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 lustre_mdt_attrs *lma;
 227         struct md_attr          *ma = &info->mti_attr;
 228         struct lu_buf           *buf = &info->mti_buf;
 229         struct md_object        *next = mdt_object_child(obj);
 230         struct mdt_device       *mdt = info->mti_mdt;
 231         struct lu_attr          *la = &ma->ma_attr;
 232         int                      rc;
 233         ENTRY;
 234
 235         CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
 236                " on "DFID".\n", enable ? "update" : "disabling",
 237                ioepoch, PFID(mdt_object_fid(obj)));
 238
 239         lma = (struct lustre_mdt_attrs *) info->mti_xattr_buf;
 240         CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*lma));
 241
 242         buf->lb_buf = lma;
 243         buf->lb_len = sizeof(info->mti_xattr_buf);
 244         rc = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_LMA);
 245         if (rc > 0) {
 246                 lustre_lma_swab(lma);
 247         } else if (rc == -ENODATA) {
 248                 memset(lma, 0, sizeof(*lma));
 249         } else {
 250                 RETURN(rc);
 251         }
 252
 253         /* Copy FID */
 254         memcpy(&lma->lma_self_fid, mdt_object_fid(obj), sizeof(lma->lma_self_fid));
 255
 256         /* Copy SOM data */
 257         lma->lma_ioepoch     = ioepoch;
 258         lma->lma_som_size    = la->la_valid & LA_SIZE ? la->la_size : 0;
 259         lma->lma_som_blocks  = la->la_valid & LA_BLOCKS ?  la->la_blocks : 0;
 260         lma->lma_som_mountid = mdt->mdt_lut.lut_obd->u.obt.obt_mount_count;
 261         if (enable)
 262                 lma->lma_compat |= LMAC_SOM;
 263         else
 264                 lma->lma_compat &= ~LMAC_SOM;
 265
 266         rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_LMA, 0);
 267
 268         RETURN(rc);
 269 }
 270
 271 /** Perform the eviction specific actions on ioepoch close. */
 272 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
 273                                                 struct mdt_object *o)
 274 {
 275         int rc = 0;
 276
 277         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 278         CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
 279                "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
 280                o->mot_ioepoch_count);
 281         o->mot_ioepoch_count--;
 282
 283         /* If eviction occured set MOF_SOM_RECOV,
 284          * if no other epoch holders, disable SOM on disk. */
 285         o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
 286         if (!mdt_ioepoch_opened(o)) {
 287                 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
 288                 mdt_object_som_enable(o, o->mot_ioepoch);
 289         }
 290         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 291         RETURN(rc);
 292 }
 293
 294 /**
 295  * Perform the replay specific actions on ioepoch close.
 296  * Skip SOM attribute update if obtained and just forget about the inode state
 297  * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
 298  */
 299 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
 300                                               struct mdt_object *o)
 301 {
 302         int rc = MDT_IOEPOCH_CLOSED;
 303         ENTRY;
 304
 305         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 306         CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
 307                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
 308         o->mot_ioepoch_count--;
 309
 310         /* Get an info from the replayed request if client is supposed
 311          * to send an Attibute Update, reconstruct @rc if so */
 312         if (info->mti_ioepoch->flags & MF_SOM_AU)
 313                 rc = MDT_IOEPOCH_GETATTR;
 314
 315         if (!mdt_ioepoch_opened(o))
 316                 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
 317         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 318
 319         RETURN(rc);
 320 }
 321
 322 /**
 323  * Regular file IOepoch close.
 324  * Closes the ioepoch, checks the object state, apply obtained attributes and
 325  * re-enable SOM on the object, if possible. Also checks if the recovery is
 326  * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
 327  * to obtain SOM attributes under the server-side OST locks.
 328  *
 329  * Return value:
 330  * MDT_IOEPOCH_CLOSED if ioepoch is closed.
 331  * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
 332  */
 333 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
 334                                         struct mdt_object *o)
 335 {
 336         struct md_attr *tmp_ma;
 337         struct lu_attr *la;
 338         int achange, opened;
 339         int recovery = 0;
 340         int rc = 0, ret = MDT_IOEPOCH_CLOSED;
 341         ENTRY;
 342
 343         la = &info->mti_attr.ma_attr;
 344         achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
 345
 346         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 347         o->mot_ioepoch_count--;
 348
 349         tmp_ma = &info->mti_u.som.attr;
 350         tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
 351         tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
 352         tmp_ma->ma_som = &info->mti_u.som.data;
 353         tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
 354         tmp_ma->ma_valid = 0;
 355         rc = mdt_attr_get_complex(info, o, tmp_ma);
 356         if (rc)
 357                 GOTO(error_up, rc);
 358
 359         /* Check the on-disk SOM state. */
 360         if (o->mot_flags & MOF_SOM_RECOV)
 361                 recovery = 1;
 362         else if (!(o->mot_flags & MOF_SOM_CREATED) &&
 363                  !(tmp_ma->ma_valid & MA_SOM))
 364                 recovery = 1;
 365
 366         CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
 367                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
 368
 369         opened = mdt_ioepoch_opened(o);
 370         /**
 371          * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
 372          * Skip the check for file with no LOV  or for unlink files.
 373          */
 374         if (!opened && tmp_ma->ma_valid & MA_LOV &&
 375             !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
 376                 if (recovery)
 377                         /* If some previous writer was evicted, re-ask the
 378                          * client for attributes. Even if attributes are
 379                          * provided, we cannot believe in them.
 380                          * Another use case is that there is no SOM cache on
 381                          * disk -- first access with SOM or there was an MDS
 382                          * failure. */
 383                         ret = MDT_IOEPOCH_GETATTR;
 384                 else if (o->mot_flags & MOF_SOM_CHANGE)
 385                         /* Some previous writer changed the attribute.
 386                          * Do not believe to the current Size-on-MDS
 387                          * update, re-ask client. */
 388                         ret = MDT_IOEPOCH_GETATTR;
 389                 else if (!(la->la_valid & LA_SIZE) && achange)
 390                         /* Attributes were changed by the last writer
 391                          * only but no Size-on-MDS update is received.*/
 392                         ret = MDT_IOEPOCH_GETATTR;
 393         }
 394
 395         if (achange || ret == MDT_IOEPOCH_GETATTR)
 396                 o->mot_flags |= MOF_SOM_CHANGE;
 397
 398         /* If epoch ends and relable SOM attributes are obtained, update them.
