4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Lustre Unified Target
37 * These are common function to work with last_received file
39 * Author: Mikhail Pershin <mike.pershin@intel.com>
42 #include <obd_class.h>
43 #include <lustre_fid.h>
45 #include "tgt_internal.h"
48 /* Allocate a bitmap for a chunk of reply data slots */
49 static int tgt_bitmap_chunk_alloc(struct lu_target *lut, int chunk)
53 OBD_ALLOC(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) * sizeof(long));
57 spin_lock(&lut->lut_client_bitmap_lock);
59 if (lut->lut_reply_bitmap[chunk] != NULL) {
60 /* someone else already allocated the bitmap for this chunk */
61 spin_unlock(&lut->lut_client_bitmap_lock);
62 OBD_FREE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
67 lut->lut_reply_bitmap[chunk] = bm;
69 spin_unlock(&lut->lut_client_bitmap_lock);
74 /* Look for an available reply data slot in the bitmap
76 * Allocate bitmap chunk when first used
77 * XXX algo could be improved if this routine limits performance
79 static int tgt_find_free_reply_slot(struct lu_target *lut)
86 for (chunk = 0; chunk < LUT_REPLY_SLOTS_MAX_CHUNKS; chunk++) {
87 /* allocate the bitmap chunk if necessary */
88 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
89 rc = tgt_bitmap_chunk_alloc(lut, chunk);
93 bmp = lut->lut_reply_bitmap[chunk];
95 /* look for an available slot in this chunk */
97 b = find_first_zero_bit(bmp, LUT_REPLY_SLOTS_PER_CHUNK);
98 if (b >= LUT_REPLY_SLOTS_PER_CHUNK)
102 if (test_and_set_bit(b, bmp) == 0)
103 return chunk * LUT_REPLY_SLOTS_PER_CHUNK + b;
110 /* Mark the reply data slot @idx 'used' in the corresponding bitmap chunk
112 * Allocate the bitmap chunk if necessary
114 static int tgt_set_reply_slot(struct lu_target *lut, int idx)
120 chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
121 b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
123 LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
124 LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
126 /* allocate the bitmap chunk if necessary */
127 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
128 rc = tgt_bitmap_chunk_alloc(lut, chunk);
133 /* mark the slot 'used' in this chunk */
134 if (test_and_set_bit(b, lut->lut_reply_bitmap[chunk]) != 0) {
135 CERROR("%s: slot %d already set in bitmap\n",
144 /* Mark the reply data slot @idx 'unused' in the corresponding bitmap chunk
147 static int tgt_clear_reply_slot(struct lu_target *lut, int idx)
152 chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
153 b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
155 LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
156 LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
158 if (test_and_clear_bit(b, lut->lut_reply_bitmap[chunk]) == 0) {
159 CERROR("%s: slot %d already clear in bitmap\n",
168 /* Read header of reply_data file of target @tgt into structure @lrh */
169 static int tgt_reply_header_read(const struct lu_env *env,
170 struct lu_target *tgt,
171 struct lsd_reply_header *lrh)
174 struct lsd_reply_header buf;
175 struct tgt_thread_info *tti = tgt_th_info(env);
178 tti->tti_buf.lb_buf = &buf;
179 tti->tti_buf.lb_len = sizeof(buf);
181 rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
186 lrh->lrh_magic = le32_to_cpu(buf.lrh_magic);
187 lrh->lrh_header_size = le32_to_cpu(buf.lrh_header_size);
188 lrh->lrh_reply_size = le32_to_cpu(buf.lrh_reply_size);
190 CDEBUG(D_HA, "%s: read %s header. magic=0x%08x "
191 "header_size=%d reply_size=%d\n",
192 tgt->lut_obd->obd_name, REPLY_DATA,
193 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
198 /* Write header into replay_data file of target @tgt from structure @lrh */
199 static int tgt_reply_header_write(const struct lu_env *env,
200 struct lu_target *tgt,
201 struct lsd_reply_header *lrh)
204 struct lsd_reply_header buf;
205 struct tgt_thread_info *tti = tgt_th_info(env);
207 struct dt_object *dto;
209 CDEBUG(D_HA, "%s: write %s header. magic=0x%08x "
210 "header_size=%d reply_size=%d\n",
211 tgt->lut_obd->obd_name, REPLY_DATA,
212 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
214 buf.lrh_magic = cpu_to_le32(lrh->lrh_magic);
215 buf.lrh_header_size = cpu_to_le32(lrh->lrh_header_size);
216 buf.lrh_reply_size = cpu_to_le32(lrh->lrh_reply_size);
218 th = dt_trans_create(env, tgt->lut_bottom);
224 tti->tti_buf.lb_buf = &buf;
225 tti->tti_buf.lb_len = sizeof(buf);
227 rc = dt_declare_record_write(env, tgt->lut_reply_data,
228 &tti->tti_buf, tti->tti_off, th);
232 rc = dt_trans_start(env, tgt->lut_bottom, th);
236 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
237 rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
239 dt_trans_stop(env, tgt->lut_bottom, th);
243 /* Write the reply data @lrd into reply_data file of target @tgt
246 static int tgt_reply_data_write(const struct lu_env *env, struct lu_target *tgt,
247 struct lsd_reply_data *lrd, loff_t off,
250 struct tgt_thread_info *tti = tgt_th_info(env);
251 struct dt_object *dto;
252 struct lsd_reply_data *buf = &tti->tti_lrd;
254 lrd->lrd_result = ptlrpc_status_hton(lrd->lrd_result);
256 buf->lrd_transno = cpu_to_le64(lrd->lrd_transno);
257 buf->lrd_xid = cpu_to_le64(lrd->lrd_xid);
258 buf->lrd_data = cpu_to_le64(lrd->lrd_data);
259 buf->lrd_result = cpu_to_le32(lrd->lrd_result);
260 buf->lrd_client_gen = cpu_to_le32(lrd->lrd_client_gen);
262 lrd->lrd_result = ptlrpc_status_ntoh(lrd->lrd_result);
265 tti->tti_buf.lb_buf = buf;
266 tti->tti_buf.lb_len = sizeof(*buf);
268 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
269 return dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
272 /* Read the reply data from reply_data file of target @tgt at offset @off
273 * into structure @lrd
275 static int tgt_reply_data_read(const struct lu_env *env, struct lu_target *tgt,
276 struct lsd_reply_data *lrd, loff_t off)
279 struct tgt_thread_info *tti = tgt_th_info(env);
280 struct lsd_reply_data *buf = &tti->tti_lrd;
283 tti->tti_buf.lb_buf = buf;
284 tti->tti_buf.lb_len = sizeof(*buf);
286 rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
291 lrd->lrd_transno = le64_to_cpu(buf->lrd_transno);
292 lrd->lrd_xid = le64_to_cpu(buf->lrd_xid);
293 lrd->lrd_data = le64_to_cpu(buf->lrd_data);
294 lrd->lrd_result = le32_to_cpu(buf->lrd_result);
