lustre/mdt/mdt_recovery.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  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 #define DEBUG_SUBSYSTEM S_MDS
  45
  46 #include "mdt_internal.h"
  47
  48 struct lu_buf *mdt_buf(const struct lu_env *env, void *area, ssize_t len)
  49 {
  50         struct lu_buf *buf;
  51         struct mdt_thread_info *mti;
  52
  53         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
  54         buf = &mti->mti_buf;
  55         buf->lb_buf = area;
  56         buf->lb_len = len;
  57         return buf;
  58 }
  59
  60 const struct lu_buf *mdt_buf_const(const struct lu_env *env,
  61                                    const void *area, ssize_t len)
  62 {
  63         struct lu_buf *buf;
  64         struct mdt_thread_info *mti;
  65
  66         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
  67         buf = &mti->mti_buf;
  68
  69         buf->lb_buf = (void *)area;
  70         buf->lb_len = len;
  71         return buf;
  72 }
  73
  74 void mdt_trans_stop(const struct lu_env *env,
  75                     struct mdt_device *mdt, struct thandle *th)
  76 {
  77         dt_trans_stop(env, mdt->mdt_bottom, th);
  78 }
  79
  80 static int mdt_clients_data_init(const struct lu_env *env,
  81                                  struct mdt_device *mdt,
  82                                  unsigned long last_size)
  83 {
  84         struct lr_server_data  *lsd = &mdt->mdt_lut.lut_lsd;
  85         struct lsd_client_data *lcd;
  86         struct obd_device      *obd = mdt2obd_dev(mdt);
  87         loff_t off;
  88         int cl_idx;
  89         int rc = 0;
  90         ENTRY;
  91
  92         OBD_ALLOC_PTR(lcd);
  93         if (!lcd)
  94                 RETURN(-ENOMEM);
  95
  96         /* When we do a clean MDS shutdown, we save the last_transno into
  97          * the header.  If we find clients with higher last_transno values
  98          * then those clients may need recovery done. */
  99         LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
 100         for (cl_idx = 0, off = lsd->lsd_client_start;
 101              off < last_size; cl_idx++) {
 102                 __u64 last_transno;
 103                 struct obd_export *exp;
 104                 struct mdt_thread_info *mti;
 105
 106                 off = lsd->lsd_client_start +
 107                         cl_idx * lsd->lsd_client_size;
 108
 109                 rc = lut_client_data_read(env, &mdt->mdt_lut, lcd, &off, cl_idx);
 110                 if (rc) {
 111                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 112                                LAST_RCVD, cl_idx, off, rc);
 113                         rc = 0;
 114                         break; /* read error shouldn't cause startup to fail */
 115                 }
 116
 117                 if (lcd->lcd_uuid[0] == '\0') {
 118                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 119                                cl_idx);
 120                         continue;
 121                 }
 122
 123                 last_transno = lcd_last_transno(lcd);
 124
 125                 /* These exports are cleaned up by mdt_obd_disconnect(), so
 126                  * they need to be set up like real exports as
 127                  * mdt_obd_connect() does.
 128                  */
 129                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 130                        " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
 131                        last_transno, lsd->lsd_last_transno,
 132                        lcd_last_xid(lcd));
 133
 134                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
 135                 if (IS_ERR(exp)) {
 136                         if (PTR_ERR(exp) == -EALREADY) {
 137                                 /* export already exists, zero out this one */
 138                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
 139                                 continue;
 140                         }
 141                         GOTO(err_client, rc = PTR_ERR(exp));
 142                 }
 143
 144                 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 145                 LASSERT(mti != NULL);
 146                 mti->mti_exp = exp;
 147                 /* copy on-disk lcd to the export */
 148                 *exp->exp_target_data.ted_lcd = *lcd;
 149                 rc = lut_client_add(env, exp, cl_idx);
 150                 /* can't fail existing */
 151                 LASSERTF(rc == 0, "rc = %d\n", rc);
 152                 /* VBR: set export last committed version */
 153                 exp->exp_last_committed = last_transno;
 154                 cfs_spin_lock(&exp->exp_lock);
 155                 exp->exp_connecting = 0;
 156                 exp->exp_in_recovery = 0;
 157                 cfs_spin_unlock(&exp->exp_lock);
 158                 obd->obd_max_recoverable_clients++;
 159                 class_export_put(exp);
 160
 161                 CDEBUG(D_OTHER, "client at idx %d has last_transno="LPU64"\n",
