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/
31 * Lustre Unified Target
32 * These are common function to work with last_received file
34 * Author: Mikhail Pershin <mike.pershin@intel.com>
37 #include <obd_class.h>
38 #include <lustre_fid.h>
40 #include "tgt_internal.h"
42 /** version recovery epoch */
43 #define LR_EPOCH_BITS 32
45 /* Allocate a bitmap for a chunk of reply data slots */
46 static int tgt_bitmap_chunk_alloc(struct lu_target *lut, int chunk)
50 OBD_ALLOC_LARGE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
55 spin_lock(&lut->lut_client_bitmap_lock);
57 if (lut->lut_reply_bitmap[chunk] != NULL) {
58 /* someone else already allocated the bitmap for this chunk */
59 spin_unlock(&lut->lut_client_bitmap_lock);
60 OBD_FREE_LARGE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
65 lut->lut_reply_bitmap[chunk] = bm;
67 spin_unlock(&lut->lut_client_bitmap_lock);
72 /* Look for an available reply data slot in the bitmap
74 * Allocate bitmap chunk when first used
75 * XXX algo could be improved if this routine limits performance
77 static int tgt_find_free_reply_slot(struct lu_target *lut)
84 for (chunk = 0; chunk < LUT_REPLY_SLOTS_MAX_CHUNKS; chunk++) {
85 /* allocate the bitmap chunk if necessary */
86 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
87 rc = tgt_bitmap_chunk_alloc(lut, chunk);
91 bmp = lut->lut_reply_bitmap[chunk];
93 /* look for an available slot in this chunk */
95 b = find_first_zero_bit(bmp, LUT_REPLY_SLOTS_PER_CHUNK);
96 if (b >= LUT_REPLY_SLOTS_PER_CHUNK)
100 if (test_and_set_bit(b, bmp) == 0)
101 return chunk * LUT_REPLY_SLOTS_PER_CHUNK + b;
108 /* Mark the reply data slot @idx 'used' in the corresponding bitmap chunk
110 * Allocate the bitmap chunk if necessary
112 static int tgt_set_reply_slot(struct lu_target *lut, int idx)
118 chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
119 b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
121 LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
122 LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
124 /* allocate the bitmap chunk if necessary */
125 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
126 rc = tgt_bitmap_chunk_alloc(lut, chunk);
131 /* mark the slot 'used' in this chunk */
132 if (test_and_set_bit(b, lut->lut_reply_bitmap[chunk]) != 0) {
133 CERROR("%s: slot %d already set in bitmap\n",
142 /* Mark the reply data slot @idx 'unused' in the corresponding bitmap chunk
145 static int tgt_clear_reply_slot(struct lu_target *lut, int idx)
150 if (lut->lut_obd->obd_stopping)
152 * in case of failover keep the bit set in order to
153 * avoid overwriting slots in reply_data which might
154 * be required by resent rpcs
157 chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
158 b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
160 LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
161 LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
163 if (lut->lut_reply_bitmap[chunk] == NULL) {
164 CERROR("%s: slot %d not allocated\n",
169 if (test_and_clear_bit(b, lut->lut_reply_bitmap[chunk]) == 0) {
170 CERROR("%s: slot %d already clear in bitmap\n",
179 /* Read header of reply_data file of target @tgt into structure @lrh */
180 static int tgt_reply_header_read(const struct lu_env *env,
181 struct lu_target *tgt,
182 struct lsd_reply_header *lrh)
185 struct lsd_reply_header buf;
186 struct tgt_thread_info *tti = tgt_th_info(env);
189 tti->tti_buf.lb_buf = &buf;
190 tti->tti_buf.lb_len = sizeof(buf);
192 rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
197 lrh->lrh_magic = le32_to_cpu(buf.lrh_magic);
198 lrh->lrh_header_size = le32_to_cpu(buf.lrh_header_size);
199 lrh->lrh_reply_size = le32_to_cpu(buf.lrh_reply_size);
201 CDEBUG(D_HA, "%s: read %s header. magic=0x%08x "
202 "header_size=%d reply_size=%d\n",
203 tgt->lut_obd->obd_name, REPLY_DATA,
204 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
209 /* Write header into replay_data file of target @tgt from structure @lrh */
210 static int tgt_reply_header_write(const struct lu_env *env,
211 struct lu_target *tgt,
212 struct lsd_reply_header *lrh)
215 struct lsd_reply_header buf;
216 struct tgt_thread_info *tti = tgt_th_info(env);
218 struct dt_object *dto;
220 CDEBUG(D_HA, "%s: write %s header. magic=0x%08x "
221 "header_size=%d reply_size=%d\n",
222 tgt->lut_obd->obd_name, REPLY_DATA,
223 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
225 if (tgt->lut_bottom->dd_rdonly)
228 buf.lrh_magic = cpu_to_le32(lrh->lrh_magic);
229 buf.lrh_header_size = cpu_to_le32(lrh->lrh_header_size);
230 buf.lrh_reply_size = cpu_to_le32(lrh->lrh_reply_size);
232 th = dt_trans_create(env, tgt->lut_bottom);
238 tti->tti_buf.lb_buf = &buf;
239 tti->tti_buf.lb_len = sizeof(buf);
241 rc = dt_declare_record_write(env, tgt->lut_reply_data,
242 &tti->tti_buf, tti->tti_off, th);
246 rc = dt_trans_start(env, tgt->lut_bottom, th);
250 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
251 rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
253 dt_trans_stop(env, tgt->lut_bottom, th);
257 /* Write the reply data @lrd into reply_data file of target @tgt
260 static int tgt_reply_data_write(const struct lu_env *env, struct lu_target *tgt,
261 struct lsd_reply_data *lrd, loff_t off,
264 struct tgt_thread_info *tti = tgt_th_info(env);
265 struct dt_object *dto;
266 struct lsd_reply_data *buf = &tti->tti_lrd;
268 lrd->lrd_result = ptlrpc_status_hton(lrd->lrd_result);
270 buf->lrd_transno = cpu_to_le64(lrd->lrd_transno);
271 buf->lrd_xid = cpu_to_le64(lrd->lrd_xid);
272 buf->lrd_data = cpu_to_le64(lrd->lrd_data);
273 buf->lrd_result = cpu_to_le32(lrd->lrd_result);
274 buf->lrd_client_gen = cpu_to_le32(lrd->lrd_client_gen);
276 lrd->lrd_result = ptlrpc_status_ntoh(lrd->lrd_result);
279 tti->tti_buf.lb_buf = buf;
280 tti->tti_buf.lb_len = sizeof(*buf);
282 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
283 return dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
286 /* Read the reply data from reply_data file of target @tgt at offset @off
287 * into structure @lrd
289 static int tgt_reply_data_read(const struct lu_env *env, struct lu_target *tgt,
290 struct lsd_reply_data *lrd, loff_t off)
293 struct tgt_thread_info *tti = tgt_th_info(env);
294 struct lsd_reply_data *buf = &tti->tti_lrd;
297 tti->tti_buf.lb_buf = buf;
298 tti->tti_buf.lb_len = sizeof(*buf);
300 rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
305 lrd->lrd_transno = le64_to_cpu(buf->lrd_transno);
306 lrd->lrd_xid = le64_to_cpu(buf->lrd_xid);
307 lrd->lrd_data = le64_to_cpu(buf->lrd_data);
308 lrd->lrd_result = le32_to_cpu(buf->lrd_result);
309 lrd->lrd_client_gen = le32_to_cpu(buf->lrd_client_gen);
315 /* Free the in-memory reply data structure @trd and release
316 * the corresponding slot in the reply_data file of target @lut
317 * Called with ted_lcd_lock held
319 static void tgt_free_reply_data(struct lu_target *lut,
320 struct tg_export_data *ted,
321 struct tg_reply_data *trd)
323 CDEBUG(D_TRACE, "%s: free reply data %p: xid %llu, transno %llu, "
324 "client gen %u, slot idx %d\n",
325 lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
326 trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
329 LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
331 list_del(&trd->trd_list);
332 ted->ted_reply_cnt--;
333 if (lut != NULL && trd->trd_index != TRD_INDEX_MEMORY)
334 tgt_clear_reply_slot(lut, trd->trd_index);
338 /* Release the reply data @trd from target @lut
339 * The reply data with the highest transno for this export
340 * is retained to ensure correctness of target recovery
341 * Called with ted_lcd_lock held
343 static void tgt_release_reply_data(struct lu_target *lut,
344 struct tg_export_data *ted,
345 struct tg_reply_data *trd)
347 CDEBUG(D_TRACE, "%s: release reply data %p: xid %llu, transno %llu, "
348 "client gen %u, slot idx %d\n",
349 lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
350 trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
353 LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
355 /* Do not free the reply data corresponding to the
356 * highest transno of this export.
