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         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         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         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                 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                 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         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         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         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         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         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         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         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         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         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         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         mutex_lock(&o->mot_ioepoch_mutex);
 517         rc = o->mot_writecount;
 518         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         mutex_lock(&o->mot_ioepoch_mutex);
 527         if (o->mot_writecount < 0)
 528                 rc = -ETXTBSY;
 529         else
 530                 o->mot_writecount++;
 531         mutex_unlock(&o->mot_ioepoch_mutex);
 532         RETURN(rc);
 533 }
 534
 535 void mdt_write_put(struct mdt_object *o)
 536 {
 537         ENTRY;
 538         mutex_lock(&o->mot_ioepoch_mutex);
 539         o->mot_writecount--;
 540         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         mutex_lock(&o->mot_ioepoch_mutex);
 549         if (o->mot_writecount > 0)
 550                 rc = -ETXTBSY;
 551         else
 552                 o->mot_writecount--;
 553         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         mutex_lock(&o->mot_ioepoch_mutex);
 561         o->mot_writecount++;
 562         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         spin_lock(&mdt->mdt_lut.lut_translock);
 580         if (rc != 0) {
 581                 if (info->mti_transno != 0) {
 582                         struct obd_export *exp = req->rq_export;
 583
 584                         CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
 585                                 mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name,
 586                                 info->mti_transno,
 587                                 libcfs_nid2str(exp->exp_connection->c_peer.nid),
 588                                 rc);
 589                         RETURN_EXIT;
 590                 }
 591         } else if (info->mti_transno == 0) {
 592                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
 593         } else {
 594                 /* should be replay */
 595                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 596                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 597         }
 598         spin_unlock(&mdt->mdt_lut.lut_translock);
 599
 600         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 601                         info->mti_transno,
 602                         req->rq_export->exp_obd->obd_last_committed);
 603
 604         req->rq_transno = info->mti_transno;
 605         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 606
 607         /* update lcd in memory only for resent cases */
 608         ted = &req->rq_export->exp_target_data;
 609         LASSERT(ted);
 610         mutex_lock(&ted->ted_lcd_lock);
 611         lcd = ted->ted_lcd;
 612         if (info->mti_transno < lcd->lcd_last_transno &&
 613             info->mti_transno != 0) {
 614                 /* This should happen during replay. Do not update
 615                  * last rcvd info if replay req transno < last transno,
 616                  * otherwise the following resend(after replay) can not
 617                  * be checked correctly by xid */
 618                 mutex_unlock(&ted->ted_lcd_lock);
 619                 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
 620                         mdt->mdt_md_dev.md_lu_dev.ld_obd->obd_name,
 621                         info->mti_transno, lcd->lcd_last_transno);
 622                 RETURN_EXIT;
 623         }
 624
 625         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 626             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 627                 if (info->mti_transno != 0)
 628                         lcd->lcd_last_close_transno = info->mti_transno;
 629                 lcd->lcd_last_close_xid = req->rq_xid;
 630                 lcd->lcd_last_close_result = rc;
 631         } else {
 632                 /* VBR: save versions in last_rcvd for reconstruct. */
 633                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 634                 if (pre_versions) {
 635                         lcd->lcd_pre_versions[0] = pre_versions[0];
 636                         lcd->lcd_pre_versions[1] = pre_versions[1];
 637                         lcd->lcd_pre_versions[2] = pre_versions[2];
 638                         lcd->lcd_pre_versions[3] = pre_versions[3];
 639                 }
 640                 if (info->mti_transno != 0)
 641                         lcd->lcd_last_transno = info->mti_transno;
 642
 643                 lcd->lcd_last_xid = req->rq_xid;
 644                 lcd->lcd_last_result = rc;
 645                 lcd->lcd_last_data = info->mti_opdata;
 646         }
 647         mutex_unlock(&ted->ted_lcd_lock);
 648
 649         EXIT;
 650 }
 651
 652 void mdt_mfd_set_mode(struct mdt_file_data *mfd, int mode)
 653 {
 654         LASSERT(mfd != NULL);
 655
 656         CDEBUG(D_HA, "Change mfd %p mode 0x%x->0x%x\n",
 657                mfd, (unsigned int)mfd->mfd_mode, (unsigned int)mode);
 658
 659         mfd->mfd_mode = mode;
 660 }
 661
 662 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 663                         struct mdt_object *o, __u64 flags, int created)
 664 {
 665         struct ptlrpc_request   *req = mdt_info_req(info);
 666         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 667         struct mdt_file_data    *mfd;
 668         struct md_attr          *ma  = &info->mti_attr;
 669         struct lu_attr          *la  = &ma->ma_attr;
 670         struct mdt_body         *repbody;
 671         int                      rc = 0, isdir, isreg;
 672         ENTRY;
 673
 674         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 675
 676         isreg = S_ISREG(la->la_mode);
 677         isdir = S_ISDIR(la->la_mode);
 678         if (isreg && !(ma->ma_valid & MA_LOV)) {
 679                 /*
 680                  * No EA, check whether it is will set regEA and dirEA since in
 681                  * above attr get, these size might be zero, so reset it, to
 682                  * retrieve the MD after create obj.
