lustre/mdt/mdt_open.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/mdt/mdt_open.c
  37  *
  38  * Lustre Metadata Target (mdt) open/close file handling
  39  *
  40  * Author: Huang Hua <huanghua@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_MDS
  44
  45 #include <lustre_acl.h>
  46 #include <lustre_mds.h>
  47 #include "mdt_internal.h"
  48
  49 /* we do nothing because we do not have refcount now */
  50 static void mdt_mfd_get(void *mfdp)
  51 {
  52 }
  53
  54 static struct portals_handle_ops mfd_handle_ops = {
  55         .hop_addref = mdt_mfd_get,
  56         .hop_free   = NULL,
  57 };
  58
  59 /* Create a new mdt_file_data struct, initialize it,
  60  * and insert it to global hash table */
  61 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
  62 {
  63         struct mdt_file_data *mfd;
  64         ENTRY;
  65
  66         OBD_ALLOC_PTR(mfd);
  67         if (mfd != NULL) {
  68                 CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  69                 mfd->mfd_handle.h_owner = med;
  70                 CFS_INIT_LIST_HEAD(&mfd->mfd_list);
  71                 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
  72         }
  73
  74         RETURN(mfd);
  75 }
  76
  77 /*
  78  * Find the mfd pointed to by handle in global hash table.
  79  * In case of replay the handle is obsoleted
  80  * but mfd can be found in mfd list by that handle.
  81  * Callers need to be holding med_open_lock.
  82  */
  83 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
  84                                      const struct lustre_handle *handle,
  85                                      bool is_replay_or_resent)
  86 {
  87         struct mdt_file_data   *mfd;
  88         ENTRY;
  89
  90         LASSERT(handle != NULL);
  91         mfd = class_handle2object(handle->cookie, med);
  92         /* during dw/setattr replay the mfd can be found by old handle */
  93         if (mfd == NULL && is_replay_or_resent) {
  94                 cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  95                         if (mfd->mfd_old_handle.cookie == handle->cookie)
  96                                 RETURN(mfd);
  97                 }
  98                 mfd = NULL;
  99         }
 100
 101         RETURN(mfd);
 102 }
 103
 104 /* free mfd */
 105 void mdt_mfd_free(struct mdt_file_data *mfd)
 106 {
 107         LASSERT(cfs_list_empty(&mfd->mfd_list));
 108         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 109 }
 110
 111 static int mdt_create_data(struct mdt_thread_info *info,
 112                            struct mdt_object *p, struct mdt_object *o)
 113 {
 114         struct md_op_spec     *spec = &info->mti_spec;
 115         struct md_attr        *ma   = &info->mti_attr;
 116         int                    rc   = 0;
 117         ENTRY;
 118
 119         if (!md_should_create(spec->sp_cr_flags))
 120                 RETURN(0);
 121
 122         ma->ma_need = MA_INODE | MA_LOV;
 123         ma->ma_valid = 0;
 124         mutex_lock(&o->mot_lov_mutex);
 125         if (!(o->mot_flags & MOF_LOV_CREATED)) {
 126                 if (p != NULL && (fid_is_obf(mdt_object_fid(p)) ||
 127                                   fid_is_dot_lustre(mdt_object_fid(p))))
 128                         GOTO(unlock, rc = -EPERM);
 129
 130                 rc = mdo_create_data(info->mti_env,
 131                                      p ? mdt_object_child(p) : NULL,
 132                                      mdt_object_child(o), spec, ma);
 133                 if (rc == 0)
 134                         rc = mdt_attr_get_complex(info, o, ma);
 135
 136                 if (rc == 0 && ma->ma_valid & MA_LOV)
 137                         o->mot_flags |= MOF_LOV_CREATED;
 138         }
 139 unlock:
 140         mutex_unlock(&o->mot_lov_mutex);
 141         RETURN(rc);
 142 }
 143
 144 static int mdt_ioepoch_opened(struct mdt_object *mo)
 145 {
 146         return mo->mot_ioepoch_count;
 147 }
 148
 149 int mdt_object_is_som_enabled(struct mdt_object *mo)
 150 {
 151         return !mo->mot_ioepoch;
 152 }
 153
 154 /**
 155  * Re-enable Size-on-MDS.
 156  * Call under ->mot_ioepoch_mutex.
