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, 2015, 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 lut == NULL ? "" : 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--;
320 tgt_clear_reply_slot(lut, trd->trd_index);
324 /* Release the reply data @trd from target @lut
325 * The reply data with the highest transno for this export
326 * is retained to ensure correctness of target recovery
327 * Called with ted_lcd_lock held
329 static void tgt_release_reply_data(struct lu_target *lut,
330 struct tg_export_data *ted,
331 struct tg_reply_data *trd)
333 CDEBUG(D_TRACE, "%s: release reply data %p: xid %llu, transno %llu, "
334 "client gen %u, slot idx %d\n",
335 lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
336 trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
339 LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
341 /* Do not free the reply data corresponding to the
342 * highest transno of this export.
343 * This ensures on-disk reply data is kept and
344 * last committed transno can be restored from disk in case
347 if (trd->trd_reply.lrd_transno == ted->ted_lcd->lcd_last_transno) {
348 /* free previous retained reply */
349 if (ted->ted_reply_last != NULL)
350 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
351 /* retain the reply */
352 list_del_init(&trd->trd_list);
353 ted->ted_reply_last = trd;
355 tgt_free_reply_data(lut, ted, trd);
359 static inline struct lu_buf *tti_buf_lsd(struct tgt_thread_info *tti)
361 tti->tti_buf.lb_buf = &tti->tti_lsd;
362 tti->tti_buf.lb_len = sizeof(tti->tti_lsd);
363 return &tti->tti_buf;
366 static inline struct lu_buf *tti_buf_lcd(struct tgt_thread_info *tti)
368 tti->tti_buf.lb_buf = &tti->tti_lcd;
369 tti->tti_buf.lb_len = sizeof(tti->tti_lcd);
370 return &tti->tti_buf;
374 * Allocate in-memory data for client slot related to export.
376 int tgt_client_alloc(struct obd_export *exp)
379 LASSERT(exp != exp->exp_obd->obd_self_export);
381 spin_lock_init(&exp->exp_target_data.ted_nodemap_lock);
382 INIT_LIST_HEAD(&exp->exp_target_data.ted_nodemap_member);
384 OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
385 if (exp->exp_target_data.ted_lcd == NULL)
387 /* Mark that slot is not yet valid, 0 doesn't work here */
388 exp->exp_target_data.ted_lr_idx = -1;
389 INIT_LIST_HEAD(&exp->exp_target_data.ted_reply_list);
390 mutex_init(&exp->exp_target_data.ted_lcd_lock);
393 EXPORT_SYMBOL(tgt_client_alloc);
396 * Free in-memory data for client slot related to export.
398 void tgt_client_free(struct obd_export *exp)
400 struct tg_export_data *ted = &exp->exp_target_data;
401 struct lu_target *lut = class_exp2tgt(exp);
402 struct tg_reply_data *trd, *tmp;
404 LASSERT(exp != exp->exp_obd->obd_self_export);
406 /* free reply data */
407 mutex_lock(&ted->ted_lcd_lock);
408 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
409 tgt_release_reply_data(lut, ted, trd);
411 if (ted->ted_reply_last != NULL) {
412 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
413 ted->ted_reply_last = NULL;
415 mutex_unlock(&ted->ted_lcd_lock);
417 if (!hlist_unhashed(&exp->exp_gen_hash))
418 cfs_hash_del(exp->exp_obd->obd_gen_hash,
419 &ted->ted_lcd->lcd_generation,
422 OBD_FREE_PTR(ted->ted_lcd);
425 /* Target may have been freed (see LU-7430)
426 * Slot may be not yet assigned */
427 if (exp->exp_obd->u.obt.obt_magic != OBT_MAGIC ||
431 /* Clear bit when lcd is freed */
432 LASSERT(lut && lut->lut_client_bitmap);
433 if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
434 CERROR("%s: client %u bit already clear in bitmap\n",
435 exp->exp_obd->obd_name, ted->ted_lr_idx);
439 if (tgt_is_multimodrpcs_client(exp) && !exp->exp_obd->obd_stopping)
440 atomic_dec(&lut->lut_num_clients);
442 EXPORT_SYMBOL(tgt_client_free);
444 int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
445 struct lsd_client_data *lcd, loff_t *off, int index)
447 struct tgt_thread_info *tti = tgt_th_info(env);
451 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
453 check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
454 lcd_le_to_cpu(&tti->tti_lcd, lcd);
455 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
456 lcd->lcd_last_close_result =
457 ptlrpc_status_ntoh(lcd->lcd_last_close_result);
460 CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = "LPU64
461 ", last_xid = "LPU64", last_result = %u, last_data = %u, "
462 "last_close_transno = "LPU64", last_close_xid = "LPU64", "
463 "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
464 *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
465 lcd->lcd_last_result, lcd->lcd_last_data,
466 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
467 lcd->lcd_last_close_result, rc);
471 int tgt_client_data_write(const struct lu_env *env, struct lu_target *tgt,
472 struct lsd_client_data *lcd, loff_t *off,
475 struct tgt_thread_info *tti = tgt_th_info(env);
476 struct dt_object *dto;
478 lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
479 lcd->lcd_last_close_result =
480 ptlrpc_status_hton(lcd->lcd_last_close_result);
481 lcd_cpu_to_le(lcd, &tti->tti_lcd);
484 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
485 return dt_record_write(env, dto, &tti->tti_buf, off, th);
489 * Update client data in last_rcvd
491 static int tgt_client_data_update(const struct lu_env *env,
492 struct obd_export *exp)
494 struct tg_export_data *ted = &exp->exp_target_data;
495 struct lu_target *tgt = class_exp2tgt(exp);
496 struct tgt_thread_info *tti = tgt_th_info(env);
502 if (unlikely(tgt == NULL)) {
503 CDEBUG(D_ERROR, "%s: No target for connected export\n",
504 class_exp2obd(exp)->obd_name);
508 th = dt_trans_create(env, tgt->lut_bottom);
513 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
515 ted->ted_lr_off, th);
519 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
523 * Until this operations will be committed the sync is needed
524 * for this export. This should be done _after_ starting the
525 * transaction so that many connecting clients will not bring
526 * server down with lots of sync writes.
