lustre/mdt/mdt_open.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_open.c
  37  *
  38  * Lustre Metadata Target (mdt) open/close file handling
  39  *
  40  * Author: Huang Hua <huanghua@clusterfs.com>
  41  */
  42
  43 #ifndef EXPORT_SYMTAB
  44 # define EXPORT_SYMTAB
  45 #endif
  46 #define DEBUG_SUBSYSTEM S_MDS
  47
  48 #include <lustre_acl.h>
  49 #include <lustre_mds.h>
  50 #include "mdt_internal.h"
  51
  52 /* we do nothing because we do not have refcount now */
  53 static void mdt_mfd_get(void *mfdp)
  54 {
  55 }
  56
  57 /* Create a new mdt_file_data struct, initialize it,
  58  * and insert it to global hash table */
  59 struct mdt_file_data *mdt_mfd_new(void)
  60 {
  61         struct mdt_file_data *mfd;
  62         ENTRY;
  63
  64         OBD_ALLOC_PTR(mfd);
  65         if (mfd != NULL) {
  66                 CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  67                 CFS_INIT_LIST_HEAD(&mfd->mfd_list);
  68                 class_handle_hash(&mfd->mfd_handle, mdt_mfd_get);
  69         }
  70         RETURN(mfd);
  71 }
  72
  73 /*
  74  * Find the mfd pointed to by handle in global hash table.
  75  * In case of replay the handle is obsoleted
  76  * but mfd can be found in mfd list by that handle
  77  */
  78 struct mdt_file_data *mdt_handle2mfd(struct mdt_thread_info *info,
  79                                      const struct lustre_handle *handle)
  80 {
  81         struct ptlrpc_request *req = mdt_info_req(info);
  82         struct mdt_file_data  *mfd;
  83         ENTRY;
  84
  85         LASSERT(handle != NULL);
  86         mfd = class_handle2object(handle->cookie);
  87         /* during dw/setattr replay the mfd can be found by old handle */
  88         if (mfd == NULL && req_is_replay(req)) {
  89                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
  90                 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  91                         if (mfd->mfd_old_handle.cookie == handle->cookie)
  92                                 RETURN (mfd);
  93                 }
  94                 mfd = NULL;
  95         }
  96         RETURN (mfd);
  97 }
  98
  99 /* free mfd */
 100 void mdt_mfd_free(struct mdt_file_data *mfd)
 101 {
 102         LASSERT(cfs_list_empty(&mfd->mfd_list));
 103         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 104 }
 105
 106 static int mdt_create_data(struct mdt_thread_info *info,
 107                            struct mdt_object *p, struct mdt_object *o)
 108 {
 109         struct md_op_spec     *spec = &info->mti_spec;
 110         struct md_attr        *ma   = &info->mti_attr;
 111         int                    rc   = 0;
 112         ENTRY;
 113
 114         if (!md_should_create(spec->sp_cr_flags))
 115                 RETURN(0);
 116
 117         ma->ma_need = MA_INODE | MA_LOV;
 118         ma->ma_valid = 0;
 119         cfs_mutex_lock(&o->mot_lov_mutex);
 120         if (!(o->mot_flags & MOF_LOV_CREATED)) {
 121                 rc = mdo_create_data(info->mti_env,
 122                                      p ? mdt_object_child(p) : NULL,
 123                                      mdt_object_child(o), spec, ma);
 124                 if (rc == 0 && ma->ma_valid & MA_LOV)
 125                         o->mot_flags |= MOF_LOV_CREATED;
 126         }
 127         cfs_mutex_unlock(&o->mot_lov_mutex);
 128         RETURN(rc);
 129 }
 130
 131 static int mdt_ioepoch_opened(struct mdt_object *mo)
 132 {
 133         return mo->mot_ioepoch_count;
 134 }
 135
 136 int mdt_object_is_som_enabled(struct mdt_object *mo)
 137 {
 138         return !mo->mot_ioepoch;
 139 }
 140
 141 /**
 142  * Re-enable Size-on-MDS.
 143  * Call under ->mot_ioepoch_mutex.
 144  */
 145 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
 146 {
 147         if (ioepoch == mo->mot_ioepoch) {
 148                 LASSERT(!mdt_ioepoch_opened(mo));
 149                 mo->mot_ioepoch = 0;
 150                 mo->mot_flags = 0;
 151         }
 152 }
 153
 154 /**
 155  * Open the IOEpoch. It is allowed if @writecount is not negative.
 156  * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
 157  */
 158 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
 159                      int created)
 160 {
 161         struct mdt_device *mdt = info->mti_mdt;
 162         int cancel = 0;
 163         int rc = 0;
 164         ENTRY;
 165
 166         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 167             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
 168                 RETURN(0);
 169
 170         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 171         if (mdt_ioepoch_opened(o)) {
 172                 /* Epoch continues even if there is no writers yet. */
 173                 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
 174                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 175         } else {
 176                 /* XXX: ->mdt_ioepoch is not initialized at the mount */
 177                 cfs_spin_lock(&mdt->mdt_ioepoch_lock);
 178                 if (mdt->mdt_ioepoch < info->mti_replayepoch)
 179                         mdt->mdt_ioepoch = info->mti_replayepoch;
 180
 181                 if (info->mti_replayepoch)
 182                         o->mot_ioepoch = info->mti_replayepoch;
 183                 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
 184                         o->mot_ioepoch = ++mdt->mdt_ioepoch;
 185                 else
 186                         o->mot_ioepoch = mdt->mdt_ioepoch;
 187
 188                 cfs_spin_unlock(&mdt->mdt_ioepoch_lock);
 189
 190                 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
 191                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 192                 if (created)
 193                         o->mot_flags |= MOF_SOM_CREATED;
 194                 cancel = 1;
 195         }
 196         o->mot_ioepoch_count++;
 197         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 198
 199         /* Cancel Size-on-MDS attributes cached on clients for the open case.
 200          * In the truncate case, see mdt_reint_setattr(). */
 201         if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
 202                 struct mdt_lock_handle  *lh = &info->mti_lh[MDT_LH_CHILD];
 203                 mdt_lock_reg_init(lh, LCK_EX);
 204                 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
 205                                      MDT_LOCAL_LOCK);
 206                 if (rc == 0)
 207                         mdt_object_unlock(info, o, lh, 1);
 208         }
 209         RETURN(rc);
 210 }
 211
 212 /**
 213  * Update SOM on-disk attributes.
 214  * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
 215  * mountid and attributes. If disabling, zero IOEpoch id in lustre-ea.
 216  * Call under ->mot_ioepoch_mutex.
 217  */
 218 static int mdt_som_attr_set(struct mdt_thread_info *info,
 219                             struct mdt_object *obj, __u64 ioepoch, int enable)
 220 {
 221         struct md_attr *ma = &info->mti_attr;
 222         int rc;
 223         ENTRY;
 224
 225         CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
 226                " on "DFID".\n", enable ? "update" : "disabling",
 227                ioepoch, PFID(mdt_object_fid(obj)));
 228
 229         ma->ma_valid |= MA_SOM;
 230         ma->ma_som = &info->mti_u.som.data;
 231         if (enable) {
 232                 struct mdt_device *mdt = info->mti_mdt;
 233                 struct lu_attr *la = &ma->ma_attr;
 234
 235                 ma->ma_som->msd_ioepoch = ioepoch;
 236                 ma->ma_som->msd_size = la->la_valid & LA_SIZE ? la->la_size : 0;
 237                 ma->ma_som->msd_blocks = la->la_valid & LA_BLOCKS ?
