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 #define DEBUG_SUBSYSTEM S_MDS
  44
  45 #include <lustre_acl.h>
  46 #include <lustre_mds.h>
  47 #include "mdt_internal.h"
  48
  49 /* we do nothing because we do not have refcount now */
  50 static void mdt_mfd_get(void *mfdp)
  51 {
  52 }
  53
  54 static struct portals_handle_ops mfd_handle_ops = {
  55         .hop_addref = mdt_mfd_get,
  56         .hop_free   = NULL,
  57 };
  58
  59 /* Create a new mdt_file_data struct, initialize it,
  60  * and insert it to global hash table */
  61 struct mdt_file_data *mdt_mfd_new(void)
  62 {
  63         struct mdt_file_data *mfd;
  64         ENTRY;
  65
  66         OBD_ALLOC_PTR(mfd);
  67         if (mfd != NULL) {
  68                 CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  69                 CFS_INIT_LIST_HEAD(&mfd->mfd_list);
  70                 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
  71         }
  72         RETURN(mfd);
  73 }
  74
  75 /*
  76  * Find the mfd pointed to by handle in global hash table.
  77  * In case of replay the handle is obsoleted
  78  * but mfd can be found in mfd list by that handle
  79  */
  80 struct mdt_file_data *mdt_handle2mfd(struct mdt_thread_info *info,
  81                                      const struct lustre_handle *handle)
  82 {
  83         struct ptlrpc_request *req = mdt_info_req(info);
  84         struct mdt_file_data  *mfd;
  85         ENTRY;
  86
  87         LASSERT(handle != NULL);
  88         mfd = class_handle2object(handle->cookie);
  89         /* during dw/setattr replay the mfd can be found by old handle */
  90         if (mfd == NULL && req_is_replay(req)) {
  91                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
  92                 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  93                         if (mfd->mfd_old_handle.cookie == handle->cookie)
  94                                 RETURN (mfd);
  95                 }
  96                 mfd = NULL;
  97         }
  98         RETURN (mfd);
  99 }
 100
 101 /* free mfd */
 102 void mdt_mfd_free(struct mdt_file_data *mfd)
 103 {
 104         LASSERT(cfs_list_empty(&mfd->mfd_list));
 105         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 106 }
 107
 108 static int mdt_create_data(struct mdt_thread_info *info,
 109                            struct mdt_object *p, struct mdt_object *o)
 110 {
 111         struct md_op_spec     *spec = &info->mti_spec;
 112         struct md_attr        *ma   = &info->mti_attr;
 113         int                    rc   = 0;
 114         ENTRY;
 115
 116         if (!md_should_create(spec->sp_cr_flags))
 117                 RETURN(0);
 118
 119         ma->ma_need = MA_INODE | MA_LOV;
 120         ma->ma_valid = 0;
 121         cfs_mutex_lock(&o->mot_lov_mutex);
 122         if (!(o->mot_flags & MOF_LOV_CREATED)) {
 123                 rc = mdo_create_data(info->mti_env,
 124                                      p ? mdt_object_child(p) : NULL,
 125                                      mdt_object_child(o), spec, ma);
 126                 if (rc == 0)
 127                         rc = mdt_attr_get_complex(info, o, ma);
 128
 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 = mdt_attr_get_complex(info, 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
1007                         mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA);
1008                         rc = mdt_attr_get_complex(info, child, ma);
1009                         if (rc == 0)
1010                               rc = mdt_finish_open(info, parent, child,
1011                                                    flags, 1, ldlm_rep);
1012                 } else if (rc < 0) {
1013                         /* the child object was created on remote server */
1014                         repbody->fid1 = *rr->rr_fid2;
1015                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1016                         rc = 0;
1017                 } else if (rc == 0) {
1018                         /* the child does not exist, we should do regular open */
1019                         mdt_object_put(env, parent);
1020                         mdt_object_put(env, child);
1021                         GOTO(regular_open, 0);
1022                 }
1023                 mdt_object_put(env, parent);
1024                 mdt_object_put(env, child);
1025                 GOTO(out, rc);
1026         } else {
1027 regular_open:
1028                 /* We did not try to create, so we are a pure open */
1029                 rc = mdt_reint_open(info, lhc);
1030         }
1031
1032         EXIT;
1033 out:
1034         req->rq_status = rc;
1035         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1036         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1037 }
1038
1039 int mdt_open_by_fid(struct mdt_thread_info* info,
1040                     struct ldlm_reply *rep)
1041 {
1042         __u32                    flags = info->mti_spec.sp_cr_flags;
1043         struct mdt_reint_record *rr = &info->mti_rr;
1044         struct md_attr          *ma = &info->mti_attr;
1045         struct mdt_object       *o;
1046         int                      rc;
1047         ENTRY;
