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 (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011 Whamcloud, Inc.
  33  *
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/mdt/mdt_recovery.c
  40  *
  41  * Lustre Metadata Target (mdt) recovery-related methods
  42  *
  43  * Author: Huang Hua <huanghua@clusterfs.com>
  44  * Author: Pershin Mike <tappro@clusterfs.com>
  45  */
  46
  47 #ifndef EXPORT_SYMTAB
  48 # define EXPORT_SYMTAB
  49 #endif
  50 #define DEBUG_SUBSYSTEM S_MDS
  51
  52 #include "mdt_internal.h"
  53
  54 static int mdt_server_data_update(const struct lu_env *env,
  55                                   struct mdt_device *mdt);
  56
  57 struct lu_buf *mdt_buf(const struct lu_env *env, void *area, ssize_t len)
  58 {
  59         struct lu_buf *buf;
  60         struct mdt_thread_info *mti;
  61
  62         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
  63         buf = &mti->mti_buf;
  64         buf->lb_buf = area;
  65         buf->lb_len = len;
  66         return buf;
  67 }
  68
  69 const struct lu_buf *mdt_buf_const(const struct lu_env *env,
  70                                    const void *area, ssize_t len)
  71 {
  72         struct lu_buf *buf;
  73         struct mdt_thread_info *mti;
  74
  75         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
  76         buf = &mti->mti_buf;
  77
  78         buf->lb_buf = (void *)area;
  79         buf->lb_len = len;
  80         return buf;
  81 }
  82
  83 static inline int mdt_trans_credit_get(const struct lu_env *env,
  84                                        struct mdt_device *mdt,
  85                                        enum mdt_txn_op op)
  86 {
  87         struct dt_device *dev = mdt->mdt_bottom;
  88         int cr;
  89         switch (op) {
  90                 case MDT_TXN_CAPA_KEYS_WRITE_OP:
  91                 case MDT_TXN_LAST_RCVD_WRITE_OP:
  92                         cr = dev->dd_ops->dt_credit_get(env,
  93                                                         dev,
  94                                                         DTO_WRITE_BLOCK);
  95                 break;
  96                 default:
  97                         LBUG();
  98         }
  99         return cr;
 100 }
 101
 102 void mdt_trans_credit_init(const struct lu_env *env,
 103                            struct mdt_device *mdt,
 104                            enum mdt_txn_op op)
 105 {
 106         struct mdt_thread_info *mti;
 107         struct txn_param *p;
 108         int cr;
 109
 110         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 111         p = &mti->mti_txn_param;
 112
 113         cr = mdt_trans_credit_get(env, mdt, op);
 114         txn_param_init(p, cr);
 115 }
 116
 117 struct thandle* mdt_trans_start(const struct lu_env *env,
 118                                 struct mdt_device *mdt)
 119 {
 120         struct mdt_thread_info *mti;
 121         struct txn_param *p;
 122
 123         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 124         p = &mti->mti_txn_param;
 125
 126         /* export can require sync operations */
 127         if (mti->mti_exp != NULL)
 128                 p->tp_sync = mti->mti_exp->exp_need_sync;
 129
 130         return mdt->mdt_bottom->dd_ops->dt_trans_start(env, mdt->mdt_bottom, p);
 131 }
 132
 133 void mdt_trans_stop(const struct lu_env *env,
 134                     struct mdt_device *mdt, struct thandle *th)
 135 {
 136         mdt->mdt_bottom->dd_ops->dt_trans_stop(env, th);
 137 }
 138
 139 static inline int mdt_last_rcvd_header_read(const struct lu_env *env,
 140                                             struct mdt_device *mdt)
 141 {
 142         struct mdt_thread_info *mti;
 143         int rc;
 144
 145         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 146
 147         mti->mti_off = 0;
 148         rc = dt_record_read(env, mdt->mdt_lut.lut_last_rcvd,
 149                             mdt_buf(env, &mti->mti_lsd, sizeof(mti->mti_lsd)),
 150                             &mti->mti_off);
 151         if (rc == 0)
 152                 lsd_le_to_cpu(&mti->mti_lsd, &mdt->mdt_lut.lut_lsd);
 153
 154         CDEBUG(D_INFO, "read last_rcvd header rc = %d, uuid = %s, "
 155                "last_transno = "LPU64"\n", rc, mdt->mdt_lut.lut_lsd.lsd_uuid,
 156                mdt->mdt_lut.lut_lsd.lsd_last_transno);
 157         return rc;
 158 }
 159
 160 static int mdt_last_rcvd_header_write(const struct lu_env *env,
 161                                       struct mdt_device *mdt,
 162                                       struct thandle *th)
 163 {
 164         struct mdt_thread_info *mti;
 165         int rc;
 166         ENTRY;
 167
 168         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 169
 170         mti->mti_off = 0;
 171         lsd_cpu_to_le(&mdt->mdt_lut.lut_lsd, &mti->mti_lsd);
 172
 173         rc = dt_record_write(env, mdt->mdt_lut.lut_last_rcvd,
 174                              mdt_buf_const(env, &mti->mti_lsd,
 175                                            sizeof(mti->mti_lsd)),
