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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Lustre Unified Target
37 * These are common function to work with last_received file
39 * Author: Mikhail Pershin <mike.pershin@intel.com>
42 #include <obd_class.h>
43 #include <lustre_fid.h>
45 #include "tgt_internal.h"
47 static inline struct lu_buf *tti_buf_lsd(struct tgt_thread_info *tti)
49 tti->tti_buf.lb_buf = &tti->tti_lsd;
50 tti->tti_buf.lb_len = sizeof(tti->tti_lsd);
54 static inline struct lu_buf *tti_buf_lcd(struct tgt_thread_info *tti)
56 tti->tti_buf.lb_buf = &tti->tti_lcd;
57 tti->tti_buf.lb_len = sizeof(tti->tti_lcd);
62 * Allocate in-memory data for client slot related to export.
64 int tgt_client_alloc(struct obd_export *exp)
67 LASSERT(exp != exp->exp_obd->obd_self_export);
69 OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
70 if (exp->exp_target_data.ted_lcd == NULL)
72 /* Mark that slot is not yet valid, 0 doesn't work here */
73 exp->exp_target_data.ted_lr_idx = -1;
76 EXPORT_SYMBOL(tgt_client_alloc);
79 * Free in-memory data for client slot related to export.
81 void tgt_client_free(struct obd_export *exp)
83 struct tg_export_data *ted = &exp->exp_target_data;
84 struct lu_target *lut = class_exp2tgt(exp);
86 LASSERT(exp != exp->exp_obd->obd_self_export);
88 OBD_FREE_PTR(ted->ted_lcd);
91 /* Slot may be not yet assigned */
92 if (ted->ted_lr_idx < 0)
94 /* Clear bit when lcd is freed */
95 LASSERT(lut->lut_client_bitmap);
96 if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
97 CERROR("%s: client %u bit already clear in bitmap\n",
98 exp->exp_obd->obd_name, ted->ted_lr_idx);
102 EXPORT_SYMBOL(tgt_client_free);
104 int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
105 struct lsd_client_data *lcd, loff_t *off, int index)
107 struct tgt_thread_info *tti = tgt_th_info(env);
111 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
113 check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
114 lcd_le_to_cpu(&tti->tti_lcd, lcd);
115 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
116 lcd->lcd_last_close_result =
117 ptlrpc_status_ntoh(lcd->lcd_last_close_result);
120 CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = "LPU64
121 ", last_xid = "LPU64", last_result = %u, last_data = %u, "
122 "last_close_transno = "LPU64", last_close_xid = "LPU64", "
123 "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
124 *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
125 lcd->lcd_last_result, lcd->lcd_last_data,
126 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
127 lcd->lcd_last_close_result, rc);
130 EXPORT_SYMBOL(tgt_client_data_read);
132 int tgt_client_data_write(const struct lu_env *env, struct lu_target *tgt,
133 struct lsd_client_data *lcd, loff_t *off,
136 struct tgt_thread_info *tti = tgt_th_info(env);
138 lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
139 lcd->lcd_last_close_result =
140 ptlrpc_status_hton(lcd->lcd_last_close_result);
141 lcd_cpu_to_le(lcd, &tti->tti_lcd);
144 return dt_record_write(env, tgt->lut_last_rcvd, &tti->tti_buf, off, th);
146 EXPORT_SYMBOL(tgt_client_data_write);
149 * Update client data in last_rcvd
151 int tgt_client_data_update(const struct lu_env *env, struct obd_export *exp)
153 struct tg_export_data *ted = &exp->exp_target_data;
154 struct lu_target *tgt = class_exp2tgt(exp);
155 struct tgt_thread_info *tti = tgt_th_info(env);
161 th = dt_trans_create(env, tgt->lut_bottom);
166 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
168 ted->ted_lr_off, th);
172 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
176 * Until this operations will be committed the sync is needed
177 * for this export. This should be done _after_ starting the
178 * transaction so that many connecting clients will not bring
179 * server down with lots of sync writes.
