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) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2015, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Lustre Unified Target
33 * These are common function to work with last_received file
35 * Author: Mikhail Pershin <mike.pershin@intel.com>
38 #include <obd_class.h>
39 #include <lustre_fid.h>
41 #include "tgt_internal.h"
44 /* Allocate a bitmap for a chunk of reply data slots */
45 static int tgt_bitmap_chunk_alloc(struct lu_target *lut, int chunk)
49 OBD_ALLOC_LARGE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
54 spin_lock(&lut->lut_client_bitmap_lock);
56 if (lut->lut_reply_bitmap[chunk] != NULL) {
57 /* someone else already allocated the bitmap for this chunk */
58 spin_unlock(&lut->lut_client_bitmap_lock);
59 OBD_FREE_LARGE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
64 lut->lut_reply_bitmap[chunk] = bm;
66 spin_unlock(&lut->lut_client_bitmap_lock);
71 /* Look for an available reply data slot in the bitmap
73 * Allocate bitmap chunk when first used
74 * XXX algo could be improved if this routine limits performance
76 static int tgt_find_free_reply_slot(struct lu_target *lut)
83 for (chunk = 0; chunk < LUT_REPLY_SLOTS_MAX_CHUNKS; chunk++) {
84 /* allocate the bitmap chunk if necessary */
85 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
86 rc = tgt_bitmap_chunk_alloc(lut, chunk);
90 bmp = lut->lut_reply_bitmap[chunk];
92 /* look for an available slot in this chunk */
94 b = find_first_zero_bit(bmp, LUT_REPLY_SLOTS_PER_CHUNK);
95 if (b >= LUT_REPLY_SLOTS_PER_CHUNK)
99 if (test_and_set_bit(b, bmp) == 0)
100 return chunk * LUT_REPLY_SLOTS_PER_CHUNK + b;
107 /* Mark the reply data slot @idx 'used' in the corresponding bitmap chunk
109 * Allocate the bitmap chunk if necessary
111 static int tgt_set_reply_slot(struct lu_target *lut, int idx)
117 chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
118 b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
120 LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
121 LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
123 /* allocate the bitmap chunk if necessary */
124 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
125 rc = tgt_bitmap_chunk_alloc(lut, chunk);
130 /* mark the slot 'used' in this chunk */
131 if (test_and_set_bit(b, lut->lut_reply_bitmap[chunk]) != 0) {
132 CERROR("%s: slot %d already set in bitmap\n",
141 /* Mark the reply data slot @idx 'unused' in the corresponding bitmap chunk
144 static int tgt_clear_reply_slot(struct lu_target *lut, int idx)
149 chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
150 b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
152 LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
153 LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
155 if (lut->lut_reply_bitmap[chunk] == NULL) {
156 CERROR("%s: slot %d not allocated\n",
161 if (test_and_clear_bit(b, lut->lut_reply_bitmap[chunk]) == 0) {
162 CERROR("%s: slot %d already clear in bitmap\n",
171 /* Read header of reply_data file of target @tgt into structure @lrh */
172 static int tgt_reply_header_read(const struct lu_env *env,
173 struct lu_target *tgt,
174 struct lsd_reply_header *lrh)
177 struct lsd_reply_header buf;
178 struct tgt_thread_info *tti = tgt_th_info(env);
181 tti->tti_buf.lb_buf = &buf;
182 tti->tti_buf.lb_len = sizeof(buf);
184 rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
189 lrh->lrh_magic = le32_to_cpu(buf.lrh_magic);
190 lrh->lrh_header_size = le32_to_cpu(buf.lrh_header_size);
191 lrh->lrh_reply_size = le32_to_cpu(buf.lrh_reply_size);
193 CDEBUG(D_HA, "%s: read %s header. magic=0x%08x "
194 "header_size=%d reply_size=%d\n",
195 tgt->lut_obd->obd_name, REPLY_DATA,
196 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
201 /* Write header into replay_data file of target @tgt from structure @lrh */
202 static int tgt_reply_header_write(const struct lu_env *env,
203 struct lu_target *tgt,
204 struct lsd_reply_header *lrh)
207 struct lsd_reply_header buf;
208 struct tgt_thread_info *tti = tgt_th_info(env);
210 struct dt_object *dto;
212 CDEBUG(D_HA, "%s: write %s header. magic=0x%08x "
213 "header_size=%d reply_size=%d\n",
214 tgt->lut_obd->obd_name, REPLY_DATA,
215 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
217 buf.lrh_magic = cpu_to_le32(lrh->lrh_magic);
218 buf.lrh_header_size = cpu_to_le32(lrh->lrh_header_size);
219 buf.lrh_reply_size = cpu_to_le32(lrh->lrh_reply_size);
221 th = dt_trans_create(env, tgt->lut_bottom);
227 tti->tti_buf.lb_buf = &buf;
228 tti->tti_buf.lb_len = sizeof(buf);
230 rc = dt_declare_record_write(env, tgt->lut_reply_data,
231 &tti->tti_buf, tti->tti_off, th);
235 rc = dt_trans_start(env, tgt->lut_bottom, th);
239 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
240 rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
242 dt_trans_stop(env, tgt->lut_bottom, th);
246 /* Write the reply data @lrd into reply_data file of target @tgt
249 static int tgt_reply_data_write(const struct lu_env *env, struct lu_target *tgt,
250 struct lsd_reply_data *lrd, loff_t off,
253 struct tgt_thread_info *tti = tgt_th_info(env);
254 struct dt_object *dto;
255 struct lsd_reply_data *buf = &tti->tti_lrd;
257 lrd->lrd_result = ptlrpc_status_hton(lrd->lrd_result);
259 buf->lrd_transno = cpu_to_le64(lrd->lrd_transno);
260 buf->lrd_xid = cpu_to_le64(lrd->lrd_xid);
261 buf->lrd_data = cpu_to_le64(lrd->lrd_data);
262 buf->lrd_result = cpu_to_le32(lrd->lrd_result);
263 buf->lrd_client_gen = cpu_to_le32(lrd->lrd_client_gen);
265 lrd->lrd_result = ptlrpc_status_ntoh(lrd->lrd_result);
268 tti->tti_buf.lb_buf = buf;
269 tti->tti_buf.lb_len = sizeof(*buf);
271 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
272 return dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
275 /* Read the reply data from reply_data file of target @tgt at offset @off
276 * into structure @lrd
278 static int tgt_reply_data_read(const struct lu_env *env, struct lu_target *tgt,
279 struct lsd_reply_data *lrd, loff_t off)
282 struct tgt_thread_info *tti = tgt_th_info(env);
283 struct lsd_reply_data *buf = &tti->tti_lrd;
286 tti->tti_buf.lb_buf = buf;
287 tti->tti_buf.lb_len = sizeof(*buf);
289 rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
294 lrd->lrd_transno = le64_to_cpu(buf->lrd_transno);
295 lrd->lrd_xid = le64_to_cpu(buf->lrd_xid);
296 lrd->lrd_data = le64_to_cpu(buf->lrd_data);
297 lrd->lrd_result = le32_to_cpu(buf->lrd_result);
298 lrd->lrd_client_gen = le32_to_cpu(buf->lrd_client_gen);
304 /* Free the in-memory reply data structure @trd and release
305 * the corresponding slot in the reply_data file of target @lut
306 * Called with ted_lcd_lock held
308 static void tgt_free_reply_data(struct lu_target *lut,
309 struct tg_export_data *ted,
310 struct tg_reply_data *trd)
312 CDEBUG(D_TRACE, "%s: free reply data %p: xid %llu, transno %llu, "
313 "client gen %u, slot idx %d\n",
314 lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
315 trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
318 LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
320 list_del(&trd->trd_list);
321 ted->ted_reply_cnt--;
323 tgt_clear_reply_slot(lut, trd->trd_index);
327 /* Release the reply data @trd from target @lut
328 * The reply data with the highest transno for this export
329 * is retained to ensure correctness of target recovery
330 * Called with ted_lcd_lock held
332 static void tgt_release_reply_data(struct lu_target *lut,
333 struct tg_export_data *ted,
334 struct tg_reply_data *trd)
336 CDEBUG(D_TRACE, "%s: release reply data %p: xid %llu, transno %llu, "
337 "client gen %u, slot idx %d\n",
338 lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
339 trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
342 LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
344 /* Do not free the reply data corresponding to the
345 * highest transno of this export.
