lustre/mdt/mdt_recovery.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  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/mdt/mdt_recovery.c
  37  *
  38  * Lustre Metadata Target (mdt) recovery-related methods
  39  *
  40  * Author: Huang Hua <huanghua@clusterfs.com>
  41  * Author: Pershin Mike <tappro@clusterfs.com>
  42  */
  43
  44 #ifndef EXPORT_SYMTAB
  45 # define EXPORT_SYMTAB
  46 #endif
  47 #define DEBUG_SUBSYSTEM S_MDS
  48
  49 #include "mdt_internal.h"
  50
  51 static int mdt_server_data_update(const struct lu_env *env,
  52                                   struct mdt_device *mdt,
  53                                   int need_sync);
  54
  55 struct lu_buf *mdt_buf(const struct lu_env *env, void *area, ssize_t len)
  56 {
  57         struct lu_buf *buf;
  58         struct mdt_thread_info *mti;
  59
  60         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
  61         buf = &mti->mti_buf;
  62         buf->lb_buf = area;
  63         buf->lb_len = len;
  64         return buf;
  65 }
  66
  67 const struct lu_buf *mdt_buf_const(const struct lu_env *env,
  68                                    const void *area, ssize_t len)
  69 {
  70         struct lu_buf *buf;
  71         struct mdt_thread_info *mti;
  72
  73         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
  74         buf = &mti->mti_buf;
  75
  76         buf->lb_buf = (void *)area;
  77         buf->lb_len = len;
  78         return buf;
  79 }
  80
  81 static inline int mdt_trans_credit_get(const struct lu_env *env,
  82                                        struct mdt_device *mdt,
  83                                        enum mdt_txn_op op)
  84 {
  85         struct dt_device *dev = mdt->mdt_bottom;
  86         int cr;
  87         switch (op) {
  88                 case MDT_TXN_CAPA_KEYS_WRITE_OP:
  89                 case MDT_TXN_LAST_RCVD_WRITE_OP:
  90                         cr = dev->dd_ops->dt_credit_get(env,
  91                                                         dev,
  92                                                         DTO_WRITE_BLOCK);
  93                 break;
  94                 default:
  95                         LBUG();
  96         }
  97         return cr;
  98 }
  99
 100 void mdt_trans_credit_init(const struct lu_env *env,
 101                            struct mdt_device *mdt,
 102                            enum mdt_txn_op op)
 103 {
 104         struct mdt_thread_info *mti;
 105         struct txn_param *p;
 106         int cr;
 107
 108         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 109         p = &mti->mti_txn_param;
 110
 111         cr = mdt_trans_credit_get(env, mdt, op);
 112         txn_param_init(p, cr);
 113 }
 114
 115 struct thandle* mdt_trans_start(const struct lu_env *env,
 116                                 struct mdt_device *mdt)
 117 {
 118         struct mdt_thread_info *mti;
 119         struct txn_param *p;
 120
 121         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 122         p = &mti->mti_txn_param;
 123
 124         /* export can require sync operations */
 125         if (mti->mti_exp != NULL)
 126                 p->tp_sync = mti->mti_exp->exp_need_sync;
 127
 128         return mdt->mdt_bottom->dd_ops->dt_trans_start(env, mdt->mdt_bottom, p);
 129 }
 130
 131 void mdt_trans_stop(const struct lu_env *env,
 132                     struct mdt_device *mdt, struct thandle *th)
 133 {
 134         mdt->mdt_bottom->dd_ops->dt_trans_stop(env, th);
 135 }
 136
 137 static inline int mdt_last_rcvd_header_read(const struct lu_env *env,
 138                                             struct mdt_device *mdt)
 139 {
 140         struct mdt_thread_info *mti;
 141         int rc;
 142
 143         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 144
 145         mti->mti_off = 0;
 146         rc = dt_record_read(env, mdt->mdt_last_rcvd,
 147                             mdt_buf(env, &mti->mti_lsd, sizeof(mti->mti_lsd)),
 148                             &mti->mti_off);
 149         if (rc == 0)
 150                 lsd_le_to_cpu(&mti->mti_lsd, &mdt->mdt_lsd);
 151
 152         CDEBUG(D_INFO, "read last_rcvd header rc = %d:\n"
 153                        "uuid = %s\n"
 154                        "last_transno = "LPU64"\n",
 155                         rc, mdt->mdt_lsd.lsd_uuid,
 156                         mdt->mdt_lsd.lsd_last_transno);
 157         return rc;
 158 }
 159
 160 static inline int mdt_last_rcvd_header_write(const struct lu_env *env,
 161                                              struct mdt_device *mdt,
 162                                              int need_sync)
 163 {
 164         struct mdt_thread_info *mti;
 165         struct thandle *th;
 166         int rc;
 167         ENTRY;
 168
 169         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 170
 171         if (mti->mti_exp) {
 172                 spin_lock(&mti->mti_exp->exp_lock);
 173                 mti->mti_exp->exp_need_sync = need_sync;
 174                 spin_unlock(&mti->mti_exp->exp_lock);
 175         }
 176         mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
 177         th = mdt_trans_start(env, mdt);
 178         if (IS_ERR(th))
 179                 RETURN(PTR_ERR(th));
 180
 181         mti->mti_off = 0;
 182         lsd_cpu_to_le(&mdt->mdt_lsd, &mti->mti_lsd);
 183
 184         if (need_sync && mti->mti_exp)
 185                 mdt_trans_add_cb(th, lut_cb_client, mti->mti_exp);
 186
 187         rc = dt_record_write(env, mdt->mdt_last_rcvd,
 188                              mdt_buf_const(env, &mti->mti_lsd,
 189                                            sizeof(mti->mti_lsd)),
 190                              &mti->mti_off, th);
 191
