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, 2013, Intel Corporation.
  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(const struct mdt_export_data *med)
  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                 mfd->mfd_handle.h_owner = med;
  70                 CFS_INIT_LIST_HEAD(&mfd->mfd_list);
  71                 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
  72         }
  73
  74         RETURN(mfd);
  75 }
  76
  77 /*
  78  * Find the mfd pointed to by handle in global hash table.
  79  * In case of replay the handle is obsoleted
  80  * but mfd can be found in mfd list by that handle.
  81  * Callers need to be holding med_open_lock.
  82  */
  83 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
  84                                      const struct lustre_handle *handle,
  85                                      bool is_replay_or_resent)
  86 {
  87         struct mdt_file_data   *mfd;
  88         ENTRY;
  89
  90         LASSERT(handle != NULL);
  91         mfd = class_handle2object(handle->cookie, med);
  92         /* during dw/setattr replay the mfd can be found by old handle */
  93         if (mfd == NULL && is_replay_or_resent) {
  94                 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  95                         if (mfd->mfd_old_handle.cookie == handle->cookie)
  96                                 RETURN(mfd);
  97                 }
  98                 mfd = NULL;
  99         }
 100
 101         RETURN(mfd);
 102 }
 103
 104 /* free mfd */
 105 void mdt_mfd_free(struct mdt_file_data *mfd)
 106 {
 107         LASSERT(cfs_list_empty(&mfd->mfd_list));
 108         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 109 }
 110
 111 static int mdt_create_data(struct mdt_thread_info *info,
 112                            struct mdt_object *p, struct mdt_object *o)
 113 {
 114         struct md_op_spec     *spec = &info->mti_spec;
 115         struct md_attr        *ma   = &info->mti_attr;
 116         int                    rc   = 0;
 117         ENTRY;
 118
 119         if (!md_should_create(spec->sp_cr_flags))
 120                 RETURN(0);
 121
 122         ma->ma_need = MA_INODE | MA_LOV;
 123         ma->ma_valid = 0;
 124         mutex_lock(&o->mot_lov_mutex);
 125         if (!(o->mot_flags & MOF_LOV_CREATED)) {
 126                 if (p != NULL && (fid_is_obf(mdt_object_fid(p)) ||
 127                                   fid_is_dot_lustre(mdt_object_fid(p))))
 128                         GOTO(unlock, rc = -EPERM);
 129
 130                 rc = mdo_create_data(info->mti_env,
 131                                      p ? mdt_object_child(p) : NULL,
 132                                      mdt_object_child(o), spec, ma);
 133                 if (rc == 0)
 134                         rc = mdt_attr_get_complex(info, o, ma);
 135
 136                 if (rc == 0 && ma->ma_valid & MA_LOV)
 137                         o->mot_flags |= MOF_LOV_CREATED;
 138         }
 139 unlock:
 140         mutex_unlock(&o->mot_lov_mutex);
 141         RETURN(rc);
 142 }
 143
 144 static int mdt_ioepoch_opened(struct mdt_object *mo)
 145 {
 146         return mo->mot_ioepoch_count;
 147 }
 148
 149 int mdt_object_is_som_enabled(struct mdt_object *mo)
 150 {
 151         return !mo->mot_ioepoch;
 152 }
 153
 154 /**
 155  * Re-enable Size-on-MDS.
 156  * Call under ->mot_ioepoch_mutex.
 157  */
 158 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
 159 {
 160         if (ioepoch == mo->mot_ioepoch) {
 161                 LASSERT(!mdt_ioepoch_opened(mo));
 162                 mo->mot_ioepoch = 0;
 163                 mo->mot_flags = 0;
 164         }
 165 }
 166
 167 /**
 168  * Open the IOEpoch. It is allowed if @writecount is not negative.
 169  * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
 170  */
 171 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
 172                      int created)
 173 {
 174         struct mdt_device *mdt = info->mti_mdt;
 175         int cancel = 0;
 176         int rc = 0;
 177         ENTRY;
 178
 179         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 180             !S_ISREG(lu_object_attr(&o->mot_obj)))
 181                 RETURN(0);
 182
 183         mutex_lock(&o->mot_ioepoch_mutex);
 184         if (mdt_ioepoch_opened(o)) {
 185                 /* Epoch continues even if there is no writers yet. */
 186                 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
 187                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 188         } else {
 189                 /* XXX: ->mdt_ioepoch is not initialized at the mount */
 190                 spin_lock(&mdt->mdt_ioepoch_lock);
 191                 if (mdt->mdt_ioepoch < info->mti_replayepoch)
 192                         mdt->mdt_ioepoch = info->mti_replayepoch;
 193
 194                 if (info->mti_replayepoch)
 195                         o->mot_ioepoch = info->mti_replayepoch;
 196                 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
 197                         o->mot_ioepoch = ++mdt->mdt_ioepoch;
 198                 else
 199                         o->mot_ioepoch = mdt->mdt_ioepoch;
 200
 201                 spin_unlock(&mdt->mdt_ioepoch_lock);
 202
 203                 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
 204                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 205                 if (created)
 206                         o->mot_flags |= MOF_SOM_CREATED;
 207                 cancel = 1;
 208         }
 209         o->mot_ioepoch_count++;
 210         mutex_unlock(&o->mot_ioepoch_mutex);
 211
 212         /* Cancel Size-on-MDS attributes cached on clients for the open case.
 213          * In the truncate case, see mdt_reint_setattr(). */
 214         if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
 215                 struct mdt_lock_handle  *lh = &info->mti_lh[MDT_LH_CHILD];
 216                 mdt_lock_reg_init(lh, LCK_EX);
 217                 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
 218                                      MDT_LOCAL_LOCK);
 219                 if (rc == 0)
 220                         mdt_object_unlock(info, o, lh, 1);
 221         }
 222         RETURN(rc);
 223 }
 224
 225 /**
 226  * Update SOM on-disk attributes.
 227  * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
 228  * mountid and attributes. If disabling, clean SOM xattr.
 229  * Call under ->mot_ioepoch_mutex.
 230  */
 231 static int mdt_som_attr_set(struct mdt_thread_info *info,
 232                             struct mdt_object *obj, __u64 ioepoch, bool enable)
 233 {
 234         struct md_object        *next = mdt_object_child(obj);
 235         int                      rc;
 236         ENTRY;
 237
 238         CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
 239                " on "DFID".\n", enable ? "update" : "disabling",
 240                ioepoch, PFID(mdt_object_fid(obj)));
 241
 242         if (enable) {
 243                 struct lu_buf           *buf = &info->mti_buf;
 244                 struct som_attrs        *attrs;
 245                 struct md_attr          *ma = &info->mti_attr;
 246                 struct lu_attr          *la = &ma->ma_attr;
 247                 struct obd_device       *obd = info->mti_mdt->mdt_lut.lut_obd;
 248
 249                 attrs = (struct som_attrs *)info->mti_xattr_buf;
 250                 CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*attrs));
 251
 252                 /* pack SOM attributes */
 253                 memset(attrs, 0, sizeof(*attrs));
 254                 attrs->som_ioepoch = ioepoch;
 255                 attrs->som_mountid = obd->u.obt.obt_mount_count;
 256                 if ((la->la_valid & LA_SIZE) != 0)
 257                         attrs->som_size = la->la_size;
 258                 if ((la->la_valid & LA_BLOCKS) != 0)
 259                         attrs->som_blocks = la->la_blocks;
 260                 lustre_som_swab(attrs);
 261
 262                 /* update SOM attributes */
 263                 buf->lb_buf = attrs;
 264                 buf->lb_len = sizeof(*attrs);
 265                 rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_SOM, 0);
 266         } else {
 267                 /* delete SOM attributes */
 268                 rc = mo_xattr_del(info->mti_env, next, XATTR_NAME_SOM);
 269         }
 270
 271         RETURN(rc);
 272 }
 273
 274 /** Perform the eviction specific actions on ioepoch close. */
 275 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
 276                                                 struct mdt_object *o)
 277 {
 278         int rc = 0;
 279
 280         mutex_lock(&o->mot_ioepoch_mutex);
 281         CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
 282                "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
 283                o->mot_ioepoch_count);
 284         o->mot_ioepoch_count--;
 285
 286         /* If eviction occured set MOF_SOM_RECOV,
 287          * if no other epoch holders, disable SOM on disk. */
 288         o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
 289         if (!mdt_ioepoch_opened(o)) {
 290                 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
 291                 mdt_object_som_enable(o, o->mot_ioepoch);
 292         }
 293         mutex_unlock(&o->mot_ioepoch_mutex);
 294         RETURN(rc);
 295 }
 296
 297 /**
 298  * Perform the replay specific actions on ioepoch close.
 299  * Skip SOM attribute update if obtained and just forget about the inode state
 300  * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
 301  */
 302 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
 303                                               struct mdt_object *o)
 304 {
 305         int rc = MDT_IOEPOCH_CLOSED;
 306         ENTRY;
 307
 308         mutex_lock(&o->mot_ioepoch_mutex);
 309         CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
 310                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
 311         o->mot_ioepoch_count--;
 312
 313         /* Get an info from the replayed request if client is supposed
 314          * to send an Attibute Update, reconstruct @rc if so */
 315         if (info->mti_ioepoch->flags & MF_SOM_AU)
 316                 rc = MDT_IOEPOCH_GETATTR;
 317
 318         if (!mdt_ioepoch_opened(o))
 319                 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
 320         mutex_unlock(&o->mot_ioepoch_mutex);
 321
 322         RETURN(rc);
 323 }
 324
 325 /**
 326  * Regular file IOepoch close.
 327  * Closes the ioepoch, checks the object state, apply obtained attributes and
 328  * re-enable SOM on the object, if possible. Also checks if the recovery is
 329  * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
 330  * to obtain SOM attributes under the server-side OST locks.
 331  *
 332  * Return value:
 333  * MDT_IOEPOCH_CLOSED if ioepoch is closed.