346 * This ensures on-disk reply data is kept and
347 * last committed transno can be restored from disk in case
350 if (trd->trd_reply.lrd_transno == ted->ted_lcd->lcd_last_transno) {
351 /* free previous retained reply */
352 if (ted->ted_reply_last != NULL)
353 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
354 /* retain the reply */
355 list_del_init(&trd->trd_list);
356 ted->ted_reply_last = trd;
358 tgt_free_reply_data(lut, ted, trd);
362 static inline struct lu_buf *tti_buf_lsd(struct tgt_thread_info *tti)
364 tti->tti_buf.lb_buf = &tti->tti_lsd;
365 tti->tti_buf.lb_len = sizeof(tti->tti_lsd);
366 return &tti->tti_buf;
369 static inline struct lu_buf *tti_buf_lcd(struct tgt_thread_info *tti)
371 tti->tti_buf.lb_buf = &tti->tti_lcd;
372 tti->tti_buf.lb_len = sizeof(tti->tti_lcd);
373 return &tti->tti_buf;
377 * Allocate in-memory data for client slot related to export.
379 int tgt_client_alloc(struct obd_export *exp)
382 LASSERT(exp != exp->exp_obd->obd_self_export);
384 spin_lock_init(&exp->exp_target_data.ted_nodemap_lock);
385 INIT_LIST_HEAD(&exp->exp_target_data.ted_nodemap_member);
387 OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
388 if (exp->exp_target_data.ted_lcd == NULL)
390 /* Mark that slot is not yet valid, 0 doesn't work here */
391 exp->exp_target_data.ted_lr_idx = -1;
392 INIT_LIST_HEAD(&exp->exp_target_data.ted_reply_list);
393 mutex_init(&exp->exp_target_data.ted_lcd_lock);
396 EXPORT_SYMBOL(tgt_client_alloc);
399 * Free in-memory data for client slot related to export.
401 void tgt_client_free(struct obd_export *exp)
403 struct tg_export_data *ted = &exp->exp_target_data;
404 struct lu_target *lut = class_exp2tgt(exp);
405 struct tg_reply_data *trd, *tmp;
407 LASSERT(exp != exp->exp_obd->obd_self_export);
409 /* free reply data */
410 mutex_lock(&ted->ted_lcd_lock);
411 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
412 tgt_release_reply_data(lut, ted, trd);
414 if (ted->ted_reply_last != NULL) {
415 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
416 ted->ted_reply_last = NULL;
418 mutex_unlock(&ted->ted_lcd_lock);
420 if (!hlist_unhashed(&exp->exp_gen_hash))
421 cfs_hash_del(exp->exp_obd->obd_gen_hash,
422 &ted->ted_lcd->lcd_generation,
425 OBD_FREE_PTR(ted->ted_lcd);
428 /* Target may have been freed (see LU-7430)
429 * Slot may be not yet assigned */
430 if (exp->exp_obd->u.obt.obt_magic != OBT_MAGIC ||
434 /* Clear bit when lcd is freed */
435 LASSERT(lut && lut->lut_client_bitmap);
436 if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
437 CERROR("%s: client %u bit already clear in bitmap\n",
438 exp->exp_obd->obd_name, ted->ted_lr_idx);
442 if (tgt_is_multimodrpcs_client(exp) && !exp->exp_obd->obd_stopping)
443 atomic_dec(&lut->lut_num_clients);
445 EXPORT_SYMBOL(tgt_client_free);
447 int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
448 struct lsd_client_data *lcd, loff_t *off, int index)
450 struct tgt_thread_info *tti = tgt_th_info(env);
454 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
456 check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
457 lcd_le_to_cpu(&tti->tti_lcd, lcd);
458 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
459 lcd->lcd_last_close_result =
460 ptlrpc_status_ntoh(lcd->lcd_last_close_result);
463 CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = %llu"
464 ", last_xid = %llu, last_result = %u, last_data = %u, "
465 "last_close_transno = %llu, last_close_xid = %llu, "
466 "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
467 *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
468 lcd->lcd_last_result, lcd->lcd_last_data,
469 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
470 lcd->lcd_last_close_result, rc);
474 int tgt_client_data_write(const struct lu_env *env, struct lu_target *tgt,
475 struct lsd_client_data *lcd, loff_t *off,
478 struct tgt_thread_info *tti = tgt_th_info(env);
479 struct dt_object *dto;
481 lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
482 lcd->lcd_last_close_result =
483 ptlrpc_status_hton(lcd->lcd_last_close_result);
484 lcd_cpu_to_le(lcd, &tti->tti_lcd);
487 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
488 return dt_record_write(env, dto, &tti->tti_buf, off, th);
492 * Update client data in last_rcvd
494 static int tgt_client_data_update(const struct lu_env *env,
495 struct obd_export *exp)
497 struct tg_export_data *ted = &exp->exp_target_data;
498 struct lu_target *tgt = class_exp2tgt(exp);
499 struct tgt_thread_info *tti = tgt_th_info(env);
505 if (unlikely(tgt == NULL)) {
506 CDEBUG(D_ERROR, "%s: No target for connected export\n",
507 class_exp2obd(exp)->obd_name);
511 th = dt_trans_create(env, tgt->lut_bottom);
516 mutex_lock(&ted->ted_lcd_lock);
517 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
519 ted->ted_lr_off, th);
523 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
527 * Until this operations will be committed the sync is needed
528 * for this export. This should be done _after_ starting the
529 * transaction so that many connecting clients will not bring
530 * server down with lots of sync writes.
