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 static struct portals_handle_ops mfd_handle_ops = {
  58         .hop_addref = mdt_mfd_get,
  59         .hop_free   = NULL,
  60 };
  61
  62 /* Create a new mdt_file_data struct, initialize it,
  63  * and insert it to global hash table */
  64 struct mdt_file_data *mdt_mfd_new(void)
  65 {
  66         struct mdt_file_data *mfd;
  67         ENTRY;
  68
  69         OBD_ALLOC_PTR(mfd);
  70         if (mfd != NULL) {
  71                 CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  72                 CFS_INIT_LIST_HEAD(&mfd->mfd_list);
  73                 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
  74         }
  75         RETURN(mfd);
  76 }
  77
  78 /*
  79  * Find the mfd pointed to by handle in global hash table.
  80  * In case of replay the handle is obsoleted
  81  * but mfd can be found in mfd list by that handle
  82  */
  83 struct mdt_file_data *mdt_handle2mfd(struct mdt_thread_info *info,
  84                                      const struct lustre_handle *handle)
  85 {
  86         struct ptlrpc_request *req = mdt_info_req(info);
  87         struct mdt_file_data  *mfd;
  88         ENTRY;
  89
  90         LASSERT(handle != NULL);
  91         mfd = class_handle2object(handle->cookie);
  92         /* during dw/setattr replay the mfd can be found by old handle */
  93         if (mfd == NULL && req_is_replay(req)) {
  94                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
  95                 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  96                         if (mfd->mfd_old_handle.cookie == handle->cookie)
  97                                 RETURN (mfd);
  98                 }
  99                 mfd = NULL;
 100         }
 101         RETURN (mfd);
 102 }
 103
 104 /* free mfd */
 105 void mdt_mfd_free(struct mdt_file_data *mfd)
 106 {
 107         LASSERT(cfs_list_empty(&mfd->mfd_list));
 108         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 109 }
 110
 111 static int mdt_create_data(struct mdt_thread_info *info,
 112                            struct mdt_object *p, struct mdt_object *o)
 113 {
 114         struct md_op_spec     *spec = &info->mti_spec;
 115         struct md_attr        *ma   = &info->mti_attr;
 116         int                    rc   = 0;
 117         ENTRY;
 118
 119         if (!md_should_create(spec->sp_cr_flags))
 120                 RETURN(0);
 121
 122         ma->ma_need = MA_INODE | MA_LOV;
 123         ma->ma_valid = 0;
 124         cfs_mutex_lock(&o->mot_lov_mutex);
 125         if (!(o->mot_flags & MOF_LOV_CREATED)) {
 126                 rc = mdo_create_data(info->mti_env,
 127                                      p ? mdt_object_child(p) : NULL,
 128                                      mdt_object_child(o), spec, ma);
 129                 if (rc == 0 && ma->ma_valid & MA_LOV)
 130                         o->mot_flags |= MOF_LOV_CREATED;
 131         }
 132         cfs_mutex_unlock(&o->mot_lov_mutex);
 133         RETURN(rc);
 134 }
 135
 136 static int mdt_ioepoch_opened(struct mdt_object *mo)
 137 {
 138         return mo->mot_ioepoch_count;
 139 }
 140
 141 int mdt_object_is_som_enabled(struct mdt_object *mo)
 142 {
 143         return !mo->mot_ioepoch;
 144 }
 145
 146 /**
 147  * Re-enable Size-on-MDS.
 148  * Call under ->mot_ioepoch_mutex.
 149  */
 150 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
 151 {
 152         if (ioepoch == mo->mot_ioepoch) {
 153                 LASSERT(!mdt_ioepoch_opened(mo));
 154                 mo->mot_ioepoch = 0;
 155                 mo->mot_flags = 0;
 156         }
 157 }
 158
 159 /**
 160  * Open the IOEpoch. It is allowed if @writecount is not negative.
 161  * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
 162  */
 163 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
 164                      int created)
 165 {
 166         struct mdt_device *mdt = info->mti_mdt;
 167         int cancel = 0;
 168         int rc = 0;
 169         ENTRY;
 170
 171         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 172             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
 173                 RETURN(0);
 174
 175         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 176         if (mdt_ioepoch_opened(o)) {
 177                 /* Epoch continues even if there is no writers yet. */
 178                 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
 179                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 180         } else {
 181                 /* XXX: ->mdt_ioepoch is not initialized at the mount */
 182                 cfs_spin_lock(&mdt->mdt_ioepoch_lock);
 183                 if (mdt->mdt_ioepoch < info->mti_replayepoch)
 184                         mdt->mdt_ioepoch = info->mti_replayepoch;
 185
 186                 if (info->mti_replayepoch)
 187                         o->mot_ioepoch = info->mti_replayepoch;
 188                 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
 189                         o->mot_ioepoch = ++mdt->mdt_ioepoch;
 190                 else
 191                         o->mot_ioepoch = mdt->mdt_ioepoch;
 192
 193                 cfs_spin_unlock(&mdt->mdt_ioepoch_lock);
 194
 195                 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
 196                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 197                 if (created)
 198                         o->mot_flags |= MOF_SOM_CREATED;
 199                 cancel = 1;
 200         }
 201         o->mot_ioepoch_count++;
 202         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 203
 204         /* Cancel Size-on-MDS attributes cached on clients for the open case.
 205          * In the truncate case, see mdt_reint_setattr(). */
 206         if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
 207                 struct mdt_lock_handle  *lh = &info->mti_lh[MDT_LH_CHILD];
 208                 mdt_lock_reg_init(lh, LCK_EX);
 209                 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
 210                                      MDT_LOCAL_LOCK);
 211                 if (rc == 0)
 212                         mdt_object_unlock(info, o, lh, 1);
 213         }
 214         RETURN(rc);
 215 }
 216
 217 /**
 218  * Update SOM on-disk attributes.
 219  * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
 220  * mountid and attributes. If disabling, zero IOEpoch id in lustre-ea.
 221  * Call under ->mot_ioepoch_mutex.
 222  */
 223 static int mdt_som_attr_set(struct mdt_thread_info *info,
 224                             struct mdt_object *obj, __u64 ioepoch, int enable)
 225 {
 226         struct md_attr *ma = &info->mti_attr;
 227         int rc;
 228         ENTRY;
 229
 230         CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
 231                " on "DFID".\n", enable ? "update" : "disabling",
 232                ioepoch, PFID(mdt_object_fid(obj)));
 233
 234         ma->ma_valid |= MA_SOM;
 235         ma->ma_som = &info->mti_u.som.data;
 236         if (enable) {
 237                 struct mdt_device *mdt = info->mti_mdt;
 238                 struct lu_attr *la = &ma->ma_attr;
 239
 240                 ma->ma_som->msd_ioepoch = ioepoch;
 241                 ma->ma_som->msd_size = la->la_valid & LA_SIZE ? la->la_size : 0;
 242                 ma->ma_som->msd_blocks = la->la_valid & LA_BLOCKS ?
