_lock);
387 if (record->ur_batchid >= tdtd->tdtd_batchid) {
388 CDEBUG(D_HA, "%s update batchid from "LPU64 " to "LPU64"\n",
389 tdtd->tdtd_lut->lut_obd->obd_name,
390 tdtd->tdtd_batchid, record->ur_batchid);
391 tdtd->tdtd_batchid = record->ur_batchid + 1;
393 spin_unlock(&tdtd->tdtd_batchid_lock);
396 spin_lock(&tdtd->tdtd_replay_list_lock);
397 /* First try to build the replay update request with the records */
398 dtrq = dtrq_lookup(tdtd, record->ur_batchid);
400 spin_unlock(&tdtd->tdtd_replay_list_lock);
401 dtrq = dtrq_create(tdtd, lur);
403 RETURN(PTR_ERR(dtrq));
405 spin_lock(&tdtd->tdtd_replay_list_lock);
406 rc = dtrq_insert(tdtd, dtrq);
408 spin_unlock(&tdtd->tdtd_replay_list_lock);
415 /* If the master transno in update header is not
416 * matched with the one in the record, then it means
417 * the dtrq is originally created by master record,
418 * so we need update master transno and reposition
419 * the dtrq(by master transno) in the list and also
420 * replace update record */
421 if (record->ur_master_transno != 0 &&
422 dtrq->dtrq_master_transno != record->ur_master_transno &&
423 dtrq->dtrq_lur != NULL) {
424 list_del_init(&dtrq->dtrq_list);
425 dtrq->dtrq_lur->lur_update_rec.ur_master_transno =
426 record->ur_master_transno;
428 dtrq->dtrq_master_transno = record->ur_master_transno;
429 replace_record = true;
430 /* try to insert again */
431 rc = dtrq_insert(tdtd, dtrq);
433 spin_unlock(&tdtd->tdtd_replay_list_lock);
439 spin_unlock(&tdtd->tdtd_replay_list_lock);
441 /* Because there should be only thread access the update record, so
442 * we do not need lock here */
443 if (replace_record) {
444 /* Replace the update record and master transno */
445 OBD_FREE(dtrq->dtrq_lur, dtrq->dtrq_lur_size);
446 dtrq->dtrq_lur = NULL;
447 dtrq->dtrq_lur_size = llog_update_record_size(lur);
448 OBD_ALLOC_LARGE(dtrq->dtrq_lur, dtrq->dtrq_lur_size);
449 if (dtrq->dtrq_lur == NULL)
452 memcpy(dtrq->dtrq_lur, lur, dtrq->dtrq_lur_size);
455 /* This is a partial update records, let's try to append
456 * the record to the current replay request */
457 if (record->ur_flags & UPDATE_RECORD_CONTINUE)
458 rc = dtrq_append_updates(dtrq, record);
460 /* Then create and add sub update request */
461 rc = dtrq_sub_create_and_insert(dtrq, cookie, mdt_index);
465 EXPORT_SYMBOL(insert_update_records_to_replay_list);
468 * Dump updates of distribute txns.
470 * Output all of recovery updates in the distribute txn list to the
473 * \param[in] tdtd distribute txn data where all of distribute txn
475 * \param[in] mask debug mask
477 void dtrq_list_dump(struct target_distribute_txn_data *tdtd, unsigned int mask)
479 struct distribute_txn_replay_req *dtrq;
481 spin_lock(&tdtd->tdtd_replay_list_lock);
482 list_for_each_entry(dtrq, &tdtd->tdtd_replay_list, dtrq_list)
483 update_records_dump(&dtrq->dtrq_lur->lur_update_rec, mask,
485 spin_unlock(&tdtd->tdtd_replay_list_lock);
487 EXPORT_SYMBOL(dtrq_list_dump);
490 * Destroy distribute txn replay req
492 * Destroy distribute txn replay req and all of subs.
494 * \param[in] dtrq distribute txn replqy req to be destroyed.