 162                        cl_idx, last_transno);
 163                 /* protect __u64 value update */
 164                 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 165                 mdt->mdt_lut.lut_last_transno = max(last_transno,
 166                                                 mdt->mdt_lut.lut_last_transno);
 167                 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 168         }
 169
 170 err_client:
 171         OBD_FREE_PTR(lcd);
 172         RETURN(rc);
 173 }
 174
 175 static int mdt_server_data_init(const struct lu_env *env,
 176                                 struct mdt_device *mdt,
 177                                 struct lustre_sb_info *lsi)
 178 {
 179         struct lr_server_data  *lsd = &mdt->mdt_lut.lut_lsd;
 180         struct lsd_client_data *lcd = NULL;
 181         struct obd_device      *obd = mdt2obd_dev(mdt);
 182         struct mdt_thread_info *mti;
 183         struct dt_object       *obj;
 184         struct lu_attr         *la;
 185         unsigned long last_rcvd_size;
 186         __u64 mount_count;
 187         int rc;
 188         ENTRY;
 189
 190         /* ensure padding in the struct is the correct size */
 191         CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
 192                 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
 193         CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
 194                 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
 195
 196         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 197         LASSERT(mti != NULL);
 198         la = &mti->mti_attr.ma_attr;
 199
 200         obj = mdt->mdt_lut.lut_last_rcvd;
 201         rc = dt_attr_get(env, obj, la, BYPASS_CAPA);
 202         if (rc)
 203                 RETURN(rc);
 204
 205         last_rcvd_size = (unsigned long)la->la_size;
 206
 207         if (last_rcvd_size == 0) {
 208                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
 209
 210                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,
 211                        sizeof(lsd->lsd_uuid));
 212                 lsd->lsd_last_transno = 0;
 213                 lsd->lsd_mount_count = 0;
 214                 lsd->lsd_server_size = LR_SERVER_SIZE;
 215                 lsd->lsd_client_start = LR_CLIENT_START;
 216                 lsd->lsd_client_size = LR_CLIENT_SIZE;
 217                 lsd->lsd_feature_compat = OBD_COMPAT_MDT;
 218                 lsd->lsd_feature_rocompat = OBD_ROCOMPAT_LOVOBJID;
 219                 lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT |
 220                                             OBD_INCOMPAT_COMMON_LR |
 221                                             OBD_INCOMPAT_MULTI_OI;
 222         } else {
 223                 LCONSOLE_WARN("%s: used disk, loading\n", obd->obd_name);
 224                 rc = lut_server_data_read(env, &mdt->mdt_lut);
 225                 if (rc) {
 226                         CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
 227                         GOTO(out, rc);
 228                 }
 229                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
 230                         LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s using"
 231                                            "the wrong disk %s. Were the /dev/ "
 232                                            "assignments rearranged?\n",
 233                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
 234                         GOTO(out, rc = -EINVAL);
 235                 }
 236                 lsd->lsd_feature_compat |= OBD_COMPAT_MDT;
 237                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_MDT |
 238                                              OBD_INCOMPAT_COMMON_LR;
 239         }
 240         mount_count = lsd->lsd_mount_count;
 241
 242         if (lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP) {
 243                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
 244                        obd->obd_name,
 245                        lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP);
 246                 GOTO(out, rc = -EINVAL);
 247         }
 248         if (lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP) {
 249                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
 250                        obd->obd_name,
 251                        lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP);
 252                 /* XXX: Do something like remount filesystem read-only */
 253                 GOTO(out, rc = -EINVAL);
 254         }
 255         /** Interop: evict all clients at first boot with 1.8 last_rcvd */
 256         if (!(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
 257                 if (last_rcvd_size > lsd->lsd_client_start) {
 258                         LCONSOLE_WARN("Mounting %s at first time on 1.8 FS, "
 259                                       "remove all clients for interop needs\n",
 260                                       obd->obd_name);