295 lrd->lrd_client_gen = le32_to_cpu(buf->lrd_client_gen);
301 /* Free the in-memory reply data structure @trd and release
302 * the corresponding slot in the reply_data file of target @lut
303 * Called with ted_lcd_lock held
305 static void tgt_free_reply_data(struct lu_target *lut,
306 struct tg_export_data *ted,
307 struct tg_reply_data *trd)
309 CDEBUG(D_TRACE, "%s: free reply data %p: xid %llu, transno %llu, "
310 "client gen %u, slot idx %d\n",
311 tgt_name(lut), trd, trd->trd_reply.lrd_xid,
312 trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
315 LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
317 list_del(&trd->trd_list);
318 ted->ted_reply_cnt--;
319 tgt_clear_reply_slot(lut, trd->trd_index);
323 /* Release the reply data @trd from target @lut
324 * The reply data with the highest transno for this export
325 * is retained to ensure correctness of target recovery
326 * Called with ted_lcd_lock held
328 static void tgt_release_reply_data(struct lu_target *lut,
329 struct tg_export_data *ted,
330 struct tg_reply_data *trd)
332 CDEBUG(D_TRACE, "%s: release reply data %p: xid %llu, transno %llu, "
333 "client gen %u, slot idx %d\n",
334 tgt_name(lut), trd, trd->trd_reply.lrd_xid,
335 trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
338 LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
340 /* Do not free the reply data corresponding to the
341 * highest transno of this export.
342 * This ensures on-disk reply data is kept and
343 * last committed transno can be restored from disk in case
346 if (trd->trd_reply.lrd_transno == ted->ted_lcd->lcd_last_transno) {
347 /* free previous retained reply */
348 if (ted->ted_reply_last != NULL)
349 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
350 /* retain the reply */
351 list_del_init(&trd->trd_list);
352 ted->ted_reply_last = trd;
354 tgt_free_reply_data(lut, ted, trd);
358 static inline struct lu_buf *tti_buf_lsd(struct tgt_thread_info *tti)
360 tti->tti_buf.lb_buf = &tti->tti_lsd;
361 tti->tti_buf.lb_len = sizeof(tti->tti_lsd);
362 return &tti->tti_buf;
365 static inline struct lu_buf *tti_buf_lcd(struct tgt_thread_info *tti)
367 tti->tti_buf.lb_buf = &tti->tti_lcd;
368 tti->tti_buf.lb_len = sizeof(tti->tti_lcd);
369 return &tti->tti_buf;
373 * Allocate in-memory data for client slot related to export.
375 int tgt_client_alloc(struct obd_export *exp)
378 LASSERT(exp != exp->exp_obd->obd_self_export);
380 OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
381 if (exp->exp_target_data.ted_lcd == NULL)
383 /* Mark that slot is not yet valid, 0 doesn't work here */
384 exp->exp_target_data.ted_lr_idx = -1;
385 INIT_LIST_HEAD(&exp->exp_target_data.ted_reply_list);
388 EXPORT_SYMBOL(tgt_client_alloc);
391 * Free in-memory data for client slot related to export.
393 void tgt_client_free(struct obd_export *exp)
395 struct tg_export_data *ted = &exp->exp_target_data;
396 struct lu_target *lut = class_exp2tgt(exp);
397 struct tg_reply_data *trd, *tmp;
399 LASSERT(exp != exp->exp_obd->obd_self_export);
401 /* free reply data */
402 mutex_lock(&ted->ted_lcd_lock);
403 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
404 tgt_release_reply_data(lut, ted, trd);
406 if (ted->ted_reply_last != NULL) {
407 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
408 ted->ted_reply_last = NULL;
410 mutex_unlock(&ted->ted_lcd_lock);
412 if (!hlist_unhashed(&exp->exp_gen_hash))
413 cfs_hash_del(exp->exp_obd->obd_gen_hash,
414 &ted->ted_lcd->lcd_generation,
417 OBD_FREE_PTR(ted->ted_lcd);
420 /* Slot may be not yet assigned */
421 if (ted->ted_lr_idx < 0)
423 /* Clear bit when lcd is freed */
424 LASSERT(lut && lut->lut_client_bitmap);
425 if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
426 CERROR("%s: client %u bit already clear in bitmap\n",
427 exp->exp_obd->obd_name, ted->ted_lr_idx);
431 if (tgt_is_multimodrpcs_client(exp) && !exp->exp_obd->obd_stopping)
432 atomic_dec(&lut->lut_num_clients);
434 EXPORT_SYMBOL(tgt_client_free);
436 int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
437 struct lsd_client_data *lcd, loff_t *off, int index)
439 struct tgt_thread_info *tti = tgt_th_info(env);
443 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
445 check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
446 lcd_le_to_cpu(&tti->tti_lcd, lcd);
447 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
448 lcd->lcd_last_close_result =
449 ptlrpc_status_ntoh(lcd->lcd_last_close_result);
452 CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = "LPU64
453 ", last_xid = "LPU64", last_result = %u, last_data = %u, "
454 "last_close_transno = "LPU64", last_close_xid = "LPU64", "
455 "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
456 *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
457 lcd->lcd_last_result, lcd->lcd_last_data,
458 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
459 lcd->lcd_last_close_result, rc);
463 int tgt_client_data_write(const struct lu_env *env, struct lu_target *tgt,
464 struct lsd_client_data *lcd, loff_t *off,
467 struct tgt_thread_info *tti = tgt_th_info(env);
468 struct dt_object *dto;
470 lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
471 lcd->lcd_last_close_result =
472 ptlrpc_status_hton(lcd->lcd_last_close_result);
473 lcd_cpu_to_le(lcd, &tti->tti_lcd);
476 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
477 return dt_record_write(env, dto, &tti->tti_buf, off, th);
481 * Update client data in last_rcvd
483 static int tgt_client_data_update(const struct lu_env *env,
484 struct obd_export *exp)
486 struct tg_export_data *ted = &exp->exp_target_data;
487 struct lu_target *tgt = class_exp2tgt(exp);
488 struct tgt_thread_info *tti = tgt_th_info(env);
494 if (unlikely(tgt == NULL)) {
495 CDEBUG(D_ERROR, "%s: No target for connected export\n",
496 class_exp2obd(exp)->obd_name);
500 th = dt_trans_create(env, tgt->lut_bottom);
505 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
507 ted->ted_lr_off, th);
511 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
515 * Until this operations will be committed the sync is needed
516 * for this export. This should be done _after_ starting the
517 * transaction so that many connecting clients will not bring
518 * server down with lots of sync writes.