 243                                          la->la_blocks : 0;
 244                 ma->ma_som->msd_mountid = mdt->mdt_lut.lut_obd->u.obt.obt_mount_count;
 245                 ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
 246         } else {
 247                 ma->ma_som->msd_ioepoch = IOEPOCH_INVAL;
 248                 ma->ma_attr.la_valid &= LA_ATIME;
 249         }
 250
 251         /* Since we have opened the file, it is unnecessary
 252          * to check permission when close it. Between the "open"
 253          * and "close", maybe someone has changed the file mode
 254          * or flags, or the file created mode do not permit wirte,
 255          * and so on. Just set MDS_PERM_BYPASS for all the cases. */
 256         ma->ma_attr_flags |= MDS_PERM_BYPASS | MDS_SOM;
 257
 258         rc = mdt_attr_set(info, obj, ma, 0);
 259         RETURN(rc);
 260 }
 261
 262 /** Perform the eviction specific actions on ioepoch close. */
 263 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
 264                                                 struct mdt_object *o)
 265 {
 266         int rc = 0;
 267
 268         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 269         CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
 270                "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
 271                o->mot_ioepoch_count);
 272         o->mot_ioepoch_count--;
 273
 274         /* If eviction occured set MOF_SOM_RECOV,
 275          * if no other epoch holders, disable SOM on disk. */
 276         o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
 277         if (!mdt_ioepoch_opened(o)) {
 278                 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
 279                 mdt_object_som_enable(o, o->mot_ioepoch);
 280         }
 281         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 282         RETURN(rc);
 283 }
 284
 285 /**
 286  * Perform the replay specific actions on ioepoch close.
 287  * Skip SOM attribute update if obtained and just forget about the inode state
 288  * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
 289  */
 290 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
 291                                               struct mdt_object *o)
 292 {
 293         int rc = MDT_IOEPOCH_CLOSED;
 294         ENTRY;
 295
 296         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 297         CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
 298                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
 299         o->mot_ioepoch_count--;
 300
 301         /* Get an info from the replayed request if client is supposed
 302          * to send an Attibute Update, reconstruct @rc if so */
 303         if (info->mti_ioepoch->flags & MF_SOM_AU)
 304                 rc = MDT_IOEPOCH_GETATTR;
 305
 306         if (!mdt_ioepoch_opened(o))
 307                 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
 308         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 309
 310         RETURN(rc);
 311 }
 312
 313 /**
 314  * Regular file IOepoch close.
 315  * Closes the ioepoch, checks the object state, apply obtained attributes and
 316  * re-enable SOM on the object, if possible. Also checks if the recovery is
 317  * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
 318  * to obtain SOM attributes under the server-side OST locks.
 319  *
 320  * Return value:
 321  * MDT_IOEPOCH_CLOSED if ioepoch is closed.
 322  * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
 323  */
 324 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
 325                                         struct mdt_object *o)
 326 {
 327         struct md_attr *tmp_ma;
 328         struct lu_attr *la;
 329         int achange, opened;
 330         int recovery = 0;
 331         int rc = 0, ret = MDT_IOEPOCH_CLOSED;
 332         ENTRY;
 333
 334         la = &info->mti_attr.ma_attr;
 335         achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
 336
 337         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 338         o->mot_ioepoch_count--;
 339
 340         tmp_ma = &info->mti_u.som.attr;
 341         tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
 342         tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
 343         tmp_ma->ma_som = &info->mti_u.som.data;
 344         tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
 345         tmp_ma->ma_valid = 0;
 346         rc = mo_attr_get(info->mti_env, mdt_object_child(o), tmp_ma);
 347         if (rc)
 348                 GOTO(error_up, rc);
 349
 350         /* Check the on-disk SOM state. */
 351         if (o->mot_flags & MOF_SOM_RECOV)
 352                 recovery = 1;
 353         else if (!(o->mot_flags & MOF_SOM_CREATED) &&
 354                  !(tmp_ma->ma_valid & MA_SOM))
 355                 recovery = 1;
 356
 357         CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
 358                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
 359
 360         opened = mdt_ioepoch_opened(o);
 361         /**
 362          * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
 363          * Skip the check for file with no LOV  or for unlink files.
 364          */
 365         if (!opened && tmp_ma->ma_valid & MA_LOV &&
 366             !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
 367                 if (recovery)
 368                         /* If some previous writer was evicted, re-ask the
 369                          * client for attributes. Even if attributes are
 370                          * provided, we cannot believe in them.
 371                          * Another use case is that there is no SOM cache on
 372                          * disk -- first access with SOM or there was an MDS
 373                          * failure. */
 374                         ret = MDT_IOEPOCH_GETATTR;
 375                 else if (o->mot_flags & MOF_SOM_CHANGE)
 376                         /* Some previous writer changed the attribute.
 377                          * Do not believe to the current Size-on-MDS
 378                          * update, re-ask client. */
 379                         ret = MDT_IOEPOCH_GETATTR;
 380                 else if (!(la->la_valid & LA_SIZE) && achange)
 381                         /* Attributes were changed by the last writer
 382                          * only but no Size-on-MDS update is received.*/
 383                         ret = MDT_IOEPOCH_GETATTR;
 384         }
 385
 386         if (achange || ret == MDT_IOEPOCH_GETATTR)
 387                 o->mot_flags |= MOF_SOM_CHANGE;
 388
 389         /* If epoch ends and relable SOM attributes are obtained, update them.
 390          * Create SOM ea for new files even if there is no attributes obtained
 391          * (0-length file). */
 392         if (ret == MDT_IOEPOCH_CLOSED && !opened) {
 393                 if (achange || o->mot_flags & MOF_SOM_CREATED) {
 394                         LASSERT(achange || !(la->la_valid & LA_SIZE));
 395                         rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
 396                                               MDT_SOM_ENABLE);
 397                         /* Avoid the following setattrs of these attributes,
 398                          * e.g. for atime update. */
 399                         info->mti_attr.ma_valid = 0;
 400                 }
 401                 mdt_object_som_enable(o, o->mot_ioepoch);
 402         }
 403
 404         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 405         /* If recovery is needed, tell the client to perform GETATTR under
 406          * the lock. */
 407         if (ret == MDT_IOEPOCH_GETATTR && recovery) {
 408                 struct mdt_body *rep;
 409                 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 410                 rep->valid |= OBD_MD_FLGETATTRLOCK;
 411         }
 412
 413         RETURN(rc ? : ret);
 414
 415 error_up:
 416         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 417         return rc;
 418 }
 419
 420 /**
 421  * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
 422  * - a client closes the IOEpoch;
 423  * - a client eviction occured.
 424  * Return values:
 425  * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
 426  * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
 427  * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
 428  * update is needed.
 429  */
 430 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
 431 {
 432         struct ptlrpc_request *req = mdt_info_req(info);
 433         ENTRY;
 434
 435         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 436             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
 437                 RETURN(0);
 438
 439         LASSERT(o->mot_ioepoch_count);
 440         LASSERT(info->mti_ioepoch == NULL ||
 441                 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
 442
 443         /* IOEpoch is closed only if client tells about it or eviction occures.
 444          * In the replay case, always close the epoch. */
 445         if (req == NULL)
 446                 RETURN(mdt_ioepoch_close_on_eviction(info, o));
 447         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
 448                 RETURN(mdt_ioepoch_close_on_replay(info, o));
 449         if (info->mti_ioepoch->flags & MF_EPOCH_CLOSE)
 450                 RETURN(mdt_ioepoch_close_reg(info, o));
 451         /* IO epoch is not closed. */
 452         RETURN(MDT_IOEPOCH_OPENED);
 453 }
 454
 455 /**
 456  * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
 457  * for attribute update. It happens if:
 458  * - SOM Attribute Update is obtained;
 459  * - the client failed to obtain it and informs MDS about it;
 460  * - a client eviction occured.
 461  * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
 462  * cache otherwise.
