lustre/target/tgt_lastrcvd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Lustre Unified Target
  33  * These are common function to work with last_received file
  34  *
  35  * Author: Mikhail Pershin <mike.pershin@intel.com>
  36  */
  37 #include <obd.h>
  38 #include <obd_class.h>
  39 #include <lustre_fid.h>
  40
  41 #include "tgt_internal.h"
  42
  43 /** version recovery epoch */
  44 #define LR_EPOCH_BITS   32
  45
  46 /* Allocate a bitmap for a chunk of reply data slots */
  47 static int tgt_bitmap_chunk_alloc(struct lu_target *lut, int chunk)
  48 {
  49         unsigned long *bm;
  50
  51         OBD_ALLOC_LARGE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
  52                         sizeof(long));
  53         if (bm == NULL)
  54                 return -ENOMEM;
  55
  56         spin_lock(&lut->lut_client_bitmap_lock);
  57
  58         if (lut->lut_reply_bitmap[chunk] != NULL) {
  59                 /* someone else already allocated the bitmap for this chunk */
  60                 spin_unlock(&lut->lut_client_bitmap_lock);
  61                 OBD_FREE_LARGE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
  62                          sizeof(long));
  63                 return 0;
  64         }
  65
  66         lut->lut_reply_bitmap[chunk] = bm;
  67
  68         spin_unlock(&lut->lut_client_bitmap_lock);
  69
  70         return 0;
  71 }
  72
  73 /* Look for an available reply data slot in the bitmap
  74  * of the target @lut
  75  * Allocate bitmap chunk when first used
  76  * XXX algo could be improved if this routine limits performance
  77  */
  78 static int tgt_find_free_reply_slot(struct lu_target *lut)
  79 {
  80         unsigned long *bmp;
  81         int chunk = 0;
  82         int rc;
  83         int b;
  84
  85         for (chunk = 0; chunk < LUT_REPLY_SLOTS_MAX_CHUNKS; chunk++) {
  86                 /* allocate the bitmap chunk if necessary */
  87                 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
  88                         rc = tgt_bitmap_chunk_alloc(lut, chunk);
  89                         if (rc != 0)
  90                                 return rc;
  91                 }
  92                 bmp = lut->lut_reply_bitmap[chunk];
  93
  94                 /* look for an available slot in this chunk */
  95                 do {
  96                         b = find_first_zero_bit(bmp, LUT_REPLY_SLOTS_PER_CHUNK);
  97                         if (b >= LUT_REPLY_SLOTS_PER_CHUNK)
  98                                 break;
  99
 100                         /* found one */
 101                         if (test_and_set_bit(b, bmp) == 0)
 102                                 return chunk * LUT_REPLY_SLOTS_PER_CHUNK + b;
 103                 } while (true);
 104         }
 105
 106         return -ENOSPC;
 107 }
 108
 109 /* Mark the reply data slot @idx 'used' in the corresponding bitmap chunk
 110  * of the target @lut
 111  * Allocate the bitmap chunk if necessary
 112  */
 113 static int tgt_set_reply_slot(struct lu_target *lut, int idx)
 114 {
 115         int chunk;
 116         int b;
 117         int rc;
 118
 119         chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
 120         b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
 121
 122         LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
 123         LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
 124
 125         /* allocate the bitmap chunk if necessary */
 126         if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
 127                 rc = tgt_bitmap_chunk_alloc(lut, chunk);
 128                 if (rc != 0)
 129                         return rc;
 130         }
 131
 132         /* mark the slot 'used' in this chunk */
 133         if (test_and_set_bit(b, lut->lut_reply_bitmap[chunk]) != 0) {
 134                 CERROR("%s: slot %d already set in bitmap\n",
 135                        tgt_name(lut), idx);
 136                 return -EALREADY;
 137         }
 138
 139         return 0;
 140 }
 141
 142
 143 /* Mark the reply data slot @idx 'unused' in the corresponding bitmap chunk
 144  * of the target @lut
 145  */
 146 static int tgt_clear_reply_slot(struct lu_target *lut, int idx)
 147 {
 148         int chunk;
 149         int b;
 150
 151         if (lut->lut_obd->obd_stopping)
 152                 /*
 153                  * in case of failover keep the bit set in order to
 154                  * avoid overwriting slots in reply_data which might
 155                  * be required by resent rpcs
 156                  */
 157                 return 0;
 158         chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
 159         b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
 160
 161         LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
 162         LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
 163
 164         if (lut->lut_reply_bitmap[chunk] == NULL) {
 165                 CERROR("%s: slot %d not allocated\n",
 166                        tgt_name(lut), idx);
 167                 return -ENOENT;
 168         }
 169
 170         if (test_and_clear_bit(b, lut->lut_reply_bitmap[chunk]) == 0) {
 171                 CERROR("%s: slot %d already clear in bitmap\n",
 172                        tgt_name(lut), idx);
 173                 return -EALREADY;
 174         }
 175
 176         return 0;
 177 }
 178
 179
 180 /* Read header of reply_data file of target @tgt into structure @lrh */
 181 static int tgt_reply_header_read(const struct lu_env *env,
 182                                  struct lu_target *tgt,
 183                                  struct lsd_reply_header *lrh)
 184 {
 185         int                      rc;
 186         struct lsd_reply_header  buf;
 187         struct tgt_thread_info  *tti = tgt_th_info(env);
 188
 189         tti->tti_off = 0;
 190         tti->tti_buf.lb_buf = &buf;
 191         tti->tti_buf.lb_len = sizeof(buf);
 192
 193         rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
 194                             &tti->tti_off);
 195         if (rc != 0)
 196                 return rc;
 197
 198         lrh->lrh_magic = le32_to_cpu(buf.lrh_magic);
 199         lrh->lrh_header_size = le32_to_cpu(buf.lrh_header_size);
 200         lrh->lrh_reply_size = le32_to_cpu(buf.lrh_reply_size);
 201
 202         CDEBUG(D_HA, "%s: read %s header. magic=0x%08x "
 203                "header_size=%d reply_size=%d\n",
 204                 tgt->lut_obd->obd_name, REPLY_DATA,
 205                 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
 206
 207         return 0;
 208 }
 209
 210 /* Write header into replay_data file of target @tgt from structure @lrh */
 211 static int tgt_reply_header_write(const struct lu_env *env,
 212                                   struct lu_target *tgt,
 213                                   struct lsd_reply_header *lrh)
 214 {
 215         int                      rc;
 216         struct lsd_reply_header  buf;
 217         struct tgt_thread_info  *tti = tgt_th_info(env);
 218         struct thandle          *th;
 219         struct dt_object        *dto;
 220
 221         CDEBUG(D_HA, "%s: write %s header. magic=0x%08x "
 222                "header_size=%d reply_size=%d\n",
 223                 tgt->lut_obd->obd_name, REPLY_DATA,
 224                 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
 225
 226         if (tgt->lut_bottom->dd_rdonly)
 227                 RETURN(0);
 228
 229         buf.lrh_magic = cpu_to_le32(lrh->lrh_magic);
 230         buf.lrh_header_size = cpu_to_le32(lrh->lrh_header_size);
 231         buf.lrh_reply_size = cpu_to_le32(lrh->lrh_reply_size);
 232
 233         th = dt_trans_create(env, tgt->lut_bottom);
 234         if (IS_ERR(th))
 235                 return PTR_ERR(th);
 236         th->th_sync = 1;
 237
 238         tti->tti_off = 0;
 239         tti->tti_buf.lb_buf = &buf;
 240         tti->tti_buf.lb_len = sizeof(buf);
 241
 242         rc = dt_declare_record_write(env, tgt->lut_reply_data,
 243                                      &tti->tti_buf, tti->tti_off, th);
 244         if (rc)
 245                 GOTO(out, rc);
 246
 247         rc = dt_trans_start(env, tgt->lut_bottom, th);
 248         if (rc)
 249                 GOTO(out, rc);
 250
 251         dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
 252         rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 253 out:
 254         dt_trans_stop(env, tgt->lut_bottom, th);
 255         return rc;
 256 }
 257
 258 /* Write the reply data @lrd into reply_data file of target @tgt
 259  * at offset @off
 260  */
 261 static int tgt_reply_data_write(const struct lu_env *env, struct lu_target *tgt,
 262                                 struct lsd_reply_data *lrd, loff_t off,
 263                                 struct thandle *th)
 264 {
 265         struct tgt_thread_info  *tti = tgt_th_info(env);
 266         struct dt_object        *dto;
 267         struct lsd_reply_data   *buf = &tti->tti_lrd;
 268
 269         lrd->lrd_result = ptlrpc_status_hton(lrd->lrd_result);
 270
 271         buf->lrd_transno         = cpu_to_le64(lrd->lrd_transno);
 272         buf->lrd_xid             = cpu_to_le64(lrd->lrd_xid);
 273         buf->lrd_data            = cpu_to_le64(lrd->lrd_data);
 274         buf->lrd_result          = cpu_to_le32(lrd->lrd_result);
 275         buf->lrd_client_gen      = cpu_to_le32(lrd->lrd_client_gen);
 276
 277         lrd->lrd_result = ptlrpc_status_ntoh(lrd->lrd_result);
 278
 279         tti->tti_off = off;
 280         tti->tti_buf.lb_buf = buf;
 281         tti->tti_buf.lb_len = sizeof(*buf);
 282
 283         dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
 284         return dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 285 }
 286
 287 /* Read the reply data from reply_data file of target @tgt at offset @off
 288  * into structure @lrd
 289  */
 290 static int tgt_reply_data_read(const struct lu_env *env, struct lu_target *tgt,
 291                                struct lsd_reply_data *lrd, loff_t off)
 292 {
 293         int                      rc;
 294         struct tgt_thread_info  *tti = tgt_th_info(env);
 295         struct lsd_reply_data   *buf = &tti->tti_lrd;
 296
 297         tti->tti_off = off;
 298         tti->tti_buf.lb_buf = buf;
 299         tti->tti_buf.lb_len = sizeof(*buf);
 300
 301         rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
 302                             &tti->tti_off);
 303         if (rc != 0)
 304                 return rc;
 305
 306         lrd->lrd_transno         = le64_to_cpu(buf->lrd_transno);
 307         lrd->lrd_xid             = le64_to_cpu(buf->lrd_xid);
 308         lrd->lrd_data            = le64_to_cpu(buf->lrd_data);
 309         lrd->lrd_result          = le32_to_cpu(buf->lrd_result);
 310         lrd->lrd_client_gen      = le32_to_cpu(buf->lrd_client_gen);
 311
 312         return 0;
 313 }
 314
 315
 316 /* Free the in-memory reply data structure @trd and release
 317  * the corresponding slot in the reply_data file of target @lut
 318  * Called with ted_lcd_lock held
 319  */
 320 static void tgt_free_reply_data(struct lu_target *lut,
 321                                 struct tg_export_data *ted,
 322                                 struct tg_reply_data *trd)
 323 {
 324         CDEBUG(D_TRACE, "%s: free reply data %p: xid %llu, transno %llu, "
 325                "client gen %u, slot idx %d\n",
 326                lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
 327                trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
 328                trd->trd_index);
 329
 330         LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
 331
 332         list_del(&trd->trd_list);
 333         ted->ted_reply_cnt--;
 334         if (lut != NULL && trd->trd_index != TRD_INDEX_MEMORY)
 335                 tgt_clear_reply_slot(lut, trd->trd_index);
 336         OBD_FREE_PTR(trd);
 337 }
 338
 339 /* Release the reply data @trd from target @lut
 340  * The reply data with the highest transno for this export
 341  * is retained to ensure correctness of target recovery
 342  * Called with ted_lcd_lock held
 343  */
 344 static void tgt_release_reply_data(struct lu_target *lut,
 345                                    struct tg_export_data *ted,
 346                                    struct tg_reply_data *trd)
 347 {
 348         CDEBUG(D_TRACE, "%s: release reply data %p: xid %llu, transno %llu, "
 349                "client gen %u, slot idx %d\n",
 350                lut == NULL ? "" : tgt_name(lut), trd, trd->trd_reply.lrd_xid,
 351                trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
 352                trd->trd_index);
 353
 354         LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
 355
 356         /* Do not free the reply data corresponding to the
 357          * highest transno of this export.