496 void dtrq_destroy(struct distribute_txn_replay_req *dtrq)
498 struct distribute_txn_replay_req_sub *dtrqs;
499 struct distribute_txn_replay_req_sub *tmp;
501 LASSERT(list_empty(&dtrq->dtrq_list));
502 spin_lock(&dtrq->dtrq_sub_list_lock);
503 list_for_each_entry_safe(dtrqs, tmp, &dtrq->dtrq_sub_list, dtrqs_list) {
504 struct sub_thandle_cookie *stc;
505 struct sub_thandle_cookie *tmp;
507 list_del(&dtrqs->dtrqs_list);
508 list_for_each_entry_safe(stc, tmp, &dtrqs->dtrqs_cookie_list,
510 list_del(&stc->stc_list);
515 spin_unlock(&dtrq->dtrq_sub_list_lock);
517 if (dtrq->dtrq_lur != NULL)
518 OBD_FREE_LARGE(dtrq->dtrq_lur, dtrq->dtrq_lur_size);
522 EXPORT_SYMBOL(dtrq_destroy);
525 * Destroy all of replay req.
527 * Destroy all of replay req in the replay list.
529 * \param[in] tdtd target distribute txn data where the replay list is.
531 void dtrq_list_destroy(struct target_distribute_txn_data *tdtd)
533 struct distribute_txn_replay_req *dtrq;
534 struct distribute_txn_replay_req *tmp;
536 spin_lock(&tdtd->tdtd_replay_list_lock);
537 list_for_each_entry_safe(dtrq, tmp, &tdtd->tdtd_replay_list,
539 list_del_init(&dtrq->dtrq_list);
542 list_for_each_entry_safe(dtrq, tmp, &tdtd->tdtd_replay_finish_list,
544 list_del_init(&dtrq->dtrq_list);
547 spin_unlock(&tdtd->tdtd_replay_list_lock);
549 EXPORT_SYMBOL(dtrq_list_destroy);
552 * Get next req in the replay list
554 * Get next req needs to be replayed, since it is a sorted list
555 * (by master MDT transno)
557 * \param[in] tdtd distribute txn data where the replay list is
559 * \retval the pointer of update recovery header
561 struct distribute_txn_replay_req *
562 distribute_txn_get_next_req(struct target_distribute_txn_data *tdtd)
564 struct distribute_txn_replay_req *dtrq = NULL;
566 spin_lock(&tdtd->tdtd_replay_list_lock);
567 if (!list_empty(&tdtd->tdtd_replay_list)) {
568 dtrq = list_entry(tdtd->tdtd_replay_list.next,
569 struct distribute_txn_replay_req, dtrq_list);
570 list_del_init(&dtrq->dtrq_list);
572 spin_unlock(&tdtd->tdtd_replay_list_lock);
576 EXPORT_SYMBOL(distribute_txn_get_next_req);
579 * Get next transno in the replay list, because this is the sorted
580 * list, so it will return the transno of next req in the list.
582 * \param[in] tdtd distribute txn data where the replay list is
584 * \retval the transno of next update in the list
586 __u64 distribute_txn_get_next_transno(struct target_distribute_txn_data *tdtd)
588 struct distribute_txn_replay_req *dtrq = NULL;
591 spin_lock(&tdtd->tdtd_replay_list_lock);
592 if (!list_empty(&tdtd->tdtd_replay_list)) {
593 dtrq = list_entry(tdtd->tdtd_replay_list.next,
594 struct distribute_txn_replay_req, dtrq_list);
595 transno = dtrq->dtrq_master_transno;
597 spin_unlock(&tdtd->tdtd_replay_list_lock);
599 CDEBUG(D_HA, "%s: Next update transno "LPU64"\n",
600 tdtd->tdtd_lut->lut_obd->obd_name, transno);
603 EXPORT_SYMBOL(distribute_txn_get_next_transno);
605 struct distribute_txn_replay_req *
606 distribute_txn_lookup_finish_list(struct target_distribute_txn_data *tdtd,
609 struct distribute_txn_replay_req *dtrq = NULL;
610 struct distribute_txn_replay_req *iter;
612 spin_lock(&tdtd->tdtd_replay_list_lock);
613 list_for_each_entry(iter, &tdtd->tdtd_replay_finish_list, dtrq_list) {
614 if (iter->dtrq_xid == xid) {
619 spin_unlock(&tdtd->tdtd_replay_list_lock);
623 bool is_req_replayed_by_update(struct ptlrpc_request *req)
625 struct lu_target *tgt = class_exp2tgt(req->rq_export);
626 struct distribute_txn_replay_req *dtrq;
628 if (tgt->lut_tdtd == NULL)
631 dtrq = distribute_txn_lookup_finish_list(tgt->lut_tdtd, req->rq_xid);
637 EXPORT_SYMBOL(is_req_replayed_by_update);
640 * Check if the update of one object is committed
642 * Check whether the update for the object is committed by checking whether
643 * the correspondent sub exists in the replay req. If it is committed, mark
644 * the committed flag in correspondent the sub thandle.