 176                              &mti->mti_off, th);
 177
 178         CDEBUG(D_INFO, "write last_rcvd header rc = %d, uuid = %s, "
 179                "last_transno = "LPU64"\n", rc, mdt->mdt_lut.lut_lsd.lsd_uuid,
 180                mdt->mdt_lut.lut_lsd.lsd_last_transno);
 181
 182         RETURN(rc);
 183 }
 184
 185 static int mdt_last_rcvd_read(const struct lu_env *env, struct mdt_device *mdt,
 186                               struct lsd_client_data *lcd, loff_t *off,
 187                               int index)
 188 {
 189         struct mdt_thread_info *mti;
 190         struct lsd_client_data *tmp;
 191         int rc;
 192
 193         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 194         tmp = &mti->mti_lcd;
 195         rc = dt_record_read(env, mdt->mdt_lut.lut_last_rcvd,
 196                             mdt_buf(env, tmp, sizeof(*tmp)), off);
 197         if (rc == 0) {
 198                 check_lcd(mdt2obd_dev(mdt)->obd_name, index, tmp);
 199                 lcd_le_to_cpu(tmp, lcd);
 200         }
 201
 202         CDEBUG(D_INFO, "read lcd @%d rc = %d, uuid = %s, last_transno = "LPU64
 203                ", last_xid = "LPU64", last_result = %u, last_data = %u, "
 204                "last_close_transno = "LPU64", last_close_xid = "LPU64", "
 205                "last_close_result = %u\n", (int)(*off - sizeof(*tmp)),
 206                rc, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
 207                lcd->lcd_last_result, lcd->lcd_last_data,
 208                lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
 209                lcd->lcd_last_close_result);
 210         return rc;
 211 }
 212
 213 static int mdt_last_rcvd_write(const struct lu_env *env,
 214                                struct mdt_device *mdt,
 215                                struct lsd_client_data *lcd,
 216                                loff_t *off, struct thandle *th)
 217 {
 218         struct mdt_thread_info *mti;
 219         struct lsd_client_data *tmp;
 220         int rc;
 221
 222         LASSERT(th != NULL);
 223         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 224         tmp = &mti->mti_lcd;
 225
 226         lcd_cpu_to_le(lcd, tmp);
 227
 228         rc = dt_record_write(env, mdt->mdt_lut.lut_last_rcvd,
 229                              mdt_buf_const(env, tmp, sizeof(*tmp)), off, th);
 230
 231         CDEBUG(D_INFO, "write lcd @%d rc = %d, uuid = %s, last_transno = "LPU64
 232                ", last_xid = "LPU64", last_result = %u, last_data = %u, "
 233                "last_close_transno = "LPU64", last_close_xid = "LPU64" ,"
 234                "last_close_result = %u\n", (int)(*off - sizeof(*tmp)),
 235                rc, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
 236                lcd->lcd_last_result, lcd->lcd_last_data,
 237                lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
 238                lcd->lcd_last_close_result);
 239         return rc;
 240 }
 241
 242 static int mdt_clients_data_init(const struct lu_env *env,
 243                                  struct mdt_device *mdt,
 244                                  unsigned long last_size)
 245 {
 246         struct lr_server_data  *lsd = &mdt->mdt_lut.lut_lsd;
 247         struct lsd_client_data *lcd;
 248         struct obd_device      *obd = mdt2obd_dev(mdt);
 249         loff_t off;
 250         int cl_idx;
 251         int rc = 0;
 252         ENTRY;
 253
 254         OBD_ALLOC_PTR(lcd);
 255         if (!lcd)
 256                 RETURN(-ENOMEM);
 257
 258         /* When we do a clean MDS shutdown, we save the last_transno into
 259          * the header.  If we find clients with higher last_transno values
 260          * then those clients may need recovery done. */
 261         LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
 262         for (cl_idx = 0, off = lsd->lsd_client_start;
 263              off < last_size; cl_idx++) {
 264                 __u64 last_transno;
 265                 struct obd_export *exp;
 266                 struct mdt_thread_info *mti;
 267
 268                 off = lsd->lsd_client_start +
 269                         cl_idx * lsd->lsd_client_size;
 270
 271                 rc = mdt_last_rcvd_read(env, mdt, lcd, &off, cl_idx);
 272                 if (rc) {
 273                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 274                                LAST_RCVD, cl_idx, off, rc);
 275                         rc = 0;
 276                         break; /* read error shouldn't cause startup to fail */
 277                 }
 278
 279                 if (lcd->lcd_uuid[0] == '\0') {
 280                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 281                                cl_idx);
 282                         continue;
 283                 }
 284
 285                 last_transno = lcd_last_transno(lcd);
 286
 287                 /* These exports are cleaned up by mdt_obd_disconnect(), so
 288                  * they need to be set up like real exports as
 289                  * mdt_obd_connect() does.
