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