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, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/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 = tgt_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 = tgt_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                 spin_lock(&exp->exp_lock);
 155                 exp->exp_connecting = 0;
 156                 exp->exp_in_recovery = 0;
 157                 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                 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                 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         __u32                   index;
 187         __u64 mount_count;
 188         int rc;
 189         ENTRY;
 190
 191         /* ensure padding in the struct is the correct size */
 192         CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
 193                 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
 194         CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
 195                 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
 196
 197         rc = server_name2index(mdt_obd_name(mdt), &index, NULL);
 198         if (rc < 0) {
 199                 CERROR("%s: Can not get index from obd_name: rc = %d\n",
 200                        mdt_obd_name(mdt), rc);
 201                 RETURN(rc);
 202         }
 203
 204         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 205         LASSERT(mti != NULL);
 206         la = &mti->mti_attr.ma_attr;
 207
 208         obj = mdt->mdt_lut.lut_last_rcvd;
 209         rc = dt_attr_get(env, obj, la, BYPASS_CAPA);
 210         if (rc)
 211                 RETURN(rc);
 212
 213         last_rcvd_size = (unsigned long)la->la_size;
 214
 215         if (last_rcvd_size == 0) {
 216                 LCONSOLE_WARN("%s: new disk, initializing\n",
 217                               mdt_obd_name(mdt));
 218
 219                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,
 220                        sizeof(lsd->lsd_uuid));
 221                 lsd->lsd_last_transno = 0;
 222                 lsd->lsd_mount_count = 0;
 223                 lsd->lsd_server_size = LR_SERVER_SIZE;
 224                 lsd->lsd_client_start = LR_CLIENT_START;
 225                 lsd->lsd_client_size = LR_CLIENT_SIZE;
 226                 lsd->lsd_feature_compat = OBD_COMPAT_MDT;
 227                 lsd->lsd_feature_rocompat = OBD_ROCOMPAT_LOVOBJID;
 228                 lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT |
 229                                             OBD_INCOMPAT_COMMON_LR |
 230                                             OBD_INCOMPAT_MULTI_OI;
 231                 lsd->lsd_osd_index = index;
 232         } else {
 233                 LCONSOLE_WARN("%s: used disk, loading\n", mdt_obd_name(mdt));
 234                 rc = tgt_server_data_read(env, &mdt->mdt_lut);
 235                 if (rc) {
 236                         CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
 237                         GOTO(out, rc);
 238                 }
 239                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
 240                         LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s using"
 241                                            "the wrong disk %s. Were the /dev/ "
 242                                            "assignments rearranged?\n",
 243                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
 244                         GOTO(out, rc = -EINVAL);
 245                 }
 246                 lsd->lsd_feature_compat |= OBD_COMPAT_MDT;
 247                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_MDT |
 248                                                 OBD_INCOMPAT_COMMON_LR;
 249                 if (lsd->lsd_osd_index != index) {
 250                         LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd is"
 251                                            "different with the index %d in"
 252                                            "config log, It might be disk"
 253                                            "corruption!\n", mdt_obd_name(mdt),
 254                                            lsd->lsd_osd_index, index);
 255                         GOTO(out, rc = -EINVAL);
 256                 }
 257         }
 258         mount_count = lsd->lsd_mount_count;
 259
 260         if (lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP) {