 358          * This ensures on-disk reply data is kept and
 359          * last committed transno can be restored from disk in case
 360          * of target recovery
 361          */
 362         if (trd->trd_reply.lrd_transno == ted->ted_lcd->lcd_last_transno) {
 363                 /* free previous retained reply */
 364                 if (ted->ted_reply_last != NULL)
 365                         tgt_free_reply_data(lut, ted, ted->ted_reply_last);
 366                 /* retain the reply */
 367                 list_del_init(&trd->trd_list);
 368                 ted->ted_reply_last = trd;
 369         } else {
 370                 tgt_free_reply_data(lut, ted, trd);
 371         }
 372 }
 373
 374 static inline struct lu_buf *tti_buf_lsd(struct tgt_thread_info *tti)
 375 {
 376         tti->tti_buf.lb_buf = &tti->tti_lsd;
 377         tti->tti_buf.lb_len = sizeof(tti->tti_lsd);
 378         return &tti->tti_buf;
 379 }
 380
 381 static inline struct lu_buf *tti_buf_lcd(struct tgt_thread_info *tti)
 382 {
 383         tti->tti_buf.lb_buf = &tti->tti_lcd;
 384         tti->tti_buf.lb_len = sizeof(tti->tti_lcd);
 385         return &tti->tti_buf;
 386 }
 387
 388 /**
 389  * Allocate in-memory data for client slot related to export.
 390  */
 391 int tgt_client_alloc(struct obd_export *exp)
 392 {
 393         ENTRY;
 394         LASSERT(exp != exp->exp_obd->obd_self_export);
 395
 396         spin_lock_init(&exp->exp_target_data.ted_nodemap_lock);
 397         INIT_LIST_HEAD(&exp->exp_target_data.ted_nodemap_member);
 398         spin_lock_init(&exp->exp_target_data.ted_fmd_lock);
 399         INIT_LIST_HEAD(&exp->exp_target_data.ted_fmd_list);
 400
 401         OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
 402         if (exp->exp_target_data.ted_lcd == NULL)
 403                 RETURN(-ENOMEM);
 404         /* Mark that slot is not yet valid, 0 doesn't work here */
 405         exp->exp_target_data.ted_lr_idx = -1;
 406         INIT_LIST_HEAD(&exp->exp_target_data.ted_reply_list);
 407         mutex_init(&exp->exp_target_data.ted_lcd_lock);
 408         RETURN(0);
 409 }
 410 EXPORT_SYMBOL(tgt_client_alloc);
 411
 412 /**
 413  * Free in-memory data for client slot related to export.
 414  */
 415 void tgt_client_free(struct obd_export *exp)
 416 {
 417         struct tg_export_data   *ted = &exp->exp_target_data;
 418         struct lu_target        *lut = class_exp2tgt(exp);
 419         struct tg_reply_data    *trd, *tmp;
 420
 421         LASSERT(exp != exp->exp_obd->obd_self_export);
 422
 423         tgt_fmd_cleanup(exp);
 424
 425         /* free reply data */
 426         mutex_lock(&ted->ted_lcd_lock);
 427         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
 428                 tgt_release_reply_data(lut, ted, trd);
 429         }
 430         if (ted->ted_reply_last != NULL) {
 431                 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
 432                 ted->ted_reply_last = NULL;
 433         }
 434         mutex_unlock(&ted->ted_lcd_lock);
 435
 436         if (!hlist_unhashed(&exp->exp_gen_hash))
 437                 cfs_hash_del(exp->exp_obd->obd_gen_hash,
 438                              &ted->ted_lcd->lcd_generation,
 439                              &exp->exp_gen_hash);
 440
 441         OBD_FREE_PTR(ted->ted_lcd);
 442         ted->ted_lcd = NULL;
 443
 444         /* Target may have been freed (see LU-7430)
 445          * Slot may be not yet assigned */
 446         if (exp->exp_obd->u.obt.obt_magic != OBT_MAGIC ||
 447             ted->ted_lr_idx < 0)
 448                 return;
 449
 450         /* Clear bit when lcd is freed */
 451         LASSERT(lut && lut->lut_client_bitmap);
 452         if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
 453                 CERROR("%s: client %u bit already clear in bitmap\n",
 454                        exp->exp_obd->obd_name, ted->ted_lr_idx);
 455                 LBUG();
 456         }
 457
 458         if (tgt_is_multimodrpcs_client(exp) && !exp->exp_obd->obd_stopping)
 459                 atomic_dec(&lut->lut_num_clients);
 460 }
 461 EXPORT_SYMBOL(tgt_client_free);
 462
 463 static inline void tgt_check_lcd(const char *obd_name, int index,
 464                                  struct lsd_client_data *lcd)
 465 {
 466         size_t uuid_size = sizeof(lcd->lcd_uuid);
 467
 468         if (strnlen((char*)lcd->lcd_uuid, uuid_size) == uuid_size) {
 469                 lcd->lcd_uuid[uuid_size - 1] = '\0';
 470
 471                 LCONSOLE_ERROR("the client UUID (%s) on %s for exports stored in last_rcvd(index = %d) is bad!\n",
 472                                lcd->lcd_uuid, obd_name, index);
 473         }
 474 }
 475
 476 static int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
 477                                 struct lsd_client_data *lcd,
 478                                 loff_t *off, int index)
 479 {
 480         struct tgt_thread_info  *tti = tgt_th_info(env);
 481         int                      rc;
 482
 483         tti_buf_lcd(tti);
 484         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
 485         if (rc == 0) {
 486                 tgt_check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
 487                 lcd_le_to_cpu(&tti->tti_lcd, lcd);
 488                 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
 489                 lcd->lcd_last_close_result =
 490                         ptlrpc_status_ntoh(lcd->lcd_last_close_result);
 491         }
 492
 493         CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = %llu"
 494                ", last_xid = %llu, last_result = %u, last_data = %u, "
 495                "last_close_transno = %llu, last_close_xid = %llu, "
 496                "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
 497                *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
 498                lcd->lcd_last_result, lcd->lcd_last_data,
 499                lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
 500                lcd->lcd_last_close_result, rc);
 501         return rc;
 502 }
 503
 504 static int tgt_client_data_write(const struct lu_env *env,
 505                                  struct lu_target *tgt,
 506                                  struct lsd_client_data *lcd,
 507                                  loff_t *off, struct thandle *th)
 508 {
 509         struct tgt_thread_info *tti = tgt_th_info(env);
 510         struct dt_object        *dto;
 511
 512         lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
 513         lcd->lcd_last_close_result =
 514                 ptlrpc_status_hton(lcd->lcd_last_close_result);
 515         lcd_cpu_to_le(lcd, &tti->tti_lcd);
 516         tti_buf_lcd(tti);
 517
 518         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
 519         return dt_record_write(env, dto, &tti->tti_buf, off, th);
 520 }
 521
 522 struct tgt_new_client_callback {
 523         struct dt_txn_commit_cb  lncc_cb;
 524         struct obd_export       *lncc_exp;
 525 };
 526
 527 static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
 528                               struct dt_txn_commit_cb *cb, int err)
 529 {
 530         struct tgt_new_client_callback *ccb;
 531
 532         ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
 533
 534         LASSERT(ccb->lncc_exp->exp_obd);
 535
 536         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 537                ccb->lncc_exp->exp_obd->obd_name,
 538                ccb->lncc_exp->exp_client_uuid.uuid);
 539
 540         spin_lock(&ccb->lncc_exp->exp_lock);
 541
 542         ccb->lncc_exp->exp_need_sync = 0;
 543
 544         spin_unlock(&ccb->lncc_exp->exp_lock);
 545         class_export_cb_put(ccb->lncc_exp);
 546
 547         OBD_FREE_PTR(ccb);
 548 }
 549
 550 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
 551 {
 552         struct tgt_new_client_callback  *ccb;
 553         struct dt_txn_commit_cb         *dcb;
 554         int                              rc;
 555
 556         OBD_ALLOC_PTR(ccb);
 557         if (ccb == NULL)
 558                 return -ENOMEM;
 559
 560         ccb->lncc_exp = class_export_cb_get(exp);
 561
 562         dcb = &ccb->lncc_cb;
 563         dcb->dcb_func = tgt_cb_new_client;
 564         INIT_LIST_HEAD(&dcb->dcb_linkage);
 565         strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
 566
 567         rc = dt_trans_cb_add(th, dcb);
 568         if (rc) {
 569                 class_export_cb_put(exp);
 570                 OBD_FREE_PTR(ccb);
 571         }
 572         return rc;
 573 }
 574
 575 /**
 576  * Update client data in last_rcvd
 577  */
 578 static int tgt_client_data_update(const struct lu_env *env,
 579                                   struct obd_export *exp)
 580 {
 581         struct tg_export_data   *ted = &exp->exp_target_data;
 582         struct lu_target        *tgt = class_exp2tgt(exp);
 583         struct tgt_thread_info  *tti = tgt_th_info(env);
 584         struct thandle          *th;
 585         int                      rc = 0;
 586
 587         ENTRY;
 588
 589         if (unlikely(tgt == NULL)) {
 590                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
 591                           class_exp2obd(exp)->obd_name);
 592                 RETURN(-EINVAL);
 593         }
 594
 595         if (tgt->lut_bottom->dd_rdonly)
 596                 RETURN(0);
 597
 598         th = dt_trans_create(env, tgt->lut_bottom);
 599         if (IS_ERR(th))
 600                 RETURN(PTR_ERR(th));
 601
 602         tti_buf_lcd(tti);
 603         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 604                                      &tti->tti_buf,
 605                                      ted->ted_lr_off, th);
 606         if (rc)
 607                 GOTO(out, rc);
 608
 609         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 610         if (rc)
 611                 GOTO(out, rc);
 612
 613         mutex_lock(&ted->ted_lcd_lock);
 614
 615         /*
 616          * Until this operations will be committed the sync is needed
 617          * for this export. This should be done _after_ starting the
 618          * transaction so that many connecting clients will not bring
 619          * server down with lots of sync writes.