 157  */
 158 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
 159 {
 160         if (ioepoch == mo->mot_ioepoch) {
 161                 LASSERT(!mdt_ioepoch_opened(mo));
 162                 mo->mot_ioepoch = 0;
 163                 mo->mot_flags = 0;
 164         }
 165 }
 166
 167 /**
 168  * Open the IOEpoch. It is allowed if @writecount is not negative.
 169  * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
 170  */
 171 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
 172                      int created)
 173 {
 174         struct mdt_device *mdt = info->mti_mdt;
 175         int cancel = 0;
 176         int rc = 0;
 177         ENTRY;
 178
 179         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 180             !S_ISREG(lu_object_attr(&o->mot_obj)))
 181                 RETURN(0);
 182
 183         mutex_lock(&o->mot_ioepoch_mutex);
 184         if (mdt_ioepoch_opened(o)) {
 185                 /* Epoch continues even if there is no writers yet. */
 186                 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
 187                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 188         } else {
 189                 /* XXX: ->mdt_ioepoch is not initialized at the mount */
 190                 spin_lock(&mdt->mdt_ioepoch_lock);
 191                 if (mdt->mdt_ioepoch < info->mti_replayepoch)
 192                         mdt->mdt_ioepoch = info->mti_replayepoch;
 193
 194                 if (info->mti_replayepoch)
 195                         o->mot_ioepoch = info->mti_replayepoch;
 196                 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
 197                         o->mot_ioepoch = ++mdt->mdt_ioepoch;
 198                 else
 199                         o->mot_ioepoch = mdt->mdt_ioepoch;
 200
 201                 spin_unlock(&mdt->mdt_ioepoch_lock);
 202
 203                 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
 204                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 205                 if (created)
 206                         o->mot_flags |= MOF_SOM_CREATED;
 207                 cancel = 1;
 208         }
 209         o->mot_ioepoch_count++;
 210         mutex_unlock(&o->mot_ioepoch_mutex);
 211
 212         /* Cancel Size-on-MDS attributes cached on clients for the open case.
 213          * In the truncate case, see mdt_reint_setattr(). */
 214         if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
 215                 struct mdt_lock_handle  *lh = &info->mti_lh[MDT_LH_CHILD];
 216                 mdt_lock_reg_init(lh, LCK_EX);
 217                 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
 218                                      MDT_LOCAL_LOCK);
 219                 if (rc == 0)
 220                         mdt_object_unlock(info, o, lh, 1);
 221         }
 222         RETURN(rc);
 223 }
 224
 225 /**
 226  * Update SOM on-disk attributes.
 227  * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
 228  * mountid and attributes. If disabling, clean SOM xattr.
 229  * Call under ->mot_ioepoch_mutex.
 230  */
 231 static int mdt_som_attr_set(struct mdt_thread_info *info,
 232                             struct mdt_object *obj, __u64 ioepoch, bool enable)
 233 {
 234         struct md_object        *next = mdt_object_child(obj);
 235         int                      rc;
 236         ENTRY;
 237
 238         CDEBUG(D_INODE, "Size-on-MDS attribute %s for epoch "LPU64
 239                " on "DFID".\n", enable ? "update" : "disabling",
 240                ioepoch, PFID(mdt_object_fid(obj)));
 241
 242         if (enable) {
 243                 struct lu_buf           *buf = &info->mti_buf;
 244                 struct som_attrs        *attrs;
 245                 struct md_attr          *ma = &info->mti_attr;
 246                 struct lu_attr          *la = &ma->ma_attr;