 399          * Create SOM ea for new files even if there is no attributes obtained
 400          * (0-length file). */
 401         if (ret == MDT_IOEPOCH_CLOSED && !opened) {
 402                 if (achange || o->mot_flags & MOF_SOM_CREATED) {
 403                         LASSERT(achange || !(la->la_valid & LA_SIZE));
 404                         rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
 405                                               MDT_SOM_ENABLE);
 406                         /* Avoid the following setattrs of these attributes,
 407                          * e.g. for atime update. */
 408                         info->mti_attr.ma_valid = 0;
 409                 }
 410                 mdt_object_som_enable(o, o->mot_ioepoch);
 411         }
 412
 413         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 414         /* If recovery is needed, tell the client to perform GETATTR under
 415          * the lock. */
 416         if (ret == MDT_IOEPOCH_GETATTR && recovery) {
 417                 struct mdt_body *rep;
 418                 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 419                 rep->valid |= OBD_MD_FLGETATTRLOCK;
 420         }
 421
 422         RETURN(rc ? : ret);
 423
 424 error_up:
 425         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 426         return rc;
 427 }
 428
 429 /**
 430  * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
 431  * - a client closes the IOEpoch;
 432  * - a client eviction occured.
 433  * Return values:
 434  * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
 435  * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
 436  * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
 437  * update is needed.
 438  */
 439 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
 440 {
 441         struct ptlrpc_request *req = mdt_info_req(info);
 442         ENTRY;
 443
 444         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 445             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
 446                 RETURN(0);
 447
 448         LASSERT(o->mot_ioepoch_count);
 449         LASSERT(info->mti_ioepoch == NULL ||
 450                 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
 451
 452         /* IOEpoch is closed only if client tells about it or eviction occures.
 453          * In the replay case, always close the epoch. */
 454         if (req == NULL)
 455                 RETURN(mdt_ioepoch_close_on_eviction(info, o));
 456         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
 457                 RETURN(mdt_ioepoch_close_on_replay(info, o));
 458         if (info->mti_ioepoch->flags & MF_EPOCH_CLOSE)
 459                 RETURN(mdt_ioepoch_close_reg(info, o));
 460         /* IO epoch is not closed. */
 461         RETURN(MDT_IOEPOCH_OPENED);
 462 }
 463
 464 /**
 465  * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
 466  * for attribute update. It happens if:
 467  * - SOM Attribute Update is obtained;
 468  * - the client failed to obtain it and informs MDS about it;
 469  * - a client eviction occured.
 470  * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
 471  * cache otherwise.
 472  */
 473 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
 474 {
 475         struct ptlrpc_request *req = mdt_info_req(info);
 476         __u64 ioepoch = 0;
 477         int act = MDT_SOM_ENABLE;
 478         int rc = 0;
 479         ENTRY;
 480
 481         LASSERT(!req || info->mti_ioepoch);
 482         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 483             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
 484                 RETURN(0);
 485
 486         /* No size whereas MF_SOM_CHANGE is set means client failed to
 487          * obtain ost attributes, drop the SOM cache on disk if so. */
 488         if (!req ||
 489             (info->mti_ioepoch &&
 490              info->mti_ioepoch->flags & MF_SOM_CHANGE &&
 491              !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
 492                 act = MDT_SOM_DISABLE;
 493
 494         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 495         /* Mark the object it is the recovery state if we failed to obtain
 496          * SOM attributes. */
 497         if (act == MDT_SOM_DISABLE)
 498                 o->mot_flags |= MOF_SOM_RECOV;
 499
 500         if (!mdt_ioepoch_opened(o)) {
 501                 ioepoch =  info->mti_ioepoch ?
 502                         info->mti_ioepoch->ioepoch : o->mot_ioepoch;
 503
 504                 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 505                         rc = mdt_som_attr_set(info, o, ioepoch, act);
 506                 mdt_object_som_enable(o, ioepoch);
 507         }
 508         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 509         RETURN(rc);
 510 }
 511
 512 int mdt_write_read(struct mdt_object *o)
 513 {
 514         int rc = 0;
 515         ENTRY;
 516         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 517         rc = o->mot_writecount;
 518         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 519         RETURN(rc);
 520 }
 521
 522 int mdt_write_get(struct mdt_object *o)
 523 {
 524         int rc = 0;
 525         ENTRY;
 526         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 527         if (o->mot_writecount < 0)
 528                 rc = -ETXTBSY;
 529         else
 530                 o->mot_writecount++;
 531         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 532         RETURN(rc);
 533 }
 534
 535 void mdt_write_put(struct mdt_object *o)
 536 {
 537         ENTRY;
 538         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 539         o->mot_writecount--;
 540         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 541         EXIT;
 542 }
 543
 544 static int mdt_write_deny(struct mdt_object *o)
 545 {
 546         int rc = 0;
 547         ENTRY;
 548         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 549         if (o->mot_writecount > 0)
 550                 rc = -ETXTBSY;
 551         else
 552                 o->mot_writecount--;
 553         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 554         RETURN(rc);
 555 }
 556
 557 static void mdt_write_allow(struct mdt_object *o)
 558 {
 559         ENTRY;
 560         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 561         o->mot_writecount++;
 562         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 563         EXIT;
 564 }
 565
 566 /* there can be no real transaction so prepare the fake one */
 567 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 568 {
 569         struct mdt_device      *mdt = info->mti_mdt;
 570         struct ptlrpc_request  *req = mdt_info_req(info);
 571         struct tg_export_data  *ted;
 572         struct lsd_client_data *lcd;
 573
 574         ENTRY;
 575         /* transaction has occurred already */
 576         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 577                 RETURN_EXIT;
 578
 579         cfs_spin_lock(&mdt->mdt_lut.lut_translock);
 580         if (info->mti_transno == 0) {
 581                 info->mti_transno = ++ mdt->mdt_lut.lut_last_transno;
 582         } else {
 583                 /* should be replay */
 584                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 585                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 586         }
 587         cfs_spin_unlock(&mdt->mdt_lut.lut_translock);
 588
 589         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 590                         info->mti_transno,
 591                         req->rq_export->exp_obd->obd_last_committed);
 592
 593         req->rq_transno = info->mti_transno;
 594         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 595
 596         /* update lcd in memory only for resent cases */
 597         ted = &req->rq_export->exp_target_data;
 598         LASSERT(ted);
 599         cfs_mutex_lock(&ted->ted_lcd_lock);
 600         lcd = ted->ted_lcd;
 601         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 602             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 603                 if (info->mti_transno != 0)
 604                         lcd->lcd_last_close_transno = info->mti_transno;
 605                 lcd->lcd_last_close_xid = req->rq_xid;
 606                 lcd->lcd_last_close_result = rc;
 607         } else {
 608                 /* VBR: save versions in last_rcvd for reconstruct. */
 609                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 610                 if (pre_versions) {
 611                         lcd->lcd_pre_versions[0] = pre_versions[0];
 612                         lcd->lcd_pre_versions[1] = pre_versions[1];
 613                         lcd->lcd_pre_versions[2] = pre_versions[2];
 614                         lcd->lcd_pre_versions[3] = pre_versions[3];
 615                 }
 616                 if (info->mti_transno != 0)
 617                         lcd->lcd_last_transno = info->mti_transno;
 618                 lcd->lcd_last_xid = req->rq_xid;
 619                 lcd->lcd_last_result = rc;
 620                 lcd->lcd_last_data = info->mti_opdata;
 621         }
 622         cfs_mutex_unlock(&ted->ted_lcd_lock);
 623
 624         EXIT;
 625 }
 626
 627 void mdt_mfd_set_mode(struct mdt_file_data *mfd, int mode)
 628 {
 629         LASSERT(mfd != NULL);
 630
 631         CDEBUG(D_HA, "Change mfd %p mode 0x%x->0x%x\n",
 632                mfd, (unsigned int)mfd->mfd_mode, (unsigned int)mode);
 633
 634         mfd->mfd_mode = mode;
 635 }
 636
 637 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 638                         struct mdt_object *o, __u64 flags, int created)
 639 {
 640         struct ptlrpc_request   *req = mdt_info_req(info);
 641         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 642         struct mdt_file_data    *mfd;
 643         struct md_attr          *ma  = &info->mti_attr;
 644         struct lu_attr          *la  = &ma->ma_attr;
 645         struct mdt_body         *repbody;
 646         int                      rc = 0, isdir, isreg;
 647         ENTRY;
 648
 649         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 650
 651         isreg = S_ISREG(la->la_mode);
 652         isdir = S_ISDIR(la->la_mode);
 653         if (isreg && !(ma->ma_valid & MA_LOV)) {
 654                 /*
 655                  * No EA, check whether it is will set regEA and dirEA since in
 656                  * above attr get, these size might be zero, so reset it, to
 657                  * retrieve the MD after create obj.