532 rc = tgt_new_client_cb_add(th, exp);
534 /* can't add callback, do sync now */
537 spin_lock(&exp->exp_lock);
538 exp->exp_need_sync = 1;
539 spin_unlock(&exp->exp_lock);
542 tti->tti_off = ted->ted_lr_off;
543 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
546 mutex_unlock(&ted->ted_lcd_lock);
547 dt_trans_stop(env, tgt->lut_bottom, th);
548 CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
549 "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
550 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
555 int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
557 struct tgt_thread_info *tti = tgt_th_info(env);
562 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
565 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
567 CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
568 "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
569 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
573 int tgt_server_data_write(const struct lu_env *env, struct lu_target *tgt,
576 struct tgt_thread_info *tti = tgt_th_info(env);
577 struct dt_object *dto;
584 lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
586 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
587 rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
589 CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
590 "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
591 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
597 * Update server data in last_rcvd
599 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
602 struct tgt_thread_info *tti = tgt_th_info(env);
609 "%s: mount_count is %llu, last_transno is %llu\n",
610 tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
611 tgt->lut_last_transno);
613 /* Always save latest transno to keep it fresh */
614 spin_lock(&tgt->lut_translock);
615 tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
616 spin_unlock(&tgt->lut_translock);
618 th = dt_trans_create(env, tgt->lut_bottom);
625 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
626 &tti->tti_buf, tti->tti_off, th);
630 rc = dt_trans_start(env, tgt->lut_bottom, th);
634 rc = tgt_server_data_write(env, tgt, th);
636 dt_trans_stop(env, tgt->lut_bottom, th);
638 CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
639 "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
640 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
643 EXPORT_SYMBOL(tgt_server_data_update);
645 int tgt_truncate_last_rcvd(const struct lu_env *env, struct lu_target *tgt,
648 struct dt_object *dt = tgt->lut_last_rcvd;
656 attr.la_valid = LA_SIZE;
658 th = dt_trans_create(env, tgt->lut_bottom);
661 rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
664 rc = dt_declare_attr_set(env, dt, &attr, th);
667 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
671 rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
673 rc = dt_attr_set(env, dt, &attr, th);
676 dt_trans_stop(env, tgt->lut_bottom, th);
681 static void tgt_client_epoch_update(const struct lu_env *env,
682 struct obd_export *exp)
684 struct lsd_client_data *lcd = exp->exp_target_data.ted_lcd;
685 struct lu_target *tgt = class_exp2tgt(exp);
687 LASSERT(tgt && tgt->lut_bottom);
688 /** VBR: set client last_epoch to current epoch */
689 if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
691 lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
692 tgt_client_data_update(env, exp);
696 * Update boot epoch when recovery ends
698 void tgt_boot_epoch_update(struct lu_target *tgt)
701 struct ptlrpc_request *req;
703 struct list_head client_list;
706 if (tgt->lut_obd->obd_stopping)
709 rc = lu_env_init(&env, LCT_DT_THREAD);
711 CERROR("%s: can't initialize environment: rc = %d\n",
712 tgt->lut_obd->obd_name, rc);
716 spin_lock(&tgt->lut_translock);
717 start_epoch = lr_epoch(tgt->lut_last_transno) + 1;
718 tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
719 tgt->lut_lsd.lsd_start_epoch = start_epoch;
720 spin_unlock(&tgt->lut_translock);
722 INIT_LIST_HEAD(&client_list);
724 * The recovery is not yet finished and final queue can still be updated
725 * with resend requests. Move final list to separate one for processing
727 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
728 list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
729 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
732 * go through list of exports participated in recovery and
733 * set new epoch for them
735 list_for_each_entry(req, &client_list, rq_list) {
736 LASSERT(!req->rq_export->exp_delayed);
737 if (!req->rq_export->exp_vbr_failed)
738 tgt_client_epoch_update(&env, req->rq_export);
740 /** return list back at once */
741 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
742 list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
743 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
745 /** Clear MULTI RPCS incompatibility flag if
746 * - target is MDT and
747 * - there is no client to recover or the recovery was aborted
749 if (!strncmp(tgt->lut_obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) &&
750 (tgt->lut_obd->obd_max_recoverable_clients == 0 ||
751 tgt->lut_obd->obd_abort_recovery))
752 tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
754 /** update server epoch */
755 tgt_server_data_update(&env, tgt, 1);
760 * commit callback, need to update last_committed value
762 struct tgt_last_committed_callback {
763 struct dt_txn_commit_cb llcc_cb;
764 struct lu_target *llcc_tgt;
765 struct obd_export *llcc_exp;
769 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
770 struct dt_txn_commit_cb *cb, int err)
772 struct tgt_last_committed_callback *ccb;
774 ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
776 LASSERT(ccb->llcc_exp);
777 LASSERT(ccb->llcc_tgt != NULL);
778 LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
780 /* error hit, don't update last committed to provide chance to
781 * replay data after fail */
785 /* Fast path w/o spinlock, if exp_last_committed was updated
786 * with higher transno, no need to take spinlock and check,
787 * also no need to update obd_last_committed. */
788 if (ccb->llcc_transno <= ccb->llcc_exp->exp_last_committed)
790 spin_lock(&ccb->llcc_tgt->lut_translock);
791 if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
792 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
794 if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
795 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
796 spin_unlock(&ccb->llcc_tgt->lut_translock);
798 ptlrpc_commit_replies(ccb->llcc_exp);
799 tgt_cancel_slc_locks(ccb->llcc_tgt, ccb->llcc_transno);
801 spin_unlock(&ccb->llcc_tgt->lut_translock);
804 class_export_cb_put(ccb->llcc_exp);
805 if (ccb->llcc_transno)
806 CDEBUG(D_HA, "%s: transno %lld is committed\n",
807 ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
811 int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
812 struct obd_export *exp, __u64 transno)
814 struct tgt_last_committed_callback *ccb;
815 struct dt_txn_commit_cb *dcb;
823 ccb->llcc_exp = class_export_cb_get(exp);
824 ccb->llcc_transno = transno;
827 dcb->dcb_func = tgt_cb_last_committed;
828 INIT_LIST_HEAD(&dcb->dcb_linkage);
829 strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
831 rc = dt_trans_cb_add(th, dcb);
833 class_export_cb_put(exp);
837 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
838 /* report failure to force synchronous operation */
844 struct tgt_new_client_callback {
845 struct dt_txn_commit_cb lncc_cb;
846 struct obd_export *lncc_exp;
849 static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
850 struct dt_txn_commit_cb *cb, int err)
852 struct tgt_new_client_callback *ccb;
854 ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
856 LASSERT(ccb->lncc_exp->exp_obd);
858 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
859 ccb->lncc_exp->exp_obd->obd_name,
860 ccb->lncc_exp->exp_client_uuid.uuid);
862 spin_lock(&ccb->lncc_exp->exp_lock);
864 ccb->lncc_exp->exp_need_sync = 0;
866 spin_unlock(&ccb->lncc_exp->exp_lock);
867 class_export_cb_put(ccb->lncc_exp);
872 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
874 struct tgt_new_client_callback *ccb;
875 struct dt_txn_commit_cb *dcb;
882 ccb->lncc_exp = class_export_cb_get(exp);
885 dcb->dcb_func = tgt_cb_new_client;
886 INIT_LIST_HEAD(&dcb->dcb_linkage);
887 strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
889 rc = dt_trans_cb_add(th, dcb);
891 class_export_cb_put(exp);
898 * Add new client to the last_rcvd upon new connection.
