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