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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Lustre Unified Target
  37  * These are common function to work with last_received file
  38  *
  39  * Author: Mikhail Pershin <mike.pershin@intel.com>
  40  */
  41 #include <obd.h>
  42 #include <obd_class.h>
  43 #include <lustre_fid.h>
  44
  45 #include "tgt_internal.h"
  46
  47
  48 /* Allocate a bitmap for a chunk of reply data slots */
  49 static int tgt_bitmap_chunk_alloc(struct lu_target *lut, int chunk)
  50 {
  51         unsigned long *bm;
  52
  53         OBD_ALLOC(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) * sizeof(long));
  54         if (bm == NULL)
  55                 return -ENOMEM;
  56
  57         spin_lock(&lut->lut_client_bitmap_lock);
  58
  59         if (lut->lut_reply_bitmap[chunk] != NULL) {
  60                 /* someone else already allocated the bitmap for this chunk */
  61                 spin_unlock(&lut->lut_client_bitmap_lock);
  62                 OBD_FREE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
  63                          sizeof(long));
  64                 return 0;
  65         }
  66
  67         lut->lut_reply_bitmap[chunk] = bm;
  68
  69         spin_unlock(&lut->lut_client_bitmap_lock);
  70
  71         return 0;
  72 }
  73
  74 /* Look for an available reply data slot in the bitmap
  75  * of the target @lut
  76  * Allocate bitmap chunk when first used
  77  * XXX algo could be improved if this routine limits performance
  78  */
  79 static int tgt_find_free_reply_slot(struct lu_target *lut)
  80 {
  81         unsigned long *bmp;
  82         int chunk = 0;
  83         int rc;
  84         int b;
  85
  86         for (chunk = 0; chunk < LUT_REPLY_SLOTS_MAX_CHUNKS; chunk++) {
  87                 /* allocate the bitmap chunk if necessary */
  88                 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
  89                         rc = tgt_bitmap_chunk_alloc(lut, chunk);
  90                         if (rc != 0)
  91                                 return rc;
  92                 }
  93                 bmp = lut->lut_reply_bitmap[chunk];
  94
  95                 /* look for an available slot in this chunk */
  96                 do {
  97                         b = find_first_zero_bit(bmp, LUT_REPLY_SLOTS_PER_CHUNK);
  98                         if (b >= LUT_REPLY_SLOTS_PER_CHUNK)
  99                                 break;
 100
 101                         /* found one */
 102                         if (test_and_set_bit(b, bmp) == 0)
 103                                 return chunk * LUT_REPLY_SLOTS_PER_CHUNK + b;
 104                 } while (true);
 105         }
 106
 107         return -ENOSPC;
 108 }
 109
 110 /* Mark the reply data slot @idx 'used' in the corresponding bitmap chunk
 111  * of the target @lut
 112  * Allocate the bitmap chunk if necessary
 113  */
 114 static int tgt_set_reply_slot(struct lu_target *lut, int idx)
 115 {
 116         int chunk;
 117         int b;
 118         int rc;
 119
 120         chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
 121         b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
 122
 123         LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
 124         LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
 125
 126         /* allocate the bitmap chunk if necessary */
 127         if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
 128                 rc = tgt_bitmap_chunk_alloc(lut, chunk);
 129                 if (rc != 0)
 130                         return rc;
 131         }
 132
 133         /* mark the slot 'used' in this chunk */
 134         if (test_and_set_bit(b, lut->lut_reply_bitmap[chunk]) != 0) {
 135                 CERROR("%s: slot %d already set in bitmap\n",
 136                        tgt_name(lut), idx);
 137                 return -EALREADY;
 138         }
 139
 140         return 0;
 141 }
 142
 143
 144 /* Mark the reply data slot @idx 'unused' in the corresponding bitmap chunk
 145  * of the target @lut
 146  */
 147 static int tgt_clear_reply_slot(struct lu_target *lut, int idx)
 148 {
 149         int chunk;
 150         int b;
 151
 152         chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
 153         b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
 154
 155         LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
 156         LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
 157
 158         if (test_and_clear_bit(b, lut->lut_reply_bitmap[chunk]) == 0) {
 159                 CERROR("%s: slot %d already clear in bitmap\n",
 160                        tgt_name(lut), idx);
 161                 return -EALREADY;
 162         }
 163
 164         return 0;
 165 }
 166
 167
 168 /* Read header of reply_data file of target @tgt into structure @lrh */
 169 static int tgt_reply_header_read(const struct lu_env *env,
 170                                  struct lu_target *tgt,
 171                                  struct lsd_reply_header *lrh)
 172 {
 173         int                      rc;
 174         struct lsd_reply_header  buf;
 175         struct tgt_thread_info  *tti = tgt_th_info(env);
 176
 177         tti->tti_off = 0;
 178         tti->tti_buf.lb_buf = &buf;
 179         tti->tti_buf.lb_len = sizeof(buf);
 180
 181         rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
 182                             &tti->tti_off);
 183         if (rc != 0)
 184                 return rc;
 185
 186         lrh->lrh_magic = le32_to_cpu(buf.lrh_magic);
 187         lrh->lrh_header_size = le32_to_cpu(buf.lrh_header_size);
 188         lrh->lrh_reply_size = le32_to_cpu(buf.lrh_reply_size);
 189
 190         CDEBUG(D_HA, "%s: read %s header. magic=0x%08x "
 191                "header_size=%d reply_size=%d\n",
 192                 tgt->lut_obd->obd_name, REPLY_DATA,
 193                 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
 194
 195         return 0;
 196 }
 197
 198 /* Write header into replay_data file of target @tgt from structure @lrh */
 199 static int tgt_reply_header_write(const struct lu_env *env,
 200                                   struct lu_target *tgt,
 201                                   struct lsd_reply_header *lrh)
 202 {
 203         int                      rc;
 204         struct lsd_reply_header  buf;
 205         struct tgt_thread_info  *tti = tgt_th_info(env);
 206         struct thandle          *th;
 207         struct dt_object        *dto;
 208
 209         CDEBUG(D_HA, "%s: write %s header. magic=0x%08x "
 210                "header_size=%d reply_size=%d\n",
 211                 tgt->lut_obd->obd_name, REPLY_DATA,
 212                 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
 213
 214         buf.lrh_magic = cpu_to_le32(lrh->lrh_magic);
 215         buf.lrh_header_size = cpu_to_le32(lrh->lrh_header_size);
 216         buf.lrh_reply_size = cpu_to_le32(lrh->lrh_reply_size);
 217
 218         th = dt_trans_create(env, tgt->lut_bottom);
 219         if (IS_ERR(th))
 220                 return PTR_ERR(th);
 221         th->th_sync = 1;
 222
 223         tti->tti_off = 0;
 224         tti->tti_buf.lb_buf = &buf;
 225         tti->tti_buf.lb_len = sizeof(buf);
 226
 227         rc = dt_declare_record_write(env, tgt->lut_reply_data,
 228                                      &tti->tti_buf, tti->tti_off, th);
 229         if (rc)
 230                 GOTO(out, rc);
 231
 232         rc = dt_trans_start(env, tgt->lut_bottom, th);
 233         if (rc)
 234                 GOTO(out, rc);
 235
 236         dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
 237         rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 238 out:
 239         dt_trans_stop(env, tgt->lut_bottom, th);
 240         return rc;
 241 }
 242
 243 /* Write the reply data @lrd into reply_data file of target @tgt
 244  * at offset @off
 245  */
 246 static int tgt_reply_data_write(const struct lu_env *env, struct lu_target *tgt,
 247                                 struct lsd_reply_data *lrd, loff_t off,
 248                                 struct thandle *th)
 249 {
 250         struct tgt_thread_info  *tti = tgt_th_info(env);
 251         struct dt_object        *dto;
 252         struct lsd_reply_data   *buf = &tti->tti_lrd;
 253
 254         lrd->lrd_result = ptlrpc_status_hton(lrd->lrd_result);
 255
 256         buf->lrd_transno         = cpu_to_le64(lrd->lrd_transno);
 257         buf->lrd_xid             = cpu_to_le64(lrd->lrd_xid);
 258         buf->lrd_data            = cpu_to_le64(lrd->lrd_data);
 259         buf->lrd_result          = cpu_to_le32(lrd->lrd_result);
 260         buf->lrd_client_gen      = cpu_to_le32(lrd->lrd_client_gen);
 261
 262         lrd->lrd_result = ptlrpc_status_ntoh(lrd->lrd_result);
 263
 264         tti->tti_off = off;
 265         tti->tti_buf.lb_buf = buf;
 266         tti->tti_buf.lb_len = sizeof(*buf);
 267
 268         dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
 269         return dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 270 }
 271
 272 /* Read the reply data from reply_data file of target @tgt at offset @off
 273  * into structure @lrd
 274  */
 275 static int tgt_reply_data_read(const struct lu_env *env, struct lu_target *tgt,
 276                                struct lsd_reply_data *lrd, loff_t off)
 277 {
 278         int                      rc;
 279         struct tgt_thread_info  *tti = tgt_th_info(env);
 280         struct lsd_reply_data   *buf = &tti->tti_lrd;
 281
 282         tti->tti_off = off;
 283         tti->tti_buf.lb_buf = buf;
 284         tti->tti_buf.lb_len = sizeof(*buf);
 285
 286         rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