 290                  */
 291                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 292                        " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
 293                        last_transno, lsd->lsd_last_transno,
 294                        lcd_last_xid(lcd));
 295
 296                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
 297                 if (IS_ERR(exp)) {
 298                         if (PTR_ERR(exp) == -EALREADY) {
 299                                 /* export already exists, zero out this one */
 300                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
 301                                 continue;
 302                         }
 303                         GOTO(err_client, rc = PTR_ERR(exp));
 304                 }
 305
 306                 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 307                 LASSERT(mti != NULL);
 308                 mti->mti_exp = exp;
 309                 /* copy on-disk lcd to the export */
 310                 *exp->exp_target_data.ted_lcd = *lcd;
 311                 rc = mdt_client_add(env, mdt, cl_idx);
 312                 /* can't fail existing */
 313                 LASSERTF(rc == 0, "rc = %d\n", rc);
 314                 /* VBR: set export last committed version */
 315                 exp->exp_last_committed = last_transno;
 316                 cfs_spin_lock(&exp->exp_lock);
 317                 exp->exp_connecting = 0;
 318                 exp->exp_in_recovery = 0;
 319                 cfs_spin_unlock(&exp->exp_lock);
 320                 obd->obd_max_recoverable_clients++;
 321                 class_export_put(exp);
 322
 323                 CDEBUG(D_OTHER, "client at idx %d has last_transno="LPU64"\n",
 324                        cl_idx, last_transno);
 325                 /* protect __u64 value update */
 326                 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 327                 mdt->mdt_lut.lut_last_transno = max(last_transno,
 328                                                 mdt->mdt_lut.lut_last_transno);
 329                 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 330         }
 331
 332 err_client:
 333         OBD_FREE_PTR(lcd);
 334         RETURN(rc);
 335 }
 336
 337 static int mdt_server_data_init(const struct lu_env *env,
 338                                 struct mdt_device *mdt,
 339                                 struct lustre_sb_info *lsi)
 340 {
 341         struct lr_server_data  *lsd = &mdt->mdt_lut.lut_lsd;
 342         struct lsd_client_data *lcd = NULL;
 343         struct obd_device      *obd = mdt2obd_dev(mdt);
 344         struct mdt_thread_info *mti;
 345         struct dt_object       *obj;
 346         struct lu_attr         *la;
 347         struct lustre_disk_data  *ldd;
 348         unsigned long last_rcvd_size;
 349         __u64 mount_count;
 350         int rc;
 351         ENTRY;
 352
 353         /* ensure padding in the struct is the correct size */
 354         CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
 355                 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
 356         CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
 357                 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
 358
 359         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 360         LASSERT(mti != NULL);
 361         la = &mti->mti_attr.ma_attr;
 362
 363         obj = mdt->mdt_lut.lut_last_rcvd;
 364         rc = obj->do_ops->do_attr_get(env, obj, la, BYPASS_CAPA);
 365         if (rc)
 366                 RETURN(rc);
 367
 368         last_rcvd_size = (unsigned long)la->la_size;
 369
 370         if (last_rcvd_size == 0) {
 371                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
 372
 373                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,
 374                        sizeof(lsd->lsd_uuid));
 375                 lsd->lsd_last_transno = 0;
 376                 lsd->lsd_mount_count = 0;
 377                 lsd->lsd_server_size = LR_SERVER_SIZE;
 378                 lsd->lsd_client_start = LR_CLIENT_START;
 379                 lsd->lsd_client_size = LR_CLIENT_SIZE;
 380                 lsd->lsd_feature_compat = OBD_COMPAT_MDT;
 381                 lsd->lsd_feature_rocompat = OBD_ROCOMPAT_LOVOBJID;
 382                 lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT |
 383                                             OBD_INCOMPAT_COMMON_LR;
 384         } else {
 385                 LCONSOLE_WARN("%s: used disk, loading\n", obd->obd_name);
 386                 rc = mdt_last_rcvd_header_read(env, mdt);
 387                 if (rc) {
 388                         CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
 389                         GOTO(out, rc);
 390                 }
 391                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
 392                         LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s using"
 393                                            "the wrong disk %s. Were the /dev/ "
 394                                            "assignments rearranged?\n",
 395                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
 396                         GOTO(out, rc = -EINVAL);
 397                 }
 398                 lsd->lsd_feature_compat |= OBD_COMPAT_MDT;
 399                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_MDT |
 400                                              OBD_INCOMPAT_COMMON_LR;
 401         }
 402         mount_count = lsd->lsd_mount_count;
 403
 404         ldd = lsi->lsi_ldd;
 405
 406         if (lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP) {
 407                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
 408                        obd->obd_name,
 409                        lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP);
 410                 GOTO(out, rc = -EINVAL);
 411         }
 412         if (lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP) {
 413                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
 414                        obd->obd_name,
 415                        lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP);
 416                 /* XXX: Do something like remount filesystem read-only */
 417                 GOTO(out, rc = -EINVAL);
 418         }
 419         /** Interop: evict all clients at first boot with 1.8 last_rcvd */
 420         if (!(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
 421                 if (last_rcvd_size > lsd->lsd_client_start) {
 422                         LCONSOLE_WARN("Mounting %s at first time on 1.8 FS, "
 423                                       "remove all clients for interop needs\n",
 424                                       obd->obd_name);
 425                         simple_truncate(lsi->lsi_srv_mnt->mnt_sb->s_root,
 426                                         lsi->lsi_srv_mnt, LAST_RCVD,
 427                                         lsd->lsd_client_start);
 428                         last_rcvd_size = lsd->lsd_client_start;
 429                 }
 430                 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
 431                 lsd->lsd_feature_compat |= OBD_COMPAT_20;
 432         }
 433
