lustre/target/tgt_lastrcvd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   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.
   9  *
  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).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Lustre Unified Target
  33  * These are common function to work with last_received file
  34  *
  35  * Author: Mikhail Pershin <mike.pershin@intel.com>
  36  */
  37 #include <obd.h>
  38 #include <obd_class.h>
  39 #include <lustre_fid.h>
  40
  41 #include "tgt_internal.h"
  42
  43 /** version recovery epoch */
  44 #define LR_EPOCH_BITS   32
  45
  46 /* Allocate a bitmap for a chunk of reply data slots */
  47 static int tgt_bitmap_chunk_alloc(struct lu_target *lut, int chunk)
  48 {
  49         unsigned long *bm;
  50
  51         OBD_ALLOC_LARGE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
  52                         sizeof(long));
  53         if (bm == NULL)
  54                 return -ENOMEM;
  55
  56         spin_lock(&lut->lut_client_bitmap_lock);
  57
  58         if (lut->lut_reply_bitmap[chunk] != NULL) {
  59                 /* someone else already allocated the bitmap for this chunk */
  60                 spin_unlock(&lut->lut_client_bitmap_lock);
  61                 OBD_FREE_LARGE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
  62                          sizeof(long));
  63                 return 0;
  64         }
  65
  66         lut->lut_reply_bitmap[chunk] = bm;
  67
  68         spin_unlock(&lut->lut_client_bitmap_lock);
  69
  70         return 0;
  71 }
  72
  73 /* Look for an available reply data slot in the bitmap
  74  * of the target @lut
  75  * Allocate bitmap chunk when first used
  76  * XXX algo could be improved if this routine limits performance
  77  */
  78 static int tgt_find_free_reply_slot(struct lu_target *lut)
  79 {
  80         unsigned long *bmp;
  81         int chunk = 0;
  82         int rc;
  83         int b;
  84
  85         for (chunk = 0; chunk < LUT_REPLY_SLOTS_MAX_CHUNKS; chunk++) {
  86                 /* allocate the bitmap chunk if necessary */
  87                 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
  88                         rc = tgt_bitmap_chunk_alloc(lut, chunk);
  89                         if (rc != 0)
  90                                 return rc;
  91                 }
  92                 bmp = lut->lut_reply_bitmap[chunk];
  93
  94                 /* look for an available slot in this chunk */
  95                 do {
  96                         b = find_first_zero_bit(bmp, LUT_REPLY_SLOTS_PER_CHUNK);
  97                         if (b >= LUT_REPLY_SLOTS_PER_CHUNK)
  98                                 break;
  99
 100                         /* found one */
 101                         if (test_and_set_bit(b, bmp) == 0)
 102                                 return chunk * LUT_REPLY_SLOTS_PER_CHUNK + b;
 103                 } while (true);
 104         }
 105
 106         return -ENOSPC;
 107 }
 108
 109 /* Mark the reply data slot @idx 'used' in the corresponding bitmap chunk
 110  * of the target @lut
 111  * Allocate the bitmap chunk if necessary
 112  */
 113 static int tgt_set_reply_slot(struct lu_target *lut, int idx)
 114 {
 115         int chunk;
 116         int b;
 117         int rc;
 118
 119         chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
 120         b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
 121
 122         LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
 123         LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
 124
 125         /* allocate the bitmap chunk if necessary */
 126         if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
 127                 rc = tgt_bitmap_chunk_alloc(lut, chunk);
 128                 if (rc != 0)
 129                         return rc;
 130         }
 131
 132         /* mark the slot 'used' in this chunk */
 133         if (test_and_set_bit(b, lut->lut_reply_bitmap[chunk]) != 0) {
 134                 CERROR("%s: slot %d already set in bitmap\n",
 135                        tgt_name(lut), idx);
 136                 return -EALREADY;
 137         }
 138
 139         return 0;
 140 }
 141
 142
 143 /* Mark the reply data slot @idx 'unused' in the corresponding bitmap chunk
 144  * of the target @lut
 145  */
 146 static int tgt_clear_reply_slot(struct lu_target *lut, int idx)
 147 {
 148         int chunk;
 149         int b;
 150
 151         if (lut->lut_obd->obd_stopping)
 152                 /*
 153                  * in case of failover keep the bit set in order to
 154                  * avoid overwriting slots in reply_data which might
 155                  * be required by resent rpcs
 156                  */
 157                 return 0;
 158         chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
 159         b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
 160
 161         LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
 162         LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
 163
 164         if (lut->lut_reply_bitmap[chunk] == NULL) {
 165                 CERROR("%s: slot %d not allocated\n",
 166                        tgt_name(lut), idx);
 167                 return -ENOENT;
 168         }
 169
 170         if (test_and_clear_bit(b, lut->lut_reply_bitmap[chunk]) == 0) {
 171                 CERROR("%s: slot %d already clear in bitmap\n",
 172                        tgt_name(lut), idx);
 173                 return -EALREADY;
 174         }
 175
 176         return 0;
 177 }
 178
 179
 180 /* Read header of reply_data file of target @tgt into structure @lrh */
 181 static int tgt_reply_header_read(const struct lu_env *env,
 182                                  struct lu_target *tgt,
 183                                  struct lsd_reply_header *lrh)
 184 {
 185         int                      rc;
 186         struct lsd_reply_header  buf;
 187         struct tgt_thread_info  *tti = tgt_th_info(env);
 188
 189         tti->tti_off = 0;
 190         tti->tti_buf.lb_buf = &buf;
 191         tti->tti_buf.lb_len = sizeof(buf);
 192
 193         rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
 194                             &tti->tti_off);
 195         if (rc != 0)
 196                 return rc;
 197
 198         lrh->lrh_magic = le32_to_cpu(buf.lrh_magic);
 199         lrh->lrh_header_size = le32_to_cpu(buf.lrh_header_size);
 200         lrh->lrh_reply_size = le32_to_cpu(buf.lrh_reply_size);
 201
 202         CDEBUG(D_HA, "%s: read %s header. magic=0x%08x "
 203                "header_size=%d reply_size=%d\n",
 204                 tgt->lut_obd->obd_name, REPLY_DATA,
 205                 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
 206
 207         return 0;
 208 }
 209
 210 /* Write header into replay_data file of target @tgt from structure @lrh */
 211 static int tgt_reply_header_write(const struct lu_env *env,
 212                                   struct lu_target *tgt,
 213                                   struct lsd_reply_header *lrh)
 214 {
 215         int                      rc;
 216         struct lsd_reply_header  buf;
 217         struct tgt_thread_info  *tti = tgt_th_info(env);
 218         struct thandle          *th;
 219         struct dt_object        *dto;
 220
 221         CDEBUG(D_HA, "%s: write %s header. magic=0x%08x "
 222                "header_size=%d reply_size=%d\n",
 223                 tgt->lut_obd->obd_name, REPLY_DATA,
 224                 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
 225
 226         if (tgt->lut_bottom->dd_rdonly)
 227                 RETURN(0);
 228
 229         buf.lrh_magic = cpu_to_le32(lrh->lrh_magic);
 230         buf.lrh_header_size = cpu_to_le32(lrh->lrh_header_size);
 231         buf.lrh_reply_size = cpu_to_le32(lrh->lrh_reply_size);
 232
 233         th = dt_trans_create(env, tgt->lut_bottom);
 234         if (IS_ERR(th))
 235                 return PTR_ERR(th);
 236         th->th_sync = 1;
 237
 238         tti->tti_off = 0;
 239         tti->tti_buf.lb_buf = &buf;
 240         tti->tti_buf.lb_len = sizeof(buf);
 241
 242         rc = dt_declare_record_write(env, tgt->lut_reply_data,
 243                                      &tti->tti_buf, tti->tti_off, th);
 244         if (rc)
 245                 GOTO(out, rc);
 246
 247         rc = dt_trans_start(env, tgt->lut_bottom, th);
 248         if (rc)
 249                 GOTO(out, rc);
 250
 251         dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
 252         rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 253 out:
 254         dt_trans_stop(env, tgt->lut_bottom, th);
 255         return rc;
 256 }
 257
 258 /* Write the reply data @lrd into reply_data file of target @tgt
 259  * at offset @off
 260  */
 261 static int tgt_reply_data_write(const struct lu_env *env, struct lu_target *tgt,
 262                                 struct lsd_reply_data *lrd, loff_t off,
 263                                 struct thandle *th)
 264 {
 265         struct tgt_thread_info  *tti = tgt_th_info(env);
 266         struct dt_object        *dto;
 267         struct lsd_reply_data   *buf = &tti->tti_lrd;
 268
 269         lrd->lrd_result = ptlrpc_status_hton(lrd->lrd_result);
 270
 271         buf->lrd_transno         = cpu_to_le64(lrd->lrd_transno);
 272         buf->lrd_xid             = cpu_to_le64(lrd->lrd_xid);
 273         buf->lrd_data            = cpu_to_le64(lrd->lrd_data);
 274         buf->lrd_result          = cpu_to_le32(lrd->lrd_result);
 275         buf->lrd_client_gen      = cpu_to_le32(lrd->lrd_client_gen);
 276
 277         lrd->lrd_result = ptlrpc_status_ntoh(lrd->lrd_result);
 278
 279         tti->tti_off = off;
 280         tti->tti_buf.lb_buf = buf;
 281         tti->tti_buf.lb_len = sizeof(*buf);
 282
 283         dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
 284         return dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 285 }
 286
 287 /* Read the reply data from reply_data file of target @tgt at offset @off
 288  * into structure @lrd
 289  */
 290 static int tgt_reply_data_read(const struct lu_env *env, struct lu_target *tgt,
 291                                struct lsd_reply_data *lrd, loff_t off)
 292 {
 293         int                      rc;
 294         struct tgt_thread_info  *tti = tgt_th_info(env);
 295         struct lsd_reply_data   *buf = &tti->tti_lrd;
 296
 297         tti->tti_off = off;
 298         tti->tti_buf.lb_buf = buf;
 299         tti->tti_buf.lb_len = sizeof(*buf);
 300
 301         rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
 302                             &tti->tti_off);
 303         if (rc != 0)
 304                 return rc;
 305
 306         lrd->lrd_transno         = le64_to_cpu(buf->lrd_transno);
 307         lrd->lrd_xid             = le64_to_cpu(buf->lrd_xid);
 308         lrd->lrd_data            = le64_to_cpu(buf->lrd_data);
 309         lrd->lrd_result          = le32_to_cpu(buf->lrd_result);
 310         lrd->lrd_client_gen      = le32_to_cpu(buf->lrd_client_gen);
 311
 312         return 0;
 313 }
 314
 315
 316 /* Free the in-memory reply data structure @trd and release
 317  * the corresponding slot in the reply_data file of target @lut
 318  * Called with ted_lcd_lock held
 319  */
 320 static void tgt_free_reply_data(struct lu_target *lut,
 321                                 struct tg_export_data *ted,
 322                                 struct tg_reply_data *trd)
 323 {
 324         CDEBUG(D_TRACE, "%s: free reply data %p: xid %llu, transno %llu, "
 325                "client gen %u, slot idx %d\n",
 326                lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
 327                trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
 328                trd->trd_index);
 329
 330         LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
 331
 332         list_del(&trd->trd_list);
 333         ted->ted_reply_cnt--;
 334         if (lut != NULL && trd->trd_index != TRD_INDEX_MEMORY)
 335                 tgt_clear_reply_slot(lut, trd->trd_index);
 336         OBD_FREE_PTR(trd);
 337 }
 338
 339 /* Release the reply data @trd from target @lut
 340  * The reply data with the highest transno for this export
 341  * is retained to ensure correctness of target recovery
 342  * Called with ted_lcd_lock held
 343  */
 344 static void tgt_release_reply_data(struct lu_target *lut,
 345                                    struct tg_export_data *ted,
 346                                    struct tg_reply_data *trd)
 347 {
 348         CDEBUG(D_TRACE, "%s: release reply data %p: xid %llu, transno %llu, "
 349                "client gen %u, slot idx %d\n",
 350                lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
 351                trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
 352                trd->trd_index);
 353
 354         LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
 355
 356         /* Do not free the reply data corresponding to the
 357          * highest transno of this export.