 192         mdt_trans_stop(env, mdt, th);
 193
 194         CDEBUG(D_INFO, "write last_rcvd header rc = %d:\n"
 195                "uuid = %s\nlast_transno = "LPU64"\n",
 196                rc, mdt->mdt_lsd.lsd_uuid, mdt->mdt_lsd.lsd_last_transno);
 197
 198         RETURN(rc);
 199 }
 200
 201 static int mdt_last_rcvd_read(const struct lu_env *env,
 202                               struct mdt_device *mdt,
 203                               struct lsd_client_data *lcd, loff_t *off)
 204 {
 205         struct mdt_thread_info *mti;
 206         struct lsd_client_data *tmp;
 207         int rc;
 208
 209         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 210         tmp = &mti->mti_lcd;
 211         rc = dt_record_read(env, mdt->mdt_last_rcvd,
 212                             mdt_buf(env, tmp, sizeof(*tmp)), off);
 213         if (rc == 0)
 214                 lcd_le_to_cpu(tmp, lcd);
 215
 216         CDEBUG(D_INFO, "read lcd @%d rc = %d:\n"
 217                        "uuid = %s\n"
 218                        "last_transno = "LPU64"\n"
 219                        "last_xid = "LPU64"\n"
 220                        "last_result = %u\n"
 221                        "last_data = %u\n"
 222                        "last_close_transno = "LPU64"\n"
 223                        "last_close_xid = "LPU64"\n"
 224                        "last_close_result = %u\n",
 225                         (int)(*off - sizeof(*tmp)),
 226                         rc,
 227                         lcd->lcd_uuid,
 228                         lcd->lcd_last_transno,
 229                         lcd->lcd_last_xid,
 230                         lcd->lcd_last_result,
 231                         lcd->lcd_last_data,
 232                         lcd->lcd_last_close_transno,
 233                         lcd->lcd_last_close_xid,
 234                         lcd->lcd_last_close_result);
 235         return rc;
 236 }
 237
 238 static int mdt_last_rcvd_write(const struct lu_env *env,
 239                                struct mdt_device *mdt,
 240                                struct lsd_client_data *lcd,
 241                                loff_t *off, struct thandle *th)
 242 {
 243         struct mdt_thread_info *mti;
 244         struct lsd_client_data *tmp;
 245         int rc;
 246
 247         LASSERT(th != NULL);
 248         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 249         tmp = &mti->mti_lcd;
 250
 251         lcd_cpu_to_le(lcd, tmp);
 252
 253         rc = dt_record_write(env, mdt->mdt_last_rcvd,
 254                              mdt_buf_const(env, tmp, sizeof(*tmp)), off, th);
 255
 256         CDEBUG(D_INFO, "write lcd @%d rc = %d:\n"
 257                        "uuid = %s\n"
 258                        "last_transno = "LPU64"\n"
 259                        "last_xid = "LPU64"\n"
 260                        "last_result = %u\n"
 261                        "last_data = %u\n"
 262                        "last_close_transno = "LPU64"\n"
 263                        "last_close_xid = "LPU64"\n"
 264                        "last_close_result = %u\n",
 265                         (int)(*off - sizeof(*tmp)),
 266                         rc,
 267                         lcd->lcd_uuid,
 268                         lcd->lcd_last_transno,
 269                         lcd->lcd_last_xid,
 270                         lcd->lcd_last_result,
 271                         lcd->lcd_last_data,
 272                         lcd->lcd_last_close_transno,
 273                         lcd->lcd_last_close_xid,
 274                         lcd->lcd_last_close_result);
 275         return rc;
 276 }
 277
 278 static int mdt_clients_data_init(const struct lu_env *env,
 279                                  struct mdt_device *mdt,
 280                                  unsigned long last_size)
 281 {
 282         struct lr_server_data  *lsd = &mdt->mdt_lsd;
 283         struct lsd_client_data *lcd = NULL;
 284         struct obd_device      *obd = mdt2obd_dev(mdt);
 285         loff_t off;
 286         int cl_idx;
 287         int rc = 0;
 288         ENTRY;
 289
 290         /* When we do a clean MDS shutdown, we save the last_transno into
 291          * the header.  If we find clients with higher last_transno values
 292          * then those clients may need recovery done. */
 293         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
 294         for (cl_idx = 0, off = lsd->lsd_client_start;
 295              off < last_size; cl_idx++) {
 296                 __u64 last_transno;
 297                 struct obd_export *exp;
 298
 299                 if (!lcd) {
 300                         OBD_ALLOC_PTR(lcd);
 301                         if (!lcd)
 302                                 RETURN(-ENOMEM);
 303                 }
 304
 305                 off = lsd->lsd_client_start +
 306                         cl_idx * lsd->lsd_client_size;
 307
 308                 rc = mdt_last_rcvd_read(env, mdt, lcd, &off);
 309                 if (rc) {
 310                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 311                                LAST_RCVD, cl_idx, off, rc);
 312                         rc = 0;
 313                         break; /* read error shouldn't cause startup to fail */
 314                 }
 315
 316                 if (lcd->lcd_uuid[0] == '\0') {
 317                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 318                                cl_idx);
 319                         continue;
 320                 }
 321
 322                 last_transno = lcd_last_transno(lcd);
 323
 324                 /* These exports are cleaned up by mdt_obd_disconnect(), so
 325                  * they need to be set up like real exports as
 326                  * mdt_obd_connect() does.
