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 OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
382 if (exp->exp_target_data.ted_lcd == NULL)
384 /* Mark that slot is not yet valid, 0 doesn't work here */
385 exp->exp_target_data.ted_lr_idx = -1;
386 INIT_LIST_HEAD(&exp->exp_target_data.ted_reply_list);
387 mutex_init(&exp->exp_target_data.ted_lcd_lock);
390 EXPORT_SYMBOL(tgt_client_alloc);
393 * Free in-memory data for client slot related to export.
395 void tgt_client_free(struct obd_export *exp)
397 struct tg_export_data *ted = &exp->exp_target_data;
398 struct lu_target *lut = class_exp2tgt(exp);
399 struct tg_reply_data *trd, *tmp;
401 LASSERT(exp != exp->exp_obd->obd_self_export);
403 /* free reply data */
404 mutex_lock(&ted->ted_lcd_lock);
405 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
406 tgt_release_reply_data(lut, ted, trd);
408 if (ted->ted_reply_last != NULL) {
409 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
410 ted->ted_reply_last = NULL;
412 mutex_unlock(&ted->ted_lcd_lock);
414 if (!hlist_unhashed(&exp->exp_gen_hash))
415 cfs_hash_del(exp->exp_obd->obd_gen_hash,
416 &ted->ted_lcd->lcd_generation,
419 OBD_FREE_PTR(ted->ted_lcd);
422 /* Target may have been freed (see LU-7430)
423 * Slot may be not yet assigned */
424 if (exp->exp_obd->u.obt.obt_magic != OBT_MAGIC ||
428 /* Clear bit when lcd is freed */
429 LASSERT(lut && lut->lut_client_bitmap);
430 if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
431 CERROR("%s: client %u bit already clear in bitmap\n",
432 exp->exp_obd->obd_name, ted->ted_lr_idx);
436 if (tgt_is_multimodrpcs_client(exp) && !exp->exp_obd->obd_stopping)
437 atomic_dec(&lut->lut_num_clients);
439 EXPORT_SYMBOL(tgt_client_free);
441 int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
442 struct lsd_client_data *lcd, loff_t *off, int index)
444 struct tgt_thread_info *tti = tgt_th_info(env);
448 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
450 check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
451 lcd_le_to_cpu(&tti->tti_lcd, lcd);
452 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
453 lcd->lcd_last_close_result =
454 ptlrpc_status_ntoh(lcd->lcd_last_close_result);
457 CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = "LPU64
458 ", last_xid = "LPU64", last_result = %u, last_data = %u, "
459 "last_close_transno = "LPU64", last_close_xid = "LPU64", "
460 "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
461 *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
462 lcd->lcd_last_result, lcd->lcd_last_data,
463 lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
464 lcd->lcd_last_close_result, rc);
468 int tgt_client_data_write(const struct lu_env *env, struct lu_target *tgt,
469 struct lsd_client_data *lcd, loff_t *off,
472 struct tgt_thread_info *tti = tgt_th_info(env);
473 struct dt_object *dto;
475 lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
476 lcd->lcd_last_close_result =
477 ptlrpc_status_hton(lcd->lcd_last_close_result);
478 lcd_cpu_to_le(lcd, &tti->tti_lcd);
481 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
482 return dt_record_write(env, dto, &tti->tti_buf, off, th);
486 * Update client data in last_rcvd
488 static int tgt_client_data_update(const struct lu_env *env,
489 struct obd_export *exp)
491 struct tg_export_data *ted = &exp->exp_target_data;
492 struct lu_target *tgt = class_exp2tgt(exp);
493 struct tgt_thread_info *tti = tgt_th_info(env);
499 if (unlikely(tgt == NULL)) {
500 CDEBUG(D_ERROR, "%s: No target for connected export\n",
501 class_exp2obd(exp)->obd_name);
505 th = dt_trans_create(env, tgt->lut_bottom);
510 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
512 ted->ted_lr_off, th);
516 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
520 * Until this operations will be committed the sync is needed
521 * for this export. This should be done _after_ starting the
522 * transaction so that many connecting clients will not bring
523 * server down with lots of sync writes.