 247                 struct obd_device       *obd = info->mti_mdt->mdt_lut.lut_obd;
 248
 249                 attrs = (struct som_attrs *)info->mti_xattr_buf;
 250                 CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*attrs));
 251
 252                 /* pack SOM attributes */
 253                 memset(attrs, 0, sizeof(*attrs));
 254                 attrs->som_ioepoch = ioepoch;
 255                 attrs->som_mountid = obd->u.obt.obt_mount_count;
 256                 if ((la->la_valid & LA_SIZE) != 0)
 257                         attrs->som_size = la->la_size;
 258                 if ((la->la_valid & LA_BLOCKS) != 0)
 259                         attrs->som_blocks = la->la_blocks;
 260                 lustre_som_swab(attrs);
 261
 262                 /* update SOM attributes */
 263                 buf->lb_buf = attrs;
 264                 buf->lb_len = sizeof(*attrs);
 265                 rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_SOM, 0);
 266         } else {
 267                 /* delete SOM attributes */
 268                 rc = mo_xattr_del(info->mti_env, next, XATTR_NAME_SOM);
 269         }
 270
 271         RETURN(rc);
 272 }
 273
 274 /** Perform the eviction specific actions on ioepoch close. */
 275 static inline int mdt_ioepoch_close_on_eviction(struct mdt_thread_info *info,
 276                                                 struct mdt_object *o)
 277 {
 278         int rc = 0;
 279
 280         mutex_lock(&o->mot_ioepoch_mutex);
 281         CDEBUG(D_INODE, "Eviction. Closing IOepoch "LPU64" on "DFID". "
 282                "Count %d\n", o->mot_ioepoch, PFID(mdt_object_fid(o)),
 283                o->mot_ioepoch_count);
 284         o->mot_ioepoch_count--;
 285
 286         /* If eviction occured set MOF_SOM_RECOV,
 287          * if no other epoch holders, disable SOM on disk. */
 288         o->mot_flags |= MOF_SOM_CHANGE | MOF_SOM_RECOV;
 289         if (!mdt_ioepoch_opened(o)) {
 290                 rc = mdt_som_attr_set(info, o, o->mot_ioepoch, MDT_SOM_DISABLE);
 291                 mdt_object_som_enable(o, o->mot_ioepoch);
 292         }
 293         mutex_unlock(&o->mot_ioepoch_mutex);
 294         RETURN(rc);
 295 }
 296
 297 /**
 298  * Perform the replay specific actions on ioepoch close.
 299  * Skip SOM attribute update if obtained and just forget about the inode state
 300  * for the last ioepoch holder. The SOM cache is invalidated on MDS failure.
 301  */
 302 static inline int mdt_ioepoch_close_on_replay(struct mdt_thread_info *info,
 303                                               struct mdt_object *o)
 304 {
 305         int rc = MDT_IOEPOCH_CLOSED;
 306         ENTRY;
 307
 308         mutex_lock(&o->mot_ioepoch_mutex);
 309         CDEBUG(D_INODE, "Replay. Closing epoch "LPU64" on "DFID". Count %d\n",
 310                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
 311         o->mot_ioepoch_count--;
 312
 313         /* Get an info from the replayed request if client is supposed
 314          * to send an Attibute Update, reconstruct @rc if so */
 315         if (info->mti_ioepoch->flags & MF_SOM_AU)
 316                 rc = MDT_IOEPOCH_GETATTR;
 317
 318         if (!mdt_ioepoch_opened(o))
 319                 mdt_object_som_enable(o, info->mti_ioepoch->ioepoch);
 320         mutex_unlock(&o->mot_ioepoch_mutex);
 321
 322         RETURN(rc);
 323 }
 324
 325 /**
 326  * Regular file IOepoch close.
 327  * Closes the ioepoch, checks the object state, apply obtained attributes and
 328  * re-enable SOM on the object, if possible. Also checks if the recovery is
 329  * needed and packs OBD_MD_FLGETATTRLOCK flag into the reply to force the client
 330  * to obtain SOM attributes under the server-side OST locks.
 331  *
 332  * Return value:
 333  * MDT_IOEPOCH_CLOSED if ioepoch is closed.
 334  * MDT_IOEPOCH_GETATTR if ioepoch is closed but another SOM update is needed.