357 * This ensures on-disk reply data is kept and
358 * last committed transno can be restored from disk in case
361 if (trd->trd_reply.lrd_transno == ted->ted_lcd->lcd_last_transno) {
362 /* free previous retained reply */
363 if (ted->ted_reply_last != NULL)
364 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
365 /* retain the reply */
366 list_del_init(&trd->trd_list);
367 ted->ted_reply_last = trd;
369 tgt_free_reply_data(lut, ted, trd);
373 static inline struct lu_buf *tti_buf_lsd(struct tgt_thread_info *tti)
375 tti->tti_buf.lb_buf = &tti->tti_lsd;
376 tti->tti_buf.lb_len = sizeof(tti->tti_lsd);
377 return &tti->tti_buf;
380 static inline struct lu_buf *tti_buf_lcd(struct tgt_thread_info *tti)
382 tti->tti_buf.lb_buf = &tti->tti_lcd;
383 tti->tti_buf.lb_len = sizeof(tti->tti_lcd);
384 return &tti->tti_buf;
387 static inline bool tgt_is_multimodrpcs_record(struct lu_target *tgt,
388 struct lsd_client_data *lcd)
390 return tgt->lut_lsd.lsd_feature_incompat & OBD_INCOMPAT_MULTI_RPCS &&
391 lcd->lcd_generation != 0;
395 * Allocate in-memory data for client slot related to export.
397 int tgt_client_alloc(struct obd_export *exp)
400 LASSERT(exp != exp->exp_obd->obd_self_export);
402 spin_lock_init(&exp->exp_target_data.ted_nodemap_lock);
403 INIT_LIST_HEAD(&exp->exp_target_data.ted_nodemap_member);
404 spin_lock_init(&exp->exp_target_data.ted_fmd_lock);
405 INIT_LIST_HEAD(&exp->exp_target_data.ted_fmd_list);
407 OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
408 if (exp->exp_target_data.ted_lcd == NULL)
410 /* Mark that slot is not yet valid, 0 doesn't work here */
411 exp->exp_target_data.ted_lr_idx = -1;
412 INIT_LIST_HEAD(&exp->exp_target_data.ted_reply_list);
413 mutex_init(&exp->exp_target_data.ted_lcd_lock);
416 EXPORT_SYMBOL(tgt_client_alloc);
419 * Free in-memory data for client slot related to export.
421 void tgt_client_free(struct obd_export *exp)
423 struct tg_export_data *ted = &exp->exp_target_data;
424 struct lu_target *lut = class_exp2tgt(exp);
425 struct tg_reply_data *trd, *tmp;
427 LASSERT(exp != exp->exp_obd->obd_self_export);
429 tgt_fmd_cleanup(exp);
431 /* free reply data */
432 mutex_lock(&ted->ted_lcd_lock);
433 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
434 tgt_release_reply_data(lut, ted, trd);
436 if (ted->ted_reply_last != NULL) {
437 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
438 ted->ted_reply_last = NULL;
440 mutex_unlock(&ted->ted_lcd_lock);
442 if (!hlist_unhashed(&exp->exp_gen_hash))
443 cfs_hash_del(exp->exp_obd->obd_gen_hash,
444 &ted->ted_lcd->lcd_generation,
447 OBD_FREE_PTR(ted->ted_lcd);
450 /* Target may have been freed (see LU-7430)
451 * Slot may be not yet assigned */
452 if (((struct obd_device_target *)(&exp->exp_obd->u))->obt_magic !=
457 /* Clear bit when lcd is freed */
458 LASSERT(lut && lut->lut_client_bitmap);
459 if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
460 CERROR("%s: client %u bit already clear in bitmap\n",
461 exp->exp_obd->obd_name, ted->ted_lr_idx);
465 EXPORT_SYMBOL(tgt_client_free);
467 static inline void tgt_check_lcd(const char *obd_name, int index,
468 struct lsd_client_data *lcd)
470 size_t uuid_size = sizeof(lcd->lcd_uuid);
472 if (strnlen((char*)lcd->lcd_uuid, uuid_size) == uuid_size) {
473 lcd->lcd_uuid[uuid_size - 1] = '\0';
475 LCONSOLE_ERROR("the client UUID (%s) on %s for exports stored in last_rcvd(index = %d) is bad!\n",
476 lcd->lcd_uuid, obd_name, index);
480 static int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
481 struct lsd_client_data *lcd,
482 loff_t *off, int index)
484 struct tgt_thread_info *tti = tgt_th_info(env);
488 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
490 tgt_check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
491 lcd_le_to_cpu(&tti->tti_lcd, lcd);
492 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
493 lcd->lcd_last_close_result =
494 ptlrpc_status_ntoh(lcd->lcd_last_close_result);
497 CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = %llu"
498 ", last_xid = %llu, last_result = %u, last_data = %u, "
499 "last_close_transno = %llu, last_close_xid = %llu, "
500 "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
501 *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
502 lcd->lcd_last_result, lcd->lcd_last_data,
503 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
504 lcd->lcd_last_close_result, rc);
508 static int tgt_client_data_write(const struct lu_env *env,
509 struct lu_target *tgt,
510 struct lsd_client_data *lcd,
511 loff_t *off, struct thandle *th)
513 struct tgt_thread_info *tti = tgt_th_info(env);
514 struct dt_object *dto;
516 lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
517 lcd->lcd_last_close_result =
518 ptlrpc_status_hton(lcd->lcd_last_close_result);
519 lcd_cpu_to_le(lcd, &tti->tti_lcd);
522 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
523 return dt_record_write(env, dto, &tti->tti_buf, off, th);
526 struct tgt_new_client_callback {
527 struct dt_txn_commit_cb lncc_cb;
528 struct obd_export *lncc_exp;
531 static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
532 struct dt_txn_commit_cb *cb, int err)
534 struct tgt_new_client_callback *ccb;
536 ccb = container_of(cb, struct tgt_new_client_callback, lncc_cb);
538 LASSERT(ccb->lncc_exp->exp_obd);
540 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
541 ccb->lncc_exp->exp_obd->obd_name,
542 ccb->lncc_exp->exp_client_uuid.uuid);
544 spin_lock(&ccb->lncc_exp->exp_lock);
546 ccb->lncc_exp->exp_need_sync = 0;
548 spin_unlock(&ccb->lncc_exp->exp_lock);
549 class_export_cb_put(ccb->lncc_exp);
554 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
556 struct tgt_new_client_callback *ccb;
557 struct dt_txn_commit_cb *dcb;
564 ccb->lncc_exp = class_export_cb_get(exp);
567 dcb->dcb_func = tgt_cb_new_client;
568 INIT_LIST_HEAD(&dcb->dcb_linkage);
569 strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
571 rc = dt_trans_cb_add(th, dcb);
573 class_export_cb_put(exp);
580 * Update client data in last_rcvd
582 static int tgt_client_data_update(const struct lu_env *env,
583 struct obd_export *exp)
585 struct tg_export_data *ted = &exp->exp_target_data;
586 struct lu_target *tgt = class_exp2tgt(exp);
587 struct tgt_thread_info *tti = tgt_th_info(env);
593 if (unlikely(tgt == NULL)) {
594 CDEBUG(D_ERROR, "%s: No target for connected export\n",
595 class_exp2obd(exp)->obd_name);
599 if (tgt->lut_bottom->dd_rdonly)
602 th = dt_trans_create(env, tgt->lut_bottom);
607 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
609 ted->ted_lr_off, th);
613 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
617 mutex_lock(&ted->ted_lcd_lock);
620 * Until this operations will be committed the sync is needed
621 * for this export. This should be done _after_ starting the
622 * transaction so that many connecting clients will not bring
623 * server down with lots of sync writes.