 620          */
 621         rc = tgt_new_client_cb_add(th, exp);
 622         if (rc) {
 623                 /* can't add callback, do sync now */
 624                 th->th_sync = 1;
 625         } else {
 626                 spin_lock(&exp->exp_lock);
 627                 exp->exp_need_sync = 1;
 628                 spin_unlock(&exp->exp_lock);
 629         }
 630
 631         tti->tti_off = ted->ted_lr_off;
 632         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
 633
 634         mutex_unlock(&ted->ted_lcd_lock);
 635
 636         EXIT;
 637 out:
 638         dt_trans_stop(env, tgt->lut_bottom, th);
 639         CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
 640                "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
 641                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 642
 643         return rc;
 644 }
 645
 646 static int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
 647 {
 648         struct tgt_thread_info  *tti = tgt_th_info(env);
 649         int                      rc;
 650
 651         tti->tti_off = 0;
 652         tti_buf_lsd(tti);
 653         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
 654                             &tti->tti_off);
 655         if (rc == 0)
 656                 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
 657
 658         CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
 659                "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
 660                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 661         return rc;
 662 }
 663
 664 static int tgt_server_data_write(const struct lu_env *env,
 665                                  struct lu_target *tgt, struct thandle *th)
 666 {
 667         struct tgt_thread_info  *tti = tgt_th_info(env);
 668         struct dt_object        *dto;
 669         int                      rc;
 670
 671         ENTRY;
 672
 673         tti->tti_off = 0;
 674         tti_buf_lsd(tti);
 675         lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
 676
 677         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
 678         rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 679
 680         CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
 681                "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
 682                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 683
 684         RETURN(rc);
 685 }
 686
 687 /**
 688  * Update server data in last_rcvd
 689  */
 690 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
 691                            int sync)
 692 {
 693         struct tgt_thread_info  *tti = tgt_th_info(env);
 694         struct thandle          *th;
 695         int                      rc = 0;
 696
 697         ENTRY;
 698
 699         CDEBUG(D_SUPER,
 700                "%s: mount_count is %llu, last_transno is %llu\n",
 701                tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
 702                tgt->lut_last_transno);
 703
 704         /* Always save latest transno to keep it fresh */
 705         spin_lock(&tgt->lut_translock);
 706         tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
 707         spin_unlock(&tgt->lut_translock);
 708
 709         if (tgt->lut_bottom->dd_rdonly)
 710                 RETURN(0);
 711
 712         th = dt_trans_create(env, tgt->lut_bottom);
 713         if (IS_ERR(th))
 714                 RETURN(PTR_ERR(th));
 715
 716         th->th_sync = sync;
 717
 718         tti_buf_lsd(tti);
 719         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 720                                      &tti->tti_buf, tti->tti_off, th);
 721         if (rc)
 722                 GOTO(out, rc);
 723
 724         rc = dt_trans_start(env, tgt->lut_bottom, th);
 725         if (rc)
 726                 GOTO(out, rc);
 727
 728         rc = tgt_server_data_write(env, tgt, th);
 729 out:
 730         dt_trans_stop(env, tgt->lut_bottom, th);
 731
 732         CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
 733                "last_transno = %llu: rc = %d\n", tgt->lut_obd->obd_name,
 734                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 735         RETURN(rc);
 736 }
 737 EXPORT_SYMBOL(tgt_server_data_update);
 738
 739 static int tgt_truncate_last_rcvd(const struct lu_env *env,
 740                                   struct lu_target *tgt, loff_t size)
 741 {
 742         struct dt_object *dt = tgt->lut_last_rcvd;
 743         struct thandle   *th;
 744         struct lu_attr    attr;
 745         int               rc;
 746
 747         ENTRY;
 748
 749         if (tgt->lut_bottom->dd_rdonly)
 750                 RETURN(0);
 751
 752         attr.la_size = size;
 753         attr.la_valid = LA_SIZE;
 754
 755         th = dt_trans_create(env, tgt->lut_bottom);
 756         if (IS_ERR(th))
 757                 RETURN(PTR_ERR(th));
 758         rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
 759         if (rc)
 760                 GOTO(cleanup, rc);
 761         rc = dt_declare_attr_set(env, dt, &attr, th);
 762         if (rc)
 763                 GOTO(cleanup, rc);
 764         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 765         if (rc)
 766                 GOTO(cleanup, rc);
 767
 768         rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
 769         if (rc == 0)
 770                 rc = dt_attr_set(env, dt, &attr, th);
 771
 772 cleanup:
 773         dt_trans_stop(env, tgt->lut_bottom, th);
 774
 775         RETURN(rc);
 776 }
 777
 778 static void tgt_client_epoch_update(const struct lu_env *env,
 779                                     struct obd_export *exp)
 780 {
 781         struct lsd_client_data  *lcd = exp->exp_target_data.ted_lcd;
 782         struct lu_target        *tgt = class_exp2tgt(exp);
 783
 784         LASSERT(tgt && tgt->lut_bottom);
 785         /** VBR: set client last_epoch to current epoch */
 786         if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
 787                 return;
 788         lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
 789         tgt_client_data_update(env, exp);
 790 }
 791
 792 /**
 793  * Update boot epoch when recovery ends
 794  */
 795 void tgt_boot_epoch_update(struct lu_target *tgt)
 796 {
 797         struct lu_env            env;
 798         struct ptlrpc_request   *req;
 799         __u32                    start_epoch;
 800         LIST_HEAD(client_list);
 801         int                      rc;
 802
 803         if (tgt->lut_obd->obd_stopping)
 804                 return;
 805
 806         rc = lu_env_init(&env, LCT_DT_THREAD);
 807         if (rc) {
 808                 CERROR("%s: can't initialize environment: rc = %d\n",
 809                         tgt->lut_obd->obd_name, rc);
 810                 return;
 811         }
 812
 813         spin_lock(&tgt->lut_translock);
 814         start_epoch = (tgt->lut_last_transno >> LR_EPOCH_BITS) + 1;
 815         tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
 816         tgt->lut_lsd.lsd_start_epoch = start_epoch;
 817         spin_unlock(&tgt->lut_translock);
 818
 819         /**
 820          * The recovery is not yet finished and final queue can still be updated
 821          * with resend requests. Move final list to separate one for processing
 822          */
 823         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
 824         list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
 825         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
 826
 827         /**
 828          * go through list of exports participated in recovery and
 829          * set new epoch for them
 830          */
 831         list_for_each_entry(req, &client_list, rq_list) {
 832                 LASSERT(!req->rq_export->exp_delayed);
 833                 if (!req->rq_export->exp_vbr_failed)
 834                         tgt_client_epoch_update(&env, req->rq_export);
 835         }
 836         /** return list back at once */
 837         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
 838         list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
 839         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
 840
 841         /** Clear MULTI RPCS incompatibility flag if
 842          * - target is MDT and
 843          * - there is no client to recover or the recovery was aborted
 844          */
 845         if (!strncmp(tgt->lut_obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) &&
 846             (atomic_read(&tgt->lut_obd->obd_max_recoverable_clients) == 0 ||
 847             tgt->lut_obd->obd_abort_recovery))
 848                 tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
 849
 850         /** update server epoch */
 851         tgt_server_data_update(&env, tgt, 1);
 852         lu_env_fini(&env);
 853 }
 854
 855 /**
 856  * commit callback, need to update last_committed value
 857  */
 858 struct tgt_last_committed_callback {
 859         struct dt_txn_commit_cb  llcc_cb;
 860         struct lu_target        *llcc_tgt;
 861         struct obd_export       *llcc_exp;
 862         __u64                    llcc_transno;
 863 };
 864
 865 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
 866                                   struct dt_txn_commit_cb *cb, int err)
 867 {
 868         struct tgt_last_committed_callback *ccb;
 869
 870         ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
 871
 872         LASSERT(ccb->llcc_exp);
 873         LASSERT(ccb->llcc_tgt != NULL);
 874         LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
 875
 876         /* error hit, don't update last committed to provide chance to
 877          * replay data after fail */
 878         if (err != 0)
 879                 goto out;
 880
 881         /* Fast path w/o spinlock, if exp_last_committed was updated
 882          * with higher transno, no need to take spinlock and check,
 883          * also no need to update obd_last_committed. */
 884         if (ccb->llcc_transno <= ccb->llcc_exp->exp_last_committed)
 885                 goto out;
 886         spin_lock(&ccb->llcc_tgt->lut_translock);
 887         if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
 888                 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
 889
 890         if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
 891                 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
 892                 spin_unlock(&ccb->llcc_tgt->lut_translock);
 893
 894                 ptlrpc_commit_replies(ccb->llcc_exp);
 895                 tgt_cancel_slc_locks(ccb->llcc_tgt, ccb->llcc_transno);
 896         } else {
 897                 spin_unlock(&ccb->llcc_tgt->lut_translock);
 898         }
 899
 900         CDEBUG(D_HA, "%s: transno %lld is committed\n",
 901                ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
 902
 903 out:
 904         class_export_cb_put(ccb->llcc_exp);
 905         OBD_FREE_PTR(ccb);
 906 }
 907
 908 /**
 909  * Add commit callback function, it returns a non-zero value to inform
 910  * caller to use sync transaction if necessary.
