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, 2017, 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"
43 /** version recovery epoch */
44 #define LR_EPOCH_BITS 32
46 /* Allocate a bitmap for a chunk of reply data slots */
47 static int tgt_bitmap_chunk_alloc(struct lu_target *lut, int chunk)
51 OBD_ALLOC_LARGE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
56 spin_lock(&lut->lut_client_bitmap_lock);
58 if (lut->lut_reply_bitmap[chunk] != NULL) {
59 /* someone else already allocated the bitmap for this chunk */
60 spin_unlock(&lut->lut_client_bitmap_lock);
61 OBD_FREE_LARGE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
66 lut->lut_reply_bitmap[chunk] = bm;
68 spin_unlock(&lut->lut_client_bitmap_lock);
73 /* Look for an available reply data slot in the bitmap
75 * Allocate bitmap chunk when first used
76 * XXX algo could be improved if this routine limits performance
78 static int tgt_find_free_reply_slot(struct lu_target *lut)
85 for (chunk = 0; chunk < LUT_REPLY_SLOTS_MAX_CHUNKS; chunk++) {
86 /* allocate the bitmap chunk if necessary */
87 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
88 rc = tgt_bitmap_chunk_alloc(lut, chunk);
92 bmp = lut->lut_reply_bitmap[chunk];
94 /* look for an available slot in this chunk */
96 b = find_first_zero_bit(bmp, LUT_REPLY_SLOTS_PER_CHUNK);
97 if (b >= LUT_REPLY_SLOTS_PER_CHUNK)
101 if (test_and_set_bit(b, bmp) == 0)
102 return chunk * LUT_REPLY_SLOTS_PER_CHUNK + b;
109 /* Mark the reply data slot @idx 'used' in the corresponding bitmap chunk
111 * Allocate the bitmap chunk if necessary
113 static int tgt_set_reply_slot(struct lu_target *lut, int idx)
119 chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
120 b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
122 LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
123 LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
125 /* allocate the bitmap chunk if necessary */
126 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
127 rc = tgt_bitmap_chunk_alloc(lut, chunk);
132 /* mark the slot 'used' in this chunk */
133 if (test_and_set_bit(b, lut->lut_reply_bitmap[chunk]) != 0) {
134 CERROR("%s: slot %d already set in bitmap\n",
143 /* Mark the reply data slot @idx 'unused' in the corresponding bitmap chunk
146 static int tgt_clear_reply_slot(struct lu_target *lut, int idx)
151 if (lut->lut_obd->obd_stopping)
153 * in case of failover keep the bit set in order to
154 * avoid overwriting slots in reply_data which might
155 * be required by resent rpcs
158 chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
159 b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
161 LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
162 LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
164 if (lut->lut_reply_bitmap[chunk] == NULL) {
165 CERROR("%s: slot %d not allocated\n",
170 if (test_and_clear_bit(b, lut->lut_reply_bitmap[chunk]) == 0) {
171 CERROR("%s: slot %d already clear in bitmap\n",
180 /* Read header of reply_data file of target @tgt into structure @lrh */
181 static int tgt_reply_header_read(const struct lu_env *env,
182 struct lu_target *tgt,
183 struct lsd_reply_header *lrh)
186 struct lsd_reply_header buf;
187 struct tgt_thread_info *tti = tgt_th_info(env);
190 tti->tti_buf.lb_buf = &buf;
191 tti->tti_buf.lb_len = sizeof(buf);
193 rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
198 lrh->lrh_magic = le32_to_cpu(buf.lrh_magic);
199 lrh->lrh_header_size = le32_to_cpu(buf.lrh_header_size);
200 lrh->lrh_reply_size = le32_to_cpu(buf.lrh_reply_size);
202 CDEBUG(D_HA, "%s: read %s header. magic=0x%08x "
203 "header_size=%d reply_size=%d\n",
204 tgt->lut_obd->obd_name, REPLY_DATA,
205 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
210 /* Write header into replay_data file of target @tgt from structure @lrh */
211 static int tgt_reply_header_write(const struct lu_env *env,
212 struct lu_target *tgt,
213 struct lsd_reply_header *lrh)
216 struct lsd_reply_header buf;
217 struct tgt_thread_info *tti = tgt_th_info(env);
219 struct dt_object *dto;
221 CDEBUG(D_HA, "%s: write %s header. magic=0x%08x "
222 "header_size=%d reply_size=%d\n",
223 tgt->lut_obd->obd_name, REPLY_DATA,
224 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
226 if (tgt->lut_bottom->dd_rdonly)
229 buf.lrh_magic = cpu_to_le32(lrh->lrh_magic);
230 buf.lrh_header_size = cpu_to_le32(lrh->lrh_header_size);
231 buf.lrh_reply_size = cpu_to_le32(lrh->lrh_reply_size);
233 th = dt_trans_create(env, tgt->lut_bottom);
239 tti->tti_buf.lb_buf = &buf;
240 tti->tti_buf.lb_len = sizeof(buf);
242 rc = dt_declare_record_write(env, tgt->lut_reply_data,
243 &tti->tti_buf, tti->tti_off, th);
247 rc = dt_trans_start(env, tgt->lut_bottom, th);
251 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
252 rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
254 dt_trans_stop(env, tgt->lut_bottom, th);
258 /* Write the reply data @lrd into reply_data file of target @tgt
261 static int tgt_reply_data_write(const struct lu_env *env, struct lu_target *tgt,
262 struct lsd_reply_data *lrd, loff_t off,
265 struct tgt_thread_info *tti = tgt_th_info(env);
266 struct dt_object *dto;
267 struct lsd_reply_data *buf = &tti->tti_lrd;
269 lrd->lrd_result = ptlrpc_status_hton(lrd->lrd_result);
271 buf->lrd_transno = cpu_to_le64(lrd->lrd_transno);
272 buf->lrd_xid = cpu_to_le64(lrd->lrd_xid);
273 buf->lrd_data = cpu_to_le64(lrd->lrd_data);
274 buf->lrd_result = cpu_to_le32(lrd->lrd_result);
275 buf->lrd_client_gen = cpu_to_le32(lrd->lrd_client_gen);
277 lrd->lrd_result = ptlrpc_status_ntoh(lrd->lrd_result);
280 tti->tti_buf.lb_buf = buf;
281 tti->tti_buf.lb_len = sizeof(*buf);
283 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
284 return dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
287 /* Read the reply data from reply_data file of target @tgt at offset @off
288 * into structure @lrd
290 static int tgt_reply_data_read(const struct lu_env *env, struct lu_target *tgt,
291 struct lsd_reply_data *lrd, loff_t off)
294 struct tgt_thread_info *tti = tgt_th_info(env);
295 struct lsd_reply_data *buf = &tti->tti_lrd;
298 tti->tti_buf.lb_buf = buf;
299 tti->tti_buf.lb_len = sizeof(*buf);
301 rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
306 lrd->lrd_transno = le64_to_cpu(buf->lrd_transno);
307 lrd->lrd_xid = le64_to_cpu(buf->lrd_xid);
308 lrd->lrd_data = le64_to_cpu(buf->lrd_data);
309 lrd->lrd_result = le32_to_cpu(buf->lrd_result);
310 lrd->lrd_client_gen = le32_to_cpu(buf->lrd_client_gen);
316 /* Free the in-memory reply data structure @trd and release
317 * the corresponding slot in the reply_data file of target @lut
318 * Called with ted_lcd_lock held
320 static void tgt_free_reply_data(struct lu_target *lut,
321 struct tg_export_data *ted,
322 struct tg_reply_data *trd)
324 CDEBUG(D_TRACE, "%s: free reply data %p: xid %llu, transno %llu, "
325 "client gen %u, slot idx %d\n",
326 lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
327 trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
330 LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
332 list_del(&trd->trd_list);
333 ted->ted_reply_cnt--;
335 tgt_clear_reply_slot(lut, trd->trd_index);
339 /* Release the reply data @trd from target @lut
340 * The reply data with the highest transno for this export
341 * is retained to ensure correctness of target recovery
342 * Called with ted_lcd_lock held
344 static void tgt_release_reply_data(struct lu_target *lut,
345 struct tg_export_data *ted,
346 struct tg_reply_data *trd)
348 CDEBUG(D_TRACE, "%s: release reply data %p: xid %llu, transno %llu, "
349 "client gen %u, slot idx %d\n",
350 lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
351 trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
354 LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
356 /* Do not free the reply data corresponding to the
357 * highest transno of this export.