 463  */
 464 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
 465 {
 466         struct ptlrpc_request *req = mdt_info_req(info);
 467         __u64 ioepoch = 0;
 468         int act = MDT_SOM_ENABLE;
 469         int rc = 0;
 470         ENTRY;
 471
 472         LASSERT(!req || info->mti_ioepoch);
 473         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 474             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
 475                 RETURN(0);
 476
 477         /* No size whereas MF_SOM_CHANGE is set means client failed to
 478          * obtain ost attributes, drop the SOM cache on disk if so. */
 479         if (!req ||
 480             (info->mti_ioepoch &&
 481              info->mti_ioepoch->flags & MF_SOM_CHANGE &&
 482              !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
 483                 act = MDT_SOM_DISABLE;
 484
 485         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 486         /* Mark the object it is the recovery state if we failed to obtain
 487          * SOM attributes. */
 488         if (act == MDT_SOM_DISABLE)
 489                 o->mot_flags |= MOF_SOM_RECOV;
 490
 491         if (!mdt_ioepoch_opened(o)) {
 492                 ioepoch =  info->mti_ioepoch ?
 493                         info->mti_ioepoch->ioepoch : o->mot_ioepoch;
 494
 495                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 496                         rc = mdt_som_attr_set(info, o, ioepoch, act);
 497                 mdt_object_som_enable(o, ioepoch);
 498         }
 499         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 500         RETURN(rc);
 501 }
 502
 503 int mdt_write_read(struct mdt_object *o)
 504 {
 505         int rc = 0;
 506         ENTRY;
 507         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 508         rc = o->mot_writecount;
 509         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 510         RETURN(rc);
 511 }
 512
 513 int mdt_write_get(struct mdt_object *o)
 514 {
 515         int rc = 0;
 516         ENTRY;
 517         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 518         if (o->mot_writecount < 0)
 519                 rc = -ETXTBSY;
 520         else
 521                 o->mot_writecount++;
 522         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 523         RETURN(rc);
 524 }
 525
 526 void mdt_write_put(struct mdt_object *o)
 527 {
 528         ENTRY;
 529         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 530         o->mot_writecount--;
 531         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 532         EXIT;
 533 }
 534
 535 static int mdt_write_deny(struct mdt_object *o)
 536 {
 537         int rc = 0;
 538         ENTRY;
 539         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 540         if (o->mot_writecount > 0)
 541                 rc = -ETXTBSY;
 542         else
 543                 o->mot_writecount--;
 544         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 545         RETURN(rc);
 546 }
 547
 548 static void mdt_write_allow(struct mdt_object *o)
 549 {
 550         ENTRY;
 551         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 552         o->mot_writecount++;
 553         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 554         EXIT;
 555 }
 556
 557 /* there can be no real transaction so prepare the fake one */
 558 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 559 {
 560         struct mdt_device      *mdt = info->mti_mdt;
 561         struct ptlrpc_request  *req = mdt_info_req(info);
 562         struct tg_export_data  *ted;
 563         struct lsd_client_data *lcd;
 564
 565         ENTRY;
 566         /* transaction has occurred already */
 567         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 568                 RETURN_EXIT;
 569
 570         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 571         if (info->mti_transno == 0) {
 572                 info->mti_transno = ++ mdt->mdt_lut.lut_last_transno;
 573         } else {
 574                 /* should be replay */
 575                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 576                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 577         }
 578         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 579
 580         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 581                         info->mti_transno,
 582                         req->rq_export->exp_obd->obd_last_committed);
 583
 584         req->rq_transno = info->mti_transno;
 585         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 586
 587         /* update lcd in memory only for resent cases */
 588         ted = &req->rq_export->exp_target_data;
 589         LASSERT(ted);
 590         cfs_mutex_lock(&ted->ted_lcd_lock);
 591         lcd = ted->ted_lcd;
 592         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 593             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 594                 if (info->mti_transno != 0)
 595                         lcd->lcd_last_close_transno = info->mti_transno;
 596                 lcd->lcd_last_close_xid = req->rq_xid;
 597                 lcd->lcd_last_close_result = rc;
 598         } else {
 599                 /* VBR: save versions in last_rcvd for reconstruct. */
 600                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 601                 if (pre_versions) {
 602                         lcd->lcd_pre_versions[0] = pre_versions[0];
 603                         lcd->lcd_pre_versions[1] = pre_versions[1];
 604                         lcd->lcd_pre_versions[2] = pre_versions[2];
 605                         lcd->lcd_pre_versions[3] = pre_versions[3];
 606                 }
 607                 if (info->mti_transno != 0)
 608                         lcd->lcd_last_transno = info->mti_transno;
 609                 lcd->lcd_last_xid = req->rq_xid;
 610                 lcd->lcd_last_result = rc;
 611                 lcd->lcd_last_data = info->mti_opdata;
 612         }
 613         cfs_mutex_unlock(&ted->ted_lcd_lock);
 614
 615         EXIT;
 616 }
 617
 618 void mdt_mfd_set_mode(struct mdt_file_data *mfd, int mode)
 619 {
 620         LASSERT(mfd != NULL);
 621
 622         CDEBUG(D_HA, "Change mfd %p mode 0x%x->0x%x\n",
 623                mfd, (unsigned int)mfd->mfd_mode, (unsigned int)mode);
 624
 625         mfd->mfd_mode = mode;
 626 }
 627
 628 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 629                         struct mdt_object *o, __u64 flags, int created)
 630 {
 631         struct ptlrpc_request   *req = mdt_info_req(info);
 632         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 633         struct mdt_file_data    *mfd;
 634         struct md_attr          *ma  = &info->mti_attr;
 635         struct lu_attr          *la  = &ma->ma_attr;
 636         struct mdt_body         *repbody;
 637         int                      rc = 0, isdir, isreg;
 638         ENTRY;
 639
 640         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 641
 642         isreg = S_ISREG(la->la_mode);
 643         isdir = S_ISDIR(la->la_mode);
 644         if (isreg && !(ma->ma_valid & MA_LOV)) {
 645                 /*
 646                  * No EA, check whether it is will set regEA and dirEA since in
 647                  * above attr get, these size might be zero, so reset it, to
 648                  * retrieve the MD after create obj.
 649                  */
 650                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 651                                                        &RMF_MDT_MD,
 652                                                        RCL_SERVER);
 653                 /* in replay case, p == NULL */
 654                 rc = mdt_create_data(info, p, o);
 655                 if (rc)
 656                         RETURN(rc);
 657         }
 658
 659         CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
 660                ma->ma_valid, ma->ma_lmm_size);
 661
 662         if (ma->ma_valid & MA_LOV) {
 663                 LASSERT(ma->ma_lmm_size != 0);
 664                 repbody->eadatasize = ma->ma_lmm_size;
 665                 if (isdir)
 666                         repbody->valid |= OBD_MD_FLDIREA;
 667                 else
 668                         repbody->valid |= OBD_MD_FLEASIZE;
 669         }
 670
 671         if (flags & FMODE_WRITE) {
 672                 rc = mdt_write_get(o);
 673                 if (rc == 0) {
 674                         mdt_ioepoch_open(info, o, created);
 675                         repbody->ioepoch = o->mot_ioepoch;
 676                 }
 677         } else if (flags & MDS_FMODE_EXEC) {
 678                 rc = mdt_write_deny(o);
 679         }
 680         if (rc)
 681                 RETURN(rc);
 682
 683         rc = mo_open(info->mti_env, mdt_object_child(o),
 684                      created ? flags | MDS_OPEN_CREATED : flags);
 685         if (rc)
 686                 GOTO(err_out, rc);
 687
 688         mfd = mdt_mfd_new();
 689         if (mfd != NULL) {
 690                 /*
 691                  * Keep a reference on this object for this open, and is
 692                  * released by mdt_mfd_close().