1048
1049         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1050         if (IS_ERR(o))
1051                 RETURN(rc = PTR_ERR(o));
1052
1053         rc = mdt_object_exists(o);
1054         if (rc > 0) {
1055                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1056                                                 DISP_LOOKUP_EXECD |
1057                                                 DISP_LOOKUP_POS));
1058
1059                 rc = mdt_attr_get_complex(info, o, ma);
1060                 if (rc == 0)
1061                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1062         } else if (rc == 0) {
1063                 rc = -ENOENT;
1064         } else  {
1065                 /* the child object was created on remote server */
1066                 struct mdt_body *repbody;
1067                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1068                 repbody->fid1 = *rr->rr_fid2;
1069                 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1070                 rc = 0;
1071         }
1072
1073         mdt_object_put(info->mti_env, o);
1074         RETURN(rc);
1075 }
1076
1077 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1078                          struct mdt_lock_handle *lhc)
1079 {
1080         const struct lu_env     *env   = info->mti_env;
1081         struct mdt_device       *mdt   = info->mti_mdt;
1082         __u32                    flags = info->mti_spec.sp_cr_flags;
1083         struct mdt_reint_record *rr    = &info->mti_rr;
1084         struct md_attr          *ma    = &info->mti_attr;
1085         struct mdt_object       *parent= NULL;
1086         struct mdt_object       *o;
1087         int                      rc;
1088         ldlm_mode_t              lm;
1089         ENTRY;
1090
1091         if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1092                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1093                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1094                         if (IS_ERR(parent)) {
1095                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1096                                        " for anonymous created %ld, try to"
1097                                        " use server-side parent.\n",
1098                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1099                                 parent = NULL;
1100                         }
1101                 }
1102                 if (parent == NULL)
1103                         ma->ma_need |= MA_PFID;
1104         }
1105
1106         o = mdt_object_find(env, mdt, rr->rr_fid2);
1107         if (IS_ERR(o))
1108                 RETURN(rc = PTR_ERR(o));
1109
1110         rc = mdt_object_exists(o);
1111         if (rc == 0) {
1112                 mdt_set_disposition(info, rep, (DISP_LOOKUP_EXECD |
1113                                     DISP_LOOKUP_NEG));
1114                 GOTO(out, rc = -ENOENT);
1115         } else if (rc < 0) {
1116                 CERROR("NFS remote open shouldn't happen.\n");
1117                 GOTO(out, rc);
1118         }
1119         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1120                                         DISP_LOOKUP_EXECD |
1121                                         DISP_LOOKUP_POS));
1122
1123         if (flags & FMODE_WRITE)
1124                 lm = LCK_CW;
1125         else if (flags & MDS_FMODE_EXEC)
1126                 lm = LCK_PR;
1127         else
1128                 lm = LCK_CR;
1129
1130         mdt_lock_handle_init(lhc);
1131         mdt_lock_reg_init(lhc, lm);
1132         rc = mdt_object_lock(info, o, lhc,
1133                              MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1134                              MDT_CROSS_LOCK);
1135         if (rc)
1136                 GOTO(out, rc);
1137
1138         rc = mdt_attr_get_complex(info, o, ma);
1139         if (rc)
1140                 GOTO(out, rc);
1141
1142         if (ma->ma_valid & MA_PFID) {
1143                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1144                 if (IS_ERR(parent)) {
1145                         CDEBUG(D_INODE, "Fail to find parent "DFID
1146                                " for anonymous created %ld, try to"
1147                                " use system default.\n",
1148                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1149                         parent = NULL;
1150                 }
1151         }
1152
1153         if (flags & MDS_OPEN_LOCK)
1154                 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1155         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1156
1157         if (!(flags & MDS_OPEN_LOCK) || rc)
1158                 mdt_object_unlock(info, o, lhc, 1);
1159
1160         GOTO(out, rc);
1161 out:
1162         mdt_object_put(env, o);
1163         if (parent != NULL)
1164                 mdt_object_put(env, parent);
1165         return rc;
1166 }
1167
1168 int mdt_pin(struct mdt_thread_info* info)
1169 {
1170         ENTRY;
1171         RETURN(err_serious(-EOPNOTSUPP));
1172 }
1173
1174 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1175 int mdt_cross_open(struct mdt_thread_info* info,