625 rc = tgt_new_client_cb_add(th, exp);
627 /* can't add callback, do sync now */
630 spin_lock(&exp->exp_lock);
631 exp->exp_need_sync = 1;
632 spin_unlock(&exp->exp_lock);
635 tti->tti_off = ted->ted_lr_off;
636 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
638 mutex_unlock(&ted->ted_lcd_lock);
642 dt_trans_stop(env, tgt->lut_bottom, th);
643 CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
644 "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
645 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
650 static int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
652 struct tgt_thread_info *tti = tgt_th_info(env);
657 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
660 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
662 CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
663 "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
664 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
668 static int tgt_server_data_write(const struct lu_env *env,
669 struct lu_target *tgt, struct thandle *th)
671 struct tgt_thread_info *tti = tgt_th_info(env);
672 struct dt_object *dto;
679 lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
681 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
682 rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
684 CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
685 "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
686 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
692 * Update server data in last_rcvd
694 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
697 struct tgt_thread_info *tti = tgt_th_info(env);
704 "%s: mount_count is %llu, last_transno is %llu\n",
705 tgt->lut_lsd.lsd_uuid, obd2obt(tgt->lut_obd)->obt_mount_count,
706 tgt->lut_last_transno);
708 /* Always save latest transno to keep it fresh */
709 spin_lock(&tgt->lut_translock);
710 tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
711 spin_unlock(&tgt->lut_translock);
713 if (tgt->lut_bottom->dd_rdonly)
716 th = dt_trans_create(env, tgt->lut_bottom);
723 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
724 &tti->tti_buf, tti->tti_off, th);
728 rc = dt_trans_start(env, tgt->lut_bottom, th);
732 rc = tgt_server_data_write(env, tgt, th);
734 dt_trans_stop(env, tgt->lut_bottom, th);
736 CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
737 "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
738 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
741 EXPORT_SYMBOL(tgt_server_data_update);
743 static int tgt_truncate_last_rcvd(const struct lu_env *env,
744 struct lu_target *tgt, loff_t size)
746 struct dt_object *dt = tgt->lut_last_rcvd;
753 if (tgt->lut_bottom->dd_rdonly)
757 attr.la_valid = LA_SIZE;
759 th = dt_trans_create(env, tgt->lut_bottom);
762 rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
765 rc = dt_declare_attr_set(env, dt, &attr, th);
768 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
772 rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
774 rc = dt_attr_set(env, dt, &attr, th);
777 dt_trans_stop(env, tgt->lut_bottom, th);
782 static void tgt_client_epoch_update(const struct lu_env *env,
783 struct obd_export *exp)
785 struct lsd_client_data *lcd = exp->exp_target_data.ted_lcd;
786 struct lu_target *tgt = class_exp2tgt(exp);
788 LASSERT(tgt && tgt->lut_bottom);
789 /** VBR: set client last_epoch to current epoch */
790 if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
792 lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
793 tgt_client_data_update(env, exp);
797 * Update boot epoch when recovery ends
799 void tgt_boot_epoch_update(struct lu_target *tgt)
802 struct ptlrpc_request *req;
804 LIST_HEAD(client_list);
807 if (tgt->lut_obd->obd_stopping)
810 rc = lu_env_init(&env, LCT_DT_THREAD);
812 CERROR("%s: can't initialize environment: rc = %d\n",
813 tgt->lut_obd->obd_name, rc);
817 spin_lock(&tgt->lut_translock);
818 start_epoch = (tgt->lut_last_transno >> LR_EPOCH_BITS) + 1;
819 tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
820 tgt->lut_lsd.lsd_start_epoch = start_epoch;
821 spin_unlock(&tgt->lut_translock);
824 * The recovery is not yet finished and final queue can still be updated
825 * with resend requests. Move final list to separate one for processing
827 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
828 list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
829 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
832 * go through list of exports participated in recovery and
833 * set new epoch for them
835 list_for_each_entry(req, &client_list, rq_list) {
836 LASSERT(!req->rq_export->exp_delayed);
837 if (!req->rq_export->exp_vbr_failed)
838 tgt_client_epoch_update(&env, req->rq_export);
840 /** return list back at once */
841 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
842 list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
843 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
846 * Clear MULTI RPCS incompatibility flag if there is no multi-rpcs
847 * client in last_rcvd file
849 if (atomic_read(&tgt->lut_num_clients) == 0)
850 tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
852 /** update server epoch */
853 tgt_server_data_update(&env, tgt, 1);
858 * commit callback, need to update last_committed value
860 struct tgt_last_committed_callback {
861 struct dt_txn_commit_cb llcc_cb;
862 struct lu_target *llcc_tgt;
863 struct obd_export *llcc_exp;
867 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
868 struct dt_txn_commit_cb *cb, int err)
870 struct tgt_last_committed_callback *ccb;
872 ccb = container_of(cb, struct tgt_last_committed_callback, llcc_cb);
874 LASSERT(ccb->llcc_exp);
875 LASSERT(ccb->llcc_tgt != NULL);
876 LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
878 if (th->th_reserved_quota.lqi_space > 0) {
879 CDEBUG(D_QUOTA, "free quota %llu %llu\n",
880 th->th_reserved_quota.lqi_id.qid_gid,
881 th->th_reserved_quota.lqi_space);
883 /* env can be NULL for freeing reserved quota */
884 th->th_reserved_quota.lqi_space *= -1;
885 dt_reserve_or_free_quota(NULL, th->th_dev,
886 &th->th_reserved_quota);
889 /* error hit, don't update last committed to provide chance to
890 * replay data after fail */
894 /* Fast path w/o spinlock, if exp_last_committed was updated
895 * with higher transno, no need to take spinlock and check,
896 * also no need to update obd_last_committed. */
897 if (ccb->llcc_transno <= ccb->llcc_exp->exp_last_committed)
899 spin_lock(&ccb->llcc_tgt->lut_translock);
900 if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
901 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
903 if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
904 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
905 spin_unlock(&ccb->llcc_tgt->lut_translock);
907 ptlrpc_commit_replies(ccb->llcc_exp);
908 tgt_cancel_slc_locks(ccb->llcc_tgt, ccb->llcc_transno);
910 spin_unlock(&ccb->llcc_tgt->lut_translock);
913 CDEBUG(D_HA, "%s: transno %lld is committed\n",
914 ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
917 class_export_cb_put(ccb->llcc_exp);
922 * Add commit callback function, it returns a non-zero value to inform
923 * caller to use sync transaction if necessary.