 683                  */
 684                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 685                                                        &RMF_MDT_MD,
 686                                                        RCL_SERVER);
 687                 /* in replay case, p == NULL */
 688                 rc = mdt_create_data(info, p, o);
 689                 if (rc)
 690                         RETURN(rc);
 691         }
 692
 693         CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
 694                ma->ma_valid, ma->ma_lmm_size);
 695
 696         if (ma->ma_valid & MA_LOV) {
 697                 LASSERT(ma->ma_lmm_size != 0);
 698                 repbody->eadatasize = ma->ma_lmm_size;
 699                 if (isdir)
 700                         repbody->valid |= OBD_MD_FLDIREA;
 701                 else
 702                         repbody->valid |= OBD_MD_FLEASIZE;
 703         }
 704
 705         if (flags & FMODE_WRITE) {
 706                 rc = mdt_write_get(o);
 707                 if (rc == 0) {
 708                         mdt_ioepoch_open(info, o, created);
 709                         repbody->ioepoch = o->mot_ioepoch;
 710                 }
 711         } else if (flags & MDS_FMODE_EXEC) {
 712                 rc = mdt_write_deny(o);
 713         }
 714         if (rc)
 715                 RETURN(rc);
 716
 717         rc = mo_open(info->mti_env, mdt_object_child(o),
 718                      created ? flags | MDS_OPEN_CREATED : flags);
 719         if (rc)
 720                 GOTO(err_out, rc);
 721
 722         mfd = mdt_mfd_new();
 723         if (mfd != NULL) {
 724                 /*
 725                  * Keep a reference on this object for this open, and is
 726                  * released by mdt_mfd_close().
 727                  */
 728                 mdt_object_get(info->mti_env, o);
 729
 730                 /*
 731                  * @flags is always not zero. At least it should be FMODE_READ,
 732                  * FMODE_WRITE or MDS_FMODE_EXEC.
 733                  */
 734                 LASSERT(flags != 0);
 735
 736                 /* Open handling. */
 737                 mdt_mfd_set_mode(mfd, flags);
 738
 739                 mfd->mfd_object = o;
 740                 mfd->mfd_xid = req->rq_xid;
 741
 742                 /* replay handle */
 743                 if (req_is_replay(req)) {
 744                         struct mdt_file_data *old_mfd;
 745                         /* Check wheather old cookie already exist in
 746                          * the list, becasue when do recovery, client
 747                          * might be disconnected from server, and
 748                          * restart replay, so there maybe some orphan
 749                          * mfd here, we should remove them */
 750                         LASSERT(info->mti_rr.rr_handle != NULL);
 751                         old_mfd = mdt_handle2mfd(info, info->mti_rr.rr_handle);
 752                         if (old_mfd) {
 753                                 CDEBUG(D_HA, "del orph mfd %p fid=("DFID") "
 754                                        "cookie=" LPX64"\n", mfd,
 755                                        PFID(mdt_object_fid(mfd->mfd_object)),
 756                                        info->mti_rr.rr_handle->cookie);
 757                                 spin_lock(&med->med_open_lock);
 758                                 class_handle_unhash(&old_mfd->mfd_handle);
 759                                 cfs_list_del_init(&old_mfd->mfd_list);
 760                                 spin_unlock(&med->med_open_lock);
 761                                 /* no attr update for that close */
 762                                 la->la_valid = 0;
 763                                 ma->ma_valid |= MA_FLAGS;
 764                                 ma->ma_attr_flags |= MDS_RECOV_OPEN;
 765                                 mdt_mfd_close(info, old_mfd);
 766                                 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 767                                 ma->ma_valid &= ~MA_FLAGS;
 768                         }
 769                         CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
 770                                info->mti_rr.rr_handle->cookie);
 771                         mfd->mfd_old_handle.cookie =
 772                                                 info->mti_rr.rr_handle->cookie;
 773                 }
 774                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 775
 776                 if (req->rq_export->exp_disconnected) {
 777                         spin_lock(&med->med_open_lock);
 778                         class_handle_unhash(&mfd->mfd_handle);
 779                         cfs_list_del_init(&mfd->mfd_list);
 780                         spin_unlock(&med->med_open_lock);
 781                         mdt_mfd_close(info, mfd);
 782                 } else {
 783                         spin_lock(&med->med_open_lock);
 784                         cfs_list_add(&mfd->mfd_list, &med->med_open_head);
 785                         spin_unlock(&med->med_open_lock);
 786                 }
 787
 788                 mdt_empty_transno(info, rc);
 789         } else {
 790                 GOTO(err_out, rc = -ENOMEM);
 791         }
 792
 793         RETURN(rc);
 794
 795 err_out:
 796         if (flags & FMODE_WRITE)
 797                         /* XXX We also need to close io epoch here.