646 * \param[in] env execution environment
647 * \param[in] dtrq replay request
648 * \param[in] dt_obj object for the update
649 * \param[in] top_th top thandle
650 * \param[in] sub_th sub thandle which the update belongs to
652 * \retval 1 if the update is not committed.
653 * \retval 0 if the update is committed.
654 * \retval negative errno if some other failures happen.
656 static int update_is_committed(const struct lu_env *env,
657 struct distribute_txn_replay_req *dtrq,
658 struct dt_object *dt_obj,
659 struct top_thandle *top_th,
660 struct sub_thandle *st)
662 struct seq_server_site *seq_site;
663 const struct lu_fid *fid = lu_object_fid(&dt_obj->do_lu);
664 struct distribute_txn_replay_req_sub *dtrqs;
668 if (st->st_sub_th != NULL)
671 if (st->st_committed)
674 seq_site = lu_site2seq(dt_obj->do_lu.lo_dev->ld_site);
675 if (fid_is_update_log(fid) || fid_is_update_log_dir(fid)) {
676 mdt_index = fid_oid(fid);
677 } else if (!fid_seq_in_fldb(fid_seq(fid))) {
678 mdt_index = seq_site->ss_node_id;
680 struct lu_server_fld *fld;
681 struct lu_seq_range range = {0};
684 fld = seq_site->ss_server_fld;
685 fld_range_set_type(&range, LU_SEQ_RANGE_MDT);
686 LASSERT(fld->lsf_seq_lookup != NULL);
687 rc = fld->lsf_seq_lookup(env, fld, fid_seq(fid),
691 mdt_index = range.lsr_index;
694 dtrqs = dtrq_sub_lookup(dtrq, mdt_index);
695 if (dtrqs != NULL || top_th->tt_multiple_thandle->tmt_committed) {
696 st->st_committed = 1;
698 struct sub_thandle_cookie *stc;
699 struct sub_thandle_cookie *tmp;
701 list_for_each_entry_safe(stc, tmp,
702 &dtrqs->dtrqs_cookie_list,
704 list_move(&stc->stc_list, &st->st_cookie_list);
709 CDEBUG(D_HA, "Update of "DFID "on MDT%u is not committed\n", PFID(fid),
716 * Implementation of different update methods for update recovery.
718 * These following functions update_recovery_$(update_name) implement
719 * different updates recovery methods. They will extract the parameters
720 * from the common parameters area and call correspondent dt API to redo
723 * \param[in] env execution environment
724 * \param[in] op update operation to be replayed
725 * \param[in] params common update parameters which holds all parameters
727 * \param[in] th transaction handle
728 * \param[in] declare indicate it will do declare or real execution, true
729 * means declare, false means real execution
731 * \retval 0 if it succeeds.
732 * \retval negative errno if it fails.