520 rc = tgt_new_client_cb_add(th, exp);
522 /* can't add callback, do sync now */
525 spin_lock(&exp->exp_lock);
526 exp->exp_need_sync = 1;
527 spin_unlock(&exp->exp_lock);
530 tti->tti_off = ted->ted_lr_off;
531 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
534 dt_trans_stop(env, tgt->lut_bottom, th);
535 CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
536 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
537 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
542 int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
544 struct tgt_thread_info *tti = tgt_th_info(env);
549 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
552 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
554 CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
555 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
556 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
560 int tgt_server_data_write(const struct lu_env *env, struct lu_target *tgt,
563 struct tgt_thread_info *tti = tgt_th_info(env);
564 struct dt_object *dto;
571 lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
573 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
574 rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
576 CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
577 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
578 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
584 * Update server data in last_rcvd
586 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
589 struct tgt_thread_info *tti = tgt_th_info(env);
596 "%s: mount_count is "LPU64", last_transno is "LPU64"\n",
597 tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
598 tgt->lut_last_transno);
600 /* Always save latest transno to keep it fresh */
601 spin_lock(&tgt->lut_translock);
602 tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
603 spin_unlock(&tgt->lut_translock);
605 th = dt_trans_create(env, tgt->lut_bottom);
612 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
613 &tti->tti_buf, tti->tti_off, th);
617 rc = dt_trans_start(env, tgt->lut_bottom, th);
621 rc = tgt_server_data_write(env, tgt, th);
623 dt_trans_stop(env, tgt->lut_bottom, th);
625 CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
626 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
627 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
630 EXPORT_SYMBOL(tgt_server_data_update);
632 int tgt_truncate_last_rcvd(const struct lu_env *env, struct lu_target *tgt,
635 struct dt_object *dt = tgt->lut_last_rcvd;
643 attr.la_valid = LA_SIZE;
645 th = dt_trans_create(env, tgt->lut_bottom);
648 rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
651 rc = dt_declare_attr_set(env, dt, &attr, th);
654 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
658 rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
660 rc = dt_attr_set(env, dt, &attr, th);
663 dt_trans_stop(env, tgt->lut_bottom, th);
668 static void tgt_client_epoch_update(const struct lu_env *env,
669 struct obd_export *exp)
671 struct lsd_client_data *lcd = exp->exp_target_data.ted_lcd;
672 struct lu_target *tgt = class_exp2tgt(exp);
674 LASSERT(tgt && tgt->lut_bottom);
675 /** VBR: set client last_epoch to current epoch */
676 if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
678 lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
679 tgt_client_data_update(env, exp);
683 * Update boot epoch when recovery ends
685 void tgt_boot_epoch_update(struct lu_target *tgt)
688 struct ptlrpc_request *req;
690 struct list_head client_list;
693 if (tgt->lut_obd->obd_stopping)
696 rc = lu_env_init(&env, LCT_DT_THREAD);
698 CERROR("%s: can't initialize environment: rc = %d\n",
699 tgt->lut_obd->obd_name, rc);
703 spin_lock(&tgt->lut_translock);
704 start_epoch = lr_epoch(tgt->lut_last_transno) + 1;
705 tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
706 tgt->lut_lsd.lsd_start_epoch = start_epoch;
707 spin_unlock(&tgt->lut_translock);
709 INIT_LIST_HEAD(&client_list);
711 * The recovery is not yet finished and final queue can still be updated
712 * with resend requests. Move final list to separate one for processing
714 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
715 list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
716 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
719 * go through list of exports participated in recovery and
720 * set new epoch for them
722 list_for_each_entry(req, &client_list, rq_list) {
723 LASSERT(!req->rq_export->exp_delayed);
724 if (!req->rq_export->exp_vbr_failed)
725 tgt_client_epoch_update(&env, req->rq_export);
727 /** return list back at once */
728 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
729 list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
730 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
732 /** Clear MULTI RPCS incompatibility flag if
733 * - target is MDT and
734 * - there is no client to recover or the recovery was aborted
736 if (!strncmp(tgt->lut_obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) &&
737 (tgt->lut_obd->obd_max_recoverable_clients == 0 ||
738 tgt->lut_obd->obd_abort_recovery))
739 tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
741 /** update server epoch */
742 tgt_server_data_update(&env, tgt, 1);
747 * commit callback, need to update last_commited value
749 struct tgt_last_committed_callback {
750 struct dt_txn_commit_cb llcc_cb;
751 struct lu_target *llcc_tgt;
752 struct obd_export *llcc_exp;
756 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
757 struct dt_txn_commit_cb *cb, int err)
759 struct tgt_last_committed_callback *ccb;
761 ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
763 LASSERT(ccb->llcc_tgt != NULL);
764 LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
766 spin_lock(&ccb->llcc_tgt->lut_translock);
767 if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
768 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
770 LASSERT(ccb->llcc_exp);
771 if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
772 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
773 spin_unlock(&ccb->llcc_tgt->lut_translock);
774 ptlrpc_commit_replies(ccb->llcc_exp);
776 spin_unlock(&ccb->llcc_tgt->lut_translock);
778 class_export_cb_put(ccb->llcc_exp);
779 if (ccb->llcc_transno)
780 CDEBUG(D_HA, "%s: transno "LPD64" is committed\n",
781 ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
785 int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
786 struct obd_export *exp, __u64 transno)
788 struct tgt_last_committed_callback *ccb;
789 struct dt_txn_commit_cb *dcb;
797 ccb->llcc_exp = class_export_cb_get(exp);
798 ccb->llcc_transno = transno;
801 dcb->dcb_func = tgt_cb_last_committed;
802 INIT_LIST_HEAD(&dcb->dcb_linkage);
803 strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
805 rc = dt_trans_cb_add(th, dcb);
807 class_export_cb_put(exp);
811 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
812 /* report failure to force synchronous operation */
818 struct tgt_new_client_callback {
819 struct dt_txn_commit_cb lncc_cb;
820 struct obd_export *lncc_exp;
823 static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
824 struct dt_txn_commit_cb *cb, int err)
826 struct tgt_new_client_callback *ccb;
828 ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
830 LASSERT(ccb->lncc_exp->exp_obd);
832 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
833 ccb->lncc_exp->exp_obd->obd_name,
834 ccb->lncc_exp->exp_client_uuid.uuid);
836 spin_lock(&ccb->lncc_exp->exp_lock);
838 ccb->lncc_exp->exp_need_sync = 0;
840 spin_unlock(&ccb->lncc_exp->exp_lock);
841 class_export_cb_put(ccb->lncc_exp);
846 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
848 struct tgt_new_client_callback *ccb;
849 struct dt_txn_commit_cb *dcb;
856 ccb->lncc_exp = class_export_cb_get(exp);
859 dcb->dcb_func = tgt_cb_new_client;
860 INIT_LIST_HEAD(&dcb->dcb_linkage);
861 strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
863 rc = dt_trans_cb_add(th, dcb);
865 class_export_cb_put(exp);
872 * Add new client to the last_rcvd upon new connection.