 238                                          la->la_blocks : 0;
 239                 ma->ma_som->msd_mountid = mdt->mdt_lut.lut_obd->u.obt.obt_mount_count;
 240                 ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
 241         } else {
 242                 ma->ma_som->msd_ioepoch = IOEPOCH_INVAL;
 243                 ma->ma_attr.la_valid &= LA_ATIME;
 244         }
 245
 246         /* Since we have opened the file, it is unnecessary
 247          * to check permission when close it. Between the "open"
 248          * and "close", maybe someone has changed the file mode
 249          * or flags, or the file created mode do not permit wirte,
 250          * and so on. Just set MDS_PERM_BYPASS for all the cases. */
 251         ma->ma_attr_flags |= MDS_PERM_BYPASS | MDS_SOM;
 252
 253         rc = mdt_attr_set(info, obj, ma, 0);
 254         RETURN(rc);
 255 }
 256
 257 /** Perform the eviction specific actions on ioepoch close. */
 258 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
 259                                                 struct mdt_object *o)
 260 {
 261         int rc = 0;
 262
 263         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 264         CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
 265                "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
 266                o->mot_ioepoch_count);
 267         o->mot_ioepoch_count--;
 268
 269         /* If eviction occured set MOF_SOM_RECOV,
 270          * if no other epoch holders, disable SOM on disk. */
 271         o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
 272         if (!mdt_ioepoch_opened(o)) {
 273                 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
 274                 mdt_object_som_enable(o, o->mot_ioepoch);
 275         }
 276         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 277         RETURN(rc);
 278 }
 279
 280 /**
 281  * Perform the replay specific actions on ioepoch close.
 282  * Skip SOM attribute update if obtained and just forget about the inode state
 283  * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
 284  */
 285 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
 286                                               struct mdt_object *o)
 287 {
 288         int rc = MDT_IOEPOCH_CLOSED;
 289         ENTRY;
 290
 291         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 292         CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
 293                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
 294         o->mot_ioepoch_count--;
 295
 296         /* Get an info from the replayed request if client is supposed
 297          * to send an Attibute Update, reconstruct @rc if so */
 298         if (info->mti_ioepoch->flags & MF_SOM_AU)
 299                 rc = MDT_IOEPOCH_GETATTR;
 300
 301         if (!mdt_ioepoch_opened(o))
 302                 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
 303         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 304
 305         RETURN(rc);
 306 }
 307
 308 /**
 309  * Regular file IOepoch close.
 310  * Closes the ioepoch, checks the object state, apply obtained attributes and
 311  * re-enable SOM on the object, if possible. Also checks if the recovery is
 312  * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
 313  * to obtain SOM attributes under the server-side OST locks.
 314  *
 315  * Return value:
 316  * MDT_IOEPOCH_CLOSED if ioepoch is closed.
 317  * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
 318  */
 319 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
 320                                         struct mdt_object *o)
 321 {
 322         struct md_attr *tmp_ma;
 323         struct lu_attr *la;
 324         int achange, opened;
 325         int recovery = 0;
 326         int rc = 0, ret = MDT_IOEPOCH_CLOSED;
 327         ENTRY;
 328
 329         la = &info->mti_attr.ma_attr;
 330         achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
 331
 332         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 333         o->mot_ioepoch_count--;
 334
 335         tmp_ma = &info->mti_u.som.attr;
 336         tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
 337         tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
 338         tmp_ma->ma_som = &info->mti_u.som.data;
 339         tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
 340         tmp_ma->ma_valid = 0;
 341         rc = mo_attr_get(info->mti_env, mdt_object_child(o), tmp_ma);
 342         if (rc)
 343                 GOTO(error_up, rc);
 344
 345         /* Check the on-disk SOM state. */
 346         if (o->mot_flags & MOF_SOM_RECOV)
 347                 recovery = 1;
 348         else if (!(o->mot_flags & MOF_SOM_CREATED) &&
 349                  !(tmp_ma->ma_valid & MA_SOM))
 350                 recovery = 1;
 351
 352         CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
 353                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
 354
 355         opened = mdt_ioepoch_opened(o);
 356         /**
 357          * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
 358          * Skip the check for file with no LOV  or for unlink files.
 359          */
 360         if (!opened && tmp_ma->ma_valid & MA_LOV &&
 361             !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
 362                 if (recovery)
 363                         /* If some previous writer was evicted, re-ask the
 364                          * client for attributes. Even if attributes are
 365                          * provided, we cannot believe in them.
 366                          * Another use case is that there is no SOM cache on
 367                          * disk -- first access with SOM or there was an MDS
 368                          * failure. */
 369                         ret = MDT_IOEPOCH_GETATTR;
 370                 else if (o->mot_flags & MOF_SOM_CHANGE)
 371                         /* Some previous writer changed the attribute.
 372                          * Do not believe to the current Size-on-MDS
 373                          * update, re-ask client. */
 374                         ret = MDT_IOEPOCH_GETATTR;
 375                 else if (!(la->la_valid & LA_SIZE) && achange)
 376                         /* Attributes were changed by the last writer
 377                          * only but no Size-on-MDS update is received.*/
 378                         ret = MDT_IOEPOCH_GETATTR;
 379         }
 380
 381         if (achange || ret == MDT_IOEPOCH_GETATTR)
 382                 o->mot_flags |= MOF_SOM_CHANGE;
 383
 384         /* If epoch ends and relable SOM attributes are obtained, update them.
 385          * Create SOM ea for new files even if there is no attributes obtained
 386          * (0-length file). */
 387         if (ret == MDT_IOEPOCH_CLOSED && !opened) {
 388                 if (achange || o->mot_flags & MOF_SOM_CREATED) {
 389                         LASSERT(achange || !(la->la_valid & LA_SIZE));
 390                         rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
 391                                               MDT_SOM_ENABLE);
 392                         /* Avoid the following setattrs of these attributes,
 393                          * e.g. for atime update. */
 394                         info->mti_attr.ma_valid = 0;
 395                 }
 396                 mdt_object_som_enable(o, o->mot_ioepoch);
 397         }
 398
 399         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 400         /* If recovery is needed, tell the client to perform GETATTR under
 401          * the lock. */
 402         if (ret == MDT_IOEPOCH_GETATTR && recovery) {
 403                 struct mdt_body *rep;
 404                 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 405                 rep->valid |= OBD_MD_FLGETATTRLOCK;
 406         }
 407
 408         RETURN(rc ? : ret);
 409
 410 error_up:
 411         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 412         return rc;
 413 }
 414
 415 /**
 416  * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
 417  * - a client closes the IOEpoch;
 418  * - a client eviction occured.
 419  * Return values:
 420  * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
 421  * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
 422  * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
 423  * update is needed.
 424  */
 425 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
 426 {
 427         struct ptlrpc_request *req = mdt_info_req(info);
 428         ENTRY;
 429
 430         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 431             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
 432                 RETURN(0);
 433
 434         LASSERT(o->mot_ioepoch_count);
 435         LASSERT(info->mti_ioepoch == NULL ||
 436                 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
 437
 438         /* IOEpoch is closed only if client tells about it or eviction occures.
 439          * In the replay case, always close the epoch. */
 440         if (req == NULL)
 441                 RETURN(mdt_ioepoch_close_on_eviction(info, o));
 442         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
 443                 RETURN(mdt_ioepoch_close_on_replay(info, o));
 444         if (info->mti_ioepoch->flags & MF_EPOCH_CLOSE)
 445                 RETURN(mdt_ioepoch_close_reg(info, o));
 446         /* IO epoch is not closed. */
 447         RETURN(MDT_IOEPOCH_OPENED);
 448 }
 449
 450 /**
 451  * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
 452  * for attribute update. It happens if:
 453  * - SOM Attribute Update is obtained;
 454  * - the client failed to obtain it and informs MDS about it;
 455  * - a client eviction occured.