181 rc = tgt_new_client_cb_add(th, exp);
183 /* can't add callback, do sync now */
186 spin_lock(&exp->exp_lock);
187 exp->exp_need_sync = 1;
188 spin_unlock(&exp->exp_lock);
191 tti->tti_off = ted->ted_lr_off;
192 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
195 dt_trans_stop(env, tgt->lut_bottom, th);
196 CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
197 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
198 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
203 int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
205 struct tgt_thread_info *tti = tgt_th_info(env);
210 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
213 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
215 CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
216 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
217 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
220 EXPORT_SYMBOL(tgt_server_data_read);
222 int tgt_server_data_write(const struct lu_env *env, struct lu_target *tgt,
225 struct tgt_thread_info *tti = tgt_th_info(env);
232 lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
234 rc = dt_record_write(env, tgt->lut_last_rcvd, &tti->tti_buf,
237 CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
238 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
239 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
243 EXPORT_SYMBOL(tgt_server_data_write);
246 * Update server data in last_rcvd
248 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
251 struct tgt_thread_info *tti = tgt_th_info(env);
258 "%s: mount_count is "LPU64", last_transno is "LPU64"\n",
259 tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
260 tgt->lut_last_transno);
262 /* Always save latest transno to keep it fresh */
263 spin_lock(&tgt->lut_translock);
264 tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
265 spin_unlock(&tgt->lut_translock);
267 th = dt_trans_create(env, tgt->lut_bottom);
274 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
275 &tti->tti_buf, tti->tti_off, th);
279 rc = dt_trans_start(env, tgt->lut_bottom, th);
283 rc = tgt_server_data_write(env, tgt, th);
285 dt_trans_stop(env, tgt->lut_bottom, th);
287 CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
288 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
289 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
292 EXPORT_SYMBOL(tgt_server_data_update);
294 int tgt_truncate_last_rcvd(const struct lu_env *env, struct lu_target *tgt,
297 struct dt_object *dt = tgt->lut_last_rcvd;
305 attr.la_valid = LA_SIZE;
307 th = dt_trans_create(env, tgt->lut_bottom);
310 rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
313 rc = dt_declare_attr_set(env, dt, &attr, th);
316 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
320 rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th, BYPASS_CAPA);
322 rc = dt_attr_set(env, dt, &attr, th, BYPASS_CAPA);
325 dt_trans_stop(env, tgt->lut_bottom, th);
329 EXPORT_SYMBOL(tgt_truncate_last_rcvd);
331 void tgt_client_epoch_update(const struct lu_env *env, struct obd_export *exp)
333 struct lsd_client_data *lcd = exp->exp_target_data.ted_lcd;
334 struct lu_target *tgt = class_exp2tgt(exp);
336 LASSERT(tgt->lut_bottom);
337 /** VBR: set client last_epoch to current epoch */
338 if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
340 lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
341 tgt_client_data_update(env, exp);
345 * Update boot epoch when recovery ends
347 void tgt_boot_epoch_update(struct lu_target *tgt)
350 struct ptlrpc_request *req;
352 cfs_list_t client_list;
355 if (tgt->lut_obd->obd_stopping)
358 rc = lu_env_init(&env, LCT_DT_THREAD);
360 CERROR("%s: can't initialize environment: rc = %d\n",
361 tgt->lut_obd->obd_name, rc);
365 spin_lock(&tgt->lut_translock);
366 start_epoch = lr_epoch(tgt->lut_last_transno) + 1;
367 tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
368 tgt->lut_lsd.lsd_start_epoch = start_epoch;
369 spin_unlock(&tgt->lut_translock);
371 CFS_INIT_LIST_HEAD(&client_list);
373 * The recovery is not yet finished and final queue can still be updated
374 * with resend requests. Move final list to separate one for processing
376 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
377 cfs_list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
378 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
381 * go through list of exports participated in recovery and
382 * set new epoch for them
384 cfs_list_for_each_entry(req, &client_list, rq_list) {
385 LASSERT(!req->rq_export->exp_delayed);
386 if (!req->rq_export->exp_vbr_failed)
387 tgt_client_epoch_update(&env, req->rq_export);
389 /** return list back at once */
390 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
391 cfs_list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
392 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
393 /** update server epoch */
394 tgt_server_data_update(&env, tgt, 1);
397 EXPORT_SYMBOL(tgt_boot_epoch_update);