925 static int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
926 struct obd_export *exp, __u64 transno)
928 struct tgt_last_committed_callback *ccb;
929 struct dt_txn_commit_cb *dcb;
937 ccb->llcc_exp = class_export_cb_get(exp);
938 ccb->llcc_transno = transno;
941 dcb->dcb_func = tgt_cb_last_committed;
942 INIT_LIST_HEAD(&dcb->dcb_linkage);
943 strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
945 rc = dt_trans_cb_add(th, dcb);
947 class_export_cb_put(exp);
951 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
952 /* report failure to force synchronous operation */
955 /* if exp_need_sync is set, return non-zero value to force
956 * a sync transaction. */
957 return rc ? rc : exp->exp_need_sync;
960 static int tgt_is_local_client(const struct lu_env *env,
961 struct obd_export *exp)
963 struct lu_target *tgt = class_exp2tgt(exp);
964 struct tgt_session_info *tsi = tgt_ses_info(env);
965 struct ptlrpc_request *req = tgt_ses_req(tsi);
967 if (exp_connect_flags(exp) & OBD_CONNECT_MDS ||
968 exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS)
970 if (tgt->lut_local_recovery)
974 if (!LNetIsPeerLocal(&req->rq_peer.nid))
981 * Add new client to the last_rcvd upon new connection.
983 * We use a bitmap to locate a free space in the last_rcvd file and initialize
986 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
988 struct tg_export_data *ted = &exp->exp_target_data;
989 struct lu_target *tgt = class_exp2tgt(exp);
994 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
995 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
998 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1001 if (tgt_is_local_client(env, exp)) {
1002 LCONSOLE_WARN("%s: local client %s w/o recovery\n",
1003 exp->exp_obd->obd_name, ted->ted_lcd->lcd_uuid);
1004 exp->exp_no_recovery = 1;
1008 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
1009 * there's no need for extra complication here
1011 idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
1013 if (idx >= LR_MAX_CLIENTS ||
1014 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
1015 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
1016 tgt->lut_obd->obd_name, idx);
1019 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
1020 idx = find_next_zero_bit(tgt->lut_client_bitmap,
1021 LR_MAX_CLIENTS, idx);
1025 ted->ted_lr_idx = idx;
1026 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
1027 idx * tgt->lut_lsd.lsd_client_size;
1029 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1031 if (tgt_is_multimodrpcs_client(exp)) {
1032 /* Set MULTI RPCS incompatibility flag to prevent previous
1033 * Lustre versions to mount a target with reply_data file */
1034 if (!(tgt->lut_lsd.lsd_feature_incompat &
1035 OBD_INCOMPAT_MULTI_RPCS)) {
1036 tgt->lut_lsd.lsd_feature_incompat |=
1037 OBD_INCOMPAT_MULTI_RPCS;
1038 rc = tgt_server_data_update(env, tgt, 1);
1040 CERROR("%s: unable to set MULTI RPCS "
1041 "incompatibility flag\n",
1042 exp->exp_obd->obd_name);
1047 /* assign client slot generation */
1048 ted->ted_lcd->lcd_generation =
1049 atomic_inc_return(&tgt->lut_client_generation);
1051 ted->ted_lcd->lcd_generation = 0;
1054 CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s' "
1056 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1057 ted->ted_lcd->lcd_uuid, ted->ted_lcd->lcd_generation);
1059 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
1062 rc = tgt_client_data_update(env, exp);
1064 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
1065 tgt->lut_obd->obd_name, idx, rc);
1069 if (tgt_is_multimodrpcs_client(exp))
1070 atomic_inc(&tgt->lut_num_clients);
1074 EXPORT_SYMBOL(tgt_client_new);
1076 /* Add an existing client to the MDS in-memory state based on
1077 * a client that was previously found in the last_rcvd file and
1078 * already has an assigned slot (idx >= 0).
1080 * It should not be possible to fail adding an existing client - otherwise
1081 * mdt_init_server_data() callsite needs to be fixed.
1083 int tgt_client_add(const struct lu_env *env, struct obd_export *exp, int idx)
1085 struct tg_export_data *ted = &exp->exp_target_data;
1086 struct lu_target *tgt = class_exp2tgt(exp);
1090 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
1091 LASSERTF(idx >= 0, "%d\n", idx);
1093 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
1094 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1097 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
1098 CERROR("%s: client %d: bit already set in bitmap!!\n",
1099 tgt->lut_obd->obd_name, idx);
1103 CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added, "
1105 tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid,
1106 ted->ted_lcd->lcd_generation);
1108 ted->ted_lr_idx = idx;
1109 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
1110 idx * tgt->lut_lsd.lsd_client_size;
1112 mutex_init(&ted->ted_lcd_lock);
1114 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1119 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
1121 struct tg_export_data *ted = &exp->exp_target_data;
1122 struct lu_target *tgt = class_exp2tgt(exp);
1127 LASSERT(ted->ted_lcd);
1129 if (unlikely(tgt == NULL)) {
1130 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1131 class_exp2obd(exp)->obd_name);
1135 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
1136 if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
1137 (char *)tgt->lut_obd->obd_uuid.uuid) ||
1138 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT ||
1139 exp->exp_no_recovery)
1142 /* Slot may be not yet assigned, use case is race between Client
1143 * reconnect and forced eviction */
1144 if (ted->ted_lr_idx < 0) {
1145 CWARN("%s: client with UUID '%s' not in bitmap\n",
1146 tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid);
1150 CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
1151 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1152 ted->ted_lcd->lcd_uuid);
1154 /* Clear the bit _after_ zeroing out the client so we don't
1155 race with filter_client_add and zero out new clients.*/
1156 if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
1157 CERROR("%s: client %u: bit already clear in bitmap!!\n",
1158 tgt->lut_obd->obd_name, ted->ted_lr_idx);
1162 /* Do not erase record for recoverable client. */
1163 if (exp->exp_flags & OBD_OPT_FAILOVER)
1166 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_DEL))
1169 /* Make sure the server's last_transno is up to date.
1170 * This should be done before zeroing client slot so last_transno will
1171 * be in server data or in client data in case of failure */
1172 rc = tgt_server_data_update(env, tgt, 0);
1174 CERROR("%s: failed to update server data, skip client %s "
1175 "zeroing, rc %d\n", tgt->lut_obd->obd_name,
1176 ted->ted_lcd->lcd_uuid, rc);
1180 /* Race between an eviction and a disconnection ?*/
1181 mutex_lock(&ted->ted_lcd_lock);
1182 if (ted->ted_lcd->lcd_uuid[0] == '\0') {
1183 mutex_unlock(&ted->ted_lcd_lock);
1187 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
1188 mutex_unlock(&ted->ted_lcd_lock);
1190 rc = tgt_client_data_update(env, exp);
1192 if (!rc && tgt_is_multimodrpcs_record(tgt, ted->ted_lcd))
1193 atomic_dec(&tgt->lut_num_clients);
1195 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
1196 "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
1197 tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
1198 ted->ted_lr_idx, ted->ted_lr_off, rc);
1201 EXPORT_SYMBOL(tgt_client_del);
1203 static void tgt_clean_by_tag(struct obd_export *exp, __u64 xid, __u16 tag)
1205 struct tg_export_data *ted = &exp->exp_target_data;
1206 struct lu_target *lut = class_exp2tgt(exp);
1207 struct tg_reply_data *trd, *tmp;
1212 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
1213 if (trd->trd_tag != tag)
1216 LASSERT(ergo(tgt_is_increasing_xid_client(exp),
1217 trd->trd_reply.lrd_xid <= xid));
1219 ted->ted_release_tag++;
1220 tgt_release_reply_data(lut, ted, trd);
1224 static int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
1225 struct tg_export_data *ted, struct tg_reply_data *trd,
1226 struct ptlrpc_request *req,
1227 struct thandle *th, bool update_lrd_file)
1229 struct lsd_reply_data *lrd;
1233 lrd = &trd->trd_reply;
1234 /* update export last transno */
1235 mutex_lock(&ted->ted_lcd_lock);
1236 if (lrd->lrd_transno > ted->ted_lcd->lcd_last_transno)
1237 ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
1238 mutex_unlock(&ted->ted_lcd_lock);
1241 /* find a empty slot */
1242 i = tgt_find_free_reply_slot(tgt);
1243 if (unlikely(i < 0)) {
1244 CERROR("%s: couldn't find a slot for reply data: "
1245 "rc = %d\n", tgt_name(tgt), i);
1250 if (update_lrd_file) {
1253 /* write reply data to disk */
1254 off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
1255 rc = tgt_reply_data_write(env, tgt, lrd, off, th);
1256 if (unlikely(rc != 0)) {
1257 CERROR("%s: can't update %s file: rc = %d\n",
1258 tgt_name(tgt), REPLY_DATA, rc);
1259 GOTO(free_slot, rc);
1263 trd->trd_index = TRD_INDEX_MEMORY;
1266 /* add reply data to target export's reply list */
1267 mutex_lock(&ted->ted_lcd_lock);
1269 int exclude = tgt_is_increasing_xid_client(req->rq_export) ?