900 * We use a bitmap to locate a free space in the last_rcvd file and initialize
903 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
905 struct tg_export_data *ted = &exp->exp_target_data;
906 struct lu_target *tgt = class_exp2tgt(exp);
911 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
912 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
915 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
918 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
919 * there's no need for extra complication here
921 idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
923 if (idx >= LR_MAX_CLIENTS ||
924 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
925 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
926 tgt->lut_obd->obd_name, idx);
929 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
930 idx = find_next_zero_bit(tgt->lut_client_bitmap,
931 LR_MAX_CLIENTS, idx);
935 ted->ted_lr_idx = idx;
936 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
937 idx * tgt->lut_lsd.lsd_client_size;
939 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
941 if (tgt_is_multimodrpcs_client(exp)) {
942 /* Set MULTI RPCS incompatibility flag to prevent previous
943 * Lustre versions to mount a target with reply_data file */
944 atomic_inc(&tgt->lut_num_clients);
945 if (!(tgt->lut_lsd.lsd_feature_incompat &
946 OBD_INCOMPAT_MULTI_RPCS)) {
947 tgt->lut_lsd.lsd_feature_incompat |=
948 OBD_INCOMPAT_MULTI_RPCS;
949 rc = tgt_server_data_update(env, tgt, 1);
951 CERROR("%s: unable to set MULTI RPCS "
952 "incompatibility flag\n",
953 exp->exp_obd->obd_name);
958 /* assign client slot generation */
959 ted->ted_lcd->lcd_generation =
960 atomic_inc_return(&tgt->lut_client_generation);
962 ted->ted_lcd->lcd_generation = 0;
965 CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s' "
967 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
968 ted->ted_lcd->lcd_uuid, ted->ted_lcd->lcd_generation);
970 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
973 rc = tgt_client_data_update(env, exp);
975 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
976 tgt->lut_obd->obd_name, idx, rc);
980 EXPORT_SYMBOL(tgt_client_new);
982 /* Add an existing client to the MDS in-memory state based on
983 * a client that was previously found in the last_rcvd file and
984 * already has an assigned slot (idx >= 0).
986 * It should not be possible to fail adding an existing client - otherwise
987 * mdt_init_server_data() callsite needs to be fixed.
989 int tgt_client_add(const struct lu_env *env, struct obd_export *exp, int idx)
991 struct tg_export_data *ted = &exp->exp_target_data;
992 struct lu_target *tgt = class_exp2tgt(exp);
996 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
997 LASSERTF(idx >= 0, "%d\n", idx);
999 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
1000 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1003 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
1004 CERROR("%s: client %d: bit already set in bitmap!!\n",
1005 tgt->lut_obd->obd_name, idx);
1008 atomic_inc(&tgt->lut_num_clients);
1010 CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added, "
1012 tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid,
1013 ted->ted_lcd->lcd_generation);
1015 ted->ted_lr_idx = idx;
1016 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
1017 idx * tgt->lut_lsd.lsd_client_size;
1019 mutex_init(&ted->ted_lcd_lock);
1021 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1026 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
1028 struct tg_export_data *ted = &exp->exp_target_data;
1029 struct lu_target *tgt = class_exp2tgt(exp);
1034 LASSERT(ted->ted_lcd);
1036 if (unlikely(tgt == NULL)) {
1037 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1038 class_exp2obd(exp)->obd_name);
1042 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
1043 if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
1044 (char *)tgt->lut_obd->obd_uuid.uuid) ||
1045 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1048 /* Slot may be not yet assigned, use case is race between Client
1049 * reconnect and forced eviction */
1050 if (ted->ted_lr_idx < 0) {
1051 CWARN("%s: client with UUID '%s' not in bitmap\n",
1052 tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid);
1056 CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
1057 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1058 ted->ted_lcd->lcd_uuid);
1060 /* Clear the bit _after_ zeroing out the client so we don't
1061 race with filter_client_add and zero out new clients.*/
1062 if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
1063 CERROR("%s: client %u: bit already clear in bitmap!!\n",
1064 tgt->lut_obd->obd_name, ted->ted_lr_idx);
1068 /* Do not erase record for recoverable client. */
1069 if (exp->exp_flags & OBD_OPT_FAILOVER)
1072 /* Make sure the server's last_transno is up to date.
