lustre/mdt/mdt_open.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/mdt/mdt_open.c
  37  *
  38  * Lustre Metadata Target (mdt) open/close file handling
  39  *
  40  * Author: Huang Hua <huanghua@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_MDS
  44
  45 #include <lustre_acl.h>
  46 #include <lustre_mds.h>
  47 #include "mdt_internal.h"
  48
  49 /* we do nothing because we do not have refcount now */
  50 static void mdt_mfd_get(void *mfdp)
  51 {
  52 }
  53
  54 static struct portals_handle_ops mfd_handle_ops = {
  55         .hop_addref = mdt_mfd_get,
  56         .hop_free   = NULL,
  57 };
  58
  59 /* Create a new mdt_file_data struct, initialize it,
  60  * and insert it to global hash table */
  61 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
  62 {
  63         struct mdt_file_data *mfd;
  64         ENTRY;
  65
  66         OBD_ALLOC_PTR(mfd);
  67         if (mfd != NULL) {
  68                 CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  69                 mfd->mfd_handle.h_owner = med;
  70                 CFS_INIT_LIST_HEAD(&mfd->mfd_list);
  71                 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
  72         }
  73
  74         RETURN(mfd);
  75 }
  76
  77 /*
  78  * Find the mfd pointed to by handle in global hash table.
  79  * In case of replay the handle is obsoleted
  80  * but mfd can be found in mfd list by that handle.
  81  * Callers need to be holding med_open_lock.
  82  */
  83 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
  84                                      const struct lustre_handle *handle,
  85                                      bool is_replay_or_resent)
  86 {
  87         struct mdt_file_data   *mfd;
  88         ENTRY;
  89
  90         LASSERT(handle != NULL);
  91         mfd = class_handle2object(handle->cookie, med);
  92         /* during dw/setattr replay the mfd can be found by old handle */
  93         if (mfd == NULL && is_replay_or_resent) {
  94                 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  95                         if (mfd->mfd_old_handle.cookie == handle->cookie)
  96                                 RETURN(mfd);
  97                 }
  98                 mfd = NULL;
  99         }
 100
 101         RETURN(mfd);
 102 }
 103
 104 /* free mfd */
 105 void mdt_mfd_free(struct mdt_file_data *mfd)
 106 {
 107         LASSERT(cfs_list_empty(&mfd->mfd_list));
 108         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 109 }
 110
 111 static int mdt_create_data(struct mdt_thread_info *info,
 112                            struct mdt_object *p, struct mdt_object *o)
 113 {
 114         struct md_op_spec     *spec = &info->mti_spec;
 115         struct md_attr        *ma   = &info->mti_attr;
 116         int                    rc   = 0;
 117         ENTRY;
 118
 119         if (!md_should_create(spec->sp_cr_flags))
 120                 RETURN(0);
 121
 122         ma->ma_need = MA_INODE | MA_LOV;
 123         ma->ma_valid = 0;
 124         mutex_lock(&o->mot_lov_mutex);
 125         if (!(o->mot_flags & MOF_LOV_CREATED)) {
 126                 if (p != NULL && !fid_is_md_operative(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 /**
 666  * prep ma_lmm/ma_lmv for md_attr from reply
 667  */
 668 void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
 669                               struct mdt_object *obj,
 670                               struct md_attr *ma)
 671 {
 672         LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
 673         if (S_ISDIR(obj->mot_header.loh_attr)) {
 674                 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
 675                                                     &RMF_MDT_MD);
 676                 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
 677                                                        &RMF_MDT_MD,
 678                                                        RCL_SERVER);
 679                 if (ma->ma_lmv_size > 0)
 680                         ma->ma_need |= MA_LMV;
 681         } else {
 682                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
 683                                                     &RMF_MDT_MD);
 684                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 685                                                        &RMF_MDT_MD,
 686                                                        RCL_SERVER);
 687                 if (ma->ma_lmm_size > 0)
 688                         ma->ma_need |= MA_LOV;
 689         }
 690 }
 691
 692 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 693                         struct mdt_object *o, __u64 flags, int created,
 694                         struct ldlm_reply *rep)
 695 {
 696         struct ptlrpc_request   *req = mdt_info_req(info);
 697         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 698         struct mdt_file_data    *mfd;
 699         struct md_attr          *ma  = &info->mti_attr;
 700         struct lu_attr          *la  = &ma->ma_attr;
 701         struct mdt_body         *repbody;
 702         int                      rc = 0, isdir, isreg;
 703         ENTRY;
 704
 705         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 706
 707         isreg = S_ISREG(la->la_mode);
 708         isdir = S_ISDIR(la->la_mode);
 709         if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
 710                 /*
 711                  * No EA, check whether it is will set regEA and dirEA since in
 712                  * above attr get, these size might be zero, so reset it, to
 713                  * retrieve the MD after create obj.
 714                  */
 715                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 716                                                        &RMF_MDT_MD,
 717                                                        RCL_SERVER);
 718                 /* in replay case, p == NULL */
 719                 rc = mdt_create_data(info, p, o);
 720                 if (rc)
 721                         RETURN(rc);
 722
 723                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
 724                         mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
 725         }
 726
 727         CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
 728                ma->ma_valid, ma->ma_lmm_size);
 729
 730         if (ma->ma_valid & MA_LOV) {
 731                 LASSERT(ma->ma_lmm_size != 0);
 732                 repbody->eadatasize = ma->ma_lmm_size;
 733                 if (isdir)
 734                         repbody->valid |= OBD_MD_FLDIREA;
 735                 else
 736                         repbody->valid |= OBD_MD_FLEASIZE;
 737         }
 738
 739         if (ma->ma_valid & MA_LMV) {
 740                 LASSERT(ma->ma_lmv_size != 0);
 741                 repbody->eadatasize = ma->ma_lmv_size;
 742                 LASSERT(isdir);
 743                 repbody->valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
 744         }
 745
 746         if (flags & FMODE_WRITE) {
 747                 rc = mdt_write_get(o);
 748                 if (rc == 0) {
 749                         mdt_ioepoch_open(info, o, created);
 750                         repbody->ioepoch = o->mot_ioepoch;
 751                 }
 752         } else if (flags & MDS_FMODE_EXEC) {
 753                 rc = mdt_write_deny(o);
 754         }
 755         if (rc)
 756                 RETURN(rc);
 757
 758         rc = mo_open(info->mti_env, mdt_object_child(o),
 759                      created ? flags | MDS_OPEN_CREATED : flags);
 760         if (rc)
 761                 GOTO(err_out, rc);
 762
 763         mfd = mdt_mfd_new(med);
 764         if (mfd == NULL)
 765                 GOTO(err_out, rc = -ENOMEM);
 766
 767         /*
 768          * Keep a reference on this object for this open, and is
 769          * released by mdt_mfd_close().
 770          */
 771         mdt_object_get(info->mti_env, o);
 772         mfd->mfd_object = o;
 773         mfd->mfd_xid = req->rq_xid;
 774
 775         /*
 776          * @flags is always not zero. At least it should be FMODE_READ,
 777          * FMODE_WRITE or MDS_FMODE_EXEC.