 327                  */
 328                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 329                        " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
 330                        last_transno, lsd->lsd_last_transno,
 331                        lcd_last_xid(lcd));
 332
 333                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
 334                 if (IS_ERR(exp)) {
 335                         if (PTR_ERR(exp) == -EALREADY) {
 336                                 /* export already exists, zero out this one */
 337                                 lcd->lcd_uuid[0] = '\0';
 338                         } else
 339                                 GOTO(err_client, rc = PTR_ERR(exp));
 340                 } else {
 341                         struct mdt_thread_info *mti;
 342                         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 343                         LASSERT(mti != NULL);
 344                         mti->mti_exp = exp;
 345                         exp->exp_mdt_data.med_lcd = lcd;
 346                         rc = mdt_client_add(env, mdt, cl_idx);
 347                         /* can't fail existing */
 348                         LASSERTF(rc == 0, "rc = %d\n", rc);
 349                         /* VBR: set export last committed version */
 350                         exp->exp_last_committed = last_transno;
 351                         lcd = NULL;
 352                         spin_lock(&exp->exp_lock);
 353                         exp->exp_connecting = 0;
 354                         exp->exp_in_recovery = 0;
 355                         spin_unlock(&exp->exp_lock);
 356                         obd->obd_max_recoverable_clients++;
 357                         class_export_put(exp);
 358                 }
 359
 360                 CDEBUG(D_OTHER, "client at idx %d has last_transno="LPU64"\n",
 361                        cl_idx, last_transno);
 362                 /* protect __u64 value update */
 363                 spin_lock(&mdt->mdt_transno_lock);
 364                 mdt->mdt_last_transno = max(last_transno,
 365                                             mdt->mdt_last_transno);
 366                 spin_unlock(&mdt->mdt_transno_lock);
 367         }
 368
 369 err_client:
 370         if (lcd)
 371                 OBD_FREE_PTR(lcd);
 372         RETURN(rc);
 373 }
 374
 375 static int mdt_server_data_init(const struct lu_env *env,
 376                                 struct mdt_device *mdt,
 377                                 struct lustre_sb_info *lsi)
 378 {
 379         struct lr_server_data  *lsd = &mdt->mdt_lsd;
 380         struct lsd_client_data *lcd = NULL;
 381         struct obd_device      *obd = mdt2obd_dev(mdt);
 382         struct mdt_thread_info *mti;
 383         struct dt_object       *obj;
 384         struct lu_attr         *la;
 385         struct lustre_disk_data  *ldd;
 386         unsigned long last_rcvd_size;
 387         __u64 mount_count;
 388         int rc;
 389         ENTRY;
 390
 391         /* ensure padding in the struct is the correct size */
 392         CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
 393                 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
 394         CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
 395                 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
 396
 397         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 398         LASSERT(mti != NULL);
 399         la = &mti->mti_attr.ma_attr;
 400
 401         obj = mdt->mdt_last_rcvd;
 402         rc = obj->do_ops->do_attr_get(env, mdt->mdt_last_rcvd, la, BYPASS_CAPA);
 403         if (rc)
 404                 RETURN(rc);
 405
 406         last_rcvd_size = (unsigned long)la->la_size;
 407
 408         if (last_rcvd_size == 0) {
 409                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
 410
 411                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,
 412                        sizeof(lsd->lsd_uuid));
 413                 lsd->lsd_last_transno = 0;
 414                 lsd->lsd_mount_count = 0;
 415                 lsd->lsd_server_size = LR_SERVER_SIZE;
 416                 lsd->lsd_client_start = LR_CLIENT_START;
 417                 lsd->lsd_client_size = LR_CLIENT_SIZE;
 418                 lsd->lsd_feature_rocompat = OBD_ROCOMPAT_LOVOBJID;
 419                 lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT |
 420                                                        OBD_INCOMPAT_COMMON_LR;
 421         } else {
 422                 LCONSOLE_WARN("%s: used disk, loading\n", obd->obd_name);
 423                 rc = mdt_last_rcvd_header_read(env, mdt);
 424                 if (rc) {
 425                         CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
 426                         GOTO(out, rc);
 427                 }
 428                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
 429                         LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s using"
 430                                            "the wrong disk %s. Were the /dev/ "
 431                                            "assignments rearranged?\n",
 432                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
 433                         GOTO(out, rc = -EINVAL);
 434                 }
 435         }
 436         mount_count = lsd->lsd_mount_count;
 437
 438         ldd = lsi->lsi_ldd;
 439
 440         if (ldd->ldd_flags & LDD_F_IAM_DIR)
 441                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_IAM_DIR;
 442
 443         lsd->lsd_feature_compat = OBD_COMPAT_MDT;
 444         lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
 445
 446         spin_lock(&mdt->mdt_transno_lock);
 447         mdt->mdt_last_transno = lsd->lsd_last_transno;
 448         spin_unlock(&mdt->mdt_transno_lock);
 449
 450         CDEBUG(D_INODE, "========BEGIN DUMPING LAST_RCVD========\n");
 451         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 452                obd->obd_name, mdt->mdt_last_transno);
 453         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 454                obd->obd_name, mount_count + 1);
 455         CDEBUG(D_INODE, "%s: server data size: %u\n",
 456                obd->obd_name, lsd->lsd_server_size);
 457         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 458                obd->obd_name, lsd->lsd_client_start);
 459         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 460                obd->obd_name, lsd->lsd_client_size);
 461         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 462                obd->obd_name, last_rcvd_size);
 463         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 464                last_rcvd_size <= lsd->lsd_client_start ? 0 :
 465                (last_rcvd_size - lsd->lsd_client_start) /
 466                 lsd->lsd_client_size);
 467         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
 468
 469         if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
 470             !lsd->lsd_client_size) {
 471                 CERROR("Bad last_rcvd contents!\n");
 472                 GOTO(out, rc = -EINVAL);
 473         }
 474
 475         rc = mdt_clients_data_init(env, mdt, last_rcvd_size);
 476         if (rc)
 477                 GOTO(err_client, rc);
 478
 479         spin_lock(&mdt->mdt_transno_lock);
 480         /* obd_last_committed is used for compatibility
 481          * with other lustre recovery code */
 482         obd->obd_last_committed = mdt->mdt_last_transno;
 483         spin_unlock(&mdt->mdt_transno_lock);
 484
 485         mdt->mdt_mount_count++;
 486         lsd->lsd_mount_count = mdt->mdt_mount_count;
 487
 488         /* save it, so mount count and last_transno is current */
 489         rc = mdt_server_data_update(env, mdt, (mti->mti_exp &&
 490                                                mti->mti_exp->exp_need_sync));
 491         if (rc)
 492                 GOTO(err_client, rc);
 493
 494         RETURN(0);
 495
 496 err_client:
 497         class_disconnect_exports(obd);
 498 out:
 499         return rc;
 500 }
 501
 502 static int mdt_server_data_update(const struct lu_env *env,
 503                                   struct mdt_device *mdt,
 504                                   int need_sync)
 505 {
 506         int rc = 0;
 507         ENTRY;
 508
 509         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
 510                mdt->mdt_mount_count, mdt->mdt_last_transno);
 511
 512         spin_lock(&mdt->mdt_transno_lock);
 513         mdt->mdt_lsd.lsd_last_transno = mdt->mdt_last_transno;
 514         spin_unlock(&mdt->mdt_transno_lock);
 515
 516         /*
 517          * This may be called from difficult reply handler and
 518          * mdt->mdt_last_rcvd may be NULL that time.