 287                             &tti->tti_off);
 288         if (rc != 0)
 289                 return rc;
 290
 291         lrd->lrd_transno         = le64_to_cpu(buf->lrd_transno);
 292         lrd->lrd_xid             = le64_to_cpu(buf->lrd_xid);
 293         lrd->lrd_data            = le64_to_cpu(buf->lrd_data);
 294         lrd->lrd_result          = le32_to_cpu(buf->lrd_result);
 295         lrd->lrd_client_gen      = le32_to_cpu(buf->lrd_client_gen);
 296
 297         return 0;
 298 }
 299
 300
 301 /* Free the in-memory reply data structure @trd and release
 302  * the corresponding slot in the reply_data file of target @lut
 303  * Called with ted_lcd_lock held
 304  */
 305 static void tgt_free_reply_data(struct lu_target *lut,
 306                                 struct tg_export_data *ted,
 307                                 struct tg_reply_data *trd)
 308 {
 309         CDEBUG(D_TRACE, "%s: free reply data %p: xid %llu, transno %llu, "
 310                "client gen %u, slot idx %d\n",
 311                tgt_name(lut), trd, trd->trd_reply.lrd_xid,
 312                trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
 313                trd->trd_index);
 314
 315         LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
 316
 317         list_del(&trd->trd_list);
 318         ted->ted_reply_cnt--;
 319         tgt_clear_reply_slot(lut, trd->trd_index);
 320         OBD_FREE_PTR(trd);
 321 }
 322
 323 /* Release the reply data @trd from target @lut
 324  * The reply data with the highest transno for this export
 325  * is retained to ensure correctness of target recovery
 326  * Called with ted_lcd_lock held
 327  */
 328 static void tgt_release_reply_data(struct lu_target *lut,
 329                                    struct tg_export_data *ted,
 330                                    struct tg_reply_data *trd)
 331 {
 332         CDEBUG(D_TRACE, "%s: release reply data %p: xid %llu, transno %llu, "
 333                "client gen %u, slot idx %d\n",
 334                tgt_name(lut), trd, trd->trd_reply.lrd_xid,
 335                trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
 336                trd->trd_index);
 337
 338         LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
 339
 340         /* Do not free the reply data corresponding to the
 341          * highest transno of this export.
 342          * This ensures on-disk reply data is kept and
 343          * last committed transno can be restored from disk in case
 344          * of target recovery
 345          */
 346         if (trd->trd_reply.lrd_transno == ted->ted_lcd->lcd_last_transno) {
 347                 /* free previous retained reply */
 348                 if (ted->ted_reply_last != NULL)
 349                         tgt_free_reply_data(lut, ted, ted->ted_reply_last);
 350                 /* retain the reply */
 351                 list_del_init(&trd->trd_list);
 352                 ted->ted_reply_last = trd;
 353         } else {
 354                 tgt_free_reply_data(lut, ted, trd);
 355         }
 356 }
 357
 358 static inline struct lu_buf *tti_buf_lsd(struct tgt_thread_info *tti)
 359 {
 360         tti->tti_buf.lb_buf = &tti->tti_lsd;
 361         tti->tti_buf.lb_len = sizeof(tti->tti_lsd);
 362         return &tti->tti_buf;
 363 }
 364
 365 static inline struct lu_buf *tti_buf_lcd(struct tgt_thread_info *tti)
 366 {
 367         tti->tti_buf.lb_buf = &tti->tti_lcd;
 368         tti->tti_buf.lb_len = sizeof(tti->tti_lcd);
 369         return &tti->tti_buf;
 370 }
 371
 372 /**
 373  * Allocate in-memory data for client slot related to export.
 374  */
 375 int tgt_client_alloc(struct obd_export *exp)
 376 {
 377         ENTRY;
 378         LASSERT(exp != exp->exp_obd->obd_self_export);
 379
 380         OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
 381         if (exp->exp_target_data.ted_lcd == NULL)
 382                 RETURN(-ENOMEM);
 383         /* Mark that slot is not yet valid, 0 doesn't work here */
 384         exp->exp_target_data.ted_lr_idx = -1;
 385         INIT_LIST_HEAD(&exp->exp_target_data.ted_reply_list);
 386         mutex_init(&exp->exp_target_data.ted_lcd_lock);
 387         RETURN(0);
 388 }
 389 EXPORT_SYMBOL(tgt_client_alloc);
 390
 391 /**
 392  * Free in-memory data for client slot related to export.
 393  */
 394 void tgt_client_free(struct obd_export *exp)
 395 {
 396         struct tg_export_data   *ted = &exp->exp_target_data;
 397         struct lu_target        *lut = class_exp2tgt(exp);
 398         struct tg_reply_data    *trd, *tmp;
 399
 400         LASSERT(exp != exp->exp_obd->obd_self_export);
 401
 402         /* free reply data */
 403         mutex_lock(&ted->ted_lcd_lock);
 404         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
 405                 tgt_release_reply_data(lut, ted, trd);
 406         }
 407         if (ted->ted_reply_last != NULL) {
 408                 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
 409                 ted->ted_reply_last = NULL;
 410         }
 411         mutex_unlock(&ted->ted_lcd_lock);
 412
 413         if (!hlist_unhashed(&exp->exp_gen_hash))
 414                 cfs_hash_del(exp->exp_obd->obd_gen_hash,
 415                              &ted->ted_lcd->lcd_generation,
 416                              &exp->exp_gen_hash);
 417
 418         OBD_FREE_PTR(ted->ted_lcd);
 419         ted->ted_lcd = NULL;
 420
 421         /* Slot may be not yet assigned */
 422         if (ted->ted_lr_idx < 0)
 423                 return;
 424         /* Clear bit when lcd is freed */
 425         LASSERT(lut && lut->lut_client_bitmap);
 426         if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
 427                 CERROR("%s: client %u bit already clear in bitmap\n",
 428                        exp->exp_obd->obd_name, ted->ted_lr_idx);
 429                 LBUG();
 430         }
 431
 432         if (tgt_is_multimodrpcs_client(exp) && !exp->exp_obd->obd_stopping)
 433                 atomic_dec(&lut->lut_num_clients);
 434 }
 435 EXPORT_SYMBOL(tgt_client_free);
 436
 437 int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
 438                          struct lsd_client_data *lcd, loff_t *off, int index)
 439 {
 440         struct tgt_thread_info  *tti = tgt_th_info(env);
 441         int                      rc;
 442
 443         tti_buf_lcd(tti);
 444         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
 445         if (rc == 0) {
 446                 check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
 447                 lcd_le_to_cpu(&tti->tti_lcd, lcd);
 448                 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
 449                 lcd->lcd_last_close_result =
 450                         ptlrpc_status_ntoh(lcd->lcd_last_close_result);
 451         }
 452
 453         CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = "LPU64
 454                ", last_xid = "LPU64", last_result = %u, last_data = %u, "
 455                "last_close_transno = "LPU64", last_close_xid = "LPU64", "
 456                "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
 457                *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
 458                lcd->lcd_last_result, lcd->lcd_last_data,
 459                lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
 460                lcd->lcd_last_close_result, rc);
 461         return rc;
 462 }
 463
 464 int tgt_client_data_write(const struct lu_env *env, struct lu_target *tgt,
 465                           struct lsd_client_data *lcd, loff_t *off,
 466                           struct thandle *th)
 467 {
 468         struct tgt_thread_info *tti = tgt_th_info(env);
 469         struct dt_object        *dto;
 470
 471         lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
 472         lcd->lcd_last_close_result =
 473                 ptlrpc_status_hton(lcd->lcd_last_close_result);
 474         lcd_cpu_to_le(lcd, &tti->tti_lcd);
 475         tti_buf_lcd(tti);
 476
 477         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
 478         return dt_record_write(env, dto, &tti->tti_buf, off, th);
 479 }
 480
 481 /**
 482  * Update client data in last_rcvd
 483  */
 484 static int tgt_client_data_update(const struct lu_env *env,
 485                                   struct obd_export *exp)
 486 {
 487         struct tg_export_data   *ted = &exp->exp_target_data;
 488         struct lu_target        *tgt = class_exp2tgt(exp);
 489         struct tgt_thread_info  *tti = tgt_th_info(env);
 490         struct thandle          *th;
 491         int                      rc = 0;
 492
 493         ENTRY;
 494
 495         if (unlikely(tgt == NULL)) {
 496                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
 497                           class_exp2obd(exp)->obd_name);
 498                 RETURN(-EINVAL);
 499         }
 500
 501         th = dt_trans_create(env, tgt->lut_bottom);
 502         if (IS_ERR(th))
 503                 RETURN(PTR_ERR(th));
 504
 505         tti_buf_lcd(tti);
 506         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 507                                      &tti->tti_buf,
 508                                      ted->ted_lr_off, th);
 509         if (rc)
 510                 GOTO(out, rc);
 511
 512         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 513         if (rc)
 514                 GOTO(out, rc);
 515         /*
 516          * Until this operations will be committed the sync is needed
 517          * for this export. This should be done _after_ starting the
 518          * transaction so that many connecting clients will not bring
 519          * server down with lots of sync writes.