1270 MSG_REPLAY : MSG_REPLAY|MSG_RESENT;
1272 if (req->rq_obsolete) {
1274 "drop reply data update for obsolete req xid=%llu,"
1275 "transno=%llu, tag=%hu\n", req->rq_xid,
1276 lrd->lrd_transno, trd->trd_tag);
1277 mutex_unlock(&ted->ted_lcd_lock);
1278 GOTO(free_slot, rc = -EBADR);
1281 if (!(lustre_msg_get_flags(req->rq_reqmsg) & exclude))
1282 tgt_clean_by_tag(req->rq_export, req->rq_xid,
1285 list_add(&trd->trd_list, &ted->ted_reply_list);
1286 ted->ted_reply_cnt++;
1287 if (ted->ted_reply_cnt > ted->ted_reply_max)
1288 ted->ted_reply_max = ted->ted_reply_cnt;
1289 mutex_unlock(&ted->ted_lcd_lock);
1291 CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
1292 "tag %hu, client gen %u, slot idx %d\n",
1293 trd, lrd->lrd_xid, lrd->lrd_transno,
1294 trd->trd_tag, lrd->lrd_client_gen, trd->trd_index);
1300 tgt_clear_reply_slot(tgt, trd->trd_index);
1304 int tgt_mk_reply_data(const struct lu_env *env,
1305 struct lu_target *tgt,
1306 struct tg_export_data *ted,
1307 struct ptlrpc_request *req,
1313 struct tg_reply_data *trd;
1314 struct lsd_reply_data *lrd;
1315 __u64 *pre_versions = NULL;
1317 struct tgt_session_info *tsi = NULL;
1320 if (unlikely(trd == NULL))
1324 tsi = tgt_ses_info(env);
1326 /* fill reply data information */
1327 lrd = &trd->trd_reply;
1328 lrd->lrd_transno = transno;
1330 lrd->lrd_xid = req->rq_xid;
1331 trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
1332 lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
1334 pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1335 lrd->lrd_result = th->th_result;
1338 LASSERT(env != NULL);
1339 LASSERT(tsi->tsi_xid != 0);
1341 lrd->lrd_xid = tsi->tsi_xid;
1342 lrd->lrd_result = tsi->tsi_result;
1343 lrd->lrd_client_gen = tsi->tsi_client_gen;
1346 lrd->lrd_data = opdata;
1348 trd->trd_pre_versions[0] = pre_versions[0];
1349 trd->trd_pre_versions[1] = pre_versions[1];
1350 trd->trd_pre_versions[2] = pre_versions[2];
1351 trd->trd_pre_versions[3] = pre_versions[3];
1354 if (tsi && tsi->tsi_open_obj)
1355 trd->trd_object = *lu_object_fid(&tsi->tsi_open_obj->do_lu);
1357 rc = tgt_add_reply_data(env, tgt, ted, trd, req,
1367 EXPORT_SYMBOL(tgt_mk_reply_data);
1370 * last_rcvd & last_committed update callbacks
1372 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
1373 struct dt_object *obj, __u64 opdata,
1374 struct thandle *th, struct ptlrpc_request *req)
1376 struct tgt_thread_info *tti = tgt_th_info(env);
1377 struct tgt_session_info *tsi = tgt_ses_info(env);
1378 struct obd_export *exp = tsi->tsi_exp;
1379 struct tg_export_data *ted;
1387 LASSERT(exp != NULL);
1388 ted = &exp->exp_target_data;
1390 /* Some clients don't support recovery, and they don't have last_rcvd
1392 * 1. lightweight clients.
1393 * 2. local clients on MDS which doesn't enable "localrecov".
1394 * 3. OFD connect may cause transaction before export has last_rcvd
1397 if (ted->ted_lr_idx < 0)
1401 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
1403 /* From update replay, tti_transno should be set already */
1404 LASSERT(tti->tti_transno != 0);
1406 spin_lock(&tgt->lut_translock);
1407 if (th->th_result != 0) {
1408 if (tti->tti_transno != 0) {
1409 CERROR("%s: replay transno %llu failed: rc = %d\n",
1410 tgt_name(tgt), tti->tti_transno, th->th_result);
1412 } else if (tti->tti_transno == 0) {
1413 tti->tti_transno = ++tgt->lut_last_transno;
1415 /* should be replay */
1416 if (tti->tti_transno > tgt->lut_last_transno)
1417 tgt->lut_last_transno = tti->tti_transno;
1419 spin_unlock(&tgt->lut_translock);
1421 /** VBR: set new versions */
1422 if (th->th_result == 0 && obj != NULL) {
1423 struct dt_object *dto = dt_object_locate(obj, th->th_dev);
1424 dt_version_set(env, dto, tti->tti_transno, th);
1427 /* filling reply data */
1428 CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
1429 tti->tti_transno, tgt->lut_obd->obd_last_committed);
1432 req->rq_transno = tti->tti_transno;
1433 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
1436 /* if can't add callback, do sync write */
1437 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp, tti->tti_transno);
1440 /* store transno in the last_rcvd header */
1441 spin_lock(&tgt->lut_translock);
1442 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
1443 tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
1444 spin_unlock(&tgt->lut_translock);
1445 /* Although current connection doesn't have slot
1446 * in the last_rcvd, we still want to maintain
1447 * the in-memory lsd_client_data structure in order to
1448 * properly handle reply reconstruction. */
1449 rc = tgt_server_data_write(env, tgt, th);
1451 spin_unlock(&tgt->lut_translock);
1453 } else if (ted->ted_lr_off == 0) {
1454 CERROR("%s: client idx %d has offset %lld\n",
1455 tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
1459 /* Target that supports multiple reply data */
1460 if (tgt_is_multimodrpcs_client(exp)) {
1461 return tgt_mk_reply_data(env, tgt, ted, req, opdata, th,
1462 !!(req != NULL), tti->tti_transno);
1465 /* Enough for update replay, let's return */
1469 mutex_lock(&ted->ted_lcd_lock);
1470 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1471 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
1472 transno_p = &ted->ted_lcd->lcd_last_close_transno;
1473 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
1474 ted->ted_lcd->lcd_last_close_result = th->th_result;
1476 /* VBR: save versions in last_rcvd for reconstruct. */
1477 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1480 ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
1481 ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
1482 ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
1483 ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
1485 transno_p = &ted->ted_lcd->lcd_last_transno;
1486 ted->ted_lcd->lcd_last_xid = req->rq_xid;
1487 ted->ted_lcd->lcd_last_result = th->th_result;
1488 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
1489 * see struct ldlm_reply->lock_policy_res1; */
1490 ted->ted_lcd->lcd_last_data = opdata;
1493 /* Update transno in slot only if non-zero number, i.e. no errors */
1494 if (likely(tti->tti_transno != 0)) {
1495 /* Don't overwrite bigger transaction number with lower one.