 519          */
 520         if (mdt->mdt_last_rcvd != NULL)
 521                 rc = mdt_last_rcvd_header_write(env, mdt, need_sync);
 522         RETURN(rc);
 523 }
 524
 525 int mdt_client_new(const struct lu_env *env, struct mdt_device *mdt)
 526 {
 527         unsigned long *bitmap = mdt->mdt_client_bitmap;
 528         struct mdt_thread_info *mti;
 529         struct mdt_export_data *med;
 530         struct lsd_client_data *lcd;
 531         struct lr_server_data  *lsd = &mdt->mdt_lsd;
 532         struct obd_device *obd = mdt2obd_dev(mdt);
 533         struct thandle *th;
 534         loff_t off;
 535         int rc;
 536         int cl_idx;
 537         ENTRY;
 538
 539         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 540         LASSERT(mti != NULL);
 541
 542         med = &mti->mti_exp->exp_mdt_data;
 543         lcd = med->med_lcd;
 544
 545         LASSERT(bitmap != NULL);
 546         if (!strcmp(med->med_lcd->lcd_uuid, obd->obd_uuid.uuid))
 547                 RETURN(0);
 548
 549         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 550          * there's no need for extra complication here
 551          */
 552         spin_lock(&mdt->mdt_client_bitmap_lock);
 553         cl_idx = find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
 554         if (cl_idx >= LR_MAX_CLIENTS ||
 555             OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
 556                 CERROR("no room for %u clients - fix LR_MAX_CLIENTS\n",
 557                        cl_idx);
 558                 spin_unlock(&mdt->mdt_client_bitmap_lock);
 559                 RETURN(-EOVERFLOW);
 560         }
 561         set_bit(cl_idx, bitmap);
 562         spin_unlock(&mdt->mdt_client_bitmap_lock);
 563
 564         CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
 565                cl_idx, med->med_lcd->lcd_uuid);
 566
 567         med->med_lr_idx = cl_idx;
 568         med->med_lr_off = lsd->lsd_client_start +
 569                           (cl_idx * lsd->lsd_client_size);
 570         init_mutex(&med->med_lcd_lock);
 571
 572         LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off);
 573
 574         /* Write new client data. */
 575         off = med->med_lr_off;
 576         mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
 577
 578         th = mdt_trans_start(env, mdt);
 579         if (IS_ERR(th))
 580                 RETURN(PTR_ERR(th));
 581
 582         /*
 583          * Until this operations will be committed the sync is needed
 584          * for this export. This should be done _after_ starting the
 585          * transaction so that many connecting clients will not bring
 586          * server down with lots of sync writes.
 587          */
 588         mdt_trans_add_cb(th, lut_cb_client, mti->mti_exp);
 589         spin_lock(&mti->mti_exp->exp_lock);
 590         mti->mti_exp->exp_need_sync = 1;
 591         spin_unlock(&mti->mti_exp->exp_lock);
 592
 593         rc = mdt_last_rcvd_write(env, mdt, lcd, &off, th);
 594         CDEBUG(D_INFO, "wrote client lcd at idx %u off %llu (len "LPSZ")\n",
 595                cl_idx, med->med_lr_off, sizeof(*lcd));
 596         mdt_trans_stop(env, mdt, th);
 597
 598         RETURN(rc);
 599 }
 600
 601 /* Add client data to the MDS.  We use a bitmap to locate a free space
 602  * in the last_rcvd file if cl_off is -1 (i.e. a new client).
 603  * Otherwise, we just have to read the data from the last_rcvd file and
 604  * we know its offset.
 605  *
 606  * It should not be possible to fail adding an existing client - otherwise
 607  * mdt_init_server_data() callsite needs to be fixed.
