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, 2014, 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->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);
131 int tgt_client_data_write(const struct lu_env *env, struct lu_target *tgt,
132 struct lsd_client_data *lcd, loff_t *off,
135 struct tgt_thread_info *tti = tgt_th_info(env);
137 lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
138 lcd->lcd_last_close_result =
139 ptlrpc_status_hton(lcd->lcd_last_close_result);
140 lcd_cpu_to_le(lcd, &tti->tti_lcd);
143 return dt_record_write(env, tgt->lut_last_rcvd, &tti->tti_buf, off, th);
147 * Update client data in last_rcvd
149 static int tgt_client_data_update(const struct lu_env *env,
150 struct obd_export *exp)
152 struct tg_export_data *ted = &exp->exp_target_data;
153 struct lu_target *tgt = class_exp2tgt(exp);
154 struct tgt_thread_info *tti = tgt_th_info(env);
160 if (unlikely(tgt == NULL)) {
161 CDEBUG(D_ERROR, "%s: No target for connected export\n",
162 class_exp2obd(exp)->obd_name);
166 th = dt_trans_create(env, tgt->lut_bottom);
171 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
173 ted->ted_lr_off, th);
177 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
181 * Until this operations will be committed the sync is needed
182 * for this export. This should be done _after_ starting the
183 * transaction so that many connecting clients will not bring
184 * server down with lots of sync writes.
186 rc = tgt_new_client_cb_add(th, exp);
188 /* can't add callback, do sync now */
191 spin_lock(&exp->exp_lock);
192 exp->exp_need_sync = 1;
193 spin_unlock(&exp->exp_lock);
196 tti->tti_off = ted->ted_lr_off;
197 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
200 dt_trans_stop(env, tgt->lut_bottom, th);
201 CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
202 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
203 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
208 int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
210 struct tgt_thread_info *tti = tgt_th_info(env);
215 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
218 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
220 CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
221 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
222 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
226 int tgt_server_data_write(const struct lu_env *env, struct lu_target *tgt,
229 struct tgt_thread_info *tti = tgt_th_info(env);
236 lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
238 rc = dt_record_write(env, tgt->lut_last_rcvd, &tti->tti_buf,
241 CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
242 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
243 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
249 * Update server data in last_rcvd
251 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
254 struct tgt_thread_info *tti = tgt_th_info(env);
261 "%s: mount_count is "LPU64", last_transno is "LPU64"\n",
262 tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
263 tgt->lut_last_transno);
265 /* Always save latest transno to keep it fresh */
266 spin_lock(&tgt->lut_translock);
267 tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
268 spin_unlock(&tgt->lut_translock);
270 th = dt_trans_create(env, tgt->lut_bottom);
277 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
278 &tti->tti_buf, tti->tti_off, th);
282 rc = dt_trans_start(env, tgt->lut_bottom, th);
286 rc = tgt_server_data_write(env, tgt, th);
288 dt_trans_stop(env, tgt->lut_bottom, th);
290 CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
291 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
292 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
295 EXPORT_SYMBOL(tgt_server_data_update);
297 int tgt_truncate_last_rcvd(const struct lu_env *env, struct lu_target *tgt,
300 struct dt_object *dt = tgt->lut_last_rcvd;
308 attr.la_valid = LA_SIZE;
310 th = dt_trans_create(env, tgt->lut_bottom);
313 rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
316 rc = dt_declare_attr_set(env, dt, &attr, th);
319 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
323 rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
325 rc = dt_attr_set(env, dt, &attr, th);
328 dt_trans_stop(env, tgt->lut_bottom, th);
333 static void tgt_client_epoch_update(const struct lu_env *env,
334 struct obd_export *exp)
336 struct lsd_client_data *lcd = exp->exp_target_data.ted_lcd;
337 struct lu_target *tgt = class_exp2tgt(exp);
339 LASSERT(tgt && tgt->lut_bottom);
340 /** VBR: set client last_epoch to current epoch */
341 if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
343 lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
344 tgt_client_data_update(env, exp);
348 * Update boot epoch when recovery ends
350 void tgt_boot_epoch_update(struct lu_target *tgt)
353 struct ptlrpc_request *req;
355 struct list_head client_list;
358 if (tgt->lut_obd->obd_stopping)
361 rc = lu_env_init(&env, LCT_DT_THREAD);
363 CERROR("%s: can't initialize environment: rc = %d\n",
364 tgt->lut_obd->obd_name, rc);
