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 #ifndef EXPORT_SYMTAB
  45 # define EXPORT_SYMTAB
  46 #endif
  47 #define DEBUG_SUBSYSTEM S_MDS
  48
  49 #include "mdt_internal.h"
  50
  51 struct lu_buf *mdt_buf(const struct lu_env *env, void *area, ssize_t len)
  52 {
  53         struct lu_buf *buf;
  54         struct mdt_thread_info *mti;
  55
  56         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
  57         buf = &mti->mti_buf;
  58         buf->lb_buf = area;
  59         buf->lb_len = len;
  60         return buf;
  61 }
  62
  63 const struct lu_buf *mdt_buf_const(const struct lu_env *env,
  64                                    const void *area, ssize_t len)
  65 {
  66         struct lu_buf *buf;
  67         struct mdt_thread_info *mti;
  68
  69         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
  70         buf = &mti->mti_buf;
  71
  72         buf->lb_buf = (void *)area;
  73         buf->lb_len = len;
  74         return buf;
  75 }
  76
  77 void mdt_trans_stop(const struct lu_env *env,
  78                     struct mdt_device *mdt, struct thandle *th)
  79 {
  80         dt_trans_stop(env, mdt->mdt_bottom, th);
  81 }
  82
  83 static int mdt_clients_data_init(const struct lu_env *env,
  84                                  struct mdt_device *mdt,
  85                                  unsigned long last_size)
  86 {
  87         struct lr_server_data  *lsd = &mdt->mdt_lut.lut_lsd;
  88         struct lsd_client_data *lcd;
  89         struct obd_device      *obd = mdt2obd_dev(mdt);
  90         loff_t off;
  91         int cl_idx;
  92         int rc = 0;
  93         ENTRY;
  94
  95         OBD_ALLOC_PTR(lcd);
  96         if (!lcd)
  97                 RETURN(-ENOMEM);
  98
  99         /* When we do a clean MDS shutdown, we save the last_transno into
 100          * the header.  If we find clients with higher last_transno values
 101          * then those clients may need recovery done. */
 102         LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
 103         for (cl_idx = 0, off = lsd->lsd_client_start;
 104              off < last_size; cl_idx++) {
 105                 __u64 last_transno;
 106                 struct obd_export *exp;
 107                 struct mdt_thread_info *mti;
 108
 109                 off = lsd->lsd_client_start +
 110                         cl_idx * lsd->lsd_client_size;
 111
 112                 rc = lut_client_data_read(env, &mdt->mdt_lut, lcd, &off, cl_idx);
 113                 if (rc) {
 114                         CERROR("error reading MDS %s idx %d, off %llu: rc %d\n",
 115                                LAST_RCVD, cl_idx, off, rc);
 116                         rc = 0;
 117                         break; /* read error shouldn't cause startup to fail */
 118                 }
 119
 120                 if (lcd->lcd_uuid[0] == '\0') {
 121                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 122                                cl_idx);
 123                         continue;
 124                 }
 125
 126                 last_transno = lcd_last_transno(lcd);
 127
 128                 /* These exports are cleaned up by mdt_obd_disconnect(), so
 129                  * they need to be set up like real exports as
 130                  * mdt_obd_connect() does.