 335  */
 336 static inline int mdt_ioepoch_close_reg(struct mdt_thread_info *info,
 337                                         struct mdt_object *o)
 338 {
 339         struct md_attr *tmp_ma;
 340         struct lu_attr *la;
 341         int achange, opened;
 342         int recovery = 0;
 343         int rc = 0, ret = MDT_IOEPOCH_CLOSED;
 344         ENTRY;
 345
 346         la = &info->mti_attr.ma_attr;
 347         achange = (info->mti_ioepoch->flags & MF_SOM_CHANGE);
 348
 349         mutex_lock(&o->mot_ioepoch_mutex);
 350         o->mot_ioepoch_count--;
 351
 352         tmp_ma = &info->mti_u.som.attr;
 353         tmp_ma->ma_lmm = info->mti_attr.ma_lmm;
 354         tmp_ma->ma_lmm_size = info->mti_attr.ma_lmm_size;
 355         tmp_ma->ma_som = &info->mti_u.som.data;
 356         tmp_ma->ma_need = MA_INODE | MA_LOV | MA_SOM;
 357         tmp_ma->ma_valid = 0;
 358         rc = mdt_attr_get_complex(info, o, tmp_ma);
 359         if (rc)
 360                 GOTO(error_up, rc);
 361
 362         /* Check the on-disk SOM state. */
 363         if (o->mot_flags & MOF_SOM_RECOV)
 364                 recovery = 1;
 365         else if (!(o->mot_flags & MOF_SOM_CREATED) &&
 366                  !(tmp_ma->ma_valid & MA_SOM))
 367                 recovery = 1;
 368
 369         CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
 370                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_ioepoch_count);
 371
 372         opened = mdt_ioepoch_opened(o);
 373         /**
 374          * If IOEpoch is not opened, check if a Size-on-MDS update is needed.
 375          * Skip the check for file with no LOV  or for unlink files.
 376          */
 377         if (!opened && tmp_ma->ma_valid & MA_LOV &&
 378             !(tmp_ma->ma_valid & MA_INODE && tmp_ma->ma_attr.la_nlink == 0)) {
 379                 if (recovery)
 380                         /* If some previous writer was evicted, re-ask the
 381                          * client for attributes. Even if attributes are
 382                          * provided, we cannot believe in them.
 383                          * Another use case is that there is no SOM cache on
 384                          * disk -- first access with SOM or there was an MDS
 385                          * failure. */
 386                         ret = MDT_IOEPOCH_GETATTR;
 387                 else if (o->mot_flags & MOF_SOM_CHANGE)
 388                         /* Some previous writer changed the attribute.
 389                          * Do not believe to the current Size-on-MDS
 390                          * update, re-ask client. */
 391                         ret = MDT_IOEPOCH_GETATTR;
 392                 else if (!(la->la_valid & LA_SIZE) && achange)
 393                         /* Attributes were changed by the last writer
 394                          * only but no Size-on-MDS update is received.*/
 395                         ret = MDT_IOEPOCH_GETATTR;
 396         }
 397
 398         if (achange || ret == MDT_IOEPOCH_GETATTR)
 399                 o->mot_flags |= MOF_SOM_CHANGE;
 400
 401         /* If epoch ends and relable SOM attributes are obtained, update them.
 402          * Create SOM ea for new files even if there is no attributes obtained
 403          * (0-length file). */
 404         if (ret == MDT_IOEPOCH_CLOSED && !opened) {
 405                 if (achange || o->mot_flags & MOF_SOM_CREATED) {
 406                         LASSERT(achange || !(la->la_valid & LA_SIZE));
 407                         rc = mdt_som_attr_set(info, o, o->mot_ioepoch,
 408                                               MDT_SOM_ENABLE);
 409                         /* Avoid the following setattrs of these attributes,
 410                          * e.g. for atime update. */
 411                         info->mti_attr.ma_valid = 0;
 412                 }
 413                 mdt_object_som_enable(o, o->mot_ioepoch);
 414         }
 415
 416         mutex_unlock(&o->mot_ioepoch_mutex);
 417         /* If recovery is needed, tell the client to perform GETATTR under
 418          * the lock. */
 419         if (ret == MDT_IOEPOCH_GETATTR && recovery) {
 420                 struct mdt_body *rep;
 421                 rep = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 422                 rep->valid |= OBD_MD_FLGETATTRLOCK;
 423         }
 424
 425         RETURN(rc ? : ret);
 426
 427 error_up:
 428         mutex_unlock(&o->mot_ioepoch_mutex);
 429         return rc;
 430 }
 431
 432 /**
 433  * Close IOEpoch (opened file or MDS_FMODE_EPOCH state). It happens if:
 434  * - a client closes the IOEpoch;
 435  * - a client eviction occured.
 436  * Return values:
 437  * MDT_IOEPOCH_OPENED if the client does not close IOEpoch.
 438  * MDT_IOEPOCH_CLOSED if the client closes IOEpoch.
 439  * MDT_IOEPOCH_GETATTR if the client closes IOEpoch but another SOM attribute
 440  * update is needed.