734 static int update_recovery_create(const struct lu_env *env,
735 struct dt_object *dt_obj,
736 const struct update_op *op,
737 const struct update_params *params,
738 struct thandle_exec_args *ta,
741 struct update_thread_info *uti = update_env_info(env);
742 struct llog_update_record *lur = uti->uti_dtrq->dtrq_lur;
743 struct lu_attr *attr = &uti->uti_attr;
745 struct obdo *lobdo = &uti->uti_obdo;
746 struct dt_object_format dof;
748 unsigned int param_count;
752 if (dt_object_exists(dt_obj))
755 param_count = lur->lur_update_rec.ur_param_count;
756 wobdo = update_params_get_param_buf(params, op->uop_params_off[0],
760 if (size != sizeof(*wobdo))
763 if (LLOG_REC_HDR_NEEDS_SWABBING(&lur->lur_hdr))
764 lustre_swab_obdo(wobdo);
766 lustre_get_wire_obdo(NULL, lobdo, wobdo);
767 la_from_obdo(attr, lobdo, lobdo->o_valid);
769 dof.dof_type = dt_mode_to_dft(attr->la_mode);
771 rc = out_tx_create(env, dt_obj, attr, NULL, &dof,
777 static int update_recovery_destroy(const struct lu_env *env,
778 struct dt_object *dt_obj,
779 const struct update_op *op,
780 const struct update_params *params,
781 struct thandle_exec_args *ta,
787 rc = out_tx_destroy(env, dt_obj, ta, th, NULL, 0);
792 static int update_recovery_ref_add(const struct lu_env *env,
793 struct dt_object *dt_obj,
794 const struct update_op *op,
795 const struct update_params *params,
796 struct thandle_exec_args *ta,
802 rc = out_tx_ref_add(env, dt_obj, ta, th, NULL, 0);
807 static int update_recovery_ref_del(const struct lu_env *env,
808 struct dt_object *dt_obj,
809 const struct update_op *op,
810 const struct update_params *params,
811 struct thandle_exec_args *ta,
817 rc = out_tx_ref_del(env, dt_obj, ta, th, NULL, 0);
822 static int update_recovery_attr_set(const struct lu_env *env,
823 struct dt_object *dt_obj,
824 const struct update_op *op,
825 const struct update_params *params,
826 struct thandle_exec_args *ta,
829 struct update_thread_info *uti = update_env_info(env);
830 struct llog_update_record *lur = uti->uti_dtrq->dtrq_lur;
832 struct obdo *lobdo = &uti->uti_obdo;
833 struct lu_attr *attr = &uti->uti_attr;
835 unsigned int param_count;
839 param_count = lur->lur_update_rec.ur_param_count;
840 wobdo = update_params_get_param_buf(params, op->uop_params_off[0],
844 if (size != sizeof(*wobdo))
847 if (LLOG_REC_HDR_NEEDS_SWABBING(&lur->lur_hdr))
848 lustre_swab_obdo(wobdo);
850 lustre_get_wire_obdo(NULL, lobdo, wobdo);
851 la_from_obdo(attr, lobdo, lobdo->o_valid);
853 rc = out_tx_attr_set(env, dt_obj, attr, ta, th, NULL, 0);
858 static int update_recovery_xattr_set(const struct lu_env *env,
859 struct dt_object *dt_obj,
860 const struct update_op *op,
861 const struct update_params *params,
862 struct thandle_exec_args *ta,
865 struct update_thread_info *uti = update_env_info(env);
874 param_count = uti->uti_dtrq->dtrq_lur->lur_update_rec.ur_param_count;
875 name = update_params_get_param_buf(params,
876 op->uop_params_off[0],
881 buf = update_params_get_param_buf(params,
882 op->uop_params_off[1],
887 uti->uti_buf.lb_buf = buf;
888 uti->uti_buf.lb_len = (size_t)size;
890 buf = update_params_get_param_buf(params, op->uop_params_off[2],