 608  */
 609 int mdt_client_add(const struct lu_env *env,
 610                    struct mdt_device *mdt, int cl_idx)
 611 {
 612         struct mdt_thread_info *mti;
 613         struct mdt_export_data *med;
 614         unsigned long *bitmap = mdt->mdt_client_bitmap;
 615         struct obd_device *obd = mdt2obd_dev(mdt);
 616         struct lr_server_data *lsd = &mdt->mdt_lsd;
 617         int rc = 0;
 618         ENTRY;
 619
 620         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 621         LASSERT(mti != NULL);
 622
 623         med = &mti->mti_exp->exp_mdt_data;
 624
 625         LASSERT(bitmap != NULL);
 626         LASSERTF(cl_idx >= 0, "%d\n", cl_idx);
 627
 628         if (!strcmp(med->med_lcd->lcd_uuid, obd->obd_uuid.uuid))
 629                 RETURN(0);
 630
 631         spin_lock(&mdt->mdt_client_bitmap_lock);
 632         if (test_and_set_bit(cl_idx, bitmap)) {
 633                 CERROR("MDS client %d: bit already set in bitmap!!\n",
 634                        cl_idx);
 635                 LBUG();
 636         }
 637         spin_unlock(&mdt->mdt_client_bitmap_lock);
 638
 639         CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
 640                cl_idx, med->med_lcd->lcd_uuid);
 641
 642         med->med_lr_idx = cl_idx;
 643         med->med_lr_off = lsd->lsd_client_start +
 644                           (cl_idx * lsd->lsd_client_size);
 645         init_mutex(&med->med_lcd_lock);
 646
 647         LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off);
 648
 649         RETURN(rc);
 650 }
 651
 652 int mdt_client_del(const struct lu_env *env, struct mdt_device *mdt)
 653 {
 654         struct mdt_thread_info *mti;
 655         struct mdt_export_data *med;
 656         struct lsd_client_data *lcd;
 657         struct obd_device      *obd = mdt2obd_dev(mdt);
 658         struct obd_export      *exp;
 659         struct thandle         *th;
 660         int                     need_sync;
 661         loff_t                  off;
 662         int                     rc = 0;
 663         ENTRY;
 664
 665         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 666         LASSERT(mti != NULL);
 667
 668         exp = mti->mti_exp;
 669         med = &exp->exp_mdt_data;
 670         lcd = med->med_lcd;
 671         if (!lcd)
 672                 RETURN(0);
 673
 674         /* XXX: If lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 675         if (!strcmp(lcd->lcd_uuid, obd->obd_uuid.uuid))
 676                 GOTO(free, 0);
 677
 678         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld\n",
 679                med->med_lr_idx, med->med_lr_off);
 680
 681         off = med->med_lr_off;
 682
 683         /*
 684          * Don't clear med_lr_idx here as it is likely also unset.  At worst we
 685          * leak a client slot that will be cleaned on the next recovery.
 686          */
 687         if (off <= 0) {
 688                 CERROR("client idx %d has offset %lld\n",
 689                         med->med_lr_idx, off);
 690                 GOTO(free, rc = -EINVAL);
 691         }
 692
 693         /*
 694          * Clear the bit _after_ zeroing out the client so we don't race with
 695          * mdt_client_add and zero out new clients.
 696          */
 697         if (!test_bit(med->med_lr_idx, mdt->mdt_client_bitmap)) {
 698                 CERROR("MDT client %u: bit already clear in bitmap!!\n",
 699                        med->med_lr_idx);
 700                 LBUG();
 701         }
 702
 703         /* Don't force sync on disconnect if aborting recovery,
 704          * or it does num_clients * num_osts.  b=17194 */
 705         need_sync = (!exp->exp_libclient || exp->exp_need_sync) &&
 706                      !(exp->exp_flags & OBD_OPT_ABORT_RECOV);
 707
 708         /*
 709          * This may be called from difficult reply handler path and
 710          * mdt->mdt_last_rcvd may be NULL that time.
 711          */
 712         if (mdt->mdt_last_rcvd != NULL) {
 713                 mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
 714
 715                 spin_lock(&exp->exp_lock);
 716                 exp->exp_need_sync = need_sync;
 717                 spin_unlock(&exp->exp_lock);
 718
 719                 th = mdt_trans_start(env, mdt);
 720                 if (IS_ERR(th))
 721                         GOTO(free, rc = PTR_ERR(th));
 722
 723                 if (need_sync) {
 724                         /*
 725                          * Until this operations will be committed the sync
 726                          * is needed for this export.
 727                          */
 728                         mdt_trans_add_cb(th, lut_cb_client, exp);
 729                 }
 730
 731                 mutex_down(&med->med_lcd_lock);
 732                 memset(lcd, 0, sizeof *lcd);
 733
 734                 rc = mdt_last_rcvd_write(env, mdt, lcd, &off, th);
 735                 mutex_up(&med->med_lcd_lock);
 736                 mdt_trans_stop(env, mdt, th);
 737         }
 738
 739         CDEBUG(rc == 0 ? D_INFO : D_ERROR, "Zeroing out client idx %u in "
 740                "%s %ssync rc %d\n",  med->med_lr_idx, LAST_RCVD,
 741                need_sync ? "" : "a", rc);
 742
 743         spin_lock(&mdt->mdt_client_bitmap_lock);
 744         clear_bit(med->med_lr_idx, mdt->mdt_client_bitmap);
 745         spin_unlock(&mdt->mdt_client_bitmap_lock);
 746
 747         /*
 748          * Make sure the server's last_transno is up to date. Do this
 749          * after the client is freed so we know all the client's
 750          * transactions have been committed.
 751          */
 752         mdt_server_data_update(env, mdt, need_sync);
 753
 754         EXIT;
 755 free:
 756         OBD_FREE_PTR(lcd);
 757         med->med_lcd = NULL;
 758         return 0;
 759 }
 760
 761 /*
 762  * last_rcvd & last_committed update callbacks
 763  */
 764 static int mdt_last_rcvd_update(struct mdt_thread_info *mti,
 765                                 struct thandle *th)
 766 {
 767         struct mdt_device *mdt = mti->mti_mdt;
 768         struct ptlrpc_request *req = mdt_info_req(mti);
 769         struct mdt_export_data *med;
 770         struct lsd_client_data *lcd;
 771         loff_t off;
 772         int err;
 773         __s32 rc = th->th_result;
 774
 775         ENTRY;
 776         LASSERT(req);
 777         LASSERT(req->rq_export);
 778         LASSERT(mdt);
 779         med = &req->rq_export->exp_mdt_data;
 780         LASSERT(med);
 781         lcd = med->med_lcd;
 782         /* if the export has already been failed, we have no last_rcvd slot */
 783         if (req->rq_export->exp_failed) {
 784                 CWARN("commit transaction for disconnected client %s: rc %d\n",
 785                       req->rq_export->exp_client_uuid.uuid, rc);
 786                 if (rc == 0)
 787                         rc = -ENOTCONN;
 788                 RETURN(rc);
 789         }
 790
 791         off = med->med_lr_off;
 792         LASSERT(ergo(mti->mti_transno == 0, rc != 0));
 793         mutex_down(&med->med_lcd_lock);
 794         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 795             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 796                 if (mti->mti_transno != 0)
 797                         lcd->lcd_last_close_transno = mti->mti_transno;
 798                 lcd->lcd_last_close_xid = req->rq_xid;
 799                 lcd->lcd_last_close_result = rc;
 800         } else {
 801                 /* VBR: save versions in last_rcvd for reconstruct. */
 802                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 803                 if (pre_versions) {
 804                         lcd->lcd_pre_versions[0] = pre_versions[0];
 805                         lcd->lcd_pre_versions[1] = pre_versions[1];
 806                         lcd->lcd_pre_versions[2] = pre_versions[2];
 807                         lcd->lcd_pre_versions[3] = pre_versions[3];
 808                 }
 809                 if (mti->mti_transno != 0)
 810                         lcd->lcd_last_transno = mti->mti_transno;
 811                 lcd->lcd_last_xid = req->rq_xid;
 812                 lcd->lcd_last_result = rc;
 813                 /*XXX: save intent_disposition in mdt_thread_info?