 911  */
 912 static int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
 913                                   struct obd_export *exp, __u64 transno)
 914 {
 915         struct tgt_last_committed_callback      *ccb;
 916         struct dt_txn_commit_cb                 *dcb;
 917         int                                      rc;
 918
 919         OBD_ALLOC_PTR(ccb);
 920         if (ccb == NULL)
 921                 return -ENOMEM;
 922
 923         ccb->llcc_tgt = tgt;
 924         ccb->llcc_exp = class_export_cb_get(exp);
 925         ccb->llcc_transno = transno;
 926
 927         dcb = &ccb->llcc_cb;
 928         dcb->dcb_func = tgt_cb_last_committed;
 929         INIT_LIST_HEAD(&dcb->dcb_linkage);
 930         strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
 931
 932         rc = dt_trans_cb_add(th, dcb);
 933         if (rc) {
 934                 class_export_cb_put(exp);
 935                 OBD_FREE_PTR(ccb);
 936         }
 937
 938         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 939                 /* report failure to force synchronous operation */
 940                 return -EPERM;
 941
 942         /* if exp_need_sync is set, return non-zero value to force
 943          * a sync transaction. */
 944         return rc ? rc : exp->exp_need_sync;
 945 }
 946
 947 static int tgt_is_local_client(const struct lu_env *env,
 948                                       struct obd_export *exp)
 949 {
 950         struct lu_target        *tgt = class_exp2tgt(exp);
 951         struct tgt_session_info *tsi = tgt_ses_info(env);
 952         struct ptlrpc_request   *req = tgt_ses_req(tsi);
 953
 954         if (tgt->lut_local_recovery)
 955                 return 0;
 956         if (!req)
 957                 return 0;
 958         if (!LNetIsPeerLocal(req->rq_peer.nid))
 959                 return 0;
 960         if (exp_connect_flags(exp) & OBD_CONNECT_MDS)
 961                 return 0;
 962
 963         return 1;
 964 }
 965
 966 /**
 967  * Add new client to the last_rcvd upon new connection.
 968  *
 969  * We use a bitmap to locate a free space in the last_rcvd file and initialize
 970  * tg_export_data.
 971  */
 972 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
 973 {
 974         struct tg_export_data   *ted = &exp->exp_target_data;
 975         struct lu_target        *tgt = class_exp2tgt(exp);
 976         int                      rc = 0, idx;
 977
 978         ENTRY;
 979
 980         LASSERT(tgt && tgt->lut_client_bitmap != NULL);
 981         if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
 982                 RETURN(0);
 983
 984         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 985                 RETURN(0);
 986
 987         if (tgt_is_local_client(env, exp)) {
 988                 LCONSOLE_WARN("%s: local client %s w/o recovery\n",
 989                               exp->exp_obd->obd_name, ted->ted_lcd->lcd_uuid);
 990                 exp->exp_no_recovery = 1;
 991                 RETURN(0);
 992         }
 993
 994         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 995          * there's no need for extra complication here
 996          */
 997         idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
 998 repeat:
 999         if (idx >= LR_MAX_CLIENTS ||
1000             OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
1001                 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
1002                        tgt->lut_obd->obd_name,  idx);
1003                 RETURN(-EOVERFLOW);
1004         }
1005         if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
1006                 idx = find_next_zero_bit(tgt->lut_client_bitmap,
1007                                              LR_MAX_CLIENTS, idx);
1008                 goto repeat;
1009         }
1010
1011         ted->ted_lr_idx = idx;
1012         ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
1013                           idx * tgt->lut_lsd.lsd_client_size;
1014
1015         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1016
1017         if (tgt_is_multimodrpcs_client(exp)) {
1018                 /* Set MULTI RPCS incompatibility flag to prevent previous
1019                  * Lustre versions to mount a target with reply_data file */
1020                 atomic_inc(&tgt->lut_num_clients);
1021                 if (!(tgt->lut_lsd.lsd_feature_incompat &
1022                       OBD_INCOMPAT_MULTI_RPCS)) {
1023                         tgt->lut_lsd.lsd_feature_incompat |=
1024                                                         OBD_INCOMPAT_MULTI_RPCS;
1025                         rc = tgt_server_data_update(env, tgt, 1);
1026                         if (rc < 0) {
1027                                 CERROR("%s: unable to set MULTI RPCS "
1028                                        "incompatibility flag\n",
1029                                        exp->exp_obd->obd_name);
1030                                 RETURN(rc);
1031                         }
1032                 }
1033
1034                 /* assign client slot generation */
1035                 ted->ted_lcd->lcd_generation =
1036                                 atomic_inc_return(&tgt->lut_client_generation);
1037         } else {
1038                 ted->ted_lcd->lcd_generation = 0;
1039         }
1040
1041         CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s' "
1042                "generation %d\n",
1043                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1044                ted->ted_lcd->lcd_uuid, ted->ted_lcd->lcd_generation);
1045
1046         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
1047                 RETURN(-ENOSPC);
1048
1049         rc = tgt_client_data_update(env, exp);
1050         if (rc)
1051                 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
1052                        tgt->lut_obd->obd_name, idx, rc);
1053
1054         RETURN(rc);
1055 }
1056 EXPORT_SYMBOL(tgt_client_new);
1057
1058 /* Add an existing client to the MDS in-memory state based on
1059  * a client that was previously found in the last_rcvd file and
1060  * already has an assigned slot (idx >= 0).
1061  *
1062  * It should not be possible to fail adding an existing client - otherwise
1063  * mdt_init_server_data() callsite needs to be fixed.
1064  */
1065 int tgt_client_add(const struct lu_env *env,  struct obd_export *exp, int idx)
1066 {
1067         struct tg_export_data   *ted = &exp->exp_target_data;
1068         struct lu_target        *tgt = class_exp2tgt(exp);
1069
1070         ENTRY;
1071
1072         LASSERT(tgt && tgt->lut_client_bitmap != NULL);
1073         LASSERTF(idx >= 0, "%d\n", idx);
1074
1075         if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
1076             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1077                 RETURN(0);
1078
1079         if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
1080                 CERROR("%s: client %d: bit already set in bitmap!!\n",
1081                        tgt->lut_obd->obd_name,  idx);
1082                 LBUG();
1083         }
1084         atomic_inc(&tgt->lut_num_clients);
1085
1086         CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added, "
1087                "generation %d\n",
1088                tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid,
1089                ted->ted_lcd->lcd_generation);
1090
1091         ted->ted_lr_idx = idx;
1092         ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
1093                           idx * tgt->lut_lsd.lsd_client_size;
1094
1095         mutex_init(&ted->ted_lcd_lock);
1096
1097         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1098
1099         RETURN(0);
1100 }
1101
1102 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
1103 {
1104         struct tg_export_data   *ted = &exp->exp_target_data;
1105         struct lu_target        *tgt = class_exp2tgt(exp);
1106         int                      rc;
1107
1108         ENTRY;
1109
1110         LASSERT(ted->ted_lcd);
1111
1112         if (unlikely(tgt == NULL)) {
1113                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1114                        class_exp2obd(exp)->obd_name);
1115                 RETURN(-EINVAL);
1116         }
1117
1118         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
1119         if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
1120                     (char *)tgt->lut_obd->obd_uuid.uuid) ||
1121             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT ||
1122             exp->exp_no_recovery)
1123                 RETURN(0);
1124
1125         /* Slot may be not yet assigned, use case is race between Client
1126          * reconnect and forced eviction */
1127         if (ted->ted_lr_idx < 0) {
1128                 CWARN("%s: client with UUID '%s' not in bitmap\n",
1129                       tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid);
1130                 RETURN(0);
1131         }
1132
1133         CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
1134                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1135                ted->ted_lcd->lcd_uuid);
1136
1137         /* Clear the bit _after_ zeroing out the client so we don't
1138            race with filter_client_add and zero out new clients.*/
1139         if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
1140                 CERROR("%s: client %u: bit already clear in bitmap!!\n",
1141                        tgt->lut_obd->obd_name, ted->ted_lr_idx);
1142                 LBUG();
1143         }
1144
1145         /* Do not erase record for recoverable client. */
1146         if (exp->exp_flags & OBD_OPT_FAILOVER)
1147                 RETURN(0);
1148
1149         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_DEL))
1150                 RETURN(0);
1151
1152         /* Make sure the server's last_transno is up to date.