528 rc = tgt_new_client_cb_add(th, exp);
530 /* can't add callback, do sync now */
533 spin_lock(&exp->exp_lock);
534 exp->exp_need_sync = 1;
535 spin_unlock(&exp->exp_lock);
538 tti->tti_off = ted->ted_lr_off;
539 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
542 dt_trans_stop(env, tgt->lut_bottom, th);
543 CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
544 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
545 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
550 int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
552 struct tgt_thread_info *tti = tgt_th_info(env);
557 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
560 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
562 CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
563 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
564 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
568 int tgt_server_data_write(const struct lu_env *env, struct lu_target *tgt,
571 struct tgt_thread_info *tti = tgt_th_info(env);
572 struct dt_object *dto;
579 lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
581 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
582 rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
584 CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
585 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
586 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
592 * Update server data in last_rcvd
594 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
597 struct tgt_thread_info *tti = tgt_th_info(env);
604 "%s: mount_count is "LPU64", last_transno is "LPU64"\n",
605 tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
606 tgt->lut_last_transno);
608 /* Always save latest transno to keep it fresh */
609 spin_lock(&tgt->lut_translock);
610 tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
611 spin_unlock(&tgt->lut_translock);
613 th = dt_trans_create(env, tgt->lut_bottom);
620 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
621 &tti->tti_buf, tti->tti_off, th);
625 rc = dt_trans_start(env, tgt->lut_bottom, th);
629 rc = tgt_server_data_write(env, tgt, th);
631 dt_trans_stop(env, tgt->lut_bottom, th);
633 CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
634 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
635 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
638 EXPORT_SYMBOL(tgt_server_data_update);
640 int tgt_truncate_last_rcvd(const struct lu_env *env, struct lu_target *tgt,
643 struct dt_object *dt = tgt->lut_last_rcvd;
651 attr.la_valid = LA_SIZE;
653 th = dt_trans_create(env, tgt->lut_bottom);
656 rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
659 rc = dt_declare_attr_set(env, dt, &attr, th);
662 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
666 rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
668 rc = dt_attr_set(env, dt, &attr, th);
671 dt_trans_stop(env, tgt->lut_bottom, th);
676 static void tgt_client_epoch_update(const struct lu_env *env,
677 struct obd_export *exp)
679 struct lsd_client_data *lcd = exp->exp_target_data.ted_lcd;
680 struct lu_target *tgt = class_exp2tgt(exp);
682 LASSERT(tgt && tgt->lut_bottom);
683 /** VBR: set client last_epoch to current epoch */
684 if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
686 lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
687 tgt_client_data_update(env, exp);
691 * Update boot epoch when recovery ends
693 void tgt_boot_epoch_update(struct lu_target *tgt)
696 struct ptlrpc_request *req;
698 struct list_head client_list;
701 if (tgt->lut_obd->obd_stopping)
704 rc = lu_env_init(&env, LCT_DT_THREAD);
706 CERROR("%s: can't initialize environment: rc = %d\n",
707 tgt->lut_obd->obd_name, rc);
711 spin_lock(&tgt->lut_translock);
712 start_epoch = lr_epoch(tgt->lut_last_transno) + 1;
713 tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
714 tgt->lut_lsd.lsd_start_epoch = start_epoch;
715 spin_unlock(&tgt->lut_translock);
717 INIT_LIST_HEAD(&client_list);
719 * The recovery is not yet finished and final queue can still be updated
720 * with resend requests. Move final list to separate one for processing
722 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
723 list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
724 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
727 * go through list of exports participated in recovery and
728 * set new epoch for them
730 list_for_each_entry(req, &client_list, rq_list) {
731 LASSERT(!req->rq_export->exp_delayed);
732 if (!req->rq_export->exp_vbr_failed)
733 tgt_client_epoch_update(&env, req->rq_export);
735 /** return list back at once */
736 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
737 list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
738 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
740 /** Clear MULTI RPCS incompatibility flag if
741 * - target is MDT and
742 * - there is no client to recover or the recovery was aborted
744 if (!strncmp(tgt->lut_obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) &&
745 (tgt->lut_obd->obd_max_recoverable_clients == 0 ||
746 tgt->lut_obd->obd_abort_recovery))
747 tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
749 /** update server epoch */
750 tgt_server_data_update(&env, tgt, 1);
755 * commit callback, need to update last_committed value
757 struct tgt_last_committed_callback {
758 struct dt_txn_commit_cb llcc_cb;
759 struct lu_target *llcc_tgt;
760 struct obd_export *llcc_exp;
764 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
765 struct dt_txn_commit_cb *cb, int err)
767 struct tgt_last_committed_callback *ccb;
769 ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
771 LASSERT(ccb->llcc_exp);
772 LASSERT(ccb->llcc_tgt != NULL);
773 LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
775 /* Fast path w/o spinlock, if exp_last_committed was updated
776 * with higher transno, no need to take spinlock and check,
777 * also no need to update obd_last_committed. */
778 if (ccb->llcc_transno <= ccb->llcc_exp->exp_last_committed)
780 spin_lock(&ccb->llcc_tgt->lut_translock);
781 if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
782 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
784 if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
785 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
786 spin_unlock(&ccb->llcc_tgt->lut_translock);
788 ptlrpc_commit_replies(ccb->llcc_exp);
789 tgt_cancel_slc_locks(ccb->llcc_transno);
791 spin_unlock(&ccb->llcc_tgt->lut_translock);
