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