 798                          * See LU-1220 - green */
 799                 mdt_write_put(o);
 800         else if (flags & FMODE_EXEC)
 801                 mdt_write_allow(o);
 802         return rc;
 803 }
 804
 805 int mdt_finish_open(struct mdt_thread_info *info,
 806                     struct mdt_object *p, struct mdt_object *o,
 807                     __u64 flags, int created, struct ldlm_reply *rep)
 808 {
 809         struct ptlrpc_request   *req = mdt_info_req(info);
 810         struct obd_export       *exp = req->rq_export;
 811         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 812         struct md_attr          *ma  = &info->mti_attr;
 813         struct lu_attr          *la  = &ma->ma_attr;
 814         struct mdt_file_data    *mfd;
 815         struct mdt_body         *repbody;
 816         int                      rc = 0;
 817         int                      isreg, isdir, islnk;
 818         cfs_list_t              *t;
 819         ENTRY;
 820
 821         LASSERT(ma->ma_valid & MA_INODE);
 822
 823         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 824
 825         isreg = S_ISREG(la->la_mode);
 826         isdir = S_ISDIR(la->la_mode);
 827         islnk = S_ISLNK(la->la_mode);
 828         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 829
 830         /* LU-2275, simulate broken behaviour (esp. prevalent in
 831          * pre-2.4 servers where a very strange reply is sent on error
 832          * that looks like it was actually almost succesful and a failure at the
 833          * same time */
 834         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 835                 mdt_set_disposition(info, rep, DISP_OPEN_LOCK | \
 836                                     DISP_OPEN_OPEN | DISP_LOOKUP_NEG | \
 837                                     DISP_LOOKUP_POS);
 838                 RETURN(-ENOENT);
 839         }
 840
 841         if (exp_connect_rmtclient(exp)) {
 842                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 843
 844                 rc = mdt_pack_remote_perm(info, o, buf);
 845                 if (rc) {
 846                         repbody->valid &= ~OBD_MD_FLRMTPERM;
 847                         repbody->aclsize = 0;
 848                 } else {
 849                         repbody->valid |= OBD_MD_FLRMTPERM;
 850                         repbody->aclsize = sizeof(struct mdt_remote_perm);
 851                 }
 852         }
 853 #ifdef CONFIG_FS_POSIX_ACL
 854         else if (exp->exp_connect_flags & OBD_CONNECT_ACL) {
 855                 const struct lu_env *env = info->mti_env;
 856                 struct md_object *next = mdt_object_child(o);
 857                 struct lu_buf *buf = &info->mti_buf;
 858
 859                 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 860                 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
 861                                                    RCL_SERVER);
 862                 if (buf->lb_len > 0) {
 863                         rc = mo_xattr_get(env, next, buf,
 864                                           XATTR_NAME_ACL_ACCESS);
 865                         if (rc < 0) {
 866                                 if (rc == -ENODATA) {
 867                                         repbody->aclsize = 0;
 868                                         repbody->valid |= OBD_MD_FLACL;
 869                                         rc = 0;
 870                                 } else if (rc == -EOPNOTSUPP) {
 871                                         rc = 0;
 872                                 } else {
 873                                         CERROR("got acl size: %d\n", rc);
 874                                 }
 875                         } else {
 876                                 repbody->aclsize = rc;
 877                                 repbody->valid |= OBD_MD_FLACL;
 878                                 rc = 0;
 879                         }
 880                 }
 881         }
 882 #endif
 883
 884         if (info->mti_mdt->mdt_opts.mo_mds_capa &&
 885             exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) {
 886                 struct lustre_capa *capa;
 887
 888                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
 889                 LASSERT(capa);
 890                 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
 891                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 892                 if (rc)
 893                         RETURN(rc);
 894                 repbody->valid |= OBD_MD_FLMDSCAPA;
 895         }
 896         if (info->mti_mdt->mdt_opts.mo_oss_capa &&
 897             exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA &&
 898             S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) {
 899                 struct lustre_capa *capa;
 900
 901                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
 902                 LASSERT(capa);
 903                 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
 904                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 905                 if (rc)
 906                         RETURN(rc);
 907                 repbody->valid |= OBD_MD_FLOSSCAPA;
 908         }
 909
 910         /*
 911          * If we are following a symlink, don't open; and do not return open
 912          * handle for special nodes as client required.
 913          */
 914         if (islnk || (!isreg && !isdir &&
 915             (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH))) {
 916                 lustre_msg_set_transno(req->rq_repmsg, 0);
 917                 RETURN(0);
 918         }
 919
 920         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 921
 922         /*
 923          * We need to return the existing object's fid back, so it is done here,
 924          * after preparing the reply.
 925          */
 926         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 927                 RETURN(-EEXIST);
 928
 929         /* This can't be done earlier, we need to return reply body */
 930         if (isdir) {
 931                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 932                         /* We are trying to create or write an existing dir. */
 933                         RETURN(-EISDIR);
 934                 }
 935         } else if (flags & MDS_OPEN_DIRECTORY)
 936                 RETURN(-ENOTDIR);
 937
 938         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 939                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 940                 RETURN(-EAGAIN);
 941         }
 942
 943         mfd = NULL;
 944         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 945                 spin_lock(&med->med_open_lock);
 946                 cfs_list_for_each(t, &med->med_open_head) {
 947                         mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
 948                         if (mfd->mfd_xid == req->rq_xid)
 949                                 break;
 950                         mfd = NULL;
 951                 }
 952                 spin_unlock(&med->med_open_lock);
 953
 954                 if (mfd != NULL) {
 955                         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 956                         /*set repbody->ea_size for resent case*/
 957                         if (ma->ma_valid & MA_LOV) {
 958                                 LASSERT(ma->ma_lmm_size != 0);
 959                                 repbody->eadatasize = ma->ma_lmm_size;
 960                                 if (isdir)
 961                                         repbody->valid |= OBD_MD_FLDIREA;
 962                                 else
 963                                         repbody->valid |= OBD_MD_FLEASIZE;
 964                         }
 965                         RETURN(0);
 966                 }
 967         }
 968
 969         rc = mdt_mfd_open(info, p, o, flags, created);
 970         RETURN(rc);
 971 }
 972
 973 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
 974                              struct lsd_client_data *lcd);
 975
 976 void mdt_reconstruct_open(struct mdt_thread_info *info,
 977                           struct mdt_lock_handle *lhc)
 978 {
 979         const struct lu_env *env = info->mti_env;
 980         struct mdt_device       *mdt  = info->mti_mdt;
 981         struct req_capsule      *pill = info->mti_pill;
 982         struct ptlrpc_request   *req  = mdt_info_req(info);
 983         struct tg_export_data   *ted  = &req->rq_export->exp_target_data;
 984         struct lsd_client_data  *lcd  = ted->ted_lcd;
 985         struct md_attr          *ma   = &info->mti_attr;
 986         struct mdt_reint_record *rr   = &info->mti_rr;
 987         __u32                   flags = info->mti_spec.sp_cr_flags;
 988         struct ldlm_reply       *ldlm_rep;
 989         struct mdt_object       *parent;
 990         struct mdt_object       *child;
 991         struct mdt_body         *repbody;
 992         int                      rc;
 993         ENTRY;
 994
 995         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 996         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 997         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 998
 999         ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
1000         ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
1001                                                RCL_SERVER);
1002         ma->ma_need = MA_INODE;
1003         if (ma->ma_lmm_size > 0)
1004                 ma->ma_need |= MA_LOV;
1005
1006         ma->ma_valid = 0;
1007
1008         mdt_req_from_lcd(req, lcd);
1009         mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1010
1011         CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1012                ldlm_rep->lock_policy_res1, req->rq_status);
1013
1014         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1015             req->rq_status != 0)
1016                 /* We did not create successfully, return error to client. */
1017                 GOTO(out, rc = req->rq_status);
1018
1019         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1020                 struct obd_export *exp = req->rq_export;
1021                 /*
1022                  * We failed after creation, but we do not know in which step
1023                  * we failed. So try to check the child object.