 520          */
 521         rc = tgt_new_client_cb_add(th, exp);
 522         if (rc) {
 523                 /* can't add callback, do sync now */
 524                 th->th_sync = 1;
 525         } else {
 526                 spin_lock(&exp->exp_lock);
 527                 exp->exp_need_sync = 1;
 528                 spin_unlock(&exp->exp_lock);
 529         }
 530
 531         tti->tti_off = ted->ted_lr_off;
 532         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
 533         EXIT;
 534 out:
 535         dt_trans_stop(env, tgt->lut_bottom, th);
 536         CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
 537                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 538                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 539
 540         return rc;
 541 }
 542
 543 int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
 544 {
 545         struct tgt_thread_info  *tti = tgt_th_info(env);
 546         int                      rc;
 547
 548         tti->tti_off = 0;
 549         tti_buf_lsd(tti);
 550         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
 551                             &tti->tti_off);
 552         if (rc == 0)
 553                 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
 554
 555         CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
 556                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 557                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 558         return rc;
 559 }
 560
 561 int tgt_server_data_write(const struct lu_env *env, struct lu_target *tgt,
 562                           struct thandle *th)
 563 {
 564         struct tgt_thread_info  *tti = tgt_th_info(env);
 565         struct dt_object        *dto;
 566         int                      rc;
 567
 568         ENTRY;
 569
 570         tti->tti_off = 0;
 571         tti_buf_lsd(tti);
 572         lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
 573
 574         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
 575         rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 576
 577         CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
 578                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 579                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 580
 581         RETURN(rc);
 582 }
 583
 584 /**
 585  * Update server data in last_rcvd
 586  */
 587 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
 588                            int sync)
 589 {
 590         struct tgt_thread_info  *tti = tgt_th_info(env);
 591         struct thandle          *th;
 592         int                      rc = 0;
 593
 594         ENTRY;
 595
 596         CDEBUG(D_SUPER,
 597                "%s: mount_count is "LPU64", last_transno is "LPU64"\n",
 598                tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
 599                tgt->lut_last_transno);
 600
 601         /* Always save latest transno to keep it fresh */
 602         spin_lock(&tgt->lut_translock);
 603         tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
 604         spin_unlock(&tgt->lut_translock);
 605
 606         th = dt_trans_create(env, tgt->lut_bottom);
 607         if (IS_ERR(th))
 608                 RETURN(PTR_ERR(th));
 609
 610         th->th_sync = sync;
 611
 612         tti_buf_lsd(tti);
 613         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 614                                      &tti->tti_buf, tti->tti_off, th);
 615         if (rc)
 616                 GOTO(out, rc);
 617
 618         rc = dt_trans_start(env, tgt->lut_bottom, th);
 619         if (rc)
 620                 GOTO(out, rc);
 621
 622         rc = tgt_server_data_write(env, tgt, th);
 623 out:
 624         dt_trans_stop(env, tgt->lut_bottom, th);
 625
 626         CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
 627                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 628                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 629         RETURN(rc);
 630 }
 631 EXPORT_SYMBOL(tgt_server_data_update);
 632
 633 int tgt_truncate_last_rcvd(const struct lu_env *env, struct lu_target *tgt,
 634                            loff_t size)
 635 {
 636         struct dt_object *dt = tgt->lut_last_rcvd;
 637         struct thandle   *th;
 638         struct lu_attr    attr;
 639         int               rc;
 640
 641         ENTRY;
 642
 643         attr.la_size = size;
 644         attr.la_valid = LA_SIZE;
 645
 646         th = dt_trans_create(env, tgt->lut_bottom);
 647         if (IS_ERR(th))
 648                 RETURN(PTR_ERR(th));
 649         rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
 650         if (rc)
 651                 GOTO(cleanup, rc);
 652         rc = dt_declare_attr_set(env, dt, &attr, th);
 653         if (rc)
 654                 GOTO(cleanup, rc);
 655         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 656         if (rc)
 657                 GOTO(cleanup, rc);
 658
 659         rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
 660         if (rc == 0)
 661                 rc = dt_attr_set(env, dt, &attr, th);
 662
 663 cleanup:
 664         dt_trans_stop(env, tgt->lut_bottom, th);
 665
 666         RETURN(rc);
 667 }
 668
 669 static void tgt_client_epoch_update(const struct lu_env *env,
 670                                     struct obd_export *exp)
 671 {
 672         struct lsd_client_data  *lcd = exp->exp_target_data.ted_lcd;
 673         struct lu_target        *tgt = class_exp2tgt(exp);
 674
 675         LASSERT(tgt && tgt->lut_bottom);
 676         /** VBR: set client last_epoch to current epoch */
 677         if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
 678                 return;
 679         lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
 680         tgt_client_data_update(env, exp);
 681 }
 682
 683 /**
 684  * Update boot epoch when recovery ends
 685  */
 686 void tgt_boot_epoch_update(struct lu_target *tgt)
 687 {
 688         struct lu_env            env;
 689         struct ptlrpc_request   *req;
 690         __u32                    start_epoch;
 691         struct list_head         client_list;
 692         int                      rc;
 693
 694         if (tgt->lut_obd->obd_stopping)
 695                 return;
 696
 697         rc = lu_env_init(&env, LCT_DT_THREAD);
 698         if (rc) {
 699                 CERROR("%s: can't initialize environment: rc = %d\n",
 700                         tgt->lut_obd->obd_name, rc);
 701                 return;
 702         }
 703
 704         spin_lock(&tgt->lut_translock);
 705         start_epoch = lr_epoch(tgt->lut_last_transno) + 1;
 706         tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
 707         tgt->lut_lsd.lsd_start_epoch = start_epoch;
 708         spin_unlock(&tgt->lut_translock);
 709
 710         INIT_LIST_HEAD(&client_list);
 711         /**
 712          * The recovery is not yet finished and final queue can still be updated
 713          * with resend requests. Move final list to separate one for processing
 714          */
 715         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
 716         list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
 717         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
 718
 719         /**
 720          * go through list of exports participated in recovery and
 721          * set new epoch for them
 722          */
 723         list_for_each_entry(req, &client_list, rq_list) {
 724                 LASSERT(!req->rq_export->exp_delayed);
 725                 if (!req->rq_export->exp_vbr_failed)
 726                         tgt_client_epoch_update(&env, req->rq_export);
 727         }
 728         /** return list back at once */
 729         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
 730         list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
 731         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
 732
 733         /** Clear MULTI RPCS incompatibility flag if
 734          * - target is MDT and
 735          * - there is no client to recover or the recovery was aborted
 736          */
 737         if (!strncmp(tgt->lut_obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) &&
 738             (tgt->lut_obd->obd_max_recoverable_clients == 0 ||
 739             tgt->lut_obd->obd_abort_recovery))
 740                 tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
 741
 742         /** update server epoch */
 743         tgt_server_data_update(&env, tgt, 1);
 744         lu_env_fini(&env);
 745 }
 746
 747 /**
 748  * commit callback, need to update last_commited value
 749  */
 750 struct tgt_last_committed_callback {
 751         struct dt_txn_commit_cb  llcc_cb;
 752         struct lu_target        *llcc_tgt;
 753         struct obd_export       *llcc_exp;
 754         __u64                    llcc_transno;
 755 };
 756
 757 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
 758                                   struct dt_txn_commit_cb *cb, int err)
 759 {
 760         struct tgt_last_committed_callback *ccb;
 761
 762         ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
 763
 764         LASSERT(ccb->llcc_tgt != NULL);
 765         LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
 766
 767         spin_lock(&ccb->llcc_tgt->lut_translock);
 768         if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
 769                 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
 770
 771         LASSERT(ccb->llcc_exp);
 772         if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
 773                 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
 774                 spin_unlock(&ccb->llcc_tgt->lut_translock);
 775                 ptlrpc_commit_replies(ccb->llcc_exp);
 776         } else {
 777                 spin_unlock(&ccb->llcc_tgt->lut_translock);
 778         }
 779         class_export_cb_put(ccb->llcc_exp);
 780         if (ccb->llcc_transno)
 781                 CDEBUG(D_HA, "%s: transno "LPD64" is committed\n",