794 class_export_cb_put(ccb->llcc_exp);
795 if (ccb->llcc_transno)
796 CDEBUG(D_HA, "%s: transno "LPD64" is committed\n",
797 ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
801 int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
802 struct obd_export *exp, __u64 transno)
804 struct tgt_last_committed_callback *ccb;
805 struct dt_txn_commit_cb *dcb;
813 ccb->llcc_exp = class_export_cb_get(exp);
814 ccb->llcc_transno = transno;
817 dcb->dcb_func = tgt_cb_last_committed;
818 INIT_LIST_HEAD(&dcb->dcb_linkage);
819 strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
821 rc = dt_trans_cb_add(th, dcb);
823 class_export_cb_put(exp);
827 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
828 /* report failure to force synchronous operation */
834 struct tgt_new_client_callback {
835 struct dt_txn_commit_cb lncc_cb;
836 struct obd_export *lncc_exp;
839 static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
840 struct dt_txn_commit_cb *cb, int err)
842 struct tgt_new_client_callback *ccb;
844 ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
846 LASSERT(ccb->lncc_exp->exp_obd);
848 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
849 ccb->lncc_exp->exp_obd->obd_name,
850 ccb->lncc_exp->exp_client_uuid.uuid);
852 spin_lock(&ccb->lncc_exp->exp_lock);
854 ccb->lncc_exp->exp_need_sync = 0;
856 spin_unlock(&ccb->lncc_exp->exp_lock);
857 class_export_cb_put(ccb->lncc_exp);
862 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
864 struct tgt_new_client_callback *ccb;
865 struct dt_txn_commit_cb *dcb;
872 ccb->lncc_exp = class_export_cb_get(exp);
875 dcb->dcb_func = tgt_cb_new_client;
876 INIT_LIST_HEAD(&dcb->dcb_linkage);
877 strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
879 rc = dt_trans_cb_add(th, dcb);
881 class_export_cb_put(exp);
888 * Add new client to the last_rcvd upon new connection.
890 * We use a bitmap to locate a free space in the last_rcvd file and initialize
893 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
895 struct tg_export_data *ted = &exp->exp_target_data;
896 struct lu_target *tgt = class_exp2tgt(exp);
901 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
902 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
905 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
908 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
909 * there's no need for extra complication here
911 idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
913 if (idx >= LR_MAX_CLIENTS ||
914 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
915 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
916 tgt->lut_obd->obd_name, idx);
919 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
920 idx = find_next_zero_bit(tgt->lut_client_bitmap,
921 LR_MAX_CLIENTS, idx);
925 ted->ted_lr_idx = idx;
926 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
927 idx * tgt->lut_lsd.lsd_client_size;
929 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
931 if (tgt_is_multimodrpcs_client(exp)) {
932 /* Set MULTI RPCS incompatibility flag to prevent previous
933 * Lustre versions to mount a target with reply_data file */
934 atomic_inc(&tgt->lut_num_clients);
935 if (!(tgt->lut_lsd.lsd_feature_incompat &
936 OBD_INCOMPAT_MULTI_RPCS)) {
937 tgt->lut_lsd.lsd_feature_incompat |=
938 OBD_INCOMPAT_MULTI_RPCS;
939 rc = tgt_server_data_update(env, tgt, 1);
941 CERROR("%s: unable to set MULTI RPCS "
942 "incompatibility flag\n",
943 exp->exp_obd->obd_name);
948 /* assign client slot generation */
949 ted->ted_lcd->lcd_generation =
950 atomic_inc_return(&tgt->lut_client_generation);
952 ted->ted_lcd->lcd_generation = 0;
955 CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s' "
957 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
958 ted->ted_lcd->lcd_uuid, ted->ted_lcd->lcd_generation);
960 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
963 rc = tgt_client_data_update(env, exp);
965 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
966 tgt->lut_obd->obd_name, idx, rc);
970 EXPORT_SYMBOL(tgt_client_new);
972 /* Add an existing client to the MDS in-memory state based on
973 * a client that was previously found in the last_rcvd file and
974 * already has an assigned slot (idx >= 0).
976 * It should not be possible to fail adding an existing client - otherwise
977 * mdt_init_server_data() callsite needs to be fixed.
979 int tgt_client_add(const struct lu_env *env, struct obd_export *exp, int idx)
981 struct tg_export_data *ted = &exp->exp_target_data;
982 struct lu_target *tgt = class_exp2tgt(exp);
986 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
987 LASSERTF(idx >= 0, "%d\n", idx);
989 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
990 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
993 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
994 CERROR("%s: client %d: bit already set in bitmap!!\n",
995 tgt->lut_obd->obd_name, idx);
998 atomic_inc(&tgt->lut_num_clients);
1000 CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added, "
1002 tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid,
1003 ted->ted_lcd->lcd_generation);
1005 ted->ted_lr_idx = idx;
1006 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
1007 idx * tgt->lut_lsd.lsd_client_size;
1009 mutex_init(&ted->ted_lcd_lock);
1011 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1016 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
1018 struct tg_export_data *ted = &exp->exp_target_data;
1019 struct lu_target *tgt = class_exp2tgt(exp);
1024 LASSERT(ted->ted_lcd);
1026 if (unlikely(tgt == NULL)) {
1027 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1028 class_exp2obd(exp)->obd_name);
1032 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
1033 if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
1034 (char *)tgt->lut_obd->obd_uuid.uuid) ||
1035 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1038 CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
1039 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1040 ted->ted_lcd->lcd_uuid);
1042 /* Clear the bit _after_ zeroing out the client so we don't
1043 race with filter_client_add and zero out new clients.*/
1044 if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
1045 CERROR("%s: client %u: bit already clear in bitmap!!\n",
1046 tgt->lut_obd->obd_name, ted->ted_lr_idx);
1050 /* Do not erase record for recoverable client. */
1051 if (exp->exp_flags & OBD_OPT_FAILOVER)
1054 /* Make sure the server's last_transno is up to date.