1153          * This should be done before zeroing client slot so last_transno will
1154          * be in server data or in client data in case of failure */
1155         rc = tgt_server_data_update(env, tgt, 0);
1156         if (rc != 0) {
1157                 CERROR("%s: failed to update server data, skip client %s "
1158                        "zeroing, rc %d\n", tgt->lut_obd->obd_name,
1159                        ted->ted_lcd->lcd_uuid, rc);
1160                 RETURN(rc);
1161         }
1162
1163         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
1164         rc = tgt_client_data_update(env, exp);
1165
1166         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
1167                "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
1168                tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
1169                ted->ted_lr_idx, ted->ted_lr_off, rc);
1170         RETURN(rc);
1171 }
1172 EXPORT_SYMBOL(tgt_client_del);
1173
1174 static void tgt_clean_by_tag(struct obd_export *exp, __u64 xid, __u16 tag)
1175 {
1176         struct tg_export_data   *ted = &exp->exp_target_data;
1177         struct lu_target        *lut = class_exp2tgt(exp);
1178         struct tg_reply_data    *trd, *tmp;
1179
1180         if (tag == 0)
1181                 return;
1182
1183         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
1184                 if (trd->trd_tag != tag)
1185                         continue;
1186
1187                 LASSERT(ergo(tgt_is_increasing_xid_client(exp),
1188                              trd->trd_reply.lrd_xid <= xid));
1189
1190                 ted->ted_release_tag++;
1191                 tgt_release_reply_data(lut, ted, trd);
1192         }
1193 }
1194
1195 static int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
1196                        struct tg_export_data *ted, struct tg_reply_data *trd,
1197                        struct ptlrpc_request *req,
1198                        struct thandle *th, bool update_lrd_file)
1199 {
1200         struct lsd_reply_data   *lrd;
1201         int     i;
1202
1203         lrd = &trd->trd_reply;
1204         /* update export last transno */
1205         mutex_lock(&ted->ted_lcd_lock);
1206         if (lrd->lrd_transno > ted->ted_lcd->lcd_last_transno)
1207                 ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
1208         mutex_unlock(&ted->ted_lcd_lock);
1209
1210         if (tgt != NULL) {
1211                 /* find a empty slot */
1212                 i = tgt_find_free_reply_slot(tgt);
1213                 if (unlikely(i < 0)) {
1214                         CERROR("%s: couldn't find a slot for reply data: "
1215                                "rc = %d\n", tgt_name(tgt), i);
1216                         RETURN(i);
1217                 }
1218                 trd->trd_index = i;
1219
1220                 if (update_lrd_file) {
1221                         loff_t  off;
1222                         int     rc;
1223
1224                         /* write reply data to disk */
1225                         off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
1226                         rc = tgt_reply_data_write(env, tgt, lrd, off, th);
1227                         if (unlikely(rc != 0)) {
1228                                 CERROR("%s: can't update %s file: rc = %d\n",
1229                                        tgt_name(tgt), REPLY_DATA, rc);
1230                                 RETURN(rc);
1231                         }
1232                 }
1233         } else {
1234                 trd->trd_index = TRD_INDEX_MEMORY;
1235         }
1236
1237         /* add reply data to target export's reply list */
1238         mutex_lock(&ted->ted_lcd_lock);
1239         if (req != NULL) {
1240                 int exclude = tgt_is_increasing_xid_client(req->rq_export) ?
1241                               MSG_REPLAY : MSG_REPLAY|MSG_RESENT;
1242
1243                 if (req->rq_obsolete) {
1244                         mutex_unlock(&ted->ted_lcd_lock);
1245                         RETURN(-EALREADY);
1246                 }
1247
1248                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & exclude))
1249                         tgt_clean_by_tag(req->rq_export, req->rq_xid,
1250                                          trd->trd_tag);
1251         }
1252         list_add(&trd->trd_list, &ted->ted_reply_list);
1253         ted->ted_reply_cnt++;
1254         if (ted->ted_reply_cnt > ted->ted_reply_max)
1255                 ted->ted_reply_max = ted->ted_reply_cnt;
1256         mutex_unlock(&ted->ted_lcd_lock);
1257
1258         CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
1259                "tag %hu, client gen %u, slot idx %d\n",
1260                trd, lrd->lrd_xid, lrd->lrd_transno,
1261                trd->trd_tag, lrd->lrd_client_gen, trd->trd_index);
1262
1263         RETURN(0);
1264 }
1265
1266 int tgt_mk_reply_data(const struct lu_env *env,
1267                       struct lu_target *tgt,
1268                       struct tg_export_data *ted,
1269                       struct ptlrpc_request *req,
1270                       __u64 opdata,
1271                       struct thandle *th,
1272                       bool write_update,
1273                       __u64 transno)
1274 {
1275         struct tg_reply_data    *trd;
1276         struct lsd_reply_data   *lrd;
1277         __u64                   *pre_versions = NULL;
1278         int                     rc;
1279
1280         OBD_ALLOC_PTR(trd);
1281         if (unlikely(trd == NULL))
1282                 RETURN(-ENOMEM);
1283
1284         /* fill reply data information */
1285         lrd = &trd->trd_reply;
1286         lrd->lrd_transno = transno;
1287         if (req != NULL) {
1288                 lrd->lrd_xid = req->rq_xid;
1289                 trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
1290                 lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
1291                 if (write_update) {
1292                         pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1293                         lrd->lrd_result = th->th_result;
1294                 }
1295         } else {
1296                 struct tgt_session_info *tsi;
1297
1298                 LASSERT(env != NULL);
1299                 tsi = tgt_ses_info(env);
1300                 LASSERT(tsi->tsi_xid != 0);
1301
1302                 lrd->lrd_xid = tsi->tsi_xid;
1303                 lrd->lrd_result = tsi->tsi_result;
1304                 lrd->lrd_client_gen = tsi->tsi_client_gen;
1305         }
1306
1307         lrd->lrd_data = opdata;
1308         if (pre_versions) {
1309                 trd->trd_pre_versions[0] = pre_versions[0];
1310                 trd->trd_pre_versions[1] = pre_versions[1];
1311                 trd->trd_pre_versions[2] = pre_versions[2];
1312                 trd->trd_pre_versions[3] = pre_versions[3];
1313         }
1314
1315         rc = tgt_add_reply_data(env, tgt, ted, trd, req,
1316                                 th, write_update);
1317         if (rc < 0)
1318                 OBD_FREE_PTR(trd);
1319         return rc;
1320
1321 }
1322 EXPORT_SYMBOL(tgt_mk_reply_data);
1323
1324 /*
1325  * last_rcvd & last_committed update callbacks
1326  */
1327 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
1328                                 struct dt_object *obj, __u64 opdata,
1329                                 struct thandle *th, struct ptlrpc_request *req)
1330 {
1331         struct tgt_thread_info  *tti = tgt_th_info(env);
1332         struct tgt_session_info *tsi = tgt_ses_info(env);
1333         struct obd_export       *exp = tsi->tsi_exp;
1334         struct tg_export_data   *ted;
1335         __u64                   *transno_p;
1336         int                      rc = 0;
1337         bool                     lw_client;
1338
1339         ENTRY;
1340
1341
1342         LASSERT(exp != NULL);
1343         ted = &exp->exp_target_data;
1344
1345         lw_client = exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT;
1346         if (ted->ted_lr_idx < 0 && !lw_client)
1347                 /* ofd connect may cause transaction before export has
1348                  * last_rcvd slot */
1349                 RETURN(0);
1350
1351         if (req != NULL)
1352                 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
1353         else
1354                 /* From update replay, tti_transno should be set already */
1355                 LASSERT(tti->tti_transno != 0);
1356
1357         spin_lock(&tgt->lut_translock);
1358         if (th->th_result != 0) {
1359                 if (tti->tti_transno != 0) {
1360                         CERROR("%s: replay transno %llu failed: rc = %d\n",
1361                                tgt_name(tgt), tti->tti_transno, th->th_result);
1362                 }
1363         } else if (tti->tti_transno == 0) {
1364                 tti->tti_transno = ++tgt->lut_last_transno;
1365         } else {
1366                 /* should be replay */
1367                 if (tti->tti_transno > tgt->lut_last_transno)
1368                         tgt->lut_last_transno = tti->tti_transno;
1369         }
1370         spin_unlock(&tgt->lut_translock);
1371
1372         /** VBR: set new versions */
1373         if (th->th_result == 0 && obj != NULL) {
1374                 struct dt_object *dto = dt_object_locate(obj, th->th_dev);
1375                 dt_version_set(env, dto, tti->tti_transno, th);
1376         }
1377
1378         /* filling reply data */
1379         CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
1380                tti->tti_transno, tgt->lut_obd->obd_last_committed);
1381
1382         if (req != NULL) {
1383                 req->rq_transno = tti->tti_transno;
1384                 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
1385         }
1386
1387         /* if can't add callback, do sync write */
1388         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp, tti->tti_transno);
1389
1390         if (lw_client) {
1391                 /* All operations performed by LW clients are synchronous and
1392                  * we store the committed transno in the last_rcvd header */
1393                 spin_lock(&tgt->lut_translock);
1394                 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
1395                         tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
1396                         spin_unlock(&tgt->lut_translock);
1397                         /* Although lightweight (LW) connections have no slot
1398                          * in the last_rcvd, we still want to maintain
1399                          * the in-memory lsd_client_data structure in order to
1400                          * properly handle reply reconstruction. */
1401                         rc = tgt_server_data_write(env, tgt, th);
1402                 } else {
1403                         spin_unlock(&tgt->lut_translock);
1404                 }
1405         } else if (ted->ted_lr_off == 0) {
1406                 CERROR("%s: client idx %d has offset %lld\n",
1407                        tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
1408                 RETURN(-EINVAL);
1409         }
1410
1411         /* Target that supports multiple reply data */
1412         if (tgt_is_multimodrpcs_client(exp)) {
1413                 return tgt_mk_reply_data(env, tgt, ted, req, opdata, th,
1414                                          !!(req != NULL), tti->tti_transno);
1415         }
1416
1417         /* Enough for update replay, let's return */
1418         if (req == NULL)
1419                 RETURN(rc);
1420
1421         mutex_lock(&ted->ted_lcd_lock);
1422         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1423         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
1424                 transno_p = &ted->ted_lcd->lcd_last_close_transno;
1425                 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
1426                 ted->ted_lcd->lcd_last_close_result = th->th_result;
1427         } else {
1428                 /* VBR: save versions in last_rcvd for reconstruct. */
1429                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1430
1431                 if (pre_versions) {
1432                         ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
1433                         ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
1434                         ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
1435                         ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
1436                 }
1437                 transno_p = &ted->ted_lcd->lcd_last_transno;
1438                 ted->ted_lcd->lcd_last_xid = req->rq_xid;
1439                 ted->ted_lcd->lcd_last_result = th->th_result;
1440                 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
1441                  * see struct ldlm_reply->lock_policy_res1; */
1442                 ted->ted_lcd->lcd_last_data = opdata;
1443         }
1444
1445         /* Update transno in slot only if non-zero number, i.e. no errors */
1446         if (likely(tti->tti_transno != 0)) {
1447                 /* Don't overwrite bigger transaction number with lower one.