 778          */
 779         LASSERT(flags != 0);
 780
 781         /* Open handling. */
 782         mdt_mfd_set_mode(mfd, flags);
 783
 784         atomic_inc(&o->mot_open_count);
 785         if (flags & MDS_OPEN_LEASE)
 786                 atomic_inc(&o->mot_lease_count);
 787
 788         /* replay handle */
 789         if (req_is_replay(req)) {
 790                 struct mdt_file_data *old_mfd;
 791                 /* Check wheather old cookie already exist in
 792                  * the list, becasue when do recovery, client
 793                  * might be disconnected from server, and
 794                  * restart replay, so there maybe some orphan
 795                  * mfd here, we should remove them */
 796                 LASSERT(info->mti_rr.rr_handle != NULL);
 797                 spin_lock(&med->med_open_lock);
 798                 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
 799                 if (old_mfd != NULL) {
 800                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 801                                "cookie = "LPX64"\n", mfd,
 802                                PFID(mdt_object_fid(mfd->mfd_object)),
 803                                info->mti_rr.rr_handle->cookie);
 804                         class_handle_unhash(&old_mfd->mfd_handle);
 805                         cfs_list_del_init(&old_mfd->mfd_list);
 806                         spin_unlock(&med->med_open_lock);
 807                         /* no attr update for that close */
 808                         la->la_valid = 0;
 809                         ma->ma_valid |= MA_FLAGS;
 810                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 811                         mdt_mfd_close(info, old_mfd);
 812                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 813                         ma->ma_valid &= ~MA_FLAGS;
 814                 } else {
 815                         spin_unlock(&med->med_open_lock);
 816                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 817                                "cookie = "LPX64"\n",
 818                                PFID(mdt_object_fid(mfd->mfd_object)),
 819                                info->mti_rr.rr_handle->cookie);
 820                 }
 821
 822                 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
 823                        info->mti_rr.rr_handle->cookie);
 824
 825                 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
 826         }
 827
 828         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 829
 830         if (req->rq_export->exp_disconnected) {
 831                 spin_lock(&med->med_open_lock);
 832                 class_handle_unhash(&mfd->mfd_handle);
 833                 cfs_list_del_init(&mfd->mfd_list);
 834                 spin_unlock(&med->med_open_lock);
 835                 mdt_mfd_close(info, mfd);
 836         } else {
 837                 spin_lock(&med->med_open_lock);
 838                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
 839                 spin_unlock(&med->med_open_lock);
 840         }
 841
 842         mdt_empty_transno(info, rc);
 843
 844         RETURN(rc);
 845
 846 err_out:
 847         if (flags & FMODE_WRITE)
 848                 /* XXX We also need to close io epoch here.
 849                  * See LU-1220 - green */
 850                 mdt_write_put(o);
 851         else if (flags & MDS_FMODE_EXEC)
 852                 mdt_write_allow(o);
 853
 854         return rc;
 855 }
 856
 857 int mdt_finish_open(struct mdt_thread_info *info,
 858                     struct mdt_object *p, struct mdt_object *o,
 859                     __u64 flags, int created, struct ldlm_reply *rep)
 860 {
 861         struct ptlrpc_request   *req = mdt_info_req(info);
 862         struct obd_export       *exp = req->rq_export;
 863         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 864         struct md_attr          *ma  = &info->mti_attr;
 865         struct lu_attr          *la  = &ma->ma_attr;
 866         struct mdt_file_data    *mfd;
 867         struct mdt_body         *repbody;
 868         int                      rc = 0;
 869         int                      isreg, isdir, islnk;
 870         cfs_list_t              *t;
 871         ENTRY;
 872
 873         LASSERT(ma->ma_valid & MA_INODE);
 874
 875         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 876
 877         isreg = S_ISREG(la->la_mode);
 878         isdir = S_ISDIR(la->la_mode);
 879         islnk = S_ISLNK(la->la_mode);
 880         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 881
 882         /* LU-2275, simulate broken behaviour (esp. prevalent in
 883          * pre-2.4 servers where a very strange reply is sent on error
 884          * that looks like it was actually almost succesful and a failure at the
 885          * same time */
 886         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 887                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 888                                                DISP_LOOKUP_NEG |
 889                                                DISP_LOOKUP_POS);
 890
 891                 if (flags & MDS_OPEN_LOCK)
 892                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 893
 894                 RETURN(-ENOENT);
 895         }
 896
 897         if (exp_connect_rmtclient(exp)) {
 898                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 899
 900                 rc = mdt_pack_remote_perm(info, o, buf);
 901                 if (rc) {
 902                         repbody->valid &= ~OBD_MD_FLRMTPERM;
 903                         repbody->aclsize = 0;
 904                 } else {
 905                         repbody->valid |= OBD_MD_FLRMTPERM;
 906                         repbody->aclsize = sizeof(struct mdt_remote_perm);
 907                 }
 908         }
 909 #ifdef CONFIG_FS_POSIX_ACL
 910         else if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 911                 const struct lu_env *env = info->mti_env;
 912                 struct md_object *next = mdt_object_child(o);
 913                 struct lu_buf *buf = &info->mti_buf;
 914
 915                 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 916                 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
 917                                                    RCL_SERVER);
 918                 if (buf->lb_len > 0) {
 919                         rc = mo_xattr_get(env, next, buf,
 920                                           XATTR_NAME_ACL_ACCESS);
 921                         if (rc < 0) {
 922                                 if (rc == -ENODATA) {
 923                                         repbody->aclsize = 0;
 924                                         repbody->valid |= OBD_MD_FLACL;
 925                                         rc = 0;
 926                                 } else if (rc == -EOPNOTSUPP) {
 927                                         rc = 0;
 928                                 } else {
 929                                         CERROR("got acl size: %d\n", rc);
 930                                 }
 931                         } else {
 932                                 repbody->aclsize = rc;
 933                                 repbody->valid |= OBD_MD_FLACL;
 934                                 rc = 0;
 935                         }
 936                 }
 937         }
 938 #endif
 939
 940         if (info->mti_mdt->mdt_lut.lut_mds_capa &&
 941             exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) {
 942                 struct lustre_capa *capa;
 943
 944                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
 945                 LASSERT(capa);
 946                 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
 947                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 948                 if (rc)
 949                         RETURN(rc);
 950                 repbody->valid |= OBD_MD_FLMDSCAPA;
 951         }
 952         if (info->mti_mdt->mdt_lut.lut_oss_capa &&
 953             exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA &&
 954             S_ISREG(lu_object_attr(&o->mot_obj))) {
 955                 struct lustre_capa *capa;
 956
 957                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
 958                 LASSERT(capa);
 959                 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
 960                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 961                 if (rc)
 962                         RETURN(rc);
 963                 repbody->valid |= OBD_MD_FLOSSCAPA;
 964         }
 965
 966         /*
 967          * If we are following a symlink, don't open; and do not return open
 968          * handle for special nodes as client required.
 969          */
 970         if (islnk || (!isreg && !isdir &&
 971             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 972                 lustre_msg_set_transno(req->rq_repmsg, 0);
 973                 RETURN(0);
 974         }
 975
 976         /*
 977          * We need to return the existing object's fid back, so it is done here,
 978          * after preparing the reply.
 979          */
 980         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 981                 RETURN(-EEXIST);
 982
 983         /* This can't be done earlier, we need to return reply body */
 984         if (isdir) {
 985                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 986                         /* We are trying to create or write an existing dir. */
 987                         RETURN(-EISDIR);
 988                 }
 989         } else if (flags & MDS_OPEN_DIRECTORY)
 990                 RETURN(-ENOTDIR);
 991
 992         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 993                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 994                 RETURN(-EAGAIN);
 995         }
 996
 997         mfd = NULL;
 998         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 999                 spin_lock(&med->med_open_lock);
1000                 cfs_list_for_each(t, &med->med_open_head) {
1001                         mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
1002                         if (mfd->mfd_xid == req->rq_xid)
1003                                 break;
1004                         mfd = NULL;
1005                 }
1006                 spin_unlock(&med->med_open_lock);
1007
1008                 if (mfd != NULL) {
1009                         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
1010                         /*set repbody->ea_size for resent case*/
1011                         if (ma->ma_valid & MA_LOV) {
1012                                 LASSERT(ma->ma_lmm_size != 0);
1013                                 repbody->eadatasize = ma->ma_lmm_size;
1014                                 if (isdir)
1015                                         repbody->valid |= OBD_MD_FLDIREA;
1016                                 else
1017                                         repbody->valid |= OBD_MD_FLEASIZE;
1018                         }
1019                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
1020                         RETURN(0);
1021                 }
1022         }
1023
1024         rc = mdt_mfd_open(info, p, o, flags, created, rep);
1025         if (!rc)
1026                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
1027
1028         RETURN(rc);
1029 }
1030
1031 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
1032                              struct lsd_client_data *lcd);
1033
1034 void mdt_reconstruct_open(struct mdt_thread_info *info,
1035                           struct mdt_lock_handle *lhc)
1036 {
1037         const struct lu_env *env = info->mti_env;
1038         struct mdt_device       *mdt  = info->mti_mdt;
1039         struct req_capsule      *pill = info->mti_pill;
1040         struct ptlrpc_request   *req  = mdt_info_req(info);
1041         struct tg_export_data   *ted  = &req->rq_export->exp_target_data;
1042         struct lsd_client_data  *lcd  = ted->ted_lcd;
1043         struct md_attr          *ma   = &info->mti_attr;
1044         struct mdt_reint_record *rr   = &info->mti_rr;
1045         __u64                   flags = info->mti_spec.sp_cr_flags;
1046         struct ldlm_reply       *ldlm_rep;
1047         struct mdt_object       *parent;
1048         struct mdt_object       *child;
1049         struct mdt_body         *repbody;
1050         int                      rc;
1051         ENTRY;
1052
1053         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1054         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1055         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1056
1057         ma->ma_need = MA_INODE | MA_HSM;
1058         ma->ma_valid = 0;
1059
1060         mdt_req_from_lcd(req, lcd);
1061         mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1062
1063         CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1064                ldlm_rep->lock_policy_res1, req->rq_status);
1065
1066         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1067             req->rq_status != 0)
1068                 /* We did not create successfully, return error to client. */
1069                 GOTO(out, rc = req->rq_status);
1070
1071         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1072                 struct obd_export *exp = req->rq_export;
1073                 /*
1074                  * We failed after creation, but we do not know in which step
1075                  * we failed. So try to check the child object.