 261                         rc = lut_truncate_last_rcvd(env, &mdt->mdt_lut,
 262                                                     lsd->lsd_client_start);
 263                         if (rc)
 264                                 GOTO(out, rc);
 265                         last_rcvd_size = lsd->lsd_client_start;
 266                 }
 267                 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
 268                 lsd->lsd_feature_compat |= OBD_COMPAT_20;
 269         }
 270
 271         if (lsi->lsi_flags & LDD_F_IAM_DIR)
 272                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_IAM_DIR;
 273
 274         lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
 275
 276         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 277         mdt->mdt_lut.lut_last_transno = lsd->lsd_last_transno;
 278         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 279
 280         CDEBUG(D_INODE, "========BEGIN DUMPING LAST_RCVD========\n");
 281         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 282                obd->obd_name, mdt->mdt_lut.lut_last_transno);
 283         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 284                obd->obd_name, mount_count + 1);
 285         CDEBUG(D_INODE, "%s: server data size: %u\n",
 286                obd->obd_name, lsd->lsd_server_size);
 287         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 288                obd->obd_name, lsd->lsd_client_start);
 289         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 290                obd->obd_name, lsd->lsd_client_size);
 291         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 292                obd->obd_name, last_rcvd_size);
 293         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 294                last_rcvd_size <= lsd->lsd_client_start ? 0 :
 295                (last_rcvd_size - lsd->lsd_client_start) /
 296                 lsd->lsd_client_size);
 297         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
 298
 299         if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
 300             !lsd->lsd_client_size) {
 301                 CERROR("Bad last_rcvd contents!\n");
 302                 GOTO(out, rc = -EINVAL);
 303         }
 304
 305         rc = mdt_clients_data_init(env, mdt, last_rcvd_size);
 306         if (rc)
 307                 GOTO(err_client, rc);
 308
 309         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 310         /* obd_last_committed is used for compatibility
 311          * with other lustre recovery code */
 312         obd->obd_last_committed = mdt->mdt_lut.lut_last_transno;
 313         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 314
 315         obd->u.obt.obt_mount_count = mount_count + 1;
 316         obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
 317         lsd->lsd_mount_count = obd->u.obt.obt_mount_count;
 318
 319         /* save it, so mount count and last_transno is current */
 320         rc = lut_server_data_update(env, &mdt->mdt_lut, 0);
 321         if (rc)
 322                 GOTO(err_client, rc);
 323
 324         RETURN(0);
 325
 326 err_client:
 327         class_disconnect_exports(obd);
 328 out:
 329         return rc;
 330 }
 331
 332 /*
 333  * last_rcvd & last_committed update callbacks
 334  */
 335 static int mdt_last_rcvd_update(struct mdt_thread_info *mti,
 336                                 struct thandle *th)
 337 {
 338         struct mdt_device *mdt = mti->mti_mdt;
 339         struct ptlrpc_request *req = mdt_info_req(mti);
 340         struct tg_export_data *ted;
 341         struct lsd_client_data *lcd;
 342         loff_t off;
 343         int err;
 344         __s32 rc = th->th_result;
 345
 346         ENTRY;
 347         LASSERT(req);
 348         LASSERT(req->rq_export);
 349         LASSERT(mdt);
 350         ted = &req->rq_export->exp_target_data;
 351         LASSERT(ted);
 352
 353         cfs_mutex_lock(&ted->ted_lcd_lock);
 354         lcd = ted->ted_lcd;
 355         /* if the export has already been disconnected, we have no last_rcvd slot,
 356          * update server data with latest transno then */
 357         if (lcd == NULL) {
 358                 cfs_mutex_unlock(&ted->ted_lcd_lock);
 359                 CWARN("commit transaction for disconnected client %s: rc %d\n",
 360                       req->rq_export->exp_client_uuid.uuid, rc);
 361                 err = lut_server_data_write(mti->mti_env, &mdt->mdt_lut, th);
 362                 RETURN(err);
 363         }
 364
 365         off = ted->ted_lr_off;
 366         LASSERT(ergo(mti->mti_transno == 0, rc != 0));
 367         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 368             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 369                 if (mti->mti_transno != 0) {
 370                         if (lcd->lcd_last_close_transno > mti->mti_transno) {
 371                                 CERROR("Trying to overwrite bigger transno:"
 372                                        "on-disk: "LPU64", new: "LPU64" "
 373                                        "replay: %d. see LU-617.\n",