874 * We use a bitmap to locate a free space in the last_rcvd file and initialize
877 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
879 struct tg_export_data *ted = &exp->exp_target_data;
880 struct lu_target *tgt = class_exp2tgt(exp);
885 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
886 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
889 mutex_init(&ted->ted_lcd_lock);
891 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
894 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
895 * there's no need for extra complication here
897 idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
899 if (idx >= LR_MAX_CLIENTS ||
900 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
901 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
902 tgt->lut_obd->obd_name, idx);
905 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
906 idx = find_next_zero_bit(tgt->lut_client_bitmap,
907 LR_MAX_CLIENTS, idx);
911 ted->ted_lr_idx = idx;
912 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
913 idx * tgt->lut_lsd.lsd_client_size;
915 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
917 if (tgt_is_multimodrpcs_client(exp)) {
918 /* Set MULTI RPCS incompatibility flag to prevent previous
919 * Lustre versions to mount a target with reply_data file */
920 atomic_inc(&tgt->lut_num_clients);
921 if (!(tgt->lut_lsd.lsd_feature_incompat &
922 OBD_INCOMPAT_MULTI_RPCS)) {
923 tgt->lut_lsd.lsd_feature_incompat |=
924 OBD_INCOMPAT_MULTI_RPCS;
925 rc = tgt_server_data_update(env, tgt, 1);
927 CERROR("%s: unable to set MULTI RPCS "
928 "incompatibility flag\n",
929 exp->exp_obd->obd_name);
934 /* assign client slot generation */
935 ted->ted_lcd->lcd_generation =
936 atomic_inc_return(&tgt->lut_client_generation);
938 ted->ted_lcd->lcd_generation = 0;
941 CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s' "
943 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
944 ted->ted_lcd->lcd_uuid, ted->ted_lcd->lcd_generation);
946 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
949 rc = tgt_client_data_update(env, exp);
951 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
952 tgt->lut_obd->obd_name, idx, rc);
956 EXPORT_SYMBOL(tgt_client_new);
958 /* Add an existing client to the MDS in-memory state based on
959 * a client that was previously found in the last_rcvd file and
960 * already has an assigned slot (idx >= 0).
962 * It should not be possible to fail adding an existing client - otherwise
963 * mdt_init_server_data() callsite needs to be fixed.
965 int tgt_client_add(const struct lu_env *env, struct obd_export *exp, int idx)
967 struct tg_export_data *ted = &exp->exp_target_data;
968 struct lu_target *tgt = class_exp2tgt(exp);
972 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
973 LASSERTF(idx >= 0, "%d\n", idx);
975 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
976 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
979 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
980 CERROR("%s: client %d: bit already set in bitmap!!\n",
981 tgt->lut_obd->obd_name, idx);
984 atomic_inc(&tgt->lut_num_clients);
986 CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added, "
988 tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid,
989 ted->ted_lcd->lcd_generation);
991 ted->ted_lr_idx = idx;
992 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
993 idx * tgt->lut_lsd.lsd_client_size;
995 mutex_init(&ted->ted_lcd_lock);
997 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1002 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
1004 struct tg_export_data *ted = &exp->exp_target_data;
1005 struct lu_target *tgt = class_exp2tgt(exp);
1010 LASSERT(ted->ted_lcd);
1012 if (unlikely(tgt == NULL)) {
1013 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1014 class_exp2obd(exp)->obd_name);
1018 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
1019 if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
1020 (char *)tgt->lut_obd->obd_uuid.uuid) ||
1021 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1024 CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
1025 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1026 ted->ted_lcd->lcd_uuid);
1028 /* Clear the bit _after_ zeroing out the client so we don't
1029 race with filter_client_add and zero out new clients.*/
1030 if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
1031 CERROR("%s: client %u: bit already clear in bitmap!!\n",
1032 tgt->lut_obd->obd_name, ted->ted_lr_idx);
1036 /* Do not erase record for recoverable client. */
1037 if (exp->exp_flags & OBD_OPT_FAILOVER)
1040 /* Make sure the server's last_transno is up to date.