1055 * This should be done before zeroing client slot so last_transno will
1056 * be in server data or in client data in case of failure */
1057 rc = tgt_server_data_update(env, tgt, 0);
1059 CERROR("%s: failed to update server data, skip client %s "
1060 "zeroing, rc %d\n", tgt->lut_obd->obd_name,
1061 ted->ted_lcd->lcd_uuid, rc);
1065 mutex_lock(&ted->ted_lcd_lock);
1066 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
1067 rc = tgt_client_data_update(env, exp);
1068 mutex_unlock(&ted->ted_lcd_lock);
1070 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
1071 "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
1072 tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
1073 ted->ted_lr_idx, ted->ted_lr_off, rc);
1076 EXPORT_SYMBOL(tgt_client_del);
1078 int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
1079 struct tg_export_data *ted, struct tg_reply_data *trd,
1080 struct thandle *th, bool update_lrd_file)
1082 struct lsd_reply_data *lrd;
1085 lrd = &trd->trd_reply;
1086 /* update export last transno */
1087 mutex_lock(&ted->ted_lcd_lock);
1088 if (lrd->lrd_transno > ted->ted_lcd->lcd_last_transno)
1089 ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
1090 mutex_unlock(&ted->ted_lcd_lock);
1092 /* find a empty slot */
1093 i = tgt_find_free_reply_slot(tgt);
1094 if (unlikely(i < 0)) {
1095 CERROR("%s: couldn't find a slot for reply data: "
1096 "rc = %d\n", tgt_name(tgt), i);
1101 if (update_lrd_file) {
1105 /* write reply data to disk */
1106 off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
1107 rc = tgt_reply_data_write(env, tgt, lrd, off, th);
1108 if (unlikely(rc != 0)) {
1109 CERROR("%s: can't update %s file: rc = %d\n",
1110 tgt_name(tgt), REPLY_DATA, rc);
1114 /* add reply data to target export's reply list */
1115 mutex_lock(&ted->ted_lcd_lock);
1116 list_add(&trd->trd_list, &ted->ted_reply_list);
1117 ted->ted_reply_cnt++;
1118 if (ted->ted_reply_cnt > ted->ted_reply_max)
1119 ted->ted_reply_max = ted->ted_reply_cnt;
1120 mutex_unlock(&ted->ted_lcd_lock);
1122 CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
1123 "tag %hu, client gen %u, slot idx %d\n",
1124 trd, lrd->lrd_xid, lrd->lrd_transno,
1125 trd->trd_tag, lrd->lrd_client_gen, i);
1128 EXPORT_SYMBOL(tgt_add_reply_data);
1131 * last_rcvd & last_committed update callbacks
1133 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
1134 struct dt_object *obj, __u64 opdata,
1135 struct thandle *th, struct ptlrpc_request *req)
1137 struct tgt_thread_info *tti = tgt_th_info(env);
1138 struct tgt_session_info *tsi = tgt_ses_info(env);
1139 struct obd_export *exp = tsi->tsi_exp;
1140 struct tg_export_data *ted;
1148 LASSERT(exp != NULL);
1149 ted = &exp->exp_target_data;
1151 lw_client = exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT;
1152 if (ted->ted_lr_idx < 0 && !lw_client)
1153 /* ofd connect may cause transaction before export has
1158 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
1160 /* From update replay, tti_transno should be set already */
1161 LASSERT(tti->tti_transno != 0);
1163 spin_lock(&tgt->lut_translock);
1164 if (th->th_result != 0) {
1165 if (tti->tti_transno != 0) {
1166 CERROR("%s: replay transno "LPU64" failed: rc = %d\n",
1167 tgt_name(tgt), tti->tti_transno, th->th_result);
1169 } else if (tti->tti_transno == 0) {
1170 tti->tti_transno = ++tgt->lut_last_transno;
1172 /* should be replay */
1173 if (tti->tti_transno > tgt->lut_last_transno)
1174 tgt->lut_last_transno = tti->tti_transno;
1176 spin_unlock(&tgt->lut_translock);
1178 /** VBR: set new versions */
1179 if (th->th_result == 0 && obj != NULL) {
1180 struct dt_object *dto = dt_object_locate(obj, th->th_dev);
1181 dt_version_set(env, dto, tti->tti_transno, th);
1184 /* filling reply data */
1185 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
1186 tti->tti_transno, tgt->lut_obd->obd_last_committed);
1189 req->rq_transno = tti->tti_transno;
1190 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
1193 /* if can't add callback, do sync write */
1194 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp, tti->tti_transno);
1197 /* All operations performed by LW clients are synchronous and
1198 * we store the committed transno in the last_rcvd header */
1199 spin_lock(&tgt->lut_translock);
1200 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
1201 tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
1202 spin_unlock(&tgt->lut_translock);
1203 /* Although lightweight (LW) connections have no slot
1204 * in the last_rcvd, we still want to maintain
1205 * the in-memory lsd_client_data structure in order to
1206 * properly handle reply reconstruction. */
1207 rc = tgt_server_data_write(env, tgt, th);
1209 spin_unlock(&tgt->lut_translock);
1211 } else if (ted->ted_lr_off == 0) {
1212 CERROR("%s: client idx %d has offset %lld\n",
1213 tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
1217 /* Target that supports multiple reply data */
1218 if (tgt_is_multimodrpcs_client(exp)) {
1219 struct tg_reply_data *trd;
1220 struct lsd_reply_data *lrd;
1221 __u64 *pre_versions;
1225 if (unlikely(trd == NULL))
1228 /* fill reply data information */
1229 lrd = &trd->trd_reply;
1230 lrd->lrd_transno = tti->tti_transno;
1232 lrd->lrd_xid = req->rq_xid;
1233 trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