 374                                        lcd->lcd_last_close_transno,
 375                                        mti->mti_transno, req_is_replay(req));
 376                                 if (req_is_replay(req)) {
 377                                         cfs_spin_lock(&req->rq_export->exp_lock);
 378                                         req->rq_export->exp_vbr_failed = 1;
 379                                         cfs_spin_unlock(&req->rq_export->exp_lock);
 380                                 }
 381                                 cfs_mutex_unlock(&ted->ted_lcd_lock);
 382                                 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
 383                         }
 384                         lcd->lcd_last_close_transno = mti->mti_transno;
 385                 }
 386                 lcd->lcd_last_close_xid = req->rq_xid;
 387                 lcd->lcd_last_close_result = rc;
 388         } else {
 389                 /* VBR: save versions in last_rcvd for reconstruct. */
 390                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 391                 if (pre_versions) {
 392                         lcd->lcd_pre_versions[0] = pre_versions[0];
 393                         lcd->lcd_pre_versions[1] = pre_versions[1];
 394                         lcd->lcd_pre_versions[2] = pre_versions[2];
 395                         lcd->lcd_pre_versions[3] = pre_versions[3];
 396                 }
 397                 if (mti->mti_transno != 0) {
 398                         if (lcd->lcd_last_transno > mti->mti_transno) {
 399                                 CERROR("Trying to overwrite bigger transno:"
 400                                        "on-disk: "LPU64", new: "LPU64" "
 401                                        "replay: %d. see LU-617.\n",
 402                                        lcd->lcd_last_transno,
 403                                        mti->mti_transno, req_is_replay(req));
 404                                 if (req_is_replay(req)) {
 405                                         cfs_spin_lock(&req->rq_export->exp_lock);
 406                                         req->rq_export->exp_vbr_failed = 1;
 407                                         cfs_spin_unlock(&req->rq_export->exp_lock);
 408                                 }
 409                                 cfs_mutex_unlock(&ted->ted_lcd_lock);
 410                                 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
 411                         }
 412                         lcd->lcd_last_transno = mti->mti_transno;
 413                 }
 414                 lcd->lcd_last_xid = req->rq_xid;
 415                 lcd->lcd_last_result = rc;
 416                 /*XXX: save intent_disposition in mdt_thread_info?
 417                  * also there is bug - intent_dispostion is __u64,
 418                  * see struct ldlm_reply->lock_policy_res1; */
 419                 lcd->lcd_last_data = mti->mti_opdata;
 420         }
 421
 422         if ((mti->mti_exp->exp_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0) {
 423                 /* Although lightweight (LW) connections have no slot in
 424                  * last_rcvd, we still want to maintain the in-memory
 425                  * lsd_client_data structure in order to properly handle reply
 426                  * reconstruction. */
 427                 struct lu_target        *tg = &mdt->mdt_lut;
 428                 bool                     update = false;
 429
 430                 cfs_mutex_unlock(&ted->ted_lcd_lock);
 431                 err = 0;
 432
 433                 /* All operations performed by LW clients are synchronous and
 434                  * we store the committed transno in the last_rcvd header */
 435                 cfs_spin_lock(&tg->lut_translock);
 436                 if (mti->mti_transno > tg->lut_lsd.lsd_last_transno) {
 437                         tg->lut_lsd.lsd_last_transno = mti->mti_transno;
 438                         update = true;
 439                 }
 440                 cfs_spin_unlock(&tg->lut_translock);
 441
 442                 if (update)
 443                         err = lut_server_data_write(mti->mti_env, tg, th);
 444         } else if (off <= 0) {
 445                 CERROR("%s: client idx %d has offset %lld\n",
 446                        mdt2obd_dev(mdt)->obd_name, ted->ted_lr_idx, off);
 447                 cfs_mutex_unlock(&ted->ted_lcd_lock);
 448                 err = -EINVAL;
 449         } else {
 450                 err = lut_client_data_write(mti->mti_env, &mdt->mdt_lut, lcd,
 451                                             &off, th);
 452                 cfs_mutex_unlock(&ted->ted_lcd_lock);
 453         }
 454         RETURN(err);
 455 }
 456
 457 extern struct lu_context_key mdt_thread_key;
 458
 459 /* add credits for last_rcvd update */
 460 static int mdt_txn_start_cb(const struct lu_env *env,
 461                             struct thandle *th, void *cookie)
 462 {
 463         struct mdt_device *mdt = cookie;
