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