 261                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
 262                        mdt_obd_name(mdt),
 263                        lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP);
 264                 GOTO(out, rc = -EINVAL);
 265         }
 266         if (lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP) {
 267                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
 268                        mdt_obd_name(mdt),
 269                        lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP);
 270                 /* XXX: Do something like remount filesystem read-only */
 271                 GOTO(out, rc = -EINVAL);
 272         }
 273         /** Interop: evict all clients at first boot with 1.8 last_rcvd */
 274         if (!(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
 275                 if (last_rcvd_size > lsd->lsd_client_start) {
 276                         LCONSOLE_WARN("Mounting %s at first time on 1.8 FS, "
 277                                       "remove all clients for interop needs\n",
 278                                       mdt_obd_name(mdt));
 279                         rc = tgt_truncate_last_rcvd(env, &mdt->mdt_lut,
 280                                                     lsd->lsd_client_start);
 281                         if (rc)
 282                                 GOTO(out, rc);
 283                         last_rcvd_size = lsd->lsd_client_start;
 284                 }
 285                 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
 286                 lsd->lsd_feature_compat |= OBD_COMPAT_20;
 287         }
 288
 289         lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
 290
 291         spin_lock(&mdt->mdt_lut.lut_translock);
 292         mdt->mdt_lut.lut_last_transno = lsd->lsd_last_transno;
 293         spin_unlock(&mdt->mdt_lut.lut_translock);
 294
 295         CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
 296         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 297                mdt_obd_name(mdt), mdt->mdt_lut.lut_last_transno);
 298         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 299                mdt_obd_name(mdt), mount_count + 1);
 300         CDEBUG(D_INODE, "%s: server data size: %u\n",
 301                mdt_obd_name(mdt), lsd->lsd_server_size);
 302         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 303                mdt_obd_name(mdt), lsd->lsd_client_start);
 304         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 305                mdt_obd_name(mdt), lsd->lsd_client_size);
 306         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 307                mdt_obd_name(mdt), last_rcvd_size);
 308         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", mdt_obd_name(mdt),
 309                last_rcvd_size <= lsd->lsd_client_start ? 0 :
 310                (last_rcvd_size - lsd->lsd_client_start) /
 311                 lsd->lsd_client_size);
 312         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
 313
 314         if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
 315             !lsd->lsd_client_size) {
 316                 CERROR("Bad last_rcvd contents!\n");
 317                 GOTO(out, rc = -EINVAL);
 318         }
 319
 320         rc = mdt_clients_data_init(env, mdt, last_rcvd_size);
 321         if (rc)
 322                 GOTO(err_client, rc);
 323
 324         spin_lock(&mdt->mdt_lut.lut_translock);
 325         /* obd_last_committed is used for compatibility
 326          * with other lustre recovery code */
 327         obd->obd_last_committed = mdt->mdt_lut.lut_last_transno;
 328         spin_unlock(&mdt->mdt_lut.lut_translock);
 329
 330         obd->u.obt.obt_mount_count = mount_count + 1;
 331         obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
 332         lsd->lsd_mount_count = obd->u.obt.obt_mount_count;
 333
 334         /* save it, so mount count and last_transno is current */
 335         rc = tgt_server_data_update(env, &mdt->mdt_lut, 0);
 336         if (rc)
 337                 GOTO(err_client, rc);
 338
 339         RETURN(0);
 340
 341 err_client:
 342         class_disconnect_exports(obd);
 343 out:
 344         return rc;
 345 }
 346
 347 /*
 348  * last_rcvd & last_committed update callbacks
 349  */
 350 static int mdt_last_rcvd_update(struct mdt_thread_info *mti,
 351                                 struct thandle *th)