 434         if (ldd->ldd_flags & LDD_F_IAM_DIR)
 435                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_IAM_DIR;
 436
 437         lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
 438
 439         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 440         mdt->mdt_lut.lut_last_transno = lsd->lsd_last_transno;
 441         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 442
 443         CDEBUG(D_INODE, "========BEGIN DUMPING LAST_RCVD========\n");
 444         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 445                obd->obd_name, mdt->mdt_lut.lut_last_transno);
 446         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 447                obd->obd_name, mount_count + 1);
 448         CDEBUG(D_INODE, "%s: server data size: %u\n",
 449                obd->obd_name, lsd->lsd_server_size);
 450         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 451                obd->obd_name, lsd->lsd_client_start);
 452         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 453                obd->obd_name, lsd->lsd_client_size);
 454         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 455                obd->obd_name, last_rcvd_size);
 456         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 457                last_rcvd_size <= lsd->lsd_client_start ? 0 :
 458                (last_rcvd_size - lsd->lsd_client_start) /
 459                 lsd->lsd_client_size);
 460         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
 461
 462         if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
 463             !lsd->lsd_client_size) {
 464                 CERROR("Bad last_rcvd contents!\n");
 465                 GOTO(out, rc = -EINVAL);
 466         }
 467
 468         rc = mdt_clients_data_init(env, mdt, last_rcvd_size);
 469         if (rc)
 470                 GOTO(err_client, rc);
 471
 472         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 473         /* obd_last_committed is used for compatibility
 474          * with other lustre recovery code */
 475         obd->obd_last_committed = mdt->mdt_lut.lut_last_transno;
 476         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 477
 478         obd->u.obt.obt_mount_count = mount_count + 1;
 479         lsd->lsd_mount_count = obd->u.obt.obt_mount_count;
 480
 481         /* save it, so mount count and last_transno is current */
 482         rc = mdt_server_data_update(env, mdt);
 483         if (rc)
 484                 GOTO(err_client, rc);
 485
 486         RETURN(0);
 487
 488 err_client:
 489         class_disconnect_exports(obd);
 490 out:
 491         return rc;
 492 }
 493
 494 static int mdt_server_data_update(const struct lu_env *env,
 495                                   struct mdt_device *mdt)
 496 {
 497         struct mdt_thread_info *mti;
 498         struct thandle *th;
 499         int rc;
 500
 501         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 502
 503         mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
 504         th = mdt_trans_start(env, mdt);
 505         if (IS_ERR(th))
 506                 RETURN(PTR_ERR(th));
 507
 508         CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n",
 509                mdt->mdt_lut.lut_obd->u.obt.obt_mount_count,
 510                mdt->mdt_lut.lut_last_transno);
 511
 512         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 513         mdt->mdt_lut.lut_lsd.lsd_last_transno = mdt->mdt_lut.lut_last_transno;
 514         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 515
 516         rc = mdt_last_rcvd_header_write(env, mdt, th);
 517         mdt_trans_stop(env, mdt, th);
 518         return rc;
 519 }
 520
 521
 522 int mdt_client_new(const struct lu_env *env, struct mdt_device *mdt)
 523 {
 524         unsigned long *bitmap = mdt->mdt_lut.lut_client_bitmap;
 525         struct mdt_thread_info *mti;
 526         struct tg_export_data *ted;
 527         struct lr_server_data  *lsd = &mdt->mdt_lut.lut_lsd;
 528         struct obd_device *obd = mdt2obd_dev(mdt);
 529         struct thandle *th;
 530         loff_t off;
 531         int rc;
 532         int cl_idx;
 533         ENTRY;
 534
 535         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 536         LASSERT(mti != NULL);
 537
 538         ted = &mti->mti_exp->exp_target_data;
 539
 540         LASSERT(bitmap != NULL);
 541         if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid))
 542                 RETURN(0);
 543
 544         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 545          * there's no need for extra complication here
 546          */
 547         cfs_spin_lock(&mdt->mdt_lut.lut_client_bitmap_lock);
 548         cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS);
 549         if (cl_idx >= LR_MAX_CLIENTS ||
 550             OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
 551                 CERROR("no room for %u clients - fix LR_MAX_CLIENTS\n",
 552                        cl_idx);
 553                 cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock);
 554                 RETURN(-EOVERFLOW);
 555         }
 556         cfs_set_bit(cl_idx, bitmap);
 557         cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock);
 558
 559         CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
 560                cl_idx, ted->ted_lcd->lcd_uuid);
 561
 562         ted->ted_lr_idx = cl_idx;
 563         ted->ted_lr_off = lsd->lsd_client_start +
 564                           (cl_idx * lsd->lsd_client_size);
 565         cfs_init_mutex(&ted->ted_lcd_lock);
 566
 567         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
 568
 569         /* Write new client data. */
 570         off = ted->ted_lr_off;
 571
 572         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
 573                 RETURN(-ENOSPC);
 574
 575         mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
 576
 577         th = mdt_trans_start(env, mdt);
 578         if (IS_ERR(th))
 579                 RETURN(PTR_ERR(th));
 580
 581         /*
 582          * Until this operations will be committed the sync is needed
 583          * for this export. This should be done _after_ starting the
 584          * transaction so that many connecting clients will not bring
 585          * server down with lots of sync writes.
 586          */
 587         mdt_trans_add_cb(th, lut_cb_client, class_export_cb_get(mti->mti_exp));
 588         cfs_spin_lock(&mti->mti_exp->exp_lock);
 589         mti->mti_exp->exp_need_sync = 1;
 590         cfs_spin_unlock(&mti->mti_exp->exp_lock);
 591
 592         rc = mdt_last_rcvd_write(env, mdt, ted->ted_lcd, &off, th);
 593         CDEBUG(D_INFO, "wrote client lcd at idx %u off %llu (len %u)\n",
 594                cl_idx, ted->ted_lr_off, (int)sizeof(*(ted->ted_lcd)));
 595         mdt_trans_stop(env, mdt, th);
 596
 597         RETURN(rc);
 598 }
 599
 600 /* Add client data to the MDS.  We use a bitmap to locate a free space
 601  * in the last_rcvd file if cl_off is -1 (i.e. a new client).
 602  * Otherwise, we just have to read the data from the last_rcvd file and
 603  * we know its offset.
 604  *
 605  * It should not be possible to fail adding an existing client - otherwise
 606  * mdt_init_server_data() callsite needs to be fixed.