1496 * That is not sign of problem in all cases, but in any case
1497 * this value should be monotonically increased only. */
1498 if (*transno_p > tti->tti_transno) {
1499 if (!tgt->lut_no_reconstruct) {
1500 CERROR("%s: trying to overwrite bigger transno:"
1501 "on-disk: %llu, new: %llu replay: "
1502 "%d. See LU-617.\n", tgt_name(tgt),
1503 *transno_p, tti->tti_transno,
1504 req_is_replay(req));
1505 if (req_is_replay(req)) {
1506 spin_lock(&req->rq_export->exp_lock);
1507 req->rq_export->exp_vbr_failed = 1;
1508 spin_unlock(&req->rq_export->exp_lock);
1510 mutex_unlock(&ted->ted_lcd_lock);
1511 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
1514 *transno_p = tti->tti_transno;
1519 tti->tti_off = ted->ted_lr_off;
1520 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_RCVD_EIO))
1523 rc = tgt_client_data_write(env, tgt, ted->ted_lcd,
1526 mutex_unlock(&ted->ted_lcd_lock);
1530 mutex_unlock(&ted->ted_lcd_lock);
1535 * last_rcvd update for echo client simulation.
1536 * It updates last_rcvd client slot and version of object in
1537 * simple way but with all locks to simulate all drawbacks
1539 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
1540 struct lu_target *tgt,
1541 struct dt_object *obj,
1543 struct obd_export *exp)
1545 struct tgt_thread_info *tti = tgt_th_info(env);
1546 struct tg_export_data *ted = &exp->exp_target_data;
1551 tti->tti_transno = 0;
1553 spin_lock(&tgt->lut_translock);
1554 if (th->th_result == 0)
1555 tti->tti_transno = ++tgt->lut_last_transno;
1556 spin_unlock(&tgt->lut_translock);
1558 /** VBR: set new versions */
1559 if (th->th_result == 0 && obj != NULL)
1560 dt_version_set(env, obj, tti->tti_transno, th);
1562 /* if can't add callback, do sync write */
1563 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
1566 LASSERT(ted->ted_lr_off > 0);
1568 mutex_lock(&ted->ted_lcd_lock);
1569 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1570 ted->ted_lcd->lcd_last_transno = tti->tti_transno;
1571 ted->ted_lcd->lcd_last_result = th->th_result;
1573 tti->tti_off = ted->ted_lr_off;
1574 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1575 mutex_unlock(&ted->ted_lcd_lock);
1579 static int tgt_clients_data_init(const struct lu_env *env,
1580 struct lu_target *tgt,
1581 unsigned long last_size)
1583 struct obd_device *obd = tgt->lut_obd;
1584 struct lr_server_data *lsd = &tgt->lut_lsd;
1585 struct lsd_client_data *lcd = NULL;
1586 struct tg_export_data *ted;
1589 loff_t off = lsd->lsd_client_start;
1590 __u32 generation = 0;
1591 struct cfs_hash *hash = NULL;
1595 if (tgt->lut_bottom->dd_rdonly)
1598 BUILD_BUG_ON(offsetof(struct lsd_client_data, lcd_padding) +
1599 sizeof(lcd->lcd_padding) != LR_CLIENT_SIZE);
1605 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1607 GOTO(err_out, rc = -ENODEV);
1609 for (cl_idx = 0; off < last_size; cl_idx++) {
1610 struct obd_export *exp;
1613 /* Don't assume off is incremented properly by
1614 * read_record(), in case sizeof(*lcd)
1615 * isn't the same as fsd->lsd_client_size. */
1616 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
1617 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
1619 CERROR("%s: error reading last_rcvd %s idx %d off "
1620 "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
1623 break; /* read error shouldn't cause startup to fail */
1626 if (lcd->lcd_uuid[0] == '\0') {
1627 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
1632 last_transno = lcd_last_transno(lcd);
1634 /* These exports are cleaned up by disconnect, so they
1635 * need to be set up like real exports as connect does.
1637 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: %llu"
1638 " srv lr: %llu lx: %llu gen %u\n", lcd->lcd_uuid,
1639 cl_idx, last_transno, lsd->lsd_last_transno,
1640 lcd_last_xid(lcd), lcd->lcd_generation);
1642 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
1644 if (PTR_ERR(exp) == -EALREADY) {
1645 /* export already exists, zero out this one */
1646 CERROR("%s: Duplicate export %s!\n",
1647 tgt_name(tgt), lcd->lcd_uuid);
1650 GOTO(err_out, rc = PTR_ERR(exp));
1653 ted = &exp->exp_target_data;
1654 *ted->ted_lcd = *lcd;
1656 rc = tgt_client_add(env, exp, cl_idx);
1657 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
1658 /* VBR: set export last committed version */
1659 exp->exp_last_committed = last_transno;
1660 spin_lock(&exp->exp_lock);
1661 exp->exp_connecting = 0;
1662 exp->exp_in_recovery = 0;
1663 spin_unlock(&exp->exp_lock);
1664 atomic_inc(&obd->obd_max_recoverable_clients);
1666 if (tgt_is_multimodrpcs_record(tgt, lcd)) {
1667 atomic_inc(&tgt->lut_num_clients);
1669 /* compute the highest valid client generation */
1670 generation = max(generation, lcd->lcd_generation);
1671 /* fill client_generation <-> export hash table */
1672 rc = cfs_hash_add_unique(hash, &lcd->lcd_generation,
1673 &exp->exp_gen_hash);
1675 CERROR("%s: duplicate export for client "
1677 tgt_name(tgt), lcd->lcd_generation);
1678 class_export_put(exp);
1683 class_export_put(exp);
1685 rc = rev_import_init(exp);
1687 class_unlink_export(exp);
1691 /* Need to check last_rcvd even for duplicated exports. */
1692 CDEBUG(D_OTHER, "client at idx %d has last_transno = %llu\n",
1693 cl_idx, last_transno);
1695 spin_lock(&tgt->lut_translock);
1696 tgt->lut_last_transno = max(last_transno,
1697 tgt->lut_last_transno);
1698 spin_unlock(&tgt->lut_translock);
1701 /* record highest valid client generation */
1702 atomic_set(&tgt->lut_client_generation, generation);
1706 cfs_hash_putref(hash);
1711 struct server_compat_data {
1718 static struct server_compat_data tgt_scd[] = {
1719 [LDD_F_SV_TYPE_MDT] = {
1720 .rocompat = OBD_ROCOMPAT_LOVOBJID,
1721 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1722 OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
1723 OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI |
1724 OBD_INCOMPAT_MULTI_RPCS,
1725 .rocinit = OBD_ROCOMPAT_LOVOBJID,
1726 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1727 OBD_INCOMPAT_MULTI_OI,
1729 [LDD_F_SV_TYPE_OST] = {
1730 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
1731 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
1733 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
1734 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1738 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1740 struct tgt_thread_info *tti = tgt_th_info(env);
1741 struct lr_server_data *lsd = &tgt->lut_lsd;
1742 unsigned long last_rcvd_size;
1746 rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1750 last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1752 /* ensure padding in the struct is the correct size */
1753 BUILD_BUG_ON(offsetof(struct lr_server_data, lsd_padding) +
1754 sizeof(lsd->lsd_padding) != LR_SERVER_SIZE);
1756 rc = server_name2index(tgt_name(tgt), &index, NULL);
1758 CERROR("%s: Can not get index from name: rc = %d\n",
1762 /* server_name2index() returns type */
1764 if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1765 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1769 /* last_rcvd on OST doesn't provide reconstruct support because there
1770 * may be up to 8 in-flight write requests per single slot in