 358          * This ensures on-disk reply data is kept and
 359          * last committed transno can be restored from disk in case
 360          * of target recovery
 361          */
 362         if (trd->trd_reply.lrd_transno == ted->ted_lcd->lcd_last_transno) {
 363                 /* free previous retained reply */
 364                 if (ted->ted_reply_last != NULL)
 365                         tgt_free_reply_data(lut, ted, ted->ted_reply_last);
 366                 /* retain the reply */
 367                 list_del_init(&trd->trd_list);
 368                 ted->ted_reply_last = trd;
 369         } else {
 370                 tgt_free_reply_data(lut, ted, trd);
 371         }
 372 }
 373
 374 static inline struct lu_buf *tti_buf_lsd(struct tgt_thread_info *tti)
 375 {
 376         tti->tti_buf.lb_buf = &tti->tti_lsd;
 377         tti->tti_buf.lb_len = sizeof(tti->tti_lsd);
 378         return &tti->tti_buf;
 379 }
 380
 381 static inline struct lu_buf *tti_buf_lcd(struct tgt_thread_info *tti)
 382 {
 383         tti->tti_buf.lb_buf = &tti->tti_lcd;
 384         tti->tti_buf.lb_len = sizeof(tti->tti_lcd);
 385         return &tti->tti_buf;
 386 }
 387
 388 /**
 389  * Allocate in-memory data for client slot related to export.
 390  */
 391 int tgt_client_alloc(struct obd_export *exp)
 392 {
 393         ENTRY;
 394         LASSERT(exp != exp->exp_obd->obd_self_export);
 395
 396         spin_lock_init(&exp->exp_target_data.ted_nodemap_lock);
 397         INIT_LIST_HEAD(&exp->exp_target_data.ted_nodemap_member);
 398         spin_lock_init(&exp->exp_target_data.ted_fmd_lock);
 399         INIT_LIST_HEAD(&exp->exp_target_data.ted_fmd_list);
 400
 401         OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
 402         if (exp->exp_target_data.ted_lcd == NULL)
 403                 RETURN(-ENOMEM);
 404         /* Mark that slot is not yet valid, 0 doesn't work here */
 405         exp->exp_target_data.ted_lr_idx = -1;
 406         INIT_LIST_HEAD(&exp->exp_target_data.ted_reply_list);
 407         mutex_init(&exp->exp_target_data.ted_lcd_lock);
 408         RETURN(0);
 409 }
 410 EXPORT_SYMBOL(tgt_client_alloc);
 411
 412 /**
 413  * Free in-memory data for client slot related to export.
 414  */
 415 void tgt_client_free(struct obd_export *exp)
 416 {
 417         struct tg_export_data   *ted = &exp->exp_target_data;
 418         struct lu_target        *lut = class_exp2tgt(exp);
 419         struct tg_reply_data    *trd, *tmp;
 420
 421         LASSERT(exp != exp->exp_obd->obd_self_export);
 422
 423         tgt_fmd_cleanup(exp);
 424
 425         /* free reply data */
 426         mutex_lock(&ted->ted_lcd_lock);
 427         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
 428                 tgt_release_reply_data(lut, ted, trd);
 429         }
 430         if (ted->ted_reply_last != NULL) {
 431                 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
 432                 ted->ted_reply_last = NULL;
 433         }
 434         mutex_unlock(&ted->ted_lcd_lock);
 435
 436         if (!hlist_unhashed(&exp->exp_gen_hash))
 437                 cfs_hash_del(exp->exp_obd->obd_gen_hash,
 438                              &ted->ted_lcd->lcd_generation,
 439                              &exp->exp_gen_hash);
 440
 441         OBD_FREE_PTR(ted->ted_lcd);
 442         ted->ted_lcd = NULL;
 443
 444         /* Target may have been freed (see LU-7430)
 445          * Slot may be not yet assigned */
 446         if (exp->exp_obd->u.obt.obt_magic != OBT_MAGIC ||
 447             ted->ted_lr_idx < 0)
 448                 return;
 449
 450         /* Clear bit when lcd is freed */
 451         LASSERT(lut && lut->lut_client_bitmap);
 452         if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
 453                 CERROR("%s: client %u bit already clear in bitmap\n",
 454                        exp->exp_obd->obd_name, ted->ted_lr_idx);
 455                 LBUG();
 456         }
 457
 458         if (tgt_is_multimodrpcs_client(exp) && !exp->exp_obd->obd_stopping)
 459                 atomic_dec(&lut->lut_num_clients);
 460 }
 461 EXPORT_SYMBOL(tgt_client_free);
 462
 463 static inline void tgt_check_lcd(const char *obd_name, int index,
 464                                  struct lsd_client_data *lcd)
 465 {
 466         size_t uuid_size = sizeof(lcd->lcd_uuid);
 467
 468         if (strnlen((char*)lcd->lcd_uuid, uuid_size) == uuid_size) {
 469                 lcd->lcd_uuid[uuid_size - 1] = '\0';
 470
 471                 LCONSOLE_ERROR("the client UUID (%s) on %s for exports stored in last_rcvd(index = %d) is bad!\n",
 472                                lcd->lcd_uuid, obd_name, index);
 473         }
 474 }
 475
 476 static int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
 477                                 struct lsd_client_data *lcd,
 478                                 loff_t *off, int index)
 479 {
 480         struct tgt_thread_info  *tti = tgt_th_info(env);
 481         int                      rc;
 482
 483         tti_buf_lcd(tti);
 484         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
 485         if (rc == 0) {
 486                 tgt_check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
 487                 lcd_le_to_cpu(&tti->tti_lcd, lcd);
 488                 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
 489                 lcd->lcd_last_close_result =
 490                         ptlrpc_status_ntoh(lcd->lcd_last_close_result);
 491         }
 492
 493         CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = %llu"
 494                ", last_xid = %llu, last_result = %u, last_data = %u, "
 495                "last_close_transno = %llu, last_close_xid = %llu, "
 496                "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
 497                *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
 498                lcd->lcd_last_result, lcd->lcd_last_data,
 499                lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
 500                lcd->lcd_last_close_result, rc);
 501         return rc;
 502 }
 503
 504 static int tgt_client_data_write(const struct lu_env *env,
 505                                  struct lu_target *tgt,
 506                                  struct lsd_client_data *lcd,
 507                                  loff_t *off, struct thandle *th)
 508 {
 509         struct tgt_thread_info *tti = tgt_th_info(env);
 510         struct dt_object        *dto;
 511
 512         lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
 513         lcd->lcd_last_close_result =
 514                 ptlrpc_status_hton(lcd->lcd_last_close_result);
 515         lcd_cpu_to_le(lcd, &tti->tti_lcd);
 516         tti_buf_lcd(tti);
 517
 518         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
 519         return dt_record_write(env, dto, &tti->tti_buf, off, th);
 520 }
 521
 522 struct tgt_new_client_callback {
 523         struct dt_txn_commit_cb  lncc_cb;
 524         struct obd_export       *lncc_exp;
 525 };
 526
 527 static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
 528                               struct dt_txn_commit_cb *cb, int err)
 529 {
 530         struct tgt_new_client_callback *ccb;
 531
 532         ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
 533
 534         LASSERT(ccb->lncc_exp->exp_obd);
 535
 536         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 537                ccb->lncc_exp->exp_obd->obd_name,
 538                ccb->lncc_exp->exp_client_uuid.uuid);
 539
 540         spin_lock(&ccb->lncc_exp->exp_lock);
 541
 542         ccb->lncc_exp->exp_need_sync = 0;
 543
 544         spin_unlock(&ccb->lncc_exp->exp_lock);
 545         class_export_cb_put(ccb->lncc_exp);
 546
 547         OBD_FREE_PTR(ccb);
 548 }
 549
 550 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
 551 {
 552         struct tgt_new_client_callback  *ccb;
 553         struct dt_txn_commit_cb         *dcb;
 554         int                              rc;
 555
 556         OBD_ALLOC_PTR(ccb);
 557         if (ccb == NULL)
 558                 return -ENOMEM;
 559
 560         ccb->lncc_exp = class_export_cb_get(exp);
 561
 562         dcb = &ccb->lncc_cb;
 563         dcb->dcb_func = tgt_cb_new_client;
 564         INIT_LIST_HEAD(&dcb->dcb_linkage);
 565         strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
 566
 567         rc = dt_trans_cb_add(th, dcb);
 568         if (rc) {
 569                 class_export_cb_put(exp);
 570                 OBD_FREE_PTR(ccb);
 571         }
 572         return rc;
 573 }
 574
 575 /**
 576  * Update client data in last_rcvd
 577  */
 578 static int tgt_client_data_update(const struct lu_env *env,
 579                                   struct obd_export *exp)
 580 {
 581         struct tg_export_data   *ted = &exp->exp_target_data;
 582         struct lu_target        *tgt = class_exp2tgt(exp);
 583         struct tgt_thread_info  *tti = tgt_th_info(env);
 584         struct thandle          *th;
 585         int                      rc = 0;
 586
 587         ENTRY;
 588
 589         if (unlikely(tgt == NULL)) {
 590                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
 591                           class_exp2obd(exp)->obd_name);
 592                 RETURN(-EINVAL);
 593         }
 594
 595         if (tgt->lut_bottom->dd_rdonly)
 596                 RETURN(0);
 597
 598         th = dt_trans_create(env, tgt->lut_bottom);
 599         if (IS_ERR(th))
 600                 RETURN(PTR_ERR(th));
 601
 602         tti_buf_lcd(tti);
 603         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 604                                      &tti->tti_buf,
 605                                      ted->ted_lr_off, th);
 606         if (rc)
 607                 GOTO(out, rc);
 608
 609         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 610         if (rc)
 611                 GOTO(out, rc);
 612
 613         mutex_lock(&ted->ted_lcd_lock);
 614
 615         /*
 616          * Until this operations will be committed the sync is needed
 617          * for this export. This should be done _after_ starting the
 618          * transaction so that many connecting clients will not bring
 619          * server down with lots of sync writes.