525 rc = tgt_new_client_cb_add(th, exp);
527 /* can't add callback, do sync now */
530 spin_lock(&exp->exp_lock);
531 exp->exp_need_sync = 1;
532 spin_unlock(&exp->exp_lock);
535 tti->tti_off = ted->ted_lr_off;
536 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
539 dt_trans_stop(env, tgt->lut_bottom, th);
540 CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
541 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
542 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
547 int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
549 struct tgt_thread_info *tti = tgt_th_info(env);
554 rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
557 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
559 CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
560 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
561 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
565 int tgt_server_data_write(const struct lu_env *env, struct lu_target *tgt,
568 struct tgt_thread_info *tti = tgt_th_info(env);
569 struct dt_object *dto;
576 lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
578 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
579 rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
581 CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
582 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
583 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
589 * Update server data in last_rcvd
591 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
594 struct tgt_thread_info *tti = tgt_th_info(env);
601 "%s: mount_count is "LPU64", last_transno is "LPU64"\n",
602 tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
603 tgt->lut_last_transno);
605 /* Always save latest transno to keep it fresh */
606 spin_lock(&tgt->lut_translock);
607 tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
608 spin_unlock(&tgt->lut_translock);
610 th = dt_trans_create(env, tgt->lut_bottom);
617 rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
618 &tti->tti_buf, tti->tti_off, th);
622 rc = dt_trans_start(env, tgt->lut_bottom, th);
626 rc = tgt_server_data_write(env, tgt, th);
628 dt_trans_stop(env, tgt->lut_bottom, th);
630 CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
631 "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
632 tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
635 EXPORT_SYMBOL(tgt_server_data_update);
637 int tgt_truncate_last_rcvd(const struct lu_env *env, struct lu_target *tgt,
640 struct dt_object *dt = tgt->lut_last_rcvd;
648 attr.la_valid = LA_SIZE;
650 th = dt_trans_create(env, tgt->lut_bottom);
653 rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
656 rc = dt_declare_attr_set(env, dt, &attr, th);
659 rc = dt_trans_start_local(env, tgt->lut_bottom, th);
663 rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
665 rc = dt_attr_set(env, dt, &attr, th);
668 dt_trans_stop(env, tgt->lut_bottom, th);
673 static void tgt_client_epoch_update(const struct lu_env *env,
674 struct obd_export *exp)
676 struct lsd_client_data *lcd = exp->exp_target_data.ted_lcd;
677 struct lu_target *tgt = class_exp2tgt(exp);
679 LASSERT(tgt && tgt->lut_bottom);
680 /** VBR: set client last_epoch to current epoch */
681 if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
683 lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
684 tgt_client_data_update(env, exp);
688 * Update boot epoch when recovery ends
690 void tgt_boot_epoch_update(struct lu_target *tgt)
693 struct ptlrpc_request *req;
695 struct list_head client_list;
698 if (tgt->lut_obd->obd_stopping)
701 rc = lu_env_init(&env, LCT_DT_THREAD);
703 CERROR("%s: can't initialize environment: rc = %d\n",
704 tgt->lut_obd->obd_name, rc);
708 spin_lock(&tgt->lut_translock);
709 start_epoch = lr_epoch(tgt->lut_last_transno) + 1;
710 tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
711 tgt->lut_lsd.lsd_start_epoch = start_epoch;
712 spin_unlock(&tgt->lut_translock);
714 INIT_LIST_HEAD(&client_list);
716 * The recovery is not yet finished and final queue can still be updated
717 * with resend requests. Move final list to separate one for processing
719 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
720 list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
721 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
724 * go through list of exports participated in recovery and
725 * set new epoch for them
727 list_for_each_entry(req, &client_list, rq_list) {
728 LASSERT(!req->rq_export->exp_delayed);
729 if (!req->rq_export->exp_vbr_failed)
730 tgt_client_epoch_update(&env, req->rq_export);
732 /** return list back at once */
733 spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
734 list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
735 spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
737 /** Clear MULTI RPCS incompatibility flag if
738 * - target is MDT and
739 * - there is no client to recover or the recovery was aborted
741 if (!strncmp(tgt->lut_obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) &&
742 (tgt->lut_obd->obd_max_recoverable_clients == 0 ||
743 tgt->lut_obd->obd_abort_recovery))
744 tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
746 /** update server epoch */
747 tgt_server_data_update(&env, tgt, 1);
752 * commit callback, need to update last_committed value
754 struct tgt_last_committed_callback {
755 struct dt_txn_commit_cb llcc_cb;
756 struct lu_target *llcc_tgt;
757 struct obd_export *llcc_exp;
761 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
762 struct dt_txn_commit_cb *cb, int err)
764 struct tgt_last_committed_callback *ccb;
766 ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
768 LASSERT(ccb->llcc_exp);
769 LASSERT(ccb->llcc_tgt != NULL);
770 LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
772 /* Fast path w/o spinlock, if exp_last_committed was updated
773 * with higher transno, no need to take spinlock and check,
774 * also no need to update obd_last_committed. */
775 if (ccb->llcc_transno <= ccb->llcc_exp->exp_last_committed)
777 spin_lock(&ccb->llcc_tgt->lut_translock);
778 if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
779 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
781 if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
782 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
783 spin_unlock(&ccb->llcc_tgt->lut_translock);
785 ptlrpc_commit_replies(ccb->llcc_exp);
786 tgt_cancel_slc_locks(ccb->llcc_transno);
788 spin_unlock(&ccb->llcc_tgt->lut_translock);
791 class_export_cb_put(ccb->llcc_exp);
792 if (ccb->llcc_transno)
793 CDEBUG(D_HA, "%s: transno "LPD64" is committed\n",
794 ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
798 int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
799 struct obd_export *exp, __u64 transno)
801 struct tgt_last_committed_callback *ccb;
802 struct dt_txn_commit_cb *dcb;
810 ccb->llcc_exp = class_export_cb_get(exp);
811 ccb->llcc_transno = transno;
814 dcb->dcb_func = tgt_cb_last_committed;
815 INIT_LIST_HEAD(&dcb->dcb_linkage);
816 strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
818 rc = dt_trans_cb_add(th, dcb);
820 class_export_cb_put(exp);
824 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
825 /* report failure to force synchronous operation */
831 struct tgt_new_client_callback {
832 struct dt_txn_commit_cb lncc_cb;
833 struct obd_export *lncc_exp;
836 static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
837 struct dt_txn_commit_cb *cb, int err)
839 struct tgt_new_client_callback *ccb;
841 ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
843 LASSERT(ccb->lncc_exp->exp_obd);
845 CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
846 ccb->lncc_exp->exp_obd->obd_name,
847 ccb->lncc_exp->exp_client_uuid.uuid);
849 spin_lock(&ccb->lncc_exp->exp_lock);
851 ccb->lncc_exp->exp_need_sync = 0;
853 spin_unlock(&ccb->lncc_exp->exp_lock);
854 class_export_cb_put(ccb->lncc_exp);
859 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
861 struct tgt_new_client_callback *ccb;
862 struct dt_txn_commit_cb *dcb;
869 ccb->lncc_exp = class_export_cb_get(exp);
872 dcb->dcb_func = tgt_cb_new_client;
873 INIT_LIST_HEAD(&dcb->dcb_linkage);
874 strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
876 rc = dt_trans_cb_add(th, dcb);
878 class_export_cb_put(exp);
885 * Add new client to the last_rcvd upon new connection.