1234 pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1235 lrd->lrd_result = th->th_result;
1236 lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
1237 write_update = true;
1239 LASSERT(tsi->tsi_xid != 0);
1240 lrd->lrd_xid = tsi->tsi_xid;
1241 lrd->lrd_result = tsi->tsi_result;
1242 lrd->lrd_client_gen = tsi->tsi_client_gen;
1244 pre_versions = NULL;
1245 write_update = false;
1248 lrd->lrd_data = opdata;
1250 trd->trd_pre_versions[0] = pre_versions[0];
1251 trd->trd_pre_versions[1] = pre_versions[1];
1252 trd->trd_pre_versions[2] = pre_versions[2];
1253 trd->trd_pre_versions[3] = pre_versions[3];
1256 rc = tgt_add_reply_data(env, tgt, ted, trd, th, write_update);
1262 /* Enough for update replay, let's return */
1266 mutex_lock(&ted->ted_lcd_lock);
1267 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1268 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
1269 transno_p = &ted->ted_lcd->lcd_last_close_transno;
1270 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
1271 ted->ted_lcd->lcd_last_close_result = th->th_result;
1273 /* VBR: save versions in last_rcvd for reconstruct. */
1274 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1277 ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
1278 ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
1279 ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
1280 ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
1282 transno_p = &ted->ted_lcd->lcd_last_transno;
1283 ted->ted_lcd->lcd_last_xid = req->rq_xid;
1284 ted->ted_lcd->lcd_last_result = th->th_result;
1285 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
1286 * see struct ldlm_reply->lock_policy_res1; */
1287 ted->ted_lcd->lcd_last_data = opdata;
1290 /* Update transno in slot only if non-zero number, i.e. no errors */
1291 if (likely(tti->tti_transno != 0)) {
1292 /* Don't overwrite bigger transaction number with lower one.
1293 * That is not sign of problem in all cases, but in any case
1294 * this value should be monotonically increased only. */
1295 if (*transno_p > tti->tti_transno) {
1296 if (!tgt->lut_no_reconstruct) {
1297 CERROR("%s: trying to overwrite bigger transno:"
1298 "on-disk: "LPU64", new: "LPU64" replay: "
1299 "%d. See LU-617.\n", tgt_name(tgt),
1300 *transno_p, tti->tti_transno,
1301 req_is_replay(req));
1302 if (req_is_replay(req)) {
1303 spin_lock(&req->rq_export->exp_lock);
1304 req->rq_export->exp_vbr_failed = 1;
1305 spin_unlock(&req->rq_export->exp_lock);
1307 mutex_unlock(&ted->ted_lcd_lock);
1308 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
1311 *transno_p = tti->tti_transno;
1316 tti->tti_off = ted->ted_lr_off;
1317 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1319 mutex_unlock(&ted->ted_lcd_lock);
1323 mutex_unlock(&ted->ted_lcd_lock);
1328 * last_rcvd update for echo client simulation.
1329 * It updates last_rcvd client slot and version of object in
1330 * simple way but with all locks to simulate all drawbacks
1332 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
1333 struct lu_target *tgt,
1334 struct dt_object *obj,
1336 struct obd_export *exp)
1338 struct tgt_thread_info *tti = tgt_th_info(env);
1339 struct tg_export_data *ted = &exp->exp_target_data;
1344 tti->tti_transno = 0;
1346 spin_lock(&tgt->lut_translock);
1347 if (th->th_result == 0)
1348 tti->tti_transno = ++tgt->lut_last_transno;
1349 spin_unlock(&tgt->lut_translock);
1351 /** VBR: set new versions */
1352 if (th->th_result == 0 && obj != NULL)
1353 dt_version_set(env, obj, tti->tti_transno, th);
1355 /* if can't add callback, do sync write */
1356 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
1359 LASSERT(ted->ted_lr_off > 0);
1361 mutex_lock(&ted->ted_lcd_lock);
1362 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1363 ted->ted_lcd->lcd_last_transno = tti->tti_transno;
1364 ted->ted_lcd->lcd_last_result = th->th_result;
1366 tti->tti_off = ted->ted_lr_off;
1367 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1368 mutex_unlock(&ted->ted_lcd_lock);
1372 static int tgt_clients_data_init(const struct lu_env *env,
1373 struct lu_target *tgt,
1374 unsigned long last_size)
1376 struct obd_device *obd = tgt->lut_obd;
1377 struct lr_server_data *lsd = &tgt->lut_lsd;
1378 struct lsd_client_data *lcd = NULL;
1379 struct tg_export_data *ted;
1382 loff_t off = lsd->lsd_client_start;
1383 __u32 generation = 0;
1384 struct cfs_hash *hash = NULL;
1388 CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
1389 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
1395 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1397 GOTO(err_out, rc = -ENODEV);
1399 for (cl_idx = 0; off < last_size; cl_idx++) {
1400 struct obd_export *exp;
1403 /* Don't assume off is incremented properly by
1404 * read_record(), in case sizeof(*lcd)
1405 * isn't the same as fsd->lsd_client_size. */
1406 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
1407 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
1409 CERROR("%s: error reading last_rcvd %s idx %d off "
1410 "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
1413 break; /* read error shouldn't cause startup to fail */
1416 if (lcd->lcd_uuid[0] == '\0') {
1417 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
1422 last_transno = lcd_last_transno(lcd);
1424 /* These exports are cleaned up by disconnect, so they
1425 * need to be set up like real exports as connect does.