 456  * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
 457  * cache otherwise.
 458  */
 459 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
 460 {
 461         struct ptlrpc_request *req = mdt_info_req(info);
 462         __u64 ioepoch = 0;
 463         int act = MDT_SOM_ENABLE;
 464         int rc = 0;
 465         ENTRY;
 466
 467         LASSERT(!req || info->mti_ioepoch);
 468         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 469             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
 470                 RETURN(0);
 471
 472         /* No size whereas MF_SOM_CHANGE is set means client failed to
 473          * obtain ost attributes, drop the SOM cache on disk if so. */
 474         if (!req ||
 475             (info->mti_ioepoch &&
 476              info->mti_ioepoch->flags & MF_SOM_CHANGE &&
 477              !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
 478                 act = MDT_SOM_DISABLE;
 479
 480         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 481         /* Mark the object it is the recovery state if we failed to obtain
 482          * SOM attributes. */
 483         if (act == MDT_SOM_DISABLE)
 484                 o->mot_flags |= MOF_SOM_RECOV;
 485
 486         if (!mdt_ioepoch_opened(o)) {
 487                 ioepoch =  info->mti_ioepoch ?
 488                         info->mti_ioepoch->ioepoch : o->mot_ioepoch;
 489
 490                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 491                         rc = mdt_som_attr_set(info, o, ioepoch, act);
 492                 mdt_object_som_enable(o, ioepoch);
 493         }
 494         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 495         RETURN(rc);
 496 }
 497
 498 int mdt_write_read(struct mdt_object *o)
 499 {
 500         int rc = 0;
 501         ENTRY;
 502         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 503         rc = o->mot_writecount;
 504         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 505         RETURN(rc);
 506 }
 507
 508 int mdt_write_get(struct mdt_object *o)
 509 {
 510         int rc = 0;
 511         ENTRY;
 512         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 513         if (o->mot_writecount < 0)
 514                 rc = -ETXTBSY;
 515         else
 516                 o->mot_writecount++;
 517         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 518         RETURN(rc);
 519 }
 520
 521 void mdt_write_put(struct mdt_object *o)
 522 {
 523         ENTRY;
 524         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 525         o->mot_writecount--;
 526         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 527         EXIT;
 528 }
 529
 530 static int mdt_write_deny(struct mdt_object *o)
 531 {
 532         int rc = 0;
 533         ENTRY;
 534         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 535         if (o->mot_writecount > 0)
 536                 rc = -ETXTBSY;
 537         else
 538                 o->mot_writecount--;
 539         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 540         RETURN(rc);
 541 }
 542
 543 static void mdt_write_allow(struct mdt_object *o)
 544 {
 545         ENTRY;
 546         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 547         o->mot_writecount++;
 548         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 549         EXIT;
 550 }
 551
 552 /* there can be no real transaction so prepare the fake one */
 553 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 554 {
 555         struct mdt_device      *mdt = info->mti_mdt;
 556         struct ptlrpc_request  *req = mdt_info_req(info);
 557         struct tg_export_data  *ted;
 558         struct lsd_client_data *lcd;
 559
 560         ENTRY;
 561         /* transaction has occurred already */
 562         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 563                 RETURN_EXIT;
 564
 565         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 566         if (info->mti_transno == 0) {
 567                 info->mti_transno = ++ mdt->mdt_lut.lut_last_transno;
 568         } else {
 569                 /* should be replay */
 570                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 571                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 572         }
 573         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 574
 575         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 576                         info->mti_transno,
 577                         req->rq_export->exp_obd->obd_last_committed);
 578
 579         req->rq_transno = info->mti_transno;
 580         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 581
 582         /* update lcd in memory only for resent cases */
 583         ted = &req->rq_export->exp_target_data;
 584         LASSERT(ted);
 585         cfs_mutex_lock(&ted->ted_lcd_lock);
 586         lcd = ted->ted_lcd;
 587         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 588             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 589                 if (info->mti_transno != 0)
 590                         lcd->lcd_last_close_transno = info->mti_transno;
 591                 lcd->lcd_last_close_xid = req->rq_xid;
 592                 lcd->lcd_last_close_result = rc;
 593         } else {
 594                 /* VBR: save versions in last_rcvd for reconstruct. */
 595                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 596                 if (pre_versions) {
 597                         lcd->lcd_pre_versions[0] = pre_versions[0];
 598                         lcd->lcd_pre_versions[1] = pre_versions[1];
 599                         lcd->lcd_pre_versions[2] = pre_versions[2];
 600                         lcd->lcd_pre_versions[3] = pre_versions[3];
 601                 }
 602                 if (info->mti_transno != 0)
 603                         lcd->lcd_last_transno = info->mti_transno;
 604                 lcd->lcd_last_xid = req->rq_xid;
 605                 lcd->lcd_last_result = rc;
 606                 lcd->lcd_last_data = info->mti_opdata;
 607         }
 608         cfs_mutex_unlock(&ted->ted_lcd_lock);
 609
 610         EXIT;
 611 }
 612
 613 void mdt_mfd_set_mode(struct mdt_file_data *mfd, int mode)
 614 {
 615         LASSERT(mfd != NULL);
 616
 617         CDEBUG(D_HA, "Change mfd %p mode 0x%x->0x%x\n",
 618                mfd, (unsigned int)mfd->mfd_mode, (unsigned int)mode);
 619
 620         mfd->mfd_mode = mode;
 621 }
 622
 623 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 624                         struct mdt_object *o, __u64 flags, int created)
 625 {
 626         struct ptlrpc_request   *req = mdt_info_req(info);
 627         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 628         struct mdt_file_data    *mfd;
 629         struct md_attr          *ma  = &info->mti_attr;
 630         struct lu_attr          *la  = &ma->ma_attr;
 631         struct mdt_body         *repbody;
 632         int                      rc = 0, isdir, isreg;
 633         ENTRY;
 634
 635         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 636
 637         isreg = S_ISREG(la->la_mode);
 638         isdir = S_ISDIR(la->la_mode);
 639         if (isreg && !(ma->ma_valid & MA_LOV)) {
 640                 /*
 641                  * No EA, check whether it is will set regEA and dirEA since in
 642                  * above attr get, these size might be zero, so reset it, to
 643                  * retrieve the MD after create obj.
 644                  */
 645                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 646                                                        &RMF_MDT_MD,
 647                                                        RCL_SERVER);
 648                 /* in replay case, p == NULL */
 649                 rc = mdt_create_data(info, p, o);
 650                 if (rc)
 651                         RETURN(rc);
 652         }
 653
 654         CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
 655                ma->ma_valid, ma->ma_lmm_size);
 656
 657         if (ma->ma_valid & MA_LOV) {
 658                 LASSERT(ma->ma_lmm_size != 0);
 659                 repbody->eadatasize = ma->ma_lmm_size;
 660                 if (isdir)
 661                         repbody->valid |= OBD_MD_FLDIREA;
 662                 else
 663                         repbody->valid |= OBD_MD_FLEASIZE;
 664         }
 665
 666         if (flags & FMODE_WRITE) {
 667                 rc = mdt_write_get(o);
 668                 if (rc == 0) {
 669                         mdt_ioepoch_open(info, o, created);
 670                         repbody->ioepoch = o->mot_ioepoch;
 671                 }
 672         } else if (flags & MDS_FMODE_EXEC) {
 673                 rc = mdt_write_deny(o);
 674         }
 675         if (rc)
 676                 RETURN(rc);
 677
 678         rc = mo_open(info->mti_env, mdt_object_child(o),
 679                      created ? flags | MDS_OPEN_CREATED : flags);
 680         if (rc)
 681                 GOTO(err_out, rc);
 682
 683         mfd = mdt_mfd_new();
 684         if (mfd != NULL) {
 685                 /*
 686                  * Keep a reference on this object for this open, and is
 687                  * released by mdt_mfd_close().