 131                  */
 132                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 133                        " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
 134                        last_transno, lsd->lsd_last_transno,
 135                        lcd_last_xid(lcd));
 136
 137                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
 138                 if (IS_ERR(exp)) {
 139                         if (PTR_ERR(exp) == -EALREADY) {
 140                                 /* export already exists, zero out this one */
 141                                 CERROR("Duplicate export %s!\n", lcd->lcd_uuid);
 142                                 continue;
 143                         }
 144                         GOTO(err_client, rc = PTR_ERR(exp));
 145                 }
 146
 147                 mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 148                 LASSERT(mti != NULL);
 149                 mti->mti_exp = exp;
 150                 /* copy on-disk lcd to the export */
 151                 *exp->exp_target_data.ted_lcd = *lcd;
 152                 rc = lut_client_add(env, exp, cl_idx);
 153                 /* can't fail existing */
 154                 LASSERTF(rc == 0, "rc = %d\n", rc);
 155                 /* VBR: set export last committed version */
 156                 exp->exp_last_committed = last_transno;
 157                 cfs_spin_lock(&exp->exp_lock);
 158                 exp->exp_connecting = 0;
 159                 exp->exp_in_recovery = 0;
 160                 cfs_spin_unlock(&exp->exp_lock);
 161                 obd->obd_max_recoverable_clients++;
 162                 class_export_put(exp);
 163
 164                 CDEBUG(D_OTHER, "client at idx %d has last_transno="LPU64"\n",
 165                        cl_idx, last_transno);
 166                 /* protect __u64 value update */
 167                 cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 168                 mdt->mdt_lut.lut_last_transno = max(last_transno,
 169                                                 mdt->mdt_lut.lut_last_transno);
 170                 cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 171         }
 172
 173 err_client:
 174         OBD_FREE_PTR(lcd);
 175         RETURN(rc);
 176 }
 177
 178 static int mdt_server_data_init(const struct lu_env *env,
 179                                 struct mdt_device *mdt,
 180                                 struct lustre_sb_info *lsi)
 181 {
 182         struct lr_server_data  *lsd = &mdt->mdt_lut.lut_lsd;
 183         struct lsd_client_data *lcd = NULL;
 184         struct obd_device      *obd = mdt2obd_dev(mdt);
 185         struct mdt_thread_info *mti;
 186         struct dt_object       *obj;
 187         struct lu_attr         *la;
 188         unsigned long last_rcvd_size;
 189         __u64 mount_count;
 190         int rc;
 191         ENTRY;
 192
 193         /* ensure padding in the struct is the correct size */
 194         CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
 195                 sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
 196         CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
 197                 sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
 198
 199         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 200         LASSERT(mti != NULL);
 201         la = &mti->mti_attr.ma_attr;
 202
 203         obj = mdt->mdt_lut.lut_last_rcvd;
 204         rc = dt_attr_get(env, obj, la, BYPASS_CAPA);
 205         if (rc)
 206                 RETURN(rc);
 207
 208         last_rcvd_size = (unsigned long)la->la_size;
 209
 210         if (last_rcvd_size == 0) {
 211                 LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name);
 212
 213                 memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid,
 214                        sizeof(lsd->lsd_uuid));
 215                 lsd->lsd_last_transno = 0;
 216                 lsd->lsd_mount_count = 0;
 217                 lsd->lsd_server_size = LR_SERVER_SIZE;
 218                 lsd->lsd_client_start = LR_CLIENT_START;
 219                 lsd->lsd_client_size = LR_CLIENT_SIZE;
 220                 lsd->lsd_feature_compat = OBD_COMPAT_MDT;
 221                 lsd->lsd_feature_rocompat = OBD_ROCOMPAT_LOVOBJID;
 222                 lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT |
 223                                             OBD_INCOMPAT_COMMON_LR |
 224                                             OBD_INCOMPAT_MULTI_OI;
 225         } else {
 226                 LCONSOLE_WARN("%s: used disk, loading\n", obd->obd_name);
 227                 rc = lut_server_data_read(env, &mdt->mdt_lut);
 228                 if (rc) {
 229                         CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc);
 230                         GOTO(out, rc);
 231                 }
 232                 if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) {