 334  * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
 335  */
 336 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
 337                                         struct mdt_object *o)
 338 {
 339         struct md_attr *tmp_ma;
 340         struct lu_attr *la;
 341         int achange, opened;
 342         int recovery = 0;
 343         int rc = 0, ret = MDT_IOEPOCH_CLOSED;
 344         ENTRY;
 345
 346         la = &info->mti_attr.ma_attr;
 347         achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
 348
 349         mutex_lock(&o->mot_ioepoch_mutex);
 350         o->mot_ioepoch_count--;
 351
 352         tmp_ma = &info->mti_u.som.attr;
 353         tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
 354         tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
 355         tmp_ma->ma_som = &info->mti_u.som.data;
 356         tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
 357         tmp_ma->ma_valid = 0;
 358         rc = mdt_attr_get_complex(info, o, tmp_ma);
 359         if (rc)
 360                 GOTO(error_up, rc);
 361
 362         /* Check the on-disk SOM state. */
 363         if (o->mot_flags & MOF_SOM_RECOV)
 364                 recovery = 1;
 365         else if (!(o->mot_flags & MOF_SOM_CREATED) &&
 366                  !(tmp_ma->ma_valid & MA_SOM))
 367                 recovery = 1;
 368
 369         CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
 370                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
 371
 372         opened = mdt_ioepoch_opened(o);
 373         /**
 374          * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
 375          * Skip the check for file with no LOV  or for unlink files.
 376          */
 377         if (!opened && tmp_ma->ma_valid & MA_LOV &&
 378             !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
 379                 if (recovery)
 380                         /* If some previous writer was evicted, re-ask the
 381                          * client for attributes. Even if attributes are
 382                          * provided, we cannot believe in them.
 383                          * Another use case is that there is no SOM cache on
 384                          * disk -- first access with SOM or there was an MDS
 385                          * failure. */
 386                         ret = MDT_IOEPOCH_GETATTR;
 387                 else if (o->mot_flags & MOF_SOM_CHANGE)
 388                         /* Some previous writer changed the attribute.
 389                          * Do not believe to the current Size-on-MDS
 390                          * update, re-ask client. */
 391                         ret = MDT_IOEPOCH_GETATTR;
 392                 else if (!(la->la_valid & LA_SIZE) && achange)
 393                         /* Attributes were changed by the last writer
 394                          * only but no Size-on-MDS update is received.*/
 395                         ret = MDT_IOEPOCH_GETATTR;
 396         }
 397
 398         if (achange || ret == MDT_IOEPOCH_GETATTR)
 399                 o->mot_flags |= MOF_SOM_CHANGE;
 400
 401         /* If epoch ends and relable SOM attributes are obtained, update them.
 402          * Create SOM ea for new files even if there is no attributes obtained
 403          * (0-length file). */
 404         if (ret == MDT_IOEPOCH_CLOSED && !opened) {
 405                 if (achange || o->mot_flags & MOF_SOM_CREATED) {
 406                         LASSERT(achange || !(la->la_valid & LA_SIZE));
 407                         rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
 408                                               MDT_SOM_ENABLE);
 409                         /* Avoid the following setattrs of these attributes,
 410                          * e.g. for atime update. */
 411                         info->mti_attr.ma_valid = 0;
 412                 }
 413                 mdt_object_som_enable(o, o->mot_ioepoch);
 414         }
 415
 416         mutex_unlock(&o->mot_ioepoch_mutex);
 417         /* If recovery is needed, tell the client to perform GETATTR under
 418          * the lock. */
 419         if (ret == MDT_IOEPOCH_GETATTR && recovery) {
 420                 struct mdt_body *rep;
 421                 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 422                 rep->valid |= OBD_MD_FLGETATTRLOCK;
 423         }
 424
 425         RETURN(rc ? : ret);
 426
 427 error_up:
 428         mutex_unlock(&o->mot_ioepoch_mutex);
 429         return rc;
 430 }
 431
 432 /**
 433  * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
 434  * - a client closes the IOEpoch;
 435  * - a client eviction occured.
 436  * Return values:
 437  * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
 438  * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
 439  * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
 440  * update is needed.
 441  */
 442 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
 443 {
 444         struct ptlrpc_request *req = mdt_info_req(info);
 445         ENTRY;
 446
 447         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 448             !S_ISREG(lu_object_attr(&o->mot_obj)))
 449                 RETURN(0);
 450
 451         LASSERT(o->mot_ioepoch_count);
 452         LASSERT(info->mti_ioepoch == NULL ||
 453                 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
 454
 455         /* IOEpoch is closed only if client tells about it or eviction occures.
 456          * In the replay case, always close the epoch. */
 457         if (req == NULL)
 458                 RETURN(mdt_ioepoch_close_on_eviction(info, o));
 459         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
 460                 RETURN(mdt_ioepoch_close_on_replay(info, o));
 461         if (info->mti_ioepoch && (info->mti_ioepoch->flags & MF_EPOCH_CLOSE))
 462                 RETURN(mdt_ioepoch_close_reg(info, o));
 463         /* IO epoch is not closed. */
 464         RETURN(MDT_IOEPOCH_OPENED);
 465 }
 466
 467 /**
 468  * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
 469  * for attribute update. It happens if:
 470  * - SOM Attribute Update is obtained;
 471  * - the client failed to obtain it and informs MDS about it;
 472  * - a client eviction occured.
 473  * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
 474  * cache otherwise.
 475  */
 476 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
 477 {
 478         struct ptlrpc_request   *req = mdt_info_req(info);
 479         __u64                    ioepoch = 0;
 480         int                      act = MDT_SOM_ENABLE;
 481         int                      rc = 0;
 482         ENTRY;
 483
 484         LASSERT(!req || info->mti_ioepoch);
 485         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 486             !S_ISREG(lu_object_attr(&o->mot_obj)))
 487                 RETURN(0);
 488
 489         /* No size whereas MF_SOM_CHANGE is set means client failed to
 490          * obtain ost attributes, drop the SOM cache on disk if so. */
 491         if (!req ||
 492             (info->mti_ioepoch &&
 493              info->mti_ioepoch->flags & MF_SOM_CHANGE &&
 494              !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
 495                 act = MDT_SOM_DISABLE;
 496
 497         mutex_lock(&o->mot_ioepoch_mutex);
 498         /* Mark the object it is the recovery state if we failed to obtain
 499          * SOM attributes. */
 500         if (act == MDT_SOM_DISABLE)
 501                 o->mot_flags |= MOF_SOM_RECOV;
 502
 503         if (!mdt_ioepoch_opened(o)) {
 504                 ioepoch =  info->mti_ioepoch ?
 505                         info->mti_ioepoch->ioepoch : o->mot_ioepoch;
 506
 507                 if (req != NULL
 508                     && !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 509                         rc = mdt_som_attr_set(info, o, ioepoch, act);
 510                 mdt_object_som_enable(o, ioepoch);
 511         }
 512         mutex_unlock(&o->mot_ioepoch_mutex);
 513         RETURN(rc);
 514 }
 515
 516 int mdt_write_read(struct mdt_object *o)
 517 {
 518         int rc = 0;
 519         ENTRY;
 520         mutex_lock(&o->mot_ioepoch_mutex);
 521         rc = o->mot_writecount;
 522         mutex_unlock(&o->mot_ioepoch_mutex);
 523         RETURN(rc);
 524 }
 525
 526 int mdt_write_get(struct mdt_object *o)
 527 {
 528         int rc = 0;
 529         ENTRY;
 530         mutex_lock(&o->mot_ioepoch_mutex);
 531         if (o->mot_writecount < 0)
 532                 rc = -ETXTBSY;
 533         else
 534                 o->mot_writecount++;
 535         mutex_unlock(&o->mot_ioepoch_mutex);
 536         RETURN(rc);
 537 }
 538
 539 void mdt_write_put(struct mdt_object *o)
 540 {
 541         ENTRY;
 542         mutex_lock(&o->mot_ioepoch_mutex);
 543         o->mot_writecount--;
 544         mutex_unlock(&o->mot_ioepoch_mutex);
 545         EXIT;
 546 }
 547
 548 static int mdt_write_deny(struct mdt_object *o)
 549 {
 550         int rc = 0;
 551         ENTRY;
 552         mutex_lock(&o->mot_ioepoch_mutex);
 553         if (o->mot_writecount > 0)
 554                 rc = -ETXTBSY;
 555         else
 556                 o->mot_writecount--;
 557         mutex_unlock(&o->mot_ioepoch_mutex);
 558         RETURN(rc);
 559 }
 560
 561 static void mdt_write_allow(struct mdt_object *o)
 562 {
 563         ENTRY;
 564         mutex_lock(&o->mot_ioepoch_mutex);
 565         o->mot_writecount++;
 566         mutex_unlock(&o->mot_ioepoch_mutex);
 567         EXIT;
 568 }
 569
 570 /* there can be no real transaction so prepare the fake one */
 571 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 572 {
 573         struct mdt_device      *mdt = info->mti_mdt;
 574         struct ptlrpc_request  *req = mdt_info_req(info);
 575         struct tg_export_data  *ted;
 576         struct lsd_client_data *lcd;
 577
 578         ENTRY;
 579         /* transaction has occurred already */
 580         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 581                 RETURN_EXIT;
 582
 583         spin_lock(&mdt->mdt_lut.lut_translock);
 584         if (rc != 0) {
 585                 if (info->mti_transno != 0) {
 586                         struct obd_export *exp = req->rq_export;
 587
 588                         CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
 589                                mdt_obd_name(mdt), info->mti_transno,
 590                                libcfs_nid2str(exp->exp_connection->c_peer.nid),
 591                                rc);
 592                         spin_unlock(&mdt->mdt_lut.lut_translock);
 593                         RETURN_EXIT;
 594                 }
 595         } else if (info->mti_transno == 0) {
 596                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
 597         } else {
 598                 /* should be replay */
 599                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 600                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 601         }
 602         spin_unlock(&mdt->mdt_lut.lut_translock);
 603
 604         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 605                info->mti_transno,
 606                req->rq_export->exp_obd->obd_last_committed);
 607
 608         req->rq_transno = info->mti_transno;
 609         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 610
 611         /* update lcd in memory only for resent cases */
 612         ted = &req->rq_export->exp_target_data;
 613         LASSERT(ted);
 614         mutex_lock(&ted->ted_lcd_lock);
 615         lcd = ted->ted_lcd;
 616         if (info->mti_transno < lcd->lcd_last_transno &&
 617             info->mti_transno != 0) {
 618                 /* This should happen during replay. Do not update
 619                  * last rcvd info if replay req transno < last transno,
 620                  * otherwise the following resend(after replay) can not
 621                  * be checked correctly by xid */
 622                 mutex_unlock(&ted->ted_lcd_lock);
 623                 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
 624                        mdt_obd_name(mdt), info->mti_transno,
 625                        lcd->lcd_last_transno);
 626                 RETURN_EXIT;
 627         }
 628
 629         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 630             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 631                 if (info->mti_transno != 0)
 632                         lcd->lcd_last_close_transno = info->mti_transno;
 633                 lcd->lcd_last_close_xid = req->rq_xid;
 634                 lcd->lcd_last_close_result = rc;
 635         } else {
 636                 /* VBR: save versions in last_rcvd for reconstruct. */
 637                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 638                 if (pre_versions) {
 639                         lcd->lcd_pre_versions[0] = pre_versions[0];
 640                         lcd->lcd_pre_versions[1] = pre_versions[1];
 641                         lcd->lcd_pre_versions[2] = pre_versions[2];
 642                         lcd->lcd_pre_versions[3] = pre_versions[3];
 643                 }
 644                 if (info->mti_transno != 0)
 645                         lcd->lcd_last_transno = info->mti_transno;
 646
 647                 lcd->lcd_last_xid = req->rq_xid;
 648                 lcd->lcd_last_result = rc;
 649                 lcd->lcd_last_data = info->mti_opdata;
 650         }
 651         mutex_unlock(&ted->ted_lcd_lock);
 652
 653         EXIT;
 654 }
 655
 656 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
 657 {
 658         LASSERT(mfd != NULL);
 659
 660         CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n",
 661                PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
 662
 663         mfd->mfd_mode = mode;
 664 }
 665
 666 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 667                         struct mdt_object *o, __u64 flags, int created)
 668 {
 669         struct ptlrpc_request   *req = mdt_info_req(info);
 670         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 671         struct mdt_file_data    *mfd;
 672         struct md_attr          *ma  = &info->mti_attr;
 673         struct lu_attr          *la  = &ma->ma_attr;
 674         struct mdt_body         *repbody;
 675         int                      rc = 0, isdir, isreg;
 676         ENTRY;
 677
 678         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 679
 680         isreg = S_ISREG(la->la_mode);
 681         isdir = S_ISDIR(la->la_mode);
 682         if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
 683                 /*
 684                  * No EA, check whether it is will set regEA and dirEA since in
 685                  * above attr get, these size might be zero, so reset it, to
 686                  * retrieve the MD after create obj.