887 * We use a bitmap to locate a free space in the last_rcvd file and initialize
890 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
892 struct tg_export_data *ted = &exp->exp_target_data;
893 struct lu_target *tgt = class_exp2tgt(exp);
898 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
899 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
902 if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
905 /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
906 * there's no need for extra complication here
908 idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
910 if (idx >= LR_MAX_CLIENTS ||
911 OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
912 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
913 tgt->lut_obd->obd_name, idx);
916 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
917 idx = find_next_zero_bit(tgt->lut_client_bitmap,
918 LR_MAX_CLIENTS, idx);
922 ted->ted_lr_idx = idx;
923 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
924 idx * tgt->lut_lsd.lsd_client_size;
926 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
928 if (tgt_is_multimodrpcs_client(exp)) {
929 /* Set MULTI RPCS incompatibility flag to prevent previous
930 * Lustre versions to mount a target with reply_data file */
931 atomic_inc(&tgt->lut_num_clients);
932 if (!(tgt->lut_lsd.lsd_feature_incompat &
933 OBD_INCOMPAT_MULTI_RPCS)) {
934 tgt->lut_lsd.lsd_feature_incompat |=
935 OBD_INCOMPAT_MULTI_RPCS;
936 rc = tgt_server_data_update(env, tgt, 1);
938 CERROR("%s: unable to set MULTI RPCS "
939 "incompatibility flag\n",
940 exp->exp_obd->obd_name);
945 /* assign client slot generation */
946 ted->ted_lcd->lcd_generation =
947 atomic_inc_return(&tgt->lut_client_generation);
949 ted->ted_lcd->lcd_generation = 0;
952 CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s' "
954 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
955 ted->ted_lcd->lcd_uuid, ted->ted_lcd->lcd_generation);
957 if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
960 rc = tgt_client_data_update(env, exp);
962 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
963 tgt->lut_obd->obd_name, idx, rc);
967 EXPORT_SYMBOL(tgt_client_new);
969 /* Add an existing client to the MDS in-memory state based on
970 * a client that was previously found in the last_rcvd file and
971 * already has an assigned slot (idx >= 0).
973 * It should not be possible to fail adding an existing client - otherwise
974 * mdt_init_server_data() callsite needs to be fixed.
976 int tgt_client_add(const struct lu_env *env, struct obd_export *exp, int idx)
978 struct tg_export_data *ted = &exp->exp_target_data;
979 struct lu_target *tgt = class_exp2tgt(exp);
983 LASSERT(tgt && tgt->lut_client_bitmap != NULL);
984 LASSERTF(idx >= 0, "%d\n", idx);
986 if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
987 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
990 if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
991 CERROR("%s: client %d: bit already set in bitmap!!\n",
992 tgt->lut_obd->obd_name, idx);
995 atomic_inc(&tgt->lut_num_clients);
997 CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added, "
999 tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid,
1000 ted->ted_lcd->lcd_generation);
1002 ted->ted_lr_idx = idx;
1003 ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
1004 idx * tgt->lut_lsd.lsd_client_size;
1006 mutex_init(&ted->ted_lcd_lock);
1008 LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1013 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
1015 struct tg_export_data *ted = &exp->exp_target_data;
1016 struct lu_target *tgt = class_exp2tgt(exp);
1021 LASSERT(ted->ted_lcd);
1023 if (unlikely(tgt == NULL)) {
1024 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1025 class_exp2obd(exp)->obd_name);
1029 /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
1030 if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
1031 (char *)tgt->lut_obd->obd_uuid.uuid) ||
1032 exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1035 CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
1036 tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1037 ted->ted_lcd->lcd_uuid);
1039 /* Clear the bit _after_ zeroing out the client so we don't
1040 race with filter_client_add and zero out new clients.*/
1041 if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
1042 CERROR("%s: client %u: bit already clear in bitmap!!\n",
1043 tgt->lut_obd->obd_name, ted->ted_lr_idx);
1047 /* Do not erase record for recoverable client. */
1048 if (exp->exp_flags & OBD_OPT_FAILOVER)
1051 /* Make sure the server's last_transno is up to date.