 620          */
 621         rc = tgt_new_client_cb_add(th, exp);
 622         if (rc) {
 623                 /* can't add callback, do sync now */
 624                 th->th_sync = 1;
 625         } else {
 626                 spin_lock(&exp->exp_lock);
 627                 exp->exp_need_sync = 1;
 628                 spin_unlock(&exp->exp_lock);
 629         }
 630
 631         tti->tti_off = ted->ted_lr_off;
 632         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
 633
 634         mutex_unlock(&ted->ted_lcd_lock);
 635
 636         EXIT;
 637 out:
 638         dt_trans_stop(env, tgt->lut_bottom, th);
 639         CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
 640                "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
 641                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 642
 643         return rc;
 644 }
 645
 646 static int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
 647 {
 648         struct tgt_thread_info  *tti = tgt_th_info(env);
 649         int                      rc;
 650
 651         tti->tti_off = 0;
 652         tti_buf_lsd(tti);
 653         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
 654                             &tti->tti_off);
 655         if (rc == 0)
 656                 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
 657
 658         CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
 659                "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
 660                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 661         return rc;
 662 }
 663
 664 static int tgt_server_data_write(const struct lu_env *env,
 665                                  struct lu_target *tgt, struct thandle *th)
 666 {
 667         struct tgt_thread_info  *tti = tgt_th_info(env);
 668         struct dt_object        *dto;
 669         int                      rc;
 670
 671         ENTRY;
 672
 673         tti->tti_off = 0;
 674         tti_buf_lsd(tti);
 675         lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
 676
 677         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
 678         rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 679
 680         CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
 681                "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
 682                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 683
 684         RETURN(rc);
 685 }
 686
 687 /**
 688  * Update server data in last_rcvd
 689  */
 690 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
 691                            int sync)
 692 {
 693         struct tgt_thread_info  *tti = tgt_th_info(env);
 694         struct thandle          *th;
 695         int                      rc = 0;
 696
 697         ENTRY;
 698
 699         CDEBUG(D_SUPER,
 700                "%s: mount_count is %llu, last_transno is %llu\n",
 701                tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
 702                tgt->lut_last_transno);
 703
 704         /* Always save latest transno to keep it fresh */
 705         spin_lock(&tgt->lut_translock);
 706         tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
 707         spin_unlock(&tgt->lut_translock);
 708
 709         if (tgt->lut_bottom->dd_rdonly)
 710                 RETURN(0);
 711
 712         th = dt_trans_create(env, tgt->lut_bottom);
 713         if (IS_ERR(th))
 714                 RETURN(PTR_ERR(th));
 715
 716         th->th_sync = sync;
 717
 718         tti_buf_lsd(tti);
 719         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 720                                      &tti->tti_buf, tti->tti_off, th);
 721         if (rc)
 722                 GOTO(out, rc);
 723
 724         rc = dt_trans_start(env, tgt->lut_bottom, th);
 725         if (rc)
 726                 GOTO(out, rc);
 727
 728         rc = tgt_server_data_write(env, tgt, th);
 729 out:
 730         dt_trans_stop(env, tgt->lut_bottom, th);
 731
 732         CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
 733                "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
 734                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 735         RETURN(rc);
 736 }
 737 EXPORT_SYMBOL(tgt_server_data_update);
 738
 739 static int tgt_truncate_last_rcvd(const struct lu_env *env,
 740                                   struct lu_target *tgt, loff_t size)
 741 {
 742         struct dt_object *dt = tgt->lut_last_rcvd;
 743         struct thandle   *th;
 744         struct lu_attr    attr;
 745         int               rc;
 746
 747         ENTRY;
 748
 749         if (tgt->lut_bottom->dd_rdonly)
 750                 RETURN(0);
 751
 752         attr.la_size = size;
 753         attr.la_valid = LA_SIZE;
 754
 755         th = dt_trans_create(env, tgt->lut_bottom);
 756         if (IS_ERR(th))
 757                 RETURN(PTR_ERR(th));
 758         rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
 759         if (rc)
 760                 GOTO(cleanup, rc);
 761         rc = dt_declare_attr_set(env, dt, &attr, th);
 762         if (rc)
 763                 GOTO(cleanup, rc);
 764         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 765         if (rc)
 766                 GOTO(cleanup, rc);
 767
 768         rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
 769         if (rc == 0)
 770                 rc = dt_attr_set(env, dt, &attr, th);
 771
 772 cleanup:
 773         dt_trans_stop(env, tgt->lut_bottom, th);
 774
 775         RETURN(rc);
 776 }
 777
 778 static void tgt_client_epoch_update(const struct lu_env *env,
 779                                     struct obd_export *exp)
 780 {
 781         struct lsd_client_data  *lcd = exp->exp_target_data.ted_lcd;
 782         struct lu_target        *tgt = class_exp2tgt(exp);
 783
 784         LASSERT(tgt && tgt->lut_bottom);
 785         /** VBR: set client last_epoch to current epoch */
 786         if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
 787                 return;
 788         lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
 789         tgt_client_data_update(env, exp);
 790 }
 791
 792 /**
 793  * Update boot epoch when recovery ends
 794  */
 795 void tgt_boot_epoch_update(struct lu_target *tgt)
 796 {
 797         struct lu_env            env;
 798         struct ptlrpc_request   *req;
 799         __u32                    start_epoch;
 800         LIST_HEAD(client_list);
 801         int                      rc;
 802
 803         if (tgt->lut_obd->obd_stopping)
 804                 return;
 805
 806         rc = lu_env_init(&env, LCT_DT_THREAD);
 807         if (rc) {
 808                 CERROR("%s: can't initialize environment: rc = %d\n",
 809                         tgt->lut_obd->obd_name, rc);
 810                 return;
 811         }
 812
 813         spin_lock(&tgt->lut_translock);
 814         start_epoch = (tgt->lut_last_transno >> LR_EPOCH_BITS) + 1;
 815         tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
 816         tgt->lut_lsd.lsd_start_epoch = start_epoch;
 817         spin_unlock(&tgt->lut_translock);
 818
 819         /**
 820          * The recovery is not yet finished and final queue can still be updated
 821          * with resend requests. Move final list to separate one for processing
 822          */
 823         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
 824         list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
 825         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
 826
 827         /**
 828          * go through list of exports participated in recovery and
 829          * set new epoch for them
 830          */
 831         list_for_each_entry(req, &client_list, rq_list) {
 832                 LASSERT(!req->rq_export->exp_delayed);
 833                 if (!req->rq_export->exp_vbr_failed)
 834                         tgt_client_epoch_update(&env, req->rq_export);
 835         }
 836         /** return list back at once */
 837         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
 838         list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
 839         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
 840
 841         /** Clear MULTI RPCS incompatibility flag if
 842          * - target is MDT and
 843          * - there is no client to recover or the recovery was aborted
 844          */
 845         if (!strncmp(tgt->lut_obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) &&
 846             (atomic_read(&tgt->lut_obd->obd_max_recoverable_clients) == 0 ||
 847             tgt->lut_obd->obd_abort_recovery))
 848                 tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
 849
 850         /** update server epoch */
 851         tgt_server_data_update(&env, tgt, 1);
 852         lu_env_fini(&env);
 853 }
 854
 855 /**
 856  * commit callback, need to update last_committed value
 857  */
 858 struct tgt_last_committed_callback {
 859         struct dt_txn_commit_cb  llcc_cb;
 860         struct lu_target        *llcc_tgt;
 861         struct obd_export       *llcc_exp;
 862         __u64                    llcc_transno;
 863 };
 864
 865 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
 866                                   struct dt_txn_commit_cb *cb, int err)
 867 {
 868         struct tgt_last_committed_callback *ccb;
 869
 870         ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
 871
 872         LASSERT(ccb->llcc_exp);
 873         LASSERT(ccb->llcc_tgt != NULL);
 874         LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
 875
 876         /* error hit, don't update last committed to provide chance to
 877          * replay data after fail */
 878         if (err != 0)
 879                 goto out;
 880
 881         /* Fast path w/o spinlock, if exp_last_committed was updated
 882          * with higher transno, no need to take spinlock and check,
 883          * also no need to update obd_last_committed. */
 884         if (ccb->llcc_transno <= ccb->llcc_exp->exp_last_committed)
 885                 goto out;
 886         spin_lock(&ccb->llcc_tgt->lut_translock);
 887         if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
 888                 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
 889
 890         if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
 891                 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
 892                 spin_unlock(&ccb->llcc_tgt->lut_translock);
 893
 894                 ptlrpc_commit_replies(ccb->llcc_exp);
 895                 tgt_cancel_slc_locks(ccb->llcc_tgt, ccb->llcc_transno);
 896         } else {
 897                 spin_unlock(&ccb->llcc_tgt->lut_translock);
 898         }
 899
 900         CDEBUG(D_HA, "%s: transno %lld is committed\n",
 901                ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
 902
 903 out:
 904         class_export_cb_put(ccb->llcc_exp);
 905         OBD_FREE_PTR(ccb);
 906 }
 907
 908 /**
 909  * Add commit callback function, it returns a non-zero value to inform
 910  * caller to use sync transaction if necessary.