 687                  */
 688                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 689                                                        &RMF_MDT_MD,
 690                                                        RCL_SERVER);
 691                 /* in replay case, p == NULL */
 692                 rc = mdt_create_data(info, p, o);
 693                 if (rc)
 694                         RETURN(rc);
 695         }
 696
 697         CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
 698                ma->ma_valid, ma->ma_lmm_size);
 699
 700         if (ma->ma_valid & MA_LOV) {
 701                 LASSERT(ma->ma_lmm_size != 0);
 702                 repbody->eadatasize = ma->ma_lmm_size;
 703                 if (isdir)
 704                         repbody->valid |= OBD_MD_FLDIREA;
 705                 else
 706                         repbody->valid |= OBD_MD_FLEASIZE;
 707         }
 708
 709         if (flags & FMODE_WRITE) {
 710                 rc = mdt_write_get(o);
 711                 if (rc == 0) {
 712                         mdt_ioepoch_open(info, o, created);
 713                         repbody->ioepoch = o->mot_ioepoch;
 714                 }
 715         } else if (flags & MDS_FMODE_EXEC) {
 716                 /* if file is released, we can't deny write because we must
 717                  * restore (write) it to access it.*/
 718                 if ((ma->ma_valid & MA_HSM) &&
 719                     (ma->ma_hsm.mh_flags & HS_RELEASED))
 720                         rc = 0;
 721                 else
 722                         rc = mdt_write_deny(o);
 723         }
 724         if (rc)
 725                 RETURN(rc);
 726
 727         rc = mo_open(info->mti_env, mdt_object_child(o),
 728                      created ? flags | MDS_OPEN_CREATED : flags);
 729         if (rc)
 730                 GOTO(err_out, rc);
 731
 732         mfd = mdt_mfd_new(med);
 733         if (mfd == NULL)
 734                 GOTO(err_out, rc = -ENOMEM);
 735
 736         /*
 737          * Keep a reference on this object for this open, and is
 738          * released by mdt_mfd_close().
 739          */
 740         mdt_object_get(info->mti_env, o);
 741         mfd->mfd_object = o;
 742         mfd->mfd_xid = req->rq_xid;
 743
 744         /*
 745          * @flags is always not zero. At least it should be FMODE_READ,
 746          * FMODE_WRITE or MDS_FMODE_EXEC.
 747          */
 748         LASSERT(flags != 0);
 749
 750         /* Open handling. */
 751         mdt_mfd_set_mode(mfd, flags);
 752
 753         atomic_inc(&o->mot_open_count);
 754         if (flags & MDS_OPEN_LEASE)
 755                 atomic_inc(&o->mot_lease_count);
 756
 757         /* replay handle */
 758         if (req_is_replay(req)) {
 759                 struct mdt_file_data *old_mfd;
 760                 /* Check wheather old cookie already exist in
 761                  * the list, becasue when do recovery, client
 762                  * might be disconnected from server, and
 763                  * restart replay, so there maybe some orphan
 764                  * mfd here, we should remove them */
 765                 LASSERT(info->mti_rr.rr_handle != NULL);
 766                 spin_lock(&med->med_open_lock);
 767                 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
 768                 if (old_mfd != NULL) {
 769                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 770                                "cookie = "LPX64"\n", mfd,
 771                                PFID(mdt_object_fid(mfd->mfd_object)),
 772                                info->mti_rr.rr_handle->cookie);
 773                         class_handle_unhash(&old_mfd->mfd_handle);
 774                         cfs_list_del_init(&old_mfd->mfd_list);
 775                         spin_unlock(&med->med_open_lock);
 776                         /* no attr update for that close */
 777                         la->la_valid = 0;
 778                         ma->ma_valid |= MA_FLAGS;
 779                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 780                         mdt_mfd_close(info, old_mfd);
 781                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 782                         ma->ma_valid &= ~MA_FLAGS;
 783                 } else {
 784                         spin_unlock(&med->med_open_lock);
 785                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 786                                "cookie = "LPX64"\n",
 787                                PFID(mdt_object_fid(mfd->mfd_object)),
 788                                info->mti_rr.rr_handle->cookie);
 789                 }
 790
 791                 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
 792                        info->mti_rr.rr_handle->cookie);
 793
 794                 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
 795         }
 796
 797         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 798
 799         if (req->rq_export->exp_disconnected) {
 800                 spin_lock(&med->med_open_lock);
 801                 class_handle_unhash(&mfd->mfd_handle);
 802                 cfs_list_del_init(&mfd->mfd_list);
 803                 spin_unlock(&med->med_open_lock);
 804                 mdt_mfd_close(info, mfd);
 805         } else {
 806                 spin_lock(&med->med_open_lock);
 807                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
 808                 spin_unlock(&med->med_open_lock);
 809         }
 810
 811         mdt_empty_transno(info, rc);
 812
 813         RETURN(rc);
 814
 815 err_out:
 816         if (flags & FMODE_WRITE)
 817                         /* XXX We also need to close io epoch here.
 818                          * See LU-1220 - green */
 819                 mdt_write_put(o);
 820         else if (flags & FMODE_EXEC)
 821                 mdt_write_allow(o);
 822         return rc;
 823 }
 824
 825 int mdt_finish_open(struct mdt_thread_info *info,
 826                     struct mdt_object *p, struct mdt_object *o,
 827                     __u64 flags, int created, struct ldlm_reply *rep)
 828 {
 829         struct ptlrpc_request   *req = mdt_info_req(info);
 830         struct obd_export       *exp = req->rq_export;
 831         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 832         struct md_attr          *ma  = &info->mti_attr;
 833         struct lu_attr          *la  = &ma->ma_attr;
 834         struct mdt_file_data    *mfd;
 835         struct mdt_body         *repbody;
 836         int                      rc = 0;
 837         int                      isreg, isdir, islnk;
 838         cfs_list_t              *t;
 839         ENTRY;
 840
 841         LASSERT(ma->ma_valid & MA_INODE);
 842
 843         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 844
 845         isreg = S_ISREG(la->la_mode);
 846         isdir = S_ISDIR(la->la_mode);
 847         islnk = S_ISLNK(la->la_mode);
 848         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 849
 850         /* LU-2275, simulate broken behaviour (esp. prevalent in
 851          * pre-2.4 servers where a very strange reply is sent on error
 852          * that looks like it was actually almost succesful and a failure at the
 853          * same time */
 854         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 855                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 856                                                DISP_LOOKUP_NEG |
 857                                                DISP_LOOKUP_POS);
 858
 859                 if (flags & MDS_OPEN_LOCK)
 860                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 861
 862                 RETURN(-ENOENT);
 863         }
 864
 865         if (exp_connect_rmtclient(exp)) {
 866                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 867
 868                 rc = mdt_pack_remote_perm(info, o, buf);
 869                 if (rc) {
 870                         repbody->valid &= ~OBD_MD_FLRMTPERM;
 871                         repbody->aclsize = 0;
 872                 } else {
 873                         repbody->valid |= OBD_MD_FLRMTPERM;
 874                         repbody->aclsize = sizeof(struct mdt_remote_perm);
 875                 }
 876         }
 877 #ifdef CONFIG_FS_POSIX_ACL
 878         else if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 879                 const struct lu_env *env = info->mti_env;
 880                 struct md_object *next = mdt_object_child(o);
 881                 struct lu_buf *buf = &info->mti_buf;
 882
 883                 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 884                 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
 885                                                    RCL_SERVER);
 886                 if (buf->lb_len > 0) {
 887                         rc = mo_xattr_get(env, next, buf,
 888                                           XATTR_NAME_ACL_ACCESS);
 889                         if (rc < 0) {
 890                                 if (rc == -ENODATA) {
 891                                         repbody->aclsize = 0;
 892                                         repbody->valid |= OBD_MD_FLACL;
 893                                         rc = 0;
 894                                 } else if (rc == -EOPNOTSUPP) {
 895                                         rc = 0;
 896                                 } else {
 897                                         CERROR("got acl size: %d\n", rc);
 898                                 }
 899                         } else {
 900                                 repbody->aclsize = rc;
 901                                 repbody->valid |= OBD_MD_FLACL;
 902                                 rc = 0;
 903                         }
 904                 }
 905         }
 906 #endif
 907
 908         if (info->mti_mdt->mdt_lut.lut_mds_capa &&
 909             exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) {
 910                 struct lustre_capa *capa;
 911
 912                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
 913                 LASSERT(capa);
 914                 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
 915                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 916                 if (rc)
 917                         RETURN(rc);
 918                 repbody->valid |= OBD_MD_FLMDSCAPA;
 919         }
 920         if (info->mti_mdt->mdt_lut.lut_oss_capa &&
 921             exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA &&
 922             S_ISREG(lu_object_attr(&o->mot_obj))) {
 923                 struct lustre_capa *capa;
 924
 925                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
 926                 LASSERT(capa);
 927                 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
 928                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 929                 if (rc)
 930                         RETURN(rc);
 931                 repbody->valid |= OBD_MD_FLOSSCAPA;
 932         }
 933
 934         /*
 935          * If we are following a symlink, don't open; and do not return open
 936          * handle for special nodes as client required.