 693                  */
 694                 mdt_object_get(info->mti_env, o);
 695
 696                 /*
 697                  * @flags is always not zero. At least it should be FMODE_READ,
 698                  * FMODE_WRITE or MDS_FMODE_EXEC.
 699                  */
 700                 LASSERT(flags != 0);
 701
 702                 /* Open handling. */
 703                 mdt_mfd_set_mode(mfd, flags);
 704
 705                 mfd->mfd_object = o;
 706                 mfd->mfd_xid = req->rq_xid;
 707
 708                 /* replay handle */
 709                 if (req_is_replay(req)) {
 710                         struct mdt_file_data *old_mfd;
 711                         /* Check wheather old cookie already exist in
 712                          * the list, becasue when do recovery, client
 713                          * might be disconnected from server, and
 714                          * restart replay, so there maybe some orphan
 715                          * mfd here, we should remove them */
 716                         LASSERT(info->mti_rr.rr_handle != NULL);
 717                         old_mfd = mdt_handle2mfd(info, info->mti_rr.rr_handle);
 718                         if (old_mfd) {
 719                                 CDEBUG(D_HA, "del orph mfd %p fid=("DFID") "
 720                                        "cookie=" LPX64"\n", mfd,
 721                                        PFID(mdt_object_fid(mfd->mfd_object)),
 722                                        info->mti_rr.rr_handle->cookie);
 723                                 cfs_spin_lock(&med->med_open_lock);
 724                                 class_handle_unhash(&old_mfd->mfd_handle);
 725                                 cfs_list_del_init(&old_mfd->mfd_list);
 726                                 cfs_spin_unlock(&med->med_open_lock);
 727                                 /* no attr update for that close */
 728                                 la->la_valid = 0;
 729                                 ma->ma_valid |= MA_FLAGS;
 730                                 ma->ma_attr_flags |= MDS_RECOV_OPEN;
 731                                 mdt_mfd_close(info, old_mfd);
 732                                 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 733                                 ma->ma_valid &= ~MA_FLAGS;
 734                         }
 735                         CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
 736                                info->mti_rr.rr_handle->cookie);
 737                         mfd->mfd_old_handle.cookie =
 738                                                 info->mti_rr.rr_handle->cookie;
 739                 }
 740                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 741
 742                 if (req->rq_export->exp_disconnected) {
 743                         cfs_spin_lock(&med->med_open_lock);
 744                         class_handle_unhash(&mfd->mfd_handle);
 745                         cfs_list_del_init(&mfd->mfd_list);
 746                         cfs_spin_unlock(&med->med_open_lock);
 747                         mdt_mfd_close(info, mfd);
 748                 } else {
 749                         cfs_spin_lock(&med->med_open_lock);
 750                         cfs_list_add(&mfd->mfd_list, &med->med_open_head);
 751                         cfs_spin_unlock(&med->med_open_lock);
 752                 }
 753
 754                 mdt_empty_transno(info, rc);
 755         } else {
 756                 GOTO(err_out, rc = -ENOMEM);
 757         }
 758
 759         RETURN(rc);
 760
 761 err_out:
 762         if (flags & FMODE_WRITE)
 763                         /* XXX We also need to close io epoch here.
 764                          * See LU-1220 - green */
 765                 mdt_write_put(o);
 766         else if (flags & FMODE_EXEC)
 767                 mdt_write_allow(o);
 768         return rc;
 769 }
 770
 771 int mdt_finish_open(struct mdt_thread_info *info,
 772                     struct mdt_object *p, struct mdt_object *o,
 773                     __u64 flags, int created, struct ldlm_reply *rep)
 774 {
 775         struct ptlrpc_request   *req = mdt_info_req(info);
 776         struct obd_export       *exp = req->rq_export;
 777         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 778         struct md_attr          *ma  = &info->mti_attr;
 779         struct lu_attr          *la  = &ma->ma_attr;
 780         struct mdt_file_data    *mfd;
 781         struct mdt_body         *repbody;
 782         int                      rc = 0;
 783         int                      isreg, isdir, islnk;
 784         cfs_list_t              *t;
 785         ENTRY;
 786
 787         LASSERT(ma->ma_valid & MA_INODE);
 788
 789         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 790
 791         isreg = S_ISREG(la->la_mode);
 792         isdir = S_ISDIR(la->la_mode);
 793         islnk = S_ISLNK(la->la_mode);
 794         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 795
 796         if (exp_connect_rmtclient(exp)) {
 797                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 798
 799                 rc = mdt_pack_remote_perm(info, o, buf);
 800                 if (rc) {
 801                         repbody->valid &= ~OBD_MD_FLRMTPERM;
 802                         repbody->aclsize = 0;
 803                 } else {
 804                         repbody->valid |= OBD_MD_FLRMTPERM;
 805                         repbody->aclsize = sizeof(struct mdt_remote_perm);
 806                 }
 807         }
 808 #ifdef CONFIG_FS_POSIX_ACL
 809         else if (exp->exp_connect_flags & OBD_CONNECT_ACL) {
 810                 const struct lu_env *env = info->mti_env;
 811                 struct md_object *next = mdt_object_child(o);
 812                 struct lu_buf *buf = &info->mti_buf;
 813
 814                 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 815                 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
 816                                                    RCL_SERVER);
 817                 if (buf->lb_len > 0) {
 818                         rc = mo_xattr_get(env, next, buf,
 819                                           XATTR_NAME_ACL_ACCESS);
 820                         if (rc < 0) {
 821                                 if (rc == -ENODATA) {
 822                                         repbody->aclsize = 0;
 823                                         repbody->valid |= OBD_MD_FLACL;
 824                                         rc = 0;
 825                                 } else if (rc == -EOPNOTSUPP) {
 826                                         rc = 0;
 827                                 } else {
 828                                         CERROR("got acl size: %d\n", rc);
 829                                 }
 830                         } else {
 831                                 repbody->aclsize = rc;
 832                                 repbody->valid |= OBD_MD_FLACL;
 833                                 rc = 0;
 834                         }
 835                 }
 836         }
 837 #endif
 838
 839         if (info->mti_mdt->mdt_opts.mo_mds_capa &&
 840             exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) {
 841                 struct lustre_capa *capa;
 842
 843                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
 844                 LASSERT(capa);
 845                 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
 846                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 847                 if (rc)
 848                         RETURN(rc);
 849                 repbody->valid |= OBD_MD_FLMDSCAPA;
 850         }
 851         if (info->mti_mdt->mdt_opts.mo_oss_capa &&
 852             exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA &&
 853             S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) {
 854                 struct lustre_capa *capa;
 855
 856                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
 857                 LASSERT(capa);
 858                 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
 859                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 860                 if (rc)
 861                         RETURN(rc);
 862                 repbody->valid |= OBD_MD_FLOSSCAPA;
 863         }
 864
 865         /*
 866          * If we are following a symlink, don't open; and do not return open
 867          * handle for special nodes as client required.
 868          */
 869         if (islnk || (!isreg && !isdir &&
 870             (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH))) {
 871                 lustre_msg_set_transno(req->rq_repmsg, 0);
 872                 RETURN(0);
 873         }
 874
 875         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 876
 877         /*
 878          * We need to return the existing object's fid back, so it is done here,
 879          * after preparing the reply.