1041 * This should be done before zeroing client slot so last_transno will
1042 * be in server data or in client data in case of failure */
1043 rc = tgt_server_data_update(env, tgt, 0);
1045 CERROR("%s: failed to update server data, skip client %s "
1046 "zeroing, rc %d\n", tgt->lut_obd->obd_name,
1047 ted->ted_lcd->lcd_uuid, rc);
1051 mutex_lock(&ted->ted_lcd_lock);
1052 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
1053 rc = tgt_client_data_update(env, exp);
1054 mutex_unlock(&ted->ted_lcd_lock);
1056 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
1057 "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
1058 tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
1059 ted->ted_lr_idx, ted->ted_lr_off, rc);
1062 EXPORT_SYMBOL(tgt_client_del);
1065 * last_rcvd & last_committed update callbacks
1067 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
1068 struct dt_object *obj, __u64 opdata,
1069 struct thandle *th, struct ptlrpc_request *req)
1071 struct tgt_thread_info *tti = tgt_th_info(env);
1072 struct tg_export_data *ted;
1075 bool lw_client, update = false;
1079 ted = &req->rq_export->exp_target_data;
1081 lw_client = exp_connect_flags(req->rq_export) & OBD_CONNECT_LIGHTWEIGHT;
1082 if (ted->ted_lr_idx < 0 && !lw_client)
1083 /* ofd connect may cause transaction before export has
1087 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
1089 spin_lock(&tgt->lut_translock);
1090 if (th->th_result != 0) {
1091 if (tti->tti_transno != 0) {
1092 CERROR("%s: replay transno "LPU64" failed: rc = %d\n",
1093 tgt_name(tgt), tti->tti_transno, th->th_result);
1095 } else if (tti->tti_transno == 0) {
1096 tti->tti_transno = ++tgt->lut_last_transno;
1098 /* should be replay */
1099 if (tti->tti_transno > tgt->lut_last_transno)
1100 tgt->lut_last_transno = tti->tti_transno;
1102 spin_unlock(&tgt->lut_translock);
1104 /** VBR: set new versions */
1105 if (th->th_result == 0 && obj != NULL) {
1106 struct dt_object *dto = dt_object_locate(obj, th->th_dev);
1107 dt_version_set(env, dto, tti->tti_transno, th);
1110 /* filling reply data */
1111 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
1112 tti->tti_transno, tgt->lut_obd->obd_last_committed);
1114 req->rq_transno = tti->tti_transno;
1115 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
1117 /* if can't add callback, do sync write */
1118 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, req->rq_export,
1122 /* All operations performed by LW clients are synchronous and
1123 * we store the committed transno in the last_rcvd header */
1124 spin_lock(&tgt->lut_translock);
1125 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
1126 tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
1129 spin_unlock(&tgt->lut_translock);
1130 /* Although lightweight (LW) connections have no slot in
1131 * last_rcvd, we still want to maintain the in-memory
1132 * lsd_client_data structure in order to properly handle reply
1133 * reconstruction. */
1134 } else if (ted->ted_lr_off == 0) {
1135 CERROR("%s: client idx %d has offset %lld\n",
1136 tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
1140 /* if the export has already been disconnected, we have no last_rcvd
1141 * slot, update server data with latest transno then */
1142 if (ted->ted_lcd == NULL) {
1143 CWARN("commit transaction for disconnected client %s: rc %d\n",
1144 req->rq_export->exp_client_uuid.uuid, rc);
1145 GOTO(srv_update, rc = 0);
1148 /* Target that supports multiple reply data */
1149 if (tgt_is_multimodrpcs_client(req->rq_export)) {
1150 struct tg_reply_data *trd;
1151 struct lsd_reply_data *lrd;
1152 __u64 *pre_versions;
1157 if (unlikely(trd == NULL))
1158 GOTO(srv_update, rc = -ENOMEM);
1160 /* update export last transno */
1161 mutex_lock(&ted->ted_lcd_lock);
1162 if (tti->tti_transno > ted->ted_lcd->lcd_last_transno)
1163 ted->ted_lcd->lcd_last_transno = tti->tti_transno;
1164 mutex_unlock(&ted->ted_lcd_lock);
1166 /* fill reply data information */
1167 lrd = &trd->trd_reply;
1168 lrd->lrd_transno = tti->tti_transno;
1169 lrd->lrd_xid = req->rq_xid;
1170 lrd->lrd_result = th->th_result;
1171 lrd->lrd_data = opdata;
1172 lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
1173 trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
1174 pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1176 trd->trd_pre_versions[0] = pre_versions[0];
1177 trd->trd_pre_versions[1] = pre_versions[1];
1178 trd->trd_pre_versions[2] = pre_versions[2];
1179 trd->trd_pre_versions[3] = pre_versions[3];
1182 /* find a empty slot */
1183 i = tgt_find_free_reply_slot(tgt);
1184 if (unlikely(i < 0)) {
1185 CERROR("%s: couldn't find a slot for reply data: "
1186 "rc = %d\n", tgt_name(tgt), i);
1187 GOTO(srv_update, rc = i);
1191 /* write reply data to disk */
1192 off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
1193 rc = tgt_reply_data_write(env, tgt, lrd, off, th);
1194 if (unlikely(rc != 0)) {
1195 CERROR("%s: can't update %s file: rc = %d\n",
1196 tgt_name(tgt), REPLY_DATA, rc);
1200 /* add reply data to target export's reply list */
1201 mutex_lock(&ted->ted_lcd_lock);
1202 list_add(&trd->trd_list, &ted->ted_reply_list);
1203 ted->ted_reply_cnt++;
1204 if (ted->ted_reply_cnt > ted->ted_reply_max)
1205 ted->ted_reply_max = ted->ted_reply_cnt;
1206 mutex_unlock(&ted->ted_lcd_lock);
1208 CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
1209 "tag %hu, client gen %u, slot idx %d\n",
1210 trd, lrd->lrd_xid, lrd->lrd_transno,
1211 trd->trd_tag, lrd->lrd_client_gen, i);
1213 GOTO(srv_update, rc = 0);
1216 mutex_lock(&ted->ted_lcd_lock);
1217 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1218 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
1219 transno_p = &ted->ted_lcd->lcd_last_close_transno;
1220 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
1221 ted->ted_lcd->lcd_last_close_result = th->th_result;
1223 /* VBR: save versions in last_rcvd for reconstruct. */
1224 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1227 ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
1228 ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
1229 ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
1230 ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
1232 transno_p = &ted->ted_lcd->lcd_last_transno;
1233 ted->ted_lcd->lcd_last_xid = req->rq_xid;
1234 ted->ted_lcd->lcd_last_result = th->th_result;
1235 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
1236 * see struct ldlm_reply->lock_policy_res1; */
1237 ted->ted_lcd->lcd_last_data = opdata;
1240 /* Update transno in slot only if non-zero number, i.e. no errors */
1241 if (likely(tti->tti_transno != 0)) {
1242 if (*transno_p > tti->tti_transno &&
1243 !tgt->lut_no_reconstruct) {
1244 CERROR("%s: trying to overwrite bigger transno:"
1245 "on-disk: "LPU64", new: "LPU64" replay: %d. "
1246 "see LU-617.\n", tgt_name(tgt), *transno_p,
1247 tti->tti_transno, req_is_replay(req));
1248 if (req_is_replay(req)) {
1249 spin_lock(&req->rq_export->exp_lock);
1250 req->rq_export->exp_vbr_failed = 1;
1251 spin_unlock(&req->rq_export->exp_lock);
1253 mutex_unlock(&ted->ted_lcd_lock);
1254 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
1256 *transno_p = tti->tti_transno;
1260 tti->tti_off = ted->ted_lr_off;
1261 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1263 mutex_unlock(&ted->ted_lcd_lock);
1267 mutex_unlock(&ted->ted_lcd_lock);
1271 rc = tgt_server_data_write(env, tgt, th);
1276 * last_rcvd update for echo client simulation.
