lustre/target/tgt_lastrcvd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Lustre Unified Target
  37  * These are common function to work with last_received file
  38  *
  39  * Author: Mikhail Pershin <mike.pershin@intel.com>
  40  */
  41 #include <obd.h>
  42 #include <obd_class.h>
  43 #include <lustre_fid.h>
  44
  45 #include "tgt_internal.h"
  46
  47
  48 /* Allocate a bitmap for a chunk of reply data slots */
  49 static int tgt_bitmap_chunk_alloc(struct lu_target *lut, int chunk)
  50 {
  51         unsigned long *bm;
  52
  53         OBD_ALLOC(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) * sizeof(long));
  54         if (bm == NULL)
  55                 return -ENOMEM;
  56
  57         spin_lock(&lut->lut_client_bitmap_lock);
  58
  59         if (lut->lut_reply_bitmap[chunk] != NULL) {
  60                 /* someone else already allocated the bitmap for this chunk */
  61                 spin_unlock(&lut->lut_client_bitmap_lock);
  62                 OBD_FREE(bm, BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
  63                          sizeof(long));
  64                 return 0;
  65         }
  66
  67         lut->lut_reply_bitmap[chunk] = bm;
  68
  69         spin_unlock(&lut->lut_client_bitmap_lock);
  70
  71         return 0;
  72 }
  73
  74 /* Look for an available reply data slot in the bitmap
  75  * of the target @lut
  76  * Allocate bitmap chunk when first used
  77  * XXX algo could be improved if this routine limits performance
  78  */
  79 static int tgt_find_free_reply_slot(struct lu_target *lut)
  80 {
  81         unsigned long *bmp;
  82         int chunk = 0;
  83         int rc;
  84         int b;
  85
  86         for (chunk = 0; chunk < LUT_REPLY_SLOTS_MAX_CHUNKS; chunk++) {
  87                 /* allocate the bitmap chunk if necessary */
  88                 if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
  89                         rc = tgt_bitmap_chunk_alloc(lut, chunk);
  90                         if (rc != 0)
  91                                 return rc;
  92                 }
  93                 bmp = lut->lut_reply_bitmap[chunk];
  94
  95                 /* look for an available slot in this chunk */
  96                 do {
  97                         b = find_first_zero_bit(bmp, LUT_REPLY_SLOTS_PER_CHUNK);
  98                         if (b >= LUT_REPLY_SLOTS_PER_CHUNK)
  99                                 break;
 100
 101                         /* found one */
 102                         if (test_and_set_bit(b, bmp) == 0)
 103                                 return chunk * LUT_REPLY_SLOTS_PER_CHUNK + b;
 104                 } while (true);
 105         }
 106
 107         return -ENOSPC;
 108 }
 109
 110 /* Mark the reply data slot @idx 'used' in the corresponding bitmap chunk
 111  * of the target @lut
 112  * Allocate the bitmap chunk if necessary
 113  */
 114 static int tgt_set_reply_slot(struct lu_target *lut, int idx)
 115 {
 116         int chunk;
 117         int b;
 118         int rc;
 119
 120         chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
 121         b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
 122
 123         LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
 124         LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
 125
 126         /* allocate the bitmap chunk if necessary */
 127         if (unlikely(lut->lut_reply_bitmap[chunk] == NULL)) {
 128                 rc = tgt_bitmap_chunk_alloc(lut, chunk);
 129                 if (rc != 0)
 130                         return rc;
 131         }
 132
 133         /* mark the slot 'used' in this chunk */
 134         if (test_and_set_bit(b, lut->lut_reply_bitmap[chunk]) != 0) {
 135                 CERROR("%s: slot %d already set in bitmap\n",
 136                        tgt_name(lut), idx);
 137                 return -EALREADY;
 138         }
 139
 140         return 0;
 141 }
 142
 143
 144 /* Mark the reply data slot @idx 'unused' in the corresponding bitmap chunk
 145  * of the target @lut
 146  */
 147 static int tgt_clear_reply_slot(struct lu_target *lut, int idx)
 148 {
 149         int chunk;
 150         int b;
 151
 152         chunk = idx / LUT_REPLY_SLOTS_PER_CHUNK;
 153         b = idx % LUT_REPLY_SLOTS_PER_CHUNK;
 154
 155         LASSERT(chunk < LUT_REPLY_SLOTS_MAX_CHUNKS);
 156         LASSERT(b < LUT_REPLY_SLOTS_PER_CHUNK);
 157
 158         if (test_and_clear_bit(b, lut->lut_reply_bitmap[chunk]) == 0) {
 159                 CERROR("%s: slot %d already clear in bitmap\n",
 160                        tgt_name(lut), idx);
 161                 return -EALREADY;
 162         }
 163
 164         return 0;
 165 }
 166
 167
 168 /* Read header of reply_data file of target @tgt into structure @lrh */
 169 static int tgt_reply_header_read(const struct lu_env *env,
 170                                  struct lu_target *tgt,
 171                                  struct lsd_reply_header *lrh)
 172 {
 173         int                      rc;
 174         struct lsd_reply_header  buf;
 175         struct tgt_thread_info  *tti = tgt_th_info(env);
 176
 177         tti->tti_off = 0;
 178         tti->tti_buf.lb_buf = &buf;
 179         tti->tti_buf.lb_len = sizeof(buf);
 180
 181         rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
 182                             &tti->tti_off);
 183         if (rc != 0)
 184                 return rc;
 185
 186         lrh->lrh_magic = le32_to_cpu(buf.lrh_magic);
 187         lrh->lrh_header_size = le32_to_cpu(buf.lrh_header_size);
 188         lrh->lrh_reply_size = le32_to_cpu(buf.lrh_reply_size);
 189
 190         CDEBUG(D_HA, "%s: read %s header. magic=0x%08x "
 191                "header_size=%d reply_size=%d\n",
 192                 tgt->lut_obd->obd_name, REPLY_DATA,
 193                 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
 194
 195         return 0;
 196 }
 197
 198 /* Write header into replay_data file of target @tgt from structure @lrh */
 199 static int tgt_reply_header_write(const struct lu_env *env,
 200                                   struct lu_target *tgt,
 201                                   struct lsd_reply_header *lrh)
 202 {
 203         int                      rc;
 204         struct lsd_reply_header  buf;
 205         struct tgt_thread_info  *tti = tgt_th_info(env);
 206         struct thandle          *th;
 207         struct dt_object        *dto;
 208
 209         CDEBUG(D_HA, "%s: write %s header. magic=0x%08x "
 210                "header_size=%d reply_size=%d\n",
 211                 tgt->lut_obd->obd_name, REPLY_DATA,
 212                 lrh->lrh_magic, lrh->lrh_header_size, lrh->lrh_reply_size);
 213
 214         buf.lrh_magic = cpu_to_le32(lrh->lrh_magic);
 215         buf.lrh_header_size = cpu_to_le32(lrh->lrh_header_size);
 216         buf.lrh_reply_size = cpu_to_le32(lrh->lrh_reply_size);
 217
 218         th = dt_trans_create(env, tgt->lut_bottom);
 219         if (IS_ERR(th))
 220                 return PTR_ERR(th);
 221         th->th_sync = 1;
 222
 223         tti->tti_off = 0;
 224         tti->tti_buf.lb_buf = &buf;
 225         tti->tti_buf.lb_len = sizeof(buf);
 226
 227         rc = dt_declare_record_write(env, tgt->lut_reply_data,
 228                                      &tti->tti_buf, tti->tti_off, th);
 229         if (rc)
 230                 GOTO(out, rc);
 231
 232         rc = dt_trans_start(env, tgt->lut_bottom, th);
 233         if (rc)
 234                 GOTO(out, rc);
 235
 236         dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
 237         rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 238 out:
 239         dt_trans_stop(env, tgt->lut_bottom, th);
 240         return rc;
 241 }
 242
 243 /* Write the reply data @lrd into reply_data file of target @tgt
 244  * at offset @off
 245  */
 246 static int tgt_reply_data_write(const struct lu_env *env, struct lu_target *tgt,
 247                                 struct lsd_reply_data *lrd, loff_t off,
 248                                 struct thandle *th)
 249 {
 250         struct tgt_thread_info  *tti = tgt_th_info(env);
 251         struct dt_object        *dto;
 252         struct lsd_reply_data   *buf = &tti->tti_lrd;
 253
 254         lrd->lrd_result = ptlrpc_status_hton(lrd->lrd_result);
 255
 256         buf->lrd_transno         = cpu_to_le64(lrd->lrd_transno);
 257         buf->lrd_xid             = cpu_to_le64(lrd->lrd_xid);
 258         buf->lrd_data            = cpu_to_le64(lrd->lrd_data);
 259         buf->lrd_result          = cpu_to_le32(lrd->lrd_result);
 260         buf->lrd_client_gen      = cpu_to_le32(lrd->lrd_client_gen);
 261
 262         lrd->lrd_result = ptlrpc_status_ntoh(lrd->lrd_result);
 263
 264         tti->tti_off = off;
 265         tti->tti_buf.lb_buf = buf;
 266         tti->tti_buf.lb_len = sizeof(*buf);
 267
 268         dto = dt_object_locate(tgt->lut_reply_data, th->th_dev);
 269         return dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 270 }
 271
 272 /* Read the reply data from reply_data file of target @tgt at offset @off
 273  * into structure @lrd
 274  */
 275 static int tgt_reply_data_read(const struct lu_env *env, struct lu_target *tgt,
 276                                struct lsd_reply_data *lrd, loff_t off)
 277 {
 278         int                      rc;
 279         struct tgt_thread_info  *tti = tgt_th_info(env);
 280         struct lsd_reply_data   *buf = &tti->tti_lrd;
 281
 282         tti->tti_off = off;
 283         tti->tti_buf.lb_buf = buf;
 284         tti->tti_buf.lb_len = sizeof(*buf);
 285
 286         rc = dt_record_read(env, tgt->lut_reply_data, &tti->tti_buf,