 814                  * also there is bug - intent_dispostion is __u64,
 815                  * see struct ldlm_reply->lock_policy_res1; */
 816                 lcd->lcd_last_data = mti->mti_opdata;
 817         }
 818
 819         if (off <= 0) {
 820                 CERROR("client idx %d has offset %lld\n", med->med_lr_idx, off);
 821                 err = -EINVAL;
 822         } else {
 823                 err = mdt_last_rcvd_write(mti->mti_env, mdt, lcd, &off, th);
 824         }
 825         mutex_up(&med->med_lcd_lock);
 826         RETURN(err);
 827 }
 828
 829 extern struct lu_context_key mdt_thread_key;
 830
 831 /* add credits for last_rcvd update */
 832 static int mdt_txn_start_cb(const struct lu_env *env,
 833                             struct txn_param *param, void *cookie)
 834 {
 835         struct mdt_device *mdt = cookie;
 836
 837         param->tp_credits += mdt_trans_credit_get(env, mdt,
 838                                                   MDT_TXN_LAST_RCVD_WRITE_OP);
 839         return 0;
 840 }
 841
 842 /* Set new object versions */
 843 static void mdt_versions_set(struct mdt_thread_info *info)
 844 {
 845         int i;
 846         for (i = 0; i < PTLRPC_NUM_VERSIONS; i++)
 847                 if (info->mti_mos[i] != NULL)
 848                         mo_version_set(info->mti_env,
 849                                        mdt_object_child(info->mti_mos[i]),
 850                                        info->mti_transno);
 851 }
 852
 853 /* Update last_rcvd records with latests transaction data */
 854 static int mdt_txn_stop_cb(const struct lu_env *env,
 855                            struct thandle *txn, void *cookie)
 856 {
 857         struct mdt_device *mdt = cookie;
 858         struct mdt_txn_info *txi;
 859         struct mdt_thread_info *mti;
 860         struct ptlrpc_request *req;
 861
 862         /* transno in two contexts - for commit_cb and for thread */
 863         txi = lu_context_key_get(&txn->th_ctx, &mdt_txn_key);
 864         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 865         req = mdt_info_req(mti);
 866
 867         if (mti->mti_mdt == NULL || req == NULL || mti->mti_no_need_trans) {
 868                 txi->txi_transno = 0;
 869                 mti->mti_no_need_trans = 0;
 870                 return 0;
 871         }
 872
 873         if (mti->mti_has_trans) {
 874                 /* XXX: currently there are allowed cases, but the wrong cases
 875                  * are also possible, so better check is needed here */
 876                 CDEBUG(D_INFO, "More than one transaction "LPU64"\n",
 877                        mti->mti_transno);
 878                 return 0;
 879         }
 880
 881         mti->mti_has_trans = 1;
 882         spin_lock(&mdt->mdt_transno_lock);
 883         if (txn->th_result != 0) {
 884                 if (mti->mti_transno != 0) {
 885                         CERROR("Replay transno "LPU64" failed: rc %i\n",
 886                                mti->mti_transno, txn->th_result);
 887                 }
 888         } else if (mti->mti_transno == 0) {
 889                 mti->mti_transno = ++ mdt->mdt_last_transno;
 890         } else {
 891                 /* should be replay */
 892                 if (mti->mti_transno > mdt->mdt_last_transno)
 893                         mdt->mdt_last_transno = mti->mti_transno;
 894         }
 895         spin_unlock(&mdt->mdt_transno_lock);
 896         /* sometimes the reply message has not been successfully packed */
 897         LASSERT(req != NULL && req->rq_repmsg != NULL);
 898
 899         /** VBR: set new versions */
 900         if (txn->th_result == 0)
 901                 mdt_versions_set(mti);
 902
 903         /* filling reply data */
 904         CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
 905                mti->mti_transno, req->rq_export->exp_obd->obd_last_committed);
 906
 907         req->rq_transno = mti->mti_transno;
 908         lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno);
 909         lustre_msg_set_last_xid(req->rq_repmsg,
 910                          lcd_last_xid(req->rq_export->exp_mdt_data.med_lcd));
 911         /* save transno for the commit callback */
 912         txi->txi_transno = mti->mti_transno;
 913
 914         /* add separate commit callback for transaction handling because we need
 915          * export as parameter */
 916         mdt_trans_add_cb(txn, lut_cb_last_committed,
 917                          class_export_get(mti->mti_exp));
 918
 919         return mdt_last_rcvd_update(mti, txn);
 920 }
 921
 922 /* commit callback, need to update last_committed value */
 923 static int mdt_txn_commit_cb(const struct lu_env *env,
 924                              struct thandle *txn, void *cookie)
 925 {
 926         struct mdt_device *mdt = cookie;
 927         struct mdt_txn_info *txi;
 928         int i;
 929
 930         txi = lu_context_key_get(&txn->th_ctx, &mdt_txn_key);
 931
 932         /* iterate through all additional callbacks */
 933         for (i = 0; i < txi->txi_cb_count; i++) {
 934                 txi->txi_cb[i].lut_cb_func(&mdt->mdt_lut, txi->txi_transno,
 935                                            txi->txi_cb[i].lut_cb_data, 0);
 936         }
 937         return 0;
 938 }
 939
 940 int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt,
 941                  struct obd_device *obd,
 942                  struct lustre_sb_info *lsi)
 943 {
 944         struct lu_fid fid;
 945         struct dt_object *o;
 946         int rc = 0;
 947         ENTRY;
 948
 949         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