1073 * This should be done before zeroing client slot so last_transno will
1074 * be in server data or in client data in case of failure */
1075 rc = tgt_server_data_update(env, tgt, 0);
1077 CERROR("%s: failed to update server data, skip client %s "
1078 "zeroing, rc %d\n", tgt->lut_obd->obd_name,
1079 ted->ted_lcd->lcd_uuid, rc);
1083 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
1084 rc = tgt_client_data_update(env, exp);
1086 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
1087 "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
1088 tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
1089 ted->ted_lr_idx, ted->ted_lr_off, rc);
1092 EXPORT_SYMBOL(tgt_client_del);
1094 int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
1095 struct tg_export_data *ted, struct tg_reply_data *trd,
1096 struct thandle *th, bool update_lrd_file)
1098 struct lsd_reply_data *lrd;
1101 lrd = &trd->trd_reply;
1102 /* update export last transno */
1103 mutex_lock(&ted->ted_lcd_lock);
1104 if (lrd->lrd_transno > ted->ted_lcd->lcd_last_transno)
1105 ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
1106 mutex_unlock(&ted->ted_lcd_lock);
1108 /* find a empty slot */
1109 i = tgt_find_free_reply_slot(tgt);
1110 if (unlikely(i < 0)) {
1111 CERROR("%s: couldn't find a slot for reply data: "
1112 "rc = %d\n", tgt_name(tgt), i);
1117 if (update_lrd_file) {
1121 /* write reply data to disk */
1122 off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
1123 rc = tgt_reply_data_write(env, tgt, lrd, off, th);
1124 if (unlikely(rc != 0)) {
1125 CERROR("%s: can't update %s file: rc = %d\n",
1126 tgt_name(tgt), REPLY_DATA, rc);
1130 /* add reply data to target export's reply list */
1131 mutex_lock(&ted->ted_lcd_lock);
1132 list_add(&trd->trd_list, &ted->ted_reply_list);
1133 ted->ted_reply_cnt++;
1134 if (ted->ted_reply_cnt > ted->ted_reply_max)
1135 ted->ted_reply_max = ted->ted_reply_cnt;
1136 mutex_unlock(&ted->ted_lcd_lock);
1138 CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
1139 "tag %hu, client gen %u, slot idx %d\n",
1140 trd, lrd->lrd_xid, lrd->lrd_transno,
1141 trd->trd_tag, lrd->lrd_client_gen, i);
1144 EXPORT_SYMBOL(tgt_add_reply_data);
1147 * last_rcvd & last_committed update callbacks
1149 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
1150 struct dt_object *obj, __u64 opdata,
1151 struct thandle *th, struct ptlrpc_request *req)
1153 struct tgt_thread_info *tti = tgt_th_info(env);
1154 struct tgt_session_info *tsi = tgt_ses_info(env);
1155 struct obd_export *exp = tsi->tsi_exp;
1156 struct tg_export_data *ted;
1164 LASSERT(exp != NULL);
1165 ted = &exp->exp_target_data;
1167 lw_client = exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT;
1168 if (ted->ted_lr_idx < 0 && !lw_client)
1169 /* ofd connect may cause transaction before export has
1174 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
1176 /* From update replay, tti_transno should be set already */
1177 LASSERT(tti->tti_transno != 0);
1179 spin_lock(&tgt->lut_translock);
1180 if (th->th_result != 0) {
1181 if (tti->tti_transno != 0) {
1182 CERROR("%s: replay transno %llu failed: rc = %d\n",
1183 tgt_name(tgt), tti->tti_transno, th->th_result);
1185 } else if (tti->tti_transno == 0) {
1186 tti->tti_transno = ++tgt->lut_last_transno;
1188 /* should be replay */
1189 if (tti->tti_transno > tgt->lut_last_transno)
1190 tgt->lut_last_transno = tti->tti_transno;
1192 spin_unlock(&tgt->lut_translock);
1194 /** VBR: set new versions */
1195 if (th->th_result == 0 && obj != NULL) {
1196 struct dt_object *dto = dt_object_locate(obj, th->th_dev);
1197 dt_version_set(env, dto, tti->tti_transno, th);
1200 /* filling reply data */
1201 CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
1202 tti->tti_transno, tgt->lut_obd->obd_last_committed);
1205 req->rq_transno = tti->tti_transno;
1206 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
1209 /* if can't add callback, do sync write */
1210 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp, tti->tti_transno);
1213 /* All operations performed by LW clients are synchronous and
1214 * we store the committed transno in the last_rcvd header */
1215 spin_lock(&tgt->lut_translock);
1216 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
1217 tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
1218 spin_unlock(&tgt->lut_translock);
1219 /* Although lightweight (LW) connections have no slot
1220 * in the last_rcvd, we still want to maintain
1221 * the in-memory lsd_client_data structure in order to
1222 * properly handle reply reconstruction. */
1223 rc = tgt_server_data_write(env, tgt, th);
1225 spin_unlock(&tgt->lut_translock);
1227 } else if (ted->ted_lr_off == 0) {
1228 CERROR("%s: client idx %d has offset %lld\n",
1229 tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
1233 /* Target that supports multiple reply data */
1234 if (tgt_is_multimodrpcs_client(exp)) {
1235 struct tg_reply_data *trd;
1236 struct lsd_reply_data *lrd;
1237 __u64 *pre_versions;
1241 if (unlikely(trd == NULL))
1244 /* fill reply data information */
1245 lrd = &trd->trd_reply;
1246 lrd->lrd_transno = tti->tti_transno;
1248 lrd->lrd_xid = req->rq_xid;
1249 trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
1250 pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1251 lrd->lrd_result = th->th_result;
1252 lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
1253 write_update = true;
1255 LASSERT(tsi->tsi_xid != 0);
1256 lrd->lrd_xid = tsi->tsi_xid;
1257 lrd->lrd_result = tsi->tsi_result;
1258 lrd->lrd_client_gen = tsi->tsi_client_gen;
1260 pre_versions = NULL;
1261 write_update = false;
1264 lrd->lrd_data = opdata;
1266 trd->trd_pre_versions[0] = pre_versions[0];
1267 trd->trd_pre_versions[1] = pre_versions[1];
1268 trd->trd_pre_versions[2] = pre_versions[2];
1269 trd->trd_pre_versions[3] = pre_versions[3];
1272 rc = tgt_add_reply_data(env, tgt, ted, trd, th, write_update);
1278 /* Enough for update replay, let's return */
1282 mutex_lock(&ted->ted_lcd_lock);
1283 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1284 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
1285 transno_p = &ted->ted_lcd->lcd_last_close_transno;
1286 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
1287 ted->ted_lcd->lcd_last_close_result = th->th_result;
1289 /* VBR: save versions in last_rcvd for reconstruct. */
1290 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1293 ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
1294 ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
1295 ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
1296 ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
1298 transno_p = &ted->ted_lcd->lcd_last_transno;
1299 ted->ted_lcd->lcd_last_xid = req->rq_xid;
1300 ted->ted_lcd->lcd_last_result = th->th_result;
1301 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
1302 * see struct ldlm_reply->lock_policy_res1; */
1303 ted->ted_lcd->lcd_last_data = opdata;
1306 /* Update transno in slot only if non-zero number, i.e. no errors */
1307 if (likely(tti->tti_transno != 0)) {
1308 /* Don't overwrite bigger transaction number with lower one.