368 spin_lock(&tgt->lut_translock);
369 start_epoch = lr_epoch(tgt->lut_last_transno) + 1;
370 tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
371 tgt->lut_lsd.lsd_start_epoch = start_epoch;
372 spin_unlock(&tgt->lut_translock);
374 INIT_LIST_HEAD(&client_list);
376 * The recovery is not yet finished and final queue can still be updated
377 * with resend requests. Move final list to separate one for processing
379 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
380 list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
381 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
384 * go through list of exports participated in recovery and
385 * set new epoch for them
387 list_for_each_entry(req, &client_list, rq_list) {
388 LASSERT(!req->rq_export->exp_delayed);
389 if (!req->rq_export->exp_vbr_failed)
390 tgt_client_epoch_update(&env, req->rq_export);
392 /** return list back at once */
393 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
394 list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
395 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
396 /** update server epoch */
397 tgt_server_data_update(&env, tgt, 1);
402 * commit callback, need to update last_commited value
404 struct tgt_last_committed_callback {
405 struct dt_txn_commit_cb llcc_cb;
406 struct lu_target *llcc_tgt;
407 struct obd_export *llcc_exp;
411 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
412 struct dt_txn_commit_cb *cb, int err)
414 struct tgt_last_committed_callback *ccb;
416 ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
418 LASSERT(ccb->llcc_tgt != NULL);
419 LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
421 spin_lock(&ccb->llcc_tgt->lut_translock);
422 if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
423 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
425 LASSERT(ccb->llcc_exp);
426 if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
427 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
428 spin_unlock(&ccb->llcc_tgt->lut_translock);
429 ptlrpc_commit_replies(ccb->llcc_exp);
431 spin_unlock(&ccb->llcc_tgt->lut_translock);
433 class_export_cb_put(ccb->llcc_exp);
434 if (ccb->llcc_transno)
435 CDEBUG(D_HA, "%s: transno "LPD64" is committed\n",
436 ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
440 int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
441 struct obd_export *exp, __u64 transno)
443 struct tgt_last_committed_callback *ccb;
444 struct dt_txn_commit_cb *dcb;
452 ccb->llcc_exp = class_export_cb_get(exp);
453 ccb->llcc_transno = transno;
456 dcb->dcb_func = tgt_cb_last_committed;
457 INIT_LIST_HEAD(&dcb->dcb_linkage);
458 strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
460 rc = dt_trans_cb_add(th, dcb);
462 class_export_cb_put(exp);
466 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
467 /* report failure to force synchronous operation */
473 struct tgt_new_client_callback {
474 struct dt_txn_commit_cb lncc_cb;
475 struct obd_export *lncc_exp;
478 static 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 INIT_LIST_HEAD(&dcb->dcb_linkage);
519 strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
521 rc = dt_trans_cb_add(th, dcb);
523 class_export_cb_put(exp);
530 * Add new client to the last_rcvd upon new connection.
532 * We use a bitmap to locate a free space in the last_rcvd file and initialize
535 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
537 struct tg_export_data *ted = &exp->exp_target_data;
538 struct lu_target *tgt = class_exp2tgt(exp);
543 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
544 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
547 mutex_init(&ted->ted_lcd_lock);
549 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
552 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
553 * there's no need for extra complication here
555 idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
557 if (idx >= LR_MAX_CLIENTS ||
558 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
559 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
560 tgt->lut_obd->obd_name, idx);
563 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
564 idx = find_next_zero_bit(tgt->lut_client_bitmap,
565 LR_MAX_CLIENTS, idx);
569 CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added\n",
570 tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid);
572 ted->ted_lr_idx = idx;
573 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
574 idx * tgt->lut_lsd.lsd_client_size;
576 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
578 CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s'\n",
579 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
580 ted->ted_lcd->lcd_uuid);
582 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
585 rc = tgt_client_data_update(env, exp);
587 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
588 tgt->lut_obd->obd_name, idx, rc);
592 EXPORT_SYMBOL(tgt_client_new);
594 /* Add client data to the MDS. We use a bitmap to locate a free space
595 * in the last_rcvd file if cl_off is -1 (i.e. a new client).