 464         struct mdt_thread_info *mti;
 465         int rc;
 466         ENTRY;
 467
 468         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 469
 470         LASSERT(mdt->mdt_lut.lut_last_rcvd);
 471         if (mti->mti_exp == NULL)
 472                 RETURN(0);
 473
 474         rc = dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd,
 475                                      sizeof(struct lsd_client_data),
 476                                      mti->mti_exp->exp_target_data.ted_lr_off,
 477                                      th);
 478         if (rc)
 479                 return rc;
 480
 481         rc = dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd,
 482                                      sizeof(struct lr_server_data), 0, th);
 483         if (rc)
 484                 return rc;
 485
 486         if (mti->mti_mos != NULL)
 487                 rc = dt_declare_version_set(env, mdt_obj2dt(mti->mti_mos), th);
 488
 489         return rc;
 490 }
 491
 492 /* Update last_rcvd records with latests transaction data */
 493 static int mdt_txn_stop_cb(const struct lu_env *env,
 494                            struct thandle *txn, void *cookie)
 495 {
 496         struct mdt_device *mdt = cookie;
 497         struct mdt_thread_info *mti;
 498         struct ptlrpc_request *req;
 499
 500         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 501         req = mdt_info_req(mti);
 502
 503         if (mti->mti_mdt == NULL || req == NULL)
 504                 return 0;
 505
 506         if (mti->mti_has_trans) {
 507                 /* XXX: currently there are allowed cases, but the wrong cases
 508                  * are also possible, so better check is needed here */
 509                 CDEBUG(D_INFO, "More than one transaction "LPU64"\n",
 510                        mti->mti_transno);
 511                 return 0;
 512         }
 513
 514         mti->mti_has_trans = 1;
 515         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 516         if (txn->th_result != 0) {
 517                 if (mti->mti_transno != 0) {
 518                         CERROR("Replay transno "LPU64" failed: rc %d\n",
 519                                mti->mti_transno, txn->th_result);
 520                 }
 521         } else if (mti->mti_transno == 0) {
 522                 mti->mti_transno = ++ mdt->mdt_lut.lut_last_transno;
 523         } else {
 524                 /* should be replay */
 525                 if (mti->mti_transno > mdt->mdt_lut.lut_last_transno)
 526                         mdt->mdt_lut.lut_last_transno = mti->mti_transno;
 527         }
 528         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 529         /* sometimes the reply message has not been successfully packed */
 530         LASSERT(req != NULL && req->rq_repmsg != NULL);
 531
 532         /** VBR: set new versions */
 533         if (txn->th_result == 0 && mti->mti_mos != NULL) {
 534                 dt_version_set(env, mdt_obj2dt(mti->mti_mos),
 535                                mti->mti_transno, txn);
 536                 mti->mti_mos = NULL;
 537         }
 538
 539         /* filling reply data */
 540         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 541                mti->mti_transno, req->rq_export->exp_obd->obd_last_committed);
 542
 543         req->rq_transno = mti->mti_transno;
 544         lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno);
 545         /* if can't add callback, do sync write */
 546         txn->th_sync |= !!lut_last_commit_cb_add(txn, &mdt->mdt_lut,
 547                                                  mti->mti_exp,
 548                                                  mti->mti_transno);
 549         return mdt_last_rcvd_update(mti, txn);
 550 }
 551
 552 int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt,
 553                  struct obd_device *obd,
 554                  struct lustre_sb_info *lsi)
 555 {
 556         int rc = 0;
 557         ENTRY;
 558
 559         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
 560                 RETURN(-ENOENT);
 561
 562         /* prepare transactions callbacks */
 563         mdt->mdt_txn_cb.dtc_txn_start = mdt_txn_start_cb;
 564         mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb;
 565         mdt->mdt_txn_cb.dtc_txn_commit = NULL;
 566         mdt->mdt_txn_cb.dtc_cookie = mdt;
 567         mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD;
 568         CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage);
 569
 570         dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb);
 571
 572         rc = mdt_server_data_init(env, mdt, lsi);
 573
 574         RETURN(rc);
 575 }
 576
 577 void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt)
 578 {
 579         ENTRY;
 580
 581         /* Remove transaction callback */
 582         dt_txn_callback_del(mdt->mdt_bottom, &mdt->mdt_txn_cb);
 583         if (mdt->mdt_ck_obj)
 584                 lu_object_put(env, &mdt->mdt_ck_obj->do_lu);
 585         mdt->mdt_ck_obj = NULL;
 586         EXIT;
 587 }
 588