 352 {
 353         struct mdt_device *mdt = mti->mti_mdt;
 354         struct ptlrpc_request *req = mdt_info_req(mti);
 355         struct tg_export_data *ted;
 356         struct lsd_client_data *lcd;
 357         loff_t off;
 358         int err;
 359         __s32 rc = th->th_result;
 360
 361         ENTRY;
 362         LASSERT(req);
 363         LASSERT(req->rq_export);
 364         LASSERT(mdt);
 365         ted = &req->rq_export->exp_target_data;
 366         LASSERT(ted);
 367
 368         mutex_lock(&ted->ted_lcd_lock);
 369         lcd = ted->ted_lcd;
 370         /* if the export has already been disconnected, we have no last_rcvd
 371          * slot, update server data with latest transno then */
 372         if (lcd == NULL) {
 373                 mutex_unlock(&ted->ted_lcd_lock);
 374                 CWARN("commit transaction for disconnected client %s: rc %d\n",
 375                       req->rq_export->exp_client_uuid.uuid, rc);
 376                 err = tgt_server_data_write(mti->mti_env, &mdt->mdt_lut, th);
 377                 RETURN(err);
 378         }
 379
 380         off = ted->ted_lr_off;
 381         LASSERT(ergo(mti->mti_transno == 0, rc != 0));
 382         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 383             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 384                 if (mti->mti_transno != 0) {
 385                         if (lcd->lcd_last_close_transno > mti->mti_transno) {
 386                                 CERROR("Trying to overwrite bigger transno:"
 387                                        "on-disk: "LPU64", new: "LPU64" "
 388                                        "replay: %d. see LU-617.\n",
 389                                        lcd->lcd_last_close_transno,
 390                                        mti->mti_transno, req_is_replay(req));
 391                                 if (req_is_replay(req)) {
 392                                         spin_lock(&req->rq_export->exp_lock);
 393                                         req->rq_export->exp_vbr_failed = 1;
 394                                         spin_unlock(&req->rq_export->exp_lock);
 395                                 }
 396                                 mutex_unlock(&ted->ted_lcd_lock);
 397                                 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
 398                         }
 399                         lcd->lcd_last_close_transno = mti->mti_transno;
 400                 }
 401                 lcd->lcd_last_close_xid = req->rq_xid;
 402                 lcd->lcd_last_close_result = rc;
 403         } else {
 404                 /* VBR: save versions in last_rcvd for reconstruct. */
 405                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 406                 if (pre_versions) {
 407                         lcd->lcd_pre_versions[0] = pre_versions[0];
 408                         lcd->lcd_pre_versions[1] = pre_versions[1];
 409                         lcd->lcd_pre_versions[2] = pre_versions[2];
 410                         lcd->lcd_pre_versions[3] = pre_versions[3];
 411                 }
 412                 if (mti->mti_transno != 0) {
 413                         if (lcd->lcd_last_transno > mti->mti_transno) {
 414                                 CERROR("Trying to overwrite bigger transno:"
 415                                        "on-disk: "LPU64", new: "LPU64" "
 416                                        "replay: %d. see LU-617.\n",
 417                                        lcd->lcd_last_transno,
 418                                        mti->mti_transno, req_is_replay(req));
 419                                 if (req_is_replay(req)) {
 420                                         spin_lock(&req->rq_export->exp_lock);
 421                                         req->rq_export->exp_vbr_failed = 1;
 422                                         spin_unlock(&req->rq_export->exp_lock);
 423                                 }
 424                                 mutex_unlock(&ted->ted_lcd_lock);
 425                                 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
 426                         }
 427                         lcd->lcd_last_transno = mti->mti_transno;
 428                 }
 429                 lcd->lcd_last_xid = req->rq_xid;
 430                 lcd->lcd_last_result = rc;
 431                 /*XXX: save intent_disposition in mdt_thread_info?