 607  */
 608 int mdt_client_add(const struct lu_env *env,
 609                    struct mdt_device *mdt, int cl_idx)
 610 {
 611         struct mdt_thread_info *mti;
 612         struct tg_export_data  *ted;
 613         unsigned long *bitmap = mdt->mdt_lut.lut_client_bitmap;
 614         struct obd_device *obd = mdt2obd_dev(mdt);
 615         struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd;
 616         int rc = 0;
 617         ENTRY;
 618
 619         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 620         LASSERT(mti != NULL);
 621
 622         ted = &mti->mti_exp->exp_target_data;
 623
 624         LASSERT(bitmap != NULL);
 625         LASSERTF(cl_idx >= 0, "%d\n", cl_idx);
 626
 627         if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid))
 628                 RETURN(0);
 629
 630         cfs_spin_lock(&mdt->mdt_lut.lut_client_bitmap_lock);
 631         if (cfs_test_and_set_bit(cl_idx, bitmap)) {
 632                 CERROR("MDS client %d: bit already set in bitmap!!\n",
 633                        cl_idx);
 634                 LBUG();
 635         }
 636         cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock);
 637
 638         CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n",
 639                cl_idx, ted->ted_lcd->lcd_uuid);
 640
 641         ted->ted_lr_idx = cl_idx;
 642         ted->ted_lr_off = lsd->lsd_client_start +
 643                           (cl_idx * lsd->lsd_client_size);
 644         cfs_init_mutex(&ted->ted_lcd_lock);
 645
 646         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
 647
 648         RETURN(rc);
 649 }
 650
 651 int mdt_client_del(const struct lu_env *env, struct mdt_device *mdt)
 652 {
 653         struct mdt_thread_info *mti;
 654         struct tg_export_data  *ted;
 655         struct obd_device      *obd = mdt2obd_dev(mdt);
 656         struct obd_export      *exp;
 657         struct thandle         *th;
 658         loff_t                  off;
 659         int                     rc = 0;
 660         ENTRY;
 661
 662         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 663         LASSERT(mti != NULL);
 664
 665         exp = mti->mti_exp;
 666         ted = &exp->exp_target_data;
 667         if (!ted->ted_lcd)
 668                 RETURN(0);
 669
 670         /* XXX: If lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 671         if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid))
 672                 GOTO(free, 0);
 673
 674         CDEBUG(D_INFO, "freeing client at idx %u, offset %lld\n",
 675                ted->ted_lr_idx, ted->ted_lr_off);
 676
 677         off = ted->ted_lr_off;
 678
 679         /*
 680          * Don't clear ted_lr_idx here as it is likely also unset.  At worst we
 681          * leak a client slot that will be cleaned on the next recovery.
 682          */
 683         if (off <= 0) {
 684                 CERROR("client idx %d has offset %lld\n",
 685                         ted->ted_lr_idx, off);
 686                 GOTO(free, rc = -EINVAL);
 687         }
 688
 689         /*
 690          * Clear the bit _after_ zeroing out the client so we don't race with
 691          * mdt_client_add and zero out new clients.
 692          */
 693         if (!cfs_test_bit(ted->ted_lr_idx, mdt->mdt_lut.lut_client_bitmap)) {
 694                 CERROR("MDT client %u: bit already clear in bitmap!!\n",
 695                        ted->ted_lr_idx);
 696                 LBUG();
 697         }
 698
 699         /* Make sure the server's last_transno is up to date.
 700          * This should be done before zeroing client slot so last_transno will
 701          * be in server data or in client data in case of failure */
 702         mdt_server_data_update(env, mdt);
 703
 704         mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP);
 705         th = mdt_trans_start(env, mdt);
 706         if (IS_ERR(th))
 707                 GOTO(free, rc = PTR_ERR(th));
 708
 709         cfs_mutex_down(&ted->ted_lcd_lock);
 710         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
 711         rc = mdt_last_rcvd_write(env, mdt, ted->ted_lcd, &off, th);
 712         cfs_mutex_up(&ted->ted_lcd_lock);
 713         mdt_trans_stop(env, mdt, th);
 714
 715         CDEBUG(rc == 0 ? D_INFO : D_ERROR, "Zeroing out client idx %u in "
 716                "%s, rc %d\n",  ted->ted_lr_idx, LAST_RCVD, rc);
 717         RETURN(0);
 718 free:
 719         return 0;
 720 }
 721
 722 /*
 723  * last_rcvd & last_committed update callbacks
 724  */
 725 static int mdt_last_rcvd_update(struct mdt_thread_info *mti,
 726                                 struct thandle *th)
 727 {
 728         struct mdt_device *mdt = mti->mti_mdt;
 729         struct ptlrpc_request *req = mdt_info_req(mti);
 730         struct tg_export_data *ted;
 731         struct lsd_client_data *lcd;
 732         loff_t off;
 733         int err;
 734         __s32 rc = th->th_result;
 735
 736         ENTRY;
 737         LASSERT(req);
 738         LASSERT(req->rq_export);
 739         LASSERT(mdt);
 740         ted = &req->rq_export->exp_target_data;
 741         LASSERT(ted);
 742
 743         cfs_mutex_down(&ted->ted_lcd_lock);
 744         lcd = ted->ted_lcd;
 745         /* if the export has already been disconnected, we have no last_rcvd slot,
 746          * update server data with latest transno then */
 747         if (lcd == NULL) {
 748                 cfs_mutex_up(&ted->ted_lcd_lock);
 749                 CWARN("commit transaction for disconnected client %s: rc %d\n",
 750                       req->rq_export->exp_client_uuid.uuid, rc);
 751                 err = mdt_last_rcvd_header_write(mti->mti_env, mdt, th);
 752                 RETURN(err);
 753         }
 754
 755         off = ted->ted_lr_off;
 756         LASSERT(ergo(mti->mti_transno == 0, rc != 0));
 757         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 758             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 759                 if (mti->mti_transno != 0) {
 760                         if (lcd->lcd_last_close_transno > mti->mti_transno) {
 761                                 LASSERT(req_is_replay(req));
 762                                 CERROR("Trying to overwrite bigger transno:"
 763                                        "on-disk: "LPU64", new: "LPU64"\n",
 764                                        lcd->lcd_last_close_transno,
 765                                        mti->mti_transno);
 766                                 cfs_spin_lock(&req->rq_export->exp_lock);
 767                                 req->rq_export->exp_vbr_failed = 1;
 768                                 cfs_spin_unlock(&req->rq_export->exp_lock);
 769                                 cfs_mutex_up(&ted->ted_lcd_lock);
 770                                 RETURN(-EOVERFLOW);
 771                         }
 772                         lcd->lcd_last_close_transno = mti->mti_transno;
 773                 }
 774                 lcd->lcd_last_close_xid = req->rq_xid;
 775                 lcd->lcd_last_close_result = rc;
 776         } else {
 777                 /* VBR: save versions in last_rcvd for reconstruct. */
 778                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 779                 if (pre_versions) {
 780                         lcd->lcd_pre_versions[0] = pre_versions[0];
 781                         lcd->lcd_pre_versions[1] = pre_versions[1];
 782                         lcd->lcd_pre_versions[2] = pre_versions[2];
 783                         lcd->lcd_pre_versions[3] = pre_versions[3];
 784                 }
 785                 if (mti->mti_transno != 0) {
 786                         if (lcd->lcd_last_transno > mti->mti_transno) {
 787                                 LASSERT(req_is_replay(req));
 788                                 CERROR("Trying to overwrite bigger transno:"
 789                                        "on-disk: "LPU64", new: "LPU64"\n",
 790                                        lcd->lcd_last_transno,
 791                                        mti->mti_transno);
 792                                 cfs_spin_lock(&req->rq_export->exp_lock);
 793                                 req->rq_export->exp_vbr_failed = 1;
 794                                 cfs_spin_unlock(&req->rq_export->exp_lock);
 795                                 cfs_mutex_up(&ted->ted_lcd_lock);
 796                                 RETURN(-EOVERFLOW);
 797                         }
 798                         lcd->lcd_last_transno = mti->mti_transno;
 799                 }
 800                 lcd->lcd_last_xid = req->rq_xid;
 801                 lcd->lcd_last_result = rc;
 802                 /*XXX: save intent_disposition in mdt_thread_info?