 911  */
 912 static int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
 913                                   struct obd_export *exp, __u64 transno)
 914 {
 915         struct tgt_last_committed_callback      *ccb;
 916         struct dt_txn_commit_cb                 *dcb;
 917         int                                      rc;
 918
 919         OBD_ALLOC_PTR(ccb);
 920         if (ccb == NULL)
 921                 return -ENOMEM;
 922
 923         ccb->llcc_tgt = tgt;
 924         ccb->llcc_exp = class_export_cb_get(exp);
 925         ccb->llcc_transno = transno;
 926
 927         dcb = &ccb->llcc_cb;
 928         dcb->dcb_func = tgt_cb_last_committed;
 929         INIT_LIST_HEAD(&dcb->dcb_linkage);
 930         strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
 931
 932         rc = dt_trans_cb_add(th, dcb);
 933         if (rc) {
 934                 class_export_cb_put(exp);
 935                 OBD_FREE_PTR(ccb);
 936         }
 937
 938         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 939                 /* report failure to force synchronous operation */
 940                 return -EPERM;
 941
 942         /* if exp_need_sync is set, return non-zero value to force
 943          * a sync transaction. */
 944         return rc ? rc : exp->exp_need_sync;
 945 }
 946
 947 /**
 948  * Add new client to the last_rcvd upon new connection.
 949  *
 950  * We use a bitmap to locate a free space in the last_rcvd file and initialize
 951  * tg_export_data.
 952  */
 953 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
 954 {
 955         struct tg_export_data   *ted = &exp->exp_target_data;
 956         struct lu_target        *tgt = class_exp2tgt(exp);
 957         int                      rc = 0, idx;
 958
 959         ENTRY;
 960
 961         LASSERT(tgt && tgt->lut_client_bitmap != NULL);
 962         if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
 963                 RETURN(0);
 964
 965         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 966                 RETURN(0);
 967
 968         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 969          * there's no need for extra complication here
 970          */
 971         idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
 972 repeat:
 973         if (idx >= LR_MAX_CLIENTS ||
 974             OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
 975                 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
 976                        tgt->lut_obd->obd_name,  idx);
 977                 RETURN(-EOVERFLOW);
 978         }
 979         if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
 980                 idx = find_next_zero_bit(tgt->lut_client_bitmap,
 981                                              LR_MAX_CLIENTS, idx);
 982                 goto repeat;
 983         }
 984
 985         ted->ted_lr_idx = idx;
 986         ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
 987                           idx * tgt->lut_lsd.lsd_client_size;
 988
 989         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
 990
 991         if (tgt_is_multimodrpcs_client(exp)) {
 992                 /* Set MULTI RPCS incompatibility flag to prevent previous
 993                  * Lustre versions to mount a target with reply_data file */
 994                 atomic_inc(&tgt->lut_num_clients);
 995                 if (!(tgt->lut_lsd.lsd_feature_incompat &
 996                       OBD_INCOMPAT_MULTI_RPCS)) {
 997                         tgt->lut_lsd.lsd_feature_incompat |=
 998                                                         OBD_INCOMPAT_MULTI_RPCS;
 999                         rc = tgt_server_data_update(env, tgt, 1);
1000                         if (rc < 0) {
1001                                 CERROR("%s: unable to set MULTI RPCS "
1002                                        "incompatibility flag\n",
1003                                        exp->exp_obd->obd_name);
1004                                 RETURN(rc);
1005                         }
1006                 }
1007
1008                 /* assign client slot generation */
1009                 ted->ted_lcd->lcd_generation =
1010                                 atomic_inc_return(&tgt->lut_client_generation);
1011         } else {
1012                 ted->ted_lcd->lcd_generation = 0;
1013         }
1014
1015         CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s' "
1016                "generation %d\n",
1017                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1018                ted->ted_lcd->lcd_uuid, ted->ted_lcd->lcd_generation);
1019
1020         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
1021                 RETURN(-ENOSPC);
1022
1023         rc = tgt_client_data_update(env, exp);
1024         if (rc)
1025                 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
1026                        tgt->lut_obd->obd_name, idx, rc);
1027
1028         RETURN(rc);
1029 }
1030 EXPORT_SYMBOL(tgt_client_new);
1031
1032 /* Add an existing client to the MDS in-memory state based on
1033  * a client that was previously found in the last_rcvd file and
1034  * already has an assigned slot (idx >= 0).
1035  *
1036  * It should not be possible to fail adding an existing client - otherwise
1037  * mdt_init_server_data() callsite needs to be fixed.
1038  */
1039 int tgt_client_add(const struct lu_env *env,  struct obd_export *exp, int idx)
1040 {
1041         struct tg_export_data   *ted = &exp->exp_target_data;
1042         struct lu_target        *tgt = class_exp2tgt(exp);
1043
1044         ENTRY;
1045
1046         LASSERT(tgt && tgt->lut_client_bitmap != NULL);
1047         LASSERTF(idx >= 0, "%d\n", idx);
1048
1049         if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
1050             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1051                 RETURN(0);
1052
1053         if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
1054                 CERROR("%s: client %d: bit already set in bitmap!!\n",
1055                        tgt->lut_obd->obd_name,  idx);
1056                 LBUG();
1057         }
1058         atomic_inc(&tgt->lut_num_clients);
1059
1060         CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added, "
1061                "generation %d\n",
1062                tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid,
1063                ted->ted_lcd->lcd_generation);
1064
1065         ted->ted_lr_idx = idx;
1066         ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
1067                           idx * tgt->lut_lsd.lsd_client_size;
1068
1069         mutex_init(&ted->ted_lcd_lock);
1070
1071         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1072
1073         RETURN(0);
1074 }
1075
1076 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
1077 {
1078         struct tg_export_data   *ted = &exp->exp_target_data;
1079         struct lu_target        *tgt = class_exp2tgt(exp);
1080         int                      rc;
1081
1082         ENTRY;
1083
1084         LASSERT(ted->ted_lcd);
1085
1086         if (unlikely(tgt == NULL)) {
1087                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1088                        class_exp2obd(exp)->obd_name);
1089                 RETURN(-EINVAL);
1090         }
1091
1092         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
1093         if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
1094                     (char *)tgt->lut_obd->obd_uuid.uuid) ||
1095             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1096                 RETURN(0);
1097
1098         /* Slot may be not yet assigned, use case is race between Client
1099          * reconnect and forced eviction */
1100         if (ted->ted_lr_idx < 0) {
1101                 CWARN("%s: client with UUID '%s' not in bitmap\n",
1102                       tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid);
1103                 RETURN(0);
1104         }
1105
1106         CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
1107                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1108                ted->ted_lcd->lcd_uuid);
1109
1110         /* Clear the bit _after_ zeroing out the client so we don't
1111            race with filter_client_add and zero out new clients.*/
1112         if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
1113                 CERROR("%s: client %u: bit already clear in bitmap!!\n",
1114                        tgt->lut_obd->obd_name, ted->ted_lr_idx);
1115                 LBUG();
1116         }
1117
1118         /* Do not erase record for recoverable client. */
1119         if (exp->exp_flags & OBD_OPT_FAILOVER)
1120                 RETURN(0);
1121
1122         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_DEL))
1123                 RETURN(0);
1124
1125         /* Make sure the server's last_transno is up to date.