358 * This ensures on-disk reply data is kept and
359 * last committed transno can be restored from disk in case
362 if (trd->trd_reply.lrd_transno == ted->ted_lcd->lcd_last_transno) {
363 /* free previous retained reply */
364 if (ted->ted_reply_last != NULL)
365 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
366 /* retain the reply */
367 list_del_init(&trd->trd_list);
368 ted->ted_reply_last = trd;
370 tgt_free_reply_data(lut, ted, trd);
374 static inline struct lu_buf *tti_buf_lsd(struct tgt_thread_info *tti)
376 tti->tti_buf.lb_buf = &tti->tti_lsd;
377 tti->tti_buf.lb_len = sizeof(tti->tti_lsd);
378 return &tti->tti_buf;
381 static inline struct lu_buf *tti_buf_lcd(struct tgt_thread_info *tti)
383 tti->tti_buf.lb_buf = &tti->tti_lcd;
384 tti->tti_buf.lb_len = sizeof(tti->tti_lcd);
385 return &tti->tti_buf;
389 * Allocate in-memory data for client slot related to export.
391 int tgt_client_alloc(struct obd_export *exp)
394 LASSERT(exp != exp->exp_obd->obd_self_export);
396 spin_lock_init(&exp->exp_target_data.ted_nodemap_lock);
397 INIT_LIST_HEAD(&exp->exp_target_data.ted_nodemap_member);
399 OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
400 if (exp->exp_target_data.ted_lcd == NULL)
402 /* Mark that slot is not yet valid, 0 doesn't work here */
403 exp->exp_target_data.ted_lr_idx = -1;
404 INIT_LIST_HEAD(&exp->exp_target_data.ted_reply_list);
405 mutex_init(&exp->exp_target_data.ted_lcd_lock);
408 EXPORT_SYMBOL(tgt_client_alloc);
411 * Free in-memory data for client slot related to export.
413 void tgt_client_free(struct obd_export *exp)
415 struct tg_export_data *ted = &exp->exp_target_data;
416 struct lu_target *lut = class_exp2tgt(exp);
417 struct tg_reply_data *trd, *tmp;
419 LASSERT(exp != exp->exp_obd->obd_self_export);
421 /* free reply data */
422 mutex_lock(&ted->ted_lcd_lock);
423 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
424 tgt_release_reply_data(lut, ted, trd);
426 if (ted->ted_reply_last != NULL) {
427 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
428 ted->ted_reply_last = NULL;
430 mutex_unlock(&ted->ted_lcd_lock);
432 if (!hlist_unhashed(&exp->exp_gen_hash))
433 cfs_hash_del(exp->exp_obd->obd_gen_hash,
434 &ted->ted_lcd->lcd_generation,
437 OBD_FREE_PTR(ted->ted_lcd);
440 /* Target may have been freed (see LU-7430)
441 * Slot may be not yet assigned */
442 if (exp->exp_obd->u.obt.obt_magic != OBT_MAGIC ||
446 /* Clear bit when lcd is freed */
447 LASSERT(lut && lut->lut_client_bitmap);
448 if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
449 CERROR("%s: client %u bit already clear in bitmap\n",
450 exp->exp_obd->obd_name, ted->ted_lr_idx);
454 if (tgt_is_multimodrpcs_client(exp) && !exp->exp_obd->obd_stopping)
455 atomic_dec(&lut->lut_num_clients);
457 EXPORT_SYMBOL(tgt_client_free);
459 static inline void tgt_check_lcd(const char *obd_name, int index,
460 struct lsd_client_data *lcd)
462 size_t uuid_size = sizeof(lcd->lcd_uuid);
464 if (strnlen((char*)lcd->lcd_uuid, uuid_size) == uuid_size) {
465 lcd->lcd_uuid[uuid_size - 1] = '\0';
467 LCONSOLE_ERROR("the client UUID (%s) on %s for exports stored in last_rcvd(index = %d) is bad!\n",
468 lcd->lcd_uuid, obd_name, index);
472 static int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
473 struct lsd_client_data *lcd,
474 loff_t *off, int index)
476 struct tgt_thread_info *tti = tgt_th_info(env);
480 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
482 tgt_check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
483 lcd_le_to_cpu(&tti->tti_lcd, lcd);
484 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
485 lcd->lcd_last_close_result =
486 ptlrpc_status_ntoh(lcd->lcd_last_close_result);
489 CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = %llu"
490 ", last_xid = %llu, last_result = %u, last_data = %u, "
491 "last_close_transno = %llu, last_close_xid = %llu, "
492 "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
493 *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
494 lcd->lcd_last_result, lcd->lcd_last_data,
495 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
496 lcd->lcd_last_close_result, rc);
500 static int tgt_client_data_write(const struct lu_env *env,
501 struct lu_target *tgt,
502 struct lsd_client_data *lcd,
503 loff_t *off, struct thandle *th)
505 struct tgt_thread_info *tti = tgt_th_info(env);
506 struct dt_object *dto;
508 lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
509 lcd->lcd_last_close_result =
510 ptlrpc_status_hton(lcd->lcd_last_close_result);
511 lcd_cpu_to_le(lcd, &tti->tti_lcd);
514 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
515 return dt_record_write(env, dto, &tti->tti_buf, off, th);
518 struct tgt_new_client_callback {
519 struct dt_txn_commit_cb lncc_cb;
520 struct obd_export *lncc_exp;
523 static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
524 struct dt_txn_commit_cb *cb, int err)
526 struct tgt_new_client_callback *ccb;
528 ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
530 LASSERT(ccb->lncc_exp->exp_obd);
532 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
533 ccb->lncc_exp->exp_obd->obd_name,
534 ccb->lncc_exp->exp_client_uuid.uuid);
536 spin_lock(&ccb->lncc_exp->exp_lock);
538 ccb->lncc_exp->exp_need_sync = 0;
540 spin_unlock(&ccb->lncc_exp->exp_lock);
541 class_export_cb_put(ccb->lncc_exp);
546 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
548 struct tgt_new_client_callback *ccb;
549 struct dt_txn_commit_cb *dcb;
556 ccb->lncc_exp = class_export_cb_get(exp);
559 dcb->dcb_func = tgt_cb_new_client;
560 INIT_LIST_HEAD(&dcb->dcb_linkage);
561 strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
563 rc = dt_trans_cb_add(th, dcb);
565 class_export_cb_put(exp);
572 * Update client data in last_rcvd
574 static int tgt_client_data_update(const struct lu_env *env,
575 struct obd_export *exp)
577 struct tg_export_data *ted = &exp->exp_target_data;
578 struct lu_target *tgt = class_exp2tgt(exp);
579 struct tgt_thread_info *tti = tgt_th_info(env);
585 if (unlikely(tgt == NULL)) {
586 CDEBUG(D_ERROR, "%s: No target for connected export\n",
587 class_exp2obd(exp)->obd_name);
591 if (tgt->lut_bottom->dd_rdonly)
594 th = dt_trans_create(env, tgt->lut_bottom);
599 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
601 ted->ted_lr_off, th);
605 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
609 mutex_lock(&ted->ted_lcd_lock);
612 * Until this operations will be committed the sync is needed
613 * for this export. This should be done _after_ starting the
614 * transaction so that many connecting clients will not bring
615 * server down with lots of sync writes.