400 * commit callback, need to update last_commited value
402 struct tgt_last_committed_callback {
403 struct dt_txn_commit_cb llcc_cb;
404 struct lu_target *llcc_tgt;
405 struct obd_export *llcc_exp;
409 void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
410 struct dt_txn_commit_cb *cb, int err)
412 struct tgt_last_committed_callback *ccb;
414 ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
416 LASSERT(ccb->llcc_tgt != NULL);
417 LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
419 spin_lock(&ccb->llcc_tgt->lut_translock);
420 if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
421 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
423 LASSERT(ccb->llcc_exp);
424 if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
425 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
426 spin_unlock(&ccb->llcc_tgt->lut_translock);
427 ptlrpc_commit_replies(ccb->llcc_exp);
429 spin_unlock(&ccb->llcc_tgt->lut_translock);
431 class_export_cb_put(ccb->llcc_exp);
432 if (ccb->llcc_transno)
433 CDEBUG(D_HA, "%s: transno "LPD64" is committed\n",
434 ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
438 int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
439 struct obd_export *exp, __u64 transno)
441 struct tgt_last_committed_callback *ccb;
442 struct dt_txn_commit_cb *dcb;
450 ccb->llcc_exp = class_export_cb_get(exp);
451 ccb->llcc_transno = transno;
454 dcb->dcb_func = tgt_cb_last_committed;
455 CFS_INIT_LIST_HEAD(&dcb->dcb_linkage);
456 strncpy(dcb->dcb_name, "tgt_cb_last_committed", MAX_COMMIT_CB_STR_LEN);
457 dcb->dcb_name[MAX_COMMIT_CB_STR_LEN - 1] = '\0';
459 rc = dt_trans_cb_add(th, dcb);
461 class_export_cb_put(exp);
465 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
466 /* report failure to force synchronous operation */
471 EXPORT_SYMBOL(tgt_last_commit_cb_add);
473 struct tgt_new_client_callback {
474 struct dt_txn_commit_cb lncc_cb;
475 struct obd_export *lncc_exp;
478 void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
479 struct dt_txn_commit_cb *cb, int err)
481 struct tgt_new_client_callback *ccb;
483 ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
485 LASSERT(ccb->lncc_exp->exp_obd);
487 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
488 ccb->lncc_exp->exp_obd->obd_name,
489 ccb->lncc_exp->exp_client_uuid.uuid);
491 spin_lock(&ccb->lncc_exp->exp_lock);
492 /* XXX: Currently, we use per-export based sync/async policy for
493 * the update via OUT RPC, it is coarse-grained policy, and
494 * will be changed as per-request based by DNE II patches. */
495 if (!ccb->lncc_exp->exp_keep_sync)
496 ccb->lncc_exp->exp_need_sync = 0;
498 spin_unlock(&ccb->lncc_exp->exp_lock);
499 class_export_cb_put(ccb->lncc_exp);
504 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
506 struct tgt_new_client_callback *ccb;
507 struct dt_txn_commit_cb *dcb;
514 ccb->lncc_exp = class_export_cb_get(exp);
517 dcb->dcb_func = tgt_cb_new_client;
518 CFS_INIT_LIST_HEAD(&dcb->dcb_linkage);
519 strncpy(dcb->dcb_name, "tgt_cb_new_client", MAX_COMMIT_CB_STR_LEN);
520 dcb->dcb_name[MAX_COMMIT_CB_STR_LEN - 1] = '\0';
522 rc = dt_trans_cb_add(th, dcb);
524 class_export_cb_put(exp);
531 * Add new client to the last_rcvd upon new connection.
533 * We use a bitmap to locate a free space in the last_rcvd file and initialize
536 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
538 struct tg_export_data *ted = &exp->exp_target_data;
539 struct lu_target *tgt = class_exp2tgt(exp);
544 LASSERT(tgt->lut_client_bitmap != NULL);
545 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
548 mutex_init(&ted->ted_lcd_lock);
550 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
553 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
554 * there's no need for extra complication here
556 idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
558 if (idx >= LR_MAX_CLIENTS ||
559 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
560 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
561 tgt->lut_obd->obd_name, idx);
564 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
565 idx = find_next_zero_bit(tgt->lut_client_bitmap,
566 LR_MAX_CLIENTS, idx);
570 CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added\n",
571 tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid);
573 ted->ted_lr_idx = idx;
574 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
575 idx * tgt->lut_lsd.lsd_client_size;
577 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
579 CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s'\n",
580 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
581 ted->ted_lcd->lcd_uuid);
583 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
586 rc = tgt_client_data_update(env, exp);
588 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
589 tgt->lut_obd->obd_name, idx, rc);
593 EXPORT_SYMBOL(tgt_client_new);
595 /* Add client data to the MDS. We use a bitmap to locate a free space
596 * in the last_rcvd file if cl_off is -1 (i.e. a new client).