 803                  * also there is bug - intent_dispostion is __u64,
 804                  * see struct ldlm_reply->lock_policy_res1; */
 805                 lcd->lcd_last_data = mti->mti_opdata;
 806         }
 807
 808         if (off <= 0) {
 809                 CERROR("client idx %d has offset %lld\n", ted->ted_lr_idx, off);
 810                 err = -EINVAL;
 811         } else {
 812                 err = mdt_last_rcvd_write(mti->mti_env, mdt, lcd, &off, th);
 813         }
 814         cfs_mutex_up(&ted->ted_lcd_lock);
 815         RETURN(err);
 816 }
 817
 818 extern struct lu_context_key mdt_thread_key;
 819
 820 /* add credits for last_rcvd update */
 821 static int mdt_txn_start_cb(const struct lu_env *env,
 822                             struct txn_param *param, void *cookie)
 823 {
 824         struct mdt_device *mdt = cookie;
 825
 826         param->tp_credits += mdt_trans_credit_get(env, mdt,
 827                                                   MDT_TXN_LAST_RCVD_WRITE_OP);
 828         return 0;
 829 }
 830
 831 /* Set new object versions */
 832 static void mdt_version_set(struct mdt_thread_info *info)
 833 {
 834         if (info->mti_mos != NULL) {
 835                 mo_version_set(info->mti_env, mdt_object_child(info->mti_mos),
 836                                info->mti_transno);
 837                 info->mti_mos = NULL;
 838         }
 839 }
 840
 841 /* Update last_rcvd records with latests transaction data */
 842 static int mdt_txn_stop_cb(const struct lu_env *env,
 843                            struct thandle *txn, void *cookie)
 844 {
 845         struct mdt_device *mdt = cookie;
 846         struct mdt_txn_info *txi;
 847         struct mdt_thread_info *mti;
 848         struct ptlrpc_request *req;
 849
 850         /* transno in two contexts - for commit_cb and for thread */
 851         txi = lu_context_key_get(&txn->th_ctx, &mdt_txn_key);
 852         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 853         req = mdt_info_req(mti);
 854
 855         if (mti->mti_mdt == NULL || req == NULL || mti->mti_no_need_trans) {
 856                 txi->txi_transno = 0;
 857                 mti->mti_no_need_trans = 0;
 858                 return 0;
 859         }
 860
 861         if (mti->mti_has_trans) {
 862                 /* XXX: currently there are allowed cases, but the wrong cases
 863                  * are also possible, so better check is needed here */
 864                 CDEBUG(D_INFO, "More than one transaction "LPU64"\n",
 865                        mti->mti_transno);
 866                 return 0;
 867         }
 868
 869         mti->mti_has_trans = 1;
 870         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 871         if (txn->th_result != 0) {
 872                 if (mti->mti_transno != 0) {
 873                         CERROR("Replay transno "LPU64" failed: rc %d\n",
 874                                mti->mti_transno, txn->th_result);
 875                 }
 876         } else if (mti->mti_transno == 0) {
 877                 mti->mti_transno = ++ mdt->mdt_lut.lut_last_transno;
 878         } else {
 879                 /* should be replay */
 880                 if (mti->mti_transno > mdt->mdt_lut.lut_last_transno)
 881                         mdt->mdt_lut.lut_last_transno = mti->mti_transno;
 882         }
 883         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 884         /* sometimes the reply message has not been successfully packed */
 885         LASSERT(req != NULL && req->rq_repmsg != NULL);
 886
 887         /** VBR: set new versions */
 888         if (txn->th_result == 0)
 889                 mdt_version_set(mti);
 890
 891         /* filling reply data */
 892         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 893                mti->mti_transno, req->rq_export->exp_obd->obd_last_committed);
 894
 895         req->rq_transno = mti->mti_transno;
 896         lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno);
 897         /* save transno for the commit callback */
 898         txi->txi_transno = mti->mti_transno;
 899
 900         /* add separate commit callback for transaction handling because we need
 901          * export as parameter */
 902         mdt_trans_add_cb(txn, lut_cb_last_committed,
 903                          class_export_cb_get(mti->mti_exp));
 904
 905         return mdt_last_rcvd_update(mti, txn);
 906 }
 907
 908 /* commit callback, need to update last_committed value */
 909 static int mdt_txn_commit_cb(const struct lu_env *env,
 910                              struct thandle *txn, void *cookie)
 911 {
 912         struct mdt_device *mdt = cookie;
 913         struct mdt_txn_info *txi;
 914         int i;
 915
 916         txi = lu_context_key_get(&txn->th_ctx, &mdt_txn_key);
 917
 918         /* iterate through all additional callbacks */
 919         for (i = 0; i < txi->txi_cb_count; i++) {
 920                 txi->txi_cb[i].lut_cb_func(&mdt->mdt_lut, txi->txi_transno,
 921                                            txi->txi_cb[i].lut_cb_data, 0);
 922         }
 923         return 0;
 924 }
 925
 926 int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt,
 927                  struct obd_device *obd,
 928                  struct lustre_sb_info *lsi)
 929 {
 930         struct lu_fid fid;
 931         struct dt_object *o;
 932         int rc = 0;
 933         ENTRY;
 934
 935         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
 936                 RETURN(-ENOENT);
 937
 938         /* prepare transactions callbacks */