1076                  */
1077                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1078                 if (IS_ERR(parent)) {
1079                         rc = PTR_ERR(parent);
1080                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
1081                                       " Evicting client %s with export %s.\n",
1082                                       PFID(rr->rr_fid1), rc,
1083                                       obd_uuid2str(&exp->exp_client_uuid),
1084                                       obd_export_nid2str(exp));
1085                         mdt_export_evict(exp);
1086                         RETURN_EXIT;
1087                 }
1088                 child = mdt_object_find(env, mdt, rr->rr_fid2);
1089                 if (IS_ERR(child)) {
1090                         rc = PTR_ERR(child);
1091                         LCONSOLE_WARN("Child "DFID" lookup error %d."
1092                                       " Evicting client %s with export %s.\n",
1093                                       PFID(mdt_object_fid(child)), rc,
1094                                       obd_uuid2str(&exp->exp_client_uuid),
1095                                       obd_export_nid2str(exp));
1096                         mdt_object_put(env, parent);
1097                         mdt_export_evict(exp);
1098                         RETURN_EXIT;
1099                 }
1100
1101                 if (unlikely(mdt_object_remote(child))) {
1102                         /* the child object was created on remote server */
1103                         if (!mdt_is_dne_client(exp)) {
1104                                 /* Return -EIO for old client */
1105                                 mdt_object_put(env, parent);
1106                                 mdt_object_put(env, child);
1107                                 GOTO(out, rc = -EIO);
1108                         }
1109                         repbody->fid1 = *rr->rr_fid2;
1110                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1111                         rc = 0;
1112                 } else {
1113                         if (mdt_object_exists(child)) {
1114                                 mdt_set_capainfo(info, 1, rr->rr_fid2,
1115                                                  BYPASS_CAPA);
1116                                 mdt_prep_ma_buf_from_rep(info, child, ma);
1117                                 rc = mdt_attr_get_complex(info, child, ma);
1118                                 if (rc == 0)
1119                                         rc = mdt_finish_open(info, parent,
1120                                                              child, flags,
1121                                                              1, ldlm_rep);
1122                         } else {
1123                                 /* the child does not exist, we should do
1124                                  * regular open */
1125                                 mdt_object_put(env, parent);
1126                                 mdt_object_put(env, child);
1127                                 GOTO(regular_open, 0);
1128                         }
1129                 }
1130                 mdt_object_put(env, parent);
1131                 mdt_object_put(env, child);
1132                 GOTO(out, rc);
1133         } else {
1134 regular_open:
1135                 /* We did not try to create, so we are a pure open */
1136                 rc = mdt_reint_open(info, lhc);
1137         }
1138
1139         EXIT;
1140 out:
1141         req->rq_status = rc;
1142         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1143         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1144 }
1145
1146 int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
1147 {
1148         __u64                    flags = info->mti_spec.sp_cr_flags;
1149         struct mdt_reint_record *rr = &info->mti_rr;
1150         struct md_attr          *ma = &info->mti_attr;
1151         struct mdt_object       *o;
1152         int                      rc;
1153         ENTRY;
1154
1155         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1156         if (IS_ERR(o))
1157                 RETURN(rc = PTR_ERR(o));
1158
1159         if (unlikely(mdt_object_remote(o))) {
1160                 /* the child object was created on remote server */
1161                 struct mdt_body *repbody;
1162
1163                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1164                                                 DISP_LOOKUP_EXECD |
1165                                                 DISP_LOOKUP_POS));
1166                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1167                 repbody->fid1 = *rr->rr_fid2;
1168                 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1169                 rc = 0;
1170         } else {
1171                 if (mdt_object_exists(o)) {
1172                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1173                                                         DISP_LOOKUP_EXECD |
1174                                                         DISP_LOOKUP_POS));
1175                         mdt_prep_ma_buf_from_rep(info, o, ma);
1176                         rc = mdt_attr_get_complex(info, o, ma);
1177                         if (rc == 0)
1178                                 rc = mdt_finish_open(info, NULL, o, flags, 0,
1179                                                      rep);
1180                 } else {
1181                         rc = -ENOENT;
1182                 }
1183         }
1184
1185         mdt_object_put(info->mti_env, o);
1186         RETURN(rc);
1187 }
1188
1189 /* lock object for open */
1190 static int mdt_object_open_lock(struct mdt_thread_info *info,
1191                                 struct mdt_object *obj,
1192                                 struct mdt_lock_handle *lhc,
1193                                 __u64 *ibits)
1194 {
1195         struct md_attr  *ma = &info->mti_attr;
1196         __u64            open_flags = info->mti_spec.sp_cr_flags;
1197         ldlm_mode_t      lm = LCK_CR;
1198         bool             acq_lease = !!(open_flags & MDS_OPEN_LEASE);
1199         bool             try_layout = false;
1200         bool             create_layout = false;
1201         int              rc = 0;
1202         ENTRY;
1203
1204         *ibits = 0;
1205         mdt_lock_handle_init(lhc);
1206
1207         if (req_is_replay(mdt_info_req(info)))
1208                 RETURN(0);
1209
1210         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
1211                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
1212                     md_should_create(open_flags))
1213                         create_layout = true;
1214                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
1215                     ma->ma_need & MA_LOV)
1216                         try_layout = true;
1217         }
1218
1219         if (acq_lease) {
1220                 /* lease open, acquire write mode of open sem */
1221                 down_write(&obj->mot_open_sem);
1222
1223                 /* Lease exists and ask for new lease */
1224                 if (atomic_read(&obj->mot_lease_count) > 0) {
1225                         /* only exclusive open is supported, so lease
1226                          * are conflicted to each other */
1227                         GOTO(out, rc = -EBUSY);
1228                 }
1229
1230                 /* Lease must be with open lock */
1231                 if (!(open_flags & MDS_OPEN_LOCK)) {
1232                         CERROR("Request lease for file:"DFID ", but open lock "
1233                                 "is missed, open_flags = "LPO64".\n",
1234                                 PFID(mdt_object_fid(obj)), open_flags);
1235                         GOTO(out, rc = -EPROTO);
1236                 }
1237
1238                 /* XXX: only exclusive open is supported. */
1239                 lm = LCK_EX;
1240                 *ibits = MDS_INODELOCK_OPEN;
1241
1242                 /* never grant LCK_EX layout lock to client */
1243                 try_layout = false;
1244         } else { /* normal open */
1245                 /* normal open holds read mode of open sem */
1246                 down_read(&obj->mot_open_sem);
1247
1248                 if (open_flags & FMODE_WRITE)
1249                         lm = LCK_CW;
1250                 else if (open_flags & MDS_FMODE_EXEC)
1251                         lm = LCK_PR;
1252                 else
1253                         lm = LCK_CR;
1254
1255                 if (open_flags & MDS_OPEN_LOCK) {
1256                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
1257                 } else if (open_flags & (FMODE_WRITE | MDS_FMODE_EXEC) ||
1258                            atomic_read(&obj->mot_lease_count) > 0) {
1259                         /* We need to flush conflicting locks or revoke a lease.