 937          */
 938         if (islnk || (!isreg && !isdir &&
 939             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 940                 lustre_msg_set_transno(req->rq_repmsg, 0);
 941                 RETURN(0);
 942         }
 943
 944         /*
 945          * We need to return the existing object's fid back, so it is done here,
 946          * after preparing the reply.
 947          */
 948         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 949                 RETURN(-EEXIST);
 950
 951         /* This can't be done earlier, we need to return reply body */
 952         if (isdir) {
 953                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 954                         /* We are trying to create or write an existing dir. */
 955                         RETURN(-EISDIR);
 956                 }
 957         } else if (flags & MDS_OPEN_DIRECTORY)
 958                 RETURN(-ENOTDIR);
 959
 960         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 961                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 962                 RETURN(-EAGAIN);
 963         }
 964
 965         mfd = NULL;
 966         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 967                 spin_lock(&med->med_open_lock);
 968                 cfs_list_for_each(t, &med->med_open_head) {
 969                         mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
 970                         if (mfd->mfd_xid == req->rq_xid)
 971                                 break;
 972                         mfd = NULL;
 973                 }
 974                 spin_unlock(&med->med_open_lock);
 975
 976                 if (mfd != NULL) {
 977                         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 978                         /*set repbody->ea_size for resent case*/
 979                         if (ma->ma_valid & MA_LOV) {
 980                                 LASSERT(ma->ma_lmm_size != 0);
 981                                 repbody->eadatasize = ma->ma_lmm_size;
 982                                 if (isdir)
 983                                         repbody->valid |= OBD_MD_FLDIREA;
 984                                 else
 985                                         repbody->valid |= OBD_MD_FLEASIZE;
 986                         }
 987                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 988                         RETURN(0);
 989                 }
 990         }
 991
 992         rc = mdt_mfd_open(info, p, o, flags, created);
 993         if (!rc)
 994                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 995
 996         RETURN(rc);
 997 }
 998
 999 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
1000                              struct lsd_client_data *lcd);
1001
1002 void mdt_reconstruct_open(struct mdt_thread_info *info,
1003                           struct mdt_lock_handle *lhc)
1004 {
1005         const struct lu_env *env = info->mti_env;
1006         struct mdt_device       *mdt  = info->mti_mdt;
1007         struct req_capsule      *pill = info->mti_pill;
1008         struct ptlrpc_request   *req  = mdt_info_req(info);
1009         struct tg_export_data   *ted  = &req->rq_export->exp_target_data;
1010         struct lsd_client_data  *lcd  = ted->ted_lcd;
1011         struct md_attr          *ma   = &info->mti_attr;
1012         struct mdt_reint_record *rr   = &info->mti_rr;
1013         __u64                   flags = info->mti_spec.sp_cr_flags;
1014         struct ldlm_reply       *ldlm_rep;
1015         struct mdt_object       *parent;
1016         struct mdt_object       *child;
1017         struct mdt_body         *repbody;
1018         int                      rc;
1019         ENTRY;
1020
1021         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1022         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1023         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1024
1025         ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
1026         ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
1027                                                RCL_SERVER);
1028         ma->ma_need = MA_INODE | MA_HSM;
1029         if (ma->ma_lmm_size > 0)
1030                 ma->ma_need |= MA_LOV;
1031
1032         ma->ma_valid = 0;
1033
1034         mdt_req_from_lcd(req, lcd);
1035         mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1036
1037         CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1038                ldlm_rep->lock_policy_res1, req->rq_status);
1039
1040         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1041             req->rq_status != 0)
1042                 /* We did not create successfully, return error to client. */
1043                 GOTO(out, rc = req->rq_status);
1044
1045         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1046                 struct obd_export *exp = req->rq_export;
1047                 /*
1048                  * We failed after creation, but we do not know in which step
1049                  * we failed. So try to check the child object.
1050                  */
1051                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1052                 if (IS_ERR(parent)) {
1053                         rc = PTR_ERR(parent);
1054                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
1055                                       " Evicting client %s with export %s.\n",
1056                                       PFID(rr->rr_fid1), rc,
1057                                       obd_uuid2str(&exp->exp_client_uuid),
1058                                       obd_export_nid2str(exp));
1059                         mdt_export_evict(exp);
1060                         RETURN_EXIT;
1061                 }
1062                 child = mdt_object_find(env, mdt, rr->rr_fid2);
1063                 if (IS_ERR(child)) {
1064                         rc = PTR_ERR(child);
1065                         LCONSOLE_WARN("Child "DFID" lookup error %d."
1066                                       " Evicting client %s with export %s.\n",
1067                                       PFID(mdt_object_fid(child)), rc,
1068                                       obd_uuid2str(&exp->exp_client_uuid),
1069                                       obd_export_nid2str(exp));
1070                         mdt_object_put(env, parent);
1071                         mdt_export_evict(exp);
1072                         RETURN_EXIT;
1073                 }
1074
1075                 if (unlikely(mdt_object_remote(child))) {
1076                         /* the child object was created on remote server */
1077                         if (!mdt_is_dne_client(exp)) {
1078                                 /* Return -EIO for old client */
1079                                 mdt_object_put(env, parent);
1080                                 mdt_object_put(env, child);
1081                                 GOTO(out, rc = -EIO);
1082                         }
1083                         repbody->fid1 = *rr->rr_fid2;
1084                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1085                         rc = 0;
1086                 } else {
1087                         if (mdt_object_exists(child)) {
1088                                 mdt_set_capainfo(info, 1, rr->rr_fid2,
1089                                                  BYPASS_CAPA);
1090                                 rc = mdt_attr_get_complex(info, child, ma);
1091                                 if (rc == 0)
1092                                         rc = mdt_finish_open(info, parent,
1093                                                              child, flags,
1094                                                              1, ldlm_rep);
1095                         } else {
1096                                 /* the child does not exist, we should do
1097                                  * regular open */
1098                                 mdt_object_put(env, parent);
1099                                 mdt_object_put(env, child);
1100                                 GOTO(regular_open, 0);
1101                         }
1102                 }
1103                 mdt_object_put(env, parent);
1104                 mdt_object_put(env, child);
1105                 GOTO(out, rc);
1106         } else {
1107 regular_open:
1108                 /* We did not try to create, so we are a pure open */
1109                 rc = mdt_reint_open(info, lhc);
1110         }
1111
1112         EXIT;
1113 out:
1114         req->rq_status = rc;
1115         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1116         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1117 }
1118
1119 int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
1120 {
1121         __u64                    flags = info->mti_spec.sp_cr_flags;
1122         struct mdt_reint_record *rr = &info->mti_rr;
1123         struct md_attr          *ma = &info->mti_attr;
1124         struct mdt_object       *o;
1125         int                      rc;
1126         ENTRY;
1127
1128         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1129         if (IS_ERR(o))
1130                 RETURN(rc = PTR_ERR(o));
1131
1132         if (unlikely(mdt_object_remote(o))) {
1133                 /* the child object was created on remote server */
1134                 struct mdt_body *repbody;
1135
1136                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1137                                                 DISP_LOOKUP_EXECD |
1138                                                 DISP_LOOKUP_POS));
1139                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1140                 repbody->fid1 = *rr->rr_fid2;
1141                 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1142                 rc = 0;
1143         } else {
1144                 if (mdt_object_exists(o)) {
1145                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1146                                                         DISP_LOOKUP_EXECD |
1147                                                         DISP_LOOKUP_POS));
1148
1149                         rc = mdt_attr_get_complex(info, o, ma);
1150                         if (rc == 0)
1151                                 rc = mdt_finish_open(info, NULL, o, flags, 0,
1152                                                      rep);
1153                 } else {
1154                         rc = -ENOENT;
1155                 }
1156         }
1157
1158         mdt_object_put(info->mti_env, o);
1159         RETURN(rc);
1160 }
1161
1162 /* lock object for open */
1163 static int mdt_object_open_lock(struct mdt_thread_info *info,
1164                                 struct mdt_object *obj,
1165                                 struct mdt_lock_handle *lhc,
1166                                 __u64 *ibits)
1167 {
1168         struct md_attr  *ma = &info->mti_attr;
1169         __u64            open_flags = info->mti_spec.sp_cr_flags;
1170         ldlm_mode_t      lm = LCK_CR;
1171         bool             acq_lease = !!(open_flags & MDS_OPEN_LEASE);
1172         bool             try_layout = false;
1173         bool             create_layout = false;
1174         int              rc = 0;
1175         ENTRY;
1176
1177         *ibits = 0;
1178         mdt_lock_handle_init(lhc);
1179
1180         if (req_is_replay(mdt_info_req(info)))
1181                 RETURN(0);
1182
1183         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
1184                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
1185                     md_should_create(open_flags))
1186                         create_layout = true;
1187                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
1188                     ma->ma_need & MA_LOV)
1189                         try_layout = true;
1190         }
1191
1192         if (acq_lease) {
1193                 /* lease open, acquire write mode of open sem */
1194                 down_write(&obj->mot_open_sem);
1195
1196                 /* Lease exists and ask for new lease */
1197                 if (atomic_read(&obj->mot_lease_count) > 0) {
1198                         /* only exclusive open is supported, so lease
1199                          * are conflicted to each other */
1200                         GOTO(out, rc = -EBUSY);
1201                 }
1202
1203                 /* Lease must be with open lock */
1204                 if (!(open_flags & MDS_OPEN_LOCK)) {
1205                         CERROR("Request lease for file:"DFID ", but open lock "
1206                                 "is missed, open_flags = "LPO64".\n",
1207                                 PFID(mdt_object_fid(obj)), open_flags);
1208                         GOTO(out, rc = -EPROTO);
1209                 }
1210
1211                 /* XXX: only exclusive open is supported. */
1212                 lm = LCK_EX;
1213                 *ibits = MDS_INODELOCK_OPEN;
1214
1215                 /* never grant LCK_EX layout lock to client */
1216                 try_layout = false;
1217         } else { /* normal open */
1218                 /* normal open holds read mode of open sem */
1219                 down_read(&obj->mot_open_sem);
1220
1221                 if (open_flags & MDS_OPEN_LOCK) {
1222                         if (open_flags & FMODE_WRITE)
1223                                 lm = LCK_CW;
1224                         /* if file is released, we can't deny write because we must
1225                          * restore (write) it to access it. */
1226                         else if ((open_flags & MDS_FMODE_EXEC) &&
1227                                  !((ma->ma_valid & MA_HSM) &&
1228                                    (ma->ma_hsm.mh_flags & HS_RELEASED)))
1229                                 lm = LCK_PR;
1230                         else
1231                                 lm = LCK_CR;
1232
1233                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
1234                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
1235                         if (open_flags & FMODE_WRITE)
1236                                 lm = LCK_CW;
1237                         else
1238                                 lm = LCK_CR;
1239
1240                         /* revoke lease */
1241                         *ibits = MDS_INODELOCK_OPEN;
1242                         try_layout = false;
1243
1244                         lhc = &info->mti_lh[MDT_LH_LOCAL];
1245                 }
1246                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
1247                         PFID(mdt_object_fid(obj)),
1248                         atomic_read(&obj->mot_open_count), lm);
1249         }
1250
1251         mdt_lock_reg_init(lhc, lm);
1252
1253         /* one problem to return layout lock on open is that it may result
1254          * in too many layout locks cached on the client side. */
1255         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
1256                 /* return lookup lock to validate inode at the client side,
1257                  * this is pretty important otherwise mdt will return layout
1258                  * lock for each open.