1024                  */
1025                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1026                 if (IS_ERR(parent)) {
1027                         rc = PTR_ERR(parent);
1028                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
1029                                       " Evicting client %s with export %s.\n",
1030                                       PFID(rr->rr_fid1), rc,
1031                                       obd_uuid2str(&exp->exp_client_uuid),
1032                                       obd_export_nid2str(exp));
1033                         mdt_export_evict(exp);
1034                         RETURN_EXIT;
1035                 }
1036                 child = mdt_object_find(env, mdt, rr->rr_fid2);
1037                 if (IS_ERR(child)) {
1038                         rc = PTR_ERR(child);
1039                         LCONSOLE_WARN("Child "DFID" lookup error %d."
1040                                       " Evicting client %s with export %s.\n",
1041                                       PFID(mdt_object_fid(child)), rc,
1042                                       obd_uuid2str(&exp->exp_client_uuid),
1043                                       obd_export_nid2str(exp));
1044                         mdt_object_put(env, parent);
1045                         mdt_export_evict(exp);
1046                         RETURN_EXIT;
1047                 }
1048                 rc = mdt_object_exists(child);
1049                 if (rc > 0) {
1050
1051                         mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA);
1052                         rc = mdt_attr_get_complex(info, child, ma);
1053                         if (rc == 0)
1054                               rc = mdt_finish_open(info, parent, child,
1055                                                    flags, 1, ldlm_rep);
1056                 } else if (rc < 0) {
1057                         /* the child object was created on remote server */
1058                         repbody->fid1 = *rr->rr_fid2;
1059                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1060                         rc = 0;
1061                 } else if (rc == 0) {
1062                         /* the child does not exist, we should do regular open */
1063                         mdt_object_put(env, parent);
1064                         mdt_object_put(env, child);
1065                         GOTO(regular_open, 0);
1066                 }
1067                 mdt_object_put(env, parent);
1068                 mdt_object_put(env, child);
1069                 GOTO(out, rc);
1070         } else {
1071 regular_open:
1072                 /* We did not try to create, so we are a pure open */
1073                 rc = mdt_reint_open(info, lhc);
1074         }
1075
1076         EXIT;
1077 out:
1078         req->rq_status = rc;
1079         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1080         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1081 }
1082
1083 int mdt_open_by_fid(struct mdt_thread_info* info,
1084                     struct ldlm_reply *rep)
1085 {
1086         __u32                    flags = info->mti_spec.sp_cr_flags;
1087         struct mdt_reint_record *rr = &info->mti_rr;
1088         struct md_attr          *ma = &info->mti_attr;
1089         struct mdt_object       *o;
1090         int                      rc;
1091         ENTRY;
1092
1093         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1094         if (IS_ERR(o))
1095                 RETURN(rc = PTR_ERR(o));
1096
1097         rc = mdt_object_exists(o);
1098         if (rc > 0) {
1099                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1100                                                 DISP_LOOKUP_EXECD |
1101                                                 DISP_LOOKUP_POS));
1102
1103                 rc = mdt_attr_get_complex(info, o, ma);
1104                 if (rc == 0)
1105                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1106         } else if (rc == 0) {
1107                 rc = -ENOENT;
1108         } else  {
1109                 /* the child object was created on remote server */
1110                 struct mdt_body *repbody;
1111                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1112                 repbody->fid1 = *rr->rr_fid2;
1113                 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1114                 rc = 0;
1115         }
1116
1117         mdt_object_put(info->mti_env, o);
1118         RETURN(rc);
1119 }
1120
1121 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1122                          struct mdt_lock_handle *lhc)
1123 {
1124         const struct lu_env     *env   = info->mti_env;
1125         struct mdt_device       *mdt   = info->mti_mdt;
1126         __u32                    flags = info->mti_spec.sp_cr_flags;
1127         struct mdt_reint_record *rr    = &info->mti_rr;
1128         struct md_attr          *ma    = &info->mti_attr;
1129         struct mdt_object       *parent= NULL;
1130         struct mdt_object       *o;
1131         int                      rc;
1132         ldlm_mode_t              lm;
1133         ENTRY;
1134
1135         if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1136                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1137                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1138                         if (IS_ERR(parent)) {
1139                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1140                                        " for anonymous created %ld, try to"
1141                                        " use server-side parent.\n",
1142                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1143                                 parent = NULL;
1144                         }
1145                 }
1146                 if (parent == NULL)
1147                         ma->ma_need |= MA_PFID;
1148         }
1149
1150         o = mdt_object_find(env, mdt, rr->rr_fid2);
1151         if (IS_ERR(o))
1152                 RETURN(rc = PTR_ERR(o));
1153
1154         rc = mdt_object_exists(o);
1155         if (rc == 0) {
1156                 mdt_set_disposition(info, rep, (DISP_LOOKUP_EXECD |
1157                                     DISP_LOOKUP_NEG));
1158                 GOTO(out, rc = -ENOENT);
1159         } else if (rc < 0) {
1160                 CERROR("NFS remote open shouldn't happen.\n");
1161                 GOTO(out, rc);
1162         }
1163         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1164                                         DISP_LOOKUP_EXECD |
1165                                         DISP_LOOKUP_POS));
1166
1167         if (flags & FMODE_WRITE)
1168                 lm = LCK_CW;
1169         else if (flags & MDS_FMODE_EXEC)
1170                 lm = LCK_PR;
1171         else
1172                 lm = LCK_CR;
1173