1260                          * In either case there is no need to acquire a layout
1261                          * lock since it won't be returned to the client. */
1262                         try_layout = false;
1263                         *ibits = MDS_INODELOCK_OPEN;
1264                         lhc = &info->mti_lh[MDT_LH_LOCAL];
1265                 }
1266
1267                 CDEBUG(D_INODE, "normal open FID = "DFID", open_count = %d, "
1268                        "lm = %d\n",
1269                        PFID(mdt_object_fid(obj)),
1270                        atomic_read(&obj->mot_open_count), lm);
1271         }
1272
1273         mdt_lock_reg_init(lhc, lm);
1274
1275         /* one problem to return layout lock on open is that it may result
1276          * in too many layout locks cached on the client side. */
1277         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
1278                 /* return lookup lock to validate inode at the client side,
1279                  * this is pretty important otherwise mdt will return layout
1280                  * lock for each open.
1281                  * However this is a double-edged sword because changing
1282                  * permission will revoke huge # of LOOKUP locks. */
1283                 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
1284                 if (!mdt_object_lock_try(info, obj, lhc, *ibits,
1285                                          MDT_CROSS_LOCK)) {
1286                         *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
1287                         if (*ibits != 0)
1288                                 rc = mdt_object_lock(info, obj, lhc, *ibits,
1289                                                 MDT_CROSS_LOCK);
1290                 }
1291         } else if (*ibits != 0) {
1292                 rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK);
1293         }
1294
1295         CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
1296                 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
1297                 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
1298
1299         /* will change layout, revoke layout locks by enqueuing EX lock. */
1300         if (rc == 0 && create_layout) {
1301                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
1302
1303                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
1304                         ", open_flags = "LPO64"\n",
1305                         PFID(mdt_object_fid(obj)), open_flags);
1306
1307                 /* We cannot enqueue another lock for the same resource we
1308                  * already have a lock for, due to mechanics of waiting list
1309                  * iterating in ldlm, see LU-3601.
1310                  * As such we'll drop the open lock we just got above here,
1311                  * it's ok not to have this open lock as it's main purpose is to
1312                  * flush unused cached client open handles. */
1313                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1314                         mdt_object_unlock(info, obj, lhc, 1);
1315
1316                 LASSERT(!try_layout);
1317                 mdt_lock_handle_init(ll);
1318                 mdt_lock_reg_init(ll, LCK_EX);
1319                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
1320                                         MDT_LOCAL_LOCK);
1321
1322                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
1323         }
1324
1325         /* Check if there is any other open handles after acquiring
1326          * open lock. At this point, caching open handles have been revoked
1327          * by open lock.
1328          * XXX: Now only exclusive open is supported. Need to check the
1329          * type of open for generic lease support. */
1330         if (rc == 0 && acq_lease) {
1331                 struct ptlrpc_request *req = mdt_info_req(info);
1332                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1333                 struct mdt_file_data *mfd;
1334                 bool is_replay_or_resent;
1335                 int open_count = 0;
1336
1337                 /* For lease: application can open a file and then apply lease,
1338                  * @handle contains original open handle in that case.
1339                  * In recovery, open REQ will be replayed and the lease REQ may
1340                  * be resent that means the open handle is already stale, so we
1341                  * need to fix it up here by finding new handle. */
1342                 is_replay_or_resent = req_is_replay(req) ||
1343                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
1344
1345                 /* if the request is _not_ a replay request, rr_handle
1346                  * may be used to hold an openhandle which is issuing the
1347                  * lease request, so that this openhandle doesn't count. */
1348                 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
1349                                      is_replay_or_resent);
1350                 if (mfd != NULL)
1351                         ++open_count;
1352
1353                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
1354                         PFID(mdt_object_fid(obj)),
1355                         atomic_read(&obj->mot_open_count), open_count);
1356
1357                 if (atomic_read(&obj->mot_open_count) > open_count)
1358                         GOTO(out, rc = -EBUSY);
1359         }
1360         GOTO(out, rc);
1361
1362 out:
1363         RETURN(rc);
1364 }
1365
1366 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1367                                    struct mdt_object *obj,
1368                                    struct mdt_lock_handle *lhc,
1369                                    __u64 ibits, int rc)
1370 {
1371         __u64 open_flags = info->mti_spec.sp_cr_flags;
1372         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1373         ENTRY;
1374
1375         if (req_is_replay(mdt_info_req(info)))
1376                 RETURN_EXIT;
1377
1378         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1379          * return to client side. */
1380         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1381                 mdt_object_unlock(info, obj, ll, 1);
1382
1383         ll = &info->mti_lh[MDT_LH_LAYOUT];
1384         /* Release local layout lock, layout was created */
1385         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1386                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1387                 mdt_object_unlock(info, obj, ll, 1);
1388         }
1389
1390         if (open_flags & MDS_OPEN_LEASE)
1391                 up_write(&obj->mot_open_sem);
1392         else
1393                 up_read(&obj->mot_open_sem);
1394
1395         /* Cross-ref case, the lock should be returned to the client */
1396         if (ibits == 0 || rc == -EREMOTE)
1397                 RETURN_EXIT;
1398
1399         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
1400                 /* for the open request, the lock will only return to client
1401                  * if open or layout lock is granted. */
1402                 rc = 1;
1403         }
1404
1405         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1406                 struct ldlm_reply       *ldlm_rep;
1407
1408                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1409                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1410                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1411                         mdt_object_unlock(info, obj, lhc, 1);
1412         }
1413         RETURN_EXIT;
1414 }
1415
1416 /**
1417  * Check release is permitted for the current HSM flags.