617 rc = tgt_new_client_cb_add(th, exp);
619 /* can't add callback, do sync now */
622 spin_lock(&exp->exp_lock);
623 exp->exp_need_sync = 1;
624 spin_unlock(&exp->exp_lock);
627 tti->tti_off = ted->ted_lr_off;
628 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
630 mutex_unlock(&ted->ted_lcd_lock);
634 dt_trans_stop(env, tgt->lut_bottom, th);
635 CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
636 "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
637 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
642 static int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
644 struct tgt_thread_info *tti = tgt_th_info(env);
649 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
652 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
654 CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
655 "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
656 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
660 static int tgt_server_data_write(const struct lu_env *env,
661 struct lu_target *tgt, struct thandle *th)
663 struct tgt_thread_info *tti = tgt_th_info(env);
664 struct dt_object *dto;
671 lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
673 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
674 rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
676 CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
677 "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
678 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
684 * Update server data in last_rcvd
686 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
689 struct tgt_thread_info *tti = tgt_th_info(env);
696 "%s: mount_count is %llu, last_transno is %llu\n",
697 tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
698 tgt->lut_last_transno);
700 /* Always save latest transno to keep it fresh */
701 spin_lock(&tgt->lut_translock);
702 tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
703 spin_unlock(&tgt->lut_translock);
705 if (tgt->lut_bottom->dd_rdonly)
708 th = dt_trans_create(env, tgt->lut_bottom);
715 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
716 &tti->tti_buf, tti->tti_off, th);
720 rc = dt_trans_start(env, tgt->lut_bottom, th);
724 rc = tgt_server_data_write(env, tgt, th);
726 dt_trans_stop(env, tgt->lut_bottom, th);
728 CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
729 "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
730 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
733 EXPORT_SYMBOL(tgt_server_data_update);
735 static int tgt_truncate_last_rcvd(const struct lu_env *env,
736 struct lu_target *tgt, loff_t size)
738 struct dt_object *dt = tgt->lut_last_rcvd;
745 if (tgt->lut_bottom->dd_rdonly)
749 attr.la_valid = LA_SIZE;
751 th = dt_trans_create(env, tgt->lut_bottom);
754 rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
757 rc = dt_declare_attr_set(env, dt, &attr, th);
760 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
764 rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
766 rc = dt_attr_set(env, dt, &attr, th);
769 dt_trans_stop(env, tgt->lut_bottom, th);
774 static void tgt_client_epoch_update(const struct lu_env *env,
775 struct obd_export *exp)
777 struct lsd_client_data *lcd = exp->exp_target_data.ted_lcd;
778 struct lu_target *tgt = class_exp2tgt(exp);
780 LASSERT(tgt && tgt->lut_bottom);
781 /** VBR: set client last_epoch to current epoch */
782 if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
784 lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
785 tgt_client_data_update(env, exp);
789 * Update boot epoch when recovery ends
791 void tgt_boot_epoch_update(struct lu_target *tgt)
794 struct ptlrpc_request *req;
796 struct list_head client_list;
799 if (tgt->lut_obd->obd_stopping)
802 rc = lu_env_init(&env, LCT_DT_THREAD);
804 CERROR("%s: can't initialize environment: rc = %d\n",
805 tgt->lut_obd->obd_name, rc);
809 spin_lock(&tgt->lut_translock);
810 start_epoch = (tgt->lut_last_transno >> LR_EPOCH_BITS) + 1;
811 tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
812 tgt->lut_lsd.lsd_start_epoch = start_epoch;
813 spin_unlock(&tgt->lut_translock);
815 INIT_LIST_HEAD(&client_list);
817 * The recovery is not yet finished and final queue can still be updated
818 * with resend requests. Move final list to separate one for processing
820 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
821 list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
822 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
825 * go through list of exports participated in recovery and
826 * set new epoch for them
828 list_for_each_entry(req, &client_list, rq_list) {
829 LASSERT(!req->rq_export->exp_delayed);
830 if (!req->rq_export->exp_vbr_failed)
831 tgt_client_epoch_update(&env, req->rq_export);
833 /** return list back at once */
834 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
835 list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
836 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
838 /** Clear MULTI RPCS incompatibility flag if
839 * - target is MDT and
840 * - there is no client to recover or the recovery was aborted
842 if (!strncmp(tgt->lut_obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) &&
843 (tgt->lut_obd->obd_max_recoverable_clients == 0 ||
844 tgt->lut_obd->obd_abort_recovery))
845 tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
847 /** update server epoch */
848 tgt_server_data_update(&env, tgt, 1);
853 * commit callback, need to update last_committed value
855 struct tgt_last_committed_callback {
856 struct dt_txn_commit_cb llcc_cb;
857 struct lu_target *llcc_tgt;
858 struct obd_export *llcc_exp;
862 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
863 struct dt_txn_commit_cb *cb, int err)
865 struct tgt_last_committed_callback *ccb;
867 ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
869 LASSERT(ccb->llcc_exp);
870 LASSERT(ccb->llcc_tgt != NULL);
871 LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
873 /* error hit, don't update last committed to provide chance to
874 * replay data after fail */
878 /* Fast path w/o spinlock, if exp_last_committed was updated
879 * with higher transno, no need to take spinlock and check,
880 * also no need to update obd_last_committed. */
881 if (ccb->llcc_transno <= ccb->llcc_exp->exp_last_committed)
883 spin_lock(&ccb->llcc_tgt->lut_translock);
884 if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
885 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
887 if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
888 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
889 spin_unlock(&ccb->llcc_tgt->lut_translock);
891 ptlrpc_commit_replies(ccb->llcc_exp);
892 tgt_cancel_slc_locks(ccb->llcc_tgt, ccb->llcc_transno);
894 spin_unlock(&ccb->llcc_tgt->lut_translock);
897 CDEBUG(D_HA, "%s: transno %lld is committed\n",
898 ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
901 class_export_cb_put(ccb->llcc_exp);
906 * Add commit callback function, it returns a non-zero value to inform
907 * caller to use sync transaction if necessary.