 287                             &tti->tti_off);
 288         if (rc != 0)
 289                 return rc;
 290
 291         lrd->lrd_transno         = le64_to_cpu(buf->lrd_transno);
 292         lrd->lrd_xid             = le64_to_cpu(buf->lrd_xid);
 293         lrd->lrd_data            = le64_to_cpu(buf->lrd_data);
 294         lrd->lrd_result          = le32_to_cpu(buf->lrd_result);
 295         lrd->lrd_client_gen      = le32_to_cpu(buf->lrd_client_gen);
 296
 297         return 0;
 298 }
 299
 300
 301 /* Free the in-memory reply data structure @trd and release
 302  * the corresponding slot in the reply_data file of target @lut
 303  * Called with ted_lcd_lock held
 304  */
 305 static void tgt_free_reply_data(struct lu_target *lut,
 306                                 struct tg_export_data *ted,
 307                                 struct tg_reply_data *trd)
 308 {
 309         CDEBUG(D_TRACE, "%s: free reply data %p: xid %llu, transno %llu, "
 310                "client gen %u, slot idx %d\n",
 311                tgt_name(lut), trd, trd->trd_reply.lrd_xid,
 312                trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
 313                trd->trd_index);
 314
 315         LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
 316
 317         list_del(&trd->trd_list);
 318         ted->ted_reply_cnt--;
 319         tgt_clear_reply_slot(lut, trd->trd_index);
 320         OBD_FREE_PTR(trd);
 321 }
 322
 323 /* Release the reply data @trd from target @lut
 324  * The reply data with the highest transno for this export
 325  * is retained to ensure correctness of target recovery
 326  * Called with ted_lcd_lock held
 327  */
 328 static void tgt_release_reply_data(struct lu_target *lut,
 329                                    struct tg_export_data *ted,
 330                                    struct tg_reply_data *trd)
 331 {
 332         CDEBUG(D_TRACE, "%s: release reply data %p: xid %llu, transno %llu, "
 333                "client gen %u, slot idx %d\n",
 334                tgt_name(lut), trd, trd->trd_reply.lrd_xid,
 335                trd->trd_reply.lrd_transno, trd->trd_reply.lrd_client_gen,
 336                trd->trd_index);
 337
 338         LASSERT(mutex_is_locked(&ted->ted_lcd_lock));
 339
 340         /* Do not free the reply data corresponding to the
 341          * highest transno of this export.
 342          * This ensures on-disk reply data is kept and
 343          * last committed transno can be restored from disk in case
 344          * of target recovery
 345          */
 346         if (trd->trd_reply.lrd_transno == ted->ted_lcd->lcd_last_transno) {
 347                 /* free previous retained reply */
 348                 if (ted->ted_reply_last != NULL)
 349                         tgt_free_reply_data(lut, ted, ted->ted_reply_last);
 350                 /* retain the reply */
 351                 list_del_init(&trd->trd_list);
 352                 ted->ted_reply_last = trd;
 353         } else {
 354                 tgt_free_reply_data(lut, ted, trd);
 355         }
 356 }
 357
 358 static inline struct lu_buf *tti_buf_lsd(struct tgt_thread_info *tti)
 359 {
 360         tti->tti_buf.lb_buf = &tti->tti_lsd;
 361         tti->tti_buf.lb_len = sizeof(tti->tti_lsd);
 362         return &tti->tti_buf;
 363 }
 364
 365 static inline struct lu_buf *tti_buf_lcd(struct tgt_thread_info *tti)
 366 {
 367         tti->tti_buf.lb_buf = &tti->tti_lcd;
 368         tti->tti_buf.lb_len = sizeof(tti->tti_lcd);
 369         return &tti->tti_buf;
 370 }
 371
 372 /**
 373  * Allocate in-memory data for client slot related to export.
 374  */
 375 int tgt_client_alloc(struct obd_export *exp)
 376 {
 377         ENTRY;
 378         LASSERT(exp != exp->exp_obd->obd_self_export);
 379
 380         OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
 381         if (exp->exp_target_data.ted_lcd == NULL)
 382                 RETURN(-ENOMEM);
 383         /* Mark that slot is not yet valid, 0 doesn't work here */
 384         exp->exp_target_data.ted_lr_idx = -1;
 385         INIT_LIST_HEAD(&exp->exp_target_data.ted_reply_list);
 386         mutex_init(&exp->exp_target_data.ted_lcd_lock);
 387         RETURN(0);
 388 }
 389 EXPORT_SYMBOL(tgt_client_alloc);
 390
 391 /**
 392  * Free in-memory data for client slot related to export.
 393  */
 394 void tgt_client_free(struct obd_export *exp)
 395 {
 396         struct tg_export_data   *ted = &exp->exp_target_data;
 397         struct lu_target        *lut = class_exp2tgt(exp);
 398         struct tg_reply_data    *trd, *tmp;
 399
 400         LASSERT(exp != exp->exp_obd->obd_self_export);
 401
 402         /* free reply data */
 403         mutex_lock(&ted->ted_lcd_lock);
 404         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
 405                 tgt_release_reply_data(lut, ted, trd);
 406         }
 407         if (ted->ted_reply_last != NULL) {
 408                 tgt_free_reply_data(lut, ted, ted->ted_reply_last);
 409                 ted->ted_reply_last = NULL;
 410         }
 411         mutex_unlock(&ted->ted_lcd_lock);
 412
 413         if (!hlist_unhashed(&exp->exp_gen_hash))
 414                 cfs_hash_del(exp->exp_obd->obd_gen_hash,
 415                              &ted->ted_lcd->lcd_generation,
 416                              &exp->exp_gen_hash);
 417
 418         OBD_FREE_PTR(ted->ted_lcd);
 419         ted->ted_lcd = NULL;
 420
 421         /* Slot may be not yet assigned */
 422         if (ted->ted_lr_idx < 0)
 423                 return;
 424         /* Clear bit when lcd is freed */
 425         LASSERT(lut && lut->lut_client_bitmap);
 426         if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
 427                 CERROR("%s: client %u bit already clear in bitmap\n",
 428                        exp->exp_obd->obd_name, ted->ted_lr_idx);
 429                 LBUG();
 430         }
 431
 432         if (tgt_is_multimodrpcs_client(exp) && !exp->exp_obd->obd_stopping)
 433                 atomic_dec(&lut->lut_num_clients);
 434 }
 435 EXPORT_SYMBOL(tgt_client_free);
 436
 437 int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
 438                          struct lsd_client_data *lcd, loff_t *off, int index)
 439 {
 440         struct tgt_thread_info  *tti = tgt_th_info(env);
 441         int                      rc;
 442
 443         tti_buf_lcd(tti);
 444         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
 445         if (rc == 0) {
 446                 check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
 447                 lcd_le_to_cpu(&tti->tti_lcd, lcd);
 448                 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
 449                 lcd->lcd_last_close_result =
 450                         ptlrpc_status_ntoh(lcd->lcd_last_close_result);
 451         }
 452
 453         CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = "LPU64
 454                ", last_xid = "LPU64", last_result = %u, last_data = %u, "
 455                "last_close_transno = "LPU64", last_close_xid = "LPU64", "
 456                "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
 457                *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
 458                lcd->lcd_last_result, lcd->lcd_last_data,
 459                lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
 460                lcd->lcd_last_close_result, rc);
 461         return rc;
 462 }
 463
 464 int tgt_client_data_write(const struct lu_env *env, struct lu_target *tgt,
 465                           struct lsd_client_data *lcd, loff_t *off,
 466                           struct thandle *th)
 467 {
 468         struct tgt_thread_info *tti = tgt_th_info(env);
 469         struct dt_object        *dto;
 470
 471         lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
 472         lcd->lcd_last_close_result =
 473                 ptlrpc_status_hton(lcd->lcd_last_close_result);
 474         lcd_cpu_to_le(lcd, &tti->tti_lcd);
 475         tti_buf_lcd(tti);
 476
 477         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
 478         return dt_record_write(env, dto, &tti->tti_buf, off, th);
 479 }
 480
 481 /**
 482  * Update client data in last_rcvd
 483  */
 484 static int tgt_client_data_update(const struct lu_env *env,
 485                                   struct obd_export *exp)
 486 {
 487         struct tg_export_data   *ted = &exp->exp_target_data;
 488         struct lu_target        *tgt = class_exp2tgt(exp);
 489         struct tgt_thread_info  *tti = tgt_th_info(env);
 490         struct thandle          *th;
 491         int                      rc = 0;
 492
 493         ENTRY;
 494
 495         if (unlikely(tgt == NULL)) {
 496                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
 497                           class_exp2obd(exp)->obd_name);
 498                 RETURN(-EINVAL);
 499         }
 500
 501         th = dt_trans_create(env, tgt->lut_bottom);
 502         if (IS_ERR(th))
 503                 RETURN(PTR_ERR(th));
 504
 505         tti_buf_lcd(tti);
 506         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 507                                      &tti->tti_buf,
 508                                      ted->ted_lr_off, th);
 509         if (rc)
 510                 GOTO(out, rc);
 511
 512         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 513         if (rc)
 514                 GOTO(out, rc);
 515         /*
 516          * Until this operations will be committed the sync is needed
 517          * for this export. This should be done _after_ starting the
 518          * transaction so that many connecting clients will not bring
 519          * server down with lots of sync writes.