 939         mdt->mdt_txn_cb.dtc_txn_start = mdt_txn_start_cb;
 940         mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb;
 941         mdt->mdt_txn_cb.dtc_txn_commit = mdt_txn_commit_cb;
 942         mdt->mdt_txn_cb.dtc_cookie = mdt;
 943         mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD;
 944         CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage);
 945
 946         dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb);
 947
 948         rc = mdt_server_data_init(env, mdt, lsi);
 949         if (rc)
 950                 RETURN(rc);
 951
 952         o = dt_store_open(env, mdt->mdt_bottom, "", CAPA_KEYS, &fid);
 953         if (!IS_ERR(o)) {
 954                 mdt->mdt_ck_obj = o;
 955                 rc = mdt_capa_keys_init(env, mdt);
 956                 if (rc)
 957                         GOTO(put_ck_object, rc);
 958         } else {
 959                 rc = PTR_ERR(o);
 960                 CERROR("cannot open %s: rc = %d\n", CAPA_KEYS, rc);
 961                 GOTO(disconnect_exports, rc);
 962         }
 963         RETURN(0);
 964
 965 put_ck_object:
 966         lu_object_put(env, &o->do_lu);
 967         mdt->mdt_ck_obj = NULL;
 968 disconnect_exports:
 969         class_disconnect_exports(obd);
 970         return rc;
 971 }
 972
 973 void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt)
 974 {
 975         ENTRY;
 976
 977         /* Remove transaction callback */
 978         dt_txn_callback_del(mdt->mdt_bottom, &mdt->mdt_txn_cb);
 979         if (mdt->mdt_ck_obj)
 980                 lu_object_put(env, &mdt->mdt_ck_obj->do_lu);
 981         mdt->mdt_ck_obj = NULL;
 982         EXIT;
 983 }
 984
 985 /* reconstruction code */
 986 static void mdt_steal_ack_locks(struct ptlrpc_request *req)
 987 {
 988         struct obd_export         *exp = req->rq_export;
 989         cfs_list_t                *tmp;
 990         struct ptlrpc_reply_state *oldrep;
 991         struct ptlrpc_service     *svc;
 992         int                        i;
 993
 994         /* CAVEAT EMPTOR: spinlock order */
 995         cfs_spin_lock(&exp->exp_lock);
 996         cfs_list_for_each (tmp, &exp->exp_outstanding_replies) {
 997                 oldrep = cfs_list_entry(tmp, struct ptlrpc_reply_state,
 998                                         rs_exp_list);
 999
1000                 if (oldrep->rs_xid != req->rq_xid)
1001                         continue;
1002
1003                 if (oldrep->rs_opc != lustre_msg_get_opc(req->rq_reqmsg))
1004                         CERROR ("Resent req xid "LPU64" has mismatched opc: "
1005                                 "new %d old %d\n", req->rq_xid,
1006                                 lustre_msg_get_opc(req->rq_reqmsg),
1007                                 oldrep->rs_opc);
1008
1009                 svc = oldrep->rs_service;
1010                 cfs_spin_lock (&svc->srv_rs_lock);
1011
1012                 cfs_list_del_init (&oldrep->rs_exp_list);
1013
1014                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
1015                       " o%d NID %s\n",
1016                       oldrep->rs_nlocks, oldrep,
1017                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_opc,
1018                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
1019
1020                 for (i = 0; i < oldrep->rs_nlocks; i++)
1021                         ptlrpc_save_lock(req, &oldrep->rs_locks[i],
1022                                          oldrep->rs_modes[i], 0);
1023                 oldrep->rs_nlocks = 0;
1024
1025                 DEBUG_REQ(D_HA, req, "stole locks for");
1026                 cfs_spin_lock(&oldrep->rs_lock);
1027                 ptlrpc_schedule_difficult_reply (oldrep);
1028                 cfs_spin_unlock(&oldrep->rs_lock);
1029
1030                 cfs_spin_unlock (&svc->srv_rs_lock);
1031                 break;
1032         }
1033         cfs_spin_unlock(&exp->exp_lock);
1034 }
1035
1036 /**
1037  * VBR: restore versions
1038  */
1039 void mdt_vbr_reconstruct(struct ptlrpc_request *req,
1040                          struct lsd_client_data *lcd)
1041 {
1042         __u64 pre_versions[4] = {0};
1043         pre_versions[0] = lcd->lcd_pre_versions[0];
1044         pre_versions[1] = lcd->lcd_pre_versions[1];
1045         pre_versions[2] = lcd->lcd_pre_versions[2];
1046         pre_versions[3] = lcd->lcd_pre_versions[3];
1047         lustre_msg_set_versions(req->rq_repmsg, pre_versions);
1048 }
1049
1050 void mdt_req_from_lcd(struct ptlrpc_request *req,
1051                       struct lsd_client_data *lcd)
1052 {
1053         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
1054                   lcd->lcd_last_transno, lcd->lcd_last_result);
1055
1056         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
1057             lustre_msg_get_opc(req->rq_repmsg) == MDS_DONE_WRITING) {
1058                 req->rq_transno = lcd->lcd_last_close_transno;
1059                 req->rq_status = lcd->lcd_last_close_result;
1060         } else {
1061                 req->rq_transno = lcd->lcd_last_transno;
1062                 req->rq_status = lcd->lcd_last_result;
1063                 mdt_vbr_reconstruct(req, lcd);
1064         }
1065         if (req->rq_status != 0)
1066                 req->rq_transno = 0;
1067         lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
1068         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1069         DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
1070                   req->rq_transno, req->rq_status);