1277 * It updates last_rcvd client slot and version of object in
1278 * simple way but with all locks to simulate all drawbacks
1280 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
1281 struct lu_target *tgt,
1282 struct dt_object *obj,
1284 struct obd_export *exp)
1286 struct tgt_thread_info *tti = tgt_th_info(env);
1287 struct tg_export_data *ted = &exp->exp_target_data;
1292 tti->tti_transno = 0;
1294 spin_lock(&tgt->lut_translock);
1295 if (th->th_result == 0)
1296 tti->tti_transno = ++tgt->lut_last_transno;
1297 spin_unlock(&tgt->lut_translock);
1299 /** VBR: set new versions */
1300 if (th->th_result == 0 && obj != NULL)
1301 dt_version_set(env, obj, tti->tti_transno, th);
1303 /* if can't add callback, do sync write */
1304 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
1307 LASSERT(ted->ted_lr_off > 0);
1309 mutex_lock(&ted->ted_lcd_lock);
1310 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1311 ted->ted_lcd->lcd_last_transno = tti->tti_transno;
1312 ted->ted_lcd->lcd_last_result = th->th_result;
1314 tti->tti_off = ted->ted_lr_off;
1315 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1316 mutex_unlock(&ted->ted_lcd_lock);
1320 static int tgt_clients_data_init(const struct lu_env *env,
1321 struct lu_target *tgt,
1322 unsigned long last_size)
1324 struct obd_device *obd = tgt->lut_obd;
1325 struct lr_server_data *lsd = &tgt->lut_lsd;
1326 struct lsd_client_data *lcd = NULL;
1327 struct tg_export_data *ted;
1330 loff_t off = lsd->lsd_client_start;
1331 __u32 generation = 0;
1332 struct cfs_hash *hash = NULL;
1336 CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
1337 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
1343 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1345 GOTO(err_out, rc = -ENODEV);
1347 for (cl_idx = 0; off < last_size; cl_idx++) {
1348 struct obd_export *exp;
1351 /* Don't assume off is incremented properly by
1352 * read_record(), in case sizeof(*lcd)
1353 * isn't the same as fsd->lsd_client_size. */
1354 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
1355 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
1357 CERROR("%s: error reading last_rcvd %s idx %d off "
1358 "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
1361 break; /* read error shouldn't cause startup to fail */
1364 if (lcd->lcd_uuid[0] == '\0') {
1365 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
1370 last_transno = lcd_last_transno(lcd);
1372 /* These exports are cleaned up by disconnect, so they
1373 * need to be set up like real exports as connect does.
1375 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
1376 " srv lr: "LPU64" lx: "LPU64" gen %u\n", lcd->lcd_uuid,
1377 cl_idx, last_transno, lsd->lsd_last_transno,
1378 lcd_last_xid(lcd), lcd->lcd_generation);
1380 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
1382 if (PTR_ERR(exp) == -EALREADY) {
1383 /* export already exists, zero out this one */
1384 CERROR("%s: Duplicate export %s!\n",
1385 tgt_name(tgt), lcd->lcd_uuid);
1388 GOTO(err_out, rc = PTR_ERR(exp));
1391 ted = &exp->exp_target_data;
1392 *ted->ted_lcd = *lcd;
1394 rc = tgt_client_add(env, exp, cl_idx);
1395 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
1396 /* VBR: set export last committed version */
1397 exp->exp_last_committed = last_transno;
1398 spin_lock(&exp->exp_lock);
1399 exp->exp_connecting = 0;
1400 exp->exp_in_recovery = 0;
1401 spin_unlock(&exp->exp_lock);
1402 obd->obd_max_recoverable_clients++;
1404 if (tgt->lut_lsd.lsd_feature_incompat &
1405 OBD_INCOMPAT_MULTI_RPCS &&
1406 lcd->lcd_generation != 0) {
1407 /* compute the highest valid client generation */
1408 generation = max(generation, lcd->lcd_generation);
1410 /* fill client_generation <-> export hash table */
1411 rc = cfs_hash_add_unique(hash, &lcd->lcd_generation,
1412 &exp->exp_gen_hash);
1414 CERROR("%s: duplicate export for client "
1416 tgt_name(tgt), lcd->lcd_generation);
1417 class_export_put(exp);
1422 class_export_put(exp);
1424 /* Need to check last_rcvd even for duplicated exports. */
1425 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
1426 cl_idx, last_transno);
1428 spin_lock(&tgt->lut_translock);
1429 tgt->lut_last_transno = max(last_transno,
1430 tgt->lut_last_transno);
1431 spin_unlock(&tgt->lut_translock);
1434 /* record highest valid client generation */
1435 atomic_set(&tgt->lut_client_generation, generation);
1439 cfs_hash_putref(hash);
1444 struct server_compat_data {
1451 static struct server_compat_data tgt_scd[] = {
1452 [LDD_F_SV_TYPE_MDT] = {
1453 .rocompat = OBD_ROCOMPAT_LOVOBJID,
1454 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1455 OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
1456 OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI |
1457 OBD_INCOMPAT_MULTI_RPCS,
1458 .rocinit = OBD_ROCOMPAT_LOVOBJID,
1459 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1460 OBD_INCOMPAT_MULTI_OI,
1462 [LDD_F_SV_TYPE_OST] = {
1463 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
1464 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
1466 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
1467 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1471 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1473 struct tgt_thread_info *tti = tgt_th_info(env);
1474 struct lr_server_data *lsd = &tgt->lut_lsd;
1475 unsigned long last_rcvd_size;
1479 rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1483 last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1485 /* ensure padding in the struct is the correct size */
1486 CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
1487 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
1489 rc = server_name2index(tgt_name(tgt), &index, NULL);
1491 CERROR("%s: Can not get index from name: rc = %d\n",
1495 /* server_name2index() returns type */