1126          * This should be done before zeroing client slot so last_transno will
1127          * be in server data or in client data in case of failure */
1128         rc = tgt_server_data_update(env, tgt, 0);
1129         if (rc != 0) {
1130                 CERROR("%s: failed to update server data, skip client %s "
1131                        "zeroing, rc %d\n", tgt->lut_obd->obd_name,
1132                        ted->ted_lcd->lcd_uuid, rc);
1133                 RETURN(rc);
1134         }
1135
1136         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
1137         rc = tgt_client_data_update(env, exp);
1138
1139         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
1140                "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
1141                tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
1142                ted->ted_lr_idx, ted->ted_lr_off, rc);
1143         RETURN(rc);
1144 }
1145 EXPORT_SYMBOL(tgt_client_del);
1146
1147 static void tgt_clean_by_tag(struct obd_export *exp, __u64 xid, __u16 tag)
1148 {
1149         struct tg_export_data   *ted = &exp->exp_target_data;
1150         struct lu_target        *lut = class_exp2tgt(exp);
1151         struct tg_reply_data    *trd, *tmp;
1152
1153         if (tag == 0)
1154                 return;
1155
1156         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
1157                 if (trd->trd_tag != tag)
1158                         continue;
1159
1160                 LASSERT(ergo(tgt_is_increasing_xid_client(exp),
1161                              trd->trd_reply.lrd_xid <= xid));
1162
1163                 ted->ted_release_tag++;
1164                 tgt_release_reply_data(lut, ted, trd);
1165         }
1166 }
1167
1168 static int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
1169                        struct tg_export_data *ted, struct tg_reply_data *trd,
1170                        struct ptlrpc_request *req,
1171                        struct thandle *th, bool update_lrd_file)
1172 {
1173         struct lsd_reply_data   *lrd;
1174         int     i;
1175
1176         lrd = &trd->trd_reply;
1177         /* update export last transno */
1178         mutex_lock(&ted->ted_lcd_lock);
1179         if (lrd->lrd_transno > ted->ted_lcd->lcd_last_transno)
1180                 ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
1181         mutex_unlock(&ted->ted_lcd_lock);
1182
1183         if (tgt != NULL) {
1184                 /* find a empty slot */
1185                 i = tgt_find_free_reply_slot(tgt);
1186                 if (unlikely(i < 0)) {
1187                         CERROR("%s: couldn't find a slot for reply data: "
1188                                "rc = %d\n", tgt_name(tgt), i);
1189                         RETURN(i);
1190                 }
1191                 trd->trd_index = i;
1192
1193                 if (update_lrd_file) {
1194                         loff_t  off;
1195                         int     rc;
1196
1197                         /* write reply data to disk */
1198                         off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
1199                         rc = tgt_reply_data_write(env, tgt, lrd, off, th);
1200                         if (unlikely(rc != 0)) {
1201                                 CERROR("%s: can't update %s file: rc = %d\n",
1202                                        tgt_name(tgt), REPLY_DATA, rc);
1203                                 RETURN(rc);
1204                         }
1205                 }
1206         } else {
1207                 trd->trd_index = TRD_INDEX_MEMORY;
1208         }
1209
1210         /* add reply data to target export's reply list */
1211         mutex_lock(&ted->ted_lcd_lock);
1212         if (req != NULL) {
1213                 int exclude = tgt_is_increasing_xid_client(req->rq_export) ?
1214                               MSG_REPLAY : MSG_REPLAY|MSG_RESENT;
1215
1216                 if (req->rq_obsolete) {
1217                         mutex_unlock(&ted->ted_lcd_lock);
1218                         RETURN(-EALREADY);
1219                 }
1220
1221                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & exclude))
1222                         tgt_clean_by_tag(req->rq_export, req->rq_xid,
1223                                          trd->trd_tag);
1224         }
1225         list_add(&trd->trd_list, &ted->ted_reply_list);
1226         ted->ted_reply_cnt++;
1227         if (ted->ted_reply_cnt > ted->ted_reply_max)
1228                 ted->ted_reply_max = ted->ted_reply_cnt;
1229         mutex_unlock(&ted->ted_lcd_lock);
1230
1231         CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
1232                "tag %hu, client gen %u, slot idx %d\n",
1233                trd, lrd->lrd_xid, lrd->lrd_transno,
1234                trd->trd_tag, lrd->lrd_client_gen, trd->trd_index);
1235
1236         RETURN(0);
1237 }
1238
1239 int tgt_mk_reply_data(const struct lu_env *env,
1240                       struct lu_target *tgt,
1241                       struct tg_export_data *ted,
1242                       struct ptlrpc_request *req,
1243                       __u64 opdata,
1244                       struct thandle *th,
1245                       bool write_update,
1246                       __u64 transno)
1247 {
1248         struct tg_reply_data    *trd;
1249         struct lsd_reply_data   *lrd;
1250         __u64                   *pre_versions = NULL;
1251         int                     rc;
1252
1253         OBD_ALLOC_PTR(trd);
1254         if (unlikely(trd == NULL))
1255                 RETURN(-ENOMEM);
1256
1257         /* fill reply data information */
1258         lrd = &trd->trd_reply;
1259         lrd->lrd_transno = transno;
1260         if (req != NULL) {
1261                 lrd->lrd_xid = req->rq_xid;
1262                 trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
1263                 lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
1264                 if (write_update) {
1265                         pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1266                         lrd->lrd_result = th->th_result;
1267                 }
1268         } else {
1269                 struct tgt_session_info *tsi;
1270
1271                 LASSERT(env != NULL);
1272                 tsi = tgt_ses_info(env);
1273                 LASSERT(tsi->tsi_xid != 0);
1274
1275                 lrd->lrd_xid = tsi->tsi_xid;
1276                 lrd->lrd_result = tsi->tsi_result;
1277                 lrd->lrd_client_gen = tsi->tsi_client_gen;
1278         }
1279
1280         lrd->lrd_data = opdata;
1281         if (pre_versions) {
1282                 trd->trd_pre_versions[0] = pre_versions[0];
1283                 trd->trd_pre_versions[1] = pre_versions[1];
1284                 trd->trd_pre_versions[2] = pre_versions[2];
1285                 trd->trd_pre_versions[3] = pre_versions[3];
1286         }
1287
1288         rc = tgt_add_reply_data(env, tgt, ted, trd, req,
1289                                 th, write_update);
1290         if (rc < 0)
1291                 OBD_FREE_PTR(trd);
1292         return rc;
1293
1294 }
1295 EXPORT_SYMBOL(tgt_mk_reply_data);
1296
1297 /*
1298  * last_rcvd & last_committed update callbacks
1299  */
1300 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
1301                                 struct dt_object *obj, __u64 opdata,
1302                                 struct thandle *th, struct ptlrpc_request *req)
1303 {
1304         struct tgt_thread_info  *tti = tgt_th_info(env);
1305         struct tgt_session_info *tsi = tgt_ses_info(env);
1306         struct obd_export       *exp = tsi->tsi_exp;
1307         struct tg_export_data   *ted;
1308         __u64                   *transno_p;
1309         int                      rc = 0;
1310         bool                     lw_client;
1311
1312         ENTRY;
1313
1314
1315         LASSERT(exp != NULL);
1316         ted = &exp->exp_target_data;
1317
1318         lw_client = exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT;
1319         if (ted->ted_lr_idx < 0 && !lw_client)
1320                 /* ofd connect may cause transaction before export has
1321                  * last_rcvd slot */
1322                 RETURN(0);
1323
1324         if (req != NULL)
1325                 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
1326         else
1327                 /* From update replay, tti_transno should be set already */
1328                 LASSERT(tti->tti_transno != 0);
1329
1330         spin_lock(&tgt->lut_translock);
1331         if (th->th_result != 0) {
1332                 if (tti->tti_transno != 0) {
1333                         CERROR("%s: replay transno %llu failed: rc = %d\n",
1334                                tgt_name(tgt), tti->tti_transno, th->th_result);
1335                 }
1336         } else if (tti->tti_transno == 0) {
1337                 tti->tti_transno = ++tgt->lut_last_transno;
1338         } else {
1339                 /* should be replay */
1340                 if (tti->tti_transno > tgt->lut_last_transno)
1341                         tgt->lut_last_transno = tti->tti_transno;
1342         }
1343         spin_unlock(&tgt->lut_translock);
1344
1345         /** VBR: set new versions */
1346         if (th->th_result == 0 && obj != NULL) {
1347                 struct dt_object *dto = dt_object_locate(obj, th->th_dev);
1348                 dt_version_set(env, dto, tti->tti_transno, th);
1349         }
1350
1351         /* filling reply data */
1352         CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
1353                tti->tti_transno, tgt->lut_obd->obd_last_committed);
1354
1355         if (req != NULL) {
1356                 req->rq_transno = tti->tti_transno;
1357                 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
1358         }
1359
1360         /* if can't add callback, do sync write */
1361         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp, tti->tti_transno);
1362
1363         if (lw_client) {
1364                 /* All operations performed by LW clients are synchronous and
1365                  * we store the committed transno in the last_rcvd header */
1366                 spin_lock(&tgt->lut_translock);
1367                 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
1368                         tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
1369                         spin_unlock(&tgt->lut_translock);
1370                         /* Although lightweight (LW) connections have no slot
1371                          * in the last_rcvd, we still want to maintain
1372                          * the in-memory lsd_client_data structure in order to
1373                          * properly handle reply reconstruction. */
1374                         rc = tgt_server_data_write(env, tgt, th);
1375                 } else {
1376                         spin_unlock(&tgt->lut_translock);
1377                 }
1378         } else if (ted->ted_lr_off == 0) {
1379                 CERROR("%s: client idx %d has offset %lld\n",
1380                        tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
1381                 RETURN(-EINVAL);
1382         }
1383
1384         /* Target that supports multiple reply data */
1385         if (tgt_is_multimodrpcs_client(exp)) {
1386                 return tgt_mk_reply_data(env, tgt, ted, req, opdata, th,
1387                                          !!(req != NULL), tti->tti_transno);
1388         }
1389
1390         /* Enough for update replay, let's return */
1391         if (req == NULL)
1392                 RETURN(rc);
1393
1394         mutex_lock(&ted->ted_lcd_lock);
1395         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1396         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
1397                 transno_p = &ted->ted_lcd->lcd_last_close_transno;
1398                 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
1399                 ted->ted_lcd->lcd_last_close_result = th->th_result;
1400         } else {
1401                 /* VBR: save versions in last_rcvd for reconstruct. */
1402                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1403
1404                 if (pre_versions) {
1405                         ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
1406                         ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
1407                         ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
1408                         ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
1409                 }
1410                 transno_p = &ted->ted_lcd->lcd_last_transno;
1411                 ted->ted_lcd->lcd_last_xid = req->rq_xid;
1412                 ted->ted_lcd->lcd_last_result = th->th_result;
1413                 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
1414                  * see struct ldlm_reply->lock_policy_res1; */
1415                 ted->ted_lcd->lcd_last_data = opdata;
1416         }
1417
1418         /* Update transno in slot only if non-zero number, i.e. no errors */
1419         if (likely(tti->tti_transno != 0)) {
1420                 /* Don't overwrite bigger transaction number with lower one.