 233                         LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s using"
 234                                            "the wrong disk %s. Were the /dev/ "
 235                                            "assignments rearranged?\n",
 236                                            obd->obd_uuid.uuid, lsd->lsd_uuid);
 237                         GOTO(out, rc = -EINVAL);
 238                 }
 239                 lsd->lsd_feature_compat |= OBD_COMPAT_MDT;
 240                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_MDT |
 241                                              OBD_INCOMPAT_COMMON_LR;
 242         }
 243         mount_count = lsd->lsd_mount_count;
 244
 245         if (lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP) {
 246                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
 247                        obd->obd_name,
 248                        lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP);
 249                 GOTO(out, rc = -EINVAL);
 250         }
 251         if (lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP) {
 252                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
 253                        obd->obd_name,
 254                        lsd->lsd_feature_rocompat & ~MDT_ROCOMPAT_SUPP);
 255                 /* XXX: Do something like remount filesystem read-only */
 256                 GOTO(out, rc = -EINVAL);
 257         }
 258         /** Interop: evict all clients at first boot with 1.8 last_rcvd */
 259         if (!(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
 260                 if (last_rcvd_size > lsd->lsd_client_start) {
 261                         LCONSOLE_WARN("Mounting %s at first time on 1.8 FS, "
 262                                       "remove all clients for interop needs\n",
 263                                       obd->obd_name);
 264                         rc = lut_truncate_last_rcvd(env, &mdt->mdt_lut,
 265                                                     lsd->lsd_client_start);
 266                         if (rc)
 267                                 GOTO(out, rc);
 268                         last_rcvd_size = lsd->lsd_client_start;
 269                 }
 270                 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
 271                 lsd->lsd_feature_compat |= OBD_COMPAT_20;
 272         }
 273
 274         if (lsi->lsi_ldd->ldd_flags & LDD_F_IAM_DIR)
 275                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_IAM_DIR;
 276
 277         lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
 278
 279         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 280         mdt->mdt_lut.lut_last_transno = lsd->lsd_last_transno;
 281         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 282
 283         CDEBUG(D_INODE, "========BEGIN DUMPING LAST_RCVD========\n");
 284         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
 285                obd->obd_name, mdt->mdt_lut.lut_last_transno);
 286         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
 287                obd->obd_name, mount_count + 1);
 288         CDEBUG(D_INODE, "%s: server data size: %u\n",
 289                obd->obd_name, lsd->lsd_server_size);
 290         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
 291                obd->obd_name, lsd->lsd_client_start);
 292         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
 293                obd->obd_name, lsd->lsd_client_size);
 294         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
 295                obd->obd_name, last_rcvd_size);
 296         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name,
 297                last_rcvd_size <= lsd->lsd_client_start ? 0 :
 298                (last_rcvd_size - lsd->lsd_client_start) /
 299                 lsd->lsd_client_size);
 300         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
 301
 302         if (!lsd->lsd_server_size || !lsd->lsd_client_start ||
 303             !lsd->lsd_client_size) {
 304                 CERROR("Bad last_rcvd contents!\n");
 305                 GOTO(out, rc = -EINVAL);
 306         }
 307
 308         rc = mdt_clients_data_init(env, mdt, last_rcvd_size);
 309         if (rc)
 310                 GOTO(err_client, rc);
 311
 312         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 313         /* obd_last_committed is used for compatibility
 314          * with other lustre recovery code */
 315         obd->obd_last_committed = mdt->mdt_lut.lut_last_transno;
 316         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 317
 318         obd->u.obt.obt_mount_count = mount_count + 1;
 319         obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count;
 320         lsd->lsd_mount_count = obd->u.obt.obt_mount_count;
 321
 322         /* save it, so mount count and last_transno is current */