 520          */
 521         rc = tgt_new_client_cb_add(th, exp);
 522         if (rc) {
 523                 /* can't add callback, do sync now */
 524                 th->th_sync = 1;
 525         } else {
 526                 spin_lock(&exp->exp_lock);
 527                 exp->exp_need_sync = 1;
 528                 spin_unlock(&exp->exp_lock);
 529         }
 530
 531         tti->tti_off = ted->ted_lr_off;
 532         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
 533         EXIT;
 534 out:
 535         dt_trans_stop(env, tgt->lut_bottom, th);
 536         CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
 537                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 538                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 539
 540         return rc;
 541 }
 542
 543 int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
 544 {
 545         struct tgt_thread_info  *tti = tgt_th_info(env);
 546         int                      rc;
 547
 548         tti->tti_off = 0;
 549         tti_buf_lsd(tti);
 550         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
 551                             &tti->tti_off);
 552         if (rc == 0)
 553                 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
 554
 555         CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
 556                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 557                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 558         return rc;
 559 }
 560
 561 int tgt_server_data_write(const struct lu_env *env, struct lu_target *tgt,
 562                           struct thandle *th)
 563 {
 564         struct tgt_thread_info  *tti = tgt_th_info(env);
 565         struct dt_object        *dto;
 566         int                      rc;
 567
 568         ENTRY;
 569
 570         tti->tti_off = 0;
 571         tti_buf_lsd(tti);
 572         lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
 573
 574         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
 575         rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 576
 577         CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
 578                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 579                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 580
 581         RETURN(rc);
 582 }
 583
 584 /**
 585  * Update server data in last_rcvd
 586  */
 587 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
 588                            int sync)
 589 {
 590         struct tgt_thread_info  *tti = tgt_th_info(env);
 591         struct thandle          *th;
 592         int                      rc = 0;
 593
 594         ENTRY;
 595
 596         CDEBUG(D_SUPER,
 597                "%s: mount_count is "LPU64", last_transno is "LPU64"\n",
 598                tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
 599                tgt->lut_last_transno);
 600
 601         /* Always save latest transno to keep it fresh */
 602         spin_lock(&tgt->lut_translock);
 603         tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
 604         spin_unlock(&tgt->lut_translock);
 605
 606         th = dt_trans_create(env, tgt->lut_bottom);
 607         if (IS_ERR(th))
 608                 RETURN(PTR_ERR(th));
 609
 610         th->th_sync = sync;
 611
 612         tti_buf_lsd(tti);
 613         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 614                                      &tti->tti_buf, tti->tti_off, th);
 615         if (rc)
 616                 GOTO(out, rc);
 617
 618         rc = dt_trans_start(env, tgt->lut_bottom, th);
 619         if (rc)
 620                 GOTO(out, rc);
 621
 622         rc = tgt_server_data_write(env, tgt, th);
 623 out:
 624         dt_trans_stop(env, tgt->lut_bottom, th);
 625
 626         CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
 627                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 628                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 629         RETURN(rc);
 630 }
 631 EXPORT_SYMBOL(tgt_server_data_update);
 632
 633 int tgt_truncate_last_rcvd(const struct lu_env *env, struct lu_target *tgt,
 634                            loff_t size)
 635 {
 636         struct dt_object *dt = tgt->lut_last_rcvd;
 637         struct thandle   *th;
 638         struct lu_attr    attr;
 639         int               rc;
 640
 641         ENTRY;
 642
 643         attr.la_size = size;
 644         attr.la_valid = LA_SIZE;
 645
 646         th = dt_trans_create(env, tgt->lut_bottom);
 647         if (IS_ERR(th))
 648                 RETURN(PTR_ERR(th));
 649         rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
 650         if (rc)
 651                 GOTO(cleanup, rc);
 652         rc = dt_declare_attr_set(env, dt, &attr, th);
 653         if (rc)
 654                 GOTO(cleanup, rc);
 655         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 656         if (rc)
 657                 GOTO(cleanup, rc);
 658
 659         rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
 660         if (rc == 0)
 661                 rc = dt_attr_set(env, dt, &attr, th);
 662
 663 cleanup:
 664         dt_trans_stop(env, tgt->lut_bottom, th);
 665
 666         RETURN(rc);
 667 }
 668
 669 static void tgt_client_epoch_update(const struct lu_env *env,
 670                                     struct obd_export *exp)
 671 {
 672         struct lsd_client_data  *lcd = exp->exp_target_data.ted_lcd;
 673         struct lu_target        *tgt = class_exp2tgt(exp);
 674
 675         LASSERT(tgt && tgt->lut_bottom);
 676         /** VBR: set client last_epoch to current epoch */
 677         if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
 678                 return;
 679         lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
 680         tgt_client_data_update(env, exp);
 681 }
 682
 683 /**
 684  * Update boot epoch when recovery ends
 685  */
 686 void tgt_boot_epoch_update(struct lu_target *tgt)
 687 {
 688         struct lu_env            env;
 689         struct ptlrpc_request   *req;
 690         __u32                    start_epoch;
 691         struct list_head         client_list;
 692         int                      rc;
 693
 694         if (tgt->lut_obd->obd_stopping)
 695                 return;
 696
 697         rc = lu_env_init(&env, LCT_DT_THREAD);
 698         if (rc) {
 699                 CERROR("%s: can't initialize environment: rc = %d\n",
 700                         tgt->lut_obd->obd_name, rc);
 701                 return;
 702         }
 703
 704         spin_lock(&tgt->lut_translock);
 705         start_epoch = lr_epoch(tgt->lut_last_transno) + 1;
 706         tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
 707         tgt->lut_lsd.lsd_start_epoch = start_epoch;
 708         spin_unlock(&tgt->lut_translock);
 709
 710         INIT_LIST_HEAD(&client_list);
 711         /**
 712          * The recovery is not yet finished and final queue can still be updated
 713          * with resend requests. Move final list to separate one for processing
 714          */
 715         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
 716         list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
 717         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
 718
 719         /**
 720          * go through list of exports participated in recovery and
 721          * set new epoch for them
 722          */
 723         list_for_each_entry(req, &client_list, rq_list) {
 724                 LASSERT(!req->rq_export->exp_delayed);
 725                 if (!req->rq_export->exp_vbr_failed)
 726                         tgt_client_epoch_update(&env, req->rq_export);
 727         }
 728         /** return list back at once */
 729         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
 730         list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
 731         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
 732
 733         /** Clear MULTI RPCS incompatibility flag if
 734          * - target is MDT and
 735          * - there is no client to recover or the recovery was aborted
 736          */
 737         if (!strncmp(tgt->lut_obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) &&
 738             (tgt->lut_obd->obd_max_recoverable_clients == 0 ||
 739             tgt->lut_obd->obd_abort_recovery))
 740                 tgt->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
 741
 742         /** update server epoch */
 743         tgt_server_data_update(&env, tgt, 1);
 744         lu_env_fini(&env);
 745 }
 746
 747 /**
 748  * commit callback, need to update last_commited value
 749  */
 750 struct tgt_last_committed_callback {
 751         struct dt_txn_commit_cb  llcc_cb;
 752         struct lu_target        *llcc_tgt;
 753         struct obd_export       *llcc_exp;
 754         __u64                    llcc_transno;
 755 };
 756
 757 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
 758                                   struct dt_txn_commit_cb *cb, int err)
 759 {
 760         struct tgt_last_committed_callback *ccb;
 761
 762         ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
 763
 764         LASSERT(ccb->llcc_exp);
 765         LASSERT(ccb->llcc_tgt != NULL);
 766         LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
 767
 768         /* Fast path w/o spinlock, if exp_last_committed was updated
 769          * with higher transno, no need to take spinlock and check,
 770          * also no need to update obd_last_committed. */
 771         if (ccb->llcc_transno <= ccb->llcc_exp->exp_last_committed)
 772                 goto out;
 773         spin_lock(&ccb->llcc_tgt->lut_translock);
 774         if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
 775                 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
 776
 777         if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
 778                 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
 779                 spin_unlock(&ccb->llcc_tgt->lut_translock);
 780                 ptlrpc_commit_replies(ccb->llcc_exp);
 781         } else {
 782                 spin_unlock(&ccb->llcc_tgt->lut_translock);
 783         }