 658                  */
 659                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 660                                                        &RMF_MDT_MD,
 661                                                        RCL_SERVER);
 662                 /* in replay case, p == NULL */
 663                 rc = mdt_create_data(info, p, o);
 664                 if (rc)
 665                         RETURN(rc);
 666         }
 667
 668         CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
 669                ma->ma_valid, ma->ma_lmm_size);
 670
 671         if (ma->ma_valid & MA_LOV) {
 672                 LASSERT(ma->ma_lmm_size != 0);
 673                 repbody->eadatasize = ma->ma_lmm_size;
 674                 if (isdir)
 675                         repbody->valid |= OBD_MD_FLDIREA;
 676                 else
 677                         repbody->valid |= OBD_MD_FLEASIZE;
 678         }
 679
 680         if (flags & FMODE_WRITE) {
 681                 rc = mdt_write_get(o);
 682                 if (rc == 0) {
 683                         mdt_ioepoch_open(info, o, created);
 684                         repbody->ioepoch = o->mot_ioepoch;
 685                 }
 686         } else if (flags & MDS_FMODE_EXEC) {
 687                 rc = mdt_write_deny(o);
 688         }
 689         if (rc)
 690                 RETURN(rc);
 691
 692         rc = mo_open(info->mti_env, mdt_object_child(o),
 693                      created ? flags | MDS_OPEN_CREATED : flags);
 694         if (rc)
 695                 GOTO(err_out, rc);
 696
 697         mfd = mdt_mfd_new();
 698         if (mfd != NULL) {
 699                 /*
 700                  * Keep a reference on this object for this open, and is
 701                  * released by mdt_mfd_close().
 702                  */
 703                 mdt_object_get(info->mti_env, o);
 704
 705                 /*
 706                  * @flags is always not zero. At least it should be FMODE_READ,
 707                  * FMODE_WRITE or MDS_FMODE_EXEC.
 708                  */
 709                 LASSERT(flags != 0);
 710
 711                 /* Open handling. */
 712                 mdt_mfd_set_mode(mfd, flags);
 713
 714                 mfd->mfd_object = o;
 715                 mfd->mfd_xid = req->rq_xid;
 716
 717                 /* replay handle */
 718                 if (req_is_replay(req)) {
 719                         struct mdt_file_data *old_mfd;
 720                         /* Check wheather old cookie already exist in
 721                          * the list, becasue when do recovery, client
 722                          * might be disconnected from server, and
 723                          * restart replay, so there maybe some orphan
 724                          * mfd here, we should remove them */
 725                         LASSERT(info->mti_rr.rr_handle != NULL);
 726                         old_mfd = mdt_handle2mfd(info, info->mti_rr.rr_handle);
 727                         if (old_mfd) {
 728                                 CDEBUG(D_HA, "del orph mfd %p fid=("DFID") "
 729                                        "cookie=" LPX64"\n", mfd,
 730                                        PFID(mdt_object_fid(mfd->mfd_object)),
 731                                        info->mti_rr.rr_handle->cookie);
 732                                 cfs_spin_lock(&med->med_open_lock);
 733                                 class_handle_unhash(&old_mfd->mfd_handle);
 734                                 cfs_list_del_init(&old_mfd->mfd_list);
 735                                 cfs_spin_unlock(&med->med_open_lock);
 736                                 /* no attr update for that close */
 737                                 la->la_valid = 0;
 738                                 ma->ma_valid |= MA_FLAGS;
 739                                 ma->ma_attr_flags |= MDS_RECOV_OPEN;
 740                                 mdt_mfd_close(info, old_mfd);
 741                                 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 742                                 ma->ma_valid &= ~MA_FLAGS;
 743                         }
 744                         CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
 745                                info->mti_rr.rr_handle->cookie);
 746                         mfd->mfd_old_handle.cookie =
 747                                                 info->mti_rr.rr_handle->cookie;
 748                 }
 749                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 750
 751                 if (req->rq_export->exp_disconnected) {
 752                         cfs_spin_lock(&med->med_open_lock);
 753                         class_handle_unhash(&mfd->mfd_handle);
 754                         cfs_list_del_init(&mfd->mfd_list);
 755                         cfs_spin_unlock(&med->med_open_lock);
 756                         mdt_mfd_close(info, mfd);
 757                 } else {
 758                         cfs_spin_lock(&med->med_open_lock);
 759                         cfs_list_add(&mfd->mfd_list, &med->med_open_head);
 760                         cfs_spin_unlock(&med->med_open_lock);
 761                 }
 762
 763                 mdt_empty_transno(info, rc);
 764         } else {
 765                 GOTO(err_out, rc = -ENOMEM);
 766         }
 767
 768         RETURN(rc);
 769
 770 err_out:
 771         if (flags & FMODE_WRITE)
 772                         /* XXX We also need to close io epoch here.