1448                  * That is not sign of problem in all cases, but in any case
1449                  * this value should be monotonically increased only. */
1450                 if (*transno_p > tti->tti_transno) {
1451                         if (!tgt->lut_no_reconstruct) {
1452                                 CERROR("%s: trying to overwrite bigger transno:"
1453                                        "on-disk: %llu, new: %llu replay: "
1454                                        "%d. See LU-617.\n", tgt_name(tgt),
1455                                        *transno_p, tti->tti_transno,
1456                                        req_is_replay(req));
1457                                 if (req_is_replay(req)) {
1458                                         spin_lock(&req->rq_export->exp_lock);
1459                                         req->rq_export->exp_vbr_failed = 1;
1460                                         spin_unlock(&req->rq_export->exp_lock);
1461                                 }
1462                                 mutex_unlock(&ted->ted_lcd_lock);
1463                                 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
1464                         }
1465                 } else {
1466                         *transno_p = tti->tti_transno;
1467                 }
1468         }
1469
1470         if (!lw_client) {
1471                 tti->tti_off = ted->ted_lr_off;
1472                 if (CFS_FAIL_CHECK(OBD_FAIL_TGT_RCVD_EIO))
1473                         rc = -EIO;
1474                 else
1475                         rc = tgt_client_data_write(env, tgt, ted->ted_lcd,
1476                                                    &tti->tti_off, th);
1477                 if (rc < 0) {
1478                         mutex_unlock(&ted->ted_lcd_lock);
1479                         RETURN(rc);
1480                 }
1481         }
1482         mutex_unlock(&ted->ted_lcd_lock);
1483         RETURN(rc);
1484 }
1485
1486 /*
1487  * last_rcvd update for echo client simulation.
1488  * It updates last_rcvd client slot and version of object in
1489  * simple way but with all locks to simulate all drawbacks
1490  */
1491 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
1492                                      struct lu_target *tgt,
1493                                      struct dt_object *obj,
1494                                      struct thandle *th,
1495                                      struct obd_export *exp)
1496 {
1497         struct tgt_thread_info  *tti = tgt_th_info(env);
1498         struct tg_export_data   *ted = &exp->exp_target_data;
1499         int                      rc = 0;
1500
1501         ENTRY;
1502
1503         tti->tti_transno = 0;
1504
1505         spin_lock(&tgt->lut_translock);
1506         if (th->th_result == 0)
1507                 tti->tti_transno = ++tgt->lut_last_transno;
1508         spin_unlock(&tgt->lut_translock);
1509
1510         /** VBR: set new versions */
1511         if (th->th_result == 0 && obj != NULL)
1512                 dt_version_set(env, obj, tti->tti_transno, th);
1513
1514         /* if can't add callback, do sync write */
1515         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
1516                                                 tti->tti_transno);
1517
1518         LASSERT(ted->ted_lr_off > 0);
1519
1520         mutex_lock(&ted->ted_lcd_lock);
1521         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1522         ted->ted_lcd->lcd_last_transno = tti->tti_transno;
1523         ted->ted_lcd->lcd_last_result = th->th_result;
1524
1525         tti->tti_off = ted->ted_lr_off;
1526         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1527         mutex_unlock(&ted->ted_lcd_lock);
1528         RETURN(rc);
1529 }
1530
1531 static int tgt_clients_data_init(const struct lu_env *env,
1532                                  struct lu_target *tgt,
1533                                  unsigned long last_size)
1534 {
1535         struct obd_device       *obd = tgt->lut_obd;
1536         struct lr_server_data   *lsd = &tgt->lut_lsd;
1537         struct lsd_client_data  *lcd = NULL;
1538         struct tg_export_data   *ted;
1539         int                      cl_idx;
1540         int                      rc = 0;
1541         loff_t                   off = lsd->lsd_client_start;
1542         __u32                    generation = 0;
1543         struct cfs_hash         *hash = NULL;
1544
1545         ENTRY;
1546
1547         if (tgt->lut_bottom->dd_rdonly)
1548                 RETURN(0);
1549
1550         BUILD_BUG_ON(offsetof(struct lsd_client_data, lcd_padding) +
1551                      sizeof(lcd->lcd_padding) != LR_CLIENT_SIZE);
1552
1553         OBD_ALLOC_PTR(lcd);
1554         if (lcd == NULL)
1555                 RETURN(-ENOMEM);
1556
1557         hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1558         if (hash == NULL)
1559                 GOTO(err_out, rc = -ENODEV);
1560
1561         for (cl_idx = 0; off < last_size; cl_idx++) {
1562                 struct obd_export       *exp;
1563                 __u64                    last_transno;
1564
1565                 /* Don't assume off is incremented properly by
1566                  * read_record(), in case sizeof(*lcd)
1567                  * isn't the same as fsd->lsd_client_size.  */
1568                 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
1569                 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
1570                 if (rc) {
1571                         CERROR("%s: error reading last_rcvd %s idx %d off "
1572                                "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
1573                                cl_idx, off, rc);
1574                         rc = 0;
1575                         break; /* read error shouldn't cause startup to fail */
1576                 }
1577
1578                 if (lcd->lcd_uuid[0] == '\0') {
1579                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
1580                                cl_idx);
1581                         continue;
1582                 }
1583
1584                 last_transno = lcd_last_transno(lcd);
1585
1586                 /* These exports are cleaned up by disconnect, so they
1587                  * need to be set up like real exports as connect does.
1588                  */
1589                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: %llu"
1590                        " srv lr: %llu lx: %llu gen %u\n", lcd->lcd_uuid,
1591                        cl_idx, last_transno, lsd->lsd_last_transno,
1592                        lcd_last_xid(lcd), lcd->lcd_generation);
1593
1594                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
1595                 if (IS_ERR(exp)) {
1596                         if (PTR_ERR(exp) == -EALREADY) {
1597                                 /* export already exists, zero out this one */
1598                                 CERROR("%s: Duplicate export %s!\n",
1599                                        tgt_name(tgt), lcd->lcd_uuid);
1600                                 continue;
1601                         }
1602                         GOTO(err_out, rc = PTR_ERR(exp));
1603                 }
1604
1605                 ted = &exp->exp_target_data;
1606                 *ted->ted_lcd = *lcd;
1607
1608                 rc = tgt_client_add(env, exp, cl_idx);
1609                 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
1610                 /* VBR: set export last committed version */
1611                 exp->exp_last_committed = last_transno;
1612                 spin_lock(&exp->exp_lock);
1613                 exp->exp_connecting = 0;
1614                 exp->exp_in_recovery = 0;
1615                 spin_unlock(&exp->exp_lock);
1616                 atomic_inc(&obd->obd_max_recoverable_clients);
1617
1618                 if (tgt->lut_lsd.lsd_feature_incompat &
1619                     OBD_INCOMPAT_MULTI_RPCS &&
1620                     lcd->lcd_generation != 0) {
1621                         /* compute the highest valid client generation */
1622                         generation = max(generation, lcd->lcd_generation);
1623                         /* fill client_generation <-> export hash table */
1624                         rc = cfs_hash_add_unique(hash, &lcd->lcd_generation,
1625                                                  &exp->exp_gen_hash);
1626                         if (rc != 0) {
1627                                 CERROR("%s: duplicate export for client "
1628                                        "generation %u\n",
1629                                        tgt_name(tgt), lcd->lcd_generation);
1630                                 class_export_put(exp);
1631                                 GOTO(err_out, rc);
1632                         }
1633                 }
1634
1635                 class_export_put(exp);
1636
1637                 rc = rev_import_init(exp);
1638                 if (rc != 0) {
1639                         class_unlink_export(exp);
1640                         GOTO(err_out, rc);
1641                 }
1642
1643                 /* Need to check last_rcvd even for duplicated exports. */
1644                 CDEBUG(D_OTHER, "client at idx %d has last_transno = %llu\n",
1645                        cl_idx, last_transno);
1646
1647                 spin_lock(&tgt->lut_translock);
1648                 tgt->lut_last_transno = max(last_transno,
1649                                             tgt->lut_last_transno);
1650                 spin_unlock(&tgt->lut_translock);
1651         }
1652
1653         /* record highest valid client generation */
1654         atomic_set(&tgt->lut_client_generation, generation);
1655
1656 err_out:
1657         if (hash != NULL)
1658                 cfs_hash_putref(hash);
1659         OBD_FREE_PTR(lcd);
1660         RETURN(rc);
1661 }
1662
1663 struct server_compat_data {
1664         __u32 rocompat;
1665         __u32 incompat;
1666         __u32 rocinit;
1667         __u32 incinit;
1668 };
1669
1670 static struct server_compat_data tgt_scd[] = {
1671         [LDD_F_SV_TYPE_MDT] = {
1672                 .rocompat = OBD_ROCOMPAT_LOVOBJID,
1673                 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1674                             OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
1675                             OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI |
1676                             OBD_INCOMPAT_MULTI_RPCS,
1677                 .rocinit = OBD_ROCOMPAT_LOVOBJID,
1678                 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1679                            OBD_INCOMPAT_MULTI_OI,
1680         },
1681         [LDD_F_SV_TYPE_OST] = {
1682                 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