596 * Otherwise, we just have to read the data from the last_rcvd file and
597 * we know its offset.
599 * It should not be possible to fail adding an existing client - otherwise
600 * mdt_init_server_data() callsite needs to be fixed.
602 int tgt_client_add(const struct lu_env *env, struct obd_export *exp, int idx)
604 struct tg_export_data *ted = &exp->exp_target_data;
605 struct lu_target *tgt = class_exp2tgt(exp);
609 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
610 LASSERTF(idx >= 0, "%d\n", idx);
612 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
613 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
616 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
617 CERROR("%s: client %d: bit already set in bitmap!!\n",
618 tgt->lut_obd->obd_name, idx);
622 CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added\n",
623 tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid);
625 ted->ted_lr_idx = idx;
626 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
627 idx * tgt->lut_lsd.lsd_client_size;
629 mutex_init(&ted->ted_lcd_lock);
631 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
636 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
638 struct tg_export_data *ted = &exp->exp_target_data;
639 struct lu_target *tgt = class_exp2tgt(exp);
644 LASSERT(ted->ted_lcd);
646 if (unlikely(tgt == NULL)) {
647 CDEBUG(D_ERROR, "%s: No target for connected export\n",
648 class_exp2obd(exp)->obd_name);
652 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
653 if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
654 (char *)tgt->lut_obd->obd_uuid.uuid) ||
655 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
658 CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
659 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
660 ted->ted_lcd->lcd_uuid);
662 /* Clear the bit _after_ zeroing out the client so we don't
663 race with filter_client_add and zero out new clients.*/
664 if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
665 CERROR("%s: client %u: bit already clear in bitmap!!\n",
666 tgt->lut_obd->obd_name, ted->ted_lr_idx);
670 /* Do not erase record for recoverable client. */
671 if (exp->exp_flags & OBD_OPT_FAILOVER)
674 /* Make sure the server's last_transno is up to date.
675 * This should be done before zeroing client slot so last_transno will
676 * be in server data or in client data in case of failure */
677 rc = tgt_server_data_update(env, tgt, 0);
679 CERROR("%s: failed to update server data, skip client %s "
680 "zeroing, rc %d\n", tgt->lut_obd->obd_name,
681 ted->ted_lcd->lcd_uuid, rc);
685 mutex_lock(&ted->ted_lcd_lock);
686 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
687 rc = tgt_client_data_update(env, exp);
688 mutex_unlock(&ted->ted_lcd_lock);
690 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
691 "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
692 tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
693 ted->ted_lr_idx, ted->ted_lr_off, rc);
696 EXPORT_SYMBOL(tgt_client_del);
699 * last_rcvd & last_committed update callbacks
701 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
702 struct dt_object *obj, __u64 opdata,
703 struct thandle *th, struct ptlrpc_request *req)
705 struct tgt_thread_info *tti = tgt_th_info(env);
706 struct tg_export_data *ted;
709 bool lw_client, update = false;
713 ted = &req->rq_export->exp_target_data;
715 lw_client = exp_connect_flags(req->rq_export) & OBD_CONNECT_LIGHTWEIGHT;
716 if (ted->ted_lr_idx < 0 && !lw_client)
717 /* ofd connect may cause transaction before export has
721 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
723 spin_lock(&tgt->lut_translock);
724 if (th->th_result != 0) {
725 if (tti->tti_transno != 0) {
726 CERROR("%s: replay transno "LPU64" failed: rc = %d\n",
727 tgt_name(tgt), tti->tti_transno, th->th_result);
729 } else if (tti->tti_transno == 0) {
730 tti->tti_transno = ++tgt->lut_last_transno;
732 /* should be replay */
733 if (tti->tti_transno > tgt->lut_last_transno)
734 tgt->lut_last_transno = tti->tti_transno;
736 spin_unlock(&tgt->lut_translock);
738 /** VBR: set new versions */
739 if (th->th_result == 0 && obj != NULL)
740 dt_version_set(env, obj, tti->tti_transno, th);
742 /* filling reply data */
743 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
744 tti->tti_transno, tgt->lut_obd->obd_last_committed);
746 req->rq_transno = tti->tti_transno;
747 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
749 /* if can't add callback, do sync write */
750 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, req->rq_export,
754 /* All operations performed by LW clients are synchronous and
755 * we store the committed transno in the last_rcvd header */
756 spin_lock(&tgt->lut_translock);
757 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
758 tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
761 spin_unlock(&tgt->lut_translock);