909 static int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
910 struct obd_export *exp, __u64 transno)
912 struct tgt_last_committed_callback *ccb;
913 struct dt_txn_commit_cb *dcb;
921 ccb->llcc_exp = class_export_cb_get(exp);
922 ccb->llcc_transno = transno;
925 dcb->dcb_func = tgt_cb_last_committed;
926 INIT_LIST_HEAD(&dcb->dcb_linkage);
927 strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
929 rc = dt_trans_cb_add(th, dcb);
931 class_export_cb_put(exp);
935 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
936 /* report failure to force synchronous operation */
939 /* if exp_need_sync is set, return non-zero value to force
940 * a sync transaction. */
941 return rc ? rc : exp->exp_need_sync;
945 * Add new client to the last_rcvd upon new connection.
947 * We use a bitmap to locate a free space in the last_rcvd file and initialize
950 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
952 struct tg_export_data *ted = &exp->exp_target_data;
953 struct lu_target *tgt = class_exp2tgt(exp);
958 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
959 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
962 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
965 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
966 * there's no need for extra complication here
968 idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
970 if (idx >= LR_MAX_CLIENTS ||
971 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
972 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
973 tgt->lut_obd->obd_name, idx);
976 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
977 idx = find_next_zero_bit(tgt->lut_client_bitmap,
978 LR_MAX_CLIENTS, idx);
982 ted->ted_lr_idx = idx;
983 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
984 idx * tgt->lut_lsd.lsd_client_size;
986 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
988 if (tgt_is_multimodrpcs_client(exp)) {
989 /* Set MULTI RPCS incompatibility flag to prevent previous
990 * Lustre versions to mount a target with reply_data file */
991 atomic_inc(&tgt->lut_num_clients);
992 if (!(tgt->lut_lsd.lsd_feature_incompat &
993 OBD_INCOMPAT_MULTI_RPCS)) {
994 tgt->lut_lsd.lsd_feature_incompat |=
995 OBD_INCOMPAT_MULTI_RPCS;
996 rc = tgt_server_data_update(env, tgt, 1);
998 CERROR("%s: unable to set MULTI RPCS "
999 "incompatibility flag\n",
1000 exp->exp_obd->obd_name);
1005 /* assign client slot generation */
1006 ted->ted_lcd->lcd_generation =
1007 atomic_inc_return(&tgt->lut_client_generation);
1009 ted->ted_lcd->lcd_generation = 0;
1012 CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s' "
1014 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1015 ted->ted_lcd->lcd_uuid, ted->ted_lcd->lcd_generation);
1017 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
1020 rc = tgt_client_data_update(env, exp);
1022 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
1023 tgt->lut_obd->obd_name, idx, rc);
1027 EXPORT_SYMBOL(tgt_client_new);
1029 /* Add an existing client to the MDS in-memory state based on
1030 * a client that was previously found in the last_rcvd file and
1031 * already has an assigned slot (idx >= 0).
1033 * It should not be possible to fail adding an existing client - otherwise
1034 * mdt_init_server_data() callsite needs to be fixed.
1036 int tgt_client_add(const struct lu_env *env, struct obd_export *exp, int idx)
1038 struct tg_export_data *ted = &exp->exp_target_data;
1039 struct lu_target *tgt = class_exp2tgt(exp);
1043 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
1044 LASSERTF(idx >= 0, "%d\n", idx);
1046 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
1047 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1050 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
1051 CERROR("%s: client %d: bit already set in bitmap!!\n",
1052 tgt->lut_obd->obd_name, idx);
1055 atomic_inc(&tgt->lut_num_clients);
1057 CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added, "
1059 tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid,
1060 ted->ted_lcd->lcd_generation);
1062 ted->ted_lr_idx = idx;
1063 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
1064 idx * tgt->lut_lsd.lsd_client_size;
1066 mutex_init(&ted->ted_lcd_lock);
1068 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1073 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
1075 struct tg_export_data *ted = &exp->exp_target_data;
1076 struct lu_target *tgt = class_exp2tgt(exp);
1081 LASSERT(ted->ted_lcd);
1083 if (unlikely(tgt == NULL)) {
1084 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1085 class_exp2obd(exp)->obd_name);
1089 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
1090 if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
1091 (char *)tgt->lut_obd->obd_uuid.uuid) ||
1092 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1095 /* Slot may be not yet assigned, use case is race between Client
1096 * reconnect and forced eviction */
1097 if (ted->ted_lr_idx < 0) {
1098 CWARN("%s: client with UUID '%s' not in bitmap\n",
1099 tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid);
1103 CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
1104 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1105 ted->ted_lcd->lcd_uuid);
1107 /* Clear the bit _after_ zeroing out the client so we don't
1108 race with filter_client_add and zero out new clients.*/
1109 if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
1110 CERROR("%s: client %u: bit already clear in bitmap!!\n",
1111 tgt->lut_obd->obd_name, ted->ted_lr_idx);
1115 /* Do not erase record for recoverable client. */
1116 if (exp->exp_flags & OBD_OPT_FAILOVER)
1119 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_DEL))
1122 /* Make sure the server's last_transno is up to date.