1418  */
1419 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1420 {
1421         if (!(ma->ma_valid & MA_HSM))
1422                 return false;
1423
1424         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1425                 return false;
1426
1427         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1428                 return false;
1429
1430         return true;
1431 }
1432
1433 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1434                          struct mdt_lock_handle *lhc)
1435 {
1436         const struct lu_env     *env   = info->mti_env;
1437         struct mdt_device       *mdt   = info->mti_mdt;
1438         __u64                    flags = info->mti_spec.sp_cr_flags;
1439         struct mdt_reint_record *rr    = &info->mti_rr;
1440         struct md_attr          *ma    = &info->mti_attr;
1441         struct mdt_object       *parent= NULL;
1442         struct mdt_object       *o;
1443         int                      rc;
1444         __u64                    ibits = 0;
1445         ENTRY;
1446
1447         if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1448                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1449                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1450                         if (IS_ERR(parent)) {
1451                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1452                                        " for anonymous created %ld, try to"
1453                                        " use server-side parent.\n",
1454                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1455                                 parent = NULL;
1456                         }
1457                 }
1458                 if (parent == NULL)
1459                         ma->ma_need |= MA_PFID;
1460         }
1461
1462         o = mdt_object_find(env, mdt, rr->rr_fid2);
1463         if (IS_ERR(o))
1464                 RETURN(rc = PTR_ERR(o));
1465
1466         if (mdt_object_remote(o)) {
1467                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1468                        mdt_obd_name(info->mti_mdt),
1469                        PFID(rr->rr_fid2));
1470                 GOTO(out, rc = -EREMOTE);
1471         } else if (!mdt_object_exists(o)) {
1472                 mdt_set_disposition(info, rep,
1473                                     DISP_IT_EXECD |
1474                                     DISP_LOOKUP_EXECD |
1475                                     DISP_LOOKUP_NEG);
1476                 GOTO(out, rc = -ENOENT);
1477         }
1478
1479         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1480
1481         mdt_prep_ma_buf_from_rep(info, o, ma);
1482         if (flags & MDS_OPEN_RELEASE)
1483                 ma->ma_need |= MA_HSM;
1484         rc = mdt_attr_get_complex(info, o, ma);
1485         if (rc)
1486                 GOTO(out, rc);
1487
1488         /* If a release request, check file flags are fine and ask for an
1489          * exclusive open access. */
1490         if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1491                 GOTO(out, rc = -EPERM);
1492
1493         rc = mdt_object_open_lock(info, o, lhc, &ibits);
1494         if (rc)
1495                 GOTO(out_unlock, rc);
1496
1497         if (ma->ma_valid & MA_PFID) {
1498                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1499                 if (IS_ERR(parent)) {
1500                         CDEBUG(D_INODE, "Fail to find parent "DFID
1501                                " for anonymous created %ld, try to"
1502                                " use system default.\n",
1503                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1504                         parent = NULL;
1505                 }
1506         }
1507
1508         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1509         if (!rc) {
1510                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1511                 if (flags & MDS_OPEN_LOCK)
1512                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1513                 if (flags & MDS_OPEN_LEASE)
1514                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1515         }
1516         GOTO(out_unlock, rc);
1517
1518 out_unlock:
1519         mdt_object_open_unlock(info, o, lhc, ibits, rc);
1520 out:
1521         mdt_object_put(env, o);
1522         if (parent != NULL)
1523                 mdt_object_put(env, parent);
1524         return rc;
1525 }
1526
1527 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1528 static int mdt_cross_open(struct mdt_thread_info *info,
1529                           const struct lu_fid *parent_fid,
1530                           const struct lu_fid *fid,
1531                           struct ldlm_reply *rep, __u32 flags)
1532 {
1533         struct md_attr    *ma = &info->mti_attr;
1534         struct mdt_object *o;
1535         int                rc;
1536         ENTRY;
1537
1538         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1539         if (IS_ERR(o))
1540                 RETURN(rc = PTR_ERR(o));
1541
1542         if (mdt_object_remote(o)) {
1543                 /* Something is wrong here, the object is on another MDS! */
1544                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1545                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1546                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1547                                 &o->mot_obj,
1548                                 "Object isn't on this server! FLD error?\n");
1549                 rc = -EFAULT;
1550         } else {
1551                 if (mdt_object_exists(o)) {
1552                         /* Do permission check for cross-open. */
1553                         rc = mo_permission(info->mti_env, NULL,
1554                                            mdt_object_child(o),
1555                                            NULL, flags | MDS_OPEN_CROSS);
1556                         if (rc)
1557                                 goto out;
1558
1559                         mdt_prep_ma_buf_from_rep(info, o, ma);
1560                         mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1561                         rc = mdt_attr_get_complex(info, o, ma);
1562                         if (rc != 0)
1563                                 GOTO(out, rc);
1564
1565                         /* Do not create lov object if the fid is opened
1566                          * under OBF */
1567                         if (S_ISREG(ma->ma_attr.la_mode) &&
1568                             !(ma->ma_valid & MA_LOV) && (flags & FMODE_WRITE) &&
1569                             fid_is_obf(parent_fid))
1570                                 GOTO(out, rc = -EPERM);
1571
1572                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1573                 } else {
1574                         /*
1575                          * Something is wrong here. lookup was positive but
1576                          * there is no object!
1577                          */
1578                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1579                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1580                         rc = -EFAULT;
1581                 }
1582         }
1583 out:
1584         mdt_object_put(info->mti_env, o);
1585         RETURN(rc);
1586 }
1587
1588 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1589 {
1590         struct mdt_device       *mdt = info->mti_mdt;
1591         struct ptlrpc_request   *req = mdt_info_req(info);
1592         struct mdt_object       *parent;
1593         struct mdt_object       *child;
1594         struct mdt_lock_handle  *lh;
1595         struct ldlm_reply       *ldlm_rep;
1596         struct mdt_body         *repbody;
1597         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1598         struct md_attr          *ma = &info->mti_attr;
1599         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1600         __u64                    ibits = 0;
1601         struct mdt_reint_record *rr = &info->mti_rr;
1602         int                      result, rc;
1603         int                      created = 0;
1604         __u32                    msg_flags;
1605         ENTRY;
1606
1607         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1608                                (obd_timeout + 1) / 4);
1609
1610         mdt_counter_incr(req, LPROC_MDT_OPEN);
1611         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1612
1613         ma->ma_need = MA_INODE;
1614         ma->ma_valid = 0;
1615
1616         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1617         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1618
1619         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1620                 CERROR("file join is not supported anymore.\n");
1621                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1622         }
1623         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1624
1625         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1626             info->mti_spec.u.sp_ea.eadata == NULL)
1627                 GOTO(out, result = err_serious(-EINVAL));
1628
1629         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1630                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1631                PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1632                PFID(rr->rr_fid2), create_flags,
1633                ma->ma_attr.la_mode, msg_flags);
1634
1635         if (info->mti_cross_ref) {
1636                 /* This is cross-ref open */
1637                 mdt_set_disposition(info, ldlm_rep,
1638                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1639                              DISP_LOOKUP_POS));
1640                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1641                                         ldlm_rep, create_flags);
1642                 GOTO(out, result);
1643         } else if (req_is_replay(req) ||
1644             (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1645                 /* This is a replay request or from liblustre with ea. */
1646                 result = mdt_open_by_fid(info, ldlm_rep);
1647
1648                 if (result != -ENOENT) {
1649                         if (req->rq_export->exp_libclient &&
1650                             create_flags & MDS_OPEN_HAS_EA)
1651                                 GOTO(out, result = 0);
1652                         GOTO(out, result);
1653                 }
1654                 /* We didn't find the correct object, so we need to re-create it
1655                  * via a regular replay. */
1656                 if (!(create_flags & MDS_OPEN_CREAT)) {
1657                         DEBUG_REQ(D_ERROR, req,
1658                                   "OPEN & CREAT not in open replay/by_fid.");
1659                         GOTO(out, result = -EFAULT);
1660                 }
1661                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1662         } else if ((!lu_name_is_valid(&rr->rr_name) &&
1663                     (create_flags & MDS_OPEN_LOCK)) ||
1664                    (create_flags & MDS_OPEN_BY_FID)) {
1665                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1666                 /* If result is 0 then open by FID has found the file
1667                  * and there is nothing left for us to do here.  More
1668                  * generally if it is anything other than -ENOENT or
1669                  * -EREMOTE then we return that now.  If -ENOENT and
1670                  * MDS_OPEN_CREAT is set then we must create the file
1671                  * below.  If -EREMOTE then we need to return a LOOKUP
1672                  * lock to the client, which we do below.  Hence this
1673                  * odd looking condition.  See LU-2523. */
1674                 if (!(result == -ENOENT && (create_flags & MDS_OPEN_CREAT)) &&
1675                     result != -EREMOTE)
1676                         GOTO(out, result);
1677
1678                 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1679         }
1680
1681         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1682                 GOTO(out, result = err_serious(-ENOMEM));
1683
1684         mdt_set_disposition(info, ldlm_rep,
1685                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1686
1687         if (!lu_name_is_valid(&rr->rr_name))
1688                 GOTO(out, result = -EPROTO);
1689
1690         lh = &info->mti_lh[MDT_LH_PARENT];
1691         mdt_lock_pdo_init(lh,
1692                           (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1693                           &rr->rr_name);
1694
1695         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1696                                       MDS_INODELOCK_UPDATE);
1697         if (IS_ERR(parent))
1698                 GOTO(out, result = PTR_ERR(parent));
1699
1700         /* get and check version of parent */
1701         result = mdt_version_get_check(info, parent, 0);
1702         if (result)
1703                 GOTO(out_parent, result);
1704
1705         fid_zero(child_fid);
1706
1707         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1708                             &rr->rr_name, child_fid, &info->mti_spec);
1709
1710         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1711                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1712                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1713                  PFID(child_fid));
1714
1715         if (result != 0 && result != -ENOENT && result != -ESTALE)
1716                 GOTO(out_parent, result);
1717
1718         if (result == -ENOENT || result == -ESTALE) {
1719                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1720                 if (result == -ESTALE) {
1721                         /*
1722                          * -ESTALE means the parent is a dead(unlinked) dir, so
1723                          * it should return -ENOENT to in accordance with the
1724                          * original mds implementaion.