762 /* Although lightweight (LW) connections have no slot in
763 * last_rcvd, we still want to maintain the in-memory
764 * lsd_client_data structure in order to properly handle reply
766 } else if (ted->ted_lr_off == 0) {
767 CERROR("%s: client idx %d has offset %lld\n",
768 tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
772 /* if the export has already been disconnected, we have no last_rcvd
773 * slot, update server data with latest transno then */
774 if (ted->ted_lcd == NULL) {
775 CWARN("commit transaction for disconnected client %s: rc %d\n",
776 req->rq_export->exp_client_uuid.uuid, rc);
777 GOTO(srv_update, rc = 0);
780 mutex_lock(&ted->ted_lcd_lock);
781 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
782 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
783 transno_p = &ted->ted_lcd->lcd_last_close_transno;
784 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
785 ted->ted_lcd->lcd_last_close_result = th->th_result;
787 /* VBR: save versions in last_rcvd for reconstruct. */
788 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
791 ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
792 ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
793 ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
794 ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
796 transno_p = &ted->ted_lcd->lcd_last_transno;
797 ted->ted_lcd->lcd_last_xid = req->rq_xid;
798 ted->ted_lcd->lcd_last_result = th->th_result;
799 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
800 * see struct ldlm_reply->lock_policy_res1; */
801 ted->ted_lcd->lcd_last_data = opdata;
804 /* Update transno in slot only if non-zero number, i.e. no errors */
805 if (likely(tti->tti_transno != 0)) {
806 if (*transno_p > tti->tti_transno &&
807 !tgt->lut_no_reconstruct) {
808 CERROR("%s: trying to overwrite bigger transno:"
809 "on-disk: "LPU64", new: "LPU64" replay: %d. "
810 "see LU-617.\n", tgt_name(tgt), *transno_p,
811 tti->tti_transno, req_is_replay(req));
812 if (req_is_replay(req)) {
813 spin_lock(&req->rq_export->exp_lock);
814 req->rq_export->exp_vbr_failed = 1;
815 spin_unlock(&req->rq_export->exp_lock);
817 mutex_unlock(&ted->ted_lcd_lock);
818 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
820 *transno_p = tti->tti_transno;
824 tti->tti_off = ted->ted_lr_off;
825 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
827 mutex_unlock(&ted->ted_lcd_lock);
831 mutex_unlock(&ted->ted_lcd_lock);
835 rc = tgt_server_data_write(env, tgt, th);
840 * last_rcvd update for echo client simulation.
841 * It updates last_rcvd client slot and version of object in
842 * simple way but with all locks to simulate all drawbacks
844 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
845 struct lu_target *tgt,
846 struct dt_object *obj,
848 struct obd_export *exp)
850 struct tgt_thread_info *tti = tgt_th_info(env);
851 struct tg_export_data *ted = &exp->exp_target_data;
856 tti->tti_transno = 0;
858 spin_lock(&tgt->lut_translock);
859 if (th->th_result == 0)
860 tti->tti_transno = ++tgt->lut_last_transno;
861 spin_unlock(&tgt->lut_translock);
863 /** VBR: set new versions */
864 if (th->th_result == 0 && obj != NULL)
865 dt_version_set(env, obj, tti->tti_transno, th);
867 /* if can't add callback, do sync write */
868 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
871 LASSERT(ted->ted_lr_off > 0);
873 mutex_lock(&ted->ted_lcd_lock);
874 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
875 ted->ted_lcd->lcd_last_transno = tti->tti_transno;
876 ted->ted_lcd->lcd_last_result = th->th_result;
878 tti->tti_off = ted->ted_lr_off;
879 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
880 mutex_unlock(&ted->ted_lcd_lock);
884 static int tgt_clients_data_init(const struct lu_env *env,
885 struct lu_target *tgt,
886 unsigned long last_size)
888 struct obd_device *obd = tgt->lut_obd;
889 struct lr_server_data *lsd = &tgt->lut_lsd;
890 struct lsd_client_data *lcd = NULL;
891 struct tg_export_data *ted;
894 loff_t off = lsd->lsd_client_start;
898 CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
899 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
905 for (cl_idx = 0; off < last_size; cl_idx++) {
906 struct obd_export *exp;
909 /* Don't assume off is incremented properly by
910 * read_record(), in case sizeof(*lcd)
911 * isn't the same as fsd->lsd_client_size. */
912 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
913 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
915 CERROR("%s: error reading last_rcvd %s idx %d off "
916 "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
919 break; /* read error shouldn't cause startup to fail */
922 if (lcd->lcd_uuid[0] == '\0') {
923 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
928 last_transno = lcd_last_transno(lcd);
930 /* These exports are cleaned up by disconnect, so they
931 * need to be set up like real exports as connect does.