597 * Otherwise, we just have to read the data from the last_rcvd file and
598 * we know its offset.
600 * It should not be possible to fail adding an existing client - otherwise
601 * mdt_init_server_data() callsite needs to be fixed.
603 int tgt_client_add(const struct lu_env *env, struct obd_export *exp, int idx)
605 struct tg_export_data *ted = &exp->exp_target_data;
606 struct lu_target *tgt = class_exp2tgt(exp);
610 LASSERT(tgt->lut_client_bitmap != NULL);
611 LASSERTF(idx >= 0, "%d\n", idx);
613 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
614 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
617 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
618 CERROR("%s: client %d: bit already set in bitmap!!\n",
619 tgt->lut_obd->obd_name, idx);
623 CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added\n",
624 tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid);
626 ted->ted_lr_idx = idx;
627 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
628 idx * tgt->lut_lsd.lsd_client_size;
630 mutex_init(&ted->ted_lcd_lock);
632 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
636 EXPORT_SYMBOL(tgt_client_add);
638 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
640 struct tg_export_data *ted = &exp->exp_target_data;
641 struct lu_target *tgt = class_exp2tgt(exp);
646 LASSERT(ted->ted_lcd);
648 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
649 if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
650 (char *)tgt->lut_obd->obd_uuid.uuid) ||
651 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
654 CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
655 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
656 ted->ted_lcd->lcd_uuid);
658 /* Clear the bit _after_ zeroing out the client so we don't
659 race with filter_client_add and zero out new clients.*/
660 if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
661 CERROR("%s: client %u: bit already clear in bitmap!!\n",
662 tgt->lut_obd->obd_name, ted->ted_lr_idx);
666 /* Do not erase record for recoverable client. */
667 if (exp->exp_flags & OBD_OPT_FAILOVER)
670 /* Make sure the server's last_transno is up to date.
671 * This should be done before zeroing client slot so last_transno will
672 * be in server data or in client data in case of failure */
673 rc = tgt_server_data_update(env, tgt, 0);
675 CERROR("%s: failed to update server data, skip client %s "
676 "zeroing, rc %d\n", tgt->lut_obd->obd_name,
677 ted->ted_lcd->lcd_uuid, rc);
681 mutex_lock(&ted->ted_lcd_lock);
682 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
683 rc = tgt_client_data_update(env, exp);
684 mutex_unlock(&ted->ted_lcd_lock);
686 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
687 "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
688 tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
689 ted->ted_lr_idx, ted->ted_lr_off, rc);
692 EXPORT_SYMBOL(tgt_client_del);
695 * last_rcvd & last_committed update callbacks
697 int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
698 struct dt_object *obj, __u64 opdata,
699 struct thandle *th, struct ptlrpc_request *req)
701 struct tgt_thread_info *tti = tgt_th_info(env);
702 struct tg_export_data *ted;
705 bool lw_client, update = false;
709 ted = &req->rq_export->exp_target_data;
711 lw_client = exp_connect_flags(req->rq_export) & OBD_CONNECT_LIGHTWEIGHT;
712 if (ted->ted_lr_idx < 0 && !lw_client)
713 /* ofd connect may cause transaction before export has
717 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
719 spin_lock(&tgt->lut_translock);
720 if (th->th_result != 0) {
721 if (tti->tti_transno != 0) {
722 CERROR("%s: replay transno "LPU64" failed: rc = %d\n",
723 tgt_name(tgt), tti->tti_transno, th->th_result);
725 } else if (tti->tti_transno == 0) {
726 tti->tti_transno = ++tgt->lut_last_transno;
728 /* should be replay */
729 if (tti->tti_transno > tgt->lut_last_transno)
730 tgt->lut_last_transno = tti->tti_transno;
732 spin_unlock(&tgt->lut_translock);
734 /** VBR: set new versions */
735 if (th->th_result == 0 && obj != NULL)
736 dt_version_set(env, obj, tti->tti_transno, th);
738 /* filling reply data */
739 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
740 tti->tti_transno, tgt->lut_obd->obd_last_committed);
742 req->rq_transno = tti->tti_transno;
743 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
745 /* if can't add callback, do sync write */
746 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, req->rq_export,
750 /* All operations performed by LW clients are synchronous and
751 * we store the committed transno in the last_rcvd header */
752 spin_lock(&tgt->lut_translock);
753 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
754 tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
757 spin_unlock(&tgt->lut_translock);
758 /* Although lightweight (LW) connections have no slot in