 432                  * also there is bug - intent_dispostion is __u64,
 433                  * see struct ldlm_reply->lock_policy_res1; */
 434                 lcd->lcd_last_data = mti->mti_opdata;
 435         }
 436
 437         if (exp_connect_flags(mti->mti_exp) & OBD_CONNECT_LIGHTWEIGHT) {
 438                 /* Although lightweight (LW) connections have no slot in
 439                  * last_rcvd, we still want to maintain the in-memory
 440                  * lsd_client_data structure in order to properly handle reply
 441                  * reconstruction. */
 442                 struct lu_target        *tg = &mdt->mdt_lut;
 443                 bool                     update = false;
 444
 445                 mutex_unlock(&ted->ted_lcd_lock);
 446                 err = 0;
 447
 448                 /* All operations performed by LW clients are synchronous and
 449                  * we store the committed transno in the last_rcvd header */
 450                 spin_lock(&tg->lut_translock);
 451                 if (mti->mti_transno > tg->lut_lsd.lsd_last_transno) {
 452                         tg->lut_lsd.lsd_last_transno = mti->mti_transno;
 453                         update = true;
 454                 }
 455                 spin_unlock(&tg->lut_translock);
 456
 457                 if (update)
 458                         err = tgt_server_data_write(mti->mti_env, tg, th);
 459         } else if (off <= 0) {
 460                 CERROR("%s: client idx %d has offset %lld\n",
 461                        mdt_obd_name(mdt), ted->ted_lr_idx, off);
 462                 mutex_unlock(&ted->ted_lcd_lock);
 463                 err = -EINVAL;
 464         } else {
 465                 err = tgt_client_data_write(mti->mti_env, &mdt->mdt_lut, lcd,
 466                                             &off, th);
 467                 mutex_unlock(&ted->ted_lcd_lock);
 468         }
 469         RETURN(err);
 470 }
 471
 472 extern struct lu_context_key mdt_thread_key;
 473
 474 /* add credits for last_rcvd update */
 475 static int mdt_txn_start_cb(const struct lu_env *env,
 476                             struct thandle *th, void *cookie)
 477 {
 478         struct mdt_device *mdt = cookie;
 479         struct mdt_thread_info *mti;
 480         int rc;
 481         ENTRY;
 482
 483         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 484
 485         LASSERT(mdt->mdt_lut.lut_last_rcvd);
 486         if (mti->mti_exp == NULL)
 487                 RETURN(0);
 488
 489         rc = dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd,
 490                                      sizeof(struct lsd_client_data),
 491                                      mti->mti_exp->exp_target_data.ted_lr_off,
 492                                      th);
 493         if (rc)
 494                 return rc;
 495
 496         rc = dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd,
 497                                      sizeof(struct lr_server_data), 0, th);
 498         if (rc)
 499                 return rc;
 500
 501         /* we probably should not set local transno to the remote object
 502          * on another storage, What about VBR on remote object? XXX */
 503         if (mti->mti_mos != NULL && !mdt_object_remote(mti->mti_mos))
 504                 rc = dt_declare_version_set(env, mdt_obj2dt(mti->mti_mos), th);
 505
 506         return rc;
 507 }
 508
 509 /* Update last_rcvd records with latests transaction data */
 510 static int mdt_txn_stop_cb(const struct lu_env *env,
 511                            struct thandle *txn, void *cookie)
 512 {
 513         struct mdt_device *mdt = cookie;
 514         struct mdt_thread_info *mti;
 515         struct ptlrpc_request *req;
 516
 517         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 518         req = mdt_info_req(mti);
 519
 520         if (mti->mti_mdt == NULL || req == NULL)
 521                 return 0;
 522
 523         if (mti->mti_has_trans) {
 524                 /* XXX: currently there are allowed cases, but the wrong cases
 525                  * are also possible, so better check is needed here */
 526                 CDEBUG(D_INFO, "More than one transaction "LPU64"\n",
 527                        mti->mti_transno);