 323         rc = lut_server_data_update(env, &mdt->mdt_lut, 0);
 324         if (rc)
 325                 GOTO(err_client, rc);
 326
 327         RETURN(0);
 328
 329 err_client:
 330         class_disconnect_exports(obd);
 331 out:
 332         return rc;
 333 }
 334
 335 /*
 336  * last_rcvd & last_committed update callbacks
 337  */
 338 static int mdt_last_rcvd_update(struct mdt_thread_info *mti,
 339                                 struct thandle *th)
 340 {
 341         struct mdt_device *mdt = mti->mti_mdt;
 342         struct ptlrpc_request *req = mdt_info_req(mti);
 343         struct tg_export_data *ted;
 344         struct lsd_client_data *lcd;
 345         loff_t off;
 346         int err;
 347         __s32 rc = th->th_result;
 348
 349         ENTRY;
 350         LASSERT(req);
 351         LASSERT(req->rq_export);
 352         LASSERT(mdt);
 353         ted = &req->rq_export->exp_target_data;
 354         LASSERT(ted);
 355
 356         cfs_mutex_lock(&ted->ted_lcd_lock);
 357         lcd = ted->ted_lcd;
 358         /* if the export has already been disconnected, we have no last_rcvd slot,
 359          * update server data with latest transno then */
 360         if (lcd == NULL) {
 361                 cfs_mutex_unlock(&ted->ted_lcd_lock);
 362                 CWARN("commit transaction for disconnected client %s: rc %d\n",
 363                       req->rq_export->exp_client_uuid.uuid, rc);
 364                 err = lut_server_data_write(mti->mti_env, &mdt->mdt_lut, th);
 365                 RETURN(err);
 366         }
 367
 368         off = ted->ted_lr_off;
 369         LASSERT(ergo(mti->mti_transno == 0, rc != 0));
 370         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 371             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 372                 if (mti->mti_transno != 0) {
 373                         if (lcd->lcd_last_close_transno > mti->mti_transno) {
 374                                 CERROR("Trying to overwrite bigger transno:"
 375                                        "on-disk: "LPU64", new: "LPU64" "
 376                                        "replay: %d. see LU-617.\n",
 377                                        lcd->lcd_last_close_transno,
 378                                        mti->mti_transno, req_is_replay(req));
 379                                 if (req_is_replay(req)) {
 380                                         cfs_spin_lock(&req->rq_export->exp_lock);
 381                                         req->rq_export->exp_vbr_failed = 1;
 382                                         cfs_spin_unlock(&req->rq_export->exp_lock);
 383                                 }
 384                                 cfs_mutex_unlock(&ted->ted_lcd_lock);
 385                                 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
 386                         }
 387                         lcd->lcd_last_close_transno = mti->mti_transno;
 388                 }
 389                 lcd->lcd_last_close_xid = req->rq_xid;
 390                 lcd->lcd_last_close_result = rc;
 391         } else {
 392                 /* VBR: save versions in last_rcvd for reconstruct. */
 393                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 394                 if (pre_versions) {
 395                         lcd->lcd_pre_versions[0] = pre_versions[0];
 396                         lcd->lcd_pre_versions[1] = pre_versions[1];
 397                         lcd->lcd_pre_versions[2] = pre_versions[2];
 398                         lcd->lcd_pre_versions[3] = pre_versions[3];
 399                 }
 400                 if (mti->mti_transno != 0) {
 401                         if (lcd->lcd_last_transno > mti->mti_transno) {
 402                                 CERROR("Trying to overwrite bigger transno:"
 403                                        "on-disk: "LPU64", new: "LPU64" "
 404                                        "replay: %d. see LU-617.\n",
 405                                        lcd->lcd_last_transno,
 406                                        mti->mti_transno, req_is_replay(req));
 407                                 if (req_is_replay(req)) {
 408                                         cfs_spin_lock(&req->rq_export->exp_lock);
 409                                         req->rq_export->exp_vbr_failed = 1;
 410                                         cfs_spin_unlock(&req->rq_export->exp_lock);
 411                                 }
 412                                 cfs_mutex_unlock(&ted->ted_lcd_lock);
 413                                 RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
 414                         }
 415                         lcd->lcd_last_transno = mti->mti_transno;
 416                 }
 417                 lcd->lcd_last_xid = req->rq_xid;
 418                 lcd->lcd_last_result = rc;
 419                 /*XXX: save intent_disposition in mdt_thread_info?