 784 out:
 785         class_export_cb_put(ccb->llcc_exp);
 786         if (ccb->llcc_transno)
 787                 CDEBUG(D_HA, "%s: transno "LPD64" is committed\n",
 788                        ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
 789         OBD_FREE_PTR(ccb);
 790 }
 791
 792 int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
 793                            struct obd_export *exp, __u64 transno)
 794 {
 795         struct tgt_last_committed_callback      *ccb;
 796         struct dt_txn_commit_cb                 *dcb;
 797         int                                      rc;
 798
 799         OBD_ALLOC_PTR(ccb);
 800         if (ccb == NULL)
 801                 return -ENOMEM;
 802
 803         ccb->llcc_tgt = tgt;
 804         ccb->llcc_exp = class_export_cb_get(exp);
 805         ccb->llcc_transno = transno;
 806
 807         dcb = &ccb->llcc_cb;
 808         dcb->dcb_func = tgt_cb_last_committed;
 809         INIT_LIST_HEAD(&dcb->dcb_linkage);
 810         strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
 811
 812         rc = dt_trans_cb_add(th, dcb);
 813         if (rc) {
 814                 class_export_cb_put(exp);
 815                 OBD_FREE_PTR(ccb);
 816         }
 817
 818         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 819                 /* report failure to force synchronous operation */
 820                 return -EPERM;
 821
 822         return rc;
 823 }
 824
 825 struct tgt_new_client_callback {
 826         struct dt_txn_commit_cb  lncc_cb;
 827         struct obd_export       *lncc_exp;
 828 };
 829
 830 static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
 831                               struct dt_txn_commit_cb *cb, int err)
 832 {
 833         struct tgt_new_client_callback *ccb;
 834
 835         ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
 836
 837         LASSERT(ccb->lncc_exp->exp_obd);
 838
 839         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 840                ccb->lncc_exp->exp_obd->obd_name,
 841                ccb->lncc_exp->exp_client_uuid.uuid);
 842
 843         spin_lock(&ccb->lncc_exp->exp_lock);
 844
 845         ccb->lncc_exp->exp_need_sync = 0;
 846
 847         spin_unlock(&ccb->lncc_exp->exp_lock);
 848         class_export_cb_put(ccb->lncc_exp);
 849
 850         OBD_FREE_PTR(ccb);
 851 }
 852
 853 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
 854 {
 855         struct tgt_new_client_callback  *ccb;
 856         struct dt_txn_commit_cb         *dcb;
 857         int                              rc;
 858
 859         OBD_ALLOC_PTR(ccb);
 860         if (ccb == NULL)
 861                 return -ENOMEM;
 862
 863         ccb->lncc_exp = class_export_cb_get(exp);
 864
 865         dcb = &ccb->lncc_cb;
 866         dcb->dcb_func = tgt_cb_new_client;
 867         INIT_LIST_HEAD(&dcb->dcb_linkage);
 868         strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
 869
 870         rc = dt_trans_cb_add(th, dcb);
 871         if (rc) {
 872                 class_export_cb_put(exp);
 873                 OBD_FREE_PTR(ccb);
 874         }
 875         return rc;
 876 }
 877
 878 /**
 879  * Add new client to the last_rcvd upon new connection.
 880  *
 881  * We use a bitmap to locate a free space in the last_rcvd file and initialize
 882  * tg_export_data.
 883  */
 884 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
 885 {
 886         struct tg_export_data   *ted = &exp->exp_target_data;
 887         struct lu_target        *tgt = class_exp2tgt(exp);
 888         int                      rc = 0, idx;
 889
 890         ENTRY;
 891
 892         LASSERT(tgt && tgt->lut_client_bitmap != NULL);
 893         if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
 894                 RETURN(0);
 895
 896         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 897                 RETURN(0);
 898
 899         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 900          * there's no need for extra complication here
 901          */
 902         idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
 903 repeat:
 904         if (idx >= LR_MAX_CLIENTS ||
 905             OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
 906                 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
 907                        tgt->lut_obd->obd_name,  idx);
 908                 RETURN(-EOVERFLOW);
 909         }
 910         if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
 911                 idx = find_next_zero_bit(tgt->lut_client_bitmap,
 912                                              LR_MAX_CLIENTS, idx);
 913                 goto repeat;
 914         }
 915
 916         ted->ted_lr_idx = idx;
 917         ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
 918                           idx * tgt->lut_lsd.lsd_client_size;
 919
 920         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
 921
 922         if (tgt_is_multimodrpcs_client(exp)) {
 923                 /* Set MULTI RPCS incompatibility flag to prevent previous
 924                  * Lustre versions to mount a target with reply_data file */
 925                 atomic_inc(&tgt->lut_num_clients);
 926                 if (!(tgt->lut_lsd.lsd_feature_incompat &
 927                       OBD_INCOMPAT_MULTI_RPCS)) {
 928                         tgt->lut_lsd.lsd_feature_incompat |=
 929                                                         OBD_INCOMPAT_MULTI_RPCS;
 930                         rc = tgt_server_data_update(env, tgt, 1);
 931                         if (rc < 0) {
 932                                 CERROR("%s: unable to set MULTI RPCS "
 933                                        "incompatibility flag\n",
 934                                        exp->exp_obd->obd_name);
 935                                 RETURN(rc);
 936                         }
 937                 }
 938
 939                 /* assign client slot generation */
 940                 ted->ted_lcd->lcd_generation =
 941                                 atomic_inc_return(&tgt->lut_client_generation);
 942         } else {
 943                 ted->ted_lcd->lcd_generation = 0;
 944         }
 945
 946         CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s' "
 947                "generation %d\n",
 948                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
 949                ted->ted_lcd->lcd_uuid, ted->ted_lcd->lcd_generation);
 950
 951         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
 952                 RETURN(-ENOSPC);
 953
 954         rc = tgt_client_data_update(env, exp);
 955         if (rc)
 956                 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
 957                        tgt->lut_obd->obd_name, idx, rc);
 958
 959         RETURN(rc);
 960 }
 961 EXPORT_SYMBOL(tgt_client_new);
 962
 963 /* Add an existing client to the MDS in-memory state based on
 964  * a client that was previously found in the last_rcvd file and
 965  * already has an assigned slot (idx >= 0).
 966  *
 967  * It should not be possible to fail adding an existing client - otherwise
 968  * mdt_init_server_data() callsite needs to be fixed.
 969  */
 970 int tgt_client_add(const struct lu_env *env,  struct obd_export *exp, int idx)
 971 {
 972         struct tg_export_data   *ted = &exp->exp_target_data;
 973         struct lu_target        *tgt = class_exp2tgt(exp);
 974
 975         ENTRY;
 976
 977         LASSERT(tgt && tgt->lut_client_bitmap != NULL);
 978         LASSERTF(idx >= 0, "%d\n", idx);
 979
 980         if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
 981             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 982                 RETURN(0);
 983
 984         if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
 985                 CERROR("%s: client %d: bit already set in bitmap!!\n",
 986                        tgt->lut_obd->obd_name,  idx);
 987                 LBUG();
 988         }
 989         atomic_inc(&tgt->lut_num_clients);
 990
 991         CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added, "
 992                "generation %d\n",
 993                tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid,
 994                ted->ted_lcd->lcd_generation);
 995
 996         ted->ted_lr_idx = idx;
 997         ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
 998                           idx * tgt->lut_lsd.lsd_client_size;
 999
1000         mutex_init(&ted->ted_lcd_lock);
1001
1002         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
1003
1004         RETURN(0);
1005 }
1006
1007 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
1008 {
1009         struct tg_export_data   *ted = &exp->exp_target_data;
1010         struct lu_target        *tgt = class_exp2tgt(exp);
1011         int                      rc;
1012
1013         ENTRY;
1014
1015         LASSERT(ted->ted_lcd);
1016
1017         if (unlikely(tgt == NULL)) {
1018                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
1019                        class_exp2obd(exp)->obd_name);
1020                 RETURN(-EINVAL);
1021         }
1022
1023         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
1024         if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
1025                     (char *)tgt->lut_obd->obd_uuid.uuid) ||
1026             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
1027                 RETURN(0);
1028
1029         CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
1030                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
1031                ted->ted_lcd->lcd_uuid);
1032
1033         /* Clear the bit _after_ zeroing out the client so we don't
1034            race with filter_client_add and zero out new clients.*/
1035         if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
1036                 CERROR("%s: client %u: bit already clear in bitmap!!\n",
1037                        tgt->lut_obd->obd_name, ted->ted_lr_idx);
1038                 LBUG();
1039         }
1040
1041         /* Do not erase record for recoverable client. */
1042         if (exp->exp_flags & OBD_OPT_FAILOVER)
1043                 RETURN(0);
1044
1045         /* Make sure the server's last_transno is up to date.