1123 * This should be done before zeroing client slot so last_transno will
1124 * be in server data or in client data in case of failure */
1125 rc = tgt_server_data_update(env, tgt, 0);
1127 CERROR("%s: failed to update server data, skip client %s "
1128 "zeroing, rc %d\n", tgt->lut_obd->obd_name,
1129 ted->ted_lcd->lcd_uuid, rc);
1133 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
1134 rc = tgt_client_data_update(env, exp);
1136 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
1137 "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
1138 tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
1139 ted->ted_lr_idx, ted->ted_lr_off, rc);
1142 EXPORT_SYMBOL(tgt_client_del);
1144 int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
1145 struct tg_export_data *ted, struct tg_reply_data *trd,
1146 struct thandle *th, bool update_lrd_file)
1148 struct lsd_reply_data *lrd;
1151 lrd = &trd->trd_reply;
1152 /* update export last transno */
1153 mutex_lock(&ted->ted_lcd_lock);
1154 if (lrd->lrd_transno > ted->ted_lcd->lcd_last_transno)
1155 ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
1156 mutex_unlock(&ted->ted_lcd_lock);
1158 /* find a empty slot */
1159 i = tgt_find_free_reply_slot(tgt);
1160 if (unlikely(i < 0)) {
1161 CERROR("%s: couldn't find a slot for reply data: "
1162 "rc = %d\n", tgt_name(tgt), i);
1167 if (update_lrd_file) {
1171 /* write reply data to disk */
1172 off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
1173 rc = tgt_reply_data_write(env, tgt, lrd, off, th);
1174 if (unlikely(rc != 0)) {
1175 CERROR("%s: can't update %s file: rc = %d\n",
1176 tgt_name(tgt), REPLY_DATA, rc);
1180 /* add reply data to target export's reply list */
1181 mutex_lock(&ted->ted_lcd_lock);
1182 list_add(&trd->trd_list, &ted->ted_reply_list);
1183 ted->ted_reply_cnt++;
1184 if (ted->ted_reply_cnt > ted->ted_reply_max)
1185 ted->ted_reply_max = ted->ted_reply_cnt;
1186 mutex_unlock(&ted->ted_lcd_lock);
1188 CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
1189 "tag %hu, client gen %u, slot idx %d\n",
1190 trd, lrd->lrd_xid, lrd->lrd_transno,
1191 trd->trd_tag, lrd->lrd_client_gen, i);
1194 EXPORT_SYMBOL(tgt_add_reply_data);
1197 * last_rcvd & last_committed update callbacks
1199 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
1200 struct dt_object *obj, __u64 opdata,
1201 struct thandle *th, struct ptlrpc_request *req)
1203 struct tgt_thread_info *tti = tgt_th_info(env);
1204 struct tgt_session_info *tsi = tgt_ses_info(env);
1205 struct obd_export *exp = tsi->tsi_exp;
1206 struct tg_export_data *ted;
1214 LASSERT(exp != NULL);
1215 ted = &exp->exp_target_data;
1217 lw_client = exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT;
1218 if (ted->ted_lr_idx < 0 && !lw_client)
1219 /* ofd connect may cause transaction before export has
1224 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
1226 /* From update replay, tti_transno should be set already */
1227 LASSERT(tti->tti_transno != 0);
1229 spin_lock(&tgt->lut_translock);
1230 if (th->th_result != 0) {
1231 if (tti->tti_transno != 0) {
1232 CERROR("%s: replay transno %llu failed: rc = %d\n",
1233 tgt_name(tgt), tti->tti_transno, th->th_result);
1235 } else if (tti->tti_transno == 0) {
1236 tti->tti_transno = ++tgt->lut_last_transno;
1238 /* should be replay */
1239 if (tti->tti_transno > tgt->lut_last_transno)
1240 tgt->lut_last_transno = tti->tti_transno;
1242 spin_unlock(&tgt->lut_translock);
1244 /** VBR: set new versions */
1245 if (th->th_result == 0 && obj != NULL) {
1246 struct dt_object *dto = dt_object_locate(obj, th->th_dev);
1247 dt_version_set(env, dto, tti->tti_transno, th);
1250 /* filling reply data */
1251 CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
1252 tti->tti_transno, tgt->lut_obd->obd_last_committed);
1255 req->rq_transno = tti->tti_transno;
1256 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
1259 /* if can't add callback, do sync write */
1260 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp, tti->tti_transno);
1263 /* All operations performed by LW clients are synchronous and
1264 * we store the committed transno in the last_rcvd header */
1265 spin_lock(&tgt->lut_translock);
1266 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
1267 tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
1268 spin_unlock(&tgt->lut_translock);
1269 /* Although lightweight (LW) connections have no slot
1270 * in the last_rcvd, we still want to maintain
1271 * the in-memory lsd_client_data structure in order to
1272 * properly handle reply reconstruction. */
1273 rc = tgt_server_data_write(env, tgt, th);
1275 spin_unlock(&tgt->lut_translock);
1277 } else if (ted->ted_lr_off == 0) {
1278 CERROR("%s: client idx %d has offset %lld\n",
1279 tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
1283 /* Target that supports multiple reply data */
1284 if (tgt_is_multimodrpcs_client(exp)) {
1285 struct tg_reply_data *trd;
1286 struct lsd_reply_data *lrd;
1287 __u64 *pre_versions;
1291 if (unlikely(trd == NULL))
1294 /* fill reply data information */
1295 lrd = &trd->trd_reply;
1296 lrd->lrd_transno = tti->tti_transno;
1298 lrd->lrd_xid = req->rq_xid;
1299 trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
1300 pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1301 lrd->lrd_result = th->th_result;
1302 lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
1303 write_update = true;
1305 LASSERT(tsi->tsi_xid != 0);
1306 lrd->lrd_xid = tsi->tsi_xid;
1307 lrd->lrd_result = tsi->tsi_result;
1308 lrd->lrd_client_gen = tsi->tsi_client_gen;
1310 pre_versions = NULL;
1311 write_update = false;
1314 lrd->lrd_data = opdata;
1316 trd->trd_pre_versions[0] = pre_versions[0];
1317 trd->trd_pre_versions[1] = pre_versions[1];
1318 trd->trd_pre_versions[2] = pre_versions[2];
1319 trd->trd_pre_versions[3] = pre_versions[3];
1322 rc = tgt_add_reply_data(env, tgt, ted, trd, th, write_update);
1328 /* Enough for update replay, let's return */
1332 mutex_lock(&ted->ted_lcd_lock);
1333 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1334 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
1335 transno_p = &ted->ted_lcd->lcd_last_close_transno;
1336 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
1337 ted->ted_lcd->lcd_last_close_result = th->th_result;
1339 /* VBR: save versions in last_rcvd for reconstruct. */
1340 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1343 ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
1344 ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
1345 ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
1346 ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
1348 transno_p = &ted->ted_lcd->lcd_last_transno;
1349 ted->ted_lcd->lcd_last_xid = req->rq_xid;
1350 ted->ted_lcd->lcd_last_result = th->th_result;
1351 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
1352 * see struct ldlm_reply->lock_policy_res1; */
1353 ted->ted_lcd->lcd_last_data = opdata;
1356 /* Update transno in slot only if non-zero number, i.e. no errors */
1357 if (likely(tti->tti_transno != 0)) {
1358 /* Don't overwrite bigger transaction number with lower one.