1497 if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1498 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1502 /* last_rcvd on OST doesn't provide reconstruct support because there
1503 * may be up to 8 in-flight write requests per single slot in
1504 * last_rcvd client data
1506 tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1508 if (last_rcvd_size == 0) {
1509 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1511 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1512 sizeof(lsd->lsd_uuid));
1513 lsd->lsd_last_transno = 0;
1514 lsd->lsd_mount_count = 0;
1515 lsd->lsd_server_size = LR_SERVER_SIZE;
1516 lsd->lsd_client_start = LR_CLIENT_START;
1517 lsd->lsd_client_size = LR_CLIENT_SIZE;
1518 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1519 lsd->lsd_osd_index = index;
1520 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1521 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1523 rc = tgt_server_data_read(env, tgt);
1525 CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1529 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1530 LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
1531 "using the wrong disk %s. Were the"
1532 " /dev/ assignments rearranged?\n",
1533 tgt->lut_obd->obd_uuid.uuid,
1538 if (lsd->lsd_osd_index != index) {
1539 LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
1540 "is different with the index %d in"
1541 "config log, It might be disk"
1542 "corruption!\n", tgt_name(tgt),
1543 lsd->lsd_osd_index, index);
1548 if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1549 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1551 lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1555 if (type == LDD_F_SV_TYPE_MDT)
1556 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1558 if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1559 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1561 lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1564 /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1565 if (type == LDD_F_SV_TYPE_MDT &&
1566 !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1567 if (last_rcvd_size > lsd->lsd_client_start) {
1568 LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1569 "remove all clients for interop needs\n",
1571 rc = tgt_truncate_last_rcvd(env, tgt,
1572 lsd->lsd_client_start);
1575 last_rcvd_size = lsd->lsd_client_start;
1577 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1578 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1581 spin_lock(&tgt->lut_translock);
1582 tgt->lut_last_transno = lsd->lsd_last_transno;
1583 spin_unlock(&tgt->lut_translock);
1585 lsd->lsd_mount_count++;
1587 CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1588 CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
1589 tgt_name(tgt), tgt->lut_last_transno);
1590 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
1591 tgt_name(tgt), lsd->lsd_mount_count);
1592 CDEBUG(D_INODE, "%s: server data size: %u\n",
1593 tgt_name(tgt), lsd->lsd_server_size);
1594 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1595 tgt_name(tgt), lsd->lsd_client_start);
1596 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1597 tgt_name(tgt), lsd->lsd_client_size);
1598 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1599 tgt_name(tgt), last_rcvd_size);
1600 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1601 tgt_name(tgt), lsd->lsd_subdir_count);
1602 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1603 last_rcvd_size <= lsd->lsd_client_start ? 0 :
1604 (last_rcvd_size - lsd->lsd_client_start) /
1605 lsd->lsd_client_size);
1606 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1608 if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1609 lsd->lsd_client_size == 0) {
1610 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1614 if (!tgt->lut_obd->obd_replayable)
1615 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1617 rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1619 GOTO(err_client, rc);
1621 spin_lock(&tgt->lut_translock);
1622 /* obd_last_committed is used for compatibility
1623 * with other lustre recovery code */
1624 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1625 spin_unlock(&tgt->lut_translock);
1627 tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1628 tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1630 /* save it, so mount count and last_transno is current */
1631 rc = tgt_server_data_update(env, tgt, 0);
1633 GOTO(err_client, rc);
1638 class_disconnect_exports(tgt->lut_obd);
1642 /* add credits for last_rcvd update */
1643 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1646 struct lu_target *tgt = cookie;
1647 struct tgt_session_info *tsi;
1648 struct tgt_thread_info *tti = tgt_th_info(env);
1649 struct dt_object *dto;
1652 /* if there is no session, then this transaction is not result of
1653 * request processing but some local operation */
1654 if (env->le_ses == NULL)
1657 LASSERT(tgt->lut_last_rcvd);
1658 tsi = tgt_ses_info(env);
1659 /* OFD may start transaction without export assigned */
1660 if (tsi->tsi_exp == NULL)
1663 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
1666 rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1667 tsi->tsi_exp->exp_target_data.ted_lr_off,
1673 rc = dt_declare_record_write(env, dto, &tti->tti_buf, 0, th);
1677 if (tsi->tsi_vbr_obj != NULL &&
1678 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1679 dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
1680 rc = dt_declare_version_set(env, dto, th);
1686 /* Update last_rcvd records with latests transaction data */
1687 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1690 struct lu_target *tgt = cookie;
1691 struct tgt_session_info *tsi;
1692 struct tgt_thread_info *tti = tgt_th_info(env);
1693 struct dt_object *obj = NULL;
1697 if (env->le_ses == NULL)
1700 tsi = tgt_ses_info(env);
1701 /* OFD may start transaction without export assigned */
1702 if (tsi->tsi_exp == NULL)