1174         mdt_lock_handle_init(lhc);
1175         mdt_lock_reg_init(lhc, lm);
1176         rc = mdt_object_lock(info, o, lhc,
1177                              MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1178                              MDT_CROSS_LOCK);
1179         if (rc)
1180                 GOTO(out, rc);
1181
1182         rc = mdt_attr_get_complex(info, o, ma);
1183         if (rc)
1184                 GOTO(out, rc);
1185
1186         if (ma->ma_valid & MA_PFID) {
1187                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1188                 if (IS_ERR(parent)) {
1189                         CDEBUG(D_INODE, "Fail to find parent "DFID
1190                                " for anonymous created %ld, try to"
1191                                " use system default.\n",
1192                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1193                         parent = NULL;
1194                 }
1195         }
1196
1197         if (flags & MDS_OPEN_LOCK)
1198                 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1199         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1200
1201         if (!(flags & MDS_OPEN_LOCK) || rc)
1202                 mdt_object_unlock(info, o, lhc, 1);
1203
1204         GOTO(out, rc);
1205 out:
1206         mdt_object_put(env, o);
1207         if (parent != NULL)
1208                 mdt_object_put(env, parent);
1209         return rc;
1210 }
1211
1212 int mdt_pin(struct mdt_thread_info* info)
1213 {
1214         ENTRY;
1215         RETURN(err_serious(-EOPNOTSUPP));
1216 }
1217
1218 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1219 int mdt_cross_open(struct mdt_thread_info* info,
1220                    const struct lu_fid *fid,
1221                    struct ldlm_reply *rep, __u32 flags)
1222 {
1223         struct md_attr    *ma = &info->mti_attr;
1224         struct mdt_object *o;
1225         int                rc;
1226         ENTRY;
1227
1228         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1229         if (IS_ERR(o))
1230                 RETURN(rc = PTR_ERR(o));
1231
1232         rc = mdt_object_exists(o);
1233         if (rc > 0) {
1234                 /* Do permission check for cross-open. */
1235                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o),
1236                                    NULL, flags | MDS_OPEN_CROSS);
1237                 if (rc)
1238                         goto out;
1239
1240                 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1241                 rc = mdt_attr_get_complex(info, o, ma);
1242                 if (rc == 0)
1243                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1244         } else if (rc == 0) {
1245                 /*
1246                  * Something is wrong here. lookup was positive but there is
1247                  * no object!
1248                  */
1249                 CERROR("Cross-ref object doesn't exist!\n");
1250                 rc = -EFAULT;
1251         } else  {
1252                 /* Something is wrong here, the object is on another MDS! */
1253                 CERROR("The object isn't on this server! FLD error?\n");
1254                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1255                                 &o->mot_obj.mo_lu,
1256                                 "Object isn't on this server! FLD error?\n");
1257
1258                 rc = -EFAULT;
1259         }
1260
1261 out:
1262         mdt_object_put(info->mti_env, o);
1263         RETURN(rc);
1264 }
1265
1266 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1267 {
1268         struct mdt_device       *mdt = info->mti_mdt;
1269         struct ptlrpc_request   *req = mdt_info_req(info);
1270         struct mdt_object       *parent;
1271         struct mdt_object       *child;
1272         struct mdt_lock_handle  *lh;
1273         struct ldlm_reply       *ldlm_rep;
1274         struct mdt_body         *repbody;
1275         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1276         struct md_attr          *ma = &info->mti_attr;
1277         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1278         struct mdt_reint_record *rr = &info->mti_rr;
1279         struct lu_name          *lname;
1280         int                      result, rc;
1281         int                      created = 0;
1282         __u32                    msg_flags;
1283         ENTRY;
1284
1285         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1286                                (obd_timeout + 1) / 4);
1287
1288         mdt_counter_incr(req, LPROC_MDT_OPEN);
1289         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1290
1291         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1292         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1293                                                RCL_SERVER);
1294         ma->ma_need = MA_INODE;
1295         if (ma->ma_lmm_size > 0)
1296                 ma->ma_need |= MA_LOV;
1297
1298         ma->ma_valid = 0;
1299
1300         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1301         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1302
1303         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1304                 CERROR("file join is not supported anymore.\n");
1305                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1306         }
1307         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1308
1309         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1310             info->mti_spec.u.sp_ea.eadata == NULL)
1311                 GOTO(out, result = err_serious(-EINVAL));
1312
1313         CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1314                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1315                PFID(rr->rr_fid1), rr->rr_name,
1316                PFID(rr->rr_fid2), create_flags,
1317                ma->ma_attr.la_mode, msg_flags);
1318
1319         if (req_is_replay(req) ||
1320             (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1321                 /* This is a replay request or from liblustre with ea. */
1322                 result = mdt_open_by_fid(info, ldlm_rep);
1323
1324                 if (result != -ENOENT) {
1325                         if (req->rq_export->exp_libclient &&
1326                             create_flags & MDS_OPEN_HAS_EA)
1327                                 GOTO(out, result = 0);
1328                         GOTO(out, result);
1329                 }
1330                 /* We didn't find the correct object, so we need to re-create it
1331                  * via a regular replay. */
1332                 if (!(create_flags & MDS_OPEN_CREAT)) {
1333                         DEBUG_REQ(D_ERROR, req,
1334                                   "OPEN & CREAT not in open replay/by_fid.");
1335                         GOTO(out, result = -EFAULT);
1336                 }