894 if (size != sizeof(fl))
897 fl = le32_to_cpu(*(int *)buf);
899 rc = out_tx_xattr_set(env, dt_obj, &uti->uti_buf, name, fl, ta, th,
905 static int update_recovery_index_insert(const struct lu_env *env,
906 struct dt_object *dt_obj,
907 const struct update_op *op,
908 const struct update_params *params,
909 struct thandle_exec_args *ta,
912 struct update_thread_info *uti = update_env_info(env);
922 param_count = uti->uti_dtrq->dtrq_lur->lur_update_rec.ur_param_count;
923 name = update_params_get_param_buf(params, op->uop_params_off[0],
928 fid = update_params_get_param_buf(params, op->uop_params_off[1],
932 if (size != sizeof(*fid))
935 fid_le_to_cpu(fid, fid);
937 ptype = update_params_get_param_buf(params, op->uop_params_off[2],
941 if (size != sizeof(*ptype))
943 type = le32_to_cpu(*ptype);
945 if (dt_try_as_dir(env, dt_obj) == 0)
948 uti->uti_rec.rec_fid = fid;
949 uti->uti_rec.rec_type = type;
951 rc = out_tx_index_insert(env, dt_obj,
952 (const struct dt_rec *)&uti->uti_rec,
953 (const struct dt_key *)name, ta, th,
959 static int update_recovery_index_delete(const struct lu_env *env,
960 struct dt_object *dt_obj,
961 const struct update_op *op,
962 const struct update_params *params,
963 struct thandle_exec_args *ta,
966 struct update_thread_info *uti = update_env_info(env);
973 param_count = uti->uti_dtrq->dtrq_lur->lur_update_rec.ur_param_count;
974 name = update_params_get_param_buf(params, op->uop_params_off[0],
979 if (dt_try_as_dir(env, dt_obj) == 0)
982 rc = out_tx_index_delete(env, dt_obj,
983 (const struct dt_key *)name, ta, th, NULL, 0);
988 static int update_recovery_write(const struct lu_env *env,
989 struct dt_object *dt_obj,
990 const struct update_op *op,
991 const struct update_params *params,
992 struct thandle_exec_args *ta,
995 struct update_thread_info *uti = update_env_info(env);
1003 param_count = uti->uti_dtrq->dtrq_lur->lur_update_rec.ur_param_count;
1004 buf = update_params_get_param_buf(params, op->uop_params_off[0],
1005 param_count, &size);
1009 uti->uti_buf.lb_buf = buf;
1010 uti->uti_buf.lb_len = size;
1012 buf = update_params_get_param_buf(params, op->uop_params_off[1],
1013 param_count, &size);
1017 pos = le64_to_cpu(*(__u64 *)buf);
1019 rc = out_tx_write(env, dt_obj, &uti->uti_buf, pos,
1025 static int update_recovery_xattr_del(const struct lu_env *env,
1026 struct dt_object *dt_obj,
1027 const struct update_op *op,
1028 const struct update_params *params,
1029 struct thandle_exec_args *ta,
1032 struct update_thread_info *uti = update_env_info(env);
1039 param_count = uti->uti_dtrq->dtrq_lur->lur_update_rec.ur_param_count;
1040 name = update_params_get_param_buf(params, op->uop_params_off[0],
1041 param_count, &size);
1045 rc = out_tx_xattr_del(env, dt_obj, name, ta, th, NULL, 0);
1051 * Update session information
1053 * Update session information so tgt_txn_stop_cb()->tgt_last_rcvd_update()
1054 * can be called correctly during update replay.
1056 * \param[in] env execution environment.
1057 * \param[in] tdtd distribute data structure of the recovering tgt.
1058 * \param[in] th thandle of this update replay.
1059 * \param[in] master_th master sub thandle.