1771 * last_rcvd client data
1773 tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1775 if (last_rcvd_size == 0) {
1776 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1778 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1779 sizeof(lsd->lsd_uuid));
1780 lsd->lsd_last_transno = 0;
1781 lsd->lsd_mount_count = 0;
1782 lsd->lsd_server_size = LR_SERVER_SIZE;
1783 lsd->lsd_client_start = LR_CLIENT_START;
1784 lsd->lsd_client_size = LR_CLIENT_SIZE;
1785 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1786 lsd->lsd_osd_index = index;
1787 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1788 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1790 rc = tgt_server_data_read(env, tgt);
1792 CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1796 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1797 if (tgt->lut_bottom->dd_rdonly) {
1798 /* Such difference may be caused by mounting
1799 * up snapshot with new fsname under rd_only
1800 * mode. But even if it was NOT, it will not
1801 * damage the system because of "rd_only". */
1802 memcpy(lsd->lsd_uuid,
1803 tgt->lut_obd->obd_uuid.uuid,
1804 sizeof(lsd->lsd_uuid));
1806 LCONSOLE_ERROR_MSG(0x157, "Trying to start "
1807 "OBD %s using the wrong "
1808 "disk %s. Were the /dev/ "
1809 "assignments rearranged?\n",
1810 tgt->lut_obd->obd_uuid.uuid,
1816 if (lsd->lsd_osd_index != index) {
1817 LCONSOLE_ERROR_MSG(0x157,
1818 "%s: index %d in last rcvd is different with the index %d in config log, It might be disk corruption!\n",
1820 lsd->lsd_osd_index, index);
1825 if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1826 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1828 lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1832 if (type == LDD_F_SV_TYPE_MDT)
1833 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1835 if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1836 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1838 lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1841 /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1842 if (type == LDD_F_SV_TYPE_MDT &&
1843 !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1844 if (last_rcvd_size > lsd->lsd_client_start) {
1845 LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1846 "remove all clients for interop needs\n",
1848 rc = tgt_truncate_last_rcvd(env, tgt,
1849 lsd->lsd_client_start);
1852 last_rcvd_size = lsd->lsd_client_start;
1854 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1855 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1858 spin_lock(&tgt->lut_translock);
1859 tgt->lut_last_transno = lsd->lsd_last_transno;
1860 spin_unlock(&tgt->lut_translock);
1862 lsd->lsd_mount_count++;
1864 CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1865 CDEBUG(D_INODE, "%s: server last_transno: %llu\n",
1866 tgt_name(tgt), tgt->lut_last_transno);
1867 CDEBUG(D_INODE, "%s: server mount_count: %llu\n",
1868 tgt_name(tgt), lsd->lsd_mount_count);
1869 CDEBUG(D_INODE, "%s: server data size: %u\n",
1870 tgt_name(tgt), lsd->lsd_server_size);
1871 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1872 tgt_name(tgt), lsd->lsd_client_start);
1873 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1874 tgt_name(tgt), lsd->lsd_client_size);
1875 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1876 tgt_name(tgt), last_rcvd_size);
1877 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1878 tgt_name(tgt), lsd->lsd_subdir_count);
1879 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1880 last_rcvd_size <= lsd->lsd_client_start ? 0 :
1881 (last_rcvd_size - lsd->lsd_client_start) /
1882 lsd->lsd_client_size);
1883 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1885 if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1886 lsd->lsd_client_size == 0) {
1887 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1891 if (!tgt->lut_obd->obd_replayable)
1892 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1894 rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1896 GOTO(err_client, rc);
1898 spin_lock(&tgt->lut_translock);
1899 /* obd_last_committed is used for compatibility
1900 * with other lustre recovery code */
1901 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1902 spin_unlock(&tgt->lut_translock);
1904 obd2obt(tgt->lut_obd)->obt_mount_count = lsd->lsd_mount_count;
1905 obd2obt(tgt->lut_obd)->obt_instance = (__u32)lsd->lsd_mount_count;
1907 /* save it, so mount count and last_transno is current */
1908 rc = tgt_server_data_update(env, tgt, 0);
1910 GOTO(err_client, rc);
1915 class_disconnect_exports(tgt->lut_obd);
1919 /* add credits for last_rcvd update */
1920 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1923 struct lu_target *tgt = cookie;
1924 struct tgt_session_info *tsi;
1925 struct tgt_thread_info *tti = tgt_th_info(env);
1926 struct dt_object *dto;
1929 /* For readonly case, the caller should have got failure
1930 * when start the transaction. If the logic comes here,
1931 * there must be something wrong. */
1932 if (unlikely(tgt->lut_bottom->dd_rdonly)) {
1937 /* if there is no session, then this transaction is not result of
1938 * request processing but some local operation */
1939 if (env->le_ses == NULL)
1942 LASSERT(tgt->lut_last_rcvd);
1943 tsi = tgt_ses_info(env);
1944 /* OFD may start transaction without export assigned */
1945 if (tsi->tsi_exp == NULL)
1948 if (tgt_is_multimodrpcs_client(tsi->tsi_exp)) {
1950 * Use maximum possible file offset for declaration to ensure
1951 * ZFS will reserve enough credits for a write anywhere in this
1952 * file, since we don't know where in the file the write will be
1953 * because a replay slot has not been assigned. This should be
1954 * replaced by dmu_tx_hold_append() when available.
1956 tti->tti_buf.lb_buf = NULL;
1957 tti->tti_buf.lb_len = sizeof(struct lsd_reply_data);
1958 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
1959 rc = dt_declare_record_write(env, dto, &tti->tti_buf, -1, th);
1963 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
1965 tti->tti_off = tsi->tsi_exp->exp_target_data.ted_lr_off;
1966 rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1972 if (tsi->tsi_vbr_obj != NULL &&
1973 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1974 dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
1975 rc = dt_declare_version_set(env, dto, th);
1981 /* Update last_rcvd records with latests transaction data */
1982 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1985 struct lu_target *tgt = cookie;
1986 struct tgt_session_info *tsi;
1987 struct tgt_thread_info *tti = tgt_th_info(env);
1988 struct dt_object *obj = NULL;
1992 if (env->le_ses == NULL)
1995 tsi = tgt_ses_info(env);
1996 /* OFD may start transaction without export assigned */
1997 if (tsi->tsi_exp == NULL)
2000 echo_client = (tgt_ses_req(tsi) == NULL && tsi->tsi_xid == 0);
2002 if (tsi->tsi_has_trans && !echo_client) {
2003 if (!tsi->tsi_mult_trans) {
2004 CDEBUG(D_HA, "More than one transaction %llu\n",
2007 * if RPC handler sees unexpected multiple last_rcvd
2008 * updates with transno, then it is better to return
2009 * the latest transaction number to the client.