1309 * That is not sign of problem in all cases, but in any case
1310 * this value should be monotonically increased only. */
1311 if (*transno_p > tti->tti_transno) {
1312 if (!tgt->lut_no_reconstruct) {
1313 CERROR("%s: trying to overwrite bigger transno:"
1314 "on-disk: %llu, new: %llu replay: "
1315 "%d. See LU-617.\n", tgt_name(tgt),
1316 *transno_p, tti->tti_transno,
1317 req_is_replay(req));
1318 if (req_is_replay(req)) {
1319 spin_lock(&req->rq_export->exp_lock);
1320 req->rq_export->exp_vbr_failed = 1;
1321 spin_unlock(&req->rq_export->exp_lock);
1323 mutex_unlock(&ted->ted_lcd_lock);
1324 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
1327 *transno_p = tti->tti_transno;
1332 tti->tti_off = ted->ted_lr_off;
1333 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1335 mutex_unlock(&ted->ted_lcd_lock);
1339 mutex_unlock(&ted->ted_lcd_lock);
1344 * last_rcvd update for echo client simulation.
1345 * It updates last_rcvd client slot and version of object in
1346 * simple way but with all locks to simulate all drawbacks
1348 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
1349 struct lu_target *tgt,
1350 struct dt_object *obj,
1352 struct obd_export *exp)
1354 struct tgt_thread_info *tti = tgt_th_info(env);
1355 struct tg_export_data *ted = &exp->exp_target_data;
1360 tti->tti_transno = 0;
1362 spin_lock(&tgt->lut_translock);
1363 if (th->th_result == 0)
1364 tti->tti_transno = ++tgt->lut_last_transno;
1365 spin_unlock(&tgt->lut_translock);
1367 /** VBR: set new versions */
1368 if (th->th_result == 0 && obj != NULL)
1369 dt_version_set(env, obj, tti->tti_transno, th);
1371 /* if can't add callback, do sync write */
1372 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
1375 LASSERT(ted->ted_lr_off > 0);
1377 mutex_lock(&ted->ted_lcd_lock);
1378 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1379 ted->ted_lcd->lcd_last_transno = tti->tti_transno;
1380 ted->ted_lcd->lcd_last_result = th->th_result;
1382 tti->tti_off = ted->ted_lr_off;
1383 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1384 mutex_unlock(&ted->ted_lcd_lock);
1388 static int tgt_clients_data_init(const struct lu_env *env,
1389 struct lu_target *tgt,
1390 unsigned long last_size)
1392 struct obd_device *obd = tgt->lut_obd;
1393 struct lr_server_data *lsd = &tgt->lut_lsd;
1394 struct lsd_client_data *lcd = NULL;
1395 struct tg_export_data *ted;
1398 loff_t off = lsd->lsd_client_start;
1399 __u32 generation = 0;
1400 struct cfs_hash *hash = NULL;
1404 CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
1405 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
1411 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1413 GOTO(err_out, rc = -ENODEV);
1415 for (cl_idx = 0; off < last_size; cl_idx++) {
1416 struct obd_export *exp;
1419 /* Don't assume off is incremented properly by
1420 * read_record(), in case sizeof(*lcd)
1421 * isn't the same as fsd->lsd_client_size. */
1422 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
1423 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
1425 CERROR("%s: error reading last_rcvd %s idx %d off "
1426 "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
1429 break; /* read error shouldn't cause startup to fail */
1432 if (lcd->lcd_uuid[0] == '\0') {
1433 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
1438 last_transno = lcd_last_transno(lcd);
1440 /* These exports are cleaned up by disconnect, so they
1441 * need to be set up like real exports as connect does.
1443 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: %llu"
1444 " srv lr: %llu lx: %llu gen %u\n", lcd->lcd_uuid,
1445 cl_idx, last_transno, lsd->lsd_last_transno,
1446 lcd_last_xid(lcd), lcd->lcd_generation);
1448 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
1450 if (PTR_ERR(exp) == -EALREADY) {
1451 /* export already exists, zero out this one */
1452 CERROR("%s: Duplicate export %s!\n",
1453 tgt_name(tgt), lcd->lcd_uuid);
1456 GOTO(err_out, rc = PTR_ERR(exp));
1459 ted = &exp->exp_target_data;
1460 *ted->ted_lcd = *lcd;
1462 rc = tgt_client_add(env, exp, cl_idx);
1463 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
1464 /* VBR: set export last committed version */
1465 exp->exp_last_committed = last_transno;
1466 spin_lock(&exp->exp_lock);
1467 exp->exp_connecting = 0;
1468 exp->exp_in_recovery = 0;
1469 spin_unlock(&exp->exp_lock);
1470 obd->obd_max_recoverable_clients++;
1472 if (tgt->lut_lsd.lsd_feature_incompat &
1473 OBD_INCOMPAT_MULTI_RPCS &&
1474 lcd->lcd_generation != 0) {
1475 /* compute the highest valid client generation */
1476 generation = max(generation, lcd->lcd_generation);
1477 /* fill client_generation <-> export hash table */
1478 rc = cfs_hash_add_unique(hash, &lcd->lcd_generation,
1479 &exp->exp_gen_hash);
1481 CERROR("%s: duplicate export for client "
1483 tgt_name(tgt), lcd->lcd_generation);
1484 class_export_put(exp);
1489 class_export_put(exp);
1491 rc = rev_import_init(exp);
1493 class_unlink_export(exp);
1497 /* Need to check last_rcvd even for duplicated exports. */
1498 CDEBUG(D_OTHER, "client at idx %d has last_transno = %llu\n",
1499 cl_idx, last_transno);
1501 spin_lock(&tgt->lut_translock);
1502 tgt->lut_last_transno = max(last_transno,
1503 tgt->lut_last_transno);
1504 spin_unlock(&tgt->lut_translock);
1507 /* record highest valid client generation */
1508 atomic_set(&tgt->lut_client_generation, generation);
1512 cfs_hash_putref(hash);
1517 struct server_compat_data {
1524 static struct server_compat_data tgt_scd[] = {
1525 [LDD_F_SV_TYPE_MDT] = {
1526 .rocompat = OBD_ROCOMPAT_LOVOBJID,
1527 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1528 OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
1529 OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI |
1530 OBD_INCOMPAT_MULTI_RPCS,
1531 .rocinit = OBD_ROCOMPAT_LOVOBJID,
1532 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1533 OBD_INCOMPAT_MULTI_OI,
1535 [LDD_F_SV_TYPE_OST] = {
1536 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
1537 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
1539 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