1421                  * That is not sign of problem in all cases, but in any case
1422                  * this value should be monotonically increased only. */
1423                 if (*transno_p > tti->tti_transno) {
1424                         if (!tgt->lut_no_reconstruct) {
1425                                 CERROR("%s: trying to overwrite bigger transno:"
1426                                        "on-disk: %llu, new: %llu replay: "
1427                                        "%d. See LU-617.\n", tgt_name(tgt),
1428                                        *transno_p, tti->tti_transno,
1429                                        req_is_replay(req));
1430                                 if (req_is_replay(req)) {
1431                                         spin_lock(&req->rq_export->exp_lock);
1432                                         req->rq_export->exp_vbr_failed = 1;
1433                                         spin_unlock(&req->rq_export->exp_lock);
1434                                 }
1435                                 mutex_unlock(&ted->ted_lcd_lock);
1436                                 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
1437                         }
1438                 } else {
1439                         *transno_p = tti->tti_transno;
1440                 }
1441         }
1442
1443         if (!lw_client) {
1444                 tti->tti_off = ted->ted_lr_off;
1445                 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_RCVD_EIO))
1446                         rc = -EIO;
1447                 else
1448                         rc = tgt_client_data_write(env, tgt, ted->ted_lcd,
1449                                                    &tti->tti_off, th);
1450                 if (rc < 0) {
1451                         mutex_unlock(&ted->ted_lcd_lock);
1452                         RETURN(rc);
1453                 }
1454         }
1455         mutex_unlock(&ted->ted_lcd_lock);
1456         RETURN(rc);
1457 }
1458
1459 /*
1460  * last_rcvd update for echo client simulation.
1461  * It updates last_rcvd client slot and version of object in
1462  * simple way but with all locks to simulate all drawbacks
1463  */
1464 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
1465                                      struct lu_target *tgt,
1466                                      struct dt_object *obj,
1467                                      struct thandle *th,
1468                                      struct obd_export *exp)
1469 {
1470         struct tgt_thread_info  *tti = tgt_th_info(env);
1471         struct tg_export_data   *ted = &exp->exp_target_data;
1472         int                      rc = 0;
1473
1474         ENTRY;
1475
1476         tti->tti_transno = 0;
1477
1478         spin_lock(&tgt->lut_translock);
1479         if (th->th_result == 0)
1480                 tti->tti_transno = ++tgt->lut_last_transno;
1481         spin_unlock(&tgt->lut_translock);
1482
1483         /** VBR: set new versions */
1484         if (th->th_result == 0 && obj != NULL)
1485                 dt_version_set(env, obj, tti->tti_transno, th);
1486
1487         /* if can't add callback, do sync write */
1488         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
1489                                                 tti->tti_transno);
1490
1491         LASSERT(ted->ted_lr_off > 0);
1492
1493         mutex_lock(&ted->ted_lcd_lock);
1494         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1495         ted->ted_lcd->lcd_last_transno = tti->tti_transno;
1496         ted->ted_lcd->lcd_last_result = th->th_result;
1497
1498         tti->tti_off = ted->ted_lr_off;
1499         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1500         mutex_unlock(&ted->ted_lcd_lock);
1501         RETURN(rc);
1502 }
1503
1504 static int tgt_clients_data_init(const struct lu_env *env,
1505                                  struct lu_target *tgt,
1506                                  unsigned long last_size)
1507 {
1508         struct obd_device       *obd = tgt->lut_obd;
1509         struct lr_server_data   *lsd = &tgt->lut_lsd;
1510         struct lsd_client_data  *lcd = NULL;
1511         struct tg_export_data   *ted;
1512         int                      cl_idx;
1513         int                      rc = 0;
1514         loff_t                   off = lsd->lsd_client_start;
1515         __u32                    generation = 0;
1516         struct cfs_hash         *hash = NULL;
1517
1518         ENTRY;
1519
1520         if (tgt->lut_bottom->dd_rdonly)
1521                 RETURN(0);
1522
1523         BUILD_BUG_ON(offsetof(struct lsd_client_data, lcd_padding) +
1524                      sizeof(lcd->lcd_padding) != LR_CLIENT_SIZE);
1525
1526         OBD_ALLOC_PTR(lcd);
1527         if (lcd == NULL)
1528                 RETURN(-ENOMEM);
1529
1530         hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1531         if (hash == NULL)
1532                 GOTO(err_out, rc = -ENODEV);
1533
1534         for (cl_idx = 0; off < last_size; cl_idx++) {
1535                 struct obd_export       *exp;
1536                 __u64                    last_transno;
1537
1538                 /* Don't assume off is incremented properly by
1539                  * read_record(), in case sizeof(*lcd)
1540                  * isn't the same as fsd->lsd_client_size.  */
1541                 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
1542                 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
1543                 if (rc) {
1544                         CERROR("%s: error reading last_rcvd %s idx %d off "
1545                                "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
1546                                cl_idx, off, rc);
1547                         rc = 0;
1548                         break; /* read error shouldn't cause startup to fail */
1549                 }
1550
1551                 if (lcd->lcd_uuid[0] == '\0') {
1552                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
1553                                cl_idx);
1554                         continue;
1555                 }
1556
1557                 last_transno = lcd_last_transno(lcd);
1558
1559                 /* These exports are cleaned up by disconnect, so they
1560                  * need to be set up like real exports as connect does.
1561                  */
1562                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: %llu"
1563                        " srv lr: %llu lx: %llu gen %u\n", lcd->lcd_uuid,
1564                        cl_idx, last_transno, lsd->lsd_last_transno,
1565                        lcd_last_xid(lcd), lcd->lcd_generation);
1566
1567                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
1568                 if (IS_ERR(exp)) {
1569                         if (PTR_ERR(exp) == -EALREADY) {
1570                                 /* export already exists, zero out this one */
1571                                 CERROR("%s: Duplicate export %s!\n",
1572                                        tgt_name(tgt), lcd->lcd_uuid);
1573                                 continue;
1574                         }
1575                         GOTO(err_out, rc = PTR_ERR(exp));
1576                 }
1577
1578                 ted = &exp->exp_target_data;
1579                 *ted->ted_lcd = *lcd;
1580
1581                 rc = tgt_client_add(env, exp, cl_idx);
1582                 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
1583                 /* VBR: set export last committed version */
1584                 exp->exp_last_committed = last_transno;
1585                 spin_lock(&exp->exp_lock);
1586                 exp->exp_connecting = 0;
1587                 exp->exp_in_recovery = 0;
1588                 spin_unlock(&exp->exp_lock);
1589                 atomic_inc(&obd->obd_max_recoverable_clients);
1590
1591                 if (tgt->lut_lsd.lsd_feature_incompat &
1592                     OBD_INCOMPAT_MULTI_RPCS &&
1593                     lcd->lcd_generation != 0) {
1594                         /* compute the highest valid client generation */
1595                         generation = max(generation, lcd->lcd_generation);
1596                         /* fill client_generation <-> export hash table */
1597                         rc = cfs_hash_add_unique(hash, &lcd->lcd_generation,
1598                                                  &exp->exp_gen_hash);
1599                         if (rc != 0) {
1600                                 CERROR("%s: duplicate export for client "
1601                                        "generation %u\n",
1602                                        tgt_name(tgt), lcd->lcd_generation);
1603                                 class_export_put(exp);
1604                                 GOTO(err_out, rc);
1605                         }
1606                 }
1607
1608                 class_export_put(exp);
1609
1610                 rc = rev_import_init(exp);
1611                 if (rc != 0) {
1612                         class_unlink_export(exp);
1613                         GOTO(err_out, rc);
1614                 }
1615
1616                 /* Need to check last_rcvd even for duplicated exports. */
1617                 CDEBUG(D_OTHER, "client at idx %d has last_transno = %llu\n",
1618                        cl_idx, last_transno);
1619
1620                 spin_lock(&tgt->lut_translock);
1621                 tgt->lut_last_transno = max(last_transno,
1622                                             tgt->lut_last_transno);
1623                 spin_unlock(&tgt->lut_translock);
1624         }
1625
1626         /* record highest valid client generation */
1627         atomic_set(&tgt->lut_client_generation, generation);
1628
1629 err_out:
1630         if (hash != NULL)
1631                 cfs_hash_putref(hash);
1632         OBD_FREE_PTR(lcd);
1633         RETURN(rc);
1634 }
1635
1636 struct server_compat_data {
1637         __u32 rocompat;
1638         __u32 incompat;
1639         __u32 rocinit;
1640         __u32 incinit;
1641 };
1642
1643 static struct server_compat_data tgt_scd[] = {
1644         [LDD_F_SV_TYPE_MDT] = {
1645                 .rocompat = OBD_ROCOMPAT_LOVOBJID,
1646                 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1647                             OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
1648                             OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI |
1649                             OBD_INCOMPAT_MULTI_RPCS,
1650                 .rocinit = OBD_ROCOMPAT_LOVOBJID,
1651                 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1652                            OBD_INCOMPAT_MULTI_OI,
1653         },
1654         [LDD_F_SV_TYPE_OST] = {
1655                 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