1176                    const struct lu_fid *fid,
1177                    struct ldlm_reply *rep, __u32 flags)
1178 {
1179         struct md_attr    *ma = &info->mti_attr;
1180         struct mdt_object *o;
1181         int                rc;
1182         ENTRY;
1183
1184         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1185         if (IS_ERR(o))
1186                 RETURN(rc = PTR_ERR(o));
1187
1188         rc = mdt_object_exists(o);
1189         if (rc > 0) {
1190                 /* Do permission check for cross-open. */
1191                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o),
1192                                    NULL, flags | MDS_OPEN_CROSS);
1193                 if (rc)
1194                         goto out;
1195
1196                 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1197                 rc = mdt_attr_get_complex(info, o, ma);
1198                 if (rc == 0)
1199                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1200         } else if (rc == 0) {
1201                 /*
1202                  * Something is wrong here. lookup was positive but there is
1203                  * no object!
1204                  */
1205                 CERROR("Cross-ref object doesn't exist!\n");
1206                 rc = -EFAULT;
1207         } else  {
1208                 /* Something is wrong here, the object is on another MDS! */
1209                 CERROR("The object isn't on this server! FLD error?\n");
1210                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1211                                 &o->mot_obj.mo_lu,
1212                                 "Object isn't on this server! FLD error?\n");
1213
1214                 rc = -EFAULT;
1215         }
1216
1217 out:
1218         mdt_object_put(info->mti_env, o);
1219         RETURN(rc);
1220 }
1221
1222 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1223 {
1224         struct mdt_device       *mdt = info->mti_mdt;
1225         struct ptlrpc_request   *req = mdt_info_req(info);
1226         struct mdt_object       *parent;
1227         struct mdt_object       *child;
1228         struct mdt_lock_handle  *lh;
1229         struct ldlm_reply       *ldlm_rep;
1230         struct mdt_body         *repbody;
1231         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1232         struct md_attr          *ma = &info->mti_attr;
1233         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1234         struct mdt_reint_record *rr = &info->mti_rr;
1235         struct lu_name          *lname;
1236         int                      result, rc;
1237         int                      created = 0;
1238         __u32                    msg_flags;
1239         ENTRY;
1240
1241         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1242                                (obd_timeout + 1) / 4);
1243
1244         mdt_counter_incr(req, LPROC_MDT_OPEN);
1245         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1246
1247         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1248         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1249                                                RCL_SERVER);
1250         ma->ma_need = MA_INODE;
1251         if (ma->ma_lmm_size > 0)
1252                 ma->ma_need |= MA_LOV;
1253
1254         ma->ma_valid = 0;
1255
1256         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1257         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1258
1259         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1260                 CERROR("file join is not supported anymore.\n");
1261                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1262         }
1263         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1264
1265         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1266             info->mti_spec.u.sp_ea.eadata == NULL)
1267                 GOTO(out, result = err_serious(-EINVAL));
1268
1269         CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1270                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1271                PFID(rr->rr_fid1), rr->rr_name,
1272                PFID(rr->rr_fid2), create_flags,
1273                ma->ma_attr.la_mode, msg_flags);
1274
1275         if (req_is_replay(req) ||
1276             (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1277                 /* This is a replay request or from liblustre with ea. */
1278                 result = mdt_open_by_fid(info, ldlm_rep);
1279
1280                 if (result != -ENOENT) {
1281                         if (req->rq_export->exp_libclient &&
1282                             create_flags & MDS_OPEN_HAS_EA)
1283                                 GOTO(out, result = 0);
1284                         GOTO(out, result);
1285                 }
1286                 /* We didn't find the correct object, so we need to re-create it
1287                  * via a regular replay. */
1288                 if (!(create_flags & MDS_OPEN_CREAT)) {
1289                         DEBUG_REQ(D_ERROR, req,
1290                                   "OPEN & CREAT not in open replay/by_fid.");
1291                         GOTO(out, result = -EFAULT);
1292                 }