 880          */
 881         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 882                 RETURN(-EEXIST);
 883
 884         /* This can't be done earlier, we need to return reply body */
 885         if (isdir) {
 886                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 887                         /* We are trying to create or write an existing dir. */
 888                         RETURN(-EISDIR);
 889                 }
 890         } else if (flags & MDS_OPEN_DIRECTORY)
 891                 RETURN(-ENOTDIR);
 892
 893         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 894                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 895                 RETURN(-EAGAIN);
 896         }
 897
 898         mfd = NULL;
 899         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 900                 cfs_spin_lock(&med->med_open_lock);
 901                 cfs_list_for_each(t, &med->med_open_head) {
 902                         mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
 903                         if (mfd->mfd_xid == req->rq_xid) {
 904                                 break;
 905                         }
 906                         mfd = NULL;
 907                 }
 908                 cfs_spin_unlock(&med->med_open_lock);
 909
 910                 if (mfd != NULL) {
 911                         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 912                         /*set repbody->ea_size for resent case*/
 913                         if (ma->ma_valid & MA_LOV) {
 914                                 LASSERT(ma->ma_lmm_size != 0);
 915                                 repbody->eadatasize = ma->ma_lmm_size;
 916                                 if (isdir)
 917                                         repbody->valid |= OBD_MD_FLDIREA;
 918                                 else
 919                                         repbody->valid |= OBD_MD_FLEASIZE;
 920                         }
 921                         RETURN(0);
 922                 }
 923         }
 924
 925         rc = mdt_mfd_open(info, p, o, flags, created);
 926         RETURN(rc);
 927 }
 928
 929 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
 930                              struct lsd_client_data *lcd);
 931
 932 void mdt_reconstruct_open(struct mdt_thread_info *info,
 933                           struct mdt_lock_handle *lhc)
 934 {
 935         const struct lu_env *env = info->mti_env;
 936         struct mdt_device       *mdt  = info->mti_mdt;
 937         struct req_capsule      *pill = info->mti_pill;
 938         struct ptlrpc_request   *req  = mdt_info_req(info);
 939         struct tg_export_data   *ted  = &req->rq_export->exp_target_data;
 940         struct lsd_client_data  *lcd  = ted->ted_lcd;
 941         struct md_attr          *ma   = &info->mti_attr;
 942         struct mdt_reint_record *rr   = &info->mti_rr;
 943         __u32                   flags = info->mti_spec.sp_cr_flags;
 944         struct ldlm_reply       *ldlm_rep;
 945         struct mdt_object       *parent;
 946         struct mdt_object       *child;
 947         struct mdt_body         *repbody;
 948         int                      rc;
 949         ENTRY;
 950
 951         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 952         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 953         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 954
 955         ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
 956         ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
 957                                                RCL_SERVER);
 958         ma->ma_need = MA_INODE;
 959         if (ma->ma_lmm_size > 0)
 960                 ma->ma_need |= MA_LOV;
 961
 962         ma->ma_valid = 0;
 963
 964         mdt_req_from_lcd(req, lcd);
 965         mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
 966
 967         CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
 968                ldlm_rep->lock_policy_res1, req->rq_status);
 969
 970         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
 971             req->rq_status != 0)
 972                 /* We did not create successfully, return error to client. */
 973                 GOTO(out, rc = req->rq_status);
 974
 975         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
 976                 struct obd_export *exp = req->rq_export;
 977                 /*
 978                  * We failed after creation, but we do not know in which step
 979                  * we failed. So try to check the child object.
 980                  */
 981                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
 982                 if (IS_ERR(parent)) {
 983                         rc = PTR_ERR(parent);
 984                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
 985                                       " Evicting client %s with export %s.\n",
 986                                       PFID(rr->rr_fid1), rc,
 987                                       obd_uuid2str(&exp->exp_client_uuid),
 988                                       obd_export_nid2str(exp));
 989                         mdt_export_evict(exp);
 990                         RETURN_EXIT;
 991                 }
 992                 child = mdt_object_find(env, mdt, rr->rr_fid2);
 993                 if (IS_ERR(child)) {
 994                         rc = PTR_ERR(child);
 995                         LCONSOLE_WARN("Child "DFID" lookup error %d."
 996                                       " Evicting client %s with export %s.\n",
 997                                       PFID(mdt_object_fid(child)), rc,
 998                                       obd_uuid2str(&exp->exp_client_uuid),
 999                                       obd_export_nid2str(exp));
1000                         mdt_object_put(env, parent);
1001                         mdt_export_evict(exp);
1002                         RETURN_EXIT;
1003                 }
1004                 rc = mdt_object_exists(child);
1005                 if (rc > 0) {
1006                         struct md_object *next;
1007
1008                         mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA);
1009                         next = mdt_object_child(child);
1010                         rc = mo_attr_get(env, next, ma);
1011                         if (rc == 0)
1012                               rc = mdt_finish_open(info, parent, child,
1013                                                    flags, 1, ldlm_rep);
1014                 } else if (rc < 0) {
1015                         /* the child object was created on remote server */
1016                         repbody->fid1 = *rr->rr_fid2;
1017                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1018                         rc = 0;
1019                 } else if (rc == 0) {
1020                         /* the child does not exist, we should do regular open */
1021                         mdt_object_put(env, parent);
1022                         mdt_object_put(env, child);
1023                         GOTO(regular_open, 0);
1024                 }
1025                 mdt_object_put(env, parent);
1026                 mdt_object_put(env, child);
1027                 GOTO(out, rc);
1028         } else {
1029 regular_open:
1030                 /* We did not try to create, so we are a pure open */
1031                 rc = mdt_reint_open(info, lhc);
1032         }
1033
1034         EXIT;
1035 out:
1036         req->rq_status = rc;
1037         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1038         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1039 }
1040
1041 int mdt_open_by_fid(struct mdt_thread_info* info,
1042                     struct ldlm_reply *rep)
1043 {
1044         const struct lu_env     *env = info->mti_env;
1045         __u32                    flags = info->mti_spec.sp_cr_flags;
1046         struct mdt_reint_record *rr = &info->mti_rr;
1047         struct md_attr          *ma = &info->mti_attr;
1048         struct mdt_object       *o;
1049         int                      rc;
1050         ENTRY;
1051
1052         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1053         if (IS_ERR(o))
1054                 RETURN(rc = PTR_ERR(o));
1055
1056         rc = mdt_object_exists(o);
1057         if (rc > 0) {
1058                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1059                                                 DISP_LOOKUP_EXECD |
1060                                                 DISP_LOOKUP_POS));
1061
1062                 rc = mo_attr_get(env, mdt_object_child(o), ma);
1063                 if (rc == 0)
1064                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1065         } else if (rc == 0) {
1066                 rc = -ENOENT;
1067         } else  {
1068                 /* the child object was created on remote server */
1069                 struct mdt_body *repbody;
1070                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1071                 repbody->fid1 = *rr->rr_fid2;
1072                 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1073                 rc = 0;
1074         }
1075
1076         mdt_object_put(info->mti_env, o);
1077         RETURN(rc);
1078 }
1079
1080 int mdt_open_anon_by_fid(struct mdt_thread_info *info,
1081                          struct ldlm_reply *rep,
1082                          struct mdt_lock_handle *lhc)
1083 {
1084         const struct lu_env     *env   = info->mti_env;
1085         struct mdt_device       *mdt   = info->mti_mdt;
1086         __u32                    flags = info->mti_spec.sp_cr_flags;
1087         struct mdt_reint_record *rr    = &info->mti_rr;
1088         struct md_attr          *ma    = &info->mti_attr;
1089         struct mdt_object       *parent= NULL;
1090         struct mdt_object       *o;
1091         int                      rc;
1092         ldlm_mode_t              lm;
1093         ENTRY;
1094
1095         if (md_should_create(flags)) {
1096                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1097                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1098                         if (IS_ERR(parent)) {
1099                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1100                                        " for anonymous created %ld, try to"
1101                                        " use server-side parent.\n",
1102                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1103                                 parent = NULL;
1104                         }
1105                 }
1106                 if (parent == NULL)