 420                  * also there is bug - intent_dispostion is __u64,
 421                  * see struct ldlm_reply->lock_policy_res1; */
 422                 lcd->lcd_last_data = mti->mti_opdata;
 423         }
 424
 425         if (off <= 0) {
 426                 CERROR("client idx %d has offset %lld\n", ted->ted_lr_idx, off);
 427                 err = -EINVAL;
 428         } else {
 429                 err = lut_client_data_write(mti->mti_env, &mdt->mdt_lut, lcd,
 430                                             &off, th);
 431         }
 432         cfs_mutex_unlock(&ted->ted_lcd_lock);
 433         RETURN(err);
 434 }
 435
 436 extern struct lu_context_key mdt_thread_key;
 437
 438 /* add credits for last_rcvd update */
 439 static int mdt_txn_start_cb(const struct lu_env *env,
 440                             struct thandle *th, void *cookie)
 441 {
 442         struct mdt_device *mdt = cookie;
 443         struct mdt_thread_info *mti;
 444         int rc;
 445         ENTRY;
 446
 447         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 448
 449         LASSERT(mdt->mdt_lut.lut_last_rcvd);
 450         if (mti->mti_exp == NULL)
 451                 RETURN(0);
 452
 453         rc = dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd,
 454                                      sizeof(struct lsd_client_data),
 455                                      mti->mti_exp->exp_target_data.ted_lr_off,
 456                                      th);
 457         if (rc)
 458                 return rc;
 459
 460         rc = dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd,
 461                                      sizeof(struct lr_server_data), 0, th);
 462         if (rc)
 463                 return rc;
 464
 465         if (mti->mti_mos != NULL)
 466                 rc = dt_declare_version_set(env, mdt_obj2dt(mti->mti_mos), th);
 467
 468         return rc;
 469 }
 470
 471 /* Update last_rcvd records with latests transaction data */
 472 static int mdt_txn_stop_cb(const struct lu_env *env,
 473                            struct thandle *txn, void *cookie)
 474 {
 475         struct mdt_device *mdt = cookie;
 476         struct mdt_thread_info *mti;
 477         struct ptlrpc_request *req;
 478
 479         mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key);
 480         req = mdt_info_req(mti);
 481
 482         if (mti->mti_mdt == NULL || req == NULL || mti->mti_no_need_trans) {
 483                 mti->mti_no_need_trans = 0;
 484                 return 0;
 485         }
 486
 487         if (mti->mti_has_trans) {
 488                 /* XXX: currently there are allowed cases, but the wrong cases
 489                  * are also possible, so better check is needed here */
 490                 CDEBUG(D_INFO, "More than one transaction "LPU64"\n",
 491                        mti->mti_transno);
 492                 return 0;
 493         }
 494
 495         mti->mti_has_trans = 1;
 496         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 497         if (txn->th_result != 0) {
 498                 if (mti->mti_transno != 0) {
 499                         CERROR("Replay transno "LPU64" failed: rc %d\n",
 500                                mti->mti_transno, txn->th_result);
 501                 }
 502         } else if (mti->mti_transno == 0) {
 503                 mti->mti_transno = ++ mdt->mdt_lut.lut_last_transno;
 504         } else {
 505                 /* should be replay */
 506                 if (mti->mti_transno > mdt->mdt_lut.lut_last_transno)
 507                         mdt->mdt_lut.lut_last_transno = mti->mti_transno;
 508         }
 509         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 510         /* sometimes the reply message has not been successfully packed */
 511         LASSERT(req != NULL && req->rq_repmsg != NULL);
 512
 513         /** VBR: set new versions */
 514         if (txn->th_result == 0 && mti->mti_mos != NULL) {
 515                 dt_version_set(env, mdt_obj2dt(mti->mti_mos),