1293                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1294         } else if ((rr->rr_namelen == 0 && !info->mti_cross_ref &&
1295                     create_flags & MDS_OPEN_LOCK) ||
1296                    (create_flags & MDS_OPEN_BY_FID)) {
1297                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1298                 if (result != -ENOENT && !(create_flags & MDS_OPEN_CREAT))
1299                         GOTO(out, result);
1300                 if (unlikely(rr->rr_namelen == 0))
1301                         GOTO(out, result = -EINVAL);
1302                 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1303         }
1304
1305         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1306                 GOTO(out, result = err_serious(-ENOMEM));
1307
1308         mdt_set_disposition(info, ldlm_rep,
1309                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1310
1311         if (info->mti_cross_ref) {
1312                 /* This is cross-ref open */
1313                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1314                 result = mdt_cross_open(info, rr->rr_fid1, ldlm_rep,
1315                                         create_flags);
1316                 GOTO(out, result);
1317         }
1318
1319         lh = &info->mti_lh[MDT_LH_PARENT];
1320         mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1321                           LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1322
1323         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1324                                       MDS_INODELOCK_UPDATE);
1325         if (IS_ERR(parent))
1326                 GOTO(out, result = PTR_ERR(parent));
1327
1328         /* get and check version of parent */
1329         result = mdt_version_get_check(info, parent, 0);
1330         if (result)
1331                 GOTO(out_parent, result);
1332
1333         fid_zero(child_fid);
1334
1335         lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1336         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1337                             lname, child_fid, &info->mti_spec);
1338         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1339                  "looking for "DFID"/%s, result fid="DFID"\n",
1340                  PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1341
1342         if (result != 0 && result != -ENOENT && result != -ESTALE)
1343                 GOTO(out_parent, result);
1344
1345         if (result == -ENOENT || result == -ESTALE) {
1346                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1347                 if (result == -ESTALE) {
1348                         /*
1349                          * -ESTALE means the parent is a dead(unlinked) dir, so
1350                          * it should return -ENOENT to in accordance with the
1351                          * original mds implementaion.
1352                          */
1353                         GOTO(out_parent, result = -ENOENT);
1354                 }
1355                 if (!(create_flags & MDS_OPEN_CREAT))
1356                         GOTO(out_parent, result);
1357                 *child_fid = *info->mti_rr.rr_fid2;
1358                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1359                          PFID(child_fid));
1360                 /* In the function below, .hs_keycmp resolves to
1361                  * lu_obj_hop_keycmp() */
1362                 /* coverity[overrun-buffer-val] */
1363                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1364         } else {
1365                 /*
1366                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1367                  * return FID back in that case.
1368                  */
1369                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1370                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1371         }
1372         if (IS_ERR(child))
1373                 GOTO(out_parent, result = PTR_ERR(child));
1374
1375         /** check version of child  */
1376         rc = mdt_version_get_check(info, child, 1);
1377         if (rc)
1378                 GOTO(out_child, result = rc);
1379
1380         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1381         if (result == -ENOENT) {
1382                 if (mdt_object_obf(parent))
1383                         GOTO(out_child, result = -EPERM);
1384
1385                 /* save versions in reply */
1386                 mdt_version_get_save(info, parent, 0);
1387                 mdt_version_get_save(info, child, 1);
1388
1389                 /* version of child will be changed */
1390                 info->mti_mos = child;
1391
1392                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1393                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1394
1395                 /* Let lower layers know what is lock mode on directory. */
1396                 info->mti_spec.sp_cr_mode =
1397                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1398
1399                 /*
1400                  * Do not perform lookup sanity check. We know that name does
1401                  * not exist.