1725                          */
1726                         GOTO(out_parent, result = -ENOENT);
1727                 }
1728                 if (!(create_flags & MDS_OPEN_CREAT))
1729                         GOTO(out_parent, result);
1730                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1731                         GOTO(out_parent, result = -EROFS);
1732                 *child_fid = *info->mti_rr.rr_fid2;
1733                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1734                          PFID(child_fid));
1735                 /* In the function below, .hs_keycmp resolves to
1736                  * lu_obj_hop_keycmp() */
1737                 /* coverity[overrun-buffer-val] */
1738                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1739         } else {
1740                 /*
1741                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1742                  * return FID back in that case.
1743                  */
1744                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1745                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1746         }
1747         if (IS_ERR(child))
1748                 GOTO(out_parent, result = PTR_ERR(child));
1749
1750         /** check version of child  */
1751         rc = mdt_version_get_check(info, child, 1);
1752         if (rc)
1753                 GOTO(out_child, result = rc);
1754
1755         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1756         if (result == -ENOENT) {
1757                 /* Create under OBF and .lustre is not permitted */
1758                 if (!fid_is_md_operative(rr->rr_fid1))
1759                         GOTO(out_child, result = -EPERM);
1760
1761                 /* save versions in reply */
1762                 mdt_version_get_save(info, parent, 0);
1763                 mdt_version_get_save(info, child, 1);
1764
1765                 /* version of child will be changed */
1766                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1767
1768                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1769                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1770
1771                 /* Let lower layers know what is lock mode on directory. */
1772                 info->mti_spec.sp_cr_mode =
1773                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1774
1775                 /*
1776                  * Do not perform lookup sanity check. We know that name does
1777                  * not exist.
1778                  */
1779                 info->mti_spec.sp_cr_lookup = 0;
1780                 info->mti_spec.sp_feat = &dt_directory_features;
1781
1782                 result = mdo_create(info->mti_env,
1783                                     mdt_object_child(parent),
1784                                     &rr->rr_name,
1785                                     mdt_object_child(child),
1786                                     &info->mti_spec,
1787                                     &info->mti_attr);
1788                 if (result == -ERESTART) {
1789                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1790                         GOTO(out_child, result);
1791                 } else {
1792                         mdt_prep_ma_buf_from_rep(info, child, ma);
1793                         /* XXX: we should call this once, see few lines below */
1794                         if (result == 0)
1795                                 result = mdt_attr_get_complex(info, child, ma);
1796
1797                         if (result != 0)
1798                                 GOTO(out_child, result);
1799                 }
1800                 created = 1;
1801         } else {
1802                 /*
1803                  * The object is on remote node, return its FID for remote open.
1804                  */
1805                 if (mdt_object_remote(child)) {
1806                         /*
1807                          * Check if this lock already was sent to client and
1808                          * this is resent case. For resent case do not take lock
1809                          * again, use what is already granted.
1810                          */
1811                         LASSERT(lhc != NULL);
1812
1813                         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1814                                 struct ldlm_lock *lock;
1815
1816                                 LASSERT(msg_flags & MSG_RESENT);
1817
1818                                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1819                                 if (!lock) {
1820                                         CERROR("Invalid lock handle "LPX64"\n",
1821                                                lhc->mlh_reg_lh.cookie);
1822                                         LBUG();
1823                                 }
1824                                 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1825                                                         &lock->l_resource->lr_name));
1826                                 LDLM_LOCK_PUT(lock);
1827                                 rc = 0;
1828                         } else {
1829                                 mdt_lock_handle_init(lhc);
1830                                 mdt_lock_reg_init(lhc, LCK_PR);
1831
1832                                 rc = mdt_object_lock(info, child, lhc,
1833                                                      MDS_INODELOCK_LOOKUP,
1834                                                      MDT_CROSS_LOCK);
1835                         }
1836                         repbody->fid1 = *mdt_object_fid(child);
1837                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1838                         if (rc != 0)
1839                                 result = rc;
1840                         else
1841                                 result = -EREMOTE;
1842                         GOTO(out_child, result);
1843                 } else if (mdt_object_exists(child)) {
1844                         /* We have to get attr & LOV EA & HSM for this
1845                          * object. */
1846                         mdt_prep_ma_buf_from_rep(info, child, ma);
1847                         ma->ma_need |= MA_HSM;
1848                         result = mdt_attr_get_complex(info, child, ma);
1849                 } else {
1850                         /* Object does not exist. Likely FS corruption. */
1851                         CERROR("%s: name '"DNAME"' present, but FID "
1852                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1853                                PNAME(&rr->rr_name), PFID(child_fid));
1854                         GOTO(out_child, result = -EIO);
1855                 }
1856         }
1857
1858         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1859                 /* the open lock might already be gotten in
1860                  * mdt_intent_fixup_resent */
1861                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1862                 if (create_flags & MDS_OPEN_LOCK)
1863                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1864         } else {
1865                 /* get openlock if this isn't replay and client requested it */
1866                 if (!req_is_replay(req)) {
1867                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1868                         if (rc != 0)
1869                                 GOTO(out_child_unlock, result = rc);
1870                         else if (create_flags & MDS_OPEN_LOCK)
1871                                 mdt_set_disposition(info, ldlm_rep,
1872                                                     DISP_OPEN_LOCK);
1873                 }
1874         }
1875         /* Try to open it now. */
1876         rc = mdt_finish_open(info, parent, child, create_flags,
1877                              created, ldlm_rep);
1878         if (rc) {
1879                 result = rc;
1880                 /* openlock will be released if mdt_finish_open failed */
1881                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1882
1883                 if (created && create_flags & MDS_OPEN_VOLATILE) {
1884                         CERROR("%s: cannot open volatile file "DFID", orphan "
1885                                "file will be left in PENDING directory until "
1886                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1887                                PFID(mdt_object_fid(child)), rc);
1888                         GOTO(out_child_unlock, result);
1889                 }
1890
1891                 if (created) {
1892                         ma->ma_need = 0;
1893                         ma->ma_valid = 0;
1894                         ma->ma_cookie_size = 0;
1895                         rc = mdo_unlink(info->mti_env,
1896                                         mdt_object_child(parent),
1897                                         mdt_object_child(child),
1898                                         &rr->rr_name,
1899                                         &info->mti_attr, 0);
1900                         if (rc != 0)
1901                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1902                                        mdt_obd_name(info->mti_mdt),
1903                                        PFID(mdt_object_fid(child)), rc);
1904                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1905                 }
1906         }
1907         EXIT;
1908 out_child_unlock:
1909         mdt_object_open_unlock(info, child, lhc, ibits, result);
1910 out_child:
1911         mdt_object_put(info->mti_env, child);
1912 out_parent:
1913         mdt_object_unlock_put(info, parent, lh, result || !created);
1914 out:
1915         if (result)
1916                 lustre_msg_set_transno(req->rq_repmsg, 0);
1917         return result;
1918 }
1919
1920 /**
1921  * Create an orphan object use local root.