 589 /* reconstruction code */
 590 static void mdt_steal_ack_locks(struct ptlrpc_request *req)
 591 {
 592         struct ptlrpc_service_part *svcpt;
 593         struct obd_export         *exp = req->rq_export;
 594         cfs_list_t                *tmp;
 595         struct ptlrpc_reply_state *oldrep;
 596         int                        i;
 597
 598         /* CAVEAT EMPTOR: spinlock order */
 599         cfs_spin_lock(&exp->exp_lock);
 600         cfs_list_for_each (tmp, &exp->exp_outstanding_replies) {
 601                 oldrep = cfs_list_entry(tmp, struct ptlrpc_reply_state,
 602                                         rs_exp_list);
 603
 604                 if (oldrep->rs_xid != req->rq_xid)
 605                         continue;
 606
 607                 if (oldrep->rs_opc != lustre_msg_get_opc(req->rq_reqmsg))
 608                         CERROR ("Resent req xid "LPU64" has mismatched opc: "
 609                                 "new %d old %d\n", req->rq_xid,
 610                                 lustre_msg_get_opc(req->rq_reqmsg),
 611                                 oldrep->rs_opc);
 612
 613                 svcpt = oldrep->rs_svcpt;
 614                 cfs_spin_lock(&svcpt->scp_rep_lock);
 615
 616                 cfs_list_del_init (&oldrep->rs_exp_list);
 617
 618                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 619                       " o%d NID %s\n",
 620                       oldrep->rs_nlocks, oldrep,
 621                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_opc,
 622                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
 623
 624                 for (i = 0; i < oldrep->rs_nlocks; i++)
 625                         ptlrpc_save_lock(req, &oldrep->rs_locks[i],
 626                                          oldrep->rs_modes[i], 0);
 627                 oldrep->rs_nlocks = 0;
 628
 629                 DEBUG_REQ(D_HA, req, "stole locks for");
 630                 cfs_spin_lock(&oldrep->rs_lock);
 631                 ptlrpc_schedule_difficult_reply (oldrep);
 632                 cfs_spin_unlock(&oldrep->rs_lock);
 633
 634                 cfs_spin_unlock(&svcpt->scp_rep_lock);
 635                 break;
 636         }
 637         cfs_spin_unlock(&exp->exp_lock);
 638 }
 639
 640 /**
 641  * VBR: restore versions
 642  */
 643 void mdt_vbr_reconstruct(struct ptlrpc_request *req,
 644                          struct lsd_client_data *lcd)
 645 {
 646         __u64 pre_versions[4] = {0};
 647         pre_versions[0] = lcd->lcd_pre_versions[0];
 648         pre_versions[1] = lcd->lcd_pre_versions[1];
 649         pre_versions[2] = lcd->lcd_pre_versions[2];
 650         pre_versions[3] = lcd->lcd_pre_versions[3];
 651         lustre_msg_set_versions(req->rq_repmsg, pre_versions);
 652 }
 653
 654 void mdt_req_from_lcd(struct ptlrpc_request *req,
 655                       struct lsd_client_data *lcd)
 656 {
 657         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 658                   lcd->lcd_last_transno, lcd->lcd_last_result);
 659
 660         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 661             lustre_msg_get_opc(req->rq_repmsg) == MDS_DONE_WRITING) {
 662                 req->rq_transno = lcd->lcd_last_close_transno;
 663                 req->rq_status = lcd->lcd_last_close_result;
 664         } else {
 665                 req->rq_transno = lcd->lcd_last_transno;
 666                 req->rq_status = lcd->lcd_last_result;
 667                 mdt_vbr_reconstruct(req, lcd);
 668         }
 669         if (req->rq_status != 0)
 670                 req->rq_transno = 0;
 671         lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
 672         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 673         DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
 674                   req->rq_transno, req->rq_status);
 675
 676         mdt_steal_ack_locks(req);
 677 }
 678
 679 void mdt_reconstruct_generic(struct mdt_thread_info *mti,
 680                              struct mdt_lock_handle *lhc)
 681 {
 682         struct ptlrpc_request *req = mdt_info_req(mti);
 683         struct tg_export_data *ted = &req->rq_export->exp_target_data;
 684
 685         return mdt_req_from_lcd(req, ted->ted_lcd);
 686 }
 687
 688 static void mdt_reconstruct_create(struct mdt_thread_info *mti,
 689                                    struct mdt_lock_handle *lhc)
 690 {
 691         struct ptlrpc_request  *req = mdt_info_req(mti);
 692         struct obd_export *exp = req->rq_export;
 693         struct tg_export_data *ted = &exp->exp_target_data;
 694         struct mdt_device *mdt = mti->mti_mdt;
 695         struct mdt_object *child;
 696         struct mdt_body *body;
 697         int rc;
 698
 699         mdt_req_from_lcd(req, ted->ted_lcd);
 700         if (req->rq_status)
 701                 return;
 702
 703         /* if no error, so child was created with requested fid */
 704         child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2);
 705         if (IS_ERR(child)) {
 706                 rc = PTR_ERR(child);