1046          * This should be done before zeroing client slot so last_transno will
1047          * be in server data or in client data in case of failure */
1048         rc = tgt_server_data_update(env, tgt, 0);
1049         if (rc != 0) {
1050                 CERROR("%s: failed to update server data, skip client %s "
1051                        "zeroing, rc %d\n", tgt->lut_obd->obd_name,
1052                        ted->ted_lcd->lcd_uuid, rc);
1053                 RETURN(rc);
1054         }
1055
1056         mutex_lock(&ted->ted_lcd_lock);
1057         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
1058         rc = tgt_client_data_update(env, exp);
1059         mutex_unlock(&ted->ted_lcd_lock);
1060
1061         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
1062                "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
1063                tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
1064                ted->ted_lr_idx, ted->ted_lr_off, rc);
1065         RETURN(rc);
1066 }
1067 EXPORT_SYMBOL(tgt_client_del);
1068
1069 int tgt_add_reply_data(const struct lu_env *env, struct lu_target *tgt,
1070                        struct tg_export_data *ted, struct tg_reply_data *trd,
1071                        struct thandle *th, bool update_lrd_file)
1072 {
1073         struct lsd_reply_data   *lrd;
1074         int     i;
1075
1076         lrd = &trd->trd_reply;
1077         /* update export last transno */
1078         mutex_lock(&ted->ted_lcd_lock);
1079         if (lrd->lrd_transno > ted->ted_lcd->lcd_last_transno)
1080                 ted->ted_lcd->lcd_last_transno = lrd->lrd_transno;
1081         mutex_unlock(&ted->ted_lcd_lock);
1082
1083         /* find a empty slot */
1084         i = tgt_find_free_reply_slot(tgt);
1085         if (unlikely(i < 0)) {
1086                 CERROR("%s: couldn't find a slot for reply data: "
1087                        "rc = %d\n", tgt_name(tgt), i);
1088                 RETURN(i);
1089         }
1090         trd->trd_index = i;
1091
1092         if (update_lrd_file) {
1093                 loff_t  off;
1094                 int     rc;
1095
1096                 /* write reply data to disk */
1097                 off = sizeof(struct lsd_reply_header) + sizeof(*lrd) * i;
1098                 rc = tgt_reply_data_write(env, tgt, lrd, off, th);
1099                 if (unlikely(rc != 0)) {
1100                         CERROR("%s: can't update %s file: rc = %d\n",
1101                                tgt_name(tgt), REPLY_DATA, rc);
1102                         RETURN(rc);
1103                 }
1104         }
1105         /* add reply data to target export's reply list */
1106         mutex_lock(&ted->ted_lcd_lock);
1107         list_add(&trd->trd_list, &ted->ted_reply_list);
1108         ted->ted_reply_cnt++;
1109         if (ted->ted_reply_cnt > ted->ted_reply_max)
1110                 ted->ted_reply_max = ted->ted_reply_cnt;
1111         mutex_unlock(&ted->ted_lcd_lock);
1112
1113         CDEBUG(D_TRACE, "add reply %p: xid %llu, transno %llu, "
1114                "tag %hu, client gen %u, slot idx %d\n",
1115                trd, lrd->lrd_xid, lrd->lrd_transno,
1116                trd->trd_tag, lrd->lrd_client_gen, i);
1117         RETURN(0);
1118 }
1119 EXPORT_SYMBOL(tgt_add_reply_data);
1120
1121 /*
1122  * last_rcvd & last_committed update callbacks
1123  */
1124 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
1125                                 struct dt_object *obj, __u64 opdata,
1126                                 struct thandle *th, struct ptlrpc_request *req)
1127 {
1128         struct tgt_thread_info  *tti = tgt_th_info(env);
1129         struct tgt_session_info *tsi = tgt_ses_info(env);
1130         struct obd_export       *exp = tsi->tsi_exp;
1131         struct tg_export_data   *ted;
1132         __u64                   *transno_p;
1133         int                      rc = 0;
1134         bool                     lw_client;
1135
1136         ENTRY;
1137
1138
1139         LASSERT(exp != NULL);
1140         ted = &exp->exp_target_data;
1141
1142         lw_client = exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT;
1143         if (ted->ted_lr_idx < 0 && !lw_client)
1144                 /* ofd connect may cause transaction before export has
1145                  * last_rcvd slot */
1146                 RETURN(0);
1147
1148         if (req != NULL)
1149                 tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
1150         else
1151                 /* From update replay, tti_transno should be set already */
1152                 LASSERT(tti->tti_transno != 0);
1153
1154         spin_lock(&tgt->lut_translock);
1155         if (th->th_result != 0) {
1156                 if (tti->tti_transno != 0) {
1157                         CERROR("%s: replay transno "LPU64" failed: rc = %d\n",
1158                                tgt_name(tgt), tti->tti_transno, th->th_result);
1159                 }
1160         } else if (tti->tti_transno == 0) {
1161                 tti->tti_transno = ++tgt->lut_last_transno;
1162         } else {
1163                 /* should be replay */
1164                 if (tti->tti_transno > tgt->lut_last_transno)
1165                         tgt->lut_last_transno = tti->tti_transno;
1166         }
1167         spin_unlock(&tgt->lut_translock);
1168
1169         /** VBR: set new versions */
1170         if (th->th_result == 0 && obj != NULL) {
1171                 struct dt_object *dto = dt_object_locate(obj, th->th_dev);
1172                 dt_version_set(env, dto, tti->tti_transno, th);
1173         }
1174
1175         /* filling reply data */
1176         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
1177                tti->tti_transno, tgt->lut_obd->obd_last_committed);
1178
1179         if (req != NULL) {
1180                 req->rq_transno = tti->tti_transno;
1181                 lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
1182         }
1183
1184         /* if can't add callback, do sync write */
1185         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp, tti->tti_transno);
1186
1187         if (lw_client) {
1188                 /* All operations performed by LW clients are synchronous and
1189                  * we store the committed transno in the last_rcvd header */
1190                 spin_lock(&tgt->lut_translock);
1191                 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
1192                         tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
1193                         spin_unlock(&tgt->lut_translock);
1194                         /* Although lightweight (LW) connections have no slot
1195                          * in the last_rcvd, we still want to maintain
1196                          * the in-memory lsd_client_data structure in order to
1197                          * properly handle reply reconstruction. */
1198                         rc = tgt_server_data_write(env, tgt, th);
1199                 } else {
1200                         spin_unlock(&tgt->lut_translock);
1201                 }
1202         } else if (ted->ted_lr_off == 0) {
1203                 CERROR("%s: client idx %d has offset %lld\n",
1204                        tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
1205                 RETURN(-EINVAL);
1206         }
1207
1208         /* Target that supports multiple reply data */
1209         if (tgt_is_multimodrpcs_client(exp)) {
1210                 struct tg_reply_data    *trd;
1211                 struct lsd_reply_data   *lrd;
1212                 __u64                   *pre_versions;
1213                 bool                    write_update;
1214
1215                 OBD_ALLOC_PTR(trd);
1216                 if (unlikely(trd == NULL))
1217                         RETURN(-ENOMEM);
1218
1219                 /* fill reply data information */
1220                 lrd = &trd->trd_reply;
1221                 lrd->lrd_transno = tti->tti_transno;
1222                 if (req != NULL) {
1223                         lrd->lrd_xid = req->rq_xid;
1224                         trd->trd_tag = lustre_msg_get_tag(req->rq_reqmsg);
1225                         pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1226                         lrd->lrd_result = th->th_result;
1227                         lrd->lrd_client_gen = ted->ted_lcd->lcd_generation;
1228                         write_update = true;
1229                 } else {
1230                         LASSERT(tsi->tsi_xid != 0);
1231                         lrd->lrd_xid = tsi->tsi_xid;
1232                         lrd->lrd_result = tsi->tsi_result;
1233                         lrd->lrd_client_gen = tsi->tsi_client_gen;
1234                         trd->trd_tag = 0;
1235                         pre_versions = NULL;
1236                         write_update = false;
1237                 }
1238
1239                 lrd->lrd_data = opdata;
1240                 if (pre_versions) {
1241                         trd->trd_pre_versions[0] = pre_versions[0];
1242                         trd->trd_pre_versions[1] = pre_versions[1];
1243                         trd->trd_pre_versions[2] = pre_versions[2];
1244                         trd->trd_pre_versions[3] = pre_versions[3];
1245                 }
1246
1247                 rc = tgt_add_reply_data(env, tgt, ted, trd, th, write_update);
1248                 if (rc < 0)
1249                         OBD_FREE_PTR(trd);
1250                 return rc;
1251         }
1252
1253         /* Enough for update replay, let's return */
1254         if (req == NULL)
1255                 RETURN(rc);
1256
1257         mutex_lock(&ted->ted_lcd_lock);
1258         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1259         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
1260                 transno_p = &ted->ted_lcd->lcd_last_close_transno;
1261                 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
1262                 ted->ted_lcd->lcd_last_close_result = th->th_result;
1263         } else {
1264                 /* VBR: save versions in last_rcvd for reconstruct. */
1265                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
1266
1267                 if (pre_versions) {
1268                         ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
1269                         ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
1270                         ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
1271                         ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
1272                 }
1273                 transno_p = &ted->ted_lcd->lcd_last_transno;
1274                 ted->ted_lcd->lcd_last_xid = req->rq_xid;
1275                 ted->ted_lcd->lcd_last_result = th->th_result;
1276                 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
1277                  * see struct ldlm_reply->lock_policy_res1; */
1278                 ted->ted_lcd->lcd_last_data = opdata;
1279         }
1280
1281         /* Update transno in slot only if non-zero number, i.e. no errors */
1282         if (likely(tti->tti_transno != 0)) {
1283                 /* Don't overwrite bigger transaction number with lower one.