759 * last_rcvd, we still want to maintain the in-memory
760 * lsd_client_data structure in order to properly handle reply
762 } else if (ted->ted_lr_off == 0) {
763 CERROR("%s: client idx %d has offset %lld\n",
764 tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
768 /* if the export has already been disconnected, we have no last_rcvd
769 * slot, update server data with latest transno then */
770 if (ted->ted_lcd == NULL) {
771 CWARN("commit transaction for disconnected client %s: rc %d\n",
772 req->rq_export->exp_client_uuid.uuid, rc);
773 GOTO(srv_update, rc = 0);
776 mutex_lock(&ted->ted_lcd_lock);
777 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
778 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
779 lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
780 transno_p = &ted->ted_lcd->lcd_last_close_transno;
781 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
782 ted->ted_lcd->lcd_last_close_result = th->th_result;
784 /* VBR: save versions in last_rcvd for reconstruct. */
785 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
788 ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
789 ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
790 ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
791 ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
793 transno_p = &ted->ted_lcd->lcd_last_transno;
794 ted->ted_lcd->lcd_last_xid = req->rq_xid;
795 ted->ted_lcd->lcd_last_result = th->th_result;
796 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
797 * see struct ldlm_reply->lock_policy_res1; */
798 ted->ted_lcd->lcd_last_data = opdata;
801 /* Update transno in slot only if non-zero number, i.e. no errors */
802 if (likely(tti->tti_transno != 0)) {
803 if (*transno_p > tti->tti_transno &&
804 !tgt->lut_no_reconstruct) {
805 CERROR("%s: trying to overwrite bigger transno:"
806 "on-disk: "LPU64", new: "LPU64" replay: %d. "
807 "see LU-617.\n", tgt_name(tgt), *transno_p,
808 tti->tti_transno, req_is_replay(req));
809 if (req_is_replay(req)) {
810 spin_lock(&req->rq_export->exp_lock);
811 req->rq_export->exp_vbr_failed = 1;
812 spin_unlock(&req->rq_export->exp_lock);
814 mutex_unlock(&ted->ted_lcd_lock);
815 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
817 *transno_p = tti->tti_transno;
821 tti->tti_off = ted->ted_lr_off;
822 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
824 mutex_unlock(&ted->ted_lcd_lock);
828 mutex_unlock(&ted->ted_lcd_lock);
832 rc = tgt_server_data_write(env, tgt, th);
837 * last_rcvd update for echo client simulation.
838 * It updates last_rcvd client slot and version of object in
839 * simple way but with all locks to simulate all drawbacks
841 int tgt_last_rcvd_update_echo(const struct lu_env *env, struct lu_target *tgt,
842 struct dt_object *obj, struct thandle *th,
843 struct obd_export *exp)
845 struct tgt_thread_info *tti = tgt_th_info(env);
846 struct tg_export_data *ted = &exp->exp_target_data;
851 tti->tti_transno = 0;
853 spin_lock(&tgt->lut_translock);
854 if (th->th_result == 0)
855 tti->tti_transno = ++tgt->lut_last_transno;
856 spin_unlock(&tgt->lut_translock);
858 /** VBR: set new versions */
859 if (th->th_result == 0 && obj != NULL)
860 dt_version_set(env, obj, tti->tti_transno, th);
862 /* if can't add callback, do sync write */
863 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
866 LASSERT(ted->ted_lr_off > 0);
868 mutex_lock(&ted->ted_lcd_lock);
869 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
870 ted->ted_lcd->lcd_last_transno = tti->tti_transno;
871 ted->ted_lcd->lcd_last_result = th->th_result;
873 tti->tti_off = ted->ted_lr_off;
874 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
875 mutex_unlock(&ted->ted_lcd_lock);
879 int tgt_clients_data_init(const struct lu_env *env, struct lu_target *tgt,
880 unsigned long last_size)
882 struct obd_device *obd = tgt->lut_obd;
883 struct lr_server_data *lsd = &tgt->lut_lsd;
884 struct lsd_client_data *lcd = NULL;
885 struct tg_export_data *ted;
888 loff_t off = lsd->lsd_client_start;
892 CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
893 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
899 for (cl_idx = 0; off < last_size; cl_idx++) {
900 struct obd_export *exp;
903 /* Don't assume off is incremented properly by
904 * read_record(), in case sizeof(*lcd)
905 * isn't the same as fsd->lsd_client_size. */
906 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
907 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
909 CERROR("%s: error reading last_rcvd %s idx %d off "
910 "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
913 break; /* read error shouldn't cause startup to fail */
916 if (lcd->lcd_uuid[0] == '\0') {
917 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
922 last_transno = lcd_last_transno(lcd);
924 /* These exports are cleaned up by disconnect, so they
925 * need to be set up like real exports as connect does.