1060 * \param[in] ta_arg the tx arg structure to hold the update for updating
1063 static void update_recovery_update_ses(struct lu_env *env,
1064 struct target_distribute_txn_data *tdtd,
1066 struct thandle *master_th,
1067 struct distribute_txn_replay_req *dtrq,
1068 struct tx_arg *ta_arg)
1070 struct tgt_session_info *tsi;
1071 struct lu_target *lut = tdtd->tdtd_lut;
1072 struct obd_export *export;
1073 struct cfs_hash *hash;
1074 struct top_thandle *top_th;
1075 struct lsd_reply_data *lrd;
1078 tsi = tgt_ses_info(env);
1079 if (tsi->tsi_exp != NULL)
1082 size = ta_arg->u.write.buf.lb_len;
1083 lrd = ta_arg->u.write.buf.lb_buf;
1084 if (size != sizeof(*lrd) || lrd == NULL)
1087 lrd->lrd_transno = le64_to_cpu(lrd->lrd_transno);
1088 lrd->lrd_xid = le64_to_cpu(lrd->lrd_xid);
1089 lrd->lrd_data = le64_to_cpu(lrd->lrd_data);
1090 lrd->lrd_result = le32_to_cpu(lrd->lrd_result);
1091 lrd->lrd_client_gen = le32_to_cpu(lrd->lrd_client_gen);
1093 if (lrd->lrd_transno != tgt_th_info(env)->tti_transno)
1096 hash = cfs_hash_getref(lut->lut_obd->obd_gen_hash);
1100 export = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
1101 if (export == NULL) {
1102 cfs_hash_putref(hash);
1106 tsi->tsi_exp = export;
1107 tsi->tsi_xid = lrd->lrd_xid;
1108 tsi->tsi_opdata = lrd->lrd_data;
1109 tsi->tsi_result = lrd->lrd_result;
1110 tsi->tsi_client_gen = lrd->lrd_client_gen;
1111 dtrq->dtrq_xid = lrd->lrd_xid;
1112 top_th = container_of(th, struct top_thandle, tt_super);
1113 top_th->tt_master_sub_thandle = master_th;
1114 cfs_hash_putref(hash);
1118 * Execute updates in the update replay records
1120 * Declare distribute txn replay by update records and add the updates
1121 * to the execution list. Note: it will check if the update has been
1122 * committed, and only execute the updates if it is not committed to
1125 * \param[in] env execution environment
1126 * \param[in] tdtd distribute txn replay data which hold all of replay
1127 * reqs and all replay parameters.
1128 * \param[in] dtrq distribute transaction replay req.
1129 * \param[in] ta thandle execute args.
1131 * \retval 0 if declare succeeds.
1132 * \retval negative errno if declare fails.
1134 static int update_recovery_exec(const struct lu_env *env,
1135 struct target_distribute_txn_data *tdtd,
1136 struct distribute_txn_replay_req *dtrq,
1137 struct thandle_exec_args *ta)
1139 struct llog_update_record *lur = dtrq->dtrq_lur;
1140 struct update_records *records = &lur->lur_update_rec;
1141 struct update_ops *ops = &records->ur_ops;
1142 struct update_params *params = update_records_get_params(records);
1143 struct top_thandle *top_th = container_of(ta->ta_handle,
1146 struct top_multiple_thandle *tmt = top_th->tt_multiple_thandle;
1147 struct update_op *op;
1152 /* These records have been swabbed in llog_cat_process() */
1153 for (i = 0, op = &ops->uops_op[0]; i < records->ur_update_count;
1154 i++, op = update_op_next_op(op)) {
1155 struct lu_fid *fid = &op->uop_fid;
1156 struct dt_object *dt_obj;
1157 struct dt_object *sub_dt_obj;
1158 struct dt_device *sub_dt;
1159 struct sub_thandle *st;
1161 if (op->uop_type == OUT_NOOP)
1164 dt_obj = dt_locate(env, tdtd->tdtd_dt, fid);