1052 * This should be done before zeroing client slot so last_transno will
1053 * be in server data or in client data in case of failure */
1054 rc = tgt_server_data_update(env, tgt, 0);
1056 CERROR("%s: failed to update server data, skip client %s "
1057 "zeroing, rc %d\n", tgt->lut_obd->obd_name,
1058 ted->ted_lcd->lcd_uuid, rc);
1062 mutex_lock(&ted->ted_lcd_lock);
1063 memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
1064 rc = tgt_client_data_update(env, exp);
1065 mutex_unlock(&ted->ted_lcd_lock);
1067 CDEBUG(rc == 0 ? D_INFO : D_ERROR,
1068 "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
1069 tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
1070 ted->ted_lr_idx, ted->ted_lr_off, rc);
1073 EXPORT_SYMBOL(tgt_client_del);
1075 int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
1076 struct tg_export_data *ted, struct tg_reply_data *trd,
1077 struct thandle *th, bool update_lrd_file)
1079 struct lsd_reply_data *lrd;
1082 lrd = &trd->trd_reply;
1083 /* update export last transno */
1084 mutex_lock(&ted->ted_lcd_lock);
1085 if (lrd->lrd_transno > ted->ted_lcd->lcd_last_transno)
1086 ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
1087 mutex_unlock(&ted->ted_lcd_lock);
1089 /* find a empty slot */
1090 i = tgt_find_free_reply_slot(tgt);
1091 if (unlikely(i < 0)) {
1092 CERROR("%s: couldn't find a slot for reply data: "
1093 "rc = %d\n", tgt_name(tgt), i);
1098 if (update_lrd_file) {
1102 /* write reply data to disk */
1103 off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
1104 rc = tgt_reply_data_write(env, tgt, lrd, off, th);
1105 if (unlikely(rc != 0)) {
1106 CERROR("%s: can't update %s file: rc = %d\n",
1107 tgt_name(tgt), REPLY_DATA, rc);
1111 /* add reply data to target export's reply list */
1112 mutex_lock(&ted->ted_lcd_lock);
1113 list_add(&trd->trd_list, &ted->ted_reply_list);
1114 ted->ted_reply_cnt++;
1115 if (ted->ted_reply_cnt > ted->ted_reply_max)
1116 ted->ted_reply_max = ted->ted_reply_cnt;
1117 mutex_unlock(&ted->ted_lcd_lock);
1119 CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
1120 "tag %hu, client gen %u, slot idx %d\n",
1121 trd, lrd->lrd_xid, lrd->lrd_transno,
1122 trd->trd_tag, lrd->lrd_client_gen, i);
1125 EXPORT_SYMBOL(tgt_add_reply_data);
1128 * last_rcvd & last_committed update callbacks
1130 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
1131 struct dt_object *obj, __u64 opdata,
1132 struct thandle *th, struct ptlrpc_request *req)
1134 struct tgt_thread_info *tti = tgt_th_info(env);
1135 struct tgt_session_info *tsi = tgt_ses_info(env);
1136 struct obd_export *exp = tsi->tsi_exp;
1137 struct tg_export_data *ted;
1145 LASSERT(exp != NULL);
1146 ted = &exp->exp_target_data;
1148 lw_client = exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT;
1149 if (ted->ted_lr_idx < 0 && !lw_client)
1150 /* ofd connect may cause transaction before export has
1155 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
1157 /* From update replay, tti_transno should be set already */
1158 LASSERT(tti->tti_transno != 0);
1160 spin_lock(&tgt->lut_translock);
1161 if (th->th_result != 0) {
1162 if (tti->tti_transno != 0) {
1163 CERROR("%s: replay transno "LPU64" failed: rc = %d\n",
1164 tgt_name(tgt), tti->tti_transno, th->th_result);
1166 } else if (tti->tti_transno == 0) {
1167 tti->tti_transno = ++tgt->lut_last_transno;
1169 /* should be replay */
1170 if (tti->tti_transno > tgt->lut_last_transno)
1171 tgt->lut_last_transno = tti->tti_transno;
1173 spin_unlock(&tgt->lut_translock);
1175 /** VBR: set new versions */
1176 if (th->th_result == 0 && obj != NULL) {
1177 struct dt_object *dto = dt_object_locate(obj, th->th_dev);
1178 dt_version_set(env, dto, tti->tti_transno, th);
1181 /* filling reply data */
1182 CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
1183 tti->tti_transno, tgt->lut_obd->obd_last_committed);
1186 req->rq_transno = tti->tti_transno;
1187 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
1190 /* if can't add callback, do sync write */
1191 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp, tti->tti_transno);
1194 /* All operations performed by LW clients are synchronous and
1195 * we store the committed transno in the last_rcvd header */
1196 spin_lock(&tgt->lut_translock);
1197 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
1198 tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
1199 spin_unlock(&tgt->lut_translock);
1200 /* Although lightweight (LW) connections have no slot
1201 * in the last_rcvd, we still want to maintain
1202 * the in-memory lsd_client_data structure in order to
1203 * properly handle reply reconstruction. */
1204 rc = tgt_server_data_write(env, tgt, th);
1206 spin_unlock(&tgt->lut_translock);
1208 } else if (ted->ted_lr_off == 0) {
1209 CERROR("%s: client idx %d has offset %lld\n",
1210 tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
1214 /* Target that supports multiple reply data */
1215 if (tgt_is_multimodrpcs_client(exp)) {
1216 struct tg_reply_data *trd;
1217 struct lsd_reply_data *lrd;
1218 __u64 *pre_versions;
1222 if (unlikely(trd == NULL))
1225 /* fill reply data information */
1226 lrd = &trd->trd_reply;
1227 lrd->lrd_transno = tti->tti_transno;
1229 lrd->lrd_xid = req->rq_xid;
1230 trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
1231 pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1232 lrd->lrd_result = th->th_result;