1071
1072         mdt_steal_ack_locks(req);
1073 }
1074
1075 void mdt_reconstruct_generic(struct mdt_thread_info *mti,
1076                              struct mdt_lock_handle *lhc)
1077 {
1078         struct ptlrpc_request *req = mdt_info_req(mti);
1079         struct tg_export_data *ted = &req->rq_export->exp_target_data;
1080
1081         return mdt_req_from_lcd(req, ted->ted_lcd);
1082 }
1083
1084 static void mdt_reconstruct_create(struct mdt_thread_info *mti,
1085                                    struct mdt_lock_handle *lhc)
1086 {
1087         struct ptlrpc_request  *req = mdt_info_req(mti);
1088         struct obd_export *exp = req->rq_export;
1089         struct tg_export_data *ted = &exp->exp_target_data;
1090         struct mdt_device *mdt = mti->mti_mdt;
1091         struct mdt_object *child;
1092         struct mdt_body *body;
1093         int rc;
1094
1095         mdt_req_from_lcd(req, ted->ted_lcd);
1096         if (req->rq_status)
1097                 return;
1098
1099         /* if no error, so child was created with requested fid */
1100         child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2);
1101         if (IS_ERR(child)) {
1102                 rc = PTR_ERR(child);
1103                 LCONSOLE_WARN("Child "DFID" lookup error %d."
1104                               " Evicting client %s with export %s.\n",
1105                               PFID(mdt_object_fid(child)), rc,
1106                               obd_uuid2str(&exp->exp_client_uuid),
1107                               obd_export_nid2str(exp));
1108                 mdt_export_evict(exp);
1109                 EXIT;
1110                 return;
1111         }
1112
1113         body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1114         mti->mti_attr.ma_need = MA_INODE;
1115         mti->mti_attr.ma_valid = 0;
1116         rc = mo_attr_get(mti->mti_env, mdt_object_child(child), &mti->mti_attr);
1117         if (rc == -EREMOTE) {
1118                 /* object was created on remote server */
1119                 req->rq_status = rc;
1120                 body->valid |= OBD_MD_MDS;
1121         }
1122         mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1123                            mdt_object_fid(child));
1124         mdt_object_put(mti->mti_env, child);
1125 }
1126
1127 static void mdt_reconstruct_setattr(struct mdt_thread_info *mti,
1128                                     struct mdt_lock_handle *lhc)
1129 {
1130         struct ptlrpc_request  *req = mdt_info_req(mti);
1131         struct obd_export *exp = req->rq_export;
1132         struct mdt_export_data *med = &exp->exp_mdt_data;
1133         struct mdt_device *mdt = mti->mti_mdt;
1134         struct mdt_object *obj;
1135         struct mdt_body *body;
1136
1137         mdt_req_from_lcd(req, med->med_ted.ted_lcd);
1138         if (req->rq_status)
1139                 return;
1140
1141         body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1142         obj = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid1);
1143         if (IS_ERR(obj)) {
1144                 int rc = PTR_ERR(obj);
1145                 LCONSOLE_WARN(""DFID" lookup error %d."
1146                               " Evicting client %s with export %s.\n",
1147                               PFID(mdt_object_fid(obj)), rc,
1148                               obd_uuid2str(&exp->exp_client_uuid),
1149                               obd_export_nid2str(exp));
1150                 mdt_export_evict(exp);
1151                 EXIT;
1152                 return;
1153         }
1154         mti->mti_attr.ma_need = MA_INODE;
1155         mti->mti_attr.ma_valid = 0;
1156         mo_attr_get(mti->mti_env, mdt_object_child(obj), &mti->mti_attr);
1157         mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
1158                            mdt_object_fid(obj));
1159         if (mti->mti_ioepoch && (mti->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
1160                 struct mdt_file_data *mfd;
1161                 struct mdt_body *repbody;
1162
1163                 repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
1164                 repbody->ioepoch = obj->mot_ioepoch;
1165                 cfs_spin_lock(&med->med_open_lock);
1166                 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
1167                         if (mfd->mfd_xid == req->rq_xid)
1168                                 break;
1169                 }
1170                 LASSERT(&mfd->mfd_list != &med->med_open_head);
1171                 cfs_spin_unlock(&med->med_open_lock);
1172                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
1173         }
1174
1175         mdt_object_put(mti->mti_env, obj);
1176 }
1177
1178 typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,
1179                                   struct mdt_lock_handle *lhc);
1180
1181 static mdt_reconstructor reconstructors[REINT_MAX] = {
1182         [REINT_SETATTR]  = mdt_reconstruct_setattr,
1183         [REINT_CREATE]   = mdt_reconstruct_create,
1184         [REINT_LINK]     = mdt_reconstruct_generic,
1185         [REINT_UNLINK]   = mdt_reconstruct_generic,
1186         [REINT_RENAME]   = mdt_reconstruct_generic,
1187         [REINT_OPEN]     = mdt_reconstruct_open,
1188         [REINT_SETXATTR] = mdt_reconstruct_generic
1189 };
1190
1191 void mdt_reconstruct(struct mdt_thread_info *mti,
1192                      struct mdt_lock_handle *lhc)
1193 {
1194         ENTRY;
1195         reconstructors[mti->mti_rr.rr_opcode](mti, lhc);
1196         EXIT;
1197 }