1683                 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
1684                             OBD_INCOMPAT_FID,
1685                 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
1686                 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1687         }
1688 };
1689
1690 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1691 {
1692         struct tgt_thread_info          *tti = tgt_th_info(env);
1693         struct lr_server_data           *lsd = &tgt->lut_lsd;
1694         unsigned long                    last_rcvd_size;
1695         __u32                            index;
1696         int                              rc, type;
1697
1698         rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1699         if (rc)
1700                 RETURN(rc);
1701
1702         last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1703
1704         /* ensure padding in the struct is the correct size */
1705         BUILD_BUG_ON(offsetof(struct lr_server_data, lsd_padding) +
1706                      sizeof(lsd->lsd_padding) != LR_SERVER_SIZE);
1707
1708         rc = server_name2index(tgt_name(tgt), &index, NULL);
1709         if (rc < 0) {
1710                 CERROR("%s: Can not get index from name: rc = %d\n",
1711                        tgt_name(tgt), rc);
1712                 RETURN(rc);
1713         }
1714         /* server_name2index() returns type */
1715         type = rc;
1716         if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1717                 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1718                 RETURN(-EINVAL);
1719         }
1720
1721         /* last_rcvd on OST doesn't provide reconstruct support because there
1722          * may be up to 8 in-flight write requests per single slot in
1723          * last_rcvd client data
1724          */
1725         tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1726
1727         if (last_rcvd_size == 0) {
1728                 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1729
1730                 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1731                        sizeof(lsd->lsd_uuid));
1732                 lsd->lsd_last_transno = 0;
1733                 lsd->lsd_mount_count = 0;
1734                 lsd->lsd_server_size = LR_SERVER_SIZE;
1735                 lsd->lsd_client_start = LR_CLIENT_START;
1736                 lsd->lsd_client_size = LR_CLIENT_SIZE;
1737                 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1738                 lsd->lsd_osd_index = index;
1739                 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1740                 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1741         } else {
1742                 rc = tgt_server_data_read(env, tgt);
1743                 if (rc) {
1744                         CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1745                                tgt_name(tgt), rc);
1746                         RETURN(rc);
1747                 }
1748                 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1749                         if (tgt->lut_bottom->dd_rdonly) {
1750                                 /* Such difference may be caused by mounting
1751                                  * up snapshot with new fsname under rd_only
1752                                  * mode. But even if it was NOT, it will not
1753                                  * damage the system because of "rd_only". */
1754                                 memcpy(lsd->lsd_uuid,
1755                                        tgt->lut_obd->obd_uuid.uuid,
1756                                        sizeof(lsd->lsd_uuid));
1757                         } else {
1758                                 LCONSOLE_ERROR_MSG(0x157, "Trying to start "
1759                                                    "OBD %s using the wrong "
1760                                                    "disk %s. Were the /dev/ "
1761                                                    "assignments rearranged?\n",
1762                                                    tgt->lut_obd->obd_uuid.uuid,
1763                                                    lsd->lsd_uuid);
1764                                 RETURN(-EINVAL);
1765                         }
1766                 }
1767
1768                 if (lsd->lsd_osd_index != index) {
1769                         LCONSOLE_ERROR_MSG(0x157,
1770                                            "%s: index %d in last rcvd is different with the index %d in config log, It might be disk corruption!\n",
1771                                            tgt_name(tgt),
1772                                            lsd->lsd_osd_index, index);
1773                         RETURN(-EINVAL);
1774                 }
1775         }
1776
1777         if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1778                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1779                        tgt_name(tgt),
1780                        lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1781                 RETURN(-EINVAL);
1782         }
1783
1784         if (type == LDD_F_SV_TYPE_MDT)
1785                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1786
1787         if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1788                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1789                        tgt_name(tgt),
1790                        lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1791                 RETURN(-EINVAL);
1792         }
1793         /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1794         if (type == LDD_F_SV_TYPE_MDT &&
1795             !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1796                 if (last_rcvd_size > lsd->lsd_client_start) {
1797                         LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1798                                       "remove all clients for interop needs\n",
1799                                       tgt_name(tgt));
1800                         rc = tgt_truncate_last_rcvd(env, tgt,
1801                                                     lsd->lsd_client_start);
1802                         if (rc)
1803                                 RETURN(rc);
1804                         last_rcvd_size = lsd->lsd_client_start;
1805                 }
1806                 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1807                 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1808         }
1809
1810         spin_lock(&tgt->lut_translock);
1811         tgt->lut_last_transno = lsd->lsd_last_transno;
1812         spin_unlock(&tgt->lut_translock);
1813
1814         lsd->lsd_mount_count++;
1815
1816         CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1817         CDEBUG(D_INODE, "%s: server last_transno: %llu\n",
1818                tgt_name(tgt), tgt->lut_last_transno);
1819         CDEBUG(D_INODE, "%s: server mount_count: %llu\n",
1820                tgt_name(tgt), lsd->lsd_mount_count);
1821         CDEBUG(D_INODE, "%s: server data size: %u\n",
1822                tgt_name(tgt), lsd->lsd_server_size);
1823         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1824                tgt_name(tgt), lsd->lsd_client_start);
1825         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1826                tgt_name(tgt), lsd->lsd_client_size);
1827         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1828                tgt_name(tgt), last_rcvd_size);
1829         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1830                tgt_name(tgt), lsd->lsd_subdir_count);
1831         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1832                last_rcvd_size <= lsd->lsd_client_start ? 0 :
1833                (last_rcvd_size - lsd->lsd_client_start) /
1834                 lsd->lsd_client_size);
1835         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1836
1837         if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1838             lsd->lsd_client_size == 0) {
1839                 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1840                 RETURN(-EINVAL);
1841         }
1842
1843         if (!tgt->lut_obd->obd_replayable)
1844                 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1845
1846         rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1847         if (rc < 0)
1848                 GOTO(err_client, rc);
1849
1850         spin_lock(&tgt->lut_translock);
1851         /* obd_last_committed is used for compatibility
1852          * with other lustre recovery code */
1853         tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1854         spin_unlock(&tgt->lut_translock);
1855
1856         tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1857         tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1858
1859         /* save it, so mount count and last_transno is current */
1860         rc = tgt_server_data_update(env, tgt, 0);
1861         if (rc < 0)
1862                 GOTO(err_client, rc);
1863
1864         RETURN(0);
1865
1866 err_client:
1867         class_disconnect_exports(tgt->lut_obd);
1868         return rc;
1869 }
1870
1871 /* add credits for last_rcvd update */
1872 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1873                      void *cookie)
1874 {
1875         struct lu_target        *tgt = cookie;
1876         struct tgt_session_info *tsi;
1877         struct tgt_thread_info  *tti = tgt_th_info(env);
1878         struct dt_object        *dto;
1879         int                      rc;
1880
1881         /* For readonly case, the caller should have got failure
1882          * when start the transaction. If the logic comes here,
1883          * there must be something wrong. */
1884         if (unlikely(tgt->lut_bottom->dd_rdonly)) {
1885                 dump_stack();
1886                 LBUG();
1887         }
1888
1889         /* if there is no session, then this transaction is not result of
1890          * request processing but some local operation */
1891         if (env->le_ses == NULL)
1892                 return 0;
1893
1894         LASSERT(tgt->lut_last_rcvd);
1895         tsi = tgt_ses_info(env);
1896         /* OFD may start transaction without export assigned */
1897         if (tsi->tsi_exp == NULL)
1898                 return 0;
1899
1900         if (tgt_is_multimodrpcs_client(tsi->tsi_exp)) {
1901                 /*
1902                  * Use maximum possible file offset for declaration to ensure
1903                  * ZFS will reserve enough credits for a write anywhere in this
1904                  * file, since we don't know where in the file the write will be
1905                  * because a replay slot has not been assigned.  This should be
1906                  * replaced by dmu_tx_hold_append() when available.