 707                 LCONSOLE_WARN("Child "DFID" lookup error %d."
 708                               " Evicting client %s with export %s.\n",
 709                               PFID(mdt_object_fid(child)), rc,
 710                               obd_uuid2str(&exp->exp_client_uuid),
 711                               obd_export_nid2str(exp));
 712                 mdt_export_evict(exp);
 713                 EXIT;
 714                 return;
 715         }
 716
 717         body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
 718         mti->mti_attr.ma_need = MA_INODE;
 719         mti->mti_attr.ma_valid = 0;
 720         rc = mdt_attr_get_complex(mti, child, &mti->mti_attr);
 721         if (rc == -EREMOTE) {
 722                 /* object was created on remote server */
 723                 req->rq_status = rc;
 724                 body->valid |= OBD_MD_MDS;
 725         }
 726         mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
 727                            mdt_object_fid(child));
 728         mdt_object_put(mti->mti_env, child);
 729 }
 730
 731 static void mdt_reconstruct_setattr(struct mdt_thread_info *mti,
 732                                     struct mdt_lock_handle *lhc)
 733 {
 734         struct ptlrpc_request  *req = mdt_info_req(mti);
 735         struct obd_export *exp = req->rq_export;
 736         struct mdt_export_data *med = &exp->exp_mdt_data;
 737         struct mdt_device *mdt = mti->mti_mdt;
 738         struct mdt_object *obj;
 739         struct mdt_body *body;
 740
 741         mdt_req_from_lcd(req, med->med_ted.ted_lcd);
 742         if (req->rq_status)
 743                 return;
 744
 745         body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
 746         obj = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid1);
 747         if (IS_ERR(obj)) {
 748                 int rc = PTR_ERR(obj);
 749                 LCONSOLE_WARN(""DFID" lookup error %d."
 750                               " Evicting client %s with export %s.\n",
 751                               PFID(mdt_object_fid(obj)), rc,
 752                               obd_uuid2str(&exp->exp_client_uuid),
 753                               obd_export_nid2str(exp));
 754                 mdt_export_evict(exp);
 755                 EXIT;
 756                 return;
 757         }
 758         mti->mti_attr.ma_need = MA_INODE;
 759         mti->mti_attr.ma_valid = 0;
 760         mdt_attr_get_complex(mti, obj, &mti->mti_attr);
 761         mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
 762                            mdt_object_fid(obj));
 763         if (mti->mti_ioepoch && (mti->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
 764                 struct mdt_file_data *mfd;
 765                 struct mdt_body *repbody;
 766
 767                 repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
 768                 repbody->ioepoch = obj->mot_ioepoch;
 769                 cfs_spin_lock(&med->med_open_lock);
 770                 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
 771                         if (mfd->mfd_xid == req->rq_xid)
 772                                 break;
 773                 }
 774                 LASSERT(&mfd->mfd_list != &med->med_open_head);
 775                 cfs_spin_unlock(&med->med_open_lock);
 776                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 777         }
 778
 779         mdt_object_put(mti->mti_env, obj);
 780 }
 781
 782 typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,
 783                                   struct mdt_lock_handle *lhc);
 784
 785 static mdt_reconstructor reconstructors[REINT_MAX] = {
 786         [REINT_SETATTR]  = mdt_reconstruct_setattr,
 787         [REINT_CREATE]   = mdt_reconstruct_create,
 788         [REINT_LINK]     = mdt_reconstruct_generic,
 789         [REINT_UNLINK]   = mdt_reconstruct_generic,
 790         [REINT_RENAME]   = mdt_reconstruct_generic,
 791         [REINT_OPEN]     = mdt_reconstruct_open,
 792         [REINT_SETXATTR] = mdt_reconstruct_generic
 793 };
 794
 795 void mdt_reconstruct(struct mdt_thread_info *mti,
 796                      struct mdt_lock_handle *lhc)
 797 {
 798         ENTRY;
 799         reconstructors[mti->mti_rr.rr_opcode](mti, lhc);
 800         EXIT;
 801 }