 782                        ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
 783         OBD_FREE_PTR(ccb);
 784 }
 785
 786 int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
 787                            struct obd_export *exp, __u64 transno)
 788 {
 789         struct tgt_last_committed_callback      *ccb;
 790         struct dt_txn_commit_cb                 *dcb;
 791         int                                      rc;
 792
 793         OBD_ALLOC_PTR(ccb);
 794         if (ccb == NULL)
 795                 return -ENOMEM;
 796
 797         ccb->llcc_tgt = tgt;
 798         ccb->llcc_exp = class_export_cb_get(exp);
 799         ccb->llcc_transno = transno;
 800
 801         dcb = &ccb->llcc_cb;
 802         dcb->dcb_func = tgt_cb_last_committed;
 803         INIT_LIST_HEAD(&dcb->dcb_linkage);
 804         strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
 805
 806         rc = dt_trans_cb_add(th, dcb);
 807         if (rc) {
 808                 class_export_cb_put(exp);
 809                 OBD_FREE_PTR(ccb);
 810         }
 811
 812         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 813                 /* report failure to force synchronous operation */
 814                 return -EPERM;
 815
 816         return rc;
 817 }
 818
 819 struct tgt_new_client_callback {
 820         struct dt_txn_commit_cb  lncc_cb;
 821         struct obd_export       *lncc_exp;
 822 };
 823
 824 static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
 825                               struct dt_txn_commit_cb *cb, int err)
 826 {
 827         struct tgt_new_client_callback *ccb;
 828
 829         ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
 830
 831         LASSERT(ccb->lncc_exp->exp_obd);
 832
 833         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 834                ccb->lncc_exp->exp_obd->obd_name,
 835                ccb->lncc_exp->exp_client_uuid.uuid);
 836
 837         spin_lock(&ccb->lncc_exp->exp_lock);
 838
 839         ccb->lncc_exp->exp_need_sync = 0;
 840
 841         spin_unlock(&ccb->lncc_exp->exp_lock);
 842         class_export_cb_put(ccb->lncc_exp);
 843
 844         OBD_FREE_PTR(ccb);
 845 }
 846
 847 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
 848 {
 849         struct tgt_new_client_callback  *ccb;
 850         struct dt_txn_commit_cb         *dcb;
 851         int                              rc;
 852
 853         OBD_ALLOC_PTR(ccb);
 854         if (ccb == NULL)
 855                 return -ENOMEM;
 856
 857         ccb->lncc_exp = class_export_cb_get(exp);
 858
 859         dcb = &ccb->lncc_cb;
 860         dcb->dcb_func = tgt_cb_new_client;
 861         INIT_LIST_HEAD(&dcb->dcb_linkage);
 862         strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
 863
 864         rc = dt_trans_cb_add(th, dcb);
 865         if (rc) {
 866                 class_export_cb_put(exp);
 867                 OBD_FREE_PTR(ccb);
 868         }
 869         return rc;
 870 }
 871
 872 /**
 873  * Add new client to the last_rcvd upon new connection.
 874  *
 875  * We use a bitmap to locate a free space in the last_rcvd file and initialize
 876  * tg_export_data.
 877  */
 878 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
 879 {
 880         struct tg_export_data   *ted = &exp->exp_target_data;
 881         struct lu_target        *tgt = class_exp2tgt(exp);
 882         int                      rc = 0, idx;
 883
 884         ENTRY;
 885
 886         LASSERT(tgt && tgt->lut_client_bitmap != NULL);
 887         if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
 888                 RETURN(0);
 889
 890         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 891                 RETURN(0);
 892
 893         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 894          * there's no need for extra complication here
 895          */
 896         idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
 897 repeat:
 898         if (idx >= LR_MAX_CLIENTS ||
 899             OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
 900                 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
 901                        tgt->lut_obd->obd_name,  idx);
 902                 RETURN(-EOVERFLOW);
 903         }
 904         if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
 905                 idx = find_next_zero_bit(tgt->lut_client_bitmap,
 906                                              LR_MAX_CLIENTS, idx);
 907                 goto repeat;
 908         }
 909
 910         ted->ted_lr_idx = idx;
 911         ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
 912                           idx * tgt->lut_lsd.lsd_client_size;
 913
 914         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
 915
 916         if (tgt_is_multimodrpcs_client(exp)) {
 917                 /* Set MULTI RPCS incompatibility flag to prevent previous
 918                  * Lustre versions to mount a target with reply_data file */
 919                 atomic_inc(&tgt->lut_num_clients);
 920                 if (!(tgt->lut_lsd.lsd_feature_incompat &
 921                       OBD_INCOMPAT_MULTI_RPCS)) {
 922                         tgt->lut_lsd.lsd_feature_incompat |=
 923                                                         OBD_INCOMPAT_MULTI_RPCS;
 924                         rc = tgt_server_data_update(env, tgt, 1);
 925                         if (rc < 0) {
 926                                 CERROR("%s: unable to set MULTI RPCS "
 927                                        "incompatibility flag\n",
 928                                        exp->exp_obd->obd_name);
 929                                 RETURN(rc);
 930                         }
 931                 }
 932
 933                 /* assign client slot generation */
 934                 ted->ted_lcd->lcd_generation =
 935                                 atomic_inc_return(&tgt->lut_client_generation);
 936         } else {
 937                 ted->ted_lcd->lcd_generation = 0;
 938         }
 939
 940         CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s' "
 941                "generation %d\n",
 942                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
 943                ted->ted_lcd->lcd_uuid, ted->ted_lcd->lcd_generation);
 944
 945         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
 946                 RETURN(-ENOSPC);
 947
 948         rc = tgt_client_data_update(env, exp);
 949         if (rc)
 950                 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
 951                        tgt->lut_obd->obd_name, idx, rc);
 952
 953         RETURN(rc);
 954 }
 955 EXPORT_SYMBOL(tgt_client_new);
 956
 957 /* Add an existing client to the MDS in-memory state based on
 958  * a client that was previously found in the last_rcvd file and
 959  * already has an assigned slot (idx >= 0).
 960  *
 961  * It should not be possible to fail adding an existing client - otherwise
 962  * mdt_init_server_data() callsite needs to be fixed.
 963  */
 964 int tgt_client_add(const struct lu_env *env,  struct obd_export *exp, int idx)
 965 {
 966         struct tg_export_data   *ted = &exp->exp_target_data;
 967         struct lu_target        *tgt = class_exp2tgt(exp);
 968
 969         ENTRY;
 970
 971         LASSERT(tgt && tgt->lut_client_bitmap != NULL);
 972         LASSERTF(idx >= 0, "%d\n", idx);
 973
 974         if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
 975             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 976                 RETURN(0);
 977
 978         if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
 979                 CERROR("%s: client %d: bit already set in bitmap!!\n",
 980                        tgt->lut_obd->obd_name,  idx);
 981                 LBUG();
 982         }
 983         atomic_inc(&tgt->lut_num_clients);
 984
 985         CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added, "
 986                "generation %d\n",
 987                tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid,
 988                ted->ted_lcd->lcd_generation);
 989
 990         ted->ted_lr_idx = idx;
 991         ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
 992                           idx * tgt->lut_lsd.lsd_client_size;
 993
 994         mutex_init(&ted->ted_lcd_lock);
 995
 996         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
 997
 998         RETURN(0);
 999 }
1000
1001 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
1002 {
1003         struct tg_export_data   *ted = &exp->exp_target_data;
1004         struct lu_target        *tgt = class_exp2tgt(exp);
1005         int                      rc;
1006
1007         ENTRY;
1008
1009         LASSERT(ted->ted_lcd);
1010
1011         if (unlikely(tgt == NULL)) {
1012                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1013                        class_exp2obd(exp)->obd_name);
1014                 RETURN(-EINVAL);
1015         }
1016
1017         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
1018         if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
1019                     (char *)tgt->lut_obd->obd_uuid.uuid) ||
1020             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1021                 RETURN(0);
1022
1023         CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
1024                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1025                ted->ted_lcd->lcd_uuid);
1026
1027         /* Clear the bit _after_ zeroing out the client so we don't
1028            race with filter_client_add and zero out new clients.*/
1029         if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
1030                 CERROR("%s: client %u: bit already clear in bitmap!!\n",
1031                        tgt->lut_obd->obd_name, ted->ted_lr_idx);
1032                 LBUG();
1033         }
1034
1035         /* Do not erase record for recoverable client. */
1036         if (exp->exp_flags & OBD_OPT_FAILOVER)
1037                 RETURN(0);
1038
1039         /* Make sure the server's last_transno is up to date.