 773                          * See LU-1220 - green */
 774                 mdt_write_put(o);
 775         else if (flags & FMODE_EXEC)
 776                 mdt_write_allow(o);
 777         return rc;
 778 }
 779
 780 int mdt_finish_open(struct mdt_thread_info *info,
 781                     struct mdt_object *p, struct mdt_object *o,
 782                     __u64 flags, int created, struct ldlm_reply *rep)
 783 {
 784         struct ptlrpc_request   *req = mdt_info_req(info);
 785         struct obd_export       *exp = req->rq_export;
 786         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 787         struct md_attr          *ma  = &info->mti_attr;
 788         struct lu_attr          *la  = &ma->ma_attr;
 789         struct mdt_file_data    *mfd;
 790         struct mdt_body         *repbody;
 791         int                      rc = 0;
 792         int                      isreg, isdir, islnk;
 793         cfs_list_t              *t;
 794         ENTRY;
 795
 796         LASSERT(ma->ma_valid & MA_INODE);
 797
 798         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 799
 800         isreg = S_ISREG(la->la_mode);
 801         isdir = S_ISDIR(la->la_mode);
 802         islnk = S_ISLNK(la->la_mode);
 803         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 804
 805         if (exp_connect_rmtclient(exp)) {
 806                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 807
 808                 rc = mdt_pack_remote_perm(info, o, buf);
 809                 if (rc) {
 810                         repbody->valid &= ~OBD_MD_FLRMTPERM;
 811                         repbody->aclsize = 0;
 812                 } else {
 813                         repbody->valid |= OBD_MD_FLRMTPERM;
 814                         repbody->aclsize = sizeof(struct mdt_remote_perm);
 815                 }
 816         }
 817 #ifdef CONFIG_FS_POSIX_ACL
 818         else if (exp->exp_connect_flags & OBD_CONNECT_ACL) {
 819                 const struct lu_env *env = info->mti_env;
 820                 struct md_object *next = mdt_object_child(o);
 821                 struct lu_buf *buf = &info->mti_buf;
 822
 823                 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 824                 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
 825                                                    RCL_SERVER);
 826                 if (buf->lb_len > 0) {
 827                         rc = mo_xattr_get(env, next, buf,
 828                                           XATTR_NAME_ACL_ACCESS);
 829                         if (rc < 0) {
 830                                 if (rc == -ENODATA) {
 831                                         repbody->aclsize = 0;
 832                                         repbody->valid |= OBD_MD_FLACL;
 833                                         rc = 0;
 834                                 } else if (rc == -EOPNOTSUPP) {
 835                                         rc = 0;
 836                                 } else {
 837                                         CERROR("got acl size: %d\n", rc);
 838                                 }
 839                         } else {
 840                                 repbody->aclsize = rc;
 841                                 repbody->valid |= OBD_MD_FLACL;
 842                                 rc = 0;
 843                         }
 844                 }
 845         }
 846 #endif
 847
 848         if (info->mti_mdt->mdt_opts.mo_mds_capa &&
 849             exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) {
 850                 struct lustre_capa *capa;
 851
 852                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
 853                 LASSERT(capa);
 854                 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
 855                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 856                 if (rc)
 857                         RETURN(rc);
 858                 repbody->valid |= OBD_MD_FLMDSCAPA;
 859         }
 860         if (info->mti_mdt->mdt_opts.mo_oss_capa &&
 861             exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA &&
 862             S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) {
 863                 struct lustre_capa *capa;
 864
 865                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
 866                 LASSERT(capa);
 867                 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
 868                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 869                 if (rc)
 870                         RETURN(rc);
 871                 repbody->valid |= OBD_MD_FLOSSCAPA;
 872         }
 873
 874         /*
 875          * If we are following a symlink, don't open; and do not return open
 876          * handle for special nodes as client required.
 877          */
 878         if (islnk || (!isreg && !isdir &&
 879             (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH))) {
 880                 lustre_msg_set_transno(req->rq_repmsg, 0);
 881                 RETURN(0);
 882         }
 883
 884         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 885
 886         /*
 887          * We need to return the existing object's fid back, so it is done here,
 888          * after preparing the reply.
 889          */
 890         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 891                 RETURN(-EEXIST);
 892
 893         /* This can't be done earlier, we need to return reply body */
 894         if (isdir) {
 895                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 896                         /* We are trying to create or write an existing dir. */
 897                         RETURN(-EISDIR);
 898                 }
 899         } else if (flags & MDS_OPEN_DIRECTORY)
 900                 RETURN(-ENOTDIR);
 901
 902         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 903                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 904                 RETURN(-EAGAIN);
 905         }
 906
 907         mfd = NULL;
 908         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 909                 cfs_spin_lock(&med->med_open_lock);
 910                 cfs_list_for_each(t, &med->med_open_head) {
 911                         mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
 912                         if (mfd->mfd_xid == req->rq_xid) {
 913                                 break;
 914                         }
 915                         mfd = NULL;
 916                 }
 917                 cfs_spin_unlock(&med->med_open_lock);
 918
 919                 if (mfd != NULL) {
 920                         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 921                         /*set repbody->ea_size for resent case*/
 922                         if (ma->ma_valid & MA_LOV) {
 923                                 LASSERT(ma->ma_lmm_size != 0);
 924                                 repbody->eadatasize = ma->ma_lmm_size;
 925                                 if (isdir)
 926                                         repbody->valid |= OBD_MD_FLDIREA;
 927                                 else
 928                                         repbody->valid |= OBD_MD_FLEASIZE;
 929                         }
 930                         RETURN(0);
 931                 }
 932         }
 933
 934         rc = mdt_mfd_open(info, p, o, flags, created);
 935         RETURN(rc);
 936 }
 937
 938 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
 939                              struct lsd_client_data *lcd);
 940
 941 void mdt_reconstruct_open(struct mdt_thread_info *info,
 942                           struct mdt_lock_handle *lhc)
 943 {
 944         const struct lu_env *env = info->mti_env;
 945         struct mdt_device       *mdt  = info->mti_mdt;
 946         struct req_capsule      *pill = info->mti_pill;
 947         struct ptlrpc_request   *req  = mdt_info_req(info);
 948         struct tg_export_data   *ted  = &req->rq_export->exp_target_data;
 949         struct lsd_client_data  *lcd  = ted->ted_lcd;
 950         struct md_attr          *ma   = &info->mti_attr;
 951         struct mdt_reint_record *rr   = &info->mti_rr;
 952         __u32                   flags = info->mti_spec.sp_cr_flags;
 953         struct ldlm_reply       *ldlm_rep;
 954         struct mdt_object       *parent;
 955         struct mdt_object       *child;
 956         struct mdt_body         *repbody;
 957         int                      rc;
 958         ENTRY;
 959
 960         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 961         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 962         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 963
 964         ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
 965         ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
 966                                                RCL_SERVER);
 967         ma->ma_need = MA_INODE;
 968         if (ma->ma_lmm_size > 0)
 969                 ma->ma_need |= MA_LOV;
 970
 971         ma->ma_valid = 0;
 972
 973         mdt_req_from_lcd(req, lcd);
 974         mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
 975
 976         CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
 977                ldlm_rep->lock_policy_res1, req->rq_status);
 978
 979         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
 980             req->rq_status != 0)
 981                 /* We did not create successfully, return error to client. */
 982                 GOTO(out, rc = req->rq_status);
 983
 984         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
 985                 struct obd_export *exp = req->rq_export;
 986                 /*
 987                  * We failed after creation, but we do not know in which step
 988                  * we failed. So try to check the child object.