1259                  * However this is a double-edged sword because changing
1260                  * permission will revoke huge # of LOOKUP locks. */
1261                 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
1262                 if (!mdt_object_lock_try(info, obj, lhc, *ibits,
1263                                          MDT_CROSS_LOCK)) {
1264                         *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
1265                         if (*ibits != 0)
1266                                 rc = mdt_object_lock(info, obj, lhc, *ibits,
1267                                                 MDT_CROSS_LOCK);
1268                 }
1269         } else if (*ibits != 0) {
1270                 rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK);
1271         }
1272
1273         CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
1274                 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
1275                 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
1276
1277         /* will change layout, revoke layout locks by enqueuing EX lock. */
1278         if (rc == 0 && create_layout) {
1279                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
1280
1281                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
1282                         ", open_flags = "LPO64"\n",
1283                         PFID(mdt_object_fid(obj)), open_flags);
1284
1285                 LASSERT(!try_layout);
1286                 mdt_lock_handle_init(ll);
1287                 mdt_lock_reg_init(ll, LCK_EX);
1288                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
1289                                         MDT_LOCAL_LOCK);
1290
1291                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
1292         }
1293
1294         /* Check if there is any other open handles after acquiring
1295          * open lock. At this point, caching open handles have been revoked
1296          * by open lock.
1297          * XXX: Now only exclusive open is supported. Need to check the
1298          * type of open for generic lease support. */
1299         if (rc == 0 && acq_lease) {
1300                 struct ptlrpc_request *req = mdt_info_req(info);
1301                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1302                 struct mdt_file_data *mfd;
1303                 bool is_replay_or_resent;
1304                 int open_count = 0;
1305
1306                 /* For lease: application can open a file and then apply lease,
1307                  * @handle contains original open handle in that case.
1308                  * In recovery, open REQ will be replayed and the lease REQ may
1309                  * be resent that means the open handle is already stale, so we
1310                  * need to fix it up here by finding new handle. */
1311                 is_replay_or_resent = req_is_replay(req) ||
1312                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
1313
1314                 /* if the request is _not_ a replay request, rr_handle
1315                  * may be used to hold an openhandle which is issuing the
1316                  * lease request, so that this openhandle doesn't count. */
1317                 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
1318                                      is_replay_or_resent);
1319                 if (mfd != NULL)
1320                         ++open_count;
1321
1322                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
1323                         PFID(mdt_object_fid(obj)),
1324                         atomic_read(&obj->mot_open_count), open_count);
1325
1326                 if (atomic_read(&obj->mot_open_count) > open_count)
1327                         GOTO(out, rc = -EBUSY);
1328         }
1329         GOTO(out, rc);
1330
1331 out:
1332         RETURN(rc);
1333 }
1334
1335 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1336                                    struct mdt_object *obj,
1337                                    struct mdt_lock_handle *lhc,
1338                                    __u64 ibits, int rc)
1339 {
1340         __u64 open_flags = info->mti_spec.sp_cr_flags;
1341         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1342         ENTRY;
1343
1344         if (req_is_replay(mdt_info_req(info)))
1345                 RETURN_EXIT;
1346
1347         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1348          * return to client side. */
1349         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1350                 mdt_object_unlock(info, obj, ll, 1);
1351
1352         ll = &info->mti_lh[MDT_LH_LAYOUT];
1353         /* Release local layout lock, layout was created */
1354         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1355                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1356                 mdt_object_unlock(info, obj, ll, 1);
1357         }
1358
1359         if (open_flags & MDS_OPEN_LEASE)
1360                 up_write(&obj->mot_open_sem);
1361         else
1362                 up_read(&obj->mot_open_sem);
1363
1364         /* Cross-ref case, the lock should be returned to the client */
1365         if (ibits == 0 || rc == -EREMOTE)
1366                 RETURN_EXIT;
1367
1368         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
1369                 /* for the open request, the lock will only return to client
1370                  * if open or layout lock is granted. */
1371                 rc = 1;
1372         }
1373
1374         if (rc != 0) {
1375                 struct ldlm_reply       *ldlm_rep;
1376
1377                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1378                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1379                 mdt_object_unlock(info, obj, lhc, 1);
1380         }
1381         RETURN_EXIT;
1382 }
1383
1384 /**
1385  * Check release is permitted for the current HSM flags.
1386  */
1387 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1388 {
1389         if (!(ma->ma_valid & MA_HSM))
1390                 return false;
1391
1392         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1393                 return false;
1394
1395         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1396                 return false;
1397
1398         return true;
1399 }
1400
1401 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1402                          struct mdt_lock_handle *lhc)
1403 {
1404         const struct lu_env     *env   = info->mti_env;
1405         struct mdt_device       *mdt   = info->mti_mdt;
1406         __u64                    flags = info->mti_spec.sp_cr_flags;
1407         struct mdt_reint_record *rr    = &info->mti_rr;
1408         struct md_attr          *ma    = &info->mti_attr;
1409         struct mdt_object       *parent= NULL;
1410         struct mdt_object       *o;
1411         int                      rc;
1412         __u64                    ibits = 0;
1413         ENTRY;
1414
1415         if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1416                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1417                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1418                         if (IS_ERR(parent)) {
1419                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1420                                        " for anonymous created %ld, try to"
1421                                        " use server-side parent.\n",
1422                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1423                                 parent = NULL;
1424                         }
1425                 }
1426                 if (parent == NULL)
1427                         ma->ma_need |= MA_PFID;
1428         }
1429
1430         o = mdt_object_find(env, mdt, rr->rr_fid2);
1431         if (IS_ERR(o))
1432                 RETURN(rc = PTR_ERR(o));
1433
1434         if (mdt_object_remote(o)) {
1435                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1436                        mdt_obd_name(info->mti_mdt),
1437                        PFID(rr->rr_fid2));
1438                 GOTO(out, rc = -EREMOTE);
1439         } else if (!mdt_object_exists(o)) {
1440                 mdt_set_disposition(info, rep,
1441                                     DISP_IT_EXECD |
1442                                     DISP_LOOKUP_EXECD |
1443                                     DISP_LOOKUP_NEG);
1444                 GOTO(out, rc = -ENOENT);
1445         }
1446
1447         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1448
1449         if (flags & MDS_OPEN_RELEASE)
1450                 ma->ma_need |= MA_HSM;
1451         rc = mdt_attr_get_complex(info, o, ma);
1452         if (rc)
1453                 GOTO(out, rc);
1454
1455         /* If a release request, check file flags are fine and ask for an
1456          * exclusive open access. */
1457         if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1458                 GOTO(out, rc = -EPERM);
1459
1460         rc = mdt_object_open_lock(info, o, lhc, &ibits);
1461         if (rc)
1462                 GOTO(out_unlock, rc);
1463
1464         if (ma->ma_valid & MA_PFID) {
1465                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1466                 if (IS_ERR(parent)) {
1467                         CDEBUG(D_INODE, "Fail to find parent "DFID
1468                                " for anonymous created %ld, try to"
1469                                " use system default.\n",
1470                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1471                         parent = NULL;
1472                 }
1473         }
1474
1475         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1476         if (!rc) {
1477                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1478                 if (flags & MDS_OPEN_LOCK)
1479                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1480                 if (flags & MDS_OPEN_LEASE)
1481                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1482         }
1483         GOTO(out_unlock, rc);
1484
1485 out_unlock:
1486         mdt_object_open_unlock(info, o, lhc, ibits, rc);
1487 out:
1488         mdt_object_put(env, o);
1489         if (parent != NULL)
1490                 mdt_object_put(env, parent);
1491         return rc;
1492 }
1493
1494 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1495 static int mdt_cross_open(struct mdt_thread_info *info,
1496                           const struct lu_fid *parent_fid,
1497                           const struct lu_fid *fid,
1498                           struct ldlm_reply *rep, __u32 flags)
1499 {
1500         struct md_attr    *ma = &info->mti_attr;
1501         struct mdt_object *o;
1502         int                rc;
1503         ENTRY;
1504
1505         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1506         if (IS_ERR(o))
1507                 RETURN(rc = PTR_ERR(o));
1508
1509         if (mdt_object_remote(o)) {
1510                 /* Something is wrong here, the object is on another MDS! */
1511                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1512                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1513                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1514                                 &o->mot_obj,
1515                                 "Object isn't on this server! FLD error?\n");
1516                 rc = -EFAULT;
1517         } else {
1518                 if (mdt_object_exists(o)) {
1519                         /* Do permission check for cross-open. */
1520                         rc = mo_permission(info->mti_env, NULL,
1521                                            mdt_object_child(o),
1522                                            NULL, flags | MDS_OPEN_CROSS);
1523                         if (rc)
1524                                 goto out;
1525
1526                         mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1527                         rc = mdt_attr_get_complex(info, o, ma);
1528                         if (rc != 0)
1529                                 GOTO(out, rc);
1530
1531                         /* Do not create lov object if the fid is opened
1532                          * under OBF */
1533                         if (S_ISREG(ma->ma_attr.la_mode) &&
1534                             !(ma->ma_valid & MA_LOV) && (flags & FMODE_WRITE) &&
1535                             fid_is_obf(parent_fid))
1536                                 GOTO(out, rc = -EPERM);
1537
1538                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1539                 } else {
1540                         /*
1541                          * Something is wrong here. lookup was positive but
1542                          * there is no object!