1402                  */
1403                 info->mti_spec.sp_cr_lookup = 0;
1404                 info->mti_spec.sp_feat = &dt_directory_features;
1405
1406                 result = mdo_create(info->mti_env,
1407                                     mdt_object_child(parent),
1408                                     lname,
1409                                     mdt_object_child(child),
1410                                     &info->mti_spec,
1411                                     &info->mti_attr);
1412                 if (result == -ERESTART) {
1413                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1414                         GOTO(out_child, result);
1415                 } else {
1416
1417                         /* XXX: we should call this once, see few lines below */
1418                         if (result == 0)
1419                                 result = mdt_attr_get_complex(info, child, ma);
1420
1421                         if (result != 0)
1422                                 GOTO(out_child, result);
1423                 }
1424                 created = 1;
1425         } else {
1426                 /* We have to get attr & lov ea for this object */
1427                 result = mdt_attr_get_complex(info, child, ma);
1428                 /*
1429                  * The object is on remote node, return its FID for remote open.
1430                  */
1431                 if (result == -EREMOTE) {
1432                         /*
1433                          * Check if this lock already was sent to client and
1434                          * this is resent case. For resent case do not take lock
1435                          * again, use what is already granted.
1436                          */
1437                         LASSERT(lhc != NULL);
1438
1439                         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1440                                 struct ldlm_lock *lock;
1441
1442                                 LASSERT(msg_flags & MSG_RESENT);
1443
1444                                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1445                                 if (!lock) {
1446                                         CERROR("Invalid lock handle "LPX64"\n",
1447                                                lhc->mlh_reg_lh.cookie);
1448                                         LBUG();
1449                                 }
1450                                 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1451                                                         &lock->l_resource->lr_name));
1452                                 LDLM_LOCK_PUT(lock);
1453                                 rc = 0;
1454                         } else {
1455                                 mdt_lock_handle_init(lhc);
1456                                 mdt_lock_reg_init(lhc, LCK_PR);
1457
1458                                 rc = mdt_object_lock(info, child, lhc,
1459                                                      MDS_INODELOCK_LOOKUP,
1460                                                      MDT_CROSS_LOCK);
1461                         }
1462                         repbody->fid1 = *mdt_object_fid(child);
1463                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1464                         if (rc != 0)
1465                                 result = rc;
1466                         GOTO(out_child, result);
1467                 }
1468         }
1469
1470         LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1471
1472         /* get openlock if this is not replay and if a client requested it */
1473         if (!req_is_replay(req) && create_flags & MDS_OPEN_LOCK) {
1474                 ldlm_mode_t lm;
1475
1476                 if (create_flags & FMODE_WRITE)
1477                         lm = LCK_CW;
1478                 else if (create_flags & MDS_FMODE_EXEC)
1479                         lm = LCK_PR;
1480                 else
1481                         lm = LCK_CR;
1482                 mdt_lock_handle_init(lhc);
1483                 mdt_lock_reg_init(lhc, lm);
1484                 rc = mdt_object_lock(info, child, lhc,
1485                                      MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1486                                      MDT_CROSS_LOCK);
1487                 if (rc) {
1488                         result = rc;
1489                         GOTO(out_child, result);
1490                 } else {
1491                         result = -EREMOTE;
1492                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1493                 }
1494         }
1495
1496         /* Try to open it now. */
1497         rc = mdt_finish_open(info, parent, child, create_flags,
1498                              created, ldlm_rep);
1499         if (rc) {
1500                 result = rc;
1501                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1502                         /* openlock was acquired and mdt_finish_open failed -
1503                            drop the openlock */
1504                         mdt_object_unlock(info, child, lhc, 1);
1505                 if (created) {
1506                         ma->ma_need = 0;
1507                         ma->ma_valid = 0;
1508                         ma->ma_cookie_size = 0;
1509                         info->mti_no_need_trans = 1;
1510                         rc = mdo_unlink(info->mti_env,
1511                                         mdt_object_child(parent),
1512                                         mdt_object_child(child),
1513                                         lname,