 688                  */
 689                 mdt_object_get(info->mti_env, o);
 690
 691                 /*
 692                  * @flags is always not zero. At least it should be FMODE_READ,
 693                  * FMODE_WRITE or MDS_FMODE_EXEC.
 694                  */
 695                 LASSERT(flags != 0);
 696
 697                 /* Open handling. */
 698                 mdt_mfd_set_mode(mfd, flags);
 699
 700                 mfd->mfd_object = o;
 701                 mfd->mfd_xid = req->rq_xid;
 702
 703                 /* replay handle */
 704                 if (req_is_replay(req)) {
 705                         struct mdt_file_data *old_mfd;
 706                         /* Check wheather old cookie already exist in
 707                          * the list, becasue when do recovery, client
 708                          * might be disconnected from server, and
 709                          * restart replay, so there maybe some orphan
 710                          * mfd here, we should remove them */
 711                         LASSERT(info->mti_rr.rr_handle != NULL);
 712                         old_mfd = mdt_handle2mfd(info, info->mti_rr.rr_handle);
 713                         if (old_mfd) {
 714                                 CDEBUG(D_HA, "del orph mfd %p fid=("DFID") "
 715                                        "cookie=" LPX64"\n", mfd,
 716                                        PFID(mdt_object_fid(mfd->mfd_object)),
 717                                        info->mti_rr.rr_handle->cookie);
 718                                 cfs_spin_lock(&med->med_open_lock);
 719                                 class_handle_unhash(&old_mfd->mfd_handle);
 720                                 cfs_list_del_init(&old_mfd->mfd_list);
 721                                 cfs_spin_unlock(&med->med_open_lock);
 722                                 /* no attr update for that close */
 723                                 la->la_valid = 0;
 724                                 ma->ma_valid |= MA_FLAGS;
 725                                 ma->ma_attr_flags |= MDS_RECOV_OPEN;
 726                                 mdt_mfd_close(info, old_mfd);
 727                                 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 728                                 ma->ma_valid &= ~MA_FLAGS;
 729                         }
 730                         CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
 731                                info->mti_rr.rr_handle->cookie);
 732                         mfd->mfd_old_handle.cookie =
 733                                                 info->mti_rr.rr_handle->cookie;
 734                 }
 735                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 736
 737                 if (req->rq_export->exp_disconnected) {
 738                         cfs_spin_lock(&med->med_open_lock);
 739                         class_handle_unhash(&mfd->mfd_handle);
 740                         cfs_list_del_init(&mfd->mfd_list);
 741                         cfs_spin_unlock(&med->med_open_lock);
 742                         mdt_mfd_close(info, mfd);
 743                 } else {
 744                         cfs_spin_lock(&med->med_open_lock);
 745                         cfs_list_add(&mfd->mfd_list, &med->med_open_head);
 746                         cfs_spin_unlock(&med->med_open_lock);
 747                 }
 748
 749                 mdt_empty_transno(info, rc);
 750         } else {
 751                 GOTO(err_out, rc = -ENOMEM);
 752         }
 753
 754         RETURN(rc);
 755
 756 err_out:
 757         if (flags & FMODE_WRITE)
 758                         /* XXX We also need to close io epoch here.
 759                          * See LU-1220 - green */
 760                 mdt_write_put(o);
 761         else if (flags & FMODE_EXEC)
 762                 mdt_write_allow(o);
 763         return rc;
 764 }
 765
 766 int mdt_finish_open(struct mdt_thread_info *info,
 767                     struct mdt_object *p, struct mdt_object *o,
 768                     __u64 flags, int created, struct ldlm_reply *rep)
 769 {
 770         struct ptlrpc_request   *req = mdt_info_req(info);
 771         struct obd_export       *exp = req->rq_export;
 772         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 773         struct md_attr          *ma  = &info->mti_attr;
 774         struct lu_attr          *la  = &ma->ma_attr;
 775         struct mdt_file_data    *mfd;
 776         struct mdt_body         *repbody;
 777         int                      rc = 0;
 778         int                      isreg, isdir, islnk;
 779         cfs_list_t              *t;
 780         ENTRY;
 781
 782         LASSERT(ma->ma_valid & MA_INODE);
 783
 784         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 785
 786         isreg = S_ISREG(la->la_mode);
 787         isdir = S_ISDIR(la->la_mode);
 788         islnk = S_ISLNK(la->la_mode);
 789         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 790
 791         if (exp_connect_rmtclient(exp)) {
 792                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 793
 794                 rc = mdt_pack_remote_perm(info, o, buf);
 795                 if (rc) {
 796                         repbody->valid &= ~OBD_MD_FLRMTPERM;
 797                         repbody->aclsize = 0;
 798                 } else {
 799                         repbody->valid |= OBD_MD_FLRMTPERM;
 800                         repbody->aclsize = sizeof(struct mdt_remote_perm);
 801                 }
 802         }
 803 #ifdef CONFIG_FS_POSIX_ACL
 804         else if (exp->exp_connect_flags & OBD_CONNECT_ACL) {
 805                 const struct lu_env *env = info->mti_env;
 806                 struct md_object *next = mdt_object_child(o);
 807                 struct lu_buf *buf = &info->mti_buf;
 808
 809                 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 810                 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
 811                                                    RCL_SERVER);
 812                 if (buf->lb_len > 0) {
 813                         rc = mo_xattr_get(env, next, buf,
 814                                           XATTR_NAME_ACL_ACCESS);
 815                         if (rc < 0) {
 816                                 if (rc == -ENODATA) {
 817                                         repbody->aclsize = 0;
 818                                         repbody->valid |= OBD_MD_FLACL;
 819                                         rc = 0;
 820                                 } else if (rc == -EOPNOTSUPP) {
 821                                         rc = 0;
 822                                 } else {
 823                                         CERROR("got acl size: %d\n", rc);
 824                                 }
 825                         } else {
 826                                 repbody->aclsize = rc;
 827                                 repbody->valid |= OBD_MD_FLACL;
 828                                 rc = 0;
 829                         }
 830                 }
 831         }
 832 #endif
 833
 834         if (info->mti_mdt->mdt_opts.mo_mds_capa &&
 835             exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) {
 836                 struct lustre_capa *capa;
 837
 838                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
 839                 LASSERT(capa);
 840                 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
 841                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 842                 if (rc)
 843                         RETURN(rc);
 844                 repbody->valid |= OBD_MD_FLMDSCAPA;
 845         }
 846         if (info->mti_mdt->mdt_opts.mo_oss_capa &&
 847             exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA &&
 848             S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) {
 849                 struct lustre_capa *capa;
 850
 851                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
 852                 LASSERT(capa);
 853                 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
 854                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 855                 if (rc)
 856                         RETURN(rc);
 857                 repbody->valid |= OBD_MD_FLOSSCAPA;
 858         }
 859
 860         /*
 861          * If we are following a symlink, don't open; and do not return open
 862          * handle for special nodes as client required.
 863          */
 864         if (islnk || (!isreg && !isdir &&
 865             (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH))) {
 866                 lustre_msg_set_transno(req->rq_repmsg, 0);
 867                 RETURN(0);
 868         }
 869
 870         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 871
 872         /*
 873          * We need to return the existing object's fid back, so it is done here,
 874          * after preparing the reply.