1543                          */
1544                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1545                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1546                         rc = -EFAULT;
1547                 }
1548         }
1549 out:
1550         mdt_object_put(info->mti_env, o);
1551         RETURN(rc);
1552 }
1553
1554 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1555 {
1556         struct mdt_device       *mdt = info->mti_mdt;
1557         struct ptlrpc_request   *req = mdt_info_req(info);
1558         struct mdt_object       *parent;
1559         struct mdt_object       *child;
1560         struct mdt_lock_handle  *lh;
1561         struct ldlm_reply       *ldlm_rep;
1562         struct mdt_body         *repbody;
1563         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1564         struct md_attr          *ma = &info->mti_attr;
1565         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1566         __u64                    ibits = 0;
1567         struct mdt_reint_record *rr = &info->mti_rr;
1568         struct lu_name          *lname;
1569         int                      result, rc;
1570         int                      created = 0;
1571         __u32                    msg_flags;
1572         ENTRY;
1573
1574         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1575                                (obd_timeout + 1) / 4);
1576
1577         mdt_counter_incr(req, LPROC_MDT_OPEN);
1578         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1579
1580         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1581         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1582                                                RCL_SERVER);
1583         ma->ma_need = MA_INODE;
1584         if (ma->ma_lmm_size > 0)
1585                 ma->ma_need |= MA_LOV;
1586
1587         ma->ma_valid = 0;
1588
1589         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1590         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1591
1592         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1593                 CERROR("file join is not supported anymore.\n");
1594                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1595         }
1596         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1597
1598         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1599             info->mti_spec.u.sp_ea.eadata == NULL)
1600                 GOTO(out, result = err_serious(-EINVAL));
1601
1602         CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1603                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1604                PFID(rr->rr_fid1), rr->rr_name,
1605                PFID(rr->rr_fid2), create_flags,
1606                ma->ma_attr.la_mode, msg_flags);
1607         if (info->mti_cross_ref) {
1608                 /* This is cross-ref open */
1609                 mdt_set_disposition(info, ldlm_rep,
1610                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1611                              DISP_LOOKUP_POS));
1612                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1613                                         ldlm_rep, create_flags);
1614                 GOTO(out, result);
1615         } else if (req_is_replay(req) ||
1616             (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1617                 /* This is a replay request or from liblustre with ea. */
1618                 result = mdt_open_by_fid(info, ldlm_rep);
1619
1620                 if (result != -ENOENT) {
1621                         if (req->rq_export->exp_libclient &&
1622                             create_flags & MDS_OPEN_HAS_EA)
1623                                 GOTO(out, result = 0);
1624                         GOTO(out, result);
1625                 }
1626                 /* We didn't find the correct object, so we need to re-create it
1627                  * via a regular replay. */
1628                 if (!(create_flags & MDS_OPEN_CREAT)) {
1629                         DEBUG_REQ(D_ERROR, req,
1630                                   "OPEN & CREAT not in open replay/by_fid.");
1631                         GOTO(out, result = -EFAULT);
1632                 }
1633                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1634         } else if ((rr->rr_namelen == 0 && create_flags & MDS_OPEN_LOCK) ||
1635                    (create_flags & MDS_OPEN_BY_FID)) {
1636                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1637                 /* If result is 0 then open by FID has found the file
1638                  * and there is nothing left for us to do here.  More
1639                  * generally if it is anything other than -ENOENT or
1640                  * -EREMOTE then we return that now.  If -ENOENT and
1641                  * MDS_OPEN_CREAT is set then we must create the file
1642                  * below.  If -EREMOTE then we need to return a LOOKUP
1643                  * lock to the client, which we do below.  Hence this
1644                  * odd looking condition.  See LU-2523. */
1645                 if (!(result == -ENOENT && (create_flags & MDS_OPEN_CREAT)) &&
1646                     result != -EREMOTE)
1647                         GOTO(out, result);
1648
1649                 if (unlikely(rr->rr_namelen == 0))
1650                         GOTO(out, result = -EINVAL);
1651
1652                 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1653         }
1654
1655         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1656                 GOTO(out, result = err_serious(-ENOMEM));
1657
1658         mdt_set_disposition(info, ldlm_rep,
1659                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1660
1661         lh = &info->mti_lh[MDT_LH_PARENT];
1662         mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1663                           LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1664
1665         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1666                                       MDS_INODELOCK_UPDATE);
1667         if (IS_ERR(parent))
1668                 GOTO(out, result = PTR_ERR(parent));
1669
1670         /* get and check version of parent */
1671         result = mdt_version_get_check(info, parent, 0);
1672         if (result)
1673                 GOTO(out_parent, result);
1674
1675         fid_zero(child_fid);
1676
1677         lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1678         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1679                             lname, child_fid, &info->mti_spec);
1680         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1681                  "looking for "DFID"/%s, result fid="DFID"\n",
1682                  PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1683
1684         if (result != 0 && result != -ENOENT && result != -ESTALE)
1685                 GOTO(out_parent, result);
1686
1687         if (result == -ENOENT || result == -ESTALE) {
1688                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1689                 if (result == -ESTALE) {
1690                         /*
1691                          * -ESTALE means the parent is a dead(unlinked) dir, so
1692                          * it should return -ENOENT to in accordance with the
1693                          * original mds implementaion.
1694                          */
1695                         GOTO(out_parent, result = -ENOENT);
1696                 }
1697                 if (!(create_flags & MDS_OPEN_CREAT))
1698                         GOTO(out_parent, result);
1699                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1700                         GOTO(out_parent, result = -EROFS);
1701                 *child_fid = *info->mti_rr.rr_fid2;
1702                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1703                          PFID(child_fid));
1704                 /* In the function below, .hs_keycmp resolves to
1705                  * lu_obj_hop_keycmp() */
1706                 /* coverity[overrun-buffer-val] */
1707                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1708         } else {
1709                 /*
1710                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1711                  * return FID back in that case.
1712                  */
1713                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1714                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1715         }
1716         if (IS_ERR(child))
1717                 GOTO(out_parent, result = PTR_ERR(child));
1718
1719         /** check version of child  */
1720         rc = mdt_version_get_check(info, child, 1);
1721         if (rc)
1722                 GOTO(out_child, result = rc);
1723
1724         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1725         if (result == -ENOENT) {
1726                 /* Create under OBF and .lustre is not permitted */
1727                 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
1728                         GOTO(out_child, result = -EPERM);
1729
1730                 /* save versions in reply */
1731                 mdt_version_get_save(info, parent, 0);
1732                 mdt_version_get_save(info, child, 1);
1733
1734                 /* version of child will be changed */
1735                 info->mti_mos = child;
1736
1737                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1738                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1739
1740                 /* Let lower layers know what is lock mode on directory. */
1741                 info->mti_spec.sp_cr_mode =
1742                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1743
1744                 /*
1745                  * Do not perform lookup sanity check. We know that name does
1746                  * not exist.
1747                  */
1748                 info->mti_spec.sp_cr_lookup = 0;
1749                 info->mti_spec.sp_feat = &dt_directory_features;
1750
1751                 result = mdo_create(info->mti_env,
1752                                     mdt_object_child(parent),
1753                                     lname,
1754                                     mdt_object_child(child),
1755                                     &info->mti_spec,
1756                                     &info->mti_attr);
1757                 if (result == -ERESTART) {
1758                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1759                         GOTO(out_child, result);
1760                 } else {
1761
1762                         /* XXX: we should call this once, see few lines below */
1763                         if (result == 0)
1764                                 result = mdt_attr_get_complex(info, child, ma);
1765
1766                         if (result != 0)
1767                                 GOTO(out_child, result);
1768                 }
1769                 created = 1;
1770         } else {
1771                 /*
1772                  * The object is on remote node, return its FID for remote open.
1773                  */
1774                 if (mdt_object_remote(child)) {
1775                         /*
1776                          * Check if this lock already was sent to client and
1777                          * this is resent case. For resent case do not take lock
1778                          * again, use what is already granted.