1922  */
1923 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1924                                           struct mdt_device *mdt,
1925                                           const struct lu_fid *fid,
1926                                           struct md_attr *attr, fmode_t fmode)
1927 {
1928         const struct lu_env *env = info->mti_env;
1929         struct md_op_spec *spec = &info->mti_spec;
1930         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1931         struct mdt_object *obj = NULL;
1932         struct mdt_object *local_root;
1933         static const struct lu_name lname = {
1934                 .ln_name = "i_am_nobody",
1935                 .ln_namelen = sizeof("i_am_nobody") - 1,
1936         };
1937         struct lu_ucred *uc;
1938         cfs_cap_t uc_cap_save;
1939         int rc;
1940         ENTRY;
1941
1942         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1943         if (rc != 0)
1944                 RETURN(ERR_PTR(rc));
1945
1946         local_root = mdt_object_find(env, mdt, local_root_fid);
1947         if (IS_ERR(local_root))
1948                 RETURN(local_root);
1949
1950         obj = mdt_object_new(env, mdt, fid);
1951         if (IS_ERR(obj))
1952                 GOTO(out, rc = PTR_ERR(obj));
1953
1954         spec->sp_cr_lookup = 0;
1955         spec->sp_feat = &dt_directory_features;
1956         spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1957         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1958         if (attr->ma_valid & MA_LOV) {
1959                 spec->u.sp_ea.eadata = attr->ma_lmm;
1960                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1961                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1962         } else {
1963                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1964         }
1965
1966         uc = lu_ucred(env);
1967         uc_cap_save = uc->uc_cap;
1968         uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1969         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1970                         mdt_object_child(obj), spec, attr);
1971         uc->uc_cap = uc_cap_save;
1972         if (rc < 0) {
1973                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1974                        mdt_obd_name(mdt), PFID(fid), rc);
1975                 GOTO(out, rc);
1976         }
1977
1978         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1979         if (rc < 0)
1980                 CERROR("%s: cannot open volatile file "DFID", orphan "
1981                        "file will be left in PENDING directory until "
1982                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1983                        PFID(fid), rc);
1984         GOTO(out, rc);
1985
1986 out:
1987         if (rc < 0) {
1988                 if (!IS_ERR(obj))
1989                         mdt_object_put(env, obj);
1990                 obj = ERR_PTR(rc);
1991         }
1992         mdt_object_put(env, local_root);
1993         return obj;
1994 }
1995
1996 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1997                            struct md_attr *ma)
1998 {
1999         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
2000         struct close_data      *data;
2001         struct ldlm_lock       *lease;
2002         struct mdt_object      *orphan;
2003         struct md_attr         *orp_ma;
2004         struct lu_buf          *buf;
2005         bool                    lease_broken;
2006         int                     rc;
2007         int                     rc2;
2008         ENTRY;
2009
2010         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2011                 RETURN(-EROFS);
2012
2013         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2014         if (data == NULL)
2015                 RETURN(-EPROTO);
2016
2017         lease = ldlm_handle2lock(&data->cd_handle);
2018         if (lease == NULL)
2019                 RETURN(-ESTALE);
2020
2021         /* try to hold open_sem so that nobody else can open the file */
2022         if (!down_write_trylock(&o->mot_open_sem)) {
2023                 ldlm_lock_cancel(lease);
2024                 GOTO(out_reprocess, rc = -EBUSY);
2025         }
2026
2027         /* Check if the lease open lease has already canceled */
2028         lock_res_and_lock(lease);
2029         lease_broken = ldlm_is_cancel(lease);
2030         unlock_res_and_lock(lease);
2031
2032         LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2033                    PFID(mdt_object_fid(o)), lease_broken);
2034
2035         /* Cancel server side lease. Client side counterpart should
2036          * have been cancelled. It's okay to cancel it now as we've
2037          * held mot_open_sem. */
2038         ldlm_lock_cancel(lease);
2039
2040         if (lease_broken) /* don't perform release task */
2041                 GOTO(out_unlock, rc = -ESTALE);
2042
2043         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2044                 GOTO(out_unlock, rc = -EINVAL);
2045
2046         /* ma_need was set before but it seems fine to change it in order to
2047          * avoid modifying the one from RPC */
2048         ma->ma_need = MA_HSM;
2049         rc = mdt_attr_get_complex(info, o, ma);
2050         if (rc != 0)
2051                 GOTO(out_unlock, rc);
2052
2053         if (!mdt_hsm_release_allow(ma))
2054                 GOTO(out_unlock, rc = -EPERM);
2055
2056         /* already released? */
2057         if (ma->ma_hsm.mh_flags & HS_RELEASED)
2058                 GOTO(out_unlock, rc = 0);
2059
2060         /* Compare on-disk and packed data_version */
2061         if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
2062                 CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
2063                        " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
2064                        data->cd_data_version, ma->ma_hsm.mh_arch_ver);
2065                 GOTO(out_unlock, rc = -EPERM);
2066         }
2067
2068         ma->ma_valid = MA_INODE;
2069         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
2070         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
2071         if (rc < 0)
2072                 GOTO(out_unlock, rc);
2073
2074         ma->ma_need = MA_INODE | MA_LOV;
2075         rc = mdt_attr_get_complex(info, o, ma);
2076         if (rc < 0)
2077                 GOTO(out_unlock, rc);
2078
2079         if (!(ma->ma_valid & MA_LOV)) {
2080                 /* Even empty file are released */
2081                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2082                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2083                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2084                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2085                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2086         } else {
2087                 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
2088                  * ma_lmm as lov_user_md, then it will be confused by union of
2089                  * layout_gen and stripe_offset. */
2090                 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
2091                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2092                 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
2093                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
2094                 else
2095                         GOTO(out_unlock, rc = -EINVAL);
2096         }
2097
2098         /* Set file as released */
2099         ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
2100
2101         /* Hopefully it's not used in this call path */
2102         orp_ma = &info->mti_u.som.attr;
2103         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2104         orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
2105         orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
2106         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2107         orp_ma->ma_lmm = ma->ma_lmm;
2108         orp_ma->ma_lmm_size = ma->ma_lmm_size;
2109         orp_ma->ma_valid = MA_INODE | MA_LOV;
2110         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2111                                  FMODE_WRITE);
2112         if (IS_ERR(orphan)) {
2113                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2114                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2115                        PTR_ERR(orphan));
2116                 GOTO(out_unlock, rc = PTR_ERR(orphan));
2117         }
2118
2119         /* Set up HSM attribute for orphan object */
2120         CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
2121         buf = &info->mti_buf;
2122         buf->lb_buf = info->mti_xattr_buf;
2123         buf->lb_len = sizeof(struct hsm_attrs);
2124         ma->ma_hsm.mh_flags |= HS_RELEASED;
2125         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2126         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2127
2128         mdt_lock_reg_init(lh, LCK_EX);
2129         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2130                              MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
2131         if (rc != 0)
2132                 GOTO(out_close, rc);
2133
2134         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2135                           XATTR_NAME_HSM, 0);
2136
2137         if (rc == 0)
2138                 /* Swap layout with orphan object */
2139                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2140                                      mdt_object_child(orphan),
2141                                      SWAP_LAYOUTS_MDS_HSM);
2142
2143         /* Release exclusive LL */
2144         mdt_object_unlock(info, o, lh, 1);
2145
2146         EXIT;
2147
2148 out_close:
2149         /* Close orphan object anyway */
2150         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2151                        FMODE_WRITE);
2152         if (rc2 < 0)
2153                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2154                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2155         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2156                         "object closed\n");
2157         mdt_object_put(info->mti_env, orphan);
2158
2159 out_unlock:
2160         up_write(&o->mot_open_sem);
2161
2162         if (rc == 0) { /* already released */
2163                 struct mdt_body *repbody;
2164                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2165                 LASSERT(repbody != NULL);
2166                 repbody->valid |= OBD_MD_FLRELEASED;
2167         }
2168
2169 out_reprocess:
2170         ldlm_reprocess_all(lease->l_resource);
2171         LDLM_LOCK_PUT(lease);
2172
2173         ma->ma_valid = 0;
2174         ma->ma_need = 0;
2175
2176         return rc;
2177 }
2178
2179 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
2180                                       MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
2181
2182 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2183 {
2184         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
2185 }
2186
2187 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2188 {
2189         struct mdt_object *o = mfd->mfd_object;
2190         struct md_object *next = mdt_object_child(o);
2191         struct md_attr *ma = &info->mti_attr;
2192         int ret = MDT_IOEPOCH_CLOSED;
2193         int rc = 0;
2194         __u64 mode;
2195         ENTRY;
2196
2197         mode = mfd->mfd_mode;
2198
2199         if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
2200                 rc = mdt_hsm_release(info, o, ma);
2201                 if (rc < 0) {
2202                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2203                                 mdt_obd_name(info->mti_mdt),
2204                                 PFID(mdt_object_fid(o)), rc);
2205                         /* continue to close even error occurred. */
2206                 }
2207         }
2208
2209         if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
2210                 mdt_write_put(o);
2211                 ret = mdt_ioepoch_close(info, o);
2212         } else if (mode & MDS_FMODE_EXEC) {
2213                 mdt_write_allow(o);
2214         } else if (mode & MDS_FMODE_EPOCH) {
2215                 ret = mdt_ioepoch_close(info, o);
2216         } else if (mode & MDS_FMODE_SOM) {
2217                 ret = mdt_som_au_close(info, o);
2218         }
2219
2220         /* Update atime on close only. */
2221         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
2222             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
2223                 /* Set the atime only. */
2224                 ma->ma_valid = MA_INODE;
2225                 ma->ma_attr.la_valid = LA_ATIME;
2226                 rc = mo_attr_set(info->mti_env, next, ma);
2227         }
2228
2229         /* If file data is modified, add the dirty flag. */
2230         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2231                 rc = mdt_add_dirty_flag(info, o, ma);
2232
2233         ma->ma_need |= MA_INODE;
2234         ma->ma_valid &= ~MA_INODE;
2235
2236         if (!MFD_CLOSED(mode))
2237                 rc = mo_close(info->mti_env, next, ma, mode);
2238
2239         if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
2240                 struct mdt_export_data *med;
2241
2242                 /* The IOepoch is still opened or SOM update is needed.