 875          */
 876         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 877                 RETURN(-EEXIST);
 878
 879         /* This can't be done earlier, we need to return reply body */
 880         if (isdir) {
 881                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 882                         /* We are trying to create or write an existing dir. */
 883                         RETURN(-EISDIR);
 884                 }
 885         } else if (flags & MDS_OPEN_DIRECTORY)
 886                 RETURN(-ENOTDIR);
 887
 888         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 889                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 890                 RETURN(-EAGAIN);
 891         }
 892
 893         mfd = NULL;
 894         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 895                 cfs_spin_lock(&med->med_open_lock);
 896                 cfs_list_for_each(t, &med->med_open_head) {
 897                         mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
 898                         if (mfd->mfd_xid == req->rq_xid) {
 899                                 break;
 900                         }
 901                         mfd = NULL;
 902                 }
 903                 cfs_spin_unlock(&med->med_open_lock);
 904
 905                 if (mfd != NULL) {
 906                         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 907                         /*set repbody->ea_size for resent case*/
 908                         if (ma->ma_valid & MA_LOV) {
 909                                 LASSERT(ma->ma_lmm_size != 0);
 910                                 repbody->eadatasize = ma->ma_lmm_size;
 911                                 if (isdir)
 912                                         repbody->valid |= OBD_MD_FLDIREA;
 913                                 else
 914                                         repbody->valid |= OBD_MD_FLEASIZE;
 915                         }
 916                         RETURN(0);
 917                 }
 918         }
 919
 920         rc = mdt_mfd_open(info, p, o, flags, created);
 921         RETURN(rc);
 922 }
 923
 924 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
 925                              struct lsd_client_data *lcd);
 926
 927 void mdt_reconstruct_open(struct mdt_thread_info *info,
 928                           struct mdt_lock_handle *lhc)
 929 {
 930         const struct lu_env *env = info->mti_env;
 931         struct mdt_device       *mdt  = info->mti_mdt;
 932         struct req_capsule      *pill = info->mti_pill;
 933         struct ptlrpc_request   *req  = mdt_info_req(info);
 934         struct tg_export_data   *ted  = &req->rq_export->exp_target_data;
 935         struct lsd_client_data  *lcd  = ted->ted_lcd;
 936         struct md_attr          *ma   = &info->mti_attr;
 937         struct mdt_reint_record *rr   = &info->mti_rr;
 938         __u32                   flags = info->mti_spec.sp_cr_flags;
 939         struct ldlm_reply       *ldlm_rep;
 940         struct mdt_object       *parent;
 941         struct mdt_object       *child;
 942         struct mdt_body         *repbody;
 943         int                      rc;
 944         ENTRY;
 945
 946         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 947         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 948         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 949
 950         ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
 951         ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
 952                                                RCL_SERVER);
 953         ma->ma_need = MA_INODE;
 954         if (ma->ma_lmm_size > 0)
 955                 ma->ma_need |= MA_LOV;
 956
 957         ma->ma_valid = 0;
 958
 959         mdt_req_from_lcd(req, lcd);
 960         mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
 961
 962         CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
 963                ldlm_rep->lock_policy_res1, req->rq_status);
 964
 965         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
 966             req->rq_status != 0)
 967                 /* We did not create successfully, return error to client. */
 968                 GOTO(out, rc = req->rq_status);
 969
 970         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
 971                 struct obd_export *exp = req->rq_export;
 972                 /*
 973                  * We failed after creation, but we do not know in which step
 974                  * we failed. So try to check the child object.
 975                  */
 976                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
 977                 if (IS_ERR(parent)) {
 978                         rc = PTR_ERR(parent);
 979                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
 980                                       " Evicting client %s with export %s.\n",
 981                                       PFID(rr->rr_fid1), rc,
 982                                       obd_uuid2str(&exp->exp_client_uuid),
 983                                       obd_export_nid2str(exp));
 984                         mdt_export_evict(exp);
 985                         RETURN_EXIT;
 986                 }
 987                 child = mdt_object_find(env, mdt, rr->rr_fid2);
 988                 if (IS_ERR(child)) {
 989                         rc = PTR_ERR(child);
 990                         LCONSOLE_WARN("Child "DFID" lookup error %d."
 991                                       " Evicting client %s with export %s.\n",
 992                                       PFID(mdt_object_fid(child)), rc,
 993                                       obd_uuid2str(&exp->exp_client_uuid),
 994                                       obd_export_nid2str(exp));
 995                         mdt_object_put(env, parent);
 996                         mdt_export_evict(exp);
 997                         RETURN_EXIT;
 998                 }
 999                 rc = mdt_object_exists(child);
1000                 if (rc > 0) {
1001                         struct md_object *next;
1002
1003                         mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA);
1004                         next = mdt_object_child(child);
1005                         rc = mo_attr_get(env, next, ma);
1006                         if (rc == 0)
1007                               rc = mdt_finish_open(info, parent, child,
1008                                                    flags, 1, ldlm_rep);
1009                 } else if (rc < 0) {
1010                         /* the child object was created on remote server */
1011                         repbody->fid1 = *rr->rr_fid2;
1012                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1013                         rc = 0;
1014                 } else if (rc == 0) {
1015                         /* the child does not exist, we should do regular open */
1016                         mdt_object_put(env, parent);
1017                         mdt_object_put(env, child);
1018                         GOTO(regular_open, 0);
1019                 }
1020                 mdt_object_put(env, parent);
1021                 mdt_object_put(env, child);
1022                 GOTO(out, rc);
1023         } else {
1024 regular_open:
1025                 /* We did not try to create, so we are a pure open */
1026                 rc = mdt_reint_open(info, lhc);
1027         }
1028
1029         EXIT;
1030 out:
1031         req->rq_status = rc;
1032         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1033         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1034 }
1035
1036 int mdt_open_by_fid(struct mdt_thread_info* info,
1037                     struct ldlm_reply *rep)
1038 {
1039         const struct lu_env     *env = info->mti_env;
1040         __u32                    flags = info->mti_spec.sp_cr_flags;
1041         struct mdt_reint_record *rr = &info->mti_rr;
1042         struct md_attr          *ma = &info->mti_attr;
1043         struct mdt_object       *o;
1044         int                      rc;
1045         ENTRY;
1046
1047         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1048         if (IS_ERR(o))
1049                 RETURN(rc = PTR_ERR(o));
1050
1051         rc = mdt_object_exists(o);
1052         if (rc > 0) {
1053                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1054                                                 DISP_LOOKUP_EXECD |
1055                                                 DISP_LOOKUP_POS));
1056
1057                 rc = mo_attr_get(env, mdt_object_child(o), ma);
1058                 if (rc == 0)
1059                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1060         } else if (rc == 0) {
1061                 rc = -ENOENT;
1062         } else  {
1063                 /* the child object was created on remote server */
1064                 struct mdt_body *repbody;
1065                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1066                 repbody->fid1 = *rr->rr_fid2;
1067                 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1068                 rc = 0;
1069         }
1070
1071         mdt_object_put(info->mti_env, o);
1072         RETURN(rc);
1073 }
1074
1075 int mdt_open_anon_by_fid(struct mdt_thread_info *info,
1076                          struct ldlm_reply *rep,
1077                          struct mdt_lock_handle *lhc)
1078 {
1079         const struct lu_env     *env   = info->mti_env;
1080         struct mdt_device       *mdt   = info->mti_mdt;
1081         __u32                    flags = info->mti_spec.sp_cr_flags;
1082         struct mdt_reint_record *rr    = &info->mti_rr;
1083         struct md_attr          *ma    = &info->mti_attr;
1084         struct mdt_object       *parent= NULL;
1085         struct mdt_object       *o;
1086         int                      rc;
1087         ldlm_mode_t              lm;
1088         ENTRY;
1089
1090         if (md_should_create(flags)) {
1091                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1092                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1093                         if (IS_ERR(parent)) {
1094                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1095                                        " for anonymous created %ld, try to"
1096                                        " use server-side parent.\n",
1097                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1098                                 parent = NULL;
1099                         }
1100                 }
1101                 if (parent == NULL)