1656                 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
1657                             OBD_INCOMPAT_FID,
1658                 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
1659                 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1660         }
1661 };
1662
1663 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1664 {
1665         struct tgt_thread_info          *tti = tgt_th_info(env);
1666         struct lr_server_data           *lsd = &tgt->lut_lsd;
1667         unsigned long                    last_rcvd_size;
1668         __u32                            index;
1669         int                              rc, type;
1670
1671         rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1672         if (rc)
1673                 RETURN(rc);
1674
1675         last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1676
1677         /* ensure padding in the struct is the correct size */
1678         BUILD_BUG_ON(offsetof(struct lr_server_data, lsd_padding) +
1679                      sizeof(lsd->lsd_padding) != LR_SERVER_SIZE);
1680
1681         rc = server_name2index(tgt_name(tgt), &index, NULL);
1682         if (rc < 0) {
1683                 CERROR("%s: Can not get index from name: rc = %d\n",
1684                        tgt_name(tgt), rc);
1685                 RETURN(rc);
1686         }
1687         /* server_name2index() returns type */
1688         type = rc;
1689         if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1690                 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1691                 RETURN(-EINVAL);
1692         }
1693
1694         /* last_rcvd on OST doesn't provide reconstruct support because there
1695          * may be up to 8 in-flight write requests per single slot in
1696          * last_rcvd client data
1697          */
1698         tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1699
1700         if (last_rcvd_size == 0) {
1701                 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1702
1703                 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1704                        sizeof(lsd->lsd_uuid));
1705                 lsd->lsd_last_transno = 0;
1706                 lsd->lsd_mount_count = 0;
1707                 lsd->lsd_server_size = LR_SERVER_SIZE;
1708                 lsd->lsd_client_start = LR_CLIENT_START;
1709                 lsd->lsd_client_size = LR_CLIENT_SIZE;
1710                 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1711                 lsd->lsd_osd_index = index;
1712                 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1713                 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1714         } else {
1715                 rc = tgt_server_data_read(env, tgt);
1716                 if (rc) {
1717                         CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1718                                tgt_name(tgt), rc);
1719                         RETURN(rc);
1720                 }
1721                 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1722                         if (tgt->lut_bottom->dd_rdonly) {
1723                                 /* Such difference may be caused by mounting
1724                                  * up snapshot with new fsname under rd_only
1725                                  * mode. But even if it was NOT, it will not
1726                                  * damage the system because of "rd_only". */
1727                                 memcpy(lsd->lsd_uuid,
1728                                        tgt->lut_obd->obd_uuid.uuid,
1729                                        sizeof(lsd->lsd_uuid));
1730                         } else {
1731                                 LCONSOLE_ERROR_MSG(0x157, "Trying to start "
1732                                                    "OBD %s using the wrong "
1733                                                    "disk %s. Were the /dev/ "
1734                                                    "assignments rearranged?\n",
1735                                                    tgt->lut_obd->obd_uuid.uuid,
1736                                                    lsd->lsd_uuid);
1737                                 RETURN(-EINVAL);
1738                         }
1739                 }
1740
1741                 if (lsd->lsd_osd_index != index) {
1742                         LCONSOLE_ERROR_MSG(0x157,
1743                                            "%s: index %d in last rcvd is different with the index %d in config log, It might be disk corruption!\n",
1744                                            tgt_name(tgt),
1745                                            lsd->lsd_osd_index, index);
1746                         RETURN(-EINVAL);
1747                 }
1748         }
1749
1750         if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1751                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1752                        tgt_name(tgt),
1753                        lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1754                 RETURN(-EINVAL);
1755         }
1756
1757         if (type == LDD_F_SV_TYPE_MDT)
1758                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1759
1760         if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1761                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1762                        tgt_name(tgt),
1763                        lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1764                 RETURN(-EINVAL);
1765         }
1766         /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1767         if (type == LDD_F_SV_TYPE_MDT &&
1768             !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1769                 if (last_rcvd_size > lsd->lsd_client_start) {
1770                         LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1771                                       "remove all clients for interop needs\n",
1772                                       tgt_name(tgt));
1773                         rc = tgt_truncate_last_rcvd(env, tgt,
1774                                                     lsd->lsd_client_start);
1775                         if (rc)
1776                                 RETURN(rc);
1777                         last_rcvd_size = lsd->lsd_client_start;
1778                 }
1779                 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1780                 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1781         }
1782
1783         spin_lock(&tgt->lut_translock);
1784         tgt->lut_last_transno = lsd->lsd_last_transno;
1785         spin_unlock(&tgt->lut_translock);
1786
1787         lsd->lsd_mount_count++;
1788
1789         CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1790         CDEBUG(D_INODE, "%s: server last_transno: %llu\n",
1791                tgt_name(tgt), tgt->lut_last_transno);
1792         CDEBUG(D_INODE, "%s: server mount_count: %llu\n",
1793                tgt_name(tgt), lsd->lsd_mount_count);
1794         CDEBUG(D_INODE, "%s: server data size: %u\n",
1795                tgt_name(tgt), lsd->lsd_server_size);
1796         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1797                tgt_name(tgt), lsd->lsd_client_start);
1798         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1799                tgt_name(tgt), lsd->lsd_client_size);
1800         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1801                tgt_name(tgt), last_rcvd_size);
1802         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1803                tgt_name(tgt), lsd->lsd_subdir_count);
1804         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1805                last_rcvd_size <= lsd->lsd_client_start ? 0 :
1806                (last_rcvd_size - lsd->lsd_client_start) /
1807                 lsd->lsd_client_size);
1808         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1809
1810         if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1811             lsd->lsd_client_size == 0) {
1812                 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1813                 RETURN(-EINVAL);
1814         }
1815
1816         if (!tgt->lut_obd->obd_replayable)
1817                 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1818
1819         rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1820         if (rc < 0)
1821                 GOTO(err_client, rc);
1822
1823         spin_lock(&tgt->lut_translock);
1824         /* obd_last_committed is used for compatibility
1825          * with other lustre recovery code */
1826         tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1827         spin_unlock(&tgt->lut_translock);
1828
1829         tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1830         tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1831
1832         /* save it, so mount count and last_transno is current */
1833         rc = tgt_server_data_update(env, tgt, 0);
1834         if (rc < 0)
1835                 GOTO(err_client, rc);
1836
1837         RETURN(0);
1838
1839 err_client:
1840         class_disconnect_exports(tgt->lut_obd);
1841         return rc;
1842 }
1843
1844 /* add credits for last_rcvd update */
1845 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1846                      void *cookie)
1847 {
1848         struct lu_target        *tgt = cookie;
1849         struct tgt_session_info *tsi;
1850         struct tgt_thread_info  *tti = tgt_th_info(env);
1851         struct dt_object        *dto;
1852         int                      rc;
1853
1854         /* For readonly case, the caller should have got failure
1855          * when start the transaction. If the logic comes here,
1856          * there must be something wrong. */
1857         if (unlikely(tgt->lut_bottom->dd_rdonly)) {
1858                 dump_stack();
1859                 LBUG();
1860         }
1861
1862         /* if there is no session, then this transaction is not result of
1863          * request processing but some local operation */
1864         if (env->le_ses == NULL)
1865                 return 0;
1866
1867         LASSERT(tgt->lut_last_rcvd);
1868         tsi = tgt_ses_info(env);
1869         /* OFD may start transaction without export assigned */
1870         if (tsi->tsi_exp == NULL)
1871                 return 0;
1872
1873         if (tgt_is_multimodrpcs_client(tsi->tsi_exp)) {
1874                 /*
1875                  * Use maximum possible file offset for declaration to ensure
1876                  * ZFS will reserve enough credits for a write anywhere in this
1877                  * file, since we don't know where in the file the write will be
1878                  * because a replay slot has not been assigned.  This should be
1879                  * replaced by dmu_tx_hold_append() when available.