1233 lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
1234 write_update = true;
1236 LASSERT(tsi->tsi_xid != 0);
1237 lrd->lrd_xid = tsi->tsi_xid;
1238 lrd->lrd_result = tsi->tsi_result;
1239 lrd->lrd_client_gen = tsi->tsi_client_gen;
1241 pre_versions = NULL;
1242 write_update = false;
1245 lrd->lrd_data = opdata;
1247 trd->trd_pre_versions[0] = pre_versions[0];
1248 trd->trd_pre_versions[1] = pre_versions[1];
1249 trd->trd_pre_versions[2] = pre_versions[2];
1250 trd->trd_pre_versions[3] = pre_versions[3];
1253 rc = tgt_add_reply_data(env, tgt, ted, trd, th, write_update);
1259 /* Enough for update replay, let's return */
1263 mutex_lock(&ted->ted_lcd_lock);
1264 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1265 if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
1266 transno_p = &ted->ted_lcd->lcd_last_close_transno;
1267 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
1268 ted->ted_lcd->lcd_last_close_result = th->th_result;
1270 /* VBR: save versions in last_rcvd for reconstruct. */
1271 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1274 ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
1275 ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
1276 ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
1277 ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
1279 transno_p = &ted->ted_lcd->lcd_last_transno;
1280 ted->ted_lcd->lcd_last_xid = req->rq_xid;
1281 ted->ted_lcd->lcd_last_result = th->th_result;
1282 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
1283 * see struct ldlm_reply->lock_policy_res1; */
1284 ted->ted_lcd->lcd_last_data = opdata;
1287 /* Update transno in slot only if non-zero number, i.e. no errors */
1288 if (likely(tti->tti_transno != 0)) {
1289 /* Don't overwrite bigger transaction number with lower one.
1290 * That is not sign of problem in all cases, but in any case
1291 * this value should be monotonically increased only. */
1292 if (*transno_p > tti->tti_transno) {
1293 if (!tgt->lut_no_reconstruct) {
1294 CERROR("%s: trying to overwrite bigger transno:"
1295 "on-disk: "LPU64", new: "LPU64" replay: "
1296 "%d. See LU-617.\n", tgt_name(tgt),
1297 *transno_p, tti->tti_transno,
1298 req_is_replay(req));
1299 if (req_is_replay(req)) {
1300 spin_lock(&req->rq_export->exp_lock);
1301 req->rq_export->exp_vbr_failed = 1;
1302 spin_unlock(&req->rq_export->exp_lock);
1304 mutex_unlock(&ted->ted_lcd_lock);
1305 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
1308 *transno_p = tti->tti_transno;
1313 tti->tti_off = ted->ted_lr_off;
1314 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1316 mutex_unlock(&ted->ted_lcd_lock);
1320 mutex_unlock(&ted->ted_lcd_lock);
1325 * last_rcvd update for echo client simulation.
1326 * It updates last_rcvd client slot and version of object in
1327 * simple way but with all locks to simulate all drawbacks
1329 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
1330 struct lu_target *tgt,
1331 struct dt_object *obj,
1333 struct obd_export *exp)
1335 struct tgt_thread_info *tti = tgt_th_info(env);
1336 struct tg_export_data *ted = &exp->exp_target_data;
1341 tti->tti_transno = 0;
1343 spin_lock(&tgt->lut_translock);
1344 if (th->th_result == 0)
1345 tti->tti_transno = ++tgt->lut_last_transno;
1346 spin_unlock(&tgt->lut_translock);
1348 /** VBR: set new versions */
1349 if (th->th_result == 0 && obj != NULL)
1350 dt_version_set(env, obj, tti->tti_transno, th);
1352 /* if can't add callback, do sync write */
1353 th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
1356 LASSERT(ted->ted_lr_off > 0);
1358 mutex_lock(&ted->ted_lcd_lock);
1359 LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1360 ted->ted_lcd->lcd_last_transno = tti->tti_transno;
1361 ted->ted_lcd->lcd_last_result = th->th_result;
1363 tti->tti_off = ted->ted_lr_off;
1364 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1365 mutex_unlock(&ted->ted_lcd_lock);
1369 static int tgt_clients_data_init(const struct lu_env *env,
1370 struct lu_target *tgt,
1371 unsigned long last_size)
1373 struct obd_device *obd = tgt->lut_obd;
1374 struct lr_server_data *lsd = &tgt->lut_lsd;
1375 struct lsd_client_data *lcd = NULL;
1376 struct tg_export_data *ted;
1379 loff_t off = lsd->lsd_client_start;
1380 __u32 generation = 0;
1381 struct cfs_hash *hash = NULL;
1385 CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
1386 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
1392 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1394 GOTO(err_out, rc = -ENODEV);
1396 for (cl_idx = 0; off < last_size; cl_idx++) {
1397 struct obd_export *exp;
1400 /* Don't assume off is incremented properly by
1401 * read_record(), in case sizeof(*lcd)
1402 * isn't the same as fsd->lsd_client_size. */
1403 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
1404 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
1406 CERROR("%s: error reading last_rcvd %s idx %d off "
1407 "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
1410 break; /* read error shouldn't cause startup to fail */
1413 if (lcd->lcd_uuid[0] == '\0') {
1414 CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
1419 last_transno = lcd_last_transno(lcd);
1421 /* These exports are cleaned up by disconnect, so they
1422 * need to be set up like real exports as connect does.