1284                  * That is not sign of problem in all cases, but in any case
1285                  * this value should be monotonically increased only. */
1286                 if (*transno_p > tti->tti_transno) {
1287                         if (!tgt->lut_no_reconstruct) {
1288                                 CERROR("%s: trying to overwrite bigger transno:"
1289                                        "on-disk: "LPU64", new: "LPU64" replay: "
1290                                        "%d. See LU-617.\n", tgt_name(tgt),
1291                                        *transno_p, tti->tti_transno,
1292                                        req_is_replay(req));
1293                                 if (req_is_replay(req)) {
1294                                         spin_lock(&req->rq_export->exp_lock);
1295                                         req->rq_export->exp_vbr_failed = 1;
1296                                         spin_unlock(&req->rq_export->exp_lock);
1297                                 }
1298                                 mutex_unlock(&ted->ted_lcd_lock);
1299                                 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
1300                         }
1301                 } else {
1302                         *transno_p = tti->tti_transno;
1303                 }
1304         }
1305
1306         if (!lw_client) {
1307                 tti->tti_off = ted->ted_lr_off;
1308                 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1309                 if (rc < 0) {
1310                         mutex_unlock(&ted->ted_lcd_lock);
1311                         RETURN(rc);
1312                 }
1313         }
1314         mutex_unlock(&ted->ted_lcd_lock);
1315         RETURN(rc);
1316 }
1317
1318 /*
1319  * last_rcvd update for echo client simulation.
1320  * It updates last_rcvd client slot and version of object in
1321  * simple way but with all locks to simulate all drawbacks
1322  */
1323 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
1324                                      struct lu_target *tgt,
1325                                      struct dt_object *obj,
1326                                      struct thandle *th,
1327                                      struct obd_export *exp)
1328 {
1329         struct tgt_thread_info  *tti = tgt_th_info(env);
1330         struct tg_export_data   *ted = &exp->exp_target_data;
1331         int                      rc = 0;
1332
1333         ENTRY;
1334
1335         tti->tti_transno = 0;
1336
1337         spin_lock(&tgt->lut_translock);
1338         if (th->th_result == 0)
1339                 tti->tti_transno = ++tgt->lut_last_transno;
1340         spin_unlock(&tgt->lut_translock);
1341
1342         /** VBR: set new versions */
1343         if (th->th_result == 0 && obj != NULL)
1344                 dt_version_set(env, obj, tti->tti_transno, th);
1345
1346         /* if can't add callback, do sync write */
1347         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
1348                                                 tti->tti_transno);
1349
1350         LASSERT(ted->ted_lr_off > 0);
1351
1352         mutex_lock(&ted->ted_lcd_lock);
1353         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
1354         ted->ted_lcd->lcd_last_transno = tti->tti_transno;
1355         ted->ted_lcd->lcd_last_result = th->th_result;
1356
1357         tti->tti_off = ted->ted_lr_off;
1358         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
1359         mutex_unlock(&ted->ted_lcd_lock);
1360         RETURN(rc);
1361 }
1362
1363 static int tgt_clients_data_init(const struct lu_env *env,
1364                                  struct lu_target *tgt,
1365                                  unsigned long last_size)
1366 {
1367         struct obd_device       *obd = tgt->lut_obd;
1368         struct lr_server_data   *lsd = &tgt->lut_lsd;
1369         struct lsd_client_data  *lcd = NULL;
1370         struct tg_export_data   *ted;
1371         int                      cl_idx;
1372         int                      rc = 0;
1373         loff_t                   off = lsd->lsd_client_start;
1374         __u32                    generation = 0;
1375         struct cfs_hash         *hash = NULL;
1376
1377         ENTRY;
1378
1379         CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
1380                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
1381
1382         OBD_ALLOC_PTR(lcd);
1383         if (lcd == NULL)
1384                 RETURN(-ENOMEM);
1385
1386         hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1387         if (hash == NULL)
1388                 GOTO(err_out, rc = -ENODEV);
1389
1390         for (cl_idx = 0; off < last_size; cl_idx++) {
1391                 struct obd_export       *exp;
1392                 __u64                    last_transno;
1393
1394                 /* Don't assume off is incremented properly by
1395                  * read_record(), in case sizeof(*lcd)
1396                  * isn't the same as fsd->lsd_client_size.  */
1397                 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
1398                 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
1399                 if (rc) {
1400                         CERROR("%s: error reading last_rcvd %s idx %d off "
1401                                "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
1402                                cl_idx, off, rc);
1403                         rc = 0;
1404                         break; /* read error shouldn't cause startup to fail */
1405                 }
1406
1407                 if (lcd->lcd_uuid[0] == '\0') {
1408                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
1409                                cl_idx);
1410                         continue;
1411                 }
1412
1413                 last_transno = lcd_last_transno(lcd);
1414
1415                 /* These exports are cleaned up by disconnect, so they
1416                  * need to be set up like real exports as connect does.
1417                  */
1418                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
1419                        " srv lr: "LPU64" lx: "LPU64" gen %u\n", lcd->lcd_uuid,
1420                        cl_idx, last_transno, lsd->lsd_last_transno,
1421                        lcd_last_xid(lcd), lcd->lcd_generation);
1422
1423                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
1424                 if (IS_ERR(exp)) {
1425                         if (PTR_ERR(exp) == -EALREADY) {
1426                                 /* export already exists, zero out this one */
1427                                 CERROR("%s: Duplicate export %s!\n",
1428                                        tgt_name(tgt), lcd->lcd_uuid);
1429                                 continue;
1430                         }
1431                         GOTO(err_out, rc = PTR_ERR(exp));
1432                 }
1433
1434                 ted = &exp->exp_target_data;
1435                 *ted->ted_lcd = *lcd;
1436
1437                 rc = tgt_client_add(env, exp, cl_idx);
1438                 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
1439                 /* VBR: set export last committed version */
1440                 exp->exp_last_committed = last_transno;
1441                 spin_lock(&exp->exp_lock);
1442                 exp->exp_connecting = 0;
1443                 exp->exp_in_recovery = 0;
1444                 spin_unlock(&exp->exp_lock);
1445                 obd->obd_max_recoverable_clients++;
1446
1447                 if (tgt->lut_lsd.lsd_feature_incompat &
1448                     OBD_INCOMPAT_MULTI_RPCS &&
1449                     lcd->lcd_generation != 0) {
1450                         /* compute the highest valid client generation */
1451                         generation = max(generation, lcd->lcd_generation);
1452                         /* fill client_generation <-> export hash table */
1453                         rc = cfs_hash_add_unique(hash, &lcd->lcd_generation,
1454                                                  &exp->exp_gen_hash);
1455                         if (rc != 0) {
1456                                 CERROR("%s: duplicate export for client "
1457                                        "generation %u\n",
1458                                        tgt_name(tgt), lcd->lcd_generation);
1459                                 class_export_put(exp);
1460                                 GOTO(err_out, rc);
1461                         }
1462                 }
1463
1464                 class_export_put(exp);
1465
1466                 rc = rev_import_init(exp);
1467                 if (rc != 0) {
1468                         class_unlink_export(exp);
1469                         GOTO(err_out, rc);
1470                 }
1471
1472                 /* Need to check last_rcvd even for duplicated exports. */
1473                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
1474                        cl_idx, last_transno);
1475
1476                 spin_lock(&tgt->lut_translock);
1477                 tgt->lut_last_transno = max(last_transno,
1478                                             tgt->lut_last_transno);
1479                 spin_unlock(&tgt->lut_translock);
1480         }
1481
1482         /* record highest valid client generation */
1483         atomic_set(&tgt->lut_client_generation, generation);
1484
1485 err_out:
1486         if (hash != NULL)
1487                 cfs_hash_putref(hash);
1488         OBD_FREE_PTR(lcd);
1489         RETURN(rc);
1490 }
1491
1492 struct server_compat_data {
1493         __u32 rocompat;
1494         __u32 incompat;
1495         __u32 rocinit;
1496         __u32 incinit;
1497 };
1498
1499 static struct server_compat_data tgt_scd[] = {
1500         [LDD_F_SV_TYPE_MDT] = {
1501                 .rocompat = OBD_ROCOMPAT_LOVOBJID,
1502                 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1503                             OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
1504                             OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI |
1505                             OBD_INCOMPAT_MULTI_RPCS,
1506                 .rocinit = OBD_ROCOMPAT_LOVOBJID,
1507                 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
1508                            OBD_INCOMPAT_MULTI_OI,
1509         },
1510         [LDD_F_SV_TYPE_OST] = {
1511                 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
1512                 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
1513                             OBD_INCOMPAT_FID,
1514                 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
1515                 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1516         }
1517 };
1518