1102                         ma->ma_need |= MA_PFID;
1103         }
1104
1105         o = mdt_object_find(env, mdt, rr->rr_fid2);
1106         if (IS_ERR(o))
1107                 RETURN(rc = PTR_ERR(o));
1108
1109         rc = mdt_object_exists(o);
1110         if (rc == 0) {
1111                 mdt_set_disposition(info, rep, (DISP_LOOKUP_EXECD |
1112                                     DISP_LOOKUP_NEG));
1113                 GOTO(out, rc = -ENOENT);
1114         } else if (rc < 0) {
1115                 CERROR("NFS remote open shouldn't happen.\n");
1116                 GOTO(out, rc);
1117         }
1118         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1119                                         DISP_LOOKUP_EXECD |
1120                                         DISP_LOOKUP_POS));
1121
1122         if (flags & FMODE_WRITE)
1123                 lm = LCK_CW;
1124         else if (flags & MDS_FMODE_EXEC)
1125                 lm = LCK_PR;
1126         else
1127                 lm = LCK_CR;
1128
1129         mdt_lock_handle_init(lhc);
1130         mdt_lock_reg_init(lhc, lm);
1131         rc = mdt_object_lock(info, o, lhc,
1132                              MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1133                              MDT_CROSS_LOCK);
1134         if (rc)
1135                 GOTO(out, rc);
1136
1137         rc = mo_attr_get(env, mdt_object_child(o), ma);
1138         if (rc)
1139                 GOTO(out, rc);
1140
1141         if (ma->ma_valid & MA_PFID) {
1142                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1143                 if (IS_ERR(parent)) {
1144                         CDEBUG(D_INODE, "Fail to find parent "DFID
1145                                " for anonymous created %ld, try to"
1146                                " use system default.\n",
1147                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1148                         parent = NULL;
1149                 }
1150         }
1151
1152         if (flags & MDS_OPEN_LOCK)
1153                 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1154         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1155
1156         if (!(flags & MDS_OPEN_LOCK) || rc)
1157                 mdt_object_unlock(info, o, lhc, 1);
1158
1159         GOTO(out, rc);
1160 out:
1161         mdt_object_put(env, o);
1162         if (parent != NULL)
1163                 mdt_object_put(env, parent);
1164         return rc;
1165 }
1166
1167 int mdt_pin(struct mdt_thread_info* info)
1168 {
1169         ENTRY;
1170         RETURN(err_serious(-EOPNOTSUPP));
1171 }
1172
1173 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1174 int mdt_cross_open(struct mdt_thread_info* info,
1175                    const struct lu_fid *fid,
1176                    struct ldlm_reply *rep, __u32 flags)
1177 {
1178         struct md_attr    *ma = &info->mti_attr;
1179         struct mdt_object *o;
1180         int                rc;
1181         ENTRY;
1182
1183         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1184         if (IS_ERR(o))
1185                 RETURN(rc = PTR_ERR(o));
1186
1187         rc = mdt_object_exists(o);
1188         if (rc > 0) {
1189                 /* Do permission check for cross-open. */
1190                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o),
1191                                    NULL, flags | MDS_OPEN_CROSS);
1192                 if (rc)
1193                         goto out;
1194
1195                 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1196                 rc = mo_attr_get(info->mti_env, mdt_object_child(o), ma);
1197                 if (rc == 0)
1198                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1199         } else if (rc == 0) {
1200                 /*
1201                  * Something is wrong here. lookup was positive but there is
1202                  * no object!
1203                  */
1204                 CERROR("Cross-ref object doesn't exist!\n");
1205                 rc = -EFAULT;
1206         } else  {
1207                 /* Something is wrong here, the object is on another MDS! */
1208                 CERROR("The object isn't on this server! FLD error?\n");
1209                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1210                                 &o->mot_obj.mo_lu,
1211                                 "Object isn't on this server! FLD error?\n");
1212
1213                 rc = -EFAULT;
1214         }
1215
1216 out:
1217         mdt_object_put(info->mti_env, o);
1218         RETURN(rc);
1219 }
1220
1221 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1222 {
1223         struct mdt_device       *mdt = info->mti_mdt;
1224         struct ptlrpc_request   *req = mdt_info_req(info);
1225         struct mdt_object       *parent;
1226         struct mdt_object       *child;
1227         struct mdt_lock_handle  *lh;
1228         struct ldlm_reply       *ldlm_rep;
1229         struct mdt_body         *repbody;
1230         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1231         struct md_attr          *ma = &info->mti_attr;
1232         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1233         struct mdt_reint_record *rr = &info->mti_rr;
1234         struct lu_name          *lname;
1235         int                      result, rc;
1236         int                      created = 0;
1237         __u32                    msg_flags;
1238         ENTRY;
1239
1240         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1241                                (obd_timeout + 1) / 4);
1242
1243         mdt_counter_incr(req->rq_export, LPROC_MDT_OPEN);
1244         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1245
1246         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1247         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1248                                                RCL_SERVER);
1249         ma->ma_need = MA_INODE;
1250         if (ma->ma_lmm_size > 0)
1251                 ma->ma_need |= MA_LOV;
1252
1253         ma->ma_valid = 0;
1254
1255         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1256         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1257
1258         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1259                 CERROR("file join is not supported anymore.\n");
1260                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1261         }
1262         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1263
1264         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1265             info->mti_spec.u.sp_ea.eadata == NULL)
1266                 GOTO(out, result = err_serious(-EINVAL));
1267
1268         CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1269                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1270                PFID(rr->rr_fid1), rr->rr_name,
1271                PFID(rr->rr_fid2), create_flags,
1272                ma->ma_attr.la_mode, msg_flags);
1273
1274         if (req_is_replay(req) ||
1275             (req->rq_export->exp_libclient && create_flags&MDS_OPEN_HAS_EA)) {
1276                 /* This is a replay request or from liblustre with ea. */
1277                 result = mdt_open_by_fid(info, ldlm_rep);
1278
1279                 if (result != -ENOENT) {
1280                         if (req->rq_export->exp_libclient &&
1281                             create_flags & MDS_OPEN_HAS_EA)
1282                                 GOTO(out, result = 0);
1283                         GOTO(out, result);
1284                 }
1285                 /*
1286                  * We didn't find the correct object, so we need to re-create it
1287                  * via a regular replay.
1288                  */
1289                 if (!(create_flags & MDS_OPEN_CREAT)) {
1290                         DEBUG_REQ(D_ERROR, req,
1291                                   "OPEN & CREAT not in open replay.");
1292                         GOTO(out, result = -EFAULT);
1293                 }
1294                 CDEBUG(D_INFO, "Open replay did find object, continue as "
1295                        "regular open\n");
1296         } else if (rr->rr_namelen == 0 && !info->mti_cross_ref &&
1297                    create_flags & MDS_OPEN_LOCK) {
1298                 result = mdt_open_anon_by_fid(info, ldlm_rep, lhc);
1299                 GOTO(out, result);
1300         }
1301
1302         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1303                 GOTO(out, result = err_serious(-ENOMEM));
1304
1305         mdt_set_disposition(info, ldlm_rep,
1306                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1307
1308         if (info->mti_cross_ref) {
1309                 /* This is cross-ref open */
1310                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1311                 result = mdt_cross_open(info, rr->rr_fid1, ldlm_rep,
1312                                         create_flags);
1313                 GOTO(out, result);
1314         }
1315
1316         lh = &info->mti_lh[MDT_LH_PARENT];
1317         mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1318                           LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1319
1320         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1321                                       MDS_INODELOCK_UPDATE);
1322         if (IS_ERR(parent))
1323                 GOTO(out, result = PTR_ERR(parent));
1324
1325         /* get and check version of parent */
1326         result = mdt_version_get_check(info, parent, 0);
1327         if (result)
1328                 GOTO(out_parent, result);
1329
1330         fid_zero(child_fid);
1331
1332         lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1333         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1334                             lname, child_fid, &info->mti_spec);
1335         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1336                  "looking for "DFID"/%s, result fid="DFID"\n",
1337                  PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1338
1339         if (result != 0 && result != -ENOENT && result != -ESTALE)
1340                 GOTO(out_parent, result);
1341
1342         if (result == -ENOENT || result == -ESTALE) {
1343                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1344                 if (result == -ESTALE) {
1345                         /*
1346                          * -ESTALE means the parent is a dead(unlinked) dir, so
1347                          * it should return -ENOENT to in accordance with the
1348                          * original mds implementaion.