2010 * In that case replay may fail if part of operation
2011 * was committed and can't be re-applied easily. But
2012 * that is better than report the first transno, in
2013 * which case partially committed operation would be
2014 * considered as finished so never replayed causing
2018 /* we need new transno to be assigned */
2019 tti->tti_transno = 0;
2023 tsi->tsi_has_trans++;
2025 if (tsi->tsi_vbr_obj != NULL &&
2026 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
2027 obj = tsi->tsi_vbr_obj;
2030 if (unlikely(echo_client)) /* echo client special case */
2031 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
2034 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
2039 int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
2041 struct tgt_thread_info *tti = tgt_th_info(env);
2042 struct lsd_reply_data *lrd = &tti->tti_lrd;
2043 unsigned long reply_data_size;
2045 struct lsd_reply_header *lrh = NULL;
2046 struct tg_reply_data *trd = NULL;
2049 struct cfs_hash *hash = NULL;
2050 struct obd_export *exp;
2051 struct tg_export_data *ted;
2052 int reply_data_recovered = 0;
2054 rc = dt_attr_get(env, tgt->lut_reply_data, &tti->tti_attr);
2057 reply_data_size = (unsigned long)tti->tti_attr.la_size;
2061 GOTO(out, rc = -ENOMEM);
2063 if (reply_data_size == 0) {
2064 CDEBUG(D_INFO, "%s: new reply_data file, initializing\n",
2066 lrh->lrh_magic = LRH_MAGIC;
2067 lrh->lrh_header_size = sizeof(struct lsd_reply_header);
2068 lrh->lrh_reply_size = sizeof(struct lsd_reply_data);
2069 rc = tgt_reply_header_write(env, tgt, lrh);
2071 CERROR("%s: error writing %s: rc = %d\n",
2072 tgt_name(tgt), REPLY_DATA, rc);
2076 rc = tgt_reply_header_read(env, tgt, lrh);
2078 CERROR("%s: error reading %s: rc = %d\n",
2079 tgt_name(tgt), REPLY_DATA, rc);
2082 if (lrh->lrh_magic != LRH_MAGIC ||
2083 lrh->lrh_header_size != sizeof(struct lsd_reply_header) ||
2084 lrh->lrh_reply_size != sizeof(struct lsd_reply_data)) {
2085 CERROR("%s: invalid header in %s\n",
2086 tgt_name(tgt), REPLY_DATA);
2087 GOTO(out, rc = -EINVAL);
2090 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
2092 GOTO(out, rc = -ENODEV);
2096 GOTO(out, rc = -ENOMEM);
2098 /* Load reply_data from disk */
2099 for (idx = 0, off = sizeof(struct lsd_reply_header);
2100 off < reply_data_size;
2101 idx++, off += sizeof(struct lsd_reply_data)) {
2102 rc = tgt_reply_data_read(env, tgt, lrd, off);
2104 CERROR("%s: error reading %s: rc = %d\n",
2105 tgt_name(tgt), REPLY_DATA, rc);
2109 exp = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
2111 /* old reply data from a disconnected client */
2114 ted = &exp->exp_target_data;
2115 mutex_lock(&ted->ted_lcd_lock);
2117 /* create in-memory reply_data and link it to
2118 * target export's reply list */
2119 rc = tgt_set_reply_slot(tgt, idx);
2121 mutex_unlock(&ted->ted_lcd_lock);
2124 trd->trd_reply = *lrd;
2125 trd->trd_pre_versions[0] = 0;
2126 trd->trd_pre_versions[1] = 0;
2127 trd->trd_pre_versions[2] = 0;
2128 trd->trd_pre_versions[3] = 0;
2129 trd->trd_index = idx;
2131 fid_zero(&trd->trd_object);
2132 list_add(&trd->trd_list, &ted->ted_reply_list);
2133 ted->ted_reply_cnt++;
2134 if (ted->ted_reply_cnt > ted->ted_reply_max)
2135 ted->ted_reply_max = ted->ted_reply_cnt;
2137 CDEBUG(D_HA, "%s: restore reply %p: xid %llu, "
2138 "transno %llu, client gen %u, slot idx %d\n",
2139 tgt_name(tgt), trd, lrd->lrd_xid,
2140 lrd->lrd_transno, lrd->lrd_client_gen,
2143 /* update export last committed transation */
2144 exp->exp_last_committed = max(exp->exp_last_committed,
2146 /* Update lcd_last_transno as well for check in
2147 * tgt_release_reply_data() or the latest client
2148 * transno can be lost.
2150 ted->ted_lcd->lcd_last_transno =
2151 max(ted->ted_lcd->lcd_last_transno,
2152 exp->exp_last_committed);
2154 mutex_unlock(&ted->ted_lcd_lock);
2155 class_export_put(exp);
2157 /* update target last committed transaction */
2158 spin_lock(&tgt->lut_translock);
2159 tgt->lut_last_transno = max(tgt->lut_last_transno,
2161 spin_unlock(&tgt->lut_translock);
2163 reply_data_recovered++;
2167 GOTO(out, rc = -ENOMEM);
2169 CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
2170 tgt_name(tgt), reply_data_recovered);
2173 spin_lock(&tgt->lut_translock);
2174 /* obd_last_committed is used for compatibility
2175 * with other lustre recovery code */
2176 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
2177 spin_unlock(&tgt->lut_translock);
2183 cfs_hash_putref(hash);
2191 static int tgt_check_lookup_req(struct ptlrpc_request *req, int lookup,
2192 struct tg_reply_data *trd)
2194 struct tg_export_data *ted = &req->rq_export->exp_target_data;
2195 struct lu_target *lut = class_exp2tgt(req->rq_export);
2196 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
2198 struct tg_reply_data *reply;
2199 bool check_increasing;
2204 check_increasing = tgt_is_increasing_xid_client(req->rq_export) &&
2205 !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
2206 if (!lookup && !check_increasing)
2209 list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
2210 if (lookup && reply->trd_reply.lrd_xid == req->rq_xid) {
2215 } else if (check_increasing && reply->trd_tag == tag &&
2216 reply->trd_reply.lrd_xid > req->rq_xid) {
2218 CERROR("%s: busy tag=%u req_xid=%llu, trd=%p: xid=%llu transno=%llu client_gen=%u slot_idx=%d: rc = %d\n",
2219 tgt_name(lut), tag, req->rq_xid, trd,
2220 reply->trd_reply.lrd_xid,
2221 reply->trd_reply.lrd_transno,
2222 reply->trd_reply.lrd_client_gen,
2223 reply->trd_index, rc);
2231 /* Look for a reply data matching specified request @req
2232 * A copy is returned in @trd if the pointer is not NULL
2234 int tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
2236 struct tg_export_data *ted = &req->rq_export->exp_target_data;
2238 bool not_replay = !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
2240 mutex_lock(&ted->ted_lcd_lock);
2241 if (not_replay && req->rq_xid <= req->rq_export->exp_last_xid) {
2242 /* A check for the last_xid is needed here in case there is
2243 * no reply data is left in the list. It may happen if another
2244 * RPC on another slot increased the last_xid between our
2245 * process_req_last_xid & tgt_lookup_reply calls */
2248 found = tgt_check_lookup_req(req, 1, trd);
2250 mutex_unlock(&ted->ted_lcd_lock);
2252 CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d last_xid %llu\n",
2253 tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid, found,
2254 req->rq_export->exp_last_xid);
2258 EXPORT_SYMBOL(tgt_lookup_reply);
2260 int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
2262 struct tg_export_data *ted = &exp->exp_target_data;
2263 struct lu_target *lut = class_exp2tgt(exp);
2264 struct tg_reply_data *trd, *tmp;
2267 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
2268 if (trd->trd_reply.lrd_xid > rcvd_xid)
2270 ted->ted_release_xid++;
2271 tgt_release_reply_data(lut, ted, trd);
2277 int tgt_handle_tag(struct ptlrpc_request *req)
2279 return tgt_check_lookup_req(req, 0, NULL);