1519 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1520 {
1521         struct tgt_thread_info          *tti = tgt_th_info(env);
1522         struct lr_server_data           *lsd = &tgt->lut_lsd;
1523         unsigned long                    last_rcvd_size;
1524         __u32                            index;
1525         int                              rc, type;
1526
1527         rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1528         if (rc)
1529                 RETURN(rc);
1530
1531         last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1532
1533         /* ensure padding in the struct is the correct size */
1534         CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
1535                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
1536
1537         rc = server_name2index(tgt_name(tgt), &index, NULL);
1538         if (rc < 0) {
1539                 CERROR("%s: Can not get index from name: rc = %d\n",
1540                        tgt_name(tgt), rc);
1541                 RETURN(rc);
1542         }
1543         /* server_name2index() returns type */
1544         type = rc;
1545         if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1546                 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1547                 RETURN(-EINVAL);
1548         }
1549
1550         /* last_rcvd on OST doesn't provide reconstruct support because there
1551          * may be up to 8 in-flight write requests per single slot in
1552          * last_rcvd client data
1553          */
1554         tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1555
1556         if (last_rcvd_size == 0) {
1557                 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1558
1559                 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1560                        sizeof(lsd->lsd_uuid));
1561                 lsd->lsd_last_transno = 0;
1562                 lsd->lsd_mount_count = 0;
1563                 lsd->lsd_server_size = LR_SERVER_SIZE;
1564                 lsd->lsd_client_start = LR_CLIENT_START;
1565                 lsd->lsd_client_size = LR_CLIENT_SIZE;
1566                 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1567                 lsd->lsd_osd_index = index;
1568                 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1569                 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1570         } else {
1571                 rc = tgt_server_data_read(env, tgt);
1572                 if (rc) {
1573                         CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1574                                tgt_name(tgt), rc);
1575                         RETURN(rc);
1576                 }
1577                 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1578                         LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
1579                                            "using the wrong disk %s. Were the"
1580                                            " /dev/ assignments rearranged?\n",
1581                                            tgt->lut_obd->obd_uuid.uuid,
1582                                            lsd->lsd_uuid);
1583                         RETURN(-EINVAL);
1584                 }
1585
1586                 if (lsd->lsd_osd_index != index) {
1587                         LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
1588                                            "is different with the index %d in"
1589                                            "config log, It might be disk"
1590                                            "corruption!\n", tgt_name(tgt),
1591                                            lsd->lsd_osd_index, index);
1592                         RETURN(-EINVAL);
1593                 }
1594         }
1595
1596         if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1597                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1598                        tgt_name(tgt),
1599                        lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1600                 RETURN(-EINVAL);
1601         }
1602
1603         if (type == LDD_F_SV_TYPE_MDT)
1604                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1605
1606         if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1607                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1608                        tgt_name(tgt),
1609                        lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1610                 RETURN(-EINVAL);
1611         }
1612         /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1613         if (type == LDD_F_SV_TYPE_MDT &&
1614             !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1615                 if (last_rcvd_size > lsd->lsd_client_start) {
1616                         LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1617                                       "remove all clients for interop needs\n",
1618                                       tgt_name(tgt));
1619                         rc = tgt_truncate_last_rcvd(env, tgt,
1620                                                     lsd->lsd_client_start);
1621                         if (rc)
1622                                 RETURN(rc);
1623                         last_rcvd_size = lsd->lsd_client_start;
1624                 }
1625                 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1626                 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1627         }
1628
1629         spin_lock(&tgt->lut_translock);
1630         tgt->lut_last_transno = lsd->lsd_last_transno;
1631         spin_unlock(&tgt->lut_translock);
1632
1633         lsd->lsd_mount_count++;
1634
1635         CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1636         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
1637                tgt_name(tgt), tgt->lut_last_transno);
1638         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
1639                tgt_name(tgt), lsd->lsd_mount_count);
1640         CDEBUG(D_INODE, "%s: server data size: %u\n",
1641                tgt_name(tgt), lsd->lsd_server_size);
1642         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1643                tgt_name(tgt), lsd->lsd_client_start);
1644         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1645                tgt_name(tgt), lsd->lsd_client_size);
1646         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1647                tgt_name(tgt), last_rcvd_size);
1648         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1649                tgt_name(tgt), lsd->lsd_subdir_count);
1650         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1651                last_rcvd_size <= lsd->lsd_client_start ? 0 :
1652                (last_rcvd_size - lsd->lsd_client_start) /
1653                 lsd->lsd_client_size);
1654         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1655
1656         if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1657             lsd->lsd_client_size == 0) {
1658                 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1659                 RETURN(-EINVAL);
1660         }
1661
1662         if (!tgt->lut_obd->obd_replayable)
1663                 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1664
1665         rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1666         if (rc < 0)
1667                 GOTO(err_client, rc);
1668
1669         spin_lock(&tgt->lut_translock);
1670         /* obd_last_committed is used for compatibility
1671          * with other lustre recovery code */
1672         tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1673         spin_unlock(&tgt->lut_translock);
1674
1675         tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1676         tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1677
1678         /* save it, so mount count and last_transno is current */
1679         rc = tgt_server_data_update(env, tgt, 0);
1680         if (rc < 0)
1681                 GOTO(err_client, rc);
1682
1683         RETURN(0);
1684
1685 err_client:
1686         class_disconnect_exports(tgt->lut_obd);
1687         return rc;
1688 }
1689
1690 /* add credits for last_rcvd update */
1691 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1692                      void *cookie)
1693 {
1694         struct lu_target        *tgt = cookie;
1695         struct tgt_session_info *tsi;
1696         struct tgt_thread_info  *tti = tgt_th_info(env);
1697         struct dt_object        *dto;
1698         int                      rc;
1699
1700         /* if there is no session, then this transaction is not result of
1701          * request processing but some local operation */
1702         if (env->le_ses == NULL)
1703                 return 0;
1704
1705         LASSERT(tgt->lut_last_rcvd);
1706         tsi = tgt_ses_info(env);
1707         /* OFD may start transaction without export assigned */
1708         if (tsi->tsi_exp == NULL)
1709                 return 0;
1710
1711         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
1712         tti_buf_lcd(tti);
1713
1714         rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1715                                      tsi->tsi_exp->exp_target_data.ted_lr_off,
1716                                      th);
1717         if (rc)
1718                 return rc;
1719
1720         tti_buf_lsd(tti);
1721         rc = dt_declare_record_write(env, dto, &tti->tti_buf, 0, th);
1722         if (rc)
1723                 return rc;
1724
1725         if (tsi->tsi_vbr_obj != NULL &&
1726             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1727                 dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
1728                 rc = dt_declare_version_set(env, dto, th);
1729         }
1730
1731         return rc;
1732 }
1733
1734 /* Update last_rcvd records with latests transaction data */
1735 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1736                     void *cookie)
1737 {
1738         struct lu_target        *tgt = cookie;
1739         struct tgt_session_info *tsi;
1740         struct tgt_thread_info  *tti = tgt_th_info(env);
1741         struct dt_object        *obj = NULL;
1742         int                      rc;
1743         bool                     echo_client;
1744
1745         if (env->le_ses == NULL)
1746                 return 0;
1747
1748         tsi = tgt_ses_info(env);
1749         /* OFD may start transaction without export assigned */
1750         if (tsi->tsi_exp == NULL)
1751                 return 0;
1752
1753         echo_client = (tgt_ses_req(tsi) == NULL && tsi->tsi_xid == 0);
1754
1755         if (tti->tti_has_trans && !echo_client) {
1756                 if (tti->tti_mult_trans == 0) {
1757                         CDEBUG(D_HA, "More than one transaction "LPU64"\n",