1349                          */
1350                         GOTO(out_parent, result = -ENOENT);
1351                 }
1352                 if (!(create_flags & MDS_OPEN_CREAT))
1353                         GOTO(out_parent, result);
1354                 *child_fid = *info->mti_rr.rr_fid2;
1355                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1356                          PFID(child_fid));
1357         } else {
1358                 /*
1359                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1360                  * return FID back in that case.
1361                  */
1362                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1363         }
1364
1365         child = mdt_object_find(info->mti_env, mdt, child_fid);
1366         if (IS_ERR(child))
1367                 GOTO(out_parent, result = PTR_ERR(child));
1368
1369         /** check version of child  */
1370         rc = mdt_version_get_check(info, child, 1);
1371         if (rc)
1372                 GOTO(out_child, result = rc);
1373
1374         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1375         if (result == -ENOENT) {
1376                 if (mdt_object_obf(parent))
1377                         GOTO(out_child, result = -EPERM);
1378
1379                 /* save versions in reply */
1380                 mdt_version_get_save(info, parent, 0);
1381                 mdt_version_get_save(info, child, 1);
1382
1383                 /* version of child will be changed */
1384                 info->mti_mos = child;
1385
1386                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1387                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1388
1389                 /* Let lower layers know what is lock mode on directory. */
1390                 info->mti_spec.sp_cr_mode =
1391                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1392
1393                 /*
1394                  * Do not perform lookup sanity check. We know that name does
1395                  * not exist.
1396                  */
1397                 info->mti_spec.sp_cr_lookup = 0;
1398                 info->mti_spec.sp_feat = &dt_directory_features;
1399
1400                 result = mdo_create(info->mti_env,
1401                                     mdt_object_child(parent),
1402                                     lname,
1403                                     mdt_object_child(child),
1404                                     &info->mti_spec,
1405                                     &info->mti_attr);
1406                 if (result == -ERESTART) {
1407                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1408                         GOTO(out_child, result);
1409                 } else {
1410                         if (result != 0)
1411                                 GOTO(out_child, result);
1412                 }
1413                 created = 1;
1414         } else {
1415                 /* We have to get attr & lov ea for this object */
1416                 result = mo_attr_get(info->mti_env, mdt_object_child(child),
1417                                      ma);
1418                 /*
1419                  * The object is on remote node, return its FID for remote open.
1420                  */
1421                 if (result == -EREMOTE) {
1422                         /*
1423                          * Check if this lock already was sent to client and
1424                          * this is resent case. For resent case do not take lock
1425                          * again, use what is already granted.
1426                          */
1427                         LASSERT(lhc != NULL);
1428
1429                         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1430                                 struct ldlm_lock *lock;
1431
1432                                 LASSERT(msg_flags & MSG_RESENT);
1433
1434                                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1435                                 if (!lock) {
1436                                         CERROR("Invalid lock handle "LPX64"\n",
1437                                                lhc->mlh_reg_lh.cookie);
1438                                         LBUG();
1439                                 }
1440                                 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1441                                                         &lock->l_resource->lr_name));
1442                                 LDLM_LOCK_PUT(lock);
1443                                 rc = 0;
1444                         } else {
1445                                 mdt_lock_handle_init(lhc);
1446                                 mdt_lock_reg_init(lhc, LCK_PR);
1447
1448                                 rc = mdt_object_lock(info, child, lhc,
1449                                                      MDS_INODELOCK_LOOKUP,
1450                                                      MDT_CROSS_LOCK);
1451                         }
1452                         repbody->fid1 = *mdt_object_fid(child);
1453                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1454                         if (rc != 0)
1455                                 result = rc;
1456                         GOTO(out_child, result);
1457                 }
1458         }
1459
1460         LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1461
1462         /* get openlock if this is not replay and if a client requested it */
1463         if (!req_is_replay(req) && create_flags & MDS_OPEN_LOCK) {
1464                 ldlm_mode_t lm;
1465
1466                 if (create_flags & FMODE_WRITE)
1467                         lm = LCK_CW;
1468                 else if (create_flags & MDS_FMODE_EXEC)
1469                         lm = LCK_PR;
1470                 else
1471                         lm = LCK_CR;
1472                 mdt_lock_handle_init(lhc);
1473                 mdt_lock_reg_init(lhc, lm);
1474                 rc = mdt_object_lock(info, child, lhc,
1475                                      MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1476                                      MDT_CROSS_LOCK);
1477                 if (rc) {
1478                         result = rc;
1479                         GOTO(out_child, result);
1480                 } else {
1481                         result = -EREMOTE;
1482                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1483                 }
1484         }
1485
1486         /* Try to open it now. */
1487         rc = mdt_finish_open(info, parent, child, create_flags,
1488                              created, ldlm_rep);
1489         if (rc) {
1490                 result = rc;
1491                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1492                         /* openlock was acquired and mdt_finish_open failed -
1493                            drop the openlock */
1494                         mdt_object_unlock(info, child, lhc, 1);
1495                 if (created) {
1496                         ma->ma_need = 0;
1497                         ma->ma_valid = 0;
1498                         ma->ma_cookie_size = 0;
1499                         info->mti_no_need_trans = 1;
1500                         rc = mdo_unlink(info->mti_env,
1501                                         mdt_object_child(parent),
1502                                         mdt_object_child(child),
1503                                         lname,
1504                                         &info->mti_attr);
1505                         if (rc != 0)
1506                                 CERROR("Error in cleanup of open\n");
1507                 }
1508         }
1509         EXIT;
1510 out_child:
1511         mdt_object_put(info->mti_env, child);
1512 out_parent:
1513         mdt_object_unlock_put(info, parent, lh, result || !created);
1514 out:
1515         if (result && result != -EREMOTE)
1516                 lustre_msg_set_transno(req->rq_repmsg, 0);
1517         return result;
1518 }
1519
1520 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
1521                                       MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
1522
1523 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1524 {
1525         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1526 }
1527
1528 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1529 {
1530         struct mdt_object *o = mfd->mfd_object;
1531         struct md_object *next = mdt_object_child(o);
1532         struct md_attr *ma = &info->mti_attr;
1533         int ret = MDT_IOEPOCH_CLOSED;
1534         int rc = 0;
1535         int mode;
1536         ENTRY;
1537
1538         mode = mfd->mfd_mode;
1539
1540         if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
1541                 mdt_write_put(o);
1542                 ret = mdt_ioepoch_close(info, o);
1543         } else if (mode & MDS_FMODE_EXEC) {
1544                 mdt_write_allow(o);
1545         } else if (mode & MDS_FMODE_EPOCH) {
1546                 ret = mdt_ioepoch_close(info, o);
1547         } else if (mode & MDS_FMODE_SOM) {
1548                 ret = mdt_som_au_close(info, o);
1549         }
1550
1551         /* Update atime on close only. */
1552         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1553             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1554                 /* Set the atime only. */
1555                 ma->ma_valid = MA_INODE;
1556                 ma->ma_attr.la_valid = LA_ATIME;
1557                 rc = mo_attr_set(info->mti_env, next, ma);
1558         }
1559
1560         ma->ma_need |= MA_INODE;
1561         ma->ma_valid &= ~MA_INODE;
1562
1563         if (!MFD_CLOSED(mode))
1564                 rc = mo_close(info->mti_env, next, ma, mode);
1565
1566         if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
1567                 struct mdt_export_data *med;
1568
1569                 /* The IOepoch is still opened or SOM update is needed.