1040          * This should be done before zeroing client slot so last_transno will
1041          * be in server data or in client data in case of failure */
1042         rc = tgt_server_data_update(env, tgt, 0);
1043         if (rc != 0) {
1044                 CERROR("%s: failed to update server data, skip client %s "
1045                        "zeroing, rc %d\n", tgt->lut_obd->obd_name,
1046                        ted->ted_lcd->lcd_uuid, rc);
1047                 RETURN(rc);
1048         }
1049
1050         mutex_lock(&ted->ted_lcd_lock);
1051         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
1052         rc = tgt_client_data_update(env, exp);
1053         mutex_unlock(&ted->ted_lcd_lock);
1054
1055         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
1056                "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
1057                tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
1058                ted->ted_lr_idx, ted->ted_lr_off, rc);
1059         RETURN(rc);
1060 }
1061 EXPORT_SYMBOL(tgt_client_del);
1062
1063 int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
1064                        struct tg_export_data *ted, struct tg_reply_data *trd,
1065                        struct thandle *th, bool update_lrd_file)
1066 {
1067         struct lsd_reply_data   *lrd;
1068         int     i;
1069
1070         lrd = &trd->trd_reply;
1071         /* update export last transno */
1072         mutex_lock(&ted->ted_lcd_lock);
1073         if (lrd->lrd_transno > ted->ted_lcd->lcd_last_transno)
1074                 ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
1075         mutex_unlock(&ted->ted_lcd_lock);
1076
1077         /* find a empty slot */
1078         i = tgt_find_free_reply_slot(tgt);
1079         if (unlikely(i < 0)) {
1080                 CERROR("%s: couldn't find a slot for reply data: "
1081                        "rc = %d\n", tgt_name(tgt), i);
1082                 RETURN(i);
1083         }
1084         trd->trd_index = i;
1085
1086         if (update_lrd_file) {
1087                 loff_t  off;
1088                 int     rc;
1089
1090                 /* write reply data to disk */
1091                 off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
1092                 rc = tgt_reply_data_write(env, tgt, lrd, off, th);
1093                 if (unlikely(rc != 0)) {
1094                         CERROR("%s: can't update %s file: rc = %d\n",
1095                                tgt_name(tgt), REPLY_DATA, rc);
1096                         RETURN(rc);
1097                 }
1098         }
1099         /* add reply data to target export's reply list */
1100         mutex_lock(&ted->ted_lcd_lock);
1101         list_add(&trd->trd_list, &ted->ted_reply_list);
1102         ted->ted_reply_cnt++;
1103         if (ted->ted_reply_cnt > ted->ted_reply_max)
1104                 ted->ted_reply_max = ted->ted_reply_cnt;
1105         mutex_unlock(&ted->ted_lcd_lock);
1106
1107         CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
1108                "tag %hu, client gen %u, slot idx %d\n",
1109                trd, lrd->lrd_xid, lrd->lrd_transno,
1110                trd->trd_tag, lrd->lrd_client_gen, i);
1111         RETURN(0);
1112 }
1113 EXPORT_SYMBOL(tgt_add_reply_data);
1114
1115 /*
1116  * last_rcvd & last_committed update callbacks
1117  */
1118 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
1119                                 struct dt_object *obj, __u64 opdata,
1120                                 struct thandle *th, struct ptlrpc_request *req)
1121 {
1122         struct tgt_thread_info  *tti = tgt_th_info(env);
1123         struct tgt_session_info *tsi = tgt_ses_info(env);
1124         struct obd_export       *exp = tsi->tsi_exp;
1125         struct tg_export_data   *ted;
1126         __u64                   *transno_p;
1127         int                      rc = 0;
1128         bool                     lw_client, update = false;
1129
1130         ENTRY;
1131
1132
1133         LASSERT(exp != NULL);
1134         ted = &exp->exp_target_data;
1135
1136         lw_client = exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT;
1137         if (ted->ted_lr_idx < 0 && !lw_client)
1138                 /* ofd connect may cause transaction before export has
1139                  * last_rcvd slot */
1140                 RETURN(0);
1141
1142         if (req != NULL)
1143                 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
1144         else
1145                 /* From update replay, tti_transno should be set already */
1146                 LASSERT(tti->tti_transno != 0);
1147
1148         spin_lock(&tgt->lut_translock);
1149         if (th->th_result != 0) {
1150                 if (tti->tti_transno != 0) {
1151                         CERROR("%s: replay transno "LPU64" failed: rc = %d\n",
1152                                tgt_name(tgt), tti->tti_transno, th->th_result);
1153                 }
1154         } else if (tti->tti_transno == 0) {
1155                 tti->tti_transno = ++tgt->lut_last_transno;
1156         } else {
1157                 /* should be replay */
1158                 if (tti->tti_transno > tgt->lut_last_transno)
1159                         tgt->lut_last_transno = tti->tti_transno;
1160         }
1161         spin_unlock(&tgt->lut_translock);
1162
1163         /** VBR: set new versions */
1164         if (th->th_result == 0 && obj != NULL) {
1165                 struct dt_object *dto = dt_object_locate(obj, th->th_dev);
1166                 dt_version_set(env, dto, tti->tti_transno, th);
1167         }
1168
1169         /* filling reply data */
1170         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
1171                tti->tti_transno, tgt->lut_obd->obd_last_committed);
1172
1173         if (req != NULL) {
1174                 req->rq_transno = tti->tti_transno;
1175                 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
1176         }
1177
1178         /* if can't add callback, do sync write */
1179         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp, tti->tti_transno);
1180
1181         if (lw_client) {
1182                 /* All operations performed by LW clients are synchronous and
1183                  * we store the committed transno in the last_rcvd header */
1184                 spin_lock(&tgt->lut_translock);
1185                 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
1186                         tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
1187                         update = true;
1188                 }
1189                 spin_unlock(&tgt->lut_translock);
1190                 /* Although lightweight (LW) connections have no slot in
1191                  * last_rcvd, we still want to maintain the in-memory
1192                  * lsd_client_data structure in order to properly handle reply
1193                  * reconstruction. */
1194         } else if (ted->ted_lr_off == 0) {
1195                 CERROR("%s: client idx %d has offset %lld\n",
1196                        tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
1197                 RETURN(-EINVAL);
1198         }
1199
1200         /* if the export has already been disconnected, we have no last_rcvd
1201          * slot, update server data with latest transno then */
1202         if (ted->ted_lcd == NULL) {
1203                 CWARN("commit transaction for disconnected client %s: rc %d\n",
1204                       exp->exp_client_uuid.uuid, rc);
1205                 GOTO(srv_update, rc = 0);
1206         }
1207
1208         /* Target that supports multiple reply data */
1209         if (tgt_is_multimodrpcs_client(exp)) {
1210                 struct tg_reply_data    *trd;
1211                 struct lsd_reply_data   *lrd;
1212                 __u64                   *pre_versions;
1213                 bool                    write_update;
1214
1215                 OBD_ALLOC_PTR(trd);
1216                 if (unlikely(trd == NULL))
1217                         GOTO(srv_update, rc = -ENOMEM);
1218
1219                 /* fill reply data information */
1220                 lrd = &trd->trd_reply;
1221                 lrd->lrd_transno = tti->tti_transno;
1222                 if (req != NULL) {
1223                         lrd->lrd_xid = req->rq_xid;
1224                         trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
1225                         pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1226                         lrd->lrd_result = th->th_result;
1227                         lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
1228                         write_update = true;
1229                 } else {
1230                         LASSERT(tsi->tsi_xid != 0);
1231                         lrd->lrd_xid = tsi->tsi_xid;
1232                         lrd->lrd_result = tsi->tsi_result;
1233                         lrd->lrd_client_gen = tsi->tsi_client_gen;
1234                         trd->trd_tag = 0;
1235                         pre_versions = NULL;
1236                         write_update = false;
1237                 }
1238
1239                 lrd->lrd_data = opdata;
1240                 if (pre_versions) {
1241                         trd->trd_pre_versions[0] = pre_versions[0];
1242                         trd->trd_pre_versions[1] = pre_versions[1];
1243                         trd->trd_pre_versions[2] = pre_versions[2];
1244                         trd->trd_pre_versions[3] = pre_versions[3];
1245                 }
1246
1247                 rc = tgt_add_reply_data(env, tgt, ted, trd, th, write_update);
1248                 GOTO(srv_update, rc);
1249         }
1250
1251         /* Enough for update replay, let's return */
1252         if (req == NULL)
1253                 GOTO(srv_update, rc);
1254
1255         mutex_lock(&ted->ted_lcd_lock);
1256         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1257         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
1258                 transno_p = &ted->ted_lcd->lcd_last_close_transno;
1259                 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
1260                 ted->ted_lcd->lcd_last_close_result = th->th_result;
1261         } else {
1262                 /* VBR: save versions in last_rcvd for reconstruct. */
1263                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1264
1265                 if (pre_versions) {
1266                         ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
1267                         ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
1268                         ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
1269                         ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
1270                 }
1271                 transno_p = &ted->ted_lcd->lcd_last_transno;
1272                 ted->ted_lcd->lcd_last_xid = req->rq_xid;
1273                 ted->ted_lcd->lcd_last_result = th->th_result;
1274                 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
1275                  * see struct ldlm_reply->lock_policy_res1; */
1276                 ted->ted_lcd->lcd_last_data = opdata;
1277         }
1278
1279         /* Update transno in slot only if non-zero number, i.e. no errors */
1280         if (likely(tti->tti_transno != 0)) {
1281                 if (*transno_p > tti->tti_transno &&
1282                     !tgt->lut_no_reconstruct) {
1283                         CERROR("%s: trying to overwrite bigger transno:"
1284                                "on-disk: "LPU64", new: "LPU64" replay: %d. "
1285                                "see LU-617.\n", tgt_name(tgt), *transno_p,
1286                                tti->tti_transno, req_is_replay(req));
1287                         if (req_is_replay(req)) {
1288                                 spin_lock(&req->rq_export->exp_lock);
1289                                 req->rq_export->exp_vbr_failed = 1;
1290                                 spin_unlock(&req->rq_export->exp_lock);
1291                         }
1292                         mutex_unlock(&ted->ted_lcd_lock);
1293                         RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
1294                 }
1295                 *transno_p = tti->tti_transno;
1296         }
1297
1298         if (!lw_client) {
1299                 tti->tti_off = ted->ted_lr_off;
1300                 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1301                 if (rc < 0) {
1302                         mutex_unlock(&ted->ted_lcd_lock);
1303                         RETURN(rc);
1304                 }
1305         }
1306         mutex_unlock(&ted->ted_lcd_lock);
1307         EXIT;
1308 srv_update:
1309         if (update)
1310                 rc = tgt_server_data_write(env, tgt, th);
1311         return rc;
1312 }
1313
1314 /*
1315  * last_rcvd update for echo client simulation.