1359 * That is not sign of problem in all cases, but in any case
1360 * this value should be monotonically increased only. */
1361 if (*transno_p > tti->tti_transno) {
1362 if (!tgt->lut_no_reconstruct) {
1363 CERROR("%s: trying to overwrite bigger transno:"
1364 "on-disk: %llu, new: %llu replay: "
1365 "%d. See LU-617.\n", tgt_name(tgt),
1366 *transno_p, tti->tti_transno,
1367 req_is_replay(req));
1368 if (req_is_replay(req)) {
1369 spin_lock(&req->rq_export->exp_lock);
1370 req->rq_export->exp_vbr_failed = 1;
1371 spin_unlock(&req->rq_export->exp_lock);
1373 mutex_unlock(&ted->ted_lcd_lock);
1374 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
1377 *transno_p = tti->tti_transno;
1382 tti->tti_off = ted->ted_lr_off;
1383 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_RCVD_EIO))
1386 rc = tgt_client_data_write(env, tgt, ted->ted_lcd,
1389 mutex_unlock(&ted->ted_lcd_lock);
1393 mutex_unlock(&ted->ted_lcd_lock);
1398 * last_rcvd update for echo client simulation.
1399 * It updates last_rcvd client slot and version of object in
1400 * simple way but with all locks to simulate all drawbacks
1402 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
1403 struct lu_target *tgt,
1404 struct dt_object *obj,
1406 struct obd_export *exp)
1408 struct tgt_thread_info *tti = tgt_th_info(env);
1409 struct tg_export_data *ted = &exp->exp_target_data;
1414 tti->tti_transno = 0;
1416 spin_lock(&tgt->lut_translock);
1417 if (th->th_result == 0)
1418 tti->tti_transno = ++tgt->lut_last_transno;
1419 spin_unlock(&tgt->lut_translock);
1421 /** VBR: set new versions */
1422 if (th->th_result == 0 && obj != NULL)
1423 dt_version_set(env, obj, tti->tti_transno, th);
1425 /* if can't add callback, do sync write */
1426 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
1429 LASSERT(ted->ted_lr_off > 0);
1431 mutex_lock(&ted->ted_lcd_lock);
1432 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1433 ted->ted_lcd->lcd_last_transno = tti->tti_transno;
1434 ted->ted_lcd->lcd_last_result = th->th_result;
1436 tti->tti_off = ted->ted_lr_off;
1437 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1438 mutex_unlock(&ted->ted_lcd_lock);
1442 static int tgt_clients_data_init(const struct lu_env *env,
1443 struct lu_target *tgt,
1444 unsigned long last_size)
1446 struct obd_device *obd = tgt->lut_obd;
1447 struct lr_server_data *lsd = &tgt->lut_lsd;
1448 struct lsd_client_data *lcd = NULL;
1449 struct tg_export_data *ted;
1452 loff_t off = lsd->lsd_client_start;
1453 __u32 generation = 0;
1454 struct cfs_hash *hash = NULL;
1458 if (tgt->lut_bottom->dd_rdonly)
1461 CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
1462 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
1468 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1470 GOTO(err_out, rc = -ENODEV);
1472 for (cl_idx = 0; off < last_size; cl_idx++) {
1473 struct obd_export *exp;
1476 /* Don't assume off is incremented properly by
1477 * read_record(), in case sizeof(*lcd)
1478 * isn't the same as fsd->lsd_client_size. */
1479 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
1480 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
1482 CERROR("%s: error reading last_rcvd %s idx %d off "
1483 "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
1486 break; /* read error shouldn't cause startup to fail */
1489 if (lcd->lcd_uuid[0] == '\0') {
1490 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
1495 last_transno = lcd_last_transno(lcd);
1497 /* These exports are cleaned up by disconnect, so they
1498 * need to be set up like real exports as connect does.
1500 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: %llu"
1501 " srv lr: %llu lx: %llu gen %u\n", lcd->lcd_uuid,
1502 cl_idx, last_transno, lsd->lsd_last_transno,
1503 lcd_last_xid(lcd), lcd->lcd_generation);
1505 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
1507 if (PTR_ERR(exp) == -EALREADY) {
1508 /* export already exists, zero out this one */
1509 CERROR("%s: Duplicate export %s!\n",
1510 tgt_name(tgt), lcd->lcd_uuid);
1513 GOTO(err_out, rc = PTR_ERR(exp));
1516 ted = &exp->exp_target_data;
1517 *ted->ted_lcd = *lcd;
1519 rc = tgt_client_add(env, exp, cl_idx);
1520 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
1521 /* VBR: set export last committed version */
1522 exp->exp_last_committed = last_transno;
1523 spin_lock(&exp->exp_lock);
1524 exp->exp_connecting = 0;
1525 exp->exp_in_recovery = 0;
1526 spin_unlock(&exp->exp_lock);
1527 obd->obd_max_recoverable_clients++;
1529 if (tgt->lut_lsd.lsd_feature_incompat &
1530 OBD_INCOMPAT_MULTI_RPCS &&
1531 lcd->lcd_generation != 0) {
1532 /* compute the highest valid client generation */
1533 generation = max(generation, lcd->lcd_generation);
1534 /* fill client_generation <-> export hash table */
1535 rc = cfs_hash_add_unique(hash, &lcd->lcd_generation,
1536 &exp->exp_gen_hash);
1538 CERROR("%s: duplicate export for client "
1540 tgt_name(tgt), lcd->lcd_generation);
1541 class_export_put(exp);
1546 class_export_put(exp);
1548 rc = rev_import_init(exp);
1550 class_unlink_export(exp);
1554 /* Need to check last_rcvd even for duplicated exports. */
1555 CDEBUG(D_OTHER, "client at idx %d has last_transno = %llu\n",
1556 cl_idx, last_transno);
1558 spin_lock(&tgt->lut_translock);
1559 tgt->lut_last_transno = max(last_transno,
1560 tgt->lut_last_transno);
1561 spin_unlock(&tgt->lut_translock);
1564 /* record highest valid client generation */
1565 atomic_set(&tgt->lut_client_generation, generation);
1569 cfs_hash_putref(hash);
1574 struct server_compat_data {
1581 static struct server_compat_data tgt_scd[] = {
1582 [LDD_F_SV_TYPE_MDT] = {
1583 .rocompat = OBD_ROCOMPAT_LOVOBJID,
1584 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1585 OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
1586 OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI |
1587 OBD_INCOMPAT_MULTI_RPCS,
1588 .rocinit = OBD_ROCOMPAT_LOVOBJID,
1589 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1590 OBD_INCOMPAT_MULTI_OI,
1592 [LDD_F_SV_TYPE_OST] = {
1593 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
1594 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
1596 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
1597 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1601 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1603 struct tgt_thread_info *tti = tgt_th_info(env);
1604 struct lr_server_data *lsd = &tgt->lut_lsd;
1605 unsigned long last_rcvd_size;