 989                  */
 990                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
 991                 if (IS_ERR(parent)) {
 992                         rc = PTR_ERR(parent);
 993                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
 994                                       " Evicting client %s with export %s.\n",
 995                                       PFID(rr->rr_fid1), rc,
 996                                       obd_uuid2str(&exp->exp_client_uuid),
 997                                       obd_export_nid2str(exp));
 998                         mdt_export_evict(exp);
 999                         RETURN_EXIT;
1000                 }
1001                 child = mdt_object_find(env, mdt, rr->rr_fid2);
1002                 if (IS_ERR(child)) {
1003                         rc = PTR_ERR(child);
1004                         LCONSOLE_WARN("Child "DFID" lookup error %d."
1005                                       " Evicting client %s with export %s.\n",
1006                                       PFID(mdt_object_fid(child)), rc,
1007                                       obd_uuid2str(&exp->exp_client_uuid),
1008                                       obd_export_nid2str(exp));
1009                         mdt_object_put(env, parent);
1010                         mdt_export_evict(exp);
1011                         RETURN_EXIT;
1012                 }
1013                 rc = mdt_object_exists(child);
1014                 if (rc > 0) {
1015
1016                         mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA);
1017                         rc = mdt_attr_get_complex(info, child, ma);
1018                         if (rc == 0)
1019                               rc = mdt_finish_open(info, parent, child,
1020                                                    flags, 1, ldlm_rep);
1021                 } else if (rc < 0) {
1022                         /* the child object was created on remote server */
1023                         repbody->fid1 = *rr->rr_fid2;
1024                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1025                         rc = 0;
1026                 } else if (rc == 0) {
1027                         /* the child does not exist, we should do regular open */
1028                         mdt_object_put(env, parent);
1029                         mdt_object_put(env, child);
1030                         GOTO(regular_open, 0);
1031                 }
1032                 mdt_object_put(env, parent);
1033                 mdt_object_put(env, child);
1034                 GOTO(out, rc);
1035         } else {
1036 regular_open:
1037                 /* We did not try to create, so we are a pure open */
1038                 rc = mdt_reint_open(info, lhc);
1039         }
1040
1041         EXIT;
1042 out:
1043         req->rq_status = rc;
1044         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1045         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1046 }
1047
1048 int mdt_open_by_fid(struct mdt_thread_info* info,
1049                     struct ldlm_reply *rep)
1050 {
1051         __u32                    flags = info->mti_spec.sp_cr_flags;
1052         struct mdt_reint_record *rr = &info->mti_rr;
1053         struct md_attr          *ma = &info->mti_attr;
1054         struct mdt_object       *o;
1055         int                      rc;
1056         ENTRY;
1057
1058         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1059         if (IS_ERR(o))
1060                 RETURN(rc = PTR_ERR(o));
1061
1062         rc = mdt_object_exists(o);
1063         if (rc > 0) {
1064                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1065                                                 DISP_LOOKUP_EXECD |
1066                                                 DISP_LOOKUP_POS));
1067
1068                 rc = mdt_attr_get_complex(info, o, ma);
1069                 if (rc == 0)
1070                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1071         } else if (rc == 0) {
1072                 rc = -ENOENT;
1073         } else  {
1074                 /* the child object was created on remote server */
1075                 struct mdt_body *repbody;
1076                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1077                 repbody->fid1 = *rr->rr_fid2;
1078                 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1079                 rc = 0;
1080         }
1081
1082         mdt_object_put(info->mti_env, o);
1083         RETURN(rc);
1084 }
1085
1086 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1087                          struct mdt_lock_handle *lhc)
1088 {
1089         const struct lu_env     *env   = info->mti_env;
1090         struct mdt_device       *mdt   = info->mti_mdt;
1091         __u32                    flags = info->mti_spec.sp_cr_flags;
1092         struct mdt_reint_record *rr    = &info->mti_rr;
1093         struct md_attr          *ma    = &info->mti_attr;
1094         struct mdt_object       *parent= NULL;
1095         struct mdt_object       *o;
1096         int                      rc;
1097         ldlm_mode_t              lm;
1098         ENTRY;
1099
1100         if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1101                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1102                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1103                         if (IS_ERR(parent)) {
1104                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1105                                        " for anonymous created %ld, try to"
1106                                        " use server-side parent.\n",
1107                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1108                                 parent = NULL;
1109                         }
1110                 }
1111                 if (parent == NULL)
1112                         ma->ma_need |= MA_PFID;
1113         }
1114
1115         o = mdt_object_find(env, mdt, rr->rr_fid2);
1116         if (IS_ERR(o))
1117                 RETURN(rc = PTR_ERR(o));
1118
1119         rc = mdt_object_exists(o);
1120         if (rc == 0) {
1121                 mdt_set_disposition(info, rep, (DISP_LOOKUP_EXECD |
1122                                     DISP_LOOKUP_NEG));
1123                 GOTO(out, rc = -ENOENT);
1124         } else if (rc < 0) {
1125                 CERROR("NFS remote open shouldn't happen.\n");
1126                 GOTO(out, rc);
1127         }
1128         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1129                                         DISP_LOOKUP_EXECD |
1130                                         DISP_LOOKUP_POS));
1131
1132         if (flags & FMODE_WRITE)
1133                 lm = LCK_CW;
1134         else if (flags & MDS_FMODE_EXEC)
1135                 lm = LCK_PR;
1136         else
1137                 lm = LCK_CR;
1138
1139         mdt_lock_handle_init(lhc);
1140         mdt_lock_reg_init(lhc, lm);
1141         rc = mdt_object_lock(info, o, lhc,
1142                              MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1143                              MDT_CROSS_LOCK);
1144         if (rc)
1145                 GOTO(out, rc);
1146
1147         rc = mdt_attr_get_complex(info, o, ma);
1148         if (rc)
1149                 GOTO(out, rc);
1150
1151         if (ma->ma_valid & MA_PFID) {
1152                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1153                 if (IS_ERR(parent)) {
1154                         CDEBUG(D_INODE, "Fail to find parent "DFID
1155                                " for anonymous created %ld, try to"
1156                                " use system default.\n",
1157                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1158                         parent = NULL;
1159                 }
1160         }
1161
1162         if (flags & MDS_OPEN_LOCK)
1163                 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1164         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1165
1166         if (!(flags & MDS_OPEN_LOCK) || rc)
1167                 mdt_object_unlock(info, o, lhc, 1);
1168
1169         GOTO(out, rc);
1170 out:
1171         mdt_object_put(env, o);
1172         if (parent != NULL)
1173                 mdt_object_put(env, parent);
1174         return rc;
1175 }
1176
1177 int mdt_pin(struct mdt_thread_info* info)
1178 {
1179         ENTRY;
1180         RETURN(err_serious(-EOPNOTSUPP));
1181 }
1182
1183 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1184 int mdt_cross_open(struct mdt_thread_info* info,
1185                    const struct lu_fid *fid,
1186                    struct ldlm_reply *rep, __u32 flags)
1187 {
1188         struct md_attr    *ma = &info->mti_attr;
1189         struct mdt_object *o;
1190         int                rc;
1191         ENTRY;
1192
1193         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1194         if (IS_ERR(o))
1195                 RETURN(rc = PTR_ERR(o));
1196
1197         rc = mdt_object_exists(o);
1198         if (rc > 0) {
1199                 /* Do permission check for cross-open. */
1200                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o),
1201                                    NULL, flags | MDS_OPEN_CROSS);
1202                 if (rc)
1203                         goto out;
1204
1205                 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1206                 rc = mdt_attr_get_complex(info, o, ma);
1207                 if (rc == 0)
1208                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1209         } else if (rc == 0) {
1210                 /*
1211                  * Something is wrong here. lookup was positive but there is
1212                  * no object!