933 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
934 " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
935 last_transno, lsd->lsd_last_transno, lcd_last_xid(lcd));
937 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
939 if (PTR_ERR(exp) == -EALREADY) {
940 /* export already exists, zero out this one */
941 CERROR("%s: Duplicate export %s!\n",
942 tgt_name(tgt), lcd->lcd_uuid);
945 GOTO(err_out, rc = PTR_ERR(exp));
948 ted = &exp->exp_target_data;
949 *ted->ted_lcd = *lcd;
951 rc = tgt_client_add(env, exp, cl_idx);
952 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
953 /* VBR: set export last committed version */
954 exp->exp_last_committed = last_transno;
955 spin_lock(&exp->exp_lock);
956 exp->exp_connecting = 0;
957 exp->exp_in_recovery = 0;
958 spin_unlock(&exp->exp_lock);
959 obd->obd_max_recoverable_clients++;
960 class_export_put(exp);
962 /* Need to check last_rcvd even for duplicated exports. */
963 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
964 cl_idx, last_transno);
966 spin_lock(&tgt->lut_translock);
967 tgt->lut_last_transno = max(last_transno,
968 tgt->lut_last_transno);
969 spin_unlock(&tgt->lut_translock);
977 struct server_compat_data {
984 static struct server_compat_data tgt_scd[] = {
985 [LDD_F_SV_TYPE_MDT] = {
986 .rocompat = OBD_ROCOMPAT_LOVOBJID,
987 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
988 OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
989 OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI,
990 .rocinit = OBD_ROCOMPAT_LOVOBJID,
991 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
992 OBD_INCOMPAT_MULTI_OI,
994 [LDD_F_SV_TYPE_OST] = {
995 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
996 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
998 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
999 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1003 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1005 struct tgt_thread_info *tti = tgt_th_info(env);
1006 struct lr_server_data *lsd = &tgt->lut_lsd;
1007 unsigned long last_rcvd_size;
1011 rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1015 last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1017 /* ensure padding in the struct is the correct size */
1018 CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
1019 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
1021 rc = server_name2index(tgt_name(tgt), &index, NULL);
1023 CERROR("%s: Can not get index from name: rc = %d\n",
1027 /* server_name2index() returns type */
1029 if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1030 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1034 /* last_rcvd on OST doesn't provide reconstruct support because there
1035 * may be up to 8 in-flight write requests per single slot in
1036 * last_rcvd client data
1038 tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1040 if (last_rcvd_size == 0) {
1041 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1043 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1044 sizeof(lsd->lsd_uuid));
1045 lsd->lsd_last_transno = 0;
1046 lsd->lsd_mount_count = 0;
1047 lsd->lsd_server_size = LR_SERVER_SIZE;
1048 lsd->lsd_client_start = LR_CLIENT_START;
1049 lsd->lsd_client_size = LR_CLIENT_SIZE;
1050 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1051 lsd->lsd_osd_index = index;
1052 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1053 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1055 rc = tgt_server_data_read(env, tgt);
1057 CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1061 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1062 LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
1063 "using the wrong disk %s. Were the"
1064 " /dev/ assignments rearranged?\n",
1065 tgt->lut_obd->obd_uuid.uuid,
1070 if (lsd->lsd_osd_index != index) {
1071 LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
1072 "is different with the index %d in"
1073 "config log, It might be disk"
1074 "corruption!\n", tgt_name(tgt),
1075 lsd->lsd_osd_index, index);
1080 if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1081 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1083 lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1087 if (type == LDD_F_SV_TYPE_MDT)
1088 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1090 if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1091 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1093 lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1096 /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1097 if (type == LDD_F_SV_TYPE_MDT &&