1165 if (IS_ERR(dt_obj)) {
1166 rc = PTR_ERR(dt_obj);
1169 sub_dt_obj = dt_object_child(dt_obj);
1171 /* Create sub thandle if not */
1172 sub_dt = lu2dt_dev(sub_dt_obj->do_lu.lo_dev);
1173 st = lookup_sub_thandle(tmt, sub_dt);
1175 st = create_sub_thandle(tmt, sub_dt);
1177 GOTO(next, rc = PTR_ERR(st));
1180 /* check if updates on the OSD/OSP are committed */
1181 rc = update_is_committed(env, dtrq, dt_obj, top_th, st);
1183 /* If this is committed, goto next */
1189 /* Create thandle for sub thandle if needed */
1190 if (st->st_sub_th == NULL) {
1191 rc = sub_thandle_trans_create(env, top_th, st);
1196 CDEBUG(D_HA, "replay %uth update\n", i);
1197 switch (op->uop_type) {
1199 rc = update_recovery_create(env, sub_dt_obj,
1204 rc = update_recovery_destroy(env, sub_dt_obj,
1209 rc = update_recovery_ref_add(env, sub_dt_obj,
1214 rc = update_recovery_ref_del(env, sub_dt_obj,
1219 rc = update_recovery_attr_set(env, sub_dt_obj,
1224 rc = update_recovery_xattr_set(env, sub_dt_obj,
1228 case OUT_INDEX_INSERT:
1229 rc = update_recovery_index_insert(env, sub_dt_obj,
1233 case OUT_INDEX_DELETE:
1234 rc = update_recovery_index_delete(env, sub_dt_obj,
1239 rc = update_recovery_write(env, sub_dt_obj,
1244 rc = update_recovery_xattr_del(env, sub_dt_obj,
1249 CERROR("Unknown update type %u\n", (__u32)op->uop_type);
1254 lu_object_put(env, &dt_obj->do_lu);
1259 ta->ta_handle->th_result = rc;
1264 * redo updates on MDT if needed.
1266 * During DNE recovery, the recovery thread (target_recovery_thread) will call
1267 * this function to replay distribute txn updates on all MDTs. It only replay
1268 * updates on the MDT where the update record is missing.
1270 * If the update already exists on the MDT, then it does not need replay the
1271 * updates on that MDT, and only mark the sub transaction has been committed
1274 * \param[in] env execution environment
1275 * \param[in] tdtd target distribute txn data, which holds the replay list
1276 * and all parameters needed by replay process.
1277 * \param[in] dtrq distribute txn replay req.
1279 * \retval 0 if replay succeeds.
1280 * \retval negative errno if replay failes.
1282 int distribute_txn_replay_handle(struct lu_env *env,
1283 struct target_distribute_txn_data *tdtd,
1284 struct distribute_txn_replay_req *dtrq)
1286 struct update_records *records = &dtrq->dtrq_lur->lur_update_rec;
1287 struct thandle_exec_args *ta;
1288 struct lu_context session_env;
1289 struct thandle *th = NULL;
1290 struct top_thandle *top_th;
1291 struct top_multiple_thandle *tmt;
1292 struct thandle_update_records *tur = NULL;
1297 /* initialize session, it is needed for the handler of target */
1298 rc = lu_context_init(&session_env, LCT_SERVER_SESSION | LCT_NOREF);
1300 CERROR("%s: failure to initialize session: rc = %d\n",
1301 tdtd->tdtd_lut->lut_obd->obd_name, rc);
1304 lu_context_enter(&session_env);
1305 env->le_ses = &session_env;
1307 update_records_dump(records, D_HA, true);
1308 th = top_trans_create(env, NULL);
1310 GOTO(exit_session, rc = PTR_ERR(th));
1312 ta = &update_env_info(env)->uti_tea;
1315 update_env_info(env)->uti_dtrq = dtrq;
1316 /* Create distribute transaction structure for this top thandle */