1424 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
1425 " srv lr: "LPU64" lx: "LPU64" gen %u\n", lcd->lcd_uuid,
1426 cl_idx, last_transno, lsd->lsd_last_transno,
1427 lcd_last_xid(lcd), lcd->lcd_generation);
1429 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
1431 if (PTR_ERR(exp) == -EALREADY) {
1432 /* export already exists, zero out this one */
1433 CERROR("%s: Duplicate export %s!\n",
1434 tgt_name(tgt), lcd->lcd_uuid);
1437 GOTO(err_out, rc = PTR_ERR(exp));
1440 ted = &exp->exp_target_data;
1441 *ted->ted_lcd = *lcd;
1443 rc = tgt_client_add(env, exp, cl_idx);
1444 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
1445 /* VBR: set export last committed version */
1446 exp->exp_last_committed = last_transno;
1447 spin_lock(&exp->exp_lock);
1448 exp->exp_connecting = 0;
1449 exp->exp_in_recovery = 0;
1450 spin_unlock(&exp->exp_lock);
1451 obd->obd_max_recoverable_clients++;
1453 if (tgt->lut_lsd.lsd_feature_incompat &
1454 OBD_INCOMPAT_MULTI_RPCS &&
1455 lcd->lcd_generation != 0) {
1456 /* compute the highest valid client generation */
1457 generation = max(generation, lcd->lcd_generation);
1458 /* fill client_generation <-> export hash table */
1459 rc = cfs_hash_add_unique(hash, &lcd->lcd_generation,
1460 &exp->exp_gen_hash);
1462 CERROR("%s: duplicate export for client "
1464 tgt_name(tgt), lcd->lcd_generation);
1465 class_export_put(exp);
1470 class_export_put(exp);
1472 rc = rev_import_init(exp);
1474 class_unlink_export(exp);
1478 /* Need to check last_rcvd even for duplicated exports. */
1479 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
1480 cl_idx, last_transno);
1482 spin_lock(&tgt->lut_translock);
1483 tgt->lut_last_transno = max(last_transno,
1484 tgt->lut_last_transno);
1485 spin_unlock(&tgt->lut_translock);
1488 /* record highest valid client generation */
1489 atomic_set(&tgt->lut_client_generation, generation);
1493 cfs_hash_putref(hash);
1498 struct server_compat_data {
1505 static struct server_compat_data tgt_scd[] = {
1506 [LDD_F_SV_TYPE_MDT] = {
1507 .rocompat = OBD_ROCOMPAT_LOVOBJID,
1508 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1509 OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
1510 OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI |
1511 OBD_INCOMPAT_MULTI_RPCS,
1512 .rocinit = OBD_ROCOMPAT_LOVOBJID,
1513 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1514 OBD_INCOMPAT_MULTI_OI,
1516 [LDD_F_SV_TYPE_OST] = {
1517 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
1518 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
1520 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
1521 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1525 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1527 struct tgt_thread_info *tti = tgt_th_info(env);
1528 struct lr_server_data *lsd = &tgt->lut_lsd;
1529 unsigned long last_rcvd_size;
1533 rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1537 last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1539 /* ensure padding in the struct is the correct size */
1540 CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
1541 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
1543 rc = server_name2index(tgt_name(tgt), &index, NULL);
1545 CERROR("%s: Can not get index from name: rc = %d\n",
1549 /* server_name2index() returns type */
1551 if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1552 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1556 /* last_rcvd on OST doesn't provide reconstruct support because there
1557 * may be up to 8 in-flight write requests per single slot in
1558 * last_rcvd client data
1560 tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1562 if (last_rcvd_size == 0) {
1563 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1565 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1566 sizeof(lsd->lsd_uuid));
1567 lsd->lsd_last_transno = 0;
1568 lsd->lsd_mount_count = 0;
1569 lsd->lsd_server_size = LR_SERVER_SIZE;
1570 lsd->lsd_client_start = LR_CLIENT_START;
1571 lsd->lsd_client_size = LR_CLIENT_SIZE;
1572 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1573 lsd->lsd_osd_index = index;
1574 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1575 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1577 rc = tgt_server_data_read(env, tgt);
1579 CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1583 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1584 LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
1585 "using the wrong disk %s. Were the"
1586 " /dev/ assignments rearranged?\n",
1587 tgt->lut_obd->obd_uuid.uuid,
1592 if (lsd->lsd_osd_index != index) {
1593 LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
1594 "is different with the index %d in"
1595 "config log, It might be disk"
1596 "corruption!\n", tgt_name(tgt),
1597 lsd->lsd_osd_index, index);
1602 if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1603 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1605 lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1609 if (type == LDD_F_SV_TYPE_MDT)
1610 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1612 if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1613 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1615 lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1618 /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1619 if (type == LDD_F_SV_TYPE_MDT &&