1427 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
1428 " srv lr: "LPU64" lx: "LPU64" gen %u\n", lcd->lcd_uuid,
1429 cl_idx, last_transno, lsd->lsd_last_transno,
1430 lcd_last_xid(lcd), lcd->lcd_generation);
1432 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
1434 if (PTR_ERR(exp) == -EALREADY) {
1435 /* export already exists, zero out this one */
1436 CERROR("%s: Duplicate export %s!\n",
1437 tgt_name(tgt), lcd->lcd_uuid);
1440 GOTO(err_out, rc = PTR_ERR(exp));
1443 ted = &exp->exp_target_data;
1444 *ted->ted_lcd = *lcd;
1446 rc = tgt_client_add(env, exp, cl_idx);
1447 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
1448 /* VBR: set export last committed version */
1449 exp->exp_last_committed = last_transno;
1450 spin_lock(&exp->exp_lock);
1451 exp->exp_connecting = 0;
1452 exp->exp_in_recovery = 0;
1453 spin_unlock(&exp->exp_lock);
1454 obd->obd_max_recoverable_clients++;
1456 if (tgt->lut_lsd.lsd_feature_incompat &
1457 OBD_INCOMPAT_MULTI_RPCS &&
1458 lcd->lcd_generation != 0) {
1459 /* compute the highest valid client generation */
1460 generation = max(generation, lcd->lcd_generation);
1461 /* fill client_generation <-> export hash table */
1462 rc = cfs_hash_add_unique(hash, &lcd->lcd_generation,
1463 &exp->exp_gen_hash);
1465 CERROR("%s: duplicate export for client "
1467 tgt_name(tgt), lcd->lcd_generation);
1468 class_export_put(exp);
1473 class_export_put(exp);
1475 rc = rev_import_init(exp);
1477 class_unlink_export(exp);
1481 /* Need to check last_rcvd even for duplicated exports. */
1482 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
1483 cl_idx, last_transno);
1485 spin_lock(&tgt->lut_translock);
1486 tgt->lut_last_transno = max(last_transno,
1487 tgt->lut_last_transno);
1488 spin_unlock(&tgt->lut_translock);
1491 /* record highest valid client generation */
1492 atomic_set(&tgt->lut_client_generation, generation);
1496 cfs_hash_putref(hash);
1501 struct server_compat_data {
1508 static struct server_compat_data tgt_scd[] = {
1509 [LDD_F_SV_TYPE_MDT] = {
1510 .rocompat = OBD_ROCOMPAT_LOVOBJID,
1511 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1512 OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
1513 OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI |
1514 OBD_INCOMPAT_MULTI_RPCS,
1515 .rocinit = OBD_ROCOMPAT_LOVOBJID,
1516 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1517 OBD_INCOMPAT_MULTI_OI,
1519 [LDD_F_SV_TYPE_OST] = {
1520 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
1521 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
1523 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
1524 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1528 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1530 struct tgt_thread_info *tti = tgt_th_info(env);
1531 struct lr_server_data *lsd = &tgt->lut_lsd;
1532 unsigned long last_rcvd_size;
1536 rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1540 last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1542 /* ensure padding in the struct is the correct size */
1543 CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
1544 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
1546 rc = server_name2index(tgt_name(tgt), &index, NULL);
1548 CERROR("%s: Can not get index from name: rc = %d\n",
1552 /* server_name2index() returns type */
1554 if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1555 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1559 /* last_rcvd on OST doesn't provide reconstruct support because there
1560 * may be up to 8 in-flight write requests per single slot in
1561 * last_rcvd client data
1563 tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1565 if (last_rcvd_size == 0) {
1566 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1568 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1569 sizeof(lsd->lsd_uuid));
1570 lsd->lsd_last_transno = 0;
1571 lsd->lsd_mount_count = 0;
1572 lsd->lsd_server_size = LR_SERVER_SIZE;
1573 lsd->lsd_client_start = LR_CLIENT_START;
1574 lsd->lsd_client_size = LR_CLIENT_SIZE;
1575 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1576 lsd->lsd_osd_index = index;
1577 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1578 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1580 rc = tgt_server_data_read(env, tgt);
1582 CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1586 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1587 LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
1588 "using the wrong disk %s. Were the"
1589 " /dev/ assignments rearranged?\n",
1590 tgt->lut_obd->obd_uuid.uuid,
1595 if (lsd->lsd_osd_index != index) {
1596 LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
1597 "is different with the index %d in"
1598 "config log, It might be disk"
1599 "corruption!\n", tgt_name(tgt),
1600 lsd->lsd_osd_index, index);
1605 if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1606 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1608 lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1612 if (type == LDD_F_SV_TYPE_MDT)
1613 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1615 if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1616 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1618 lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1621 /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1622 if (type == LDD_F_SV_TYPE_MDT &&