1758                                tti->tti_transno);
1759                         RETURN(0);
1760                 }
1761                 /* we need another transno to be assigned */
1762                 tti->tti_transno = 0;
1763         } else if (th->th_result == 0) {
1764                 tti->tti_has_trans = 1;
1765         }
1766
1767         if (tsi->tsi_vbr_obj != NULL &&
1768             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1769                 obj = tsi->tsi_vbr_obj;
1770         }
1771
1772         if (unlikely(echo_client)) /* echo client special case */
1773                 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1774                                                tsi->tsi_exp);
1775         else
1776                 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
1777                                           tgt_ses_req(tsi));
1778         return rc;
1779 }
1780
1781 int tgt_reply_data_init(const struct lu_env *env, struct lu_target *tgt)
1782 {
1783         struct tgt_thread_info  *tti = tgt_th_info(env);
1784         struct lsd_reply_data   *lrd = &tti->tti_lrd;
1785         unsigned long            reply_data_size;
1786         int                      rc;
1787         struct lsd_reply_header *lrh = NULL;
1788         struct lsd_client_data  *lcd = NULL;
1789         struct tg_reply_data    *trd = NULL;
1790         int                      idx;
1791         loff_t                   off;
1792         struct cfs_hash         *hash = NULL;
1793         struct obd_export       *exp;
1794         struct tg_export_data   *ted;
1795         int                      reply_data_recovered = 0;
1796
1797         rc = dt_attr_get(env, tgt->lut_reply_data, &tti->tti_attr);
1798         if (rc)
1799                 GOTO(out, rc);
1800         reply_data_size = (unsigned long)tti->tti_attr.la_size;
1801
1802         OBD_ALLOC_PTR(lrh);
1803         if (lrh == NULL)
1804                 GOTO(out, rc = -ENOMEM);
1805
1806         if (reply_data_size == 0) {
1807                 CDEBUG(D_INFO, "%s: new reply_data file, initializing\n",
1808                        tgt_name(tgt));
1809                 lrh->lrh_magic = LRH_MAGIC;
1810                 lrh->lrh_header_size = sizeof(struct lsd_reply_header);
1811                 lrh->lrh_reply_size = sizeof(struct lsd_reply_data);
1812                 rc = tgt_reply_header_write(env, tgt, lrh);
1813                 if (rc) {
1814                         CERROR("%s: error writing %s: rc = %d\n",
1815                                tgt_name(tgt), REPLY_DATA, rc);
1816                         GOTO(out, rc);
1817                 }
1818         } else {
1819                 rc = tgt_reply_header_read(env, tgt, lrh);
1820                 if (rc) {
1821                         CERROR("%s: error reading %s: rc = %d\n",
1822                                tgt_name(tgt), REPLY_DATA, rc);
1823                         GOTO(out, rc);
1824                 }
1825                 if (lrh->lrh_magic != LRH_MAGIC ||
1826                     lrh->lrh_header_size != sizeof(struct lsd_reply_header) ||
1827                     lrh->lrh_reply_size != sizeof(struct lsd_reply_data)) {
1828                         CERROR("%s: invalid header in %s\n",
1829                                tgt_name(tgt), REPLY_DATA);
1830                         GOTO(out, rc = -EINVAL);
1831                 }
1832
1833                 hash = cfs_hash_getref(tgt->lut_obd->obd_gen_hash);
1834                 if (hash == NULL)
1835                         GOTO(out, rc = -ENODEV);
1836
1837                 OBD_ALLOC_PTR(lcd);
1838                 if (lcd == NULL)
1839                         GOTO(out, rc = -ENOMEM);
1840
1841                 OBD_ALLOC_PTR(trd);
1842                 if (trd == NULL)
1843                         GOTO(out, rc = -ENOMEM);
1844
1845                 /* Load reply_data from disk */
1846                 for (idx = 0, off = sizeof(struct lsd_reply_header);
1847                      off < reply_data_size;
1848                      idx++, off += sizeof(struct lsd_reply_data)) {
1849                         rc = tgt_reply_data_read(env, tgt, lrd, off);
1850                         if (rc) {
1851                                 CERROR("%s: error reading %s: rc = %d\n",
1852                                        tgt_name(tgt), REPLY_DATA, rc);
1853                                 GOTO(out, rc);
1854                         }
1855
1856                         exp = cfs_hash_lookup(hash, &lrd->lrd_client_gen);
1857                         if (exp == NULL) {
1858                                 /* old reply data from a disconnected client */
1859                                 continue;
1860                         }
1861                         ted = &exp->exp_target_data;
1862                         mutex_lock(&ted->ted_lcd_lock);
1863
1864                         /* create in-memory reply_data and link it to
1865                          * target export's reply list */
1866                         tgt_set_reply_slot(tgt, idx);
1867                         trd->trd_reply = *lrd;
1868                         trd->trd_pre_versions[0] = 0;
1869                         trd->trd_pre_versions[1] = 0;
1870                         trd->trd_pre_versions[2] = 0;
1871                         trd->trd_pre_versions[3] = 0;
1872                         trd->trd_index = idx;
1873                         trd->trd_tag = 0;
1874                         list_add(&trd->trd_list, &ted->ted_reply_list);
1875                         ted->ted_reply_cnt++;
1876                         if (ted->ted_reply_cnt > ted->ted_reply_max)
1877                                 ted->ted_reply_max = ted->ted_reply_cnt;
1878
1879                         CDEBUG(D_HA, "%s: restore reply %p: xid %llu, "
1880                                "transno %llu, client gen %u, slot idx %d\n",
1881                                tgt_name(tgt), trd, lrd->lrd_xid,
1882                                lrd->lrd_transno, lrd->lrd_client_gen,
1883                                trd->trd_index);
1884
1885                         /* update export last committed transation */
1886                         exp->exp_last_committed = max(exp->exp_last_committed,
1887                                                       lrd->lrd_transno);
1888
1889                         mutex_unlock(&ted->ted_lcd_lock);
1890                         class_export_put(exp);
1891
1892                         /* update target last committed transaction */
1893                         spin_lock(&tgt->lut_translock);
1894                         tgt->lut_last_transno = max(tgt->lut_last_transno,
1895                                                     lrd->lrd_transno);
1896                         spin_unlock(&tgt->lut_translock);
1897
1898                         reply_data_recovered++;
1899
1900                         OBD_ALLOC_PTR(trd);
1901                         if (trd == NULL)
1902                                 GOTO(out, rc = -ENOMEM);
1903                 }
1904                 CDEBUG(D_INFO, "%s: %d reply data have been recovered\n",
1905                        tgt_name(tgt), reply_data_recovered);
1906         }
1907
1908         spin_lock(&tgt->lut_translock);
1909         /* obd_last_committed is used for compatibility
1910          * with other lustre recovery code */
1911         tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1912         spin_unlock(&tgt->lut_translock);
1913
1914         rc = 0;
1915
1916 out:
1917         if (hash != NULL)
1918                 cfs_hash_putref(hash);
1919         if (lcd != NULL)
1920                 OBD_FREE_PTR(lcd);
1921         if (trd != NULL)
1922                 OBD_FREE_PTR(trd);
1923         if (lrh != NULL)
1924                 OBD_FREE_PTR(lrh);
1925         return rc;
1926 }
1927
1928 struct tg_reply_data *tgt_lookup_reply_by_xid(struct tg_export_data *ted,
1929                                               __u64 xid)
1930 {
1931         struct tg_reply_data    *found = NULL;
1932         struct tg_reply_data    *reply;
1933
1934         mutex_lock(&ted->ted_lcd_lock);
1935         list_for_each_entry(reply, &ted->ted_reply_list, trd_list) {
1936                 if (reply->trd_reply.lrd_xid == xid) {
1937                         found = reply;
1938                         break;
1939                 }
1940         }
1941         mutex_unlock(&ted->ted_lcd_lock);
1942         return found;
1943 }
1944 EXPORT_SYMBOL(tgt_lookup_reply_by_xid);
1945
1946 /* Look for a reply data matching specified request @req
1947  * A copy is returned in @trd if the pointer is not NULL
1948  */
1949 bool tgt_lookup_reply(struct ptlrpc_request *req, struct tg_reply_data *trd)
1950 {
1951         struct tg_export_data   *ted = &req->rq_export->exp_target_data;
1952         struct tg_reply_data    *reply;
1953         bool                     found = false;
1954
1955         reply = tgt_lookup_reply_by_xid(ted, req->rq_xid);
1956         if (reply != NULL) {
1957                 found = true;
1958                 if (trd != NULL)
1959                         *trd = *reply;
1960         }
1961
1962         CDEBUG(D_TRACE, "%s: lookup reply xid %llu, found %d\n",
1963                tgt_name(class_exp2tgt(req->rq_export)), req->rq_xid,
1964                found ? 1 : 0);
1965
1966         return found;
1967 }
1968 EXPORT_SYMBOL(tgt_lookup_reply);
1969
1970 int tgt_handle_received_xid(struct obd_export *exp, __u64 rcvd_xid)
1971 {
1972         struct tg_export_data   *ted = &exp->exp_target_data;
1973         struct lu_target        *lut = class_exp2tgt(exp);
1974         struct tg_reply_data    *trd, *tmp;
1975
1976         mutex_lock(&ted->ted_lcd_lock);
1977         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
1978                 if (trd->trd_reply.lrd_xid > rcvd_xid)
1979                         continue;
1980                 ted->ted_release_xid++;
1981                 tgt_release_reply_data(lut, ted, trd);
1982         }
1983         mutex_unlock(&ted->ted_lcd_lock);
1984
1985         return 0;
1986 }
1987
1988 int tgt_handle_tag(struct obd_export *exp, __u16 tag)
1989 {
1990         struct tg_export_data   *ted = &exp->exp_target_data;
1991         struct lu_target        *lut = class_exp2tgt(exp);
1992         struct tg_reply_data    *trd, *tmp;
1993
1994         if (tag == 0)
1995                 return 0;
1996
1997         mutex_lock(&ted->ted_lcd_lock);
1998         list_for_each_entry_safe(trd, tmp, &ted->ted_reply_list, trd_list) {
1999                 if (trd->trd_tag != tag)
2000                         continue;
2001                 ted->ted_release_tag++;
2002                 tgt_release_reply_data(lut, ted, trd);
2003                 break;
2004         }
2005         mutex_unlock(&ted->ted_lcd_lock);
2006
2007         return 0;
2008 }
2009