1337                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1338         } else if ((rr->rr_namelen == 0 && !info->mti_cross_ref &&
1339                     create_flags & MDS_OPEN_LOCK) ||
1340                    (create_flags & MDS_OPEN_BY_FID)) {
1341                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1342                 if (result != -ENOENT && !(create_flags & MDS_OPEN_CREAT))
1343                         GOTO(out, result);
1344                 if (unlikely(rr->rr_namelen == 0))
1345                         GOTO(out, result = -EINVAL);
1346                 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1347         }
1348
1349         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1350                 GOTO(out, result = err_serious(-ENOMEM));
1351
1352         mdt_set_disposition(info, ldlm_rep,
1353                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1354
1355         if (info->mti_cross_ref) {
1356                 /* This is cross-ref open */
1357                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1358                 result = mdt_cross_open(info, rr->rr_fid1, ldlm_rep,
1359                                         create_flags);
1360                 GOTO(out, result);
1361         }
1362
1363         lh = &info->mti_lh[MDT_LH_PARENT];
1364         mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1365                           LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1366
1367         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1368                                       MDS_INODELOCK_UPDATE);
1369         if (IS_ERR(parent))
1370                 GOTO(out, result = PTR_ERR(parent));
1371
1372         /* get and check version of parent */
1373         result = mdt_version_get_check(info, parent, 0);
1374         if (result)
1375                 GOTO(out_parent, result);
1376
1377         fid_zero(child_fid);
1378
1379         lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1380         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1381                             lname, child_fid, &info->mti_spec);
1382         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1383                  "looking for "DFID"/%s, result fid="DFID"\n",
1384                  PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1385
1386         if (result != 0 && result != -ENOENT && result != -ESTALE)
1387                 GOTO(out_parent, result);
1388
1389         if (result == -ENOENT || result == -ESTALE) {
1390                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1391                 if (result == -ESTALE) {
1392                         /*
1393                          * -ESTALE means the parent is a dead(unlinked) dir, so
1394                          * it should return -ENOENT to in accordance with the
1395                          * original mds implementaion.
1396                          */
1397                         GOTO(out_parent, result = -ENOENT);
1398                 }
1399                 if (!(create_flags & MDS_OPEN_CREAT))
1400                         GOTO(out_parent, result);
1401                 *child_fid = *info->mti_rr.rr_fid2;
1402                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1403                          PFID(child_fid));
1404                 /* In the function below, .hs_keycmp resolves to
1405                  * lu_obj_hop_keycmp() */
1406                 /* coverity[overrun-buffer-val] */
1407                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1408         } else {
1409                 /*
1410                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1411                  * return FID back in that case.
1412                  */
1413                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1414                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1415         }
1416         if (IS_ERR(child))
1417                 GOTO(out_parent, result = PTR_ERR(child));
1418
1419         /** check version of child  */
1420         rc = mdt_version_get_check(info, child, 1);
1421         if (rc)
1422                 GOTO(out_child, result = rc);
1423
1424         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1425         if (result == -ENOENT) {
1426                 if (mdt_object_obf(parent))
1427                         GOTO(out_child, result = -EPERM);
1428
1429                 /* save versions in reply */
1430                 mdt_version_get_save(info, parent, 0);
1431                 mdt_version_get_save(info, child, 1);
1432
1433                 /* version of child will be changed */
1434                 info->mti_mos = child;
1435
1436                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1437                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1438
1439                 /* Let lower layers know what is lock mode on directory. */
1440                 info->mti_spec.sp_cr_mode =
1441                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1442
1443                 /*
1444                  * Do not perform lookup sanity check. We know that name does
1445                  * not exist.
1446                  */
1447                 info->mti_spec.sp_cr_lookup = 0;
1448                 info->mti_spec.sp_feat = &dt_directory_features;
1449
1450                 result = mdo_create(info->mti_env,
1451                                     mdt_object_child(parent),
1452                                     lname,
1453                                     mdt_object_child(child),
1454                                     &info->mti_spec,
1455                                     &info->mti_attr);
1456                 if (result == -ERESTART) {
1457                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1458                         GOTO(out_child, result);
1459                 } else {
1460
1461                         /* XXX: we should call this once, see few lines below */
1462                         if (result == 0)
1463                                 result = mdt_attr_get_complex(info, child, ma);
1464
1465                         if (result != 0)
1466                                 GOTO(out_child, result);
1467                 }
1468                 created = 1;
1469         } else {
1470                 /* We have to get attr & lov ea for this object */
1471                 result = mdt_attr_get_complex(info, child, ma);
1472                 /*
1473                  * The object is on remote node, return its FID for remote open.
1474                  */
1475                 if (result == -EREMOTE) {
1476                         /*
1477                          * Check if this lock already was sent to client and
1478                          * this is resent case. For resent case do not take lock
1479                          * again, use what is already granted.