 528                 return 0;
 529         }
 530
 531         mti->mti_has_trans = 1;
 532         spin_lock(&mdt->mdt_lut.lut_translock);
 533         if (txn->th_result != 0) {
 534                 if (mti->mti_transno != 0) {
 535                         CERROR("Replay transno "LPU64" failed: rc %d\n",
 536                                 mti->mti_transno, txn->th_result);
 537                         return 0;
 538                 }
 539         } else if (mti->mti_transno == 0) {
 540                 mti->mti_transno = ++ mdt->mdt_lut.lut_last_transno;
 541         } else {
 542                 /* should be replay */
 543                 if (mti->mti_transno > mdt->mdt_lut.lut_last_transno)
 544                         mdt->mdt_lut.lut_last_transno = mti->mti_transno;
 545         }
 546         spin_unlock(&mdt->mdt_lut.lut_translock);
 547         /* sometimes the reply message has not been successfully packed */
 548         LASSERT(req != NULL && req->rq_repmsg != NULL);
 549
 550         /** VBR: set new versions */
 551         /* we probably should not set local transno to the remote object
 552          * on another storage, What about VBR on remote object? XXX */
 553         if (txn->th_result == 0 && mti->mti_mos != NULL &&
 554             !mdt_object_remote(mti->mti_mos)) {
 555
 556                 dt_version_set(env, mdt_obj2dt(mti->mti_mos),
 557                                mti->mti_transno, txn);
 558                 mti->mti_mos = NULL;
 559         }
 560
 561         /* filling reply data */
 562         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 563                mti->mti_transno, req->rq_export->exp_obd->obd_last_committed);
 564
 565         req->rq_transno = mti->mti_transno;
 566         lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno);
 567         /* if can't add callback, do sync write */
 568         txn->th_sync |= !!tgt_last_commit_cb_add(txn, &mdt->mdt_lut,
 569                                                  mti->mti_exp,
 570                                                  mti->mti_transno);
 571         return mdt_last_rcvd_update(mti, txn);
 572 }
 573
 574 int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt,
 575                  struct obd_device *obd,
 576                  struct lustre_sb_info *lsi)
 577 {
 578         int rc = 0;
 579         ENTRY;
 580
 581         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
 582                 RETURN(-ENOENT);
 583
 584         /* prepare transactions callbacks */
 585         mdt->mdt_txn_cb.dtc_txn_start = mdt_txn_start_cb;
 586         mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb;
 587         mdt->mdt_txn_cb.dtc_txn_commit = NULL;
 588         mdt->mdt_txn_cb.dtc_cookie = mdt;
 589         mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD;
 590         CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage);
 591
 592         dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb);
 593
 594         rc = mdt_server_data_init(env, mdt, lsi);
 595
 596         RETURN(rc);
 597 }
 598
 599 void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt)
 600 {
 601         ENTRY;
 602
 603         /* Remove transaction callback */
 604         dt_txn_callback_del(mdt->mdt_bottom, &mdt->mdt_txn_cb);
 605         if (mdt->mdt_ck_obj)
 606                 lu_object_put(env, &mdt->mdt_ck_obj->do_lu);
 607         mdt->mdt_ck_obj = NULL;
 608         EXIT;
 609 }
 610
 611 /* reconstruction code */
 612 static void mdt_steal_ack_locks(struct ptlrpc_request *req)
 613 {
 614         struct ptlrpc_service_part *svcpt;
 615         struct obd_export         *exp = req->rq_export;
 616         cfs_list_t                *tmp;
 617         struct ptlrpc_reply_state *oldrep;
 618         int                        i;
 619
 620         /* CAVEAT EMPTOR: spinlock order */
 621         spin_lock(&exp->exp_lock);
 622         cfs_list_for_each (tmp, &exp->exp_outstanding_replies) {
 623                 oldrep = cfs_list_entry(tmp, struct ptlrpc_reply_state,
 624                                         rs_exp_list);
 625
 626                 if (oldrep->rs_xid != req->rq_xid)
 627                         continue;
 628
 629                 if (oldrep->rs_opc != lustre_msg_get_opc(req->rq_reqmsg))