1570                  * Put mfd back into the list. */
1571                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
1572                 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
1573                                       MDS_FMODE_EPOCH : MDS_FMODE_SOM);
1574
1575                 LASSERT(mdt_info_req(info));
1576                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
1577                 cfs_spin_lock(&med->med_open_lock);
1578                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
1579                 class_handle_hash_back(&mfd->mfd_handle);
1580                 cfs_spin_unlock(&med->med_open_lock);
1581
1582                 if (ret == MDT_IOEPOCH_OPENED) {
1583                         ret = 0;
1584                 } else {
1585                         ret = -EAGAIN;
1586                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
1587                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
1588                 }
1589         } else {
1590                 mdt_mfd_free(mfd);
1591                 mdt_object_put(info->mti_env, o);
1592         }
1593
1594         RETURN(rc ? rc : ret);
1595 }
1596
1597 int mdt_close(struct mdt_thread_info *info)
1598 {
1599         struct mdt_export_data *med;
1600         struct mdt_file_data   *mfd;
1601         struct mdt_object      *o;
1602         struct md_attr         *ma = &info->mti_attr;
1603         struct mdt_body        *repbody = NULL;
1604         struct ptlrpc_request  *req = mdt_info_req(info);
1605         int rc, ret = 0;
1606         ENTRY;
1607
1608         mdt_counter_incr(req->rq_export, LPROC_MDT_CLOSE);
1609         /* Close may come with the Size-on-MDS update. Unpack it. */
1610         rc = mdt_close_unpack(info);
1611         if (rc)
1612                 RETURN(err_serious(rc));
1613
1614         LASSERT(info->mti_ioepoch);
1615
1616         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
1617                              info->mti_mdt->mdt_max_mdsize);
1618         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
1619                              info->mti_mdt->mdt_max_cookiesize);
1620         rc = req_capsule_server_pack(info->mti_pill);
1621         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
1622                 mdt_client_compatibility(info);
1623                 if (rc == 0)
1624                         mdt_fix_reply(info);
1625                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1626         }
1627
1628         /* Continue to close handle even if we can not pack reply */
1629         if (rc == 0) {
1630                 repbody = req_capsule_server_get(info->mti_pill,
1631                                                  &RMF_MDT_BODY);
1632                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
1633                                                     &RMF_MDT_MD);
1634                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
1635                                                        &RMF_MDT_MD,
1636                                                        RCL_SERVER);
1637                 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
1638                                                        &RMF_LOGCOOKIES);
1639                 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
1640                                                           &RMF_LOGCOOKIES,
1641                                                           RCL_SERVER);
1642                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
1643                 repbody->eadatasize = 0;
1644                 repbody->aclsize = 0;
1645         } else {
1646                 rc = err_serious(rc);
1647         }
1648
1649         med = &req->rq_export->exp_mdt_data;
1650         cfs_spin_lock(&med->med_open_lock);
1651         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1652         if (mdt_mfd_closed(mfd)) {
1653                 cfs_spin_unlock(&med->med_open_lock);
1654                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
1655                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
1656                        info->mti_ioepoch->handle.cookie);
1657                 /** not serious error since bug 3633 */
1658                 rc = -ESTALE;
1659         } else {
1660                 class_handle_unhash(&mfd->mfd_handle);
1661                 cfs_list_del_init(&mfd->mfd_list);
1662                 cfs_spin_unlock(&med->med_open_lock);
1663
1664                 /* Do not lose object before last unlink. */
1665                 o = mfd->mfd_object;
1666                 mdt_object_get(info->mti_env, o);
1667                 ret = mdt_mfd_close(info, mfd);
1668                 if (repbody != NULL)
1669                         rc = mdt_handle_last_unlink(info, o, ma);
1670                 mdt_empty_transno(info, rc);
1671                 mdt_object_put(info->mti_env, o);
1672         }
1673         if (repbody != NULL) {
1674                 mdt_client_compatibility(info);
1675                 rc = mdt_fix_reply(info);
1676         }
1677
1678         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
1679                 RETURN(err_serious(-ENOMEM));
1680
1681         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
1682                                  OBD_FAIL_MDS_CLOSE_NET_REP))
1683                 info->mti_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
1684         RETURN(rc ? rc : ret);
1685 }
1686
1687 /**
1688  * DONE_WRITING rpc handler.
1689  *
1690  * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
1691  * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
1692  * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
1693  * skip attributes and reconstruct the reply here.
1694  */
1695 int mdt_done_writing(struct mdt_thread_info *info)
1696 {
1697         struct ptlrpc_request   *req = mdt_info_req(info);
1698         struct mdt_body         *repbody = NULL;
1699         struct mdt_export_data  *med;
1700         struct mdt_file_data    *mfd;
1701         int rc;
1702         ENTRY;
1703
1704         rc = req_capsule_server_pack(info->mti_pill);
1705         if (rc)
1706                 RETURN(err_serious(rc));
1707
1708         repbody = req_capsule_server_get(info->mti_pill,
1709                                          &RMF_MDT_BODY);
1710         repbody->eadatasize = 0;
1711         repbody->aclsize = 0;
1712
1713         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
1714         rc = mdt_close_unpack(info);
1715         if (rc)
1716                 RETURN(err_serious(rc));
1717
1718         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL))
1719                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1720
1721         med = &info->mti_exp->exp_mdt_data;
1722         cfs_spin_lock(&med->med_open_lock);
1723         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1724         if (mfd == NULL) {
1725                 cfs_spin_unlock(&med->med_open_lock);
1726                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
1727                        ": cookie = "LPX64" ioepoch = "LPU64"\n",
1728                        PFID(info->mti_rr.rr_fid1),
1729                        info->mti_ioepoch->handle.cookie,
1730                        info->mti_ioepoch->ioepoch);
1731                 /* If this is a replay, reconstruct the transno. */
1732                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1733                         rc = info->mti_ioepoch->flags & MF_SOM_AU ?
1734                              -EAGAIN : 0;
1735                         mdt_empty_transno(info, rc);
1736                         RETURN(rc);
1737                 }
1738                 RETURN(-ESTALE);
1739         }
1740
1741         LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
1742                 mfd->mfd_mode == MDS_FMODE_TRUNC);
1743         class_handle_unhash(&mfd->mfd_handle);
1744         cfs_list_del_init(&mfd->mfd_list);
1745         cfs_spin_unlock(&med->med_open_lock);
1746
1747         /* Set EPOCH CLOSE flag if not set by client. */
1748         info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
1749         info->mti_attr.ma_valid = 0;
1750
1751         info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
1752         OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1753         if (info->mti_attr.ma_lmm == NULL)
1754                 RETURN(-ENOMEM);
1755
1756         rc = mdt_mfd_close(info, mfd);
1757
1758         OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1759         mdt_empty_transno(info, rc);
1760         RETURN(rc);
1761 }