1480                          */
1481                         LASSERT(lhc != NULL);
1482
1483                         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1484                                 struct ldlm_lock *lock;
1485
1486                                 LASSERT(msg_flags & MSG_RESENT);
1487
1488                                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1489                                 if (!lock) {
1490                                         CERROR("Invalid lock handle "LPX64"\n",
1491                                                lhc->mlh_reg_lh.cookie);
1492                                         LBUG();
1493                                 }
1494                                 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1495                                                         &lock->l_resource->lr_name));
1496                                 LDLM_LOCK_PUT(lock);
1497                                 rc = 0;
1498                         } else {
1499                                 mdt_lock_handle_init(lhc);
1500                                 mdt_lock_reg_init(lhc, LCK_PR);
1501
1502                                 rc = mdt_object_lock(info, child, lhc,
1503                                                      MDS_INODELOCK_LOOKUP,
1504                                                      MDT_CROSS_LOCK);
1505                         }
1506                         repbody->fid1 = *mdt_object_fid(child);
1507                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1508                         if (rc != 0)
1509                                 result = rc;
1510                         GOTO(out_child, result);
1511                 }
1512         }
1513
1514         LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1515
1516         /* get openlock if this is not replay and if a client requested it */
1517         if (!req_is_replay(req) && create_flags & MDS_OPEN_LOCK) {
1518                 ldlm_mode_t lm;
1519
1520                 if (create_flags & FMODE_WRITE)
1521                         lm = LCK_CW;
1522                 else if (create_flags & MDS_FMODE_EXEC)
1523                         lm = LCK_PR;
1524                 else
1525                         lm = LCK_CR;
1526                 mdt_lock_handle_init(lhc);
1527                 mdt_lock_reg_init(lhc, lm);
1528                 rc = mdt_object_lock(info, child, lhc,
1529                                      MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1530                                      MDT_CROSS_LOCK);
1531                 if (rc) {
1532                         result = rc;
1533                         GOTO(out_child, result);
1534                 } else {
1535                         result = -EREMOTE;
1536                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1537                 }
1538         }
1539
1540         /* Try to open it now. */
1541         rc = mdt_finish_open(info, parent, child, create_flags,
1542                              created, ldlm_rep);
1543         if (rc) {
1544                 result = rc;
1545                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1546                         /* openlock was acquired and mdt_finish_open failed -
1547                            drop the openlock */
1548                         mdt_object_unlock(info, child, lhc, 1);
1549                 if (created) {
1550                         ma->ma_need = 0;
1551                         ma->ma_valid = 0;
1552                         ma->ma_cookie_size = 0;
1553                         rc = mdo_unlink(info->mti_env,
1554                                         mdt_object_child(parent),
1555                                         mdt_object_child(child),
1556                                         lname,
1557                                         &info->mti_attr);
1558                         if (rc != 0)
1559                                 CERROR("Error in cleanup of open\n");
1560                 }
1561         }
1562         EXIT;
1563 out_child:
1564         mdt_object_put(info->mti_env, child);
1565 out_parent:
1566         mdt_object_unlock_put(info, parent, lh, result || !created);
1567 out:
1568         if (result && result != -EREMOTE)
1569                 lustre_msg_set_transno(req->rq_repmsg, 0);
1570         return result;
1571 }
1572
1573 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
1574                                       MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
1575
1576 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1577 {
1578         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1579 }
1580
1581 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1582 {
1583         struct mdt_object *o = mfd->mfd_object;
1584         struct md_object *next = mdt_object_child(o);
1585         struct md_attr *ma = &info->mti_attr;
1586         int ret = MDT_IOEPOCH_CLOSED;
1587         int rc = 0;
1588         int mode;
1589         ENTRY;
1590
1591         mode = mfd->mfd_mode;
1592
1593         if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
1594                 mdt_write_put(o);
1595                 ret = mdt_ioepoch_close(info, o);
1596         } else if (mode & MDS_FMODE_EXEC) {
1597                 mdt_write_allow(o);
1598         } else if (mode & MDS_FMODE_EPOCH) {
1599                 ret = mdt_ioepoch_close(info, o);
1600         } else if (mode & MDS_FMODE_SOM) {
1601                 ret = mdt_som_au_close(info, o);
1602         }
1603
1604         /* Update atime on close only. */
1605         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1606             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1607                 /* Set the atime only. */
1608                 ma->ma_valid = MA_INODE;
1609                 ma->ma_attr.la_valid = LA_ATIME;
1610                 rc = mo_attr_set(info->mti_env, next, ma);
1611         }
1612
1613         ma->ma_need |= MA_INODE;
1614         ma->ma_valid &= ~MA_INODE;
1615
1616         if (!MFD_CLOSED(mode))
1617                 rc = mo_close(info->mti_env, next, ma, mode);
1618
1619         if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
1620                 struct mdt_export_data *med;
1621
1622                 /* The IOepoch is still opened or SOM update is needed.