1213                  */
1214                 CERROR("Cross-ref object doesn't exist!\n");
1215                 rc = -EFAULT;
1216         } else  {
1217                 /* Something is wrong here, the object is on another MDS! */
1218                 CERROR("The object isn't on this server! FLD error?\n");
1219                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1220                                 &o->mot_obj.mo_lu,
1221                                 "Object isn't on this server! FLD error?\n");
1222
1223                 rc = -EFAULT;
1224         }
1225
1226 out:
1227         mdt_object_put(info->mti_env, o);
1228         RETURN(rc);
1229 }
1230
1231 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1232 {
1233         struct mdt_device       *mdt = info->mti_mdt;
1234         struct ptlrpc_request   *req = mdt_info_req(info);
1235         struct mdt_object       *parent;
1236         struct mdt_object       *child;
1237         struct mdt_lock_handle  *lh;
1238         struct ldlm_reply       *ldlm_rep;
1239         struct mdt_body         *repbody;
1240         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1241         struct md_attr          *ma = &info->mti_attr;
1242         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1243         struct mdt_reint_record *rr = &info->mti_rr;
1244         struct lu_name          *lname;
1245         int                      result, rc;
1246         int                      created = 0;
1247         __u32                    msg_flags;
1248         ENTRY;
1249
1250         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1251                                (obd_timeout + 1) / 4);
1252
1253         mdt_counter_incr(req, LPROC_MDT_OPEN);
1254         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1255
1256         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1257         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1258                                                RCL_SERVER);
1259         ma->ma_need = MA_INODE;
1260         if (ma->ma_lmm_size > 0)
1261                 ma->ma_need |= MA_LOV;
1262
1263         ma->ma_valid = 0;
1264
1265         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1266         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1267
1268         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1269                 CERROR("file join is not supported anymore.\n");
1270                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1271         }
1272         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1273
1274         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1275             info->mti_spec.u.sp_ea.eadata == NULL)
1276                 GOTO(out, result = err_serious(-EINVAL));
1277
1278         CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1279                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1280                PFID(rr->rr_fid1), rr->rr_name,
1281                PFID(rr->rr_fid2), create_flags,
1282                ma->ma_attr.la_mode, msg_flags);
1283
1284         if (req_is_replay(req) ||
1285             (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1286                 /* This is a replay request or from liblustre with ea. */
1287                 result = mdt_open_by_fid(info, ldlm_rep);
1288
1289                 if (result != -ENOENT) {
1290                         if (req->rq_export->exp_libclient &&
1291                             create_flags & MDS_OPEN_HAS_EA)
1292                                 GOTO(out, result = 0);
1293                         GOTO(out, result);
1294                 }
1295                 /* We didn't find the correct object, so we need to re-create it
1296                  * via a regular replay. */
1297                 if (!(create_flags & MDS_OPEN_CREAT)) {
1298                         DEBUG_REQ(D_ERROR, req,
1299                                   "OPEN & CREAT not in open replay/by_fid.");
1300                         GOTO(out, result = -EFAULT);
1301                 }
1302                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1303         } else if ((rr->rr_namelen == 0 && !info->mti_cross_ref &&
1304                     create_flags & MDS_OPEN_LOCK) ||
1305                    (create_flags & MDS_OPEN_BY_FID)) {
1306                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1307                 if (result != -ENOENT && !(create_flags & MDS_OPEN_CREAT))
1308                         GOTO(out, result);
1309                 if (unlikely(rr->rr_namelen == 0))
1310                         GOTO(out, result = -EINVAL);
1311                 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1312         }
1313
1314         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1315                 GOTO(out, result = err_serious(-ENOMEM));
1316
1317         mdt_set_disposition(info, ldlm_rep,
1318                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1319
1320         if (info->mti_cross_ref) {
1321                 /* This is cross-ref open */
1322                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1323                 result = mdt_cross_open(info, rr->rr_fid1, ldlm_rep,
1324                                         create_flags);
1325                 GOTO(out, result);
1326         }
1327
1328         lh = &info->mti_lh[MDT_LH_PARENT];
1329         mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1330                           LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1331
1332         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1333                                       MDS_INODELOCK_UPDATE);
1334         if (IS_ERR(parent))
1335                 GOTO(out, result = PTR_ERR(parent));
1336
1337         /* get and check version of parent */
1338         result = mdt_version_get_check(info, parent, 0);
1339         if (result)
1340                 GOTO(out_parent, result);
1341
1342         fid_zero(child_fid);
1343
1344         lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1345         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1346                             lname, child_fid, &info->mti_spec);
1347         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1348                  "looking for "DFID"/%s, result fid="DFID"\n",
1349                  PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1350
1351         if (result != 0 && result != -ENOENT && result != -ESTALE)
1352                 GOTO(out_parent, result);
1353
1354         if (result == -ENOENT || result == -ESTALE) {
1355                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1356                 if (result == -ESTALE) {
1357                         /*
1358                          * -ESTALE means the parent is a dead(unlinked) dir, so
1359                          * it should return -ENOENT to in accordance with the
1360                          * original mds implementaion.
1361                          */
1362                         GOTO(out_parent, result = -ENOENT);
1363                 }
1364                 if (!(create_flags & MDS_OPEN_CREAT))
1365                         GOTO(out_parent, result);
1366                 *child_fid = *info->mti_rr.rr_fid2;
1367                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1368                          PFID(child_fid));
1369                 /* In the function below, .hs_keycmp resolves to
1370                  * lu_obj_hop_keycmp() */
1371                 /* coverity[overrun-buffer-val] */
1372                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1373         } else {
1374                 /*
1375                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1376                  * return FID back in that case.