1317 top_th = container_of(th, struct top_thandle, tt_super);
1318 rc = top_trans_create_tmt(env, top_th);
1320 GOTO(stop_trans, rc);
1322 th->th_dev = tdtd->tdtd_dt;
1325 /* check if the distribute transaction has been committed */
1326 tmt = top_th->tt_multiple_thandle;
1327 tmt->tmt_master_sub_dt = tdtd->tdtd_lut->lut_bottom;
1328 tmt->tmt_batchid = dtrq->dtrq_batchid;
1329 tgt_th_info(env)->tti_transno = dtrq->dtrq_master_transno;
1331 if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid)
1332 tmt->tmt_committed = 1;
1334 rc = update_recovery_exec(env, tdtd, dtrq, ta);
1336 GOTO(stop_trans, rc);
1338 /* If no updates are needed to be replayed, then mark this records as
1339 * committed, so commit thread distribute_txn_commit_thread() will
1340 * delete the record */
1341 if (ta->ta_argno == 0)
1342 tmt->tmt_committed = 1;
1344 tur = &update_env_info(env)->uti_tur;
1345 tur->tur_update_records = dtrq->dtrq_lur;
1346 tur->tur_update_records_buf_size = dtrq->dtrq_lur_size;
1347 tur->tur_update_params = NULL;
1348 tur->tur_update_param_count = 0;
1349 tmt->tmt_update_records = tur;
1351 distribute_txn_insert_by_batchid(tmt);
1352 rc = top_trans_start(env, NULL, th);
1354 GOTO(stop_trans, rc);
1356 for (i = 0; i < ta->ta_argno; i++) {
1357 struct tx_arg *ta_arg;
1358 struct dt_object *dt_obj;
1359 struct dt_device *sub_dt;
1360 struct sub_thandle *st;
1362 ta_arg = ta->ta_args[i];
1363 dt_obj = ta_arg->object;
1365 LASSERT(tmt->tmt_committed == 0);
1366 sub_dt = lu2dt_dev(dt_obj->do_lu.lo_dev);
1367 st = lookup_sub_thandle(tmt, sub_dt);
1369 LASSERT(st != NULL);
1370 LASSERT(st->st_sub_th != NULL);
1371 rc = ta->ta_args[i]->exec_fn(env, st->st_sub_th,
1374 /* If the update is to update the reply data, then
1375 * we need set the session information, so
1376 * tgt_last_rcvd_update() can be called correctly */
1377 if (rc == 0 && dt_obj == tdtd->tdtd_lut->lut_reply_data)
1378 update_recovery_update_ses(env, tdtd, th,
1379 st->st_sub_th, dtrq, ta_arg);
1381 if (unlikely(rc < 0)) {
1382 CDEBUG(D_HA, "error during execution of #%u from"
1383 " %s:%d: rc = %d\n", i, ta->ta_args[i]->file,
1384 ta->ta_args[i]->line, rc);
1386 if (ta->ta_args[i]->undo_fn != NULL) {
1387 dt_obj = ta->ta_args[i]->object;
1389 lu2dt_dev(dt_obj->do_lu.lo_dev);
1390 st = lookup_sub_thandle(tmt, sub_dt);
1391 LASSERT(st != NULL);
1392 LASSERT(st->st_sub_th != NULL);
1394 ta->ta_args[i]->undo_fn(env,
1398 CERROR("%s: undo for %s:%d: rc = %d\n",
1399 dt_obd_name(ta->ta_handle->th_dev),
1400 ta->ta_args[i]->file,
1401 ta->ta_args[i]->line, -ENOTSUPP);
1406 CDEBUG(D_HA, "%s: executed %u/%u: rc = %d\n",
1407 dt_obd_name(sub_dt), i, ta->ta_argno, rc);
1413 rc = top_trans_stop(env, tdtd->tdtd_dt, th);
1414 for (i = 0; i < ta->ta_argno; i++) {
1415 if (ta->ta_args[i]->object != NULL) {
1416 lu_object_put(env, &ta->ta_args[i]->object->do_lu);
1417 ta->ta_args[i]->object = NULL;
1422 tur->tur_update_records = NULL;
1424 if (tgt_ses_info(env)->tsi_exp != NULL) {
1425 class_export_put(tgt_ses_info(env)->tsi_exp);
1426 tgt_ses_info(env)->tsi_exp = NULL;
1429 lu_context_exit(&session_env);
1430 lu_context_fini(&session_env);
1433 EXPORT_SYMBOL(distribute_txn_replay_handle);