1107                         ma->ma_need |= MA_PFID;
1108         }
1109
1110         o = mdt_object_find(env, mdt, rr->rr_fid2);
1111         if (IS_ERR(o))
1112                 RETURN(rc = PTR_ERR(o));
1113
1114         rc = mdt_object_exists(o);
1115         if (rc == 0) {
1116                 mdt_set_disposition(info, rep, (DISP_LOOKUP_EXECD |
1117                                     DISP_LOOKUP_NEG));
1118                 GOTO(out, rc = -ENOENT);
1119         } else if (rc < 0) {
1120                 CERROR("NFS remote open shouldn't happen.\n");
1121                 GOTO(out, rc);
1122         }
1123         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1124                                         DISP_LOOKUP_EXECD |
1125                                         DISP_LOOKUP_POS));
1126
1127         if (flags & FMODE_WRITE)
1128                 lm = LCK_CW;
1129         else if (flags & MDS_FMODE_EXEC)
1130                 lm = LCK_PR;
1131         else
1132                 lm = LCK_CR;
1133
1134         mdt_lock_handle_init(lhc);
1135         mdt_lock_reg_init(lhc, lm);
1136         rc = mdt_object_lock(info, o, lhc,
1137                              MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1138                              MDT_CROSS_LOCK);
1139         if (rc)
1140                 GOTO(out, rc);
1141
1142         rc = mo_attr_get(env, mdt_object_child(o), ma);
1143         if (rc)
1144                 GOTO(out, rc);
1145
1146         if (ma->ma_valid & MA_PFID) {
1147                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1148                 if (IS_ERR(parent)) {
1149                         CDEBUG(D_INODE, "Fail to find parent "DFID
1150                                " for anonymous created %ld, try to"
1151                                " use system default.\n",
1152                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1153                         parent = NULL;
1154                 }
1155         }
1156
1157         if (flags & MDS_OPEN_LOCK)
1158                 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1159         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1160
1161         if (!(flags & MDS_OPEN_LOCK) || rc)
1162                 mdt_object_unlock(info, o, lhc, 1);
1163
1164         GOTO(out, rc);
1165 out:
1166         mdt_object_put(env, o);
1167         if (parent != NULL)
1168                 mdt_object_put(env, parent);
1169         return rc;
1170 }
1171
1172 int mdt_pin(struct mdt_thread_info* info)
1173 {
1174         ENTRY;
1175         RETURN(err_serious(-EOPNOTSUPP));
1176 }
1177
1178 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1179 int mdt_cross_open(struct mdt_thread_info* info,
1180                    const struct lu_fid *fid,
1181                    struct ldlm_reply *rep, __u32 flags)
1182 {
1183         struct md_attr    *ma = &info->mti_attr;
1184         struct mdt_object *o;
1185         int                rc;
1186         ENTRY;
1187
1188         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1189         if (IS_ERR(o))
1190                 RETURN(rc = PTR_ERR(o));
1191
1192         rc = mdt_object_exists(o);
1193         if (rc > 0) {
1194                 /* Do permission check for cross-open. */
1195                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o),
1196                                    NULL, flags | MDS_OPEN_CROSS);
1197                 if (rc)
1198                         goto out;
1199
1200                 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1201                 rc = mo_attr_get(info->mti_env, mdt_object_child(o), ma);
1202                 if (rc == 0)
1203                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1204         } else if (rc == 0) {
1205                 /*
1206                  * Something is wrong here. lookup was positive but there is
1207                  * no object!
1208                  */
1209                 CERROR("Cross-ref object doesn't exist!\n");
1210                 rc = -EFAULT;
1211         } else  {
1212                 /* Something is wrong here, the object is on another MDS! */
1213                 CERROR("The object isn't on this server! FLD error?\n");
1214                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1215                                 &o->mot_obj.mo_lu,
1216                                 "Object isn't on this server! FLD error?\n");
1217
1218                 rc = -EFAULT;
1219         }
1220
1221 out:
1222         mdt_object_put(info->mti_env, o);
1223         RETURN(rc);
1224 }
1225
1226 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1227 {
1228         struct mdt_device       *mdt = info->mti_mdt;
1229         struct ptlrpc_request   *req = mdt_info_req(info);
1230         struct mdt_object       *parent;
1231         struct mdt_object       *child;
1232         struct mdt_lock_handle  *lh;
1233         struct ldlm_reply       *ldlm_rep;
1234         struct mdt_body         *repbody;
1235         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1236         struct md_attr          *ma = &info->mti_attr;
1237         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1238         struct mdt_reint_record *rr = &info->mti_rr;
1239         struct lu_name          *lname;
1240         int                      result, rc;
1241         int                      created = 0;
1242         __u32                    msg_flags;
1243         ENTRY;
1244
1245         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1246                                (obd_timeout + 1) / 4);
1247
1248         mdt_counter_incr(req->rq_export, LPROC_MDT_OPEN);
1249         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1250
1251         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1252         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1253                                                RCL_SERVER);
1254         ma->ma_need = MA_INODE;
1255         if (ma->ma_lmm_size > 0)
1256                 ma->ma_need |= MA_LOV;
1257
1258         ma->ma_valid = 0;
1259
1260         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1261         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1262
1263         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1264                 CERROR("file join is not supported anymore.\n");
1265                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1266         }
1267         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1268
1269         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1270             info->mti_spec.u.sp_ea.eadata == NULL)
1271                 GOTO(out, result = err_serious(-EINVAL));
1272
1273         CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1274                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1275                PFID(rr->rr_fid1), rr->rr_name,
1276                PFID(rr->rr_fid2), create_flags,
1277                ma->ma_attr.la_mode, msg_flags);
1278
1279         if (req_is_replay(req) ||
1280             (req->rq_export->exp_libclient && create_flags&MDS_OPEN_HAS_EA)) {
1281                 /* This is a replay request or from liblustre with ea. */
1282                 result = mdt_open_by_fid(info, ldlm_rep);
1283
1284                 if (result != -ENOENT) {
1285                         if (req->rq_export->exp_libclient &&
1286                             create_flags & MDS_OPEN_HAS_EA)
1287                                 GOTO(out, result = 0);
1288                         GOTO(out, result);
1289                 }
1290                 /*
1291                  * We didn't find the correct object, so we need to re-create it
1292                  * via a regular replay.
1293                  */
1294                 if (!(create_flags & MDS_OPEN_CREAT)) {
1295                         DEBUG_REQ(D_ERROR, req,
1296                                   "OPEN & CREAT not in open replay.");
1297                         GOTO(out, result = -EFAULT);
1298                 }
1299                 CDEBUG(D_INFO, "Open replay did find object, continue as "
1300                        "regular open\n");
1301         } else if (rr->rr_namelen == 0 && !info->mti_cross_ref &&
1302                    create_flags & MDS_OPEN_LOCK) {
1303                 result = mdt_open_anon_by_fid(info, ldlm_rep, lhc);
1304                 GOTO(out, result);
1305         }
1306
1307         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1308                 GOTO(out, result = err_serious(-ENOMEM));
1309
1310         mdt_set_disposition(info, ldlm_rep,
1311                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1312
1313         if (info->mti_cross_ref) {
1314                 /* This is cross-ref open */
1315                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1316                 result = mdt_cross_open(info, rr->rr_fid1, ldlm_rep,
1317                                         create_flags);
1318                 GOTO(out, result);
1319         }
1320
1321         lh = &info->mti_lh[MDT_LH_PARENT];
1322         mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1323                           LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1324
1325         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1326                                       MDS_INODELOCK_UPDATE);
1327         if (IS_ERR(parent))
1328                 GOTO(out, result = PTR_ERR(parent));
1329
1330         /* get and check version of parent */
1331         result = mdt_version_get_check(info, parent, 0);
1332         if (result)
1333                 GOTO(out_parent, result);
1334
1335         fid_zero(child_fid);
1336
1337         lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1338         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1339                             lname, child_fid, &info->mti_spec);
1340         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1341                  "looking for "DFID"/%s, result fid="DFID"\n",
1342                  PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1343
1344         if (result != 0 && result != -ENOENT && result != -ESTALE)
1345                 GOTO(out_parent, result);
1346
1347         if (result == -ENOENT || result == -ESTALE) {
1348                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1349                 if (result == -ESTALE) {
1350                         /*
1351                          * -ESTALE means the parent is a dead(unlinked) dir, so
1352                          * it should return -ENOENT to in accordance with the
1353                          * original mds implementaion.