1377                  */
1378                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1379                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1380         }
1381         if (IS_ERR(child))
1382                 GOTO(out_parent, result = PTR_ERR(child));
1383
1384         /** check version of child  */
1385         rc = mdt_version_get_check(info, child, 1);
1386         if (rc)
1387                 GOTO(out_child, result = rc);
1388
1389         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1390         if (result == -ENOENT) {
1391                 if (mdt_object_obf(parent))
1392                         GOTO(out_child, result = -EPERM);
1393
1394                 /* save versions in reply */
1395                 mdt_version_get_save(info, parent, 0);
1396                 mdt_version_get_save(info, child, 1);
1397
1398                 /* version of child will be changed */
1399                 info->mti_mos = child;
1400
1401                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1402                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1403
1404                 /* Let lower layers know what is lock mode on directory. */
1405                 info->mti_spec.sp_cr_mode =
1406                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1407
1408                 /*
1409                  * Do not perform lookup sanity check. We know that name does
1410                  * not exist.
1411                  */
1412                 info->mti_spec.sp_cr_lookup = 0;
1413                 info->mti_spec.sp_feat = &dt_directory_features;
1414
1415                 result = mdo_create(info->mti_env,
1416                                     mdt_object_child(parent),
1417                                     lname,
1418                                     mdt_object_child(child),
1419                                     &info->mti_spec,
1420                                     &info->mti_attr);
1421                 if (result == -ERESTART) {
1422                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1423                         GOTO(out_child, result);
1424                 } else {
1425
1426                         /* XXX: we should call this once, see few lines below */
1427                         if (result == 0)
1428                                 result = mdt_attr_get_complex(info, child, ma);
1429
1430                         if (result != 0)
1431                                 GOTO(out_child, result);
1432                 }
1433                 created = 1;
1434         } else {
1435                 /* We have to get attr & lov ea for this object */
1436                 result = mdt_attr_get_complex(info, child, ma);
1437                 /*
1438                  * The object is on remote node, return its FID for remote open.
1439                  */
1440                 if (result == -EREMOTE) {
1441                         /*
1442                          * Check if this lock already was sent to client and
1443                          * this is resent case. For resent case do not take lock
1444                          * again, use what is already granted.
1445                          */
1446                         LASSERT(lhc != NULL);
1447
1448                         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1449                                 struct ldlm_lock *lock;
1450
1451                                 LASSERT(msg_flags & MSG_RESENT);
1452
1453                                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1454                                 if (!lock) {
1455                                         CERROR("Invalid lock handle "LPX64"\n",
1456                                                lhc->mlh_reg_lh.cookie);
1457                                         LBUG();
1458                                 }
1459                                 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1460                                                         &lock->l_resource->lr_name));
1461                                 LDLM_LOCK_PUT(lock);
1462                                 rc = 0;
1463                         } else {
1464                                 mdt_lock_handle_init(lhc);
1465                                 mdt_lock_reg_init(lhc, LCK_PR);
1466
1467                                 rc = mdt_object_lock(info, child, lhc,
1468                                                      MDS_INODELOCK_LOOKUP,
1469                                                      MDT_CROSS_LOCK);
1470                         }
1471                         repbody->fid1 = *mdt_object_fid(child);
1472                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1473                         if (rc != 0)
1474                                 result = rc;
1475                         GOTO(out_child, result);
1476                 }
1477         }
1478
1479         LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1480
1481         /* get openlock if this is not replay and if a client requested it */
1482         if (!req_is_replay(req) && create_flags & MDS_OPEN_LOCK) {
1483                 ldlm_mode_t lm;
1484
1485                 if (create_flags & FMODE_WRITE)
1486                         lm = LCK_CW;
1487                 else if (create_flags & MDS_FMODE_EXEC)
1488                         lm = LCK_PR;
1489                 else
1490                         lm = LCK_CR;
1491                 mdt_lock_handle_init(lhc);
1492                 mdt_lock_reg_init(lhc, lm);
1493                 rc = mdt_object_lock(info, child, lhc,
1494                                      MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1495                                      MDT_CROSS_LOCK);
1496                 if (rc) {
1497                         result = rc;
1498                         GOTO(out_child, result);
1499                 } else {
1500                         result = -EREMOTE;
1501                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1502                 }
1503         }
1504
1505         /* Try to open it now. */
1506         rc = mdt_finish_open(info, parent, child, create_flags,
1507                              created, ldlm_rep);
1508         if (rc) {
1509                 result = rc;
1510                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1511                         /* openlock was acquired and mdt_finish_open failed -
1512                            drop the openlock */
1513                         mdt_object_unlock(info, child, lhc, 1);
1514                 if (created) {
1515                         ma->ma_need = 0;
1516                         ma->ma_valid = 0;
1517                         ma->ma_cookie_size = 0;
1518                         info->mti_no_need_trans = 1;
1519                         rc = mdo_unlink(info->mti_env,
1520                                         mdt_object_child(parent),
1521                                         mdt_object_child(child),
1522                                         lname,
1523                                         &info->mti_attr);
1524                         if (rc != 0)
1525                                 CERROR("Error in cleanup of open\n");
1526                 }
1527         }
1528         EXIT;
1529 out_child:
1530         mdt_object_put(info->mti_env, child);
1531 out_parent:
1532         mdt_object_unlock_put(info, parent, lh, result || !created);
1533 out:
1534         if (result && result != -EREMOTE)
1535                 lustre_msg_set_transno(req->rq_repmsg, 0);
1536         return result;
1537 }
1538
1539 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
1540                                       MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
1541
1542 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1543 {
1544         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1545 }
1546
1547 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1548 {
1549         struct mdt_object *o = mfd->mfd_object;
1550         struct md_object *next = mdt_object_child(o);
1551         struct md_attr *ma = &info->mti_attr;
1552         int ret = MDT_IOEPOCH_CLOSED;
1553         int rc = 0;
1554         int mode;
1555         ENTRY;
1556
1557         mode = mfd->mfd_mode;
1558
1559         if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
1560                 mdt_write_put(o);
1561                 ret = mdt_ioepoch_close(info, o);
1562         } else if (mode & MDS_FMODE_EXEC) {
1563                 mdt_write_allow(o);
1564         } else if (mode & MDS_FMODE_EPOCH) {
1565                 ret = mdt_ioepoch_close(info, o);
1566         } else if (mode & MDS_FMODE_SOM) {
1567                 ret = mdt_som_au_close(info, o);
1568         }
1569
1570         /* Update atime on close only. */
1571         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1572             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1573                 /* Set the atime only. */
1574                 ma->ma_valid = MA_INODE;
1575                 ma->ma_attr.la_valid = LA_ATIME;
1576                 rc = mo_attr_set(info->mti_env, next, ma);
1577         }
1578
1579         ma->ma_need |= MA_INODE;
1580         ma->ma_valid &= ~MA_INODE;
1581
1582         if (!MFD_CLOSED(mode))
1583                 rc = mo_close(info->mti_env, next, ma, mode);
1584
1585         if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
1586                 struct mdt_export_data *med;
1587
1588                 /* The IOepoch is still opened or SOM update is needed.