1779                          */
1780                         LASSERT(lhc != NULL);
1781
1782                         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1783                                 struct ldlm_lock *lock;
1784
1785                                 LASSERT(msg_flags & MSG_RESENT);
1786
1787                                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1788                                 if (!lock) {
1789                                         CERROR("Invalid lock handle "LPX64"\n",
1790                                                lhc->mlh_reg_lh.cookie);
1791                                         LBUG();
1792                                 }
1793                                 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1794                                                         &lock->l_resource->lr_name));
1795                                 LDLM_LOCK_PUT(lock);
1796                                 rc = 0;
1797                         } else {
1798                                 mdt_lock_handle_init(lhc);
1799                                 mdt_lock_reg_init(lhc, LCK_PR);
1800
1801                                 rc = mdt_object_lock(info, child, lhc,
1802                                                      MDS_INODELOCK_LOOKUP,
1803                                                      MDT_CROSS_LOCK);
1804                         }
1805                         repbody->fid1 = *mdt_object_fid(child);
1806                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1807                         if (rc != 0)
1808                                 result = rc;
1809                         else
1810                                 result = -EREMOTE;
1811                         GOTO(out_child, result);
1812                 } else {
1813                         if (mdt_object_exists(child)) {
1814                                 /* We have to get attr & LOV EA & HSM for this
1815                                  * object */
1816                                 ma->ma_need |= MA_HSM;
1817                                 result = mdt_attr_get_complex(info, child, ma);
1818                         } else {
1819                                 /*object non-exist!!!*/
1820                                 LBUG();
1821                         }
1822                 }
1823         }
1824
1825         LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1826
1827         /* get openlock if this is not replay and if a client requested it */
1828         if (!req_is_replay(req)) {
1829                 rc = mdt_object_open_lock(info, child, lhc, &ibits);
1830                 if (rc != 0)
1831                         GOTO(out_child_unlock, result = rc);
1832                 else if (create_flags & MDS_OPEN_LOCK)
1833                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1834         }
1835
1836         /* Try to open it now. */
1837         rc = mdt_finish_open(info, parent, child, create_flags,
1838                              created, ldlm_rep);
1839         if (rc) {
1840                 result = rc;
1841                 /* openlock will be released if mdt_finish_open failed */
1842                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1843
1844                 if (created && create_flags & MDS_OPEN_VOLATILE) {
1845                         CERROR("%s: cannot open volatile file "DFID", orphan "
1846                                "file will be left in PENDING directory until "
1847                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1848                                PFID(mdt_object_fid(child)), rc);
1849                         GOTO(out_child_unlock, result);
1850                 }
1851
1852                 if (created) {
1853                         ma->ma_need = 0;
1854                         ma->ma_valid = 0;
1855                         ma->ma_cookie_size = 0;
1856                         rc = mdo_unlink(info->mti_env,
1857                                         mdt_object_child(parent),
1858                                         mdt_object_child(child),
1859                                         lname,
1860                                         &info->mti_attr, 0);
1861                         if (rc != 0)
1862                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1863                                        mdt_obd_name(info->mti_mdt),
1864                                        PFID(mdt_object_fid(child)), rc);
1865                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1866                 }
1867         }
1868         EXIT;
1869 out_child_unlock:
1870         mdt_object_open_unlock(info, child, lhc, ibits, result);
1871 out_child:
1872         mdt_object_put(info->mti_env, child);
1873 out_parent:
1874         mdt_object_unlock_put(info, parent, lh, result || !created);
1875 out:
1876         if (result)
1877                 lustre_msg_set_transno(req->rq_repmsg, 0);
1878         return result;
1879 }
1880
1881 /**
1882  * Create an orphan object use local root.
1883  */
1884 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1885                                           struct mdt_device *mdt,
1886                                           const struct lu_fid *fid,
1887                                           struct md_attr *attr, fmode_t fmode)
1888 {
1889         const struct lu_env *env = info->mti_env;
1890         struct md_op_spec *spec = &info->mti_spec;
1891         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1892         struct mdt_object *obj = NULL;
1893         struct mdt_object *local_root;
1894         static const char name[] = "i_am_nobody";
1895         struct lu_name *lname;
1896         struct lu_ucred *uc;
1897         cfs_cap_t uc_cap_save;
1898         int rc;
1899         ENTRY;
1900
1901         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1902         if (rc != 0)
1903                 RETURN(ERR_PTR(rc));
1904
1905         local_root = mdt_object_find(env, mdt, local_root_fid);
1906         if (IS_ERR(local_root))
1907                 RETURN(local_root);
1908
1909         obj = mdt_object_new(env, mdt, fid);
1910         if (IS_ERR(obj))
1911                 GOTO(out, rc = PTR_ERR(obj));
1912
1913         spec->sp_cr_lookup = 0;
1914         spec->sp_feat = &dt_directory_features;
1915         spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1916         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1917         if (attr->ma_valid & MA_LOV) {
1918                 spec->u.sp_ea.eadata = attr->ma_lmm;
1919                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1920                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1921         } else {
1922                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1923         }
1924
1925         lname = mdt_name(env, (char *)name, sizeof(name) - 1);
1926
1927         uc = lu_ucred(env);
1928         uc_cap_save = uc->uc_cap;
1929         uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1930         rc = mdo_create(env, mdt_object_child(local_root), lname,
1931                         mdt_object_child(obj), spec, attr);
1932         uc->uc_cap = uc_cap_save;
1933         if (rc < 0) {
1934                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1935                        mdt_obd_name(mdt), PFID(fid), rc);
1936                 GOTO(out, rc);
1937         }
1938
1939         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1940         if (rc < 0)
1941                 CERROR("%s: cannot open volatile file "DFID", orphan "
1942                        "file will be left in PENDING directory until "
1943                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1944                        PFID(fid), rc);
1945         GOTO(out, rc);
1946
1947 out:
1948         if (rc < 0) {
1949                 if (!IS_ERR(obj))
1950                         mdt_object_put(env, obj);
1951                 obj = ERR_PTR(rc);
1952         }
1953         mdt_object_put(env, local_root);
1954         return obj;
1955 }
1956
1957 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1958                            struct md_attr *ma)
1959 {
1960         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1961         struct close_data      *data;
1962         struct ldlm_lock       *lease;
1963         struct mdt_object      *orphan;
1964         struct md_attr         *orp_ma;
1965         struct lu_buf          *buf;
1966         bool                    lease_broken;
1967         int                     rc;
1968         int                     rc2;
1969         ENTRY;
1970
1971         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1972                 RETURN(-EROFS);
1973
1974         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1975         if (data == NULL)
1976                 RETURN(-EPROTO);
1977
1978         lease = ldlm_handle2lock(&data->cd_handle);
1979         if (lease == NULL)
1980                 RETURN(-ESTALE);
1981
1982         /* try to hold open_sem so that nobody else can open the file */
1983         if (!down_write_trylock(&o->mot_open_sem)) {
1984                 ldlm_lock_cancel(lease);
1985                 GOTO(out_reprocess, rc = -EBUSY);
1986         }
1987
1988         /* Check if the lease open lease has already canceled */
1989         lock_res_and_lock(lease);
1990         lease_broken = ldlm_is_cancel(lease);
1991         unlock_res_and_lock(lease);
1992
1993         LDLM_DEBUG(lease, DFID " lease broken? %d\n",
1994                    PFID(mdt_object_fid(o)), lease_broken);
1995
1996         /* Cancel server side lease. Client side counterpart should
1997          * have been cancelled. It's okay to cancel it now as we've
1998          * held mot_open_sem. */
1999         ldlm_lock_cancel(lease);
2000
2001         if (lease_broken) /* don't perform release task */
2002                 GOTO(out_unlock, rc = -ESTALE);
2003
2004         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2005                 GOTO(out_unlock, rc = -EINVAL);
2006
2007         /* ma_need was set before but it seems fine to change it in order to
2008          * avoid modifying the one from RPC */
2009         ma->ma_need = MA_HSM;
2010         rc = mdt_attr_get_complex(info, o, ma);
2011         if (rc != 0)
2012                 GOTO(out_unlock, rc);
2013
2014         if (!mdt_hsm_release_allow(ma))
2015                 GOTO(out_unlock, rc = -EPERM);
2016
2017         /* already released? */
2018         if (ma->ma_hsm.mh_flags & HS_RELEASED)
2019                 GOTO(out_unlock, rc = 0);
2020
2021         /* Compare on-disk and packed data_version */
2022         if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
2023                 CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
2024                        " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
2025                        data->cd_data_version, ma->ma_hsm.mh_arch_ver);
2026                 GOTO(out_unlock, rc = -EPERM);
2027         }
2028
2029         ma->ma_valid = MA_INODE;
2030         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
2031         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
2032         if (rc < 0)
2033                 GOTO(out_unlock, rc);
2034
2035         ma->ma_need = MA_INODE | MA_LOV;
2036         rc = mdt_attr_get_complex(info, o, ma);
2037         if (rc < 0)
2038                 GOTO(out_unlock, rc);
2039
2040         if (!(ma->ma_valid & MA_LOV)) {
2041                 /* Even empty file are released */
2042                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2043                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2044                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2045                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2046                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2047         } else {
2048                 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
2049                  * ma_lmm as lov_user_md, then it will be confused by union of
2050                  * layout_gen and stripe_offset. */
2051                 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
2052                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2053                 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
2054                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
2055                 else
2056                         GOTO(out_unlock, rc = -EINVAL);
2057         }
2058
2059         /* Set file as released */
2060         ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
2061
2062         /* Hopefully it's not used in this call path */
2063         orp_ma = &info->mti_u.som.attr;
2064         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2065         orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
2066         orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
2067         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2068         orp_ma->ma_lmm = ma->ma_lmm;
2069         orp_ma->ma_lmm_size = ma->ma_lmm_size;
2070         orp_ma->ma_valid = MA_INODE | MA_LOV;
2071         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2072                                  FMODE_WRITE);
2073         if (IS_ERR(orphan)) {
2074                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2075                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2076                        PTR_ERR(orphan));
2077                 GOTO(out_unlock, rc = PTR_ERR(orphan));
2078         }
2079
2080         /* Set up HSM attribute for orphan object */
2081         CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
2082         buf = &info->mti_buf;
2083         buf->lb_buf = info->mti_xattr_buf;
2084         buf->lb_len = sizeof(struct hsm_attrs);
2085         ma->ma_hsm.mh_flags |= HS_RELEASED;
2086         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2087         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2088
2089         mdt_lock_reg_init(lh, LCK_EX);
2090         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2091                              MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
2092         if (rc != 0)
2093                 GOTO(out_close, rc);
2094
2095         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2096                           XATTR_NAME_HSM, 0);
2097
2098         if (rc == 0)
2099                 /* Swap layout with orphan object */
2100                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2101                                      mdt_object_child(orphan),
2102                                      SWAP_LAYOUTS_MDS_HSM);
2103
2104         /* Release exclusive LL */
2105         mdt_object_unlock(info, o, lh, 1);
2106
2107         EXIT;
2108
2109 out_close:
2110         /* Close orphan object anyway */
2111         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2112                        FMODE_WRITE);
2113         if (rc2 < 0)
2114                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2115                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2116         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2117                         "object closed\n");
2118         mdt_object_put(info->mti_env, orphan);
2119
2120 out_unlock:
2121         up_write(&o->mot_open_sem);
2122
2123         if (rc == 0) { /* already released */
2124                 struct mdt_body *repbody;
2125                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2126                 LASSERT(repbody != NULL);
2127                 repbody->valid |= OBD_MD_FLRELEASED;
2128         }
2129
2130 out_reprocess:
2131         ldlm_reprocess_all(lease->l_resource);
2132         LDLM_LOCK_PUT(lease);
2133
2134         ma->ma_valid = 0;
2135         ma->ma_need = 0;
2136
2137         return rc;
2138 }
2139
2140 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
2141                                       MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
2142
2143 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2144 {
2145         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
2146 }
2147
2148 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2149 {
2150         struct mdt_object *o = mfd->mfd_object;
2151         struct md_object *next = mdt_object_child(o);
2152         struct md_attr *ma = &info->mti_attr;
2153         int ret = MDT_IOEPOCH_CLOSED;
2154         int rc = 0;
2155         __u64 mode;
2156         ENTRY;
2157
2158         mode = mfd->mfd_mode;
2159
2160         if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
2161                 rc = mdt_hsm_release(info, o, ma);
2162                 if (rc < 0) {
2163                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2164                                 mdt_obd_name(info->mti_mdt),
2165                                 PFID(mdt_object_fid(o)), rc);
2166                         /* continue to close even error occurred. */
2167                 }
2168         }
2169
2170         if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
2171                 mdt_write_put(o);
2172                 ret = mdt_ioepoch_close(info, o);
2173         } else if (mode & MDS_FMODE_EXEC) {
2174                 mdt_write_allow(o);
2175         } else if (mode & MDS_FMODE_EPOCH) {
2176                 ret = mdt_ioepoch_close(info, o);
2177         } else if (mode & MDS_FMODE_SOM) {
2178                 ret = mdt_som_au_close(info, o);
2179         }
2180
2181         /* Update atime on close only. */
2182         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
2183             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
2184                 /* Set the atime only. */
2185                 ma->ma_valid = MA_INODE;
2186                 ma->ma_attr.la_valid = LA_ATIME;
2187                 rc = mo_attr_set(info->mti_env, next, ma);
2188         }
2189
2190         /* If file data is modified, add the dirty flag. */
2191         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2192                 rc = mdt_add_dirty_flag(info, o, ma);
2193
2194         ma->ma_need |= MA_INODE;
2195         ma->ma_valid &= ~MA_INODE;
2196
2197         if (!MFD_CLOSED(mode))
2198                 rc = mo_close(info->mti_env, next, ma, mode);
2199
2200         if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
2201                 struct mdt_export_data *med;
2202
2203                 /* The IOepoch is still opened or SOM update is needed.