1316  * It updates last_rcvd client slot and version of object in
1317  * simple way but with all locks to simulate all drawbacks
1318  */
1319 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
1320                                      struct lu_target *tgt,
1321                                      struct dt_object *obj,
1322                                      struct thandle *th,
1323                                      struct obd_export *exp)
1324 {
1325         struct tgt_thread_info  *tti = tgt_th_info(env);
1326         struct tg_export_data   *ted = &exp->exp_target_data;
1327         int                      rc = 0;
1328
1329         ENTRY;
1330
1331         tti->tti_transno = 0;
1332
1333         spin_lock(&tgt->lut_translock);
1334         if (th->th_result == 0)
1335                 tti->tti_transno = ++tgt->lut_last_transno;
1336         spin_unlock(&tgt->lut_translock);
1337
1338         /** VBR: set new versions */
1339         if (th->th_result == 0 && obj != NULL)
1340                 dt_version_set(env, obj, tti->tti_transno, th);
1341
1342         /* if can't add callback, do sync write */
1343         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
1344                                                 tti->tti_transno);
1345
1346         LASSERT(ted->ted_lr_off > 0);
1347
1348         mutex_lock(&ted->ted_lcd_lock);
1349         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1350         ted->ted_lcd->lcd_last_transno = tti->tti_transno;
1351         ted->ted_lcd->lcd_last_result = th->th_result;
1352
1353         tti->tti_off = ted->ted_lr_off;
1354         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1355         mutex_unlock(&ted->ted_lcd_lock);
1356         RETURN(rc);
1357 }
1358
1359 static int tgt_clients_data_init(const struct lu_env *env,
1360                                  struct lu_target *tgt,
1361                                  unsigned long last_size)
1362 {
1363         struct obd_device       *obd = tgt->lut_obd;
1364         struct lr_server_data   *lsd = &tgt->lut_lsd;
1365         struct lsd_client_data  *lcd = NULL;
1366         struct tg_export_data   *ted;
1367         int                      cl_idx;
1368         int                      rc = 0;
1369         loff_t                   off = lsd->lsd_client_start;
1370         __u32                    generation = 0;
1371         struct cfs_hash         *hash = NULL;
1372
1373         ENTRY;
1374
1375         CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
1376                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
1377
1378         OBD_ALLOC_PTR(lcd);
1379         if (lcd == NULL)
1380                 RETURN(-ENOMEM);
1381
1382         hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1383         if (hash == NULL)
1384                 GOTO(err_out, rc = -ENODEV);
1385
1386         for (cl_idx = 0; off < last_size; cl_idx++) {
1387                 struct obd_export       *exp;
1388                 __u64                    last_transno;
1389
1390                 /* Don't assume off is incremented properly by
1391                  * read_record(), in case sizeof(*lcd)
1392                  * isn't the same as fsd->lsd_client_size.  */
1393                 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
1394                 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
1395                 if (rc) {
1396                         CERROR("%s: error reading last_rcvd %s idx %d off "
1397                                "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
1398                                cl_idx, off, rc);
1399                         rc = 0;
1400                         break; /* read error shouldn't cause startup to fail */
1401                 }
1402
1403                 if (lcd->lcd_uuid[0] == '\0') {
1404                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
1405                                cl_idx);
1406                         continue;
1407                 }
1408
1409                 last_transno = lcd_last_transno(lcd);
1410
1411                 /* These exports are cleaned up by disconnect, so they
1412                  * need to be set up like real exports as connect does.
1413                  */
1414                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
1415                        " srv lr: "LPU64" lx: "LPU64" gen %u\n", lcd->lcd_uuid,
1416                        cl_idx, last_transno, lsd->lsd_last_transno,
1417                        lcd_last_xid(lcd), lcd->lcd_generation);
1418
1419                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
1420                 if (IS_ERR(exp)) {
1421                         if (PTR_ERR(exp) == -EALREADY) {
1422                                 /* export already exists, zero out this one */
1423                                 CERROR("%s: Duplicate export %s!\n",
1424                                        tgt_name(tgt), lcd->lcd_uuid);
1425                                 continue;
1426                         }
1427                         GOTO(err_out, rc = PTR_ERR(exp));
1428                 }
1429
1430                 ted = &exp->exp_target_data;
1431                 *ted->ted_lcd = *lcd;
1432
1433                 rc = tgt_client_add(env, exp, cl_idx);
1434                 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
1435                 /* VBR: set export last committed version */
1436                 exp->exp_last_committed = last_transno;
1437                 spin_lock(&exp->exp_lock);
1438                 exp->exp_connecting = 0;
1439                 exp->exp_in_recovery = 0;
1440                 spin_unlock(&exp->exp_lock);
1441                 obd->obd_max_recoverable_clients++;
1442
1443                 if (tgt->lut_lsd.lsd_feature_incompat &
1444                     OBD_INCOMPAT_MULTI_RPCS &&
1445                     lcd->lcd_generation != 0) {
1446                         /* compute the highest valid client generation */
1447                         generation = max(generation, lcd->lcd_generation);
1448                         /* fill client_generation <-> export hash table */
1449                         rc = cfs_hash_add_unique(hash, &lcd->lcd_generation,
1450                                                  &exp->exp_gen_hash);
1451                         if (rc != 0) {
1452                                 CERROR("%s: duplicate export for client "
1453                                        "generation %u\n",
1454                                        tgt_name(tgt), lcd->lcd_generation);
1455                                 class_export_put(exp);
1456                                 GOTO(err_out, rc);
1457                         }
1458                 }
1459
1460                 class_export_put(exp);
1461
1462                 /* Need to check last_rcvd even for duplicated exports. */
1463                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
1464                        cl_idx, last_transno);
1465
1466                 spin_lock(&tgt->lut_translock);
1467                 tgt->lut_last_transno = max(last_transno,
1468                                             tgt->lut_last_transno);
1469                 spin_unlock(&tgt->lut_translock);
1470         }
1471
1472         /* record highest valid client generation */
1473         atomic_set(&tgt->lut_client_generation, generation);
1474
1475 err_out:
1476         if (hash != NULL)
1477                 cfs_hash_putref(hash);
1478         OBD_FREE_PTR(lcd);
1479         RETURN(rc);
1480 }
1481
1482 struct server_compat_data {
1483         __u32 rocompat;
1484         __u32 incompat;
1485         __u32 rocinit;
1486         __u32 incinit;
1487 };
1488
1489 static struct server_compat_data tgt_scd[] = {
1490         [LDD_F_SV_TYPE_MDT] = {
1491                 .rocompat = OBD_ROCOMPAT_LOVOBJID,
1492                 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1493                             OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
1494                             OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI |
1495                             OBD_INCOMPAT_MULTI_RPCS,
1496                 .rocinit = OBD_ROCOMPAT_LOVOBJID,
1497                 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1498                            OBD_INCOMPAT_MULTI_OI,
1499         },
1500         [LDD_F_SV_TYPE_OST] = {
1501                 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
1502                 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
1503                             OBD_INCOMPAT_FID,
1504                 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
1505                 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1506         }
1507 };
1508
1509 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1510 {
1511         struct tgt_thread_info          *tti = tgt_th_info(env);
1512         struct lr_server_data           *lsd = &tgt->lut_lsd;
1513         unsigned long                    last_rcvd_size;
1514         __u32                            index;
1515         int                              rc, type;
1516
1517         rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1518         if (rc)
1519                 RETURN(rc);
1520
1521         last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1522
1523         /* ensure padding in the struct is the correct size */
1524         CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
1525                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
1526
1527         rc = server_name2index(tgt_name(tgt), &index, NULL);
1528         if (rc < 0) {
1529                 CERROR("%s: Can not get index from name: rc = %d\n",
1530                        tgt_name(tgt), rc);
1531                 RETURN(rc);
1532         }
1533         /* server_name2index() returns type */
1534         type = rc;
1535         if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1536                 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1537                 RETURN(-EINVAL);
1538         }
1539
1540         /* last_rcvd on OST doesn't provide reconstruct support because there