1609 rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1613 last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1615 /* ensure padding in the struct is the correct size */
1616 CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
1617 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
1619 rc = server_name2index(tgt_name(tgt), &index, NULL);
1621 CERROR("%s: Can not get index from name: rc = %d\n",
1625 /* server_name2index() returns type */
1627 if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1628 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1632 /* last_rcvd on OST doesn't provide reconstruct support because there
1633 * may be up to 8 in-flight write requests per single slot in
1634 * last_rcvd client data
1636 tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1638 if (last_rcvd_size == 0) {
1639 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1641 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1642 sizeof(lsd->lsd_uuid));
1643 lsd->lsd_last_transno = 0;
1644 lsd->lsd_mount_count = 0;
1645 lsd->lsd_server_size = LR_SERVER_SIZE;
1646 lsd->lsd_client_start = LR_CLIENT_START;
1647 lsd->lsd_client_size = LR_CLIENT_SIZE;
1648 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1649 lsd->lsd_osd_index = index;
1650 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1651 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1653 rc = tgt_server_data_read(env, tgt);
1655 CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1659 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1660 if (tgt->lut_bottom->dd_rdonly) {
1661 /* Such difference may be caused by mounting
1662 * up snapshot with new fsname under rd_only
1663 * mode. But even if it was NOT, it will not
1664 * damage the system because of "rd_only". */
1665 memcpy(lsd->lsd_uuid,
1666 tgt->lut_obd->obd_uuid.uuid,
1667 sizeof(lsd->lsd_uuid));
1669 LCONSOLE_ERROR_MSG(0x157, "Trying to start "
1670 "OBD %s using the wrong "
1671 "disk %s. Were the /dev/ "
1672 "assignments rearranged?\n",
1673 tgt->lut_obd->obd_uuid.uuid,
1679 if (lsd->lsd_osd_index != index) {
1680 LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
1681 "is different with the index %d in"
1682 "config log, It might be disk"
1683 "corruption!\n", tgt_name(tgt),
1684 lsd->lsd_osd_index, index);
1689 if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1690 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1692 lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1696 if (type == LDD_F_SV_TYPE_MDT)
1697 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1699 if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1700 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1702 lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1705 /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1706 if (type == LDD_F_SV_TYPE_MDT &&
1707 !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1708 if (last_rcvd_size > lsd->lsd_client_start) {
1709 LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1710 "remove all clients for interop needs\n",
1712 rc = tgt_truncate_last_rcvd(env, tgt,
1713 lsd->lsd_client_start);
1716 last_rcvd_size = lsd->lsd_client_start;
1718 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1719 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1722 spin_lock(&tgt->lut_translock);
1723 tgt->lut_last_transno = lsd->lsd_last_transno;
1724 spin_unlock(&tgt->lut_translock);
1726 lsd->lsd_mount_count++;
1728 CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1729 CDEBUG(D_INODE, "%s: server last_transno: %llu\n",
1730 tgt_name(tgt), tgt->lut_last_transno);
1731 CDEBUG(D_INODE, "%s: server mount_count: %llu\n",
1732 tgt_name(tgt), lsd->lsd_mount_count);
1733 CDEBUG(D_INODE, "%s: server data size: %u\n",
1734 tgt_name(tgt), lsd->lsd_server_size);
1735 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1736 tgt_name(tgt), lsd->lsd_client_start);
1737 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1738 tgt_name(tgt), lsd->lsd_client_size);
1739 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1740 tgt_name(tgt), last_rcvd_size);
1741 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1742 tgt_name(tgt), lsd->lsd_subdir_count);
1743 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1744 last_rcvd_size <= lsd->lsd_client_start ? 0 :
1745 (last_rcvd_size - lsd->lsd_client_start) /
1746 lsd->lsd_client_size);
1747 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1749 if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1750 lsd->lsd_client_size == 0) {
1751 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1755 if (!tgt->lut_obd->obd_replayable)
1756 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1758 rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1760 GOTO(err_client, rc);
1762 spin_lock(&tgt->lut_translock);
1763 /* obd_last_committed is used for compatibility
1764 * with other lustre recovery code */
1765 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1766 spin_unlock(&tgt->lut_translock);
1768 tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1769 tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1771 /* save it, so mount count and last_transno is current */
1772 rc = tgt_server_data_update(env, tgt, 0);
1774 GOTO(err_client, rc);
1779 class_disconnect_exports(tgt->lut_obd);
1783 /* add credits for last_rcvd update */
1784 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1787 struct lu_target *tgt = cookie;
1788 struct tgt_session_info *tsi;
1789 struct tgt_thread_info *tti = tgt_th_info(env);
1790 struct dt_object *dto;
1793 /* For readonly case, the caller should have got failure
1794 * when start the transaction. If the logic comes here,
1795 * there must be something wrong. */
1796 if (unlikely(tgt->lut_bottom->dd_rdonly)) {
1801 /* if there is no session, then this transaction is not result of
1802 * request processing but some local operation */
1803 if (env->le_ses == NULL)
1806 LASSERT(tgt->lut_last_rcvd);
1807 tsi = tgt_ses_info(env);
1808 /* OFD may start transaction without export assigned */
1809 if (tsi->tsi_exp == NULL)
1812 if (tgt_is_multimodrpcs_client(tsi->tsi_exp)) {
1814 * Use maximum possible file offset for declaration to ensure
1815 * ZFS will reserve enough credits for a write anywhere in this
1816 * file, since we don't know where in the file the write will be
1817 * because a replay slot has not been assigned. This should be
1818 * replaced by dmu_tx_hold_append() when available.