 516                                mti->mti_transno, txn);
 517                 mti->mti_mos = NULL;
 518         }
 519
 520         /* filling reply data */
 521         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 522                mti->mti_transno, req->rq_export->exp_obd->obd_last_committed);
 523
 524         req->rq_transno = mti->mti_transno;
 525         lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno);
 526         /* if can't add callback, do sync write */
 527         txn->th_sync |= !!lut_last_commit_cb_add(txn, &mdt->mdt_lut,
 528                                                  mti->mti_exp,
 529                                                  mti->mti_transno);
 530         return mdt_last_rcvd_update(mti, txn);
 531 }
 532
 533 int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt,
 534                  struct obd_device *obd,
 535                  struct lustre_sb_info *lsi)
 536 {
 537         struct lu_fid fid;
 538         struct dt_object *o;
 539         int rc = 0;
 540         ENTRY;
 541
 542         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP))
 543                 RETURN(-ENOENT);
 544
 545         /* prepare transactions callbacks */
 546         mdt->mdt_txn_cb.dtc_txn_start = mdt_txn_start_cb;
 547         mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb;
 548         mdt->mdt_txn_cb.dtc_txn_commit = NULL;
 549         mdt->mdt_txn_cb.dtc_cookie = mdt;
 550         mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD;
 551         CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage);
 552
 553         dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb);
 554
 555         rc = mdt_server_data_init(env, mdt, lsi);
 556         if (rc)
 557                 RETURN(rc);
 558
 559         o = dt_store_open(env, mdt->mdt_bottom, "", CAPA_KEYS, &fid);
 560         if (!IS_ERR(o)) {
 561                 mdt->mdt_ck_obj = o;
 562                 rc = mdt_capa_keys_init(env, mdt);
 563                 if (rc)
 564                         GOTO(put_ck_object, rc);
 565         } else {
 566                 rc = PTR_ERR(o);
 567                 CERROR("cannot open %s: rc = %d\n", CAPA_KEYS, rc);
 568                 GOTO(disconnect_exports, rc);
 569         }
 570         RETURN(0);
 571
 572 put_ck_object:
 573         lu_object_put(env, &o->do_lu);
 574         mdt->mdt_ck_obj = NULL;
 575 disconnect_exports:
 576         class_disconnect_exports(obd);
 577         return rc;
 578 }
 579
 580 void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt)
 581 {
 582         ENTRY;
 583
 584         /* Remove transaction callback */
 585         dt_txn_callback_del(mdt->mdt_bottom, &mdt->mdt_txn_cb);
 586         if (mdt->mdt_ck_obj)
 587                 lu_object_put(env, &mdt->mdt_ck_obj->do_lu);
 588         mdt->mdt_ck_obj = NULL;
 589         EXIT;
 590 }
 591
 592 /* reconstruction code */
 593 static void mdt_steal_ack_locks(struct ptlrpc_request *req)
 594 {
 595         struct obd_export         *exp = req->rq_export;
 596         cfs_list_t                *tmp;
 597         struct ptlrpc_reply_state *oldrep;
 598         struct ptlrpc_service     *svc;
 599         int                        i;
 600
 601         /* CAVEAT EMPTOR: spinlock order */
 602         cfs_spin_lock(&exp->exp_lock);
 603         cfs_list_for_each (tmp, &exp->exp_outstanding_replies) {
 604                 oldrep = cfs_list_entry(tmp, struct ptlrpc_reply_state,
 605                                         rs_exp_list);
 606
 607                 if (oldrep->rs_xid != req->rq_xid)
 608                         continue;
 609
 610                 if (oldrep->rs_opc != lustre_msg_get_opc(req->rq_reqmsg))
 611                         CERROR ("Resent req xid "LPU64" has mismatched opc: "
 612                                 "new %d old %d\n", req->rq_xid,
 613                                 lustre_msg_get_opc(req->rq_reqmsg),
 614                                 oldrep->rs_opc);