1354                          */
1355                         GOTO(out_parent, result = -ENOENT);
1356                 }
1357                 if (!(create_flags & MDS_OPEN_CREAT))
1358                         GOTO(out_parent, result);
1359                 *child_fid = *info->mti_rr.rr_fid2;
1360                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1361                          PFID(child_fid));
1362         } else {
1363                 /*
1364                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1365                  * return FID back in that case.
1366                  */
1367                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1368         }
1369
1370         child = mdt_object_find(info->mti_env, mdt, child_fid);
1371         if (IS_ERR(child))
1372                 GOTO(out_parent, result = PTR_ERR(child));
1373
1374         /** check version of child  */
1375         rc = mdt_version_get_check(info, child, 1);
1376         if (rc)
1377                 GOTO(out_child, result = rc);
1378
1379         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1380         if (result == -ENOENT) {
1381                 if (mdt_object_obf(parent))
1382                         GOTO(out_child, result = -EPERM);
1383
1384                 /* save versions in reply */
1385                 mdt_version_get_save(info, parent, 0);
1386                 mdt_version_get_save(info, child, 1);
1387
1388                 /* version of child will be changed */
1389                 info->mti_mos = child;
1390
1391                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1392                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1393
1394                 /* Let lower layers know what is lock mode on directory. */
1395                 info->mti_spec.sp_cr_mode =
1396                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1397
1398                 /*
1399                  * Do not perform lookup sanity check. We know that name does
1400                  * not exist.
1401                  */
1402                 info->mti_spec.sp_cr_lookup = 0;
1403                 info->mti_spec.sp_feat = &dt_directory_features;
1404
1405                 result = mdo_create(info->mti_env,
1406                                     mdt_object_child(parent),
1407                                     lname,
1408                                     mdt_object_child(child),
1409                                     &info->mti_spec,
1410                                     &info->mti_attr);
1411                 if (result == -ERESTART) {
1412                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1413                         GOTO(out_child, result);
1414                 } else {
1415                         if (result != 0)
1416                                 GOTO(out_child, result);
1417                 }
1418                 created = 1;
1419         } else {
1420                 /* We have to get attr & lov ea for this object */
1421                 result = mo_attr_get(info->mti_env, mdt_object_child(child),
1422                                      ma);
1423                 /*
1424                  * The object is on remote node, return its FID for remote open.
1425                  */
1426                 if (result == -EREMOTE) {
1427                         /*
1428                          * Check if this lock already was sent to client and
1429                          * this is resent case. For resent case do not take lock
1430                          * again, use what is already granted.
1431                          */
1432                         LASSERT(lhc != NULL);
1433
1434                         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1435                                 struct ldlm_lock *lock;
1436
1437                                 LASSERT(msg_flags & MSG_RESENT);
1438
1439                                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1440                                 if (!lock) {
1441                                         CERROR("Invalid lock handle "LPX64"\n",
1442                                                lhc->mlh_reg_lh.cookie);
1443                                         LBUG();
1444                                 }
1445                                 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1446                                                         &lock->l_resource->lr_name));
1447                                 LDLM_LOCK_PUT(lock);
1448                                 rc = 0;
1449                         } else {
1450                                 mdt_lock_handle_init(lhc);
1451                                 mdt_lock_reg_init(lhc, LCK_PR);
1452
1453                                 rc = mdt_object_lock(info, child, lhc,
1454                                                      MDS_INODELOCK_LOOKUP,
1455                                                      MDT_CROSS_LOCK);
1456                         }
1457                         repbody->fid1 = *mdt_object_fid(child);
1458                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1459                         if (rc != 0)
1460                                 result = rc;
1461                         GOTO(out_child, result);
1462                 }
1463         }
1464
1465         LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1466
1467         /* get openlock if this is not replay and if a client requested it */
1468         if (!req_is_replay(req) && create_flags & MDS_OPEN_LOCK) {
1469                 ldlm_mode_t lm;
1470
1471                 if (create_flags & FMODE_WRITE)
1472                         lm = LCK_CW;
1473                 else if (create_flags & MDS_FMODE_EXEC)
1474                         lm = LCK_PR;
1475                 else
1476                         lm = LCK_CR;
1477                 mdt_lock_handle_init(lhc);
1478                 mdt_lock_reg_init(lhc, lm);
1479                 rc = mdt_object_lock(info, child, lhc,
1480                                      MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1481                                      MDT_CROSS_LOCK);
1482                 if (rc) {
1483                         result = rc;
1484                         GOTO(out_child, result);
1485                 } else {
1486                         result = -EREMOTE;
1487                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1488                 }
1489         }
1490
1491         /* Try to open it now. */
1492         rc = mdt_finish_open(info, parent, child, create_flags,
1493                              created, ldlm_rep);
1494         if (rc) {
1495                 result = rc;
1496                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1497                         /* openlock was acquired and mdt_finish_open failed -
1498                            drop the openlock */
1499                         mdt_object_unlock(info, child, lhc, 1);
1500                 if (created) {
1501                         ma->ma_need = 0;
1502                         ma->ma_valid = 0;
1503                         ma->ma_cookie_size = 0;
1504                         info->mti_no_need_trans = 1;
1505                         rc = mdo_unlink(info->mti_env,
1506                                         mdt_object_child(parent),
1507                                         mdt_object_child(child),
1508                                         lname,
1509                                         &info->mti_attr);
1510                         if (rc != 0)
1511                                 CERROR("Error in cleanup of open\n");
1512                 }
1513         }
1514         EXIT;
1515 out_child:
1516         mdt_object_put(info->mti_env, child);
1517 out_parent:
1518         mdt_object_unlock_put(info, parent, lh, result || !created);
1519 out:
1520         if (result && result != -EREMOTE)
1521                 lustre_msg_set_transno(req->rq_repmsg, 0);
1522         return result;
1523 }
1524
1525 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
1526                                       MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
1527
1528 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1529 {
1530         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1531 }
1532
1533 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1534 {
1535         struct mdt_object *o = mfd->mfd_object;
1536         struct md_object *next = mdt_object_child(o);
1537         struct md_attr *ma = &info->mti_attr;
1538         int ret = MDT_IOEPOCH_CLOSED;
1539         int rc = 0;
1540         int mode;
1541         ENTRY;
1542
1543         mode = mfd->mfd_mode;
1544
1545         if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
1546                 mdt_write_put(o);
1547                 ret = mdt_ioepoch_close(info, o);
1548         } else if (mode & MDS_FMODE_EXEC) {
1549                 mdt_write_allow(o);
1550         } else if (mode & MDS_FMODE_EPOCH) {
1551                 ret = mdt_ioepoch_close(info, o);
1552         } else if (mode & MDS_FMODE_SOM) {
1553                 ret = mdt_som_au_close(info, o);
1554         }
1555
1556         /* Update atime on close only. */
1557         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1558             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1559                 /* Set the atime only. */
1560                 ma->ma_valid = MA_INODE;
1561                 ma->ma_attr.la_valid = LA_ATIME;
1562                 rc = mo_attr_set(info->mti_env, next, ma);
1563         }
1564
1565         ma->ma_need |= MA_INODE;
1566         ma->ma_valid &= ~MA_INODE;
1567
1568         if (!MFD_CLOSED(mode))
1569                 rc = mo_close(info->mti_env, next, ma, mode);
1570
1571         if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
1572                 struct mdt_export_data *med;
1573
1574                 /* The IOepoch is still opened or SOM update is needed.