1541          * may be up to 8 in-flight write requests per single slot in
1542          * last_rcvd client data
1543          */
1544         tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1545
1546         if (last_rcvd_size == 0) {
1547                 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1548
1549                 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1550                        sizeof(lsd->lsd_uuid));
1551                 lsd->lsd_last_transno = 0;
1552                 lsd->lsd_mount_count = 0;
1553                 lsd->lsd_server_size = LR_SERVER_SIZE;
1554                 lsd->lsd_client_start = LR_CLIENT_START;
1555                 lsd->lsd_client_size = LR_CLIENT_SIZE;
1556                 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1557                 lsd->lsd_osd_index = index;
1558                 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1559                 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1560         } else {
1561                 rc = tgt_server_data_read(env, tgt);
1562                 if (rc) {
1563                         CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1564                                tgt_name(tgt), rc);
1565                         RETURN(rc);
1566                 }
1567                 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1568                         LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
1569                                            "using the wrong disk %s. Were the"
1570                                            " /dev/ assignments rearranged?\n",
1571                                            tgt->lut_obd->obd_uuid.uuid,
1572                                            lsd->lsd_uuid);
1573                         RETURN(-EINVAL);
1574                 }
1575
1576                 if (lsd->lsd_osd_index != index) {
1577                         LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
1578                                            "is different with the index %d in"
1579                                            "config log, It might be disk"
1580                                            "corruption!\n", tgt_name(tgt),
1581                                            lsd->lsd_osd_index, index);
1582                         RETURN(-EINVAL);
1583                 }
1584         }
1585
1586         if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1587                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1588                        tgt_name(tgt),
1589                        lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1590                 RETURN(-EINVAL);
1591         }
1592
1593         if (type == LDD_F_SV_TYPE_MDT)
1594                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1595
1596         if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1597                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1598                        tgt_name(tgt),
1599                        lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1600                 RETURN(-EINVAL);
1601         }
1602         /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1603         if (type == LDD_F_SV_TYPE_MDT &&
1604             !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1605                 if (last_rcvd_size > lsd->lsd_client_start) {
1606                         LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1607                                       "remove all clients for interop needs\n",
1608                                       tgt_name(tgt));
1609                         rc = tgt_truncate_last_rcvd(env, tgt,
1610                                                     lsd->lsd_client_start);
1611                         if (rc)
1612                                 RETURN(rc);
1613                         last_rcvd_size = lsd->lsd_client_start;
1614                 }
1615                 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1616                 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1617         }
1618
1619         spin_lock(&tgt->lut_translock);
1620         tgt->lut_last_transno = lsd->lsd_last_transno;
1621         spin_unlock(&tgt->lut_translock);
1622
1623         lsd->lsd_mount_count++;
1624
1625         CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1626         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
1627                tgt_name(tgt), tgt->lut_last_transno);
1628         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
1629                tgt_name(tgt), lsd->lsd_mount_count);
1630         CDEBUG(D_INODE, "%s: server data size: %u\n",
1631                tgt_name(tgt), lsd->lsd_server_size);
1632         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1633                tgt_name(tgt), lsd->lsd_client_start);
1634         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1635                tgt_name(tgt), lsd->lsd_client_size);
1636         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1637                tgt_name(tgt), last_rcvd_size);
1638         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1639                tgt_name(tgt), lsd->lsd_subdir_count);
1640         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1641                last_rcvd_size <= lsd->lsd_client_start ? 0 :
1642                (last_rcvd_size - lsd->lsd_client_start) /
1643                 lsd->lsd_client_size);
1644         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1645
1646         if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1647             lsd->lsd_client_size == 0) {
1648                 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1649                 RETURN(-EINVAL);
1650         }
1651
1652         if (!tgt->lut_obd->obd_replayable)
1653                 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1654
1655         rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1656         if (rc < 0)
1657                 GOTO(err_client, rc);
1658
1659         spin_lock(&tgt->lut_translock);
1660         /* obd_last_committed is used for compatibility
1661          * with other lustre recovery code */
1662         tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1663         spin_unlock(&tgt->lut_translock);
1664
1665         tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1666         tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1667
1668         /* save it, so mount count and last_transno is current */
1669         rc = tgt_server_data_update(env, tgt, 0);
1670         if (rc < 0)
1671                 GOTO(err_client, rc);
1672
1673         RETURN(0);
1674
1675 err_client:
1676         class_disconnect_exports(tgt->lut_obd);
1677         return rc;
1678 }
1679
1680 /* add credits for last_rcvd update */
1681 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1682                      void *cookie)
1683 {
1684         struct lu_target        *tgt = cookie;
1685         struct tgt_session_info *tsi;
1686         struct tgt_thread_info  *tti = tgt_th_info(env);
1687         struct dt_object        *dto;
1688         int                      rc;
1689
1690         /* if there is no session, then this transaction is not result of
1691          * request processing but some local operation */
1692         if (env->le_ses == NULL)
1693                 return 0;
1694
1695         LASSERT(tgt->lut_last_rcvd);
1696         tsi = tgt_ses_info(env);
1697         /* OFD may start transaction without export assigned */
1698         if (tsi->tsi_exp == NULL)
1699                 return 0;
1700
1701         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
1702         tti_buf_lcd(tti);
1703
1704         rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1705                                      tsi->tsi_exp->exp_target_data.ted_lr_off,
1706                                      th);
1707         if (rc)
1708                 return rc;
1709
1710         tti_buf_lsd(tti);
1711         rc = dt_declare_record_write(env, dto, &tti->tti_buf, 0, th);
1712         if (rc)
1713                 return rc;
1714
1715         if (tsi->tsi_vbr_obj != NULL &&
1716             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1717                 dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
1718                 rc = dt_declare_version_set(env, dto, th);
1719         }
1720
1721         return rc;
1722 }
1723
1724 /* Update last_rcvd records with latests transaction data */
1725 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1726                     void *cookie)
1727 {
1728         struct lu_target        *tgt = cookie;
1729         struct tgt_session_info *tsi;
1730         struct tgt_thread_info  *tti = tgt_th_info(env);
1731         struct dt_object        *obj = NULL;
1732         int                      rc;
1733         bool                     echo_client;
1734
1735         if (env->le_ses == NULL)
1736                 return 0;
1737
1738         tsi = tgt_ses_info(env);
1739         /* OFD may start transaction without export assigned */
1740         if (tsi->tsi_exp == NULL)
1741                 return 0;
1742
1743         echo_client = (tgt_ses_req(tsi) == NULL && tsi->tsi_xid == 0);
1744
1745         if (tti->tti_has_trans && !echo_client) {
1746                 if (tti->tti_mult_trans == 0) {
1747                         CDEBUG(D_HA, "More than one transaction "LPU64"\n",
1748                                tti->tti_transno);
1749                         RETURN(0);
1750                 }
1751                 /* we need another transno to be assigned */
1752                 tti->tti_transno = 0;
1753         } else if (th->th_result == 0) {
1754                 tti->tti_has_trans = 1;
1755         }
1756
1757         if (tsi->tsi_vbr_obj != NULL &&
1758             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1759                 obj = tsi->tsi_vbr_obj;
1760         }
1761
1762         if (unlikely(echo_client)) /* echo client special case */