 630                         CERROR ("Resent req xid "LPU64" has mismatched opc: "
 631                                 "new %d old %d\n", req->rq_xid,
 632                                 lustre_msg_get_opc(req->rq_reqmsg),
 633                                 oldrep->rs_opc);
 634
 635                 svcpt = oldrep->rs_svcpt;
 636                 spin_lock(&svcpt->scp_rep_lock);
 637
 638                 cfs_list_del_init (&oldrep->rs_exp_list);
 639
 640                 CDEBUG(D_HA, "Stealing %d locks from rs %p x"LPD64".t"LPD64
 641                        " o%d NID %s\n",
 642                        oldrep->rs_nlocks, oldrep,
 643                        oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_opc,
 644                        libcfs_nid2str(exp->exp_connection->c_peer.nid));
 645
 646                 for (i = 0; i < oldrep->rs_nlocks; i++)
 647                         ptlrpc_save_lock(req, &oldrep->rs_locks[i],
 648                                          oldrep->rs_modes[i], 0);
 649                 oldrep->rs_nlocks = 0;
 650
 651                 DEBUG_REQ(D_HA, req, "stole locks for");
 652                 spin_lock(&oldrep->rs_lock);
 653                 ptlrpc_schedule_difficult_reply(oldrep);
 654                 spin_unlock(&oldrep->rs_lock);
 655
 656                 spin_unlock(&svcpt->scp_rep_lock);
 657                 break;
 658         }
 659         spin_unlock(&exp->exp_lock);
 660 }
 661
 662 /**
 663  * VBR: restore versions
 664  */
 665 void mdt_vbr_reconstruct(struct ptlrpc_request *req,
 666                          struct lsd_client_data *lcd)
 667 {
 668         __u64 pre_versions[4] = {0};
 669         pre_versions[0] = lcd->lcd_pre_versions[0];
 670         pre_versions[1] = lcd->lcd_pre_versions[1];
 671         pre_versions[2] = lcd->lcd_pre_versions[2];
 672         pre_versions[3] = lcd->lcd_pre_versions[3];
 673         lustre_msg_set_versions(req->rq_repmsg, pre_versions);
 674 }
 675
 676 void mdt_req_from_lcd(struct ptlrpc_request *req,
 677                       struct lsd_client_data *lcd)
 678 {
 679         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 680                   lcd->lcd_last_transno, lcd->lcd_last_result);
 681
 682         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 683             lustre_msg_get_opc(req->rq_repmsg) == MDS_DONE_WRITING) {
 684                 req->rq_transno = lcd->lcd_last_close_transno;
 685                 req->rq_status = lcd->lcd_last_close_result;
 686         } else {
 687                 req->rq_transno = lcd->lcd_last_transno;
 688                 req->rq_status = lcd->lcd_last_result;
 689                 mdt_vbr_reconstruct(req, lcd);
 690         }
 691         if (req->rq_status != 0)
 692                 req->rq_transno = 0;
 693         lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
 694         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 695         DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
 696                   req->rq_transno, req->rq_status);
 697
 698         mdt_steal_ack_locks(req);
 699 }
 700
 701 void mdt_reconstruct_generic(struct mdt_thread_info *mti,
 702                              struct mdt_lock_handle *lhc)
 703 {
 704         struct ptlrpc_request *req = mdt_info_req(mti);
 705         struct tg_export_data *ted = &req->rq_export->exp_target_data;
 706
 707         return mdt_req_from_lcd(req, ted->ted_lcd);
 708 }
 709
 710 static void mdt_reconstruct_create(struct mdt_thread_info *mti,
 711                                    struct mdt_lock_handle *lhc)
 712 {
 713         struct ptlrpc_request  *req = mdt_info_req(mti);
 714         struct obd_export *exp = req->rq_export;
 715         struct tg_export_data *ted = &exp->exp_target_data;
 716         struct mdt_device *mdt = mti->mti_mdt;
 717         struct mdt_object *child;
 718         struct mdt_body *body;
 719         int rc;
 720
 721         mdt_req_from_lcd(req, ted->ted_lcd);
 722         if (req->rq_status)
 723                 return;
 724
 725         /* if no error, so child was created with requested fid */
 726         child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2);
 727         if (IS_ERR(child)) {
 728                 rc = PTR_ERR(child);