 950                 RETURN(-ENOENT);
 951
 952         /* prepare transactions callbacks */
 953         mdt->mdt_txn_cb.dtc_txn_start = mdt_txn_start_cb;
 954         mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb;
 955         mdt->mdt_txn_cb.dtc_txn_commit = mdt_txn_commit_cb;
 956         mdt->mdt_txn_cb.dtc_cookie = mdt;
 957         mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD;
 958         CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage);
 959
 960         dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb);
 961
 962         rc = mdt_server_data_init(env, mdt, lsi);
 963         if (rc)
 964                 RETURN(rc);
 965
 966         o = dt_store_open(env, mdt->mdt_bottom, "", CAPA_KEYS, &fid);
 967         if (!IS_ERR(o)) {
 968                 mdt->mdt_ck_obj = o;
 969                 rc = mdt_capa_keys_init(env, mdt);
 970                 if (rc)
 971                         GOTO(put_ck_object, rc);
 972         } else {
 973                 rc = PTR_ERR(o);
 974                 CERROR("cannot open %s: rc = %d\n", CAPA_KEYS, rc);
 975                 GOTO(disconnect_exports, rc);
 976         }
 977         RETURN(0);
 978
 979 put_ck_object:
 980         lu_object_put(env, &o->do_lu);
 981         mdt->mdt_ck_obj = NULL;
 982 disconnect_exports:
 983         class_disconnect_exports(obd);
 984         return rc;
 985 }
 986
 987 void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt)
 988 {
 989         ENTRY;
 990
 991         /* Remove transaction callback */
 992         dt_txn_callback_del(mdt->mdt_bottom, &mdt->mdt_txn_cb);
 993         if (mdt->mdt_ck_obj)
 994                 lu_object_put(env, &mdt->mdt_ck_obj->do_lu);
 995         mdt->mdt_ck_obj = NULL;
 996         EXIT;
 997 }
 998
 999 /* reconstruction code */
1000 static void mdt_steal_ack_locks(struct ptlrpc_request *req)
1001 {
1002         struct obd_export         *exp = req->rq_export;
1003         struct list_head          *tmp;
1004         struct ptlrpc_reply_state *oldrep;
1005         struct ptlrpc_service     *svc;
1006         int                        i;
1007
1008         /* CAVEAT EMPTOR: spinlock order */
1009         spin_lock(&exp->exp_lock);
1010         list_for_each (tmp, &exp->exp_outstanding_replies) {
1011                 oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list);
1012
1013                 if (oldrep->rs_xid != req->rq_xid)
1014                         continue;
1015
1016                 if (lustre_msg_get_opc(oldrep->rs_msg) !=
1017                     lustre_msg_get_opc(req->rq_reqmsg))
1018                         CERROR ("Resent req xid "LPU64" has mismatched opc: "
1019                                 "new %d old %d\n", req->rq_xid,
1020                                 lustre_msg_get_opc(req->rq_reqmsg),
1021                                 lustre_msg_get_opc(oldrep->rs_msg));
1022
1023                 svc = oldrep->rs_service;
1024                 spin_lock (&svc->srv_lock);
1025
1026                 list_del_init (&oldrep->rs_exp_list);
1027
1028                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
1029                       " o%d NID %s\n",
1030                       oldrep->rs_nlocks, oldrep,
1031                       oldrep->rs_xid, oldrep->rs_transno,
1032                       lustre_msg_get_opc(oldrep->rs_msg),
1033                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
1034
1035                 for (i = 0; i < oldrep->rs_nlocks; i++)
1036                         ptlrpc_save_lock(req, &oldrep->rs_locks[i],
1037                                          oldrep->rs_modes[i], 0);
1038                 oldrep->rs_nlocks = 0;
1039
1040                 DEBUG_REQ(D_HA, req, "stole locks for");
1041                 spin_lock(&oldrep->rs_lock);
1042                 ptlrpc_schedule_difficult_reply (oldrep);
1043                 spin_unlock(&oldrep->rs_lock);
1044
1045                 spin_unlock (&svc->srv_lock);
1046                 break;
1047         }
1048         spin_unlock(&exp->exp_lock);
1049 }
1050
1051 /**
1052  * VBR: restore versions
1053  */
1054 void mdt_vbr_reconstruct(struct ptlrpc_request *req,
1055                          struct lsd_client_data *lcd)
1056 {
1057         __u64 pre_versions[4] = {0};
1058         pre_versions[0] = lcd->lcd_pre_versions[0];
1059         pre_versions[1] = lcd->lcd_pre_versions[1];
1060         pre_versions[2] = lcd->lcd_pre_versions[2];
1061         pre_versions[3] = lcd->lcd_pre_versions[3];
1062         lustre_msg_set_versions(req->rq_repmsg, pre_versions);
1063 }
1064
1065 void mdt_req_from_lcd(struct ptlrpc_request *req,
1066                       struct lsd_client_data *lcd)
1067 {
1068         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
1069                   lcd->lcd_last_transno, lcd->lcd_last_result);
1070
1071         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
1072             lustre_msg_get_opc(req->rq_repmsg) == MDS_DONE_WRITING) {
1073                 req->rq_transno = lcd->lcd_last_close_transno;
1074                 req->rq_status = lcd->lcd_last_close_result;
1075         } else {
1076                 req->rq_transno = lcd->lcd_last_transno;
1077                 req->rq_status = lcd->lcd_last_result;
1078                 mdt_vbr_reconstruct(req, lcd);
1079         }
1080         if (req->rq_status != 0)
1081                 req->rq_transno = 0;
1082         lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
1083         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1084         DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