1623                  * Put mfd back into the list. */
1624                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
1625                 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
1626                                       MDS_FMODE_EPOCH : MDS_FMODE_SOM);
1627
1628                 LASSERT(mdt_info_req(info));
1629                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
1630                 spin_lock(&med->med_open_lock);
1631                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
1632                 class_handle_hash_back(&mfd->mfd_handle);
1633                 spin_unlock(&med->med_open_lock);
1634
1635                 if (ret == MDT_IOEPOCH_OPENED) {
1636                         ret = 0;
1637                 } else {
1638                         ret = -EAGAIN;
1639                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
1640                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
1641                 }
1642         } else {
1643                 mdt_mfd_free(mfd);
1644                 mdt_object_put(info->mti_env, o);
1645         }
1646
1647         RETURN(rc ? rc : ret);
1648 }
1649
1650 int mdt_close(struct mdt_thread_info *info)
1651 {
1652         struct mdt_export_data *med;
1653         struct mdt_file_data   *mfd;
1654         struct mdt_object      *o;
1655         struct md_attr         *ma = &info->mti_attr;
1656         struct mdt_body        *repbody = NULL;
1657         struct ptlrpc_request  *req = mdt_info_req(info);
1658         int rc, ret = 0;
1659         ENTRY;
1660
1661         mdt_counter_incr(req, LPROC_MDT_CLOSE);
1662         /* Close may come with the Size-on-MDS update. Unpack it. */
1663         rc = mdt_close_unpack(info);
1664         if (rc)
1665                 RETURN(err_serious(rc));
1666
1667         LASSERT(info->mti_ioepoch);
1668
1669         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
1670                              info->mti_mdt->mdt_max_mdsize);
1671         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
1672                              info->mti_mdt->mdt_max_cookiesize);
1673         rc = req_capsule_server_pack(info->mti_pill);
1674         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
1675                 mdt_client_compatibility(info);
1676                 if (rc == 0)
1677                         mdt_fix_reply(info);
1678                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1679         }
1680
1681         /* Continue to close handle even if we can not pack reply */
1682         if (rc == 0) {
1683                 repbody = req_capsule_server_get(info->mti_pill,
1684                                                  &RMF_MDT_BODY);
1685                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
1686                                                     &RMF_MDT_MD);
1687                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
1688                                                        &RMF_MDT_MD,
1689                                                        RCL_SERVER);
1690                 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
1691                                                        &RMF_LOGCOOKIES);
1692                 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
1693                                                           &RMF_LOGCOOKIES,
1694                                                           RCL_SERVER);
1695                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
1696                 repbody->eadatasize = 0;
1697                 repbody->aclsize = 0;
1698         } else {
1699                 rc = err_serious(rc);
1700         }
1701
1702         med = &req->rq_export->exp_mdt_data;
1703         spin_lock(&med->med_open_lock);
1704         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1705         if (mdt_mfd_closed(mfd)) {
1706                 spin_unlock(&med->med_open_lock);
1707                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
1708                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
1709                        info->mti_ioepoch->handle.cookie);
1710                 /** not serious error since bug 3633 */
1711                 rc = -ESTALE;
1712         } else {
1713                 class_handle_unhash(&mfd->mfd_handle);
1714                 cfs_list_del_init(&mfd->mfd_list);
1715                 spin_unlock(&med->med_open_lock);
1716
1717                 /* Do not lose object before last unlink. */
1718                 o = mfd->mfd_object;
1719                 mdt_object_get(info->mti_env, o);
1720                 ret = mdt_mfd_close(info, mfd);
1721                 if (repbody != NULL)
1722                         rc = mdt_handle_last_unlink(info, o, ma);
1723                 mdt_empty_transno(info, rc);
1724                 mdt_object_put(info->mti_env, o);
1725         }
1726         if (repbody != NULL) {
1727                 mdt_client_compatibility(info);
1728                 rc = mdt_fix_reply(info);
1729         }
1730
1731         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
1732                 RETURN(err_serious(-ENOMEM));
1733
1734         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
1735                                  OBD_FAIL_MDS_CLOSE_NET_REP))
1736                 info->mti_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
1737         RETURN(rc ? rc : ret);
1738 }
1739
1740 /**
1741  * DONE_WRITING rpc handler.
1742  *
1743  * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
1744  * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
1745  * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
1746  * skip attributes and reconstruct the reply here.
1747  */
1748 int mdt_done_writing(struct mdt_thread_info *info)
1749 {
1750         struct ptlrpc_request   *req = mdt_info_req(info);
1751         struct mdt_body         *repbody = NULL;
1752         struct mdt_export_data  *med;
1753         struct mdt_file_data    *mfd;
1754         int rc;
1755         ENTRY;
1756
1757         rc = req_capsule_server_pack(info->mti_pill);
1758         if (rc)
1759                 RETURN(err_serious(rc));
1760
1761         repbody = req_capsule_server_get(info->mti_pill,
1762                                          &RMF_MDT_BODY);
1763         repbody->eadatasize = 0;
1764         repbody->aclsize = 0;
1765
1766         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
1767         rc = mdt_close_unpack(info);
1768         if (rc)
1769                 RETURN(err_serious(rc));
1770
1771         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL))
1772                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1773
1774         med = &info->mti_exp->exp_mdt_data;
1775         spin_lock(&med->med_open_lock);
1776         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1777         if (mfd == NULL) {
1778                 spin_unlock(&med->med_open_lock);
1779                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
1780                        ": cookie = "LPX64" ioepoch = "LPU64"\n",
1781                        PFID(info->mti_rr.rr_fid1),
1782                        info->mti_ioepoch->handle.cookie,
1783                        info->mti_ioepoch->ioepoch);
1784                 /* If this is a replay, reconstruct the transno. */
1785                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1786                         rc = info->mti_ioepoch->flags & MF_SOM_AU ?
1787                              -EAGAIN : 0;
1788                         mdt_empty_transno(info, rc);
1789                         RETURN(rc);
1790                 }
1791                 RETURN(-ESTALE);
1792         }
1793
1794         LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
1795                 mfd->mfd_mode == MDS_FMODE_TRUNC);
1796         class_handle_unhash(&mfd->mfd_handle);
1797         cfs_list_del_init(&mfd->mfd_list);
1798         spin_unlock(&med->med_open_lock);
1799
1800         /* Set EPOCH CLOSE flag if not set by client. */
1801         info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
1802         info->mti_attr.ma_valid = 0;
1803
1804         info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
1805         OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1806         if (info->mti_attr.ma_lmm == NULL)
1807                 RETURN(-ENOMEM);
1808
1809         rc = mdt_mfd_close(info, mfd);
1810
1811         OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1812         mdt_empty_transno(info, rc);
1813         RETURN(rc);
1814 }