1589                  * Put mfd back into the list. */
1590                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
1591                 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
1592                                       MDS_FMODE_EPOCH : MDS_FMODE_SOM);
1593
1594                 LASSERT(mdt_info_req(info));
1595                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
1596                 cfs_spin_lock(&med->med_open_lock);
1597                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
1598                 class_handle_hash_back(&mfd->mfd_handle);
1599                 cfs_spin_unlock(&med->med_open_lock);
1600
1601                 if (ret == MDT_IOEPOCH_OPENED) {
1602                         ret = 0;
1603                 } else {
1604                         ret = -EAGAIN;
1605                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
1606                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
1607                 }
1608         } else {
1609                 mdt_mfd_free(mfd);
1610                 mdt_object_put(info->mti_env, o);
1611         }
1612
1613         RETURN(rc ? rc : ret);
1614 }
1615
1616 int mdt_close(struct mdt_thread_info *info)
1617 {
1618         struct mdt_export_data *med;
1619         struct mdt_file_data   *mfd;
1620         struct mdt_object      *o;
1621         struct md_attr         *ma = &info->mti_attr;
1622         struct mdt_body        *repbody = NULL;
1623         struct ptlrpc_request  *req = mdt_info_req(info);
1624         int rc, ret = 0;
1625         ENTRY;
1626
1627         mdt_counter_incr(req, LPROC_MDT_CLOSE);
1628         /* Close may come with the Size-on-MDS update. Unpack it. */
1629         rc = mdt_close_unpack(info);
1630         if (rc)
1631                 RETURN(err_serious(rc));
1632
1633         LASSERT(info->mti_ioepoch);
1634
1635         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
1636                              info->mti_mdt->mdt_max_mdsize);
1637         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
1638                              info->mti_mdt->mdt_max_cookiesize);
1639         rc = req_capsule_server_pack(info->mti_pill);
1640         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
1641                 mdt_client_compatibility(info);
1642                 if (rc == 0)
1643                         mdt_fix_reply(info);
1644                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1645         }
1646
1647         /* Continue to close handle even if we can not pack reply */
1648         if (rc == 0) {
1649                 repbody = req_capsule_server_get(info->mti_pill,
1650                                                  &RMF_MDT_BODY);
1651                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
1652                                                     &RMF_MDT_MD);
1653                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
1654                                                        &RMF_MDT_MD,
1655                                                        RCL_SERVER);
1656                 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
1657                                                        &RMF_LOGCOOKIES);
1658                 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
1659                                                           &RMF_LOGCOOKIES,
1660                                                           RCL_SERVER);
1661                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
1662                 repbody->eadatasize = 0;
1663                 repbody->aclsize = 0;
1664         } else {
1665                 rc = err_serious(rc);
1666         }
1667
1668         med = &req->rq_export->exp_mdt_data;
1669         cfs_spin_lock(&med->med_open_lock);
1670         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1671         if (mdt_mfd_closed(mfd)) {
1672                 cfs_spin_unlock(&med->med_open_lock);
1673                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
1674                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
1675                        info->mti_ioepoch->handle.cookie);
1676                 /** not serious error since bug 3633 */
1677                 rc = -ESTALE;
1678         } else {
1679                 class_handle_unhash(&mfd->mfd_handle);
1680                 cfs_list_del_init(&mfd->mfd_list);
1681                 cfs_spin_unlock(&med->med_open_lock);
1682
1683                 /* Do not lose object before last unlink. */
1684                 o = mfd->mfd_object;
1685                 mdt_object_get(info->mti_env, o);
1686                 ret = mdt_mfd_close(info, mfd);
1687                 if (repbody != NULL)
1688                         rc = mdt_handle_last_unlink(info, o, ma);
1689                 mdt_empty_transno(info, rc);
1690                 mdt_object_put(info->mti_env, o);
1691         }
1692         if (repbody != NULL) {
1693                 mdt_client_compatibility(info);
1694                 rc = mdt_fix_reply(info);
1695         }
1696
1697         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
1698                 RETURN(err_serious(-ENOMEM));
1699
1700         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
1701                                  OBD_FAIL_MDS_CLOSE_NET_REP))
1702                 info->mti_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
1703         RETURN(rc ? rc : ret);
1704 }
1705
1706 /**
1707  * DONE_WRITING rpc handler.
1708  *
1709  * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
1710  * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
1711  * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
1712  * skip attributes and reconstruct the reply here.
1713  */
1714 int mdt_done_writing(struct mdt_thread_info *info)
1715 {
1716         struct ptlrpc_request   *req = mdt_info_req(info);
1717         struct mdt_body         *repbody = NULL;
1718         struct mdt_export_data  *med;
1719         struct mdt_file_data    *mfd;
1720         int rc;
1721         ENTRY;
1722
1723         rc = req_capsule_server_pack(info->mti_pill);
1724         if (rc)
1725                 RETURN(err_serious(rc));
1726
1727         repbody = req_capsule_server_get(info->mti_pill,
1728                                          &RMF_MDT_BODY);
1729         repbody->eadatasize = 0;
1730         repbody->aclsize = 0;
1731
1732         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
1733         rc = mdt_close_unpack(info);
1734         if (rc)
1735                 RETURN(err_serious(rc));
1736
1737         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL))
1738                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1739
1740         med = &info->mti_exp->exp_mdt_data;
1741         cfs_spin_lock(&med->med_open_lock);
1742         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1743         if (mfd == NULL) {
1744                 cfs_spin_unlock(&med->med_open_lock);
1745                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
1746                        ": cookie = "LPX64" ioepoch = "LPU64"\n",
1747                        PFID(info->mti_rr.rr_fid1),
1748                        info->mti_ioepoch->handle.cookie,
1749                        info->mti_ioepoch->ioepoch);
1750                 /* If this is a replay, reconstruct the transno. */
1751                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1752                         rc = info->mti_ioepoch->flags & MF_SOM_AU ?
1753                              -EAGAIN : 0;
1754                         mdt_empty_transno(info, rc);
1755                         RETURN(rc);
1756                 }
1757                 RETURN(-ESTALE);
1758         }
1759
1760         LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
1761                 mfd->mfd_mode == MDS_FMODE_TRUNC);
1762         class_handle_unhash(&mfd->mfd_handle);
1763         cfs_list_del_init(&mfd->mfd_list);
1764         cfs_spin_unlock(&med->med_open_lock);
1765
1766         /* Set EPOCH CLOSE flag if not set by client. */
1767         info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
1768         info->mti_attr.ma_valid = 0;
1769
1770         info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
1771         OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1772         if (info->mti_attr.ma_lmm == NULL)
1773                 RETURN(-ENOMEM);
1774
1775         rc = mdt_mfd_close(info, mfd);
1776
1777         OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1778         mdt_empty_transno(info, rc);
1779         RETURN(rc);
1780 }