1620 !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1621 if (last_rcvd_size > lsd->lsd_client_start) {
1622 LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1623 "remove all clients for interop needs\n",
1625 rc = tgt_truncate_last_rcvd(env, tgt,
1626 lsd->lsd_client_start);
1629 last_rcvd_size = lsd->lsd_client_start;
1631 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1632 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1635 spin_lock(&tgt->lut_translock);
1636 tgt->lut_last_transno = lsd->lsd_last_transno;
1637 spin_unlock(&tgt->lut_translock);
1639 lsd->lsd_mount_count++;
1641 CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1642 CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
1643 tgt_name(tgt), tgt->lut_last_transno);
1644 CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
1645 tgt_name(tgt), lsd->lsd_mount_count);
1646 CDEBUG(D_INODE, "%s: server data size: %u\n",
1647 tgt_name(tgt), lsd->lsd_server_size);
1648 CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1649 tgt_name(tgt), lsd->lsd_client_start);
1650 CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1651 tgt_name(tgt), lsd->lsd_client_size);
1652 CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1653 tgt_name(tgt), last_rcvd_size);
1654 CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1655 tgt_name(tgt), lsd->lsd_subdir_count);
1656 CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1657 last_rcvd_size <= lsd->lsd_client_start ? 0 :
1658 (last_rcvd_size - lsd->lsd_client_start) /
1659 lsd->lsd_client_size);
1660 CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1662 if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1663 lsd->lsd_client_size == 0) {
1664 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1668 if (!tgt->lut_obd->obd_replayable)
1669 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1671 rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1673 GOTO(err_client, rc);
1675 spin_lock(&tgt->lut_translock);
1676 /* obd_last_committed is used for compatibility
1677 * with other lustre recovery code */
1678 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1679 spin_unlock(&tgt->lut_translock);
1681 tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1682 tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1684 /* save it, so mount count and last_transno is current */
1685 rc = tgt_server_data_update(env, tgt, 0);
1687 GOTO(err_client, rc);
1692 class_disconnect_exports(tgt->lut_obd);
1696 /* add credits for last_rcvd update */
1697 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1700 struct lu_target *tgt = cookie;
1701 struct tgt_session_info *tsi;
1702 struct tgt_thread_info *tti = tgt_th_info(env);
1703 struct dt_object *dto;
1706 /* if there is no session, then this transaction is not result of
1707 * request processing but some local operation */
1708 if (env->le_ses == NULL)
1711 LASSERT(tgt->lut_last_rcvd);
1712 tsi = tgt_ses_info(env);
1713 /* OFD may start transaction without export assigned */
1714 if (tsi->tsi_exp == NULL)
1717 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
1720 rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1721 tsi->tsi_exp->exp_target_data.ted_lr_off,
1727 rc = dt_declare_record_write(env, dto, &tti->tti_buf, 0, th);
1731 if (tgt_is_multimodrpcs_client(tsi->tsi_exp)) {
1733 * Use maximum possible file offset for declaration to ensure
1734 * ZFS will reserve enough credits for a write anywhere in this
1735 * file, since we don't know where in the file the write will be
1736 * because a replay slot has not been assigned. This should be
1737 * replaced by dmu_tx_hold_append() when available.
1739 tti->tti_off = atomic_read(&tgt->lut_num_clients) * 8 *
1740 sizeof(struct lsd_reply_data);
1741 tti->tti_buf.lb_buf = NULL;
1742 tti->tti_buf.lb_len = sizeof(struct lsd_reply_data);
1743 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
1744 rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1750 if (tsi->tsi_vbr_obj != NULL &&
1751 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1752 dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
1753 rc = dt_declare_version_set(env, dto, th);
1759 /* Update last_rcvd records with latests transaction data */
1760 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1763 struct lu_target *tgt = cookie;
1764 struct tgt_session_info *tsi;
1765 struct tgt_thread_info *tti = tgt_th_info(env);
1766 struct dt_object *obj = NULL;
1770 if (env->le_ses == NULL)
1773 tsi = tgt_ses_info(env);
1774 /* OFD may start transaction without export assigned */
1775 if (tsi->tsi_exp == NULL)
1778 echo_client = (tgt_ses_req(tsi) == NULL && tsi->tsi_xid == 0);
1780 if (tti->tti_has_trans && !echo_client) {
1781 if (tti->tti_mult_trans == 0) {
1782 CDEBUG(D_HA, "More than one transaction "LPU64"\n",
1786 /* we need another transno to be assigned */
1787 tti->tti_transno = 0;
1788 } else if (th->th_result == 0) {
1789 tti->tti_has_trans = 1;
1792 if (tsi->tsi_vbr_obj != NULL &&
1793 !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1794 obj = tsi->tsi_vbr_obj;
1797 if (unlikely(echo_client)) /* echo client special case */
1798 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1801 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