 615
 616                 svc = oldrep->rs_service;
 617                 cfs_spin_lock (&svc->srv_rs_lock);
 618
 619                 cfs_list_del_init (&oldrep->rs_exp_list);
 620
 621                 CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64
 622                       " o%d NID %s\n",
 623                       oldrep->rs_nlocks, oldrep,
 624                       oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_opc,
 625                       libcfs_nid2str(exp->exp_connection->c_peer.nid));
 626
 627                 for (i = 0; i < oldrep->rs_nlocks; i++)
 628                         ptlrpc_save_lock(req, &oldrep->rs_locks[i],
 629                                          oldrep->rs_modes[i], 0);
 630                 oldrep->rs_nlocks = 0;
 631
 632                 DEBUG_REQ(D_HA, req, "stole locks for");
 633                 cfs_spin_lock(&oldrep->rs_lock);
 634                 ptlrpc_schedule_difficult_reply (oldrep);
 635                 cfs_spin_unlock(&oldrep->rs_lock);
 636
 637                 cfs_spin_unlock (&svc->srv_rs_lock);
 638                 break;
 639         }
 640         cfs_spin_unlock(&exp->exp_lock);
 641 }
 642
 643 /**
 644  * VBR: restore versions
 645  */
 646 void mdt_vbr_reconstruct(struct ptlrpc_request *req,
 647                          struct lsd_client_data *lcd)
 648 {
 649         __u64 pre_versions[4] = {0};
 650         pre_versions[0] = lcd->lcd_pre_versions[0];
 651         pre_versions[1] = lcd->lcd_pre_versions[1];
 652         pre_versions[2] = lcd->lcd_pre_versions[2];
 653         pre_versions[3] = lcd->lcd_pre_versions[3];
 654         lustre_msg_set_versions(req->rq_repmsg, pre_versions);
 655 }
 656
 657 void mdt_req_from_lcd(struct ptlrpc_request *req,
 658                       struct lsd_client_data *lcd)
 659 {
 660         DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d",
 661                   lcd->lcd_last_transno, lcd->lcd_last_result);
 662
 663         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 664             lustre_msg_get_opc(req->rq_repmsg) == MDS_DONE_WRITING) {
 665                 req->rq_transno = lcd->lcd_last_close_transno;
 666                 req->rq_status = lcd->lcd_last_close_result;
 667         } else {
 668                 req->rq_transno = lcd->lcd_last_transno;
 669                 req->rq_status = lcd->lcd_last_result;
 670                 mdt_vbr_reconstruct(req, lcd);
 671         }
 672         if (req->rq_status != 0)
 673                 req->rq_transno = 0;
 674         lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
 675         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 676         DEBUG_REQ(D_RPCTRACE, req, "restoring transno "LPD64"/status %d",
 677                   req->rq_transno, req->rq_status);
 678
 679         mdt_steal_ack_locks(req);
 680 }
 681
 682 void mdt_reconstruct_generic(struct mdt_thread_info *mti,
 683                              struct mdt_lock_handle *lhc)
 684 {
 685         struct ptlrpc_request *req = mdt_info_req(mti);
 686         struct tg_export_data *ted = &req->rq_export->exp_target_data;
 687
 688         return mdt_req_from_lcd(req, ted->ted_lcd);
 689 }
 690
 691 static void mdt_reconstruct_create(struct mdt_thread_info *mti,
 692                                    struct mdt_lock_handle *lhc)
 693 {
 694         struct ptlrpc_request  *req = mdt_info_req(mti);
 695         struct obd_export *exp = req->rq_export;
 696         struct tg_export_data *ted = &exp->exp_target_data;
 697         struct mdt_device *mdt = mti->mti_mdt;
 698         struct mdt_object *child;
 699         struct mdt_body *body;
 700         int rc;
 701
 702         mdt_req_from_lcd(req, ted->ted_lcd);
 703         if (req->rq_status)
 704                 return;
 705
 706         /* if no error, so child was created with requested fid */
 707         child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2);
 708         if (IS_ERR(child)) {
 709                 rc = PTR_ERR(child);