 729                 LCONSOLE_WARN("Child "DFID" lookup error %d."
 730                               " Evicting client %s with export %s.\n",
 731                               PFID(mdt_object_fid(child)), rc,
 732                               obd_uuid2str(&exp->exp_client_uuid),
 733                               obd_export_nid2str(exp));
 734                 mdt_export_evict(exp);
 735                 EXIT;
 736                 return;
 737         }
 738
 739         body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
 740         mti->mti_attr.ma_need = MA_INODE;
 741         mti->mti_attr.ma_valid = 0;
 742         rc = mdt_attr_get_complex(mti, child, &mti->mti_attr);
 743         if (rc == -EREMOTE) {
 744                 /* object was created on remote server */
 745                 if (!mdt_is_dne_client(exp))
 746                         /* Return -EIO for old client */
 747                         rc = -EIO;
 748
 749                 req->rq_status = rc;
 750                 body->valid |= OBD_MD_MDS;
 751         }
 752         mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
 753                            mdt_object_fid(child));
 754         mdt_object_put(mti->mti_env, child);
 755 }
 756
 757 static void mdt_reconstruct_setattr(struct mdt_thread_info *mti,
 758                                     struct mdt_lock_handle *lhc)
 759 {
 760         struct ptlrpc_request  *req = mdt_info_req(mti);
 761         struct obd_export *exp = req->rq_export;
 762         struct mdt_export_data *med = &exp->exp_mdt_data;
 763         struct mdt_device *mdt = mti->mti_mdt;
 764         struct mdt_object *obj;
 765         struct mdt_body *body;
 766
 767         mdt_req_from_lcd(req, med->med_ted.ted_lcd);
 768         if (req->rq_status)
 769                 return;
 770
 771         body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
 772         obj = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid1);
 773         if (IS_ERR(obj)) {
 774                 int rc = PTR_ERR(obj);
 775                 LCONSOLE_WARN(""DFID" lookup error %d."
 776                               " Evicting client %s with export %s.\n",
 777                               PFID(mdt_object_fid(obj)), rc,
 778                               obd_uuid2str(&exp->exp_client_uuid),
 779                               obd_export_nid2str(exp));
 780                 mdt_export_evict(exp);
 781                 EXIT;
 782                 return;
 783         }
 784         mti->mti_attr.ma_need = MA_INODE;
 785         mti->mti_attr.ma_valid = 0;
 786         mdt_attr_get_complex(mti, obj, &mti->mti_attr);
 787         mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
 788                            mdt_object_fid(obj));
 789         if (mti->mti_ioepoch && (mti->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
 790                 struct mdt_file_data *mfd;
 791                 struct mdt_body *repbody;
 792
 793                 repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
 794                 repbody->ioepoch = obj->mot_ioepoch;
 795                 spin_lock(&med->med_open_lock);
 796                 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
 797                         if (mfd->mfd_xid == req->rq_xid)
 798                                 break;
 799                 }
 800                 LASSERT(&mfd->mfd_list != &med->med_open_head);
 801                 spin_unlock(&med->med_open_lock);
 802                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 803         }
 804
 805         mdt_object_put(mti->mti_env, obj);
 806 }
 807
 808 typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,
 809                                   struct mdt_lock_handle *lhc);
 810
 811 static mdt_reconstructor reconstructors[REINT_MAX] = {
 812         [REINT_SETATTR]  = mdt_reconstruct_setattr,
 813         [REINT_CREATE]   = mdt_reconstruct_create,
 814         [REINT_LINK]     = mdt_reconstruct_generic,
 815         [REINT_UNLINK]   = mdt_reconstruct_generic,
 816         [REINT_RENAME]   = mdt_reconstruct_generic,
 817         [REINT_OPEN]     = mdt_reconstruct_open,
 818         [REINT_SETXATTR] = mdt_reconstruct_generic
 819 };
 820
 821 void mdt_reconstruct(struct mdt_thread_info *mti,
 822                      struct mdt_lock_handle *lhc)
 823 {
 824         ENTRY;
 825         reconstructors[mti->mti_rr.rr_opcode](mti, lhc);
 826         EXIT;
 827 }