927 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
928 " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
929 last_transno, lsd->lsd_last_transno, lcd_last_xid(lcd));
931 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
933 if (PTR_ERR(exp) == -EALREADY) {
934 /* export already exists, zero out this one */
935 CERROR("%s: Duplicate export %s!\n",
936 tgt_name(tgt), lcd->lcd_uuid);
939 GOTO(err_out, rc = PTR_ERR(exp));
942 ted = &exp->exp_target_data;
943 *ted->ted_lcd = *lcd;
945 rc = tgt_client_add(env, exp, cl_idx);
946 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
947 /* VBR: set export last committed version */
948 exp->exp_last_committed = last_transno;
949 spin_lock(&exp->exp_lock);
950 exp->exp_connecting = 0;
951 exp->exp_in_recovery = 0;
952 spin_unlock(&exp->exp_lock);
953 obd->obd_max_recoverable_clients++;
954 class_export_put(exp);
956 /* Need to check last_rcvd even for duplicated exports. */
957 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
958 cl_idx, last_transno);
960 spin_lock(&tgt->lut_translock);
961 tgt->lut_last_transno = max(last_transno,
962 tgt->lut_last_transno);
963 spin_unlock(&tgt->lut_translock);
971 struct server_compat_data {
978 static struct server_compat_data tgt_scd[] = {
979 [LDD_F_SV_TYPE_MDT] = {
980 .rocompat = OBD_ROCOMPAT_LOVOBJID,
981 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
982 OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
983 OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI,
984 .rocinit = OBD_ROCOMPAT_LOVOBJID,
985 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
986 OBD_INCOMPAT_MULTI_OI,
988 [LDD_F_SV_TYPE_OST] = {
990 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
993 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
997 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
999 struct tgt_thread_info *tti = tgt_th_info(env);
1000 struct lr_server_data *lsd = &tgt->lut_lsd;
1001 unsigned long last_rcvd_size;
1005 rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr, BYPASS_CAPA);
1009 last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1011 /* ensure padding in the struct is the correct size */
1012 CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
1013 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
1015 rc = server_name2index(tgt_name(tgt), &index, NULL);
1017 CERROR("%s: Can not get index from name: rc = %d\n",
1021 /* server_name2index() returns type */
1023 if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1024 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1028 /* last_rcvd on OST doesn't provide reconstruct support because there
1029 * may be up to 8 in-flight write requests per single slot in
1030 * last_rcvd client data
1032 tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1034 if (last_rcvd_size == 0) {
1035 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1037 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1038 sizeof(lsd->lsd_uuid));
1039 lsd->lsd_last_transno = 0;
1040 lsd->lsd_mount_count = 0;
1041 lsd->lsd_server_size = LR_SERVER_SIZE;
1042 lsd->lsd_client_start = LR_CLIENT_START;
1043 lsd->lsd_client_size = LR_CLIENT_SIZE;
1044 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1045 lsd->lsd_osd_index = index;
1046 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1047 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1049 rc = tgt_server_data_read(env, tgt);
1051 CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1055 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1056 LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
1057 "using the wrong disk %s. Were the"
1058 " /dev/ assignments rearranged?\n",
1059 tgt->lut_obd->obd_uuid.uuid,
1064 if (lsd->lsd_osd_index != index) {
1065 LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
1066 "is different with the index %d in"
1067 "config log, It might be disk"
1068 "corruption!\n", tgt_name(tgt),
1069 lsd->lsd_osd_index, index);
1074 if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1075 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1077 lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1081 if (type == LDD_F_SV_TYPE_MDT)
1082 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1084 if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1085 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1087 lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1090 /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1091 if (type == LDD_F_SV_TYPE_MDT &&