1763                 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1764                                                tsi->tsi_exp);
1765         else
1766                 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
1767                                           tgt_ses_req(tsi));
1768         return rc;
1769 }
1770
1771 int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
1772 {
1773         struct tgt_thread_info  *tti = tgt_th_info(env);
1774         struct lsd_reply_data   *lrd = &tti->tti_lrd;
1775         unsigned long            reply_data_size;
1776         int                      rc;
1777         struct lsd_reply_header *lrh = NULL;
1778         struct lsd_client_data  *lcd = NULL;
1779         struct tg_reply_data    *trd = NULL;
1780         int                      idx;
1781         loff_t                   off;
1782         struct cfs_hash         *hash = NULL;
1783         struct obd_export       *exp;
1784         struct tg_export_data   *ted;
1785         int                      reply_data_recovered = 0;
1786
1787         rc = dt_attr_get(env, tgt->lut_reply_data, &tti->tti_attr);
1788         if (rc)
1789                 GOTO(out, rc);
1790         reply_data_size = (unsigned long)tti->tti_attr.la_size;
1791
1792         OBD_ALLOC_PTR(lrh);
1793         if (lrh == NULL)
1794                 GOTO(out, rc = -ENOMEM);
1795
1796         if (reply_data_size == 0) {
1797                 CDEBUG(D_INFO, "%s: new reply_data file, initializing\n",
1798                        tgt_name(tgt));
1799                 lrh->lrh_magic = LRH_MAGIC;
1800                 lrh->lrh_header_size = sizeof(struct lsd_reply_header);
1801                 lrh->lrh_reply_size = sizeof(struct lsd_reply_data);
1802                 rc = tgt_reply_header_write(env, tgt, lrh);
1803                 if (rc) {
1804                         CERROR("%s: error writing %s: rc = %d\n",
1805                                tgt_name(tgt), REPLY_DATA, rc);
1806                         GOTO(out, rc);
1807                 }
1808         } else {
1809                 rc = tgt_reply_header_read(env, tgt, lrh);
1810                 if (rc) {
1811                         CERROR("%s: error reading %s: rc = %d\n",
1812                                tgt_name(tgt), REPLY_DATA, rc);
1813                         GOTO(out, rc);
1814                 }
1815                 if (lrh->lrh_magic != LRH_MAGIC ||
1816                     lrh->lrh_header_size != sizeof(struct lsd_reply_header) ||
1817                     lrh->lrh_reply_size != sizeof(struct lsd_reply_data)) {
1818                         CERROR("%s: invalid header in %s\n",
1819                                tgt_name(tgt), REPLY_DATA);
1820                         GOTO(out, rc = -EINVAL);
1821                 }
1822
1823                 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1824                 if (hash == NULL)
1825                         GOTO(out, rc = -ENODEV);
1826
1827                 OBD_ALLOC_PTR(lcd);
1828                 if (lcd == NULL)
1829                         GOTO(out, rc = -ENOMEM);
1830
1831                 OBD_ALLOC_PTR(trd);
1832                 if (trd == NULL)
1833                         GOTO(out, rc = -ENOMEM);
1834
1835                 /* Load reply_data from disk */
1836                 for (idx = 0, off = sizeof(struct lsd_reply_header);
1837                      off < reply_data_size;
1838                      idx++, off += sizeof(struct lsd_reply_data)) {
1839                         rc = tgt_reply_data_read(env, tgt, lrd, off);
1840                         if (rc) {
1841                                 CERROR("%s: error reading %s: rc = %d\n",
1842                                        tgt_name(tgt), REPLY_DATA, rc);
1843                                 GOTO(out, rc);
1844                         }
1845
1846                         exp = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
1847                         if (exp == NULL) {
1848                                 /* old reply data from a disconnected client */
1849                                 continue;
1850                         }
1851                         ted = &exp->exp_target_data;
1852                         mutex_lock(&ted->ted_lcd_lock);
1853
1854                         /* create in-memory reply_data and link it to
1855                          * target export's reply list */
1856                         tgt_set_reply_slot(tgt, idx);
1857                         trd->trd_reply = *lrd;
1858                         trd->trd_pre_versions[0] = 0;
1859                         trd->trd_pre_versions[1] = 0;
1860                         trd->trd_pre_versions[2] = 0;
1861                         trd->trd_pre_versions[3] = 0;
1862                         trd->trd_index = idx;
1863                         trd->trd_tag = 0;
1864                         list_add(&trd->trd_list, &ted->ted_reply_list);
1865                         ted->ted_reply_cnt++;
1866                         if (ted->ted_reply_cnt > ted->ted_reply_max)
1867                                 ted->ted_reply_max = ted->ted_reply_cnt;
1868
1869                         CDEBUG(D_HA, "%s: restore reply %p: xid %llu, "
1870                                "transno %llu, client gen %u, slot idx %d\n",
1871                                tgt_name(tgt), trd, lrd->lrd_xid,
1872                                lrd->lrd_transno, lrd->lrd_client_gen,
1873                                trd->trd_index);
1874
1875                         /* update export last committed transation */
1876                         exp->exp_last_committed = max(exp->exp_last_committed,
1877                                                       lrd->lrd_transno);
1878
1879                         mutex_unlock(&ted->ted_lcd_lock);
1880                         class_export_put(exp);
1881
1882                         /* update target last committed transaction */
1883                         spin_lock(&tgt->lut_translock);
1884                         tgt->lut_last_transno = max(tgt->lut_last_transno,
1885                                                     lrd->lrd_transno);
1886                         spin_unlock(&tgt->lut_translock);
1887
1888                         reply_data_recovered++;
1889
1890                         OBD_ALLOC_PTR(trd);
1891                         if (trd == NULL)
1892                                 GOTO(out, rc = -ENOMEM);
1893                 }
1894                 CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
1895                        tgt_name(tgt), reply_data_recovered);
1896         }
1897
1898         spin_lock(&tgt->lut_translock);
1899         /* obd_last_committed is used for compatibility
1900          * with other lustre recovery code */
1901         tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1902         spin_unlock(&tgt->lut_translock);
1903
1904         rc = 0;
1905
1906 out:
1907         if (hash != NULL)
1908                 cfs_hash_putref(hash);
1909         if (lcd != NULL)
1910                 OBD_FREE_PTR(lcd);
1911         if (trd != NULL)
1912                 OBD_FREE_PTR(trd);
1913         if (lrh != NULL)
1914                 OBD_FREE_PTR(lrh);
1915         return rc;
1916 }
1917
1918 struct tg_reply_data *tgt_lookup_reply_by_xid(struct tg_export_data *ted,
1919                                               __u64 xid)
1920 {
1921         struct tg_reply_data    *found = NULL;
1922         struct tg_reply_data    *reply;
1923
1924         mutex_lock(&ted->ted_lcd_lock);
1925         list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
1926                 if (reply->trd_reply.lrd_xid == xid) {
1927                         found = reply;
1928                         break;
1929                 }
1930         }
1931         mutex_unlock(&ted->ted_lcd_lock);
1932         return found;
1933 }
1934 EXPORT_SYMBOL(tgt_lookup_reply_by_xid);
1935
1936 /* Look for a reply data matching specified request @req
1937  * A copy is returned in @trd if the pointer is not NULL
1938  */
1939 bool tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
1940 {
1941         struct tg_export_data   *ted = &req->rq_export->exp_target_data;
1942         struct tg_reply_data    *reply;
1943         bool                     found = false;
1944
1945         reply = tgt_lookup_reply_by_xid(ted, req->rq_xid);
1946         if (reply != NULL) {
1947                 found = true;
1948                 if (trd != NULL)
1949                         *trd = *reply;
1950         }
1951
1952         CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d\n",
1953                tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid,
1954                found ? 1 : 0);
1955
1956         return found;
1957 }
1958 EXPORT_SYMBOL(tgt_lookup_reply);
1959
1960 int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
1961 {
1962         struct tg_export_data   *ted = &exp->exp_target_data;
1963         struct lu_target        *lut = class_exp2tgt(exp);
1964         struct tg_reply_data    *trd, *tmp;
1965
1966         mutex_lock(&ted->ted_lcd_lock);
1967         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
1968                 if (trd->trd_reply.lrd_xid > rcvd_xid)
1969                         continue;
1970                 ted->ted_release_xid++;
1971                 tgt_release_reply_data(lut, ted, trd);
1972         }
1973         mutex_unlock(&ted->ted_lcd_lock);
1974
1975         return 0;
1976 }
1977
1978 int tgt_handle_tag(struct obd_export *exp, __u16 tag)
1979 {
1980         struct tg_export_data   *ted = &exp->exp_target_data;
1981         struct lu_target        *lut = class_exp2tgt(exp);
1982         struct tg_reply_data    *trd, *tmp;
1983
1984         if (tag == 0)
1985                 return 0;
1986
1987         mutex_lock(&ted->ted_lcd_lock);
1988         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
1989                 if (trd->trd_tag != tag)
1990                         continue;
1991                 ted->ted_release_tag++;
1992                 tgt_release_reply_data(lut, ted, trd);
1993                 break;
1994         }
1995         mutex_unlock(&ted->ted_lcd_lock);
1996
1997         return 0;
1998 }
1999