1820 tti->tti_buf.lb_buf = NULL;
1821 tti->tti_buf.lb_len = sizeof(struct lsd_reply_data);
1822 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
1823 rc = dt_declare_record_write(env, dto, &tti->tti_buf, -1, th);
1827 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
1829 tti->tti_off = tsi->tsi_exp->exp_target_data.ted_lr_off;
1830 rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1836 if (tsi->tsi_vbr_obj != NULL &&
1837 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1838 dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
1839 rc = dt_declare_version_set(env, dto, th);
1845 /* Update last_rcvd records with latests transaction data */
1846 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1849 struct lu_target *tgt = cookie;
1850 struct tgt_session_info *tsi;
1851 struct tgt_thread_info *tti = tgt_th_info(env);
1852 struct dt_object *obj = NULL;
1856 if (env->le_ses == NULL)
1859 tsi = tgt_ses_info(env);
1860 /* OFD may start transaction without export assigned */
1861 if (tsi->tsi_exp == NULL)
1864 echo_client = (tgt_ses_req(tsi) == NULL && tsi->tsi_xid == 0);
1866 if (tti->tti_has_trans && !echo_client) {
1867 if (tti->tti_mult_trans == 0) {
1868 CDEBUG(D_HA, "More than one transaction %llu\n",
1872 /* we need another transno to be assigned */
1873 tti->tti_transno = 0;
1874 } else if (th->th_result == 0) {
1875 tti->tti_has_trans = 1;
1878 if (tsi->tsi_vbr_obj != NULL &&
1879 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1880 obj = tsi->tsi_vbr_obj;
1883 if (unlikely(echo_client)) /* echo client special case */
1884 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1887 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
1892 int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
1894 struct tgt_thread_info *tti = tgt_th_info(env);
1895 struct lsd_reply_data *lrd = &tti->tti_lrd;
1896 unsigned long reply_data_size;
1898 struct lsd_reply_header *lrh = NULL;
1899 struct lsd_client_data *lcd = NULL;
1900 struct tg_reply_data *trd = NULL;
1903 struct cfs_hash *hash = NULL;
1904 struct obd_export *exp;
1905 struct tg_export_data *ted;
1906 int reply_data_recovered = 0;
1908 rc = dt_attr_get(env, tgt->lut_reply_data, &tti->tti_attr);
1911 reply_data_size = (unsigned long)tti->tti_attr.la_size;
1915 GOTO(out, rc = -ENOMEM);
1917 if (reply_data_size == 0) {
1918 CDEBUG(D_INFO, "%s: new reply_data file, initializing\n",
1920 lrh->lrh_magic = LRH_MAGIC;
1921 lrh->lrh_header_size = sizeof(struct lsd_reply_header);
1922 lrh->lrh_reply_size = sizeof(struct lsd_reply_data);
1923 rc = tgt_reply_header_write(env, tgt, lrh);
1925 CERROR("%s: error writing %s: rc = %d\n",
1926 tgt_name(tgt), REPLY_DATA, rc);
1930 rc = tgt_reply_header_read(env, tgt, lrh);
1932 CERROR("%s: error reading %s: rc = %d\n",
1933 tgt_name(tgt), REPLY_DATA, rc);
1936 if (lrh->lrh_magic != LRH_MAGIC ||
1937 lrh->lrh_header_size != sizeof(struct lsd_reply_header) ||
1938 lrh->lrh_reply_size != sizeof(struct lsd_reply_data)) {
1939 CERROR("%s: invalid header in %s\n",
1940 tgt_name(tgt), REPLY_DATA);
1941 GOTO(out, rc = -EINVAL);
1944 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1946 GOTO(out, rc = -ENODEV);
1950 GOTO(out, rc = -ENOMEM);
1954 GOTO(out, rc = -ENOMEM);
1956 /* Load reply_data from disk */
1957 for (idx = 0, off = sizeof(struct lsd_reply_header);
1958 off < reply_data_size;
1959 idx++, off += sizeof(struct lsd_reply_data)) {
1960 rc = tgt_reply_data_read(env, tgt, lrd, off);
1962 CERROR("%s: error reading %s: rc = %d\n",
1963 tgt_name(tgt), REPLY_DATA, rc);
1967 exp = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
1969 /* old reply data from a disconnected client */
1972 ted = &exp->exp_target_data;
1973 mutex_lock(&ted->ted_lcd_lock);
1975 /* create in-memory reply_data and link it to
1976 * target export's reply list */
1977 rc = tgt_set_reply_slot(tgt, idx);
1979 mutex_unlock(&ted->ted_lcd_lock);
1982 trd->trd_reply = *lrd;
1983 trd->trd_pre_versions[0] = 0;
1984 trd->trd_pre_versions[1] = 0;
1985 trd->trd_pre_versions[2] = 0;
1986 trd->trd_pre_versions[3] = 0;
1987 trd->trd_index = idx;
1989 list_add(&trd->trd_list, &ted->ted_reply_list);
1990 ted->ted_reply_cnt++;
1991 if (ted->ted_reply_cnt > ted->ted_reply_max)
1992 ted->ted_reply_max = ted->ted_reply_cnt;
1994 CDEBUG(D_HA, "%s: restore reply %p: xid %llu, "
1995 "transno %llu, client gen %u, slot idx %d\n",
1996 tgt_name(tgt), trd, lrd->lrd_xid,
1997 lrd->lrd_transno, lrd->lrd_client_gen,
2000 /* update export last committed transation */
2001 exp->exp_last_committed = max(exp->exp_last_committed,
2004 mutex_unlock(&ted->ted_lcd_lock);
2005 class_export_put(exp);
2007 /* update target last committed transaction */
2008 spin_lock(&tgt->lut_translock);
2009 tgt->lut_last_transno = max(tgt->lut_last_transno,
2011 spin_unlock(&tgt->lut_translock);
2013 reply_data_recovered++;
2017 GOTO(out, rc = -ENOMEM);
2019 CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
2020 tgt_name(tgt), reply_data_recovered);
2023 spin_lock(&tgt->lut_translock);
2024 /* obd_last_committed is used for compatibility
2025 * with other lustre recovery code */
2026 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
2027 spin_unlock(&tgt->lut_translock);
2033 cfs_hash_putref(hash);
2043 struct tg_reply_data *tgt_lookup_reply_by_xid(struct tg_export_data *ted,
2046 struct tg_reply_data *found = NULL;
2047 struct tg_reply_data *reply;
2049 mutex_lock(&ted->ted_lcd_lock);
2050 list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
2051 if (reply->trd_reply.lrd_xid == xid) {
2056 mutex_unlock(&ted->ted_lcd_lock);
2059 EXPORT_SYMBOL(tgt_lookup_reply_by_xid);
2061 /* Look for a reply data matching specified request @req
2062 * A copy is returned in @trd if the pointer is not NULL
2064 bool tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
2066 struct tg_export_data *ted = &req->rq_export->exp_target_data;
2067 struct tg_reply_data *reply;
2070 reply = tgt_lookup_reply_by_xid(ted, req->rq_xid);
2071 if (reply != NULL) {
2077 CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d\n",
2078 tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid,
2083 EXPORT_SYMBOL(tgt_lookup_reply);
2085 int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
2087 struct tg_export_data *ted = &exp->exp_target_data;
2088 struct lu_target *lut = class_exp2tgt(exp);
2089 struct tg_reply_data *trd, *tmp;
2091 mutex_lock(&ted->ted_lcd_lock);
2092 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
2093 if (trd->trd_reply.lrd_xid > rcvd_xid)
2095 ted->ted_release_xid++;
2096 tgt_release_reply_data(lut, ted, trd);
2098 mutex_unlock(&ted->ted_lcd_lock);
2103 int tgt_handle_tag(struct obd_export *exp, __u16 tag)
2105 struct tg_export_data *ted = &exp->exp_target_data;
2106 struct lu_target *lut = class_exp2tgt(exp);
2107 struct tg_reply_data *trd, *tmp;
2112 mutex_lock(&ted->ted_lcd_lock);
2113 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
2114 if (trd->trd_tag != tag)
2116 ted->ted_release_tag++;
2117 tgt_release_reply_data(lut, ted, trd);
2120 mutex_unlock(&ted->ted_lcd_lock);