1907                  */
1908                 tti->tti_buf.lb_buf = NULL;
1909                 tti->tti_buf.lb_len = sizeof(struct lsd_reply_data);
1910                 dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
1911                 rc = dt_declare_record_write(env, dto, &tti->tti_buf, -1, th);
1912                 if (rc)
1913                         return rc;
1914         } else {
1915                 dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
1916                 tti_buf_lcd(tti);
1917                 tti->tti_off = tsi->tsi_exp->exp_target_data.ted_lr_off;
1918                 rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1919                                              tti->tti_off, th);
1920                 if (rc)
1921                         return rc;
1922         }
1923
1924         if (tsi->tsi_vbr_obj != NULL &&
1925             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1926                 dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
1927                 rc = dt_declare_version_set(env, dto, th);
1928         }
1929
1930         return rc;
1931 }
1932
1933 /* Update last_rcvd records with latests transaction data */
1934 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1935                     void *cookie)
1936 {
1937         struct lu_target        *tgt = cookie;
1938         struct tgt_session_info *tsi;
1939         struct tgt_thread_info  *tti = tgt_th_info(env);
1940         struct dt_object        *obj = NULL;
1941         int                      rc;
1942         bool                     echo_client;
1943
1944         if (env->le_ses == NULL)
1945                 return 0;
1946
1947         tsi = tgt_ses_info(env);
1948         /* OFD may start transaction without export assigned */
1949         if (tsi->tsi_exp == NULL)
1950                 return 0;
1951
1952         echo_client = (tgt_ses_req(tsi) == NULL && tsi->tsi_xid == 0);
1953
1954         if (tti->tti_has_trans && !echo_client) {
1955                 if (tti->tti_mult_trans == 0) {
1956                         CDEBUG(D_HA, "More than one transaction %llu\n",
1957                                tti->tti_transno);
1958                         RETURN(0);
1959                 }
1960                 /* we need another transno to be assigned */
1961                 tti->tti_transno = 0;
1962         } else if (th->th_result == 0) {
1963                 tti->tti_has_trans = 1;
1964         }
1965
1966         if (tsi->tsi_vbr_obj != NULL &&
1967             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1968                 obj = tsi->tsi_vbr_obj;
1969         }
1970
1971         if (unlikely(echo_client)) /* echo client special case */
1972                 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1973                                                tsi->tsi_exp);
1974         else
1975                 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
1976                                           tgt_ses_req(tsi));
1977         return rc;
1978 }
1979
1980 int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
1981 {
1982         struct tgt_thread_info  *tti = tgt_th_info(env);
1983         struct lsd_reply_data   *lrd = &tti->tti_lrd;
1984         unsigned long            reply_data_size;
1985         int                      rc;
1986         struct lsd_reply_header *lrh = NULL;
1987         struct tg_reply_data    *trd = NULL;
1988         int                      idx;
1989         loff_t                   off;
1990         struct cfs_hash         *hash = NULL;
1991         struct obd_export       *exp;
1992         struct tg_export_data   *ted;
1993         int                      reply_data_recovered = 0;
1994
1995         rc = dt_attr_get(env, tgt->lut_reply_data, &tti->tti_attr);
1996         if (rc)
1997                 GOTO(out, rc);
1998         reply_data_size = (unsigned long)tti->tti_attr.la_size;
1999
2000         OBD_ALLOC_PTR(lrh);
2001         if (lrh == NULL)
2002                 GOTO(out, rc = -ENOMEM);
2003
2004         if (reply_data_size == 0) {
2005                 CDEBUG(D_INFO, "%s: new reply_data file, initializing\n",
2006                        tgt_name(tgt));
2007                 lrh->lrh_magic = LRH_MAGIC;
2008                 lrh->lrh_header_size = sizeof(struct lsd_reply_header);
2009                 lrh->lrh_reply_size = sizeof(struct lsd_reply_data);
2010                 rc = tgt_reply_header_write(env, tgt, lrh);
2011                 if (rc) {
2012                         CERROR("%s: error writing %s: rc = %d\n",
2013                                tgt_name(tgt), REPLY_DATA, rc);
2014                         GOTO(out, rc);
2015                 }
2016         } else {
2017                 rc = tgt_reply_header_read(env, tgt, lrh);
2018                 if (rc) {
2019                         CERROR("%s: error reading %s: rc = %d\n",
2020                                tgt_name(tgt), REPLY_DATA, rc);
2021                         GOTO(out, rc);
2022                 }
2023                 if (lrh->lrh_magic != LRH_MAGIC ||
2024                     lrh->lrh_header_size != sizeof(struct lsd_reply_header) ||
2025                     lrh->lrh_reply_size != sizeof(struct lsd_reply_data)) {
2026                         CERROR("%s: invalid header in %s\n",
2027                                tgt_name(tgt), REPLY_DATA);
2028                         GOTO(out, rc = -EINVAL);
2029                 }
2030
2031                 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
2032                 if (hash == NULL)
2033                         GOTO(out, rc = -ENODEV);
2034
2035                 OBD_ALLOC_PTR(trd);
2036                 if (trd == NULL)
2037                         GOTO(out, rc = -ENOMEM);
2038
2039                 /* Load reply_data from disk */
2040                 for (idx = 0, off = sizeof(struct lsd_reply_header);
2041                      off < reply_data_size;
2042                      idx++, off += sizeof(struct lsd_reply_data)) {
2043                         rc = tgt_reply_data_read(env, tgt, lrd, off);
2044                         if (rc) {
2045                                 CERROR("%s: error reading %s: rc = %d\n",
2046                                        tgt_name(tgt), REPLY_DATA, rc);
2047                                 GOTO(out, rc);
2048                         }
2049
2050                         exp = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
2051                         if (exp == NULL) {
2052                                 /* old reply data from a disconnected client */
2053                                 continue;
2054                         }
2055                         ted = &exp->exp_target_data;
2056                         mutex_lock(&ted->ted_lcd_lock);
2057
2058                         /* create in-memory reply_data and link it to
2059                          * target export's reply list */
2060                         rc = tgt_set_reply_slot(tgt, idx);
2061                         if (rc != 0) {
2062                                 mutex_unlock(&ted->ted_lcd_lock);
2063                                 GOTO(out, rc);
2064                         }
2065                         trd->trd_reply = *lrd;
2066                         trd->trd_pre_versions[0] = 0;
2067                         trd->trd_pre_versions[1] = 0;
2068                         trd->trd_pre_versions[2] = 0;
2069                         trd->trd_pre_versions[3] = 0;
2070                         trd->trd_index = idx;
2071                         trd->trd_tag = 0;
2072                         list_add(&trd->trd_list, &ted->ted_reply_list);
2073                         ted->ted_reply_cnt++;
2074                         if (ted->ted_reply_cnt > ted->ted_reply_max)
2075                                 ted->ted_reply_max = ted->ted_reply_cnt;
2076
2077                         CDEBUG(D_HA, "%s: restore reply %p: xid %llu, "
2078                                "transno %llu, client gen %u, slot idx %d\n",
2079                                tgt_name(tgt), trd, lrd->lrd_xid,
2080                                lrd->lrd_transno, lrd->lrd_client_gen,
2081                                trd->trd_index);
2082
2083                         /* update export last committed transation */
2084                         exp->exp_last_committed = max(exp->exp_last_committed,
2085                                                       lrd->lrd_transno);
2086                         /* Update lcd_last_transno as well for check in
2087                          * tgt_release_reply_data() or the latest client
2088                          * transno can be lost.
2089                          */
2090                         ted->ted_lcd->lcd_last_transno =
2091                                 max(ted->ted_lcd->lcd_last_transno,
2092                                     exp->exp_last_committed);
2093
2094                         mutex_unlock(&ted->ted_lcd_lock);
2095                         class_export_put(exp);
2096
2097                         /* update target last committed transaction */
2098                         spin_lock(&tgt->lut_translock);
2099                         tgt->lut_last_transno = max(tgt->lut_last_transno,
2100                                                     lrd->lrd_transno);
2101                         spin_unlock(&tgt->lut_translock);
2102
2103                         reply_data_recovered++;
2104
2105                         OBD_ALLOC_PTR(trd);
2106                         if (trd == NULL)
2107                                 GOTO(out, rc = -ENOMEM);
2108                 }
2109                 CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
2110                        tgt_name(tgt), reply_data_recovered);
2111         }
2112
2113         spin_lock(&tgt->lut_translock);
2114         /* obd_last_committed is used for compatibility
2115          * with other lustre recovery code */
2116         tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
2117         spin_unlock(&tgt->lut_translock);
2118
2119         rc = 0;
2120
2121 out:
2122         if (hash != NULL)
2123                 cfs_hash_putref(hash);
2124         if (trd != NULL)
2125                 OBD_FREE_PTR(trd);
2126         if (lrh != NULL)
2127                 OBD_FREE_PTR(lrh);
2128         return rc;
2129 }
2130
2131 static int tgt_check_lookup_req(struct ptlrpc_request *req, int lookup,
2132                                 struct tg_reply_data *trd)
2133 {
2134         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2135         struct lu_target *lut = class_exp2tgt(req->rq_export);
2136         __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
2137         int rc = 0;
2138         struct tg_reply_data *reply;
2139         bool check_increasing;
2140
2141         if (tag == 0)
2142                 return 0;
2143
2144         check_increasing = tgt_is_increasing_xid_client(req->rq_export) &&
2145                            !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
2146         if (!lookup && !check_increasing)
2147                 return 0;
2148
2149         list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
2150                 if (lookup && reply->trd_reply.lrd_xid == req->rq_xid) {
2151                         rc = 1;
2152                         if (trd != NULL)
2153                                 *trd = *reply;
2154                         break;
2155                 } else if (check_increasing && reply->trd_tag == tag &&
2156                            reply->trd_reply.lrd_xid > req->rq_xid) {
2157                         rc = -EPROTO;
2158                         CERROR("%s: busy tag=%u req_xid=%llu, trd=%p: xid=%llu transno=%llu client_gen=%u slot_idx=%d: rc = %d\n",
2159                                tgt_name(lut), tag, req->rq_xid, trd,
2160                                reply->trd_reply.lrd_xid,
2161                                reply->trd_reply.lrd_transno,
2162                                reply->trd_reply.lrd_client_gen,
2163                                reply->trd_index, rc);
2164                         break;
2165                 }
2166         }
2167
2168         return rc;
2169 }
2170
2171 /* Look for a reply data matching specified request @req
2172  * A copy is returned in @trd if the pointer is not NULL
2173  */
2174 int tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
2175 {
2176         struct tg_export_data *ted = &req->rq_export->exp_target_data;
2177         int found = 0;
2178         bool not_replay = !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY);
2179
2180         mutex_lock(&ted->ted_lcd_lock);
2181         if (not_replay && req->rq_xid <= req->rq_export->exp_last_xid) {
2182                 /* A check for the last_xid is needed here in case there is
2183                  * no reply data is left in the list. It may happen if another
2184                  * RPC on another slot increased the last_xid between our
2185                  * process_req_last_xid & tgt_lookup_reply calls */
2186                 found = -EPROTO;
2187         } else {
2188                 found = tgt_check_lookup_req(req, 1, trd);
2189         }
2190         mutex_unlock(&ted->ted_lcd_lock);
2191
2192         CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d last_xid %llu\n",
2193                tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid, found,
2194                req->rq_export->exp_last_xid);
2195
2196         return found;
2197 }
2198 EXPORT_SYMBOL(tgt_lookup_reply);
2199
2200 int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
2201 {
2202         struct tg_export_data   *ted = &exp->exp_target_data;
2203         struct lu_target        *lut = class_exp2tgt(exp);
2204         struct tg_reply_data    *trd, *tmp;
2205
2206
2207         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
2208                 if (trd->trd_reply.lrd_xid > rcvd_xid)
2209                         continue;
2210                 ted->ted_release_xid++;
2211                 tgt_release_reply_data(lut, ted, trd);
2212         }
2213
2214         return 0;
2215 }
2216
2217 int tgt_handle_tag(struct ptlrpc_request *req)
2218 {
2219         return tgt_check_lookup_req(req, 0, NULL);
2220 }
2221