1806 int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
1808 struct tgt_thread_info *tti = tgt_th_info(env);
1809 struct lsd_reply_data *lrd = &tti->tti_lrd;
1810 unsigned long reply_data_size;
1812 struct lsd_reply_header *lrh = NULL;
1813 struct lsd_client_data *lcd = NULL;
1814 struct tg_reply_data *trd = NULL;
1817 struct cfs_hash *hash = NULL;
1818 struct obd_export *exp;
1819 struct tg_export_data *ted;
1820 int reply_data_recovered = 0;
1822 rc = dt_attr_get(env, tgt->lut_reply_data, &tti->tti_attr);
1825 reply_data_size = (unsigned long)tti->tti_attr.la_size;
1829 GOTO(out, rc = -ENOMEM);
1831 if (reply_data_size == 0) {
1832 CDEBUG(D_INFO, "%s: new reply_data file, initializing\n",
1834 lrh->lrh_magic = LRH_MAGIC;
1835 lrh->lrh_header_size = sizeof(struct lsd_reply_header);
1836 lrh->lrh_reply_size = sizeof(struct lsd_reply_data);
1837 rc = tgt_reply_header_write(env, tgt, lrh);
1839 CERROR("%s: error writing %s: rc = %d\n",
1840 tgt_name(tgt), REPLY_DATA, rc);
1844 rc = tgt_reply_header_read(env, tgt, lrh);
1846 CERROR("%s: error reading %s: rc = %d\n",
1847 tgt_name(tgt), REPLY_DATA, rc);
1850 if (lrh->lrh_magic != LRH_MAGIC ||
1851 lrh->lrh_header_size != sizeof(struct lsd_reply_header) ||
1852 lrh->lrh_reply_size != sizeof(struct lsd_reply_data)) {
1853 CERROR("%s: invalid header in %s\n",
1854 tgt_name(tgt), REPLY_DATA);
1855 GOTO(out, rc = -EINVAL);
1858 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1860 GOTO(out, rc = -ENODEV);
1864 GOTO(out, rc = -ENOMEM);
1868 GOTO(out, rc = -ENOMEM);
1870 /* Load reply_data from disk */
1871 for (idx = 0, off = sizeof(struct lsd_reply_header);
1872 off < reply_data_size;
1873 idx++, off += sizeof(struct lsd_reply_data)) {
1874 rc = tgt_reply_data_read(env, tgt, lrd, off);
1876 CERROR("%s: error reading %s: rc = %d\n",
1877 tgt_name(tgt), REPLY_DATA, rc);
1881 exp = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
1883 /* old reply data from a disconnected client */
1886 ted = &exp->exp_target_data;
1887 mutex_lock(&ted->ted_lcd_lock);
1889 /* create in-memory reply_data and link it to
1890 * target export's reply list */
1891 tgt_set_reply_slot(tgt, idx);
1892 trd->trd_reply = *lrd;
1893 trd->trd_pre_versions[0] = 0;
1894 trd->trd_pre_versions[1] = 0;
1895 trd->trd_pre_versions[2] = 0;
1896 trd->trd_pre_versions[3] = 0;
1897 trd->trd_index = idx;
1899 list_add(&trd->trd_list, &ted->ted_reply_list);
1900 ted->ted_reply_cnt++;
1901 if (ted->ted_reply_cnt > ted->ted_reply_max)
1902 ted->ted_reply_max = ted->ted_reply_cnt;
1904 CDEBUG(D_HA, "%s: restore reply %p: xid %llu, "
1905 "transno %llu, client gen %u, slot idx %d\n",
1906 tgt_name(tgt), trd, lrd->lrd_xid,
1907 lrd->lrd_transno, lrd->lrd_client_gen,
1910 /* update export last committed transation */
1911 exp->exp_last_committed = max(exp->exp_last_committed,
1914 mutex_unlock(&ted->ted_lcd_lock);
1915 class_export_put(exp);
1917 /* update target last committed transaction */
1918 spin_lock(&tgt->lut_translock);
1919 tgt->lut_last_transno = max(tgt->lut_last_transno,
1921 spin_unlock(&tgt->lut_translock);
1923 reply_data_recovered++;
1927 GOTO(out, rc = -ENOMEM);
1929 CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
1930 tgt_name(tgt), reply_data_recovered);
1933 spin_lock(&tgt->lut_translock);
1934 /* obd_last_committed is used for compatibility
1935 * with other lustre recovery code */
1936 tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1937 spin_unlock(&tgt->lut_translock);
1943 cfs_hash_putref(hash);
1953 struct tg_reply_data *tgt_lookup_reply_by_xid(struct tg_export_data *ted,
1956 struct tg_reply_data *found = NULL;
1957 struct tg_reply_data *reply;
1959 mutex_lock(&ted->ted_lcd_lock);
1960 list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
1961 if (reply->trd_reply.lrd_xid == xid) {
1966 mutex_unlock(&ted->ted_lcd_lock);
1969 EXPORT_SYMBOL(tgt_lookup_reply_by_xid);
1971 /* Look for a reply data matching specified request @req
1972 * A copy is returned in @trd if the pointer is not NULL
1974 bool tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
1976 struct tg_export_data *ted = &req->rq_export->exp_target_data;
1977 struct tg_reply_data *reply;
1980 reply = tgt_lookup_reply_by_xid(ted, req->rq_xid);
1981 if (reply != NULL) {
1987 CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d\n",
1988 tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid,
1993 EXPORT_SYMBOL(tgt_lookup_reply);
1995 int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
1997 struct tg_export_data *ted = &exp->exp_target_data;
1998 struct lu_target *lut = class_exp2tgt(exp);
1999 struct tg_reply_data *trd, *tmp;
2001 mutex_lock(&ted->ted_lcd_lock);
2002 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
2003 if (trd->trd_reply.lrd_xid > rcvd_xid)
2005 ted->ted_release_xid++;
2006 tgt_release_reply_data(lut, ted, trd);
2008 mutex_unlock(&ted->ted_lcd_lock);
2013 int tgt_handle_tag(struct obd_export *exp, __u16 tag)
2015 struct tg_export_data *ted = &exp->exp_target_data;
2016 struct lu_target *lut = class_exp2tgt(exp);
2017 struct tg_reply_data *trd, *tmp;
2022 mutex_lock(&ted->ted_lcd_lock);
2023 list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
2024 if (trd->trd_tag != tag)
2026 ted->ted_release_tag++;
2027 tgt_release_reply_data(lut, ted, trd);
2030 mutex_unlock(&ted->ted_lcd_lock);