1514                                         &info->mti_attr);
1515                         if (rc != 0)
1516                                 CERROR("Error in cleanup of open\n");
1517                 }
1518         }
1519         EXIT;
1520 out_child:
1521         mdt_object_put(info->mti_env, child);
1522 out_parent:
1523         mdt_object_unlock_put(info, parent, lh, result || !created);
1524 out:
1525         if (result && result != -EREMOTE)
1526                 lustre_msg_set_transno(req->rq_repmsg, 0);
1527         return result;
1528 }
1529
1530 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
1531                                       MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
1532
1533 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1534 {
1535         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1536 }
1537
1538 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1539 {
1540         struct mdt_object *o = mfd->mfd_object;
1541         struct md_object *next = mdt_object_child(o);
1542         struct md_attr *ma = &info->mti_attr;
1543         int ret = MDT_IOEPOCH_CLOSED;
1544         int rc = 0;
1545         int mode;
1546         ENTRY;
1547
1548         mode = mfd->mfd_mode;
1549
1550         if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
1551                 mdt_write_put(o);
1552                 ret = mdt_ioepoch_close(info, o);
1553         } else if (mode & MDS_FMODE_EXEC) {
1554                 mdt_write_allow(o);
1555         } else if (mode & MDS_FMODE_EPOCH) {
1556                 ret = mdt_ioepoch_close(info, o);
1557         } else if (mode & MDS_FMODE_SOM) {
1558                 ret = mdt_som_au_close(info, o);
1559         }
1560
1561         /* Update atime on close only. */
1562         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1563             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1564                 /* Set the atime only. */
1565                 ma->ma_valid = MA_INODE;
1566                 ma->ma_attr.la_valid = LA_ATIME;
1567                 rc = mo_attr_set(info->mti_env, next, ma);
1568         }
1569
1570         ma->ma_need |= MA_INODE;
1571         ma->ma_valid &= ~MA_INODE;
1572
1573         if (!MFD_CLOSED(mode))
1574                 rc = mo_close(info->mti_env, next, ma, mode);
1575
1576         if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
1577                 struct mdt_export_data *med;
1578
1579                 /* The IOepoch is still opened or SOM update is needed.
1580                  * Put mfd back into the list. */
1581                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
1582                 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
1583                                       MDS_FMODE_EPOCH : MDS_FMODE_SOM);
1584
1585                 LASSERT(mdt_info_req(info));
1586                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
1587                 cfs_spin_lock(&med->med_open_lock);
1588                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
1589                 class_handle_hash_back(&mfd->mfd_handle);
1590                 cfs_spin_unlock(&med->med_open_lock);
1591
1592                 if (ret == MDT_IOEPOCH_OPENED) {
1593                         ret = 0;
1594                 } else {
1595                         ret = -EAGAIN;
1596                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
1597                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
1598                 }
1599         } else {
1600                 mdt_mfd_free(mfd);
1601                 mdt_object_put(info->mti_env, o);
1602         }
1603
1604         RETURN(rc ? rc : ret);
1605 }
1606
1607 int mdt_close(struct mdt_thread_info *info)
1608 {
1609         struct mdt_export_data *med;
1610         struct mdt_file_data   *mfd;
1611         struct mdt_object      *o;
1612         struct md_attr         *ma = &info->mti_attr;
1613         struct mdt_body        *repbody = NULL;
1614         struct ptlrpc_request  *req = mdt_info_req(info);
1615         int rc, ret = 0;
1616         ENTRY;
1617
1618         mdt_counter_incr(req, LPROC_MDT_CLOSE);
1619         /* Close may come with the Size-on-MDS update. Unpack it. */
1620         rc = mdt_close_unpack(info);
1621         if (rc)
1622                 RETURN(err_serious(rc));
1623
1624         LASSERT(info->mti_ioepoch);
1625
1626         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
1627                              info->mti_mdt->mdt_max_mdsize);
1628         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
1629                              info->mti_mdt->mdt_max_cookiesize);
1630         rc = req_capsule_server_pack(info->mti_pill);
1631         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
1632                 mdt_client_compatibility(info);
1633                 if (rc == 0)
1634                         mdt_fix_reply(info);
1635                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1636         }
1637
1638         /* Continue to close handle even if we can not pack reply */
1639         if (rc == 0) {
1640                 repbody = req_capsule_server_get(info->mti_pill,
1641                                                  &RMF_MDT_BODY);
1642                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
1643                                                     &RMF_MDT_MD);
1644                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
1645                                                        &RMF_MDT_MD,