1623 !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1624 if (last_rcvd_size > lsd->lsd_client_start) {
1625 LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1626 "remove all clients for interop needs\n",
1628 rc = tgt_truncate_last_rcvd(env, tgt,
1629 lsd->lsd_client_start);
1632 last_rcvd_size = lsd->lsd_client_start;
1634 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1635 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1638 spin_lock(&tgt->lut_translock);
1639 tgt->lut_last_transno = lsd->lsd_last_transno;
1640 spin_unlock(&tgt->lut_translock);
1642 lsd->lsd_mount_count++;
1644 CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1645 CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
1646 tgt_name(tgt), tgt->lut_last_transno);
1647 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
1648 tgt_name(tgt), lsd->lsd_mount_count);
1649 CDEBUG(D_INODE, "%s: server data size: %u\n",
1650 tgt_name(tgt), lsd->lsd_server_size);
1651 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1652 tgt_name(tgt), lsd->lsd_client_start);
1653 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1654 tgt_name(tgt), lsd->lsd_client_size);
1655 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1656 tgt_name(tgt), last_rcvd_size);
1657 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1658 tgt_name(tgt), lsd->lsd_subdir_count);
1659 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1660 last_rcvd_size <= lsd->lsd_client_start ? 0 :
1661 (last_rcvd_size - lsd->lsd_client_start) /
1662 lsd->lsd_client_size);
1663 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1665 if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1666 lsd->lsd_client_size == 0) {
1667 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1671 if (!tgt->lut_obd->obd_replayable)
1672 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1674 rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1676 GOTO(err_client, rc);
1678 spin_lock(&tgt->lut_translock);
1679 /* obd_last_committed is used for compatibility
1680 * with other lustre recovery code */
1681 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1682 spin_unlock(&tgt->lut_translock);
1684 tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1685 tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1687 /* save it, so mount count and last_transno is current */
1688 rc = tgt_server_data_update(env, tgt, 0);
1690 GOTO(err_client, rc);
1695 class_disconnect_exports(tgt->lut_obd);
1699 /* add credits for last_rcvd update */
1700 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1703 struct lu_target *tgt = cookie;
1704 struct tgt_session_info *tsi;
1705 struct tgt_thread_info *tti = tgt_th_info(env);
1706 struct dt_object *dto;
1709 /* if there is no session, then this transaction is not result of
1710 * request processing but some local operation */
1711 if (env->le_ses == NULL)
1714 LASSERT(tgt->lut_last_rcvd);
1715 tsi = tgt_ses_info(env);
1716 /* OFD may start transaction without export assigned */
1717 if (tsi->tsi_exp == NULL)
1720 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
1723 rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1724 tsi->tsi_exp->exp_target_data.ted_lr_off,
1730 rc = dt_declare_record_write(env, dto, &tti->tti_buf, 0, th);
1734 if (tgt_is_multimodrpcs_client(tsi->tsi_exp)) {
1736 * Use maximum possible file offset for declaration to ensure
1737 * ZFS will reserve enough credits for a write anywhere in this
1738 * file, since we don't know where in the file the write will be
1739 * because a replay slot has not been assigned. This should be
1740 * replaced by dmu_tx_hold_append() when available.
1742 tti->tti_off = atomic_read(&tgt->lut_num_clients) * 8 *
1743 sizeof(struct lsd_reply_data);
1744 tti->tti_buf.lb_buf = NULL;
1745 tti->tti_buf.lb_len = sizeof(struct lsd_reply_data);
1746 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
1747 rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1753 if (tsi->tsi_vbr_obj != NULL &&
1754 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1755 dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
1756 rc = dt_declare_version_set(env, dto, th);
1762 /* Update last_rcvd records with latests transaction data */
1763 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1766 struct lu_target *tgt = cookie;
1767 struct tgt_session_info *tsi;
1768 struct tgt_thread_info *tti = tgt_th_info(env);
1769 struct dt_object *obj = NULL;
1773 if (env->le_ses == NULL)
1776 tsi = tgt_ses_info(env);
1777 /* OFD may start transaction without export assigned */
1778 if (tsi->tsi_exp == NULL)
1781 echo_client = (tgt_ses_req(tsi) == NULL && tsi->tsi_xid == 0);
1783 if (tti->tti_has_trans && !echo_client) {
1784 if (tti->tti_mult_trans == 0) {
1785 CDEBUG(D_HA, "More than one transaction "LPU64"\n",
1789 /* we need another transno to be assigned */
1790 tti->tti_transno = 0;
1791 } else if (th->th_result == 0) {
1792 tti->tti_has_trans = 1;
1795 if (tsi->tsi_vbr_obj != NULL &&
1796 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1797 obj = tsi->tsi_vbr_obj;
1800 if (unlikely(echo_client)) /* echo client special case */
1801 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1804 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