 441  */
 442 static int mdt_ioepoch_close(struct mdt_thread_info *info, struct mdt_object *o)
 443 {
 444         struct ptlrpc_request *req = mdt_info_req(info);
 445         ENTRY;
 446
 447         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 448             !S_ISREG(lu_object_attr(&o->mot_obj)))
 449                 RETURN(0);
 450
 451         LASSERT(o->mot_ioepoch_count);
 452         LASSERT(info->mti_ioepoch == NULL ||
 453                 info->mti_ioepoch->ioepoch == o->mot_ioepoch);
 454
 455         /* IOEpoch is closed only if client tells about it or eviction occures.
 456          * In the replay case, always close the epoch. */
 457         if (req == NULL)
 458                 RETURN(mdt_ioepoch_close_on_eviction(info, o));
 459         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
 460                 RETURN(mdt_ioepoch_close_on_replay(info, o));
 461         if (info->mti_ioepoch && (info->mti_ioepoch->flags & MF_EPOCH_CLOSE))
 462                 RETURN(mdt_ioepoch_close_reg(info, o));
 463         /* IO epoch is not closed. */
 464         RETURN(MDT_IOEPOCH_OPENED);
 465 }
 466
 467 /**
 468  * Close MDS_FMODE_SOM state, when IOEpoch is already closed and we are waiting
 469  * for attribute update. It happens if:
 470  * - SOM Attribute Update is obtained;
 471  * - the client failed to obtain it and informs MDS about it;
 472  * - a client eviction occured.
 473  * Apply obtained attributes for the 1st case, wipe out the on-disk SOM
 474  * cache otherwise.
 475  */
 476 int mdt_som_au_close(struct mdt_thread_info *info, struct mdt_object *o)
 477 {
 478         struct ptlrpc_request   *req = mdt_info_req(info);
 479         __u64                    ioepoch = 0;
 480         int                      act = MDT_SOM_ENABLE;
 481         int                      rc = 0;
 482         ENTRY;
 483
 484         LASSERT(!req || info->mti_ioepoch);
 485         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 486             !S_ISREG(lu_object_attr(&o->mot_obj)))
 487                 RETURN(0);
 488
 489         /* No size whereas MF_SOM_CHANGE is set means client failed to
 490          * obtain ost attributes, drop the SOM cache on disk if so. */
 491         if (!req ||
 492             (info->mti_ioepoch &&
 493              info->mti_ioepoch->flags & MF_SOM_CHANGE &&
 494              !(info->mti_attr.ma_attr.la_valid & LA_SIZE)))
 495                 act = MDT_SOM_DISABLE;
 496
 497         mutex_lock(&o->mot_ioepoch_mutex);
 498         /* Mark the object it is the recovery state if we failed to obtain
 499          * SOM attributes. */
 500         if (act == MDT_SOM_DISABLE)
 501                 o->mot_flags |= MOF_SOM_RECOV;
 502
 503         if (!mdt_ioepoch_opened(o)) {
 504                 ioepoch =  info->mti_ioepoch ?