1880                  */
1881                 tti->tti_buf.lb_buf = NULL;
1882                 tti->tti_buf.lb_len = sizeof(struct lsd_reply_data);
1883                 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
1884                 rc = dt_declare_record_write(env, dto, &tti->tti_buf, -1, th);
1885                 if (rc)
1886                         return rc;
1887         } else {
1888                 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
1889                 tti_buf_lcd(tti);
1890                 tti->tti_off = tsi->tsi_exp->exp_target_data.ted_lr_off;
1891                 rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1892                                              tti->tti_off, th);
1893                 if (rc)
1894                         return rc;
1895         }
1896
1897         if (tsi->tsi_vbr_obj != NULL &&
1898             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1899                 dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
1900                 rc = dt_declare_version_set(env, dto, th);
1901         }
1902
1903         return rc;
1904 }
1905
1906 /* Update last_rcvd records with latests transaction data */
1907 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1908                     void *cookie)
1909 {
1910         struct lu_target        *tgt = cookie;
1911         struct tgt_session_info *tsi;
1912         struct tgt_thread_info  *tti = tgt_th_info(env);
1913         struct dt_object        *obj = NULL;
1914         int                      rc;
1915         bool                     echo_client;
1916
1917         if (env->le_ses == NULL)
1918                 return 0;
1919
1920         tsi = tgt_ses_info(env);
1921         /* OFD may start transaction without export assigned */
1922         if (tsi->tsi_exp == NULL)
1923                 return 0;
1924
1925         echo_client = (tgt_ses_req(tsi) == NULL && tsi->tsi_xid == 0);
1926
1927         if (tti->tti_has_trans && !echo_client) {
1928                 if (tti->tti_mult_trans == 0) {
1929                         CDEBUG(D_HA, "More than one transaction %llu\n",
1930                                tti->tti_transno);
1931                         RETURN(0);
1932                 }
1933                 /* we need another transno to be assigned */
1934                 tti->tti_transno = 0;
1935         } else if (th->th_result == 0) {
1936                 tti->tti_has_trans = 1;
1937         }
1938
1939         if (tsi->tsi_vbr_obj != NULL &&
1940             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1941                 obj = tsi->tsi_vbr_obj;
1942         }
1943
1944         if (unlikely(echo_client)) /* echo client special case */
1945                 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1946                                                tsi->tsi_exp);
1947         else
1948                 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
1949                                           tgt_ses_req(tsi));
1950         return rc;
1951 }
1952
1953 int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
1954 {
1955         struct tgt_thread_info  *tti = tgt_th_info(env);
1956         struct lsd_reply_data   *lrd = &tti->tti_lrd;
1957         unsigned long            reply_data_size;
1958         int                      rc;
1959         struct lsd_reply_header *lrh = NULL;
1960         struct tg_reply_data    *trd = NULL;
1961         int                      idx;
1962         loff_t                   off;
1963         struct cfs_hash         *hash = NULL;
1964         struct obd_export       *exp;
1965         struct tg_export_data   *ted;
1966         int                      reply_data_recovered = 0;
1967
1968         rc = dt_attr_get(env, tgt->lut_reply_data, &tti->tti_attr);
1969         if (rc)
1970                 GOTO(out, rc);
1971         reply_data_size = (unsigned long)tti->tti_attr.la_size;
1972
1973         OBD_ALLOC_PTR(lrh);
1974         if (lrh == NULL)
1975                 GOTO(out, rc = -ENOMEM);
1976
1977         if (reply_data_size == 0) {
1978                 CDEBUG(D_INFO, "%s: new reply_data file, initializing\n",
1979                        tgt_name(tgt));
1980                 lrh->lrh_magic = LRH_MAGIC;
1981                 lrh->lrh_header_size = sizeof(struct lsd_reply_header);
1982                 lrh->lrh_reply_size = sizeof(struct lsd_reply_data);
1983                 rc = tgt_reply_header_write(env, tgt, lrh);
1984                 if (rc) {
1985                         CERROR("%s: error writing %s: rc = %d\n",
1986                                tgt_name(tgt), REPLY_DATA, rc);
1987                         GOTO(out, rc);
1988                 }
1989         } else {
1990                 rc = tgt_reply_header_read(env, tgt, lrh);
1991                 if (rc) {
1992                         CERROR("%s: error reading %s: rc = %d\n",
1993                                tgt_name(tgt), REPLY_DATA, rc);
1994                         GOTO(out, rc);
1995                 }
1996                 if (lrh->lrh_magic != LRH_MAGIC ||
1997                     lrh->lrh_header_size != sizeof(struct lsd_reply_header) ||
1998                     lrh->lrh_reply_size != sizeof(struct lsd_reply_data)) {
1999                         CERROR("%s: invalid header in %s\n",
2000                                tgt_name(tgt), REPLY_DATA);
2001                         GOTO(out, rc = -EINVAL);
2002                 }
2003
2004                 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
2005                 if (hash == NULL)
2006                         GOTO(out, rc = -ENODEV);
2007
2008                 OBD_ALLOC_PTR(trd);
2009                 if (trd == NULL)
2010                         GOTO(out, rc = -ENOMEM);
2011
2012                 /* Load reply_data from disk */
2013                 for (idx = 0, off = sizeof(struct lsd_reply_header);
2014                      off < reply_data_size;
2015                      idx++, off += sizeof(struct lsd_reply_data)) {
2016                         rc = tgt_reply_data_read(env, tgt, lrd, off);
2017                         if (rc) {
2018                                 CERROR("%s: error reading %s: rc = %d\n",
2019                                        tgt_name(tgt), REPLY_DATA, rc);
2020                                 GOTO(out, rc);
2021                         }
2022
2023                         exp = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
2024                         if (exp == NULL) {
2025                                 /* old reply data from a disconnected client */
2026                                 continue;
2027                         }
2028                         ted = &exp->exp_target_data;
2029                         mutex_lock(&ted->ted_lcd_lock);
2030
2031                         /* create in-memory reply_data and link it to
2032                          * target export's reply list */
2033                         rc = tgt_set_reply_slot(tgt, idx);
2034                         if (rc != 0) {
2035                                 mutex_unlock(&ted->ted_lcd_lock);
2036                                 GOTO(out, rc);
2037                         }
2038                         trd->trd_reply = *lrd;
2039                         trd->trd_pre_versions[0] = 0;
2040                         trd->trd_pre_versions[1] = 0;
2041                         trd->trd_pre_versions[2] = 0;
2042                         trd->trd_pre_versions[3] = 0;
2043                         trd->trd_index = idx;
2044                         trd->trd_tag = 0;
2045                         list_add(&trd->trd_list, &ted->ted_reply_list);
2046                         ted->ted_reply_cnt++;
2047                         if (ted->ted_reply_cnt > ted->ted_reply_max)
2048                                 ted->ted_reply_max = ted->ted_reply_cnt;
2049
2050                         CDEBUG(D_HA, "%s: restore reply %p: xid %llu, "
2051                                "transno %llu, client gen %u, slot idx %d\n",
2052                                tgt_name(tgt), trd, lrd->lrd_xid,
2053                                lrd->lrd_transno, lrd->lrd_client_gen,
2054                                trd->trd_index);
2055
2056                         /* update export last committed transation */
2057                         exp->exp_last_committed = max(exp->exp_last_committed,
2058                                                       lrd->lrd_transno);
2059                         /* Update lcd_last_transno as well for check in
2060                          * tgt_release_reply_data() or the latest client
2061                          * transno can be lost.
2062                          */
2063                         ted->ted_lcd->lcd_last_transno =
2064                                 max(ted->ted_lcd->lcd_last_transno,
2065                                     exp->exp_last_committed);
2066
2067                         mutex_unlock(&ted->ted_lcd_lock);
2068                         class_export_put(exp);
2069
2070                         /* update target last committed transaction */
2071                         spin_lock(&tgt->lut_translock);
2072                         tgt->lut_last_transno = max(tgt->lut_last_transno,
2073                                                     lrd->lrd_transno);
2074                         spin_unlock(&tgt->lut_translock);
2075
2076                         reply_data_recovered++;
2077
2078                         OBD_ALLOC_PTR(trd);
2079                         if (trd == NULL)
2080                                 GOTO(out, rc = -ENOMEM);
2081                 }
2082                 CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
2083                        tgt_name(tgt), reply_data_recovered);
2084         }
2085
2086         spin_lock(&tgt->lut_translock);
2087         /* obd_last_committed is used for compatibility
2088          * with other lustre recovery code */
2089         tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
2090         spin_unlock(&tgt->lut_translock);
2091
2092         rc = 0;
2093
2094 out:
2095         if (hash != NULL)
2096                 cfs_hash_putref(hash);
2097         if (trd != NULL)
2098                 OBD_FREE_PTR(trd);
2099         if (lrh != NULL)
2100                 OBD_FREE_PTR(lrh);
2101         return rc;
2102 }
2103
2104 static int tgt_check_lookup_req(struct ptlrpc_request *req, int lookup,
2105                                 struct tg_reply_data *trd)
2106 {
2107         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2108         struct lu_target *lut = class_exp2tgt(req->rq_export);
2109         __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
2110         int rc = 0;
2111         struct tg_reply_data *reply;
2112         bool check_increasing;
2113
2114         if (tag == 0)
2115                 return 0;
2116
2117         check_increasing = tgt_is_increasing_xid_client(req->rq_export) &&
2118                            !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
2119         if (!lookup && !check_increasing)
2120                 return 0;
2121
2122         list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
2123                 if (lookup && reply->trd_reply.lrd_xid == req->rq_xid) {
2124                         rc = 1;
2125                         if (trd != NULL)
2126                                 *trd = *reply;
2127                         break;
2128                 } else if (check_increasing && reply->trd_tag == tag &&
2129                            reply->trd_reply.lrd_xid > req->rq_xid) {
2130                         rc = -EPROTO;
2131                         CERROR("%s: busy tag=%u req_xid=%llu, trd=%p: xid=%llu transno=%llu client_gen=%u slot_idx=%d: rc = %d\n",
2132                                tgt_name(lut), tag, req->rq_xid, trd,
2133                                reply->trd_reply.lrd_xid,
2134                                reply->trd_reply.lrd_transno,
2135                                reply->trd_reply.lrd_client_gen,
2136                                reply->trd_index, rc);
2137                         break;
2138                 }
2139         }
2140
2141         return rc;
2142 }
2143
2144 /* Look for a reply data matching specified request @req
2145  * A copy is returned in @trd if the pointer is not NULL
2146  */
2147 int tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
2148 {
2149         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2150         int found = 0;
2151         bool not_replay = !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
2152
2153         mutex_lock(&ted->ted_lcd_lock);
2154         if (not_replay && req->rq_xid <= req->rq_export->exp_last_xid) {
2155                 /* A check for the last_xid is needed here in case there is
2156                  * no reply data is left in the list. It may happen if another
2157                  * RPC on another slot increased the last_xid between our
2158                  * process_req_last_xid & tgt_lookup_reply calls */
2159                 found = -EPROTO;
2160         } else {
2161                 found = tgt_check_lookup_req(req, 1, trd);
2162         }
2163         mutex_unlock(&ted->ted_lcd_lock);
2164
2165         CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d last_xid %llu\n",
2166                tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid, found,
2167                req->rq_export->exp_last_xid);
2168
2169         return found;
2170 }
2171 EXPORT_SYMBOL(tgt_lookup_reply);
2172
2173 int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
2174 {
2175         struct tg_export_data   *ted = &exp->exp_target_data;
2176         struct lu_target        *lut = class_exp2tgt(exp);
2177         struct tg_reply_data    *trd, *tmp;
2178
2179
2180         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
2181                 if (trd->trd_reply.lrd_xid > rcvd_xid)
2182                         continue;
2183                 ted->ted_release_xid++;
2184                 tgt_release_reply_data(lut, ted, trd);
2185         }
2186
2187         return 0;
2188 }
2189
2190 int tgt_handle_tag(struct ptlrpc_request *req)
2191 {
2192         return tgt_check_lookup_req(req, 0, NULL);
2193 }
2194