1575                  * Put mfd back into the list. */
1576                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
1577                 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
1578                                       MDS_FMODE_EPOCH : MDS_FMODE_SOM);
1579
1580                 LASSERT(mdt_info_req(info));
1581                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
1582                 cfs_spin_lock(&med->med_open_lock);
1583                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
1584                 class_handle_hash_back(&mfd->mfd_handle);
1585                 cfs_spin_unlock(&med->med_open_lock);
1586
1587                 if (ret == MDT_IOEPOCH_OPENED) {
1588                         ret = 0;
1589                 } else {
1590                         ret = -EAGAIN;
1591                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
1592                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
1593                 }
1594         } else {
1595                 mdt_mfd_free(mfd);
1596                 mdt_object_put(info->mti_env, o);
1597         }
1598
1599         RETURN(rc ? rc : ret);
1600 }
1601
1602 int mdt_close(struct mdt_thread_info *info)
1603 {
1604         struct mdt_export_data *med;
1605         struct mdt_file_data   *mfd;
1606         struct mdt_object      *o;
1607         struct md_attr         *ma = &info->mti_attr;
1608         struct mdt_body        *repbody = NULL;
1609         struct ptlrpc_request  *req = mdt_info_req(info);
1610         int rc, ret = 0;
1611         ENTRY;
1612
1613         mdt_counter_incr(req->rq_export, LPROC_MDT_CLOSE);
1614         /* Close may come with the Size-on-MDS update. Unpack it. */
1615         rc = mdt_close_unpack(info);
1616         if (rc)
1617                 RETURN(err_serious(rc));
1618
1619         LASSERT(info->mti_ioepoch);
1620
1621         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
1622                              info->mti_mdt->mdt_max_mdsize);
1623         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
1624                              info->mti_mdt->mdt_max_cookiesize);
1625         rc = req_capsule_server_pack(info->mti_pill);
1626         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
1627                 mdt_client_compatibility(info);
1628                 if (rc == 0)
1629                         mdt_fix_reply(info);
1630                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1631         }
1632
1633         /* Continue to close handle even if we can not pack reply */
1634         if (rc == 0) {
1635                 repbody = req_capsule_server_get(info->mti_pill,
1636                                                  &RMF_MDT_BODY);
1637                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
1638                                                     &RMF_MDT_MD);
1639                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
1640                                                        &RMF_MDT_MD,
1641                                                        RCL_SERVER);
1642                 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
1643                                                        &RMF_LOGCOOKIES);
1644                 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
1645                                                           &RMF_LOGCOOKIES,
1646                                                           RCL_SERVER);
1647                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
1648                 repbody->eadatasize = 0;
1649                 repbody->aclsize = 0;
1650         } else {
1651                 rc = err_serious(rc);
1652         }
1653
1654         med = &req->rq_export->exp_mdt_data;
1655         cfs_spin_lock(&med->med_open_lock);
1656         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1657         if (mdt_mfd_closed(mfd)) {
1658                 cfs_spin_unlock(&med->med_open_lock);
1659                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
1660                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
1661                        info->mti_ioepoch->handle.cookie);
1662                 /** not serious error since bug 3633 */
1663                 rc = -ESTALE;
1664         } else {
1665                 class_handle_unhash(&mfd->mfd_handle);
1666                 cfs_list_del_init(&mfd->mfd_list);
1667                 cfs_spin_unlock(&med->med_open_lock);
1668
1669                 /* Do not lose object before last unlink. */
1670                 o = mfd->mfd_object;
1671                 mdt_object_get(info->mti_env, o);
1672                 ret = mdt_mfd_close(info, mfd);
1673                 if (repbody != NULL)
1674                         rc = mdt_handle_last_unlink(info, o, ma);
1675                 mdt_empty_transno(info, rc);
1676                 mdt_object_put(info->mti_env, o);
1677         }
1678         if (repbody != NULL) {
1679                 mdt_client_compatibility(info);
1680                 rc = mdt_fix_reply(info);
1681         }
1682
1683         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
1684                 RETURN(err_serious(-ENOMEM));
1685
1686         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
1687                                  OBD_FAIL_MDS_CLOSE_NET_REP))
1688                 info->mti_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
1689         RETURN(rc ? rc : ret);
1690 }
1691
1692 /**
1693  * DONE_WRITING rpc handler.
1694  *
1695  * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
1696  * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
1697  * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
1698  * skip attributes and reconstruct the reply here.
1699  */
1700 int mdt_done_writing(struct mdt_thread_info *info)
1701 {
1702         struct ptlrpc_request   *req = mdt_info_req(info);
1703         struct mdt_body         *repbody = NULL;
1704         struct mdt_export_data  *med;
1705         struct mdt_file_data    *mfd;
1706         int rc;
1707         ENTRY;
1708
1709         rc = req_capsule_server_pack(info->mti_pill);
1710         if (rc)
1711                 RETURN(err_serious(rc));
1712
1713         repbody = req_capsule_server_get(info->mti_pill,
1714                                          &RMF_MDT_BODY);
1715         repbody->eadatasize = 0;
1716         repbody->aclsize = 0;
1717
1718         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
1719         rc = mdt_close_unpack(info);
1720         if (rc)
1721                 RETURN(err_serious(rc));
1722
1723         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL))
1724                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1725
1726         med = &info->mti_exp->exp_mdt_data;
1727         cfs_spin_lock(&med->med_open_lock);
1728         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1729         if (mfd == NULL) {
1730                 cfs_spin_unlock(&med->med_open_lock);
1731                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
1732                        ": cookie = "LPX64" ioepoch = "LPU64"\n",
1733                        PFID(info->mti_rr.rr_fid1),
1734                        info->mti_ioepoch->handle.cookie,
1735                        info->mti_ioepoch->ioepoch);
1736                 /* If this is a replay, reconstruct the transno. */
1737                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1738                         rc = info->mti_ioepoch->flags & MF_SOM_AU ?
1739                              -EAGAIN : 0;
1740                         mdt_empty_transno(info, rc);
1741                         RETURN(rc);
1742                 }
1743                 RETURN(-ESTALE);
1744         }
1745
1746         LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
1747                 mfd->mfd_mode == MDS_FMODE_TRUNC);
1748         class_handle_unhash(&mfd->mfd_handle);
1749         cfs_list_del_init(&mfd->mfd_list);
1750         cfs_spin_unlock(&med->med_open_lock);
1751
1752         /* Set EPOCH CLOSE flag if not set by client. */
1753         info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
1754         info->mti_attr.ma_valid = 0;
1755
1756         info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
1757         OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1758         if (info->mti_attr.ma_lmm == NULL)
1759                 RETURN(-ENOMEM);
1760
1761         rc = mdt_mfd_close(info, mfd);
1762
1763         OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1764         mdt_empty_transno(info, rc);
1765         RETURN(rc);
1766 }