 710                 LCONSOLE_WARN("Child "DFID" lookup error %d."
 711                               " Evicting client %s with export %s.\n",
 712                               PFID(mdt_object_fid(child)), rc,
 713                               obd_uuid2str(&exp->exp_client_uuid),
 714                               obd_export_nid2str(exp));
 715                 mdt_export_evict(exp);
 716                 EXIT;
 717                 return;
 718         }
 719
 720         body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
 721         mti->mti_attr.ma_need = MA_INODE;
 722         mti->mti_attr.ma_valid = 0;
 723         rc = mo_attr_get(mti->mti_env, mdt_object_child(child), &mti->mti_attr);
 724         if (rc == -EREMOTE) {
 725                 /* object was created on remote server */
 726                 req->rq_status = rc;
 727                 body->valid |= OBD_MD_MDS;
 728         }
 729         mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
 730                            mdt_object_fid(child));
 731         mdt_object_put(mti->mti_env, child);
 732 }
 733
 734 static void mdt_reconstruct_setattr(struct mdt_thread_info *mti,
 735                                     struct mdt_lock_handle *lhc)
 736 {
 737         struct ptlrpc_request  *req = mdt_info_req(mti);
 738         struct obd_export *exp = req->rq_export;
 739         struct mdt_export_data *med = &exp->exp_mdt_data;
 740         struct mdt_device *mdt = mti->mti_mdt;
 741         struct mdt_object *obj;
 742         struct mdt_body *body;
 743
 744         mdt_req_from_lcd(req, med->med_ted.ted_lcd);
 745         if (req->rq_status)
 746                 return;
 747
 748         body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
 749         obj = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid1);
 750         if (IS_ERR(obj)) {
 751                 int rc = PTR_ERR(obj);
 752                 LCONSOLE_WARN(""DFID" lookup error %d."
 753                               " Evicting client %s with export %s.\n",
 754                               PFID(mdt_object_fid(obj)), rc,
 755                               obd_uuid2str(&exp->exp_client_uuid),
 756                               obd_export_nid2str(exp));
 757                 mdt_export_evict(exp);
 758                 EXIT;
 759                 return;
 760         }
 761         mti->mti_attr.ma_need = MA_INODE;
 762         mti->mti_attr.ma_valid = 0;
 763         mo_attr_get(mti->mti_env, mdt_object_child(obj), &mti->mti_attr);
 764         mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr,
 765                            mdt_object_fid(obj));
 766         if (mti->mti_ioepoch && (mti->mti_ioepoch->flags & MF_EPOCH_OPEN)) {
 767                 struct mdt_file_data *mfd;
 768                 struct mdt_body *repbody;
 769
 770                 repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY);
 771                 repbody->ioepoch = obj->mot_ioepoch;
 772                 cfs_spin_lock(&med->med_open_lock);
 773                 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
 774                         if (mfd->mfd_xid == req->rq_xid)
 775                                 break;
 776                 }
 777                 LASSERT(&mfd->mfd_list != &med->med_open_head);
 778                 cfs_spin_unlock(&med->med_open_lock);
 779                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 780         }
 781
 782         mdt_object_put(mti->mti_env, obj);
 783 }
 784
 785 typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,
 786                                   struct mdt_lock_handle *lhc);
 787
 788 static mdt_reconstructor reconstructors[REINT_MAX] = {
 789         [REINT_SETATTR]  = mdt_reconstruct_setattr,
 790         [REINT_CREATE]   = mdt_reconstruct_create,
 791         [REINT_LINK]     = mdt_reconstruct_generic,
 792         [REINT_UNLINK]   = mdt_reconstruct_generic,
 793         [REINT_RENAME]   = mdt_reconstruct_generic,
 794         [REINT_OPEN]     = mdt_reconstruct_open,
 795         [REINT_SETXATTR] = mdt_reconstruct_generic
 796 };
 797
 798 void mdt_reconstruct(struct mdt_thread_info *mti,
 799                      struct mdt_lock_handle *lhc)
 800 {
 801         ENTRY;
 802         reconstructors[mti->mti_rr.rr_opcode](mti, lhc);
 803         EXIT;
 804 }