1092 !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1093 if (last_rcvd_size > lsd->lsd_client_start) {
1094 LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1095 "remove all clients for interop needs\n",
1097 rc = tgt_truncate_last_rcvd(env, tgt,
1098 lsd->lsd_client_start);
1101 last_rcvd_size = lsd->lsd_client_start;
1103 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1104 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1107 spin_lock(&tgt->lut_translock);
1108 tgt->lut_last_transno = lsd->lsd_last_transno;
1109 spin_unlock(&tgt->lut_translock);
1111 lsd->lsd_mount_count++;
1113 CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1114 CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
1115 tgt_name(tgt), tgt->lut_last_transno);
1116 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
1117 tgt_name(tgt), lsd->lsd_mount_count);
1118 CDEBUG(D_INODE, "%s: server data size: %u\n",
1119 tgt_name(tgt), lsd->lsd_server_size);
1120 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1121 tgt_name(tgt), lsd->lsd_client_start);
1122 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1123 tgt_name(tgt), lsd->lsd_client_size);
1124 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1125 tgt_name(tgt), last_rcvd_size);
1126 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1127 tgt_name(tgt), lsd->lsd_subdir_count);
1128 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1129 last_rcvd_size <= lsd->lsd_client_start ? 0 :
1130 (last_rcvd_size - lsd->lsd_client_start) /
1131 lsd->lsd_client_size);
1132 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1134 if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1135 lsd->lsd_client_size == 0) {
1136 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1140 if (!tgt->lut_obd->obd_replayable)
1141 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1143 rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1145 GOTO(err_client, rc);
1147 spin_lock(&tgt->lut_translock);
1148 /* obd_last_committed is used for compatibility
1149 * with other lustre recovery code */
1150 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1151 spin_unlock(&tgt->lut_translock);
1153 tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1154 tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1156 /* save it, so mount count and last_transno is current */
1157 rc = tgt_server_data_update(env, tgt, 0);
1159 GOTO(err_client, rc);
1164 class_disconnect_exports(tgt->lut_obd);
1168 /* add credits for last_rcvd update */
1169 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1172 struct lu_target *tgt = cookie;
1173 struct tgt_session_info *tsi;
1174 struct tgt_thread_info *tti = tgt_th_info(env);
1177 /* if there is no session, then this transaction is not result of
1178 * request processing but some local operation */
1179 if (env->le_ses == NULL)
1182 LASSERT(tgt->lut_last_rcvd);
1183 tsi = tgt_ses_info(env);
1184 /* OFD may start transaction without export assigned */
1185 if (tsi->tsi_exp == NULL)
1189 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
1191 tsi->tsi_exp->exp_target_data.ted_lr_off,
1197 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
1198 &tti->tti_buf, 0, th);
1202 if (tsi->tsi_vbr_obj != NULL &&
1203 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu))
1204 rc = dt_declare_version_set(env, tsi->tsi_vbr_obj, th);
1209 /* Update last_rcvd records with latests transaction data */
1210 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1213 struct lu_target *tgt = cookie;
1214 struct tgt_session_info *tsi;
1215 struct tgt_thread_info *tti = tgt_th_info(env);
1216 struct dt_object *obj = NULL;
1220 if (env->le_ses == NULL)
1223 tsi = tgt_ses_info(env);
1224 /* OFD may start transaction without export assigned */
1225 if (tsi->tsi_exp == NULL)
1228 echo_client = (tgt_ses_req(tsi) == NULL);
1230 if (tti->tti_has_trans && !echo_client) {
1231 if (tti->tti_mult_trans == 0) {
1232 CDEBUG(D_HA, "More than one transaction "LPU64"\n",
1236 /* we need another transno to be assigned */
1237 tti->tti_transno = 0;
1238 } else if (th->th_result == 0) {
1239 tti->tti_has_trans = 1;
1242 if (tsi->tsi_vbr_obj != NULL &&
1243 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1244 obj = tsi->tsi_vbr_obj;
1247 if (unlikely(echo_client)) /* echo client special case */
1248 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1251 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,