2204                  * Put mfd back into the list. */
2205                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
2206                 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
2207                                       MDS_FMODE_EPOCH : MDS_FMODE_SOM);
2208
2209                 LASSERT(mdt_info_req(info));
2210                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
2211                 spin_lock(&med->med_open_lock);
2212                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
2213                 class_handle_hash_back(&mfd->mfd_handle);
2214                 spin_unlock(&med->med_open_lock);
2215
2216                 if (ret == MDT_IOEPOCH_OPENED) {
2217                         ret = 0;
2218                 } else {
2219                         ret = -EAGAIN;
2220                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
2221                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
2222                 }
2223         } else {
2224                 /* adjust open and lease count */
2225                 if (mode & MDS_OPEN_LEASE) {
2226                         LASSERT(atomic_read(&o->mot_lease_count) > 0);
2227                         atomic_dec(&o->mot_lease_count);
2228                 }
2229                 LASSERT(atomic_read(&o->mot_open_count) > 0);
2230                 atomic_dec(&o->mot_open_count);
2231
2232                 mdt_mfd_free(mfd);
2233                 mdt_object_put(info->mti_env, o);
2234         }
2235
2236         RETURN(rc ? rc : ret);
2237 }
2238
2239 int mdt_close(struct tgt_session_info *tsi)
2240 {
2241         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2242         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2243         struct mdt_export_data *med;
2244         struct mdt_file_data   *mfd;
2245         struct mdt_object      *o;
2246         struct md_attr         *ma = &info->mti_attr;
2247         struct mdt_body        *repbody = NULL;
2248         int rc, ret = 0;
2249         ENTRY;
2250
2251         mdt_counter_incr(req, LPROC_MDT_CLOSE);
2252         /* Close may come with the Size-on-MDS update. Unpack it. */
2253         rc = mdt_close_unpack(info);
2254         if (rc)
2255                 GOTO(out, rc = err_serious(rc));
2256
2257         LASSERT(info->mti_ioepoch);
2258
2259         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2260                              info->mti_mdt->mdt_max_mdsize);
2261         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2262                              info->mti_mdt->mdt_max_cookiesize);
2263         rc = req_capsule_server_pack(info->mti_pill);
2264         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2265                 mdt_client_compatibility(info);
2266                 if (rc == 0)
2267                         mdt_fix_reply(info);
2268                 mdt_exit_ucred(info);
2269                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2270         }
2271
2272         /* Continue to close handle even if we can not pack reply */
2273         if (rc == 0) {
2274                 repbody = req_capsule_server_get(info->mti_pill,
2275                                                  &RMF_MDT_BODY);
2276                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2277                                                     &RMF_MDT_MD);
2278                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2279                                                        &RMF_MDT_MD,
2280                                                        RCL_SERVER);
2281                 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
2282                                                        &RMF_LOGCOOKIES);
2283                 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
2284                                                           &RMF_LOGCOOKIES,
2285                                                           RCL_SERVER);
2286                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
2287                 repbody->eadatasize = 0;
2288                 repbody->aclsize = 0;
2289         } else {
2290                 rc = err_serious(rc);
2291         }
2292
2293         med = &req->rq_export->exp_mdt_data;
2294         spin_lock(&med->med_open_lock);
2295         mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2296                              req_is_replay(req));
2297         if (mdt_mfd_closed(mfd)) {
2298                 spin_unlock(&med->med_open_lock);
2299                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2300                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
2301                        info->mti_ioepoch->handle.cookie);
2302                 /** not serious error since bug 3633 */
2303                 rc = -ESTALE;
2304         } else {
2305                 class_handle_unhash(&mfd->mfd_handle);
2306                 cfs_list_del_init(&mfd->mfd_list);
2307                 spin_unlock(&med->med_open_lock);
2308
2309                 /* Do not lose object before last unlink. */
2310                 o = mfd->mfd_object;
2311                 mdt_object_get(info->mti_env, o);
2312                 ret = mdt_mfd_close(info, mfd);
2313                 if (repbody != NULL)
2314                         rc = mdt_handle_last_unlink(info, o, ma);
2315                 mdt_empty_transno(info, rc);
2316                 mdt_object_put(info->mti_env, o);
2317         }
2318         if (repbody != NULL) {
2319                 mdt_client_compatibility(info);
2320                 rc = mdt_fix_reply(info);
2321         }
2322
2323         mdt_exit_ucred(info);
2324         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2325                 GOTO(out, rc = err_serious(-ENOMEM));
2326
2327         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2328                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2329                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2330 out:
2331         mdt_thread_info_fini(info);
2332         RETURN(rc ? rc : ret);
2333 }
2334
2335 /**
2336  * DONE_WRITING rpc handler.
2337  *
2338  * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
2339  * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
2340  * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
2341  * skip attributes and reconstruct the reply here.
2342  */
2343 int mdt_done_writing(struct tgt_session_info *tsi)
2344 {
2345         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2346         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2347         struct mdt_body         *repbody = NULL;
2348         struct mdt_export_data  *med;
2349         struct mdt_file_data    *mfd;
2350         int rc;
2351         ENTRY;
2352
2353         rc = req_capsule_server_pack(tsi->tsi_pill);
2354         if (rc)
2355                 GOTO(out, rc = err_serious(rc));
2356
2357         repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2358         repbody->eadatasize = 0;
2359         repbody->aclsize = 0;
2360
2361         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
2362         rc = mdt_close_unpack(info);
2363         if (rc)
2364                 GOTO(out, rc = err_serious(rc));
2365
2366         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2367                 mdt_exit_ucred(info);
2368                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2369         }
2370
2371         med = &info->mti_exp->exp_mdt_data;
2372         spin_lock(&med->med_open_lock);
2373         mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2374                              req_is_replay(req));
2375         if (mfd == NULL) {
2376                 spin_unlock(&med->med_open_lock);
2377                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
2378                        ": cookie = "LPX64" ioepoch = "LPU64"\n",
2379                        PFID(info->mti_rr.rr_fid1),
2380                        info->mti_ioepoch->handle.cookie,
2381                        info->mti_ioepoch->ioepoch);
2382                 /* If this is a replay, reconstruct the transno. */
2383                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2384                         rc = info->mti_ioepoch->flags & MF_SOM_AU ?
2385                              -EAGAIN : 0;
2386                         mdt_empty_transno(info, rc);
2387                 } else
2388                         rc = -ESTALE;
2389                 GOTO(error_ucred, rc);
2390         }
2391
2392         LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
2393                 mfd->mfd_mode == MDS_FMODE_TRUNC);
2394         class_handle_unhash(&mfd->mfd_handle);
2395         cfs_list_del_init(&mfd->mfd_list);
2396         spin_unlock(&med->med_open_lock);
2397
2398         /* Set EPOCH CLOSE flag if not set by client. */
2399         info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
2400         info->mti_attr.ma_valid = 0;
2401
2402         info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
2403         OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2404         if (info->mti_attr.ma_lmm == NULL)
2405                 GOTO(error_ucred, rc = -ENOMEM);
2406
2407         rc = mdt_mfd_close(info, mfd);
2408
2409         OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2410         mdt_empty_transno(info, rc);
2411 error_ucred:
2412         mdt_exit_ucred(info);
2413 out:
2414         mdt_thread_info_fini(info);
2415         RETURN(rc);
2416 }