1540 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1544 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1546 struct tgt_thread_info *tti = tgt_th_info(env);
1547 struct lr_server_data *lsd = &tgt->lut_lsd;
1548 unsigned long last_rcvd_size;
1552 rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1556 last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1558 /* ensure padding in the struct is the correct size */
1559 CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
1560 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
1562 rc = server_name2index(tgt_name(tgt), &index, NULL);
1564 CERROR("%s: Can not get index from name: rc = %d\n",
1568 /* server_name2index() returns type */
1570 if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1571 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1575 /* last_rcvd on OST doesn't provide reconstruct support because there
1576 * may be up to 8 in-flight write requests per single slot in
1577 * last_rcvd client data
1579 tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1581 if (last_rcvd_size == 0) {
1582 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1584 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1585 sizeof(lsd->lsd_uuid));
1586 lsd->lsd_last_transno = 0;
1587 lsd->lsd_mount_count = 0;
1588 lsd->lsd_server_size = LR_SERVER_SIZE;
1589 lsd->lsd_client_start = LR_CLIENT_START;
1590 lsd->lsd_client_size = LR_CLIENT_SIZE;
1591 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1592 lsd->lsd_osd_index = index;
1593 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1594 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1596 rc = tgt_server_data_read(env, tgt);
1598 CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1602 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1603 LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
1604 "using the wrong disk %s. Were the"
1605 " /dev/ assignments rearranged?\n",
1606 tgt->lut_obd->obd_uuid.uuid,
1611 if (lsd->lsd_osd_index != index) {
1612 LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
1613 "is different with the index %d in"
1614 "config log, It might be disk"
1615 "corruption!\n", tgt_name(tgt),
1616 lsd->lsd_osd_index, index);
1621 if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1622 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1624 lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1628 if (type == LDD_F_SV_TYPE_MDT)
1629 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1631 if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1632 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1634 lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1637 /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1638 if (type == LDD_F_SV_TYPE_MDT &&
1639 !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1640 if (last_rcvd_size > lsd->lsd_client_start) {
1641 LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1642 "remove all clients for interop needs\n",
1644 rc = tgt_truncate_last_rcvd(env, tgt,
1645 lsd->lsd_client_start);
1648 last_rcvd_size = lsd->lsd_client_start;
1650 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1651 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1654 spin_lock(&tgt->lut_translock);
1655 tgt->lut_last_transno = lsd->lsd_last_transno;
1656 spin_unlock(&tgt->lut_translock);
1658 lsd->lsd_mount_count++;
1660 CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1661 CDEBUG(D_INODE, "%s: server last_transno: %llu\n",
1662 tgt_name(tgt), tgt->lut_last_transno);
1663 CDEBUG(D_INODE, "%s: server mount_count: %llu\n",
1664 tgt_name(tgt), lsd->lsd_mount_count);
1665 CDEBUG(D_INODE, "%s: server data size: %u\n",
1666 tgt_name(tgt), lsd->lsd_server_size);
1667 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1668 tgt_name(tgt), lsd->lsd_client_start);
1669 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1670 tgt_name(tgt), lsd->lsd_client_size);
1671 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1672 tgt_name(tgt), last_rcvd_size);
1673 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1674 tgt_name(tgt), lsd->lsd_subdir_count);
1675 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1676 last_rcvd_size <= lsd->lsd_client_start ? 0 :
1677 (last_rcvd_size - lsd->lsd_client_start) /
1678 lsd->lsd_client_size);
1679 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1681 if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1682 lsd->lsd_client_size == 0) {
1683 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1687 if (!tgt->lut_obd->obd_replayable)
1688 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1690 rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1692 GOTO(err_client, rc);
1694 spin_lock(&tgt->lut_translock);
1695 /* obd_last_committed is used for compatibility
1696 * with other lustre recovery code */
1697 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1698 spin_unlock(&tgt->lut_translock);
1700 tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1701 tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1703 /* save it, so mount count and last_transno is current */
1704 rc = tgt_server_data_update(env, tgt, 0);
1706 GOTO(err_client, rc);
1711 class_disconnect_exports(tgt->lut_obd);
1715 /* add credits for last_rcvd update */
1716 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1719 struct lu_target *tgt = cookie;
1720 struct tgt_session_info *tsi;
1721 struct tgt_thread_info *tti = tgt_th_info(env);
1722 struct dt_object *dto;
1725 /* if there is no session, then this transaction is not result of
1726 * request processing but some local operation */
1727 if (env->le_ses == NULL)
1730 LASSERT(tgt->lut_last_rcvd);
1731 tsi = tgt_ses_info(env);
1732 /* OFD may start transaction without export assigned */
1733 if (tsi->tsi_exp == NULL)
1736 if (tgt_is_multimodrpcs_client(tsi->tsi_exp)) {
1738 * Use maximum possible file offset for declaration to ensure
1739 * ZFS will reserve enough credits for a write anywhere in this
1740 * file, since we don't know where in the file the write will be
1741 * because a replay slot has not been assigned. This should be
1742 * replaced by dmu_tx_hold_append() when available.