1646                                                        RCL_SERVER);
1647                 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
1648                                                        &RMF_LOGCOOKIES);
1649                 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
1650                                                           &RMF_LOGCOOKIES,
1651                                                           RCL_SERVER);
1652                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
1653                 repbody->eadatasize = 0;
1654                 repbody->aclsize = 0;
1655         } else {
1656                 rc = err_serious(rc);
1657         }
1658
1659         med = &req->rq_export->exp_mdt_data;
1660         cfs_spin_lock(&med->med_open_lock);
1661         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1662         if (mdt_mfd_closed(mfd)) {
1663                 cfs_spin_unlock(&med->med_open_lock);
1664                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
1665                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
1666                        info->mti_ioepoch->handle.cookie);
1667                 /** not serious error since bug 3633 */
1668                 rc = -ESTALE;
1669         } else {
1670                 class_handle_unhash(&mfd->mfd_handle);
1671                 cfs_list_del_init(&mfd->mfd_list);
1672                 cfs_spin_unlock(&med->med_open_lock);
1673
1674                 /* Do not lose object before last unlink. */
1675                 o = mfd->mfd_object;
1676                 mdt_object_get(info->mti_env, o);
1677                 ret = mdt_mfd_close(info, mfd);
1678                 if (repbody != NULL)
1679                         rc = mdt_handle_last_unlink(info, o, ma);
1680                 mdt_empty_transno(info, rc);
1681                 mdt_object_put(info->mti_env, o);
1682         }
1683         if (repbody != NULL) {
1684                 mdt_client_compatibility(info);
1685                 rc = mdt_fix_reply(info);
1686         }
1687
1688         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
1689                 RETURN(err_serious(-ENOMEM));
1690
1691         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
1692                                  OBD_FAIL_MDS_CLOSE_NET_REP))
1693                 info->mti_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
1694         RETURN(rc ? rc : ret);
1695 }
1696
1697 /**
1698  * DONE_WRITING rpc handler.
1699  *
1700  * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
1701  * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
1702  * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
1703  * skip attributes and reconstruct the reply here.
1704  */
1705 int mdt_done_writing(struct mdt_thread_info *info)
1706 {
1707         struct ptlrpc_request   *req = mdt_info_req(info);
1708         struct mdt_body         *repbody = NULL;
1709         struct mdt_export_data  *med;
1710         struct mdt_file_data    *mfd;
1711         int rc;
1712         ENTRY;
1713
1714         rc = req_capsule_server_pack(info->mti_pill);
1715         if (rc)
1716                 RETURN(err_serious(rc));
1717
1718         repbody = req_capsule_server_get(info->mti_pill,
1719                                          &RMF_MDT_BODY);
1720         repbody->eadatasize = 0;
1721         repbody->aclsize = 0;
1722
1723         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
1724         rc = mdt_close_unpack(info);
1725         if (rc)
1726                 RETURN(err_serious(rc));
1727
1728         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL))
1729                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1730
1731         med = &info->mti_exp->exp_mdt_data;
1732         cfs_spin_lock(&med->med_open_lock);
1733         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1734         if (mfd == NULL) {
1735                 cfs_spin_unlock(&med->med_open_lock);
1736                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
1737                        ": cookie = "LPX64" ioepoch = "LPU64"\n",
1738                        PFID(info->mti_rr.rr_fid1),
1739                        info->mti_ioepoch->handle.cookie,
1740                        info->mti_ioepoch->ioepoch);
1741                 /* If this is a replay, reconstruct the transno. */
1742                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1743                         rc = info->mti_ioepoch->flags & MF_SOM_AU ?
1744                              -EAGAIN : 0;
1745                         mdt_empty_transno(info, rc);
1746                         RETURN(rc);
1747                 }
1748                 RETURN(-ESTALE);
1749         }
1750
1751         LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
1752                 mfd->mfd_mode == MDS_FMODE_TRUNC);
1753         class_handle_unhash(&mfd->mfd_handle);
1754         cfs_list_del_init(&mfd->mfd_list);
1755         cfs_spin_unlock(&med->med_open_lock);
1756
1757         /* Set EPOCH CLOSE flag if not set by client. */
1758         info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
1759         info->mti_attr.ma_valid = 0;
1760
1761         info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
1762         OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1763         if (info->mti_attr.ma_lmm == NULL)
1764                 RETURN(-ENOMEM);
1765
1766         rc = mdt_mfd_close(info, mfd);
1767
1768         OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1769         mdt_empty_transno(info, rc);
1770         RETURN(rc);
1771 }