1705 echo_client = (tgt_ses_req(tsi) == NULL);
1707 if (tti->tti_has_trans && !echo_client) {
1708 if (tti->tti_mult_trans == 0) {
1709 CDEBUG(D_HA, "More than one transaction "LPU64"\n",
1713 /* we need another transno to be assigned */
1714 tti->tti_transno = 0;
1715 } else if (th->th_result == 0) {
1716 tti->tti_has_trans = 1;
1719 if (tsi->tsi_vbr_obj != NULL &&
1720 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1721 obj = tsi->tsi_vbr_obj;
1724 if (unlikely(echo_client)) /* echo client special case */
1725 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1728 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
1733 int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
1735 struct tgt_thread_info *tti = tgt_th_info(env);
1736 struct lsd_reply_data *lrd = &tti->tti_lrd;
1737 unsigned long reply_data_size;
1739 struct lsd_reply_header *lrh = NULL;
1740 struct lsd_client_data *lcd = NULL;
1741 struct tg_reply_data *trd = NULL;
1744 struct cfs_hash *hash = NULL;
1745 struct obd_export *exp;
1746 struct obd_export *tmp;
1747 struct tg_export_data *ted;
1748 int reply_data_recovered = 0;
1750 rc = dt_attr_get(env, tgt->lut_reply_data, &tti->tti_attr);
1753 reply_data_size = (unsigned long)tti->tti_attr.la_size;
1757 GOTO(out, rc = -ENOMEM);
1759 if (reply_data_size == 0) {
1760 CDEBUG(D_INFO, "%s: new reply_data file, initializing\n",
1762 lrh->lrh_magic = LRH_MAGIC;
1763 lrh->lrh_header_size = sizeof(struct lsd_reply_header);
1764 lrh->lrh_reply_size = sizeof(struct lsd_reply_data);
1765 rc = tgt_reply_header_write(env, tgt, lrh);
1767 CERROR("%s: error writing %s: rc = %d\n",
1768 tgt_name(tgt), REPLY_DATA, rc);
1772 rc = tgt_reply_header_read(env, tgt, lrh);
1774 CERROR("%s: error reading %s: rc = %d\n",
1775 tgt_name(tgt), REPLY_DATA, rc);
1778 if (lrh->lrh_magic != LRH_MAGIC ||
1779 lrh->lrh_header_size != sizeof(struct lsd_reply_header) ||
1780 lrh->lrh_reply_size != sizeof(struct lsd_reply_data)) {
1781 CERROR("%s: invalid header in %s\n",
1782 tgt_name(tgt), REPLY_DATA);
1783 GOTO(out, rc = -EINVAL);
1786 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1788 GOTO(out, rc = -ENODEV);
1792 GOTO(out, rc = -ENOMEM);
1796 GOTO(out, rc = -ENOMEM);
1798 /* Load reply_data from disk */
1799 for (idx = 0, off = sizeof(struct lsd_reply_header);
1800 off < reply_data_size;
1801 idx++, off += sizeof(struct lsd_reply_data)) {
1802 rc = tgt_reply_data_read(env, tgt, lrd, off);
1804 CERROR("%s: error reading %s: rc = %d\n",
1805 tgt_name(tgt), REPLY_DATA, rc);
1809 exp = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
1811 /* old reply data from a disconnected client */
1814 ted = &exp->exp_target_data;
1815 mutex_lock(&ted->ted_lcd_lock);
1817 /* create in-memory reply_data and link it to
1818 * target export's reply list */
1819 tgt_set_reply_slot(tgt, idx);
1820 trd->trd_reply = *lrd;
1821 trd->trd_pre_versions[0] = 0;
1822 trd->trd_pre_versions[1] = 0;
1823 trd->trd_pre_versions[2] = 0;
1824 trd->trd_pre_versions[3] = 0;
1825 trd->trd_index = idx;
1827 list_add(&trd->trd_list, &ted->ted_reply_list);
1828 ted->ted_reply_cnt++;
1829 if (ted->ted_reply_cnt > ted->ted_reply_max)
1830 ted->ted_reply_max = ted->ted_reply_cnt;
1832 CDEBUG(D_HA, "%s: restore reply %p: xid %llu, "
1833 "transno %llu, client gen %u, slot idx %d\n",
1834 tgt_name(tgt), trd, lrd->lrd_xid,
1835 lrd->lrd_transno, lrd->lrd_client_gen,
1838 /* update export last committed transation */
1839 exp->exp_last_committed = max(exp->exp_last_committed,
1842 mutex_unlock(&ted->ted_lcd_lock);
1843 class_export_put(exp);
1845 /* update target last committed transaction */
1846 spin_lock(&tgt->lut_translock);
1847 tgt->lut_last_transno = max(tgt->lut_last_transno,
1849 spin_unlock(&tgt->lut_translock);
1851 reply_data_recovered++;
1855 GOTO(out, rc = -ENOMEM);
1857 CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
1858 tgt_name(tgt), reply_data_recovered);
1860 /* delete entries from client_generation<->export hash */
1861 spin_lock(&tgt->lut_obd->obd_dev_lock);
1862 list_for_each_entry_safe(exp, tmp,
1863 &tgt->lut_obd->obd_exports,
1865 struct tg_export_data *ted = &exp->exp_target_data;
1867 if (!hlist_unhashed(&exp->exp_gen_hash))
1869 &ted->ted_lcd->lcd_generation,
1870 &exp->exp_gen_hash);
1872 spin_unlock(&tgt->lut_obd->obd_dev_lock);
1879 cfs_hash_putref(hash);
1889 /* Look for a reply data matching specified request @req
1890 * A copy is returned in @trd if the pointer is not NULL
1892 bool tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
1894 struct tg_export_data *ted = &req->rq_export->exp_target_data;
1895 struct tg_reply_data *reply, *tmp;
1898 mutex_lock(&ted->ted_lcd_lock);
1899 list_for_each_entry_safe(reply, tmp, &ted->ted_reply_list, trd_list) {
1900 if (reply->trd_reply.lrd_xid == req->rq_xid) {
1905 if (found && trd != NULL)
1907 mutex_unlock(&ted->ted_lcd_lock);
1909 CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d\n",
1910 tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid,
1915 EXPORT_SYMBOL(tgt_lookup_reply);
1917 int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
1919 struct tg_export_data *ted = &exp->exp_target_data;
1920 struct lu_target *lut = class_exp2tgt(exp);
1921 struct tg_reply_data *trd, *tmp;
1923 mutex_lock(&ted->ted_lcd_lock);
1924 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
1925 if (trd->trd_reply.lrd_xid > rcvd_xid)
1927 ted->ted_release_xid++;
1928 tgt_release_reply_data(lut, ted, trd);
1930 mutex_unlock(&ted->ted_lcd_lock);
1935 int tgt_handle_tag(struct obd_export *exp, __u16 tag)
1937 struct tg_export_data *ted = &exp->exp_target_data;
1938 struct lu_target *lut = class_exp2tgt(exp);
1939 struct tg_reply_data *trd, *tmp;
1944 mutex_lock(&ted->ted_lcd_lock);
1945 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
1946 if (trd->trd_tag != tag)
1948 ted->ted_release_tag++;
1949 tgt_release_reply_data(lut, ted, trd);
1952 mutex_unlock(&ted->ted_lcd_lock);