2243                  * Put mfd back into the list. */
2244                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
2245                 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
2246                                       MDS_FMODE_EPOCH : MDS_FMODE_SOM);
2247
2248                 LASSERT(mdt_info_req(info));
2249                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
2250                 spin_lock(&med->med_open_lock);
2251                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
2252                 class_handle_hash_back(&mfd->mfd_handle);
2253                 spin_unlock(&med->med_open_lock);
2254
2255                 if (ret == MDT_IOEPOCH_OPENED) {
2256                         ret = 0;
2257                 } else {
2258                         ret = -EAGAIN;
2259                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
2260                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
2261                 }
2262         } else {
2263                 /* adjust open and lease count */
2264                 if (mode & MDS_OPEN_LEASE) {
2265                         LASSERT(atomic_read(&o->mot_lease_count) > 0);
2266                         atomic_dec(&o->mot_lease_count);
2267                 }
2268                 LASSERT(atomic_read(&o->mot_open_count) > 0);
2269                 atomic_dec(&o->mot_open_count);
2270
2271                 mdt_mfd_free(mfd);
2272                 mdt_object_put(info->mti_env, o);
2273         }
2274
2275         RETURN(rc ? rc : ret);
2276 }
2277
2278 int mdt_close(struct tgt_session_info *tsi)
2279 {
2280         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2281         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2282         struct mdt_export_data *med;
2283         struct mdt_file_data   *mfd;
2284         struct mdt_object      *o;
2285         struct md_attr         *ma = &info->mti_attr;
2286         struct mdt_body        *repbody = NULL;
2287         int rc, ret = 0;
2288         ENTRY;
2289
2290         mdt_counter_incr(req, LPROC_MDT_CLOSE);
2291         /* Close may come with the Size-on-MDS update. Unpack it. */
2292         rc = mdt_close_unpack(info);
2293         if (rc)
2294                 GOTO(out, rc = err_serious(rc));
2295
2296         LASSERT(info->mti_ioepoch);
2297
2298         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2299                              info->mti_mdt->mdt_max_mdsize);
2300         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2301                              info->mti_mdt->mdt_max_cookiesize);
2302         rc = req_capsule_server_pack(info->mti_pill);
2303         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2304                 mdt_client_compatibility(info);
2305                 if (rc == 0)
2306                         mdt_fix_reply(info);
2307                 mdt_exit_ucred(info);
2308                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2309         }
2310
2311         /* Continue to close handle even if we can not pack reply */
2312         if (rc == 0) {
2313                 repbody = req_capsule_server_get(info->mti_pill,
2314                                                  &RMF_MDT_BODY);
2315                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2316                                                     &RMF_MDT_MD);
2317                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2318                                                        &RMF_MDT_MD,
2319                                                        RCL_SERVER);
2320                 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
2321                                                        &RMF_LOGCOOKIES);
2322                 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
2323                                                           &RMF_LOGCOOKIES,
2324                                                           RCL_SERVER);
2325                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
2326                 repbody->eadatasize = 0;
2327                 repbody->aclsize = 0;
2328         } else {
2329                 rc = err_serious(rc);
2330         }
2331
2332         med = &req->rq_export->exp_mdt_data;
2333         spin_lock(&med->med_open_lock);
2334         mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2335                              req_is_replay(req));
2336         if (mdt_mfd_closed(mfd)) {
2337                 spin_unlock(&med->med_open_lock);
2338                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2339                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
2340                        info->mti_ioepoch->handle.cookie);
2341                 /** not serious error since bug 3633 */
2342                 rc = -ESTALE;
2343         } else {
2344                 class_handle_unhash(&mfd->mfd_handle);
2345                 cfs_list_del_init(&mfd->mfd_list);
2346                 spin_unlock(&med->med_open_lock);
2347
2348                 /* Do not lose object before last unlink. */
2349                 o = mfd->mfd_object;
2350                 mdt_object_get(info->mti_env, o);
2351                 ret = mdt_mfd_close(info, mfd);
2352                 if (repbody != NULL)
2353                         rc = mdt_handle_last_unlink(info, o, ma);
2354                 mdt_empty_transno(info, rc);
2355                 mdt_object_put(info->mti_env, o);
2356         }
2357         if (repbody != NULL) {
2358                 mdt_client_compatibility(info);
2359                 rc = mdt_fix_reply(info);
2360         }
2361
2362         mdt_exit_ucred(info);
2363         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2364                 GOTO(out, rc = err_serious(-ENOMEM));
2365
2366         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2367                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2368                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2369 out:
2370         mdt_thread_info_fini(info);
2371         RETURN(rc ? rc : ret);
2372 }
2373
2374 /**
2375  * DONE_WRITING rpc handler.
2376  *
2377  * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
2378  * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
2379  * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
2380  * skip attributes and reconstruct the reply here.
2381  */
2382 int mdt_done_writing(struct tgt_session_info *tsi)
2383 {
2384         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2385         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2386         struct mdt_body         *repbody = NULL;
2387         struct mdt_export_data  *med;
2388         struct mdt_file_data    *mfd;
2389         int rc;
2390         ENTRY;
2391
2392         rc = req_capsule_server_pack(tsi->tsi_pill);
2393         if (rc)
2394                 GOTO(out, rc = err_serious(rc));
2395
2396         repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2397         repbody->eadatasize = 0;
2398         repbody->aclsize = 0;
2399
2400         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
2401         rc = mdt_close_unpack(info);
2402         if (rc)
2403                 GOTO(out, rc = err_serious(rc));
2404
2405         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2406                 mdt_exit_ucred(info);
2407                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2408         }
2409
2410         med = &info->mti_exp->exp_mdt_data;
2411         spin_lock(&med->med_open_lock);
2412         mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2413                              req_is_replay(req));
2414         if (mfd == NULL) {
2415                 spin_unlock(&med->med_open_lock);
2416                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
2417                        ": cookie = "LPX64" ioepoch = "LPU64"\n",
2418                        PFID(info->mti_rr.rr_fid1),
2419                        info->mti_ioepoch->handle.cookie,
2420                        info->mti_ioepoch->ioepoch);
2421                 /* If this is a replay, reconstruct the transno. */
2422                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2423                         rc = info->mti_ioepoch->flags & MF_SOM_AU ?
2424                              -EAGAIN : 0;
2425                         mdt_empty_transno(info, rc);
2426                 } else
2427                         rc = -ESTALE;
2428                 GOTO(error_ucred, rc);
2429         }
2430
2431         LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
2432                 mfd->mfd_mode == MDS_FMODE_TRUNC);
2433         class_handle_unhash(&mfd->mfd_handle);
2434         cfs_list_del_init(&mfd->mfd_list);
2435         spin_unlock(&med->med_open_lock);
2436
2437         /* Set EPOCH CLOSE flag if not set by client. */
2438         info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
2439         info->mti_attr.ma_valid = 0;
2440
2441         info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
2442         OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2443         if (info->mti_attr.ma_lmm == NULL)
2444                 GOTO(error_ucred, rc = -ENOMEM);
2445
2446         rc = mdt_mfd_close(info, mfd);
2447
2448         OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2449         mdt_empty_transno(info, rc);
2450 error_ucred:
2451         mdt_exit_ucred(info);
2452 out:
2453         mdt_thread_info_fini(info);
2454         RETURN(rc);
2455 }