1098 !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1099 if (last_rcvd_size > lsd->lsd_client_start) {
1100 LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1101 "remove all clients for interop needs\n",
1103 rc = tgt_truncate_last_rcvd(env, tgt,
1104 lsd->lsd_client_start);
1107 last_rcvd_size = lsd->lsd_client_start;
1109 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1110 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1113 spin_lock(&tgt->lut_translock);
1114 tgt->lut_last_transno = lsd->lsd_last_transno;
1115 spin_unlock(&tgt->lut_translock);
1117 lsd->lsd_mount_count++;
1119 CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1120 CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
1121 tgt_name(tgt), tgt->lut_last_transno);
1122 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
1123 tgt_name(tgt), lsd->lsd_mount_count);
1124 CDEBUG(D_INODE, "%s: server data size: %u\n",
1125 tgt_name(tgt), lsd->lsd_server_size);
1126 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1127 tgt_name(tgt), lsd->lsd_client_start);
1128 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1129 tgt_name(tgt), lsd->lsd_client_size);
1130 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1131 tgt_name(tgt), last_rcvd_size);
1132 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1133 tgt_name(tgt), lsd->lsd_subdir_count);
1134 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1135 last_rcvd_size <= lsd->lsd_client_start ? 0 :
1136 (last_rcvd_size - lsd->lsd_client_start) /
1137 lsd->lsd_client_size);
1138 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1140 if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1141 lsd->lsd_client_size == 0) {
1142 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1146 if (!tgt->lut_obd->obd_replayable)
1147 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1149 rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1151 GOTO(err_client, rc);
1153 spin_lock(&tgt->lut_translock);
1154 /* obd_last_committed is used for compatibility
1155 * with other lustre recovery code */
1156 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1157 spin_unlock(&tgt->lut_translock);
1159 tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1160 tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1162 /* save it, so mount count and last_transno is current */
1163 rc = tgt_server_data_update(env, tgt, 0);
1165 GOTO(err_client, rc);
1170 class_disconnect_exports(tgt->lut_obd);
1174 /* add credits for last_rcvd update */
1175 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1178 struct lu_target *tgt = cookie;
1179 struct tgt_session_info *tsi;
1180 struct tgt_thread_info *tti = tgt_th_info(env);
1183 /* if there is no session, then this transaction is not result of
1184 * request processing but some local operation */
1185 if (env->le_ses == NULL)
1188 LASSERT(tgt->lut_last_rcvd);
1189 tsi = tgt_ses_info(env);
1190 /* OFD may start transaction without export assigned */
1191 if (tsi->tsi_exp == NULL)
1195 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
1197 tsi->tsi_exp->exp_target_data.ted_lr_off,
1203 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
1204 &tti->tti_buf, 0, th);
1208 if (tsi->tsi_vbr_obj != NULL &&
1209 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu))
1210 rc = dt_declare_version_set(env, tsi->tsi_vbr_obj, th);
1215 /* Update last_rcvd records with latests transaction data */
1216 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1219 struct lu_target *tgt = cookie;
1220 struct tgt_session_info *tsi;
1221 struct tgt_thread_info *tti = tgt_th_info(env);
1222 struct dt_object *obj = NULL;
1226 if (env->le_ses == NULL)
1229 tsi = tgt_ses_info(env);
1230 /* OFD may start transaction without export assigned */
1231 if (tsi->tsi_exp == NULL)
1234 echo_client = (tgt_ses_req(tsi) == NULL);
1236 if (tti->tti_has_trans && !echo_client) {
1237 if (tti->tti_mult_trans == 0) {
1238 CDEBUG(D_HA, "More than one transaction "LPU64"\n",
1242 /* we need another transno to be assigned */
1243 tti->tti_transno = 0;
1244 } else if (th->th_result == 0) {
1245 tti->tti_has_trans = 1;
1248 if (tsi->tsi_vbr_obj != NULL &&
1249 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1250 obj = tsi->tsi_vbr_obj;
1253 if (unlikely(echo_client)) /* echo client special case */
1254 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1257 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,