1744 tti->tti_buf.lb_buf = NULL;
1745 tti->tti_buf.lb_len = sizeof(struct lsd_reply_data);
1746 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
1747 rc = dt_declare_record_write(env, dto, &tti->tti_buf, -1, th);
1751 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
1753 tti->tti_off = tsi->tsi_exp->exp_target_data.ted_lr_off;
1754 rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1760 if (tsi->tsi_vbr_obj != NULL &&
1761 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1762 dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
1763 rc = dt_declare_version_set(env, dto, th);
1769 /* Update last_rcvd records with latests transaction data */
1770 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1773 struct lu_target *tgt = cookie;
1774 struct tgt_session_info *tsi;
1775 struct tgt_thread_info *tti = tgt_th_info(env);
1776 struct dt_object *obj = NULL;
1780 if (env->le_ses == NULL)
1783 tsi = tgt_ses_info(env);
1784 /* OFD may start transaction without export assigned */
1785 if (tsi->tsi_exp == NULL)
1788 echo_client = (tgt_ses_req(tsi) == NULL && tsi->tsi_xid == 0);
1790 if (tti->tti_has_trans && !echo_client) {
1791 if (tti->tti_mult_trans == 0) {
1792 CDEBUG(D_HA, "More than one transaction %llu\n",
1796 /* we need another transno to be assigned */
1797 tti->tti_transno = 0;
1798 } else if (th->th_result == 0) {
1799 tti->tti_has_trans = 1;
1802 if (tsi->tsi_vbr_obj != NULL &&
1803 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1804 obj = tsi->tsi_vbr_obj;
1807 if (unlikely(echo_client)) /* echo client special case */
1808 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1811 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
1816 int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
1818 struct tgt_thread_info *tti = tgt_th_info(env);
1819 struct lsd_reply_data *lrd = &tti->tti_lrd;
1820 unsigned long reply_data_size;
1822 struct lsd_reply_header *lrh = NULL;
1823 struct lsd_client_data *lcd = NULL;
1824 struct tg_reply_data *trd = NULL;
1827 struct cfs_hash *hash = NULL;
1828 struct obd_export *exp;
1829 struct tg_export_data *ted;
1830 int reply_data_recovered = 0;
1832 rc = dt_attr_get(env, tgt->lut_reply_data, &tti->tti_attr);
1835 reply_data_size = (unsigned long)tti->tti_attr.la_size;
1839 GOTO(out, rc = -ENOMEM);
1841 if (reply_data_size == 0) {
1842 CDEBUG(D_INFO, "%s: new reply_data file, initializing\n",
1844 lrh->lrh_magic = LRH_MAGIC;
1845 lrh->lrh_header_size = sizeof(struct lsd_reply_header);
1846 lrh->lrh_reply_size = sizeof(struct lsd_reply_data);
1847 rc = tgt_reply_header_write(env, tgt, lrh);
1849 CERROR("%s: error writing %s: rc = %d\n",
1850 tgt_name(tgt), REPLY_DATA, rc);
1854 rc = tgt_reply_header_read(env, tgt, lrh);
1856 CERROR("%s: error reading %s: rc = %d\n",
1857 tgt_name(tgt), REPLY_DATA, rc);
1860 if (lrh->lrh_magic != LRH_MAGIC ||
1861 lrh->lrh_header_size != sizeof(struct lsd_reply_header) ||
1862 lrh->lrh_reply_size != sizeof(struct lsd_reply_data)) {
1863 CERROR("%s: invalid header in %s\n",
1864 tgt_name(tgt), REPLY_DATA);
1865 GOTO(out, rc = -EINVAL);
1868 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1870 GOTO(out, rc = -ENODEV);
1874 GOTO(out, rc = -ENOMEM);
1878 GOTO(out, rc = -ENOMEM);
1880 /* Load reply_data from disk */
1881 for (idx = 0, off = sizeof(struct lsd_reply_header);
1882 off < reply_data_size;
1883 idx++, off += sizeof(struct lsd_reply_data)) {
1884 rc = tgt_reply_data_read(env, tgt, lrd, off);
1886 CERROR("%s: error reading %s: rc = %d\n",
1887 tgt_name(tgt), REPLY_DATA, rc);
1891 exp = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
1893 /* old reply data from a disconnected client */
1896 ted = &exp->exp_target_data;
1897 mutex_lock(&ted->ted_lcd_lock);
1899 /* create in-memory reply_data and link it to
1900 * target export's reply list */
1901 rc = tgt_set_reply_slot(tgt, idx);
1903 mutex_unlock(&ted->ted_lcd_lock);
1906 trd->trd_reply = *lrd;
1907 trd->trd_pre_versions[0] = 0;
1908 trd->trd_pre_versions[1] = 0;
1909 trd->trd_pre_versions[2] = 0;
1910 trd->trd_pre_versions[3] = 0;
1911 trd->trd_index = idx;
1913 list_add(&trd->trd_list, &ted->ted_reply_list);
1914 ted->ted_reply_cnt++;
1915 if (ted->ted_reply_cnt > ted->ted_reply_max)
1916 ted->ted_reply_max = ted->ted_reply_cnt;
1918 CDEBUG(D_HA, "%s: restore reply %p: xid %llu, "
1919 "transno %llu, client gen %u, slot idx %d\n",
1920 tgt_name(tgt), trd, lrd->lrd_xid,
1921 lrd->lrd_transno, lrd->lrd_client_gen,
1924 /* update export last committed transation */
1925 exp->exp_last_committed = max(exp->exp_last_committed,
1928 mutex_unlock(&ted->ted_lcd_lock);
1929 class_export_put(exp);
1931 /* update target last committed transaction */
1932 spin_lock(&tgt->lut_translock);
1933 tgt->lut_last_transno = max(tgt->lut_last_transno,
1935 spin_unlock(&tgt->lut_translock);
1937 reply_data_recovered++;
1941 GOTO(out, rc = -ENOMEM);
1943 CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
1944 tgt_name(tgt), reply_data_recovered);
1947 spin_lock(&tgt->lut_translock);
1948 /* obd_last_committed is used for compatibility
1949 * with other lustre recovery code */
1950 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1951 spin_unlock(&tgt->lut_translock);
1957 cfs_hash_putref(hash);
1967 struct tg_reply_data *tgt_lookup_reply_by_xid(struct tg_export_data *ted,
1970 struct tg_reply_data *found = NULL;
1971 struct tg_reply_data *reply;
1973 mutex_lock(&ted->ted_lcd_lock);
1974 list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
1975 if (reply->trd_reply.lrd_xid == xid) {
1980 mutex_unlock(&ted->ted_lcd_lock);
1983 EXPORT_SYMBOL(tgt_lookup_reply_by_xid);
1985 /* Look for a reply data matching specified request @req
1986 * A copy is returned in @trd if the pointer is not NULL
1988 bool tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
1990 struct tg_export_data *ted = &req->rq_export->exp_target_data;
1991 struct tg_reply_data *reply;
1994 reply = tgt_lookup_reply_by_xid(ted, req->rq_xid);
1995 if (reply != NULL) {
2001 CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d\n",
2002 tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid,
2007 EXPORT_SYMBOL(tgt_lookup_reply);
2009 int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
2011 struct tg_export_data *ted = &exp->exp_target_data;
2012 struct lu_target *lut = class_exp2tgt(exp);
2013 struct tg_reply_data *trd, *tmp;
2015 mutex_lock(&ted->ted_lcd_lock);
2016 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
2017 if (trd->trd_reply.lrd_xid > rcvd_xid)
2019 ted->ted_release_xid++;
2020 tgt_release_reply_data(lut, ted, trd);
2022 mutex_unlock(&ted->ted_lcd_lock);
2027 int tgt_handle_tag(struct obd_export *exp, __u16 tag)
2029 struct tg_export_data *ted = &exp->exp_target_data;
2030 struct lu_target *lut = class_exp2tgt(exp);
2031 struct tg_reply_data *trd, *tmp;
2036 mutex_lock(&ted->ted_lcd_lock);
2037 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
2038 if (trd->trd_tag != tag)
2040 ted->ted_release_tag++;
2041 tgt_release_reply_data(lut, ted, trd);
2044 mutex_unlock(&ted->ted_lcd_lock);