1809 int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
1811 struct tgt_thread_info *tti = tgt_th_info(env);
1812 struct lsd_reply_data *lrd = &tti->tti_lrd;
1813 unsigned long reply_data_size;
1815 struct lsd_reply_header *lrh = NULL;
1816 struct lsd_client_data *lcd = NULL;
1817 struct tg_reply_data *trd = NULL;
1820 struct cfs_hash *hash = NULL;
1821 struct obd_export *exp;
1822 struct tg_export_data *ted;
1823 int reply_data_recovered = 0;
1825 rc = dt_attr_get(env, tgt->lut_reply_data, &tti->tti_attr);
1828 reply_data_size = (unsigned long)tti->tti_attr.la_size;
1832 GOTO(out, rc = -ENOMEM);
1834 if (reply_data_size == 0) {
1835 CDEBUG(D_INFO, "%s: new reply_data file, initializing\n",
1837 lrh->lrh_magic = LRH_MAGIC;
1838 lrh->lrh_header_size = sizeof(struct lsd_reply_header);
1839 lrh->lrh_reply_size = sizeof(struct lsd_reply_data);
1840 rc = tgt_reply_header_write(env, tgt, lrh);
1842 CERROR("%s: error writing %s: rc = %d\n",
1843 tgt_name(tgt), REPLY_DATA, rc);
1847 rc = tgt_reply_header_read(env, tgt, lrh);
1849 CERROR("%s: error reading %s: rc = %d\n",
1850 tgt_name(tgt), REPLY_DATA, rc);
1853 if (lrh->lrh_magic != LRH_MAGIC ||
1854 lrh->lrh_header_size != sizeof(struct lsd_reply_header) ||
1855 lrh->lrh_reply_size != sizeof(struct lsd_reply_data)) {
1856 CERROR("%s: invalid header in %s\n",
1857 tgt_name(tgt), REPLY_DATA);
1858 GOTO(out, rc = -EINVAL);
1861 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1863 GOTO(out, rc = -ENODEV);
1867 GOTO(out, rc = -ENOMEM);
1871 GOTO(out, rc = -ENOMEM);
1873 /* Load reply_data from disk */
1874 for (idx = 0, off = sizeof(struct lsd_reply_header);
1875 off < reply_data_size;
1876 idx++, off += sizeof(struct lsd_reply_data)) {
1877 rc = tgt_reply_data_read(env, tgt, lrd, off);
1879 CERROR("%s: error reading %s: rc = %d\n",
1880 tgt_name(tgt), REPLY_DATA, rc);
1884 exp = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
1886 /* old reply data from a disconnected client */
1889 ted = &exp->exp_target_data;
1890 mutex_lock(&ted->ted_lcd_lock);
1892 /* create in-memory reply_data and link it to
1893 * target export's reply list */
1894 tgt_set_reply_slot(tgt, idx);
1895 trd->trd_reply = *lrd;
1896 trd->trd_pre_versions[0] = 0;
1897 trd->trd_pre_versions[1] = 0;
1898 trd->trd_pre_versions[2] = 0;
1899 trd->trd_pre_versions[3] = 0;
1900 trd->trd_index = idx;
1902 list_add(&trd->trd_list, &ted->ted_reply_list);
1903 ted->ted_reply_cnt++;
1904 if (ted->ted_reply_cnt > ted->ted_reply_max)
1905 ted->ted_reply_max = ted->ted_reply_cnt;
1907 CDEBUG(D_HA, "%s: restore reply %p: xid %llu, "
1908 "transno %llu, client gen %u, slot idx %d\n",
1909 tgt_name(tgt), trd, lrd->lrd_xid,
1910 lrd->lrd_transno, lrd->lrd_client_gen,
1913 /* update export last committed transation */
1914 exp->exp_last_committed = max(exp->exp_last_committed,
1917 mutex_unlock(&ted->ted_lcd_lock);
1918 class_export_put(exp);
1920 /* update target last committed transaction */
1921 spin_lock(&tgt->lut_translock);
1922 tgt->lut_last_transno = max(tgt->lut_last_transno,
1924 spin_unlock(&tgt->lut_translock);
1926 reply_data_recovered++;
1930 GOTO(out, rc = -ENOMEM);
1932 CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
1933 tgt_name(tgt), reply_data_recovered);
1936 spin_lock(&tgt->lut_translock);
1937 /* obd_last_committed is used for compatibility
1938 * with other lustre recovery code */
1939 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1940 spin_unlock(&tgt->lut_translock);
1946 cfs_hash_putref(hash);
1956 struct tg_reply_data *tgt_lookup_reply_by_xid(struct tg_export_data *ted,
1959 struct tg_reply_data *found = NULL;
1960 struct tg_reply_data *reply;
1962 mutex_lock(&ted->ted_lcd_lock);
1963 list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
1964 if (reply->trd_reply.lrd_xid == xid) {
1969 mutex_unlock(&ted->ted_lcd_lock);
1972 EXPORT_SYMBOL(tgt_lookup_reply_by_xid);
1974 /* Look for a reply data matching specified request @req
1975 * A copy is returned in @trd if the pointer is not NULL
1977 bool tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
1979 struct tg_export_data *ted = &req->rq_export->exp_target_data;
1980 struct tg_reply_data *reply;
1983 reply = tgt_lookup_reply_by_xid(ted, req->rq_xid);
1984 if (reply != NULL) {
1990 CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d\n",
1991 tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid,
1996 EXPORT_SYMBOL(tgt_lookup_reply);
1998 int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
2000 struct tg_export_data *ted = &exp->exp_target_data;
2001 struct lu_target *lut = class_exp2tgt(exp);
2002 struct tg_reply_data *trd, *tmp;
2004 mutex_lock(&ted->ted_lcd_lock);
2005 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
2006 if (trd->trd_reply.lrd_xid > rcvd_xid)
2008 ted->ted_release_xid++;
2009 tgt_release_reply_data(lut, ted, trd);
2011 mutex_unlock(&ted->ted_lcd_lock);
2016 int tgt_handle_tag(struct obd_export *exp, __u16 tag)
2018 struct tg_export_data *ted = &exp->exp_target_data;
2019 struct lu_target *lut = class_exp2tgt(exp);
2020 struct tg_reply_data *trd, *tmp;
2025 mutex_lock(&ted->ted_lcd_lock);
2026 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
2027 if (trd->trd_tag != tag)
2029 ted->ted_release_tag++;
2030 tgt_release_reply_data(lut, ted, trd);
2033 mutex_unlock(&ted->ted_lcd_lock);