1085                   req->rq_transno, req->rq_status);
1086
1087         mdt_steal_ack_locks(req);
1088 }
1089
1090 void mdt_reconstruct_generic(struct mdt_thread_info *mti,
1091                              struct mdt_lock_handle *lhc)
1092 {
1093         struct ptlrpc_request *req = mdt_info_req(mti);
1094         struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1095
1096         return mdt_req_from_lcd(req, med->med_lcd);
1097 }
1098
1099 static void mdt_reconstruct_create(struct mdt_thread_info *mti,
1100                                    struct mdt_lock_handle *lhc)
1101 {
1102         struct ptlrpc_request  *req = mdt_info_req(mti);
1103         struct obd_export *exp = req->rq_export;
1104         struct mdt_export_data *med = &exp->exp_mdt_data;
1105         struct mdt_device *mdt = mti->mti_mdt;
1106         struct mdt_object *child;
1107         struct mdt_body *body;
1108         int rc;
1109
1110         mdt_req_from_lcd(req, med->med_lcd);
1111         if (req->rq_status)
1112                 return;
1113
1114         /* if no error, so child was created with requested fid */
1115         child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2);
1116         if (IS_ERR(child)) {
1117                 rc = PTR_ERR(child);
1118                 LCONSOLE_WARN("Child "DFID" lookup error %d."
1119                               " Evicting client %s with export %s.\n",
1120                               PFID(mdt_object_fid(child)), rc,
1121                               obd_uuid2str(&exp->exp_client_uuid),
1122                               obd_export_nid2str(exp));
1123                 mdt_export_evict(exp);
1124                 EXIT;
1125                 return;
1126         }
1127
1128         body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1129         mti->mti_attr.ma_need = MA_INODE;
1130         mti->mti_attr.ma_valid = 0;
1131         rc = mo_attr_get(mti->mti_env, mdt_object_child(child), &mti->mti_attr);
1132         if (rc == -EREMOTE) {
1133                 /* object was created on remote server */
1134                 req->rq_status = rc;
1135                 body->valid |= OBD_MD_MDS;
1136         }
1137         mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1138                            mdt_object_fid(child));
1139         mdt_object_put(mti->mti_env, child);
1140 }
1141
1142 static void mdt_reconstruct_setattr(struct mdt_thread_info *mti,
1143                                     struct mdt_lock_handle *lhc)
1144 {
1145         struct ptlrpc_request  *req = mdt_info_req(mti);
1146         struct obd_export *exp = req->rq_export;
1147         struct mdt_export_data *med = &exp->exp_mdt_data;
1148         struct mdt_device *mdt = mti->mti_mdt;
1149         struct mdt_object *obj;
1150         struct mdt_body *body;
1151
1152         mdt_req_from_lcd(req, med->med_lcd);
1153         if (req->rq_status)
1154                 return;
1155
1156         body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1157         obj = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid1);
1158         if (IS_ERR(obj)) {
1159                 int rc = PTR_ERR(obj);
1160                 LCONSOLE_WARN(""DFID" lookup error %d."
1161                               " Evicting client %s with export %s.\n",
1162                               PFID(mdt_object_fid(obj)), rc,
1163                               obd_uuid2str(&exp->exp_client_uuid),
1164                               obd_export_nid2str(exp));
1165                 mdt_export_evict(exp);
1166                 EXIT;
1167                 return;
1168         }
1169         mti->mti_attr.ma_need = MA_INODE;
1170         mti->mti_attr.ma_valid = 0;
1171         mo_attr_get(mti->mti_env, mdt_object_child(obj), &mti->mti_attr);
1172         mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1173                            mdt_object_fid(obj));
1174         if (mti->mti_epoch && (mti->mti_epoch->flags & MF_EPOCH_OPEN)) {
1175                 struct mdt_file_data *mfd;
1176                 struct mdt_body *repbody;
1177
1178                 repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1179                 repbody->ioepoch = obj->mot_ioepoch;
1180                 spin_lock(&med->med_open_lock);
1181                 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
1182                         if (mfd->mfd_xid == req->rq_xid)
1183                                 break;
1184                 }
1185                 LASSERT(&mfd->mfd_list != &med->med_open_head);
1186                 spin_unlock(&med->med_open_lock);
1187                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
1188         }
1189
1190         mdt_object_put(mti->mti_env, obj);
1191 }
1192
1193 typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,
1194                                   struct mdt_lock_handle *lhc);
1195
1196 static mdt_reconstructor reconstructors[REINT_MAX] = {
1197         [REINT_SETATTR]  = mdt_reconstruct_setattr,
1198         [REINT_CREATE]   = mdt_reconstruct_create,
1199         [REINT_LINK]     = mdt_reconstruct_generic,
1200         [REINT_UNLINK]   = mdt_reconstruct_generic,
1201         [REINT_RENAME]   = mdt_reconstruct_generic,
1202         [REINT_OPEN]     = mdt_reconstruct_open,
1203         [REINT_SETXATTR] = mdt_reconstruct_generic
1204 };
1205
1206 void mdt_reconstruct(struct mdt_thread_info *mti,
1207                      struct mdt_lock_handle *lhc)
1208 {
1209         ENTRY;
1210         reconstructors[mti->mti_rr.rr_opcode](mti, lhc);
1211         EXIT;
1212 }