 505                         info->mti_ioepoch->ioepoch : o->mot_ioepoch;
 506
 507                 if (req != NULL
 508                     && !(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY))
 509                         rc = mdt_som_attr_set(info, o, ioepoch, act);
 510                 mdt_object_som_enable(o, ioepoch);
 511         }
 512         mutex_unlock(&o->mot_ioepoch_mutex);
 513         RETURN(rc);
 514 }
 515
 516 int mdt_write_read(struct mdt_object *o)
 517 {
 518         int rc = 0;
 519         ENTRY;
 520         mutex_lock(&o->mot_ioepoch_mutex);
 521         rc = o->mot_writecount;
 522         mutex_unlock(&o->mot_ioepoch_mutex);
 523         RETURN(rc);
 524 }
 525
 526 int mdt_write_get(struct mdt_object *o)
 527 {
 528         int rc = 0;
 529         ENTRY;
 530         mutex_lock(&o->mot_ioepoch_mutex);
 531         if (o->mot_writecount < 0)
 532                 rc = -ETXTBSY;
 533         else
 534                 o->mot_writecount++;
 535         mutex_unlock(&o->mot_ioepoch_mutex);
 536         RETURN(rc);
 537 }
 538
 539 void mdt_write_put(struct mdt_object *o)
 540 {
 541         ENTRY;
 542         mutex_lock(&o->mot_ioepoch_mutex);
 543         o->mot_writecount--;
 544         mutex_unlock(&o->mot_ioepoch_mutex);
 545         EXIT;
 546 }
 547
 548 static int mdt_write_deny(struct mdt_object *o)
 549 {
 550         int rc = 0;
 551         ENTRY;
 552         mutex_lock(&o->mot_ioepoch_mutex);
 553         if (o->mot_writecount > 0)
 554                 rc = -ETXTBSY;
 555         else
 556                 o->mot_writecount--;
 557         mutex_unlock(&o->mot_ioepoch_mutex);
 558         RETURN(rc);
 559 }
 560
 561 static void mdt_write_allow(struct mdt_object *o)
 562 {
 563         ENTRY;
 564         mutex_lock(&o->mot_ioepoch_mutex);
 565         o->mot_writecount++;
 566         mutex_unlock(&o->mot_ioepoch_mutex);
 567         EXIT;
 568 }
 569
 570 /* there can be no real transaction so prepare the fake one */
 571 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 572 {
 573         struct mdt_device      *mdt = info->mti_mdt;
 574         struct ptlrpc_request  *req = mdt_info_req(info);
 575         struct tg_export_data  *ted;
 576         struct lsd_client_data *lcd;
 577
 578         ENTRY;
 579         /* transaction has occurred already */
 580         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 581                 RETURN_EXIT;
 582
 583         spin_lock(&mdt->mdt_lut.lut_translock);
 584         if (rc != 0) {
 585                 if (info->mti_transno != 0) {
 586                         struct obd_export *exp = req->rq_export;
 587
 588                         CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
 589                                mdt_obd_name(mdt), info->mti_transno,
 590                                libcfs_nid2str(exp->exp_connection->c_peer.nid),
 591                                rc);
 592                         spin_unlock(&mdt->mdt_lut.lut_translock);
 593                         RETURN_EXIT;
 594                 }
 595         } else if (info->mti_transno == 0) {
 596                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
 597         } else {
 598                 /* should be replay */
 599                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 600                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 601         }
 602         spin_unlock(&mdt->mdt_lut.lut_translock);
 603
 604         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 605                info->mti_transno,
 606                req->rq_export->exp_obd->obd_last_committed);
 607
 608         req->rq_transno = info->mti_transno;
 609         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 610
 611         /* update lcd in memory only for resent cases */
 612         ted = &req->rq_export->exp_target_data;
 613         LASSERT(ted);
 614         mutex_lock(&ted->ted_lcd_lock);
 615         lcd = ted->ted_lcd;
 616         if (info->mti_transno < lcd->lcd_last_transno &&
 617             info->mti_transno != 0) {
 618                 /* This should happen during replay. Do not update
 619                  * last rcvd info if replay req transno < last transno,
 620                  * otherwise the following resend(after replay) can not
 621                  * be checked correctly by xid */
 622                 mutex_unlock(&ted->ted_lcd_lock);
 623                 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
 624                        mdt_obd_name(mdt), info->mti_transno,
 625                        lcd->lcd_last_transno);
 626                 RETURN_EXIT;
 627         }
 628
 629         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 630             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 631                 if (info->mti_transno != 0)
 632                         lcd->lcd_last_close_transno = info->mti_transno;
 633                 lcd->lcd_last_close_xid = req->rq_xid;
 634                 lcd->lcd_last_close_result = rc;
 635         } else {
 636                 /* VBR: save versions in last_rcvd for reconstruct. */
 637                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 638                 if (pre_versions) {
 639                         lcd->lcd_pre_versions[0] = pre_versions[0];
 640                         lcd->lcd_pre_versions[1] = pre_versions[1];
 641                         lcd->lcd_pre_versions[2] = pre_versions[2];
 642                         lcd->lcd_pre_versions[3] = pre_versions[3];
 643                 }
 644                 if (info->mti_transno != 0)
 645                         lcd->lcd_last_transno = info->mti_transno;
 646
 647                 lcd->lcd_last_xid = req->rq_xid;
 648                 lcd->lcd_last_result = rc;
 649                 lcd->lcd_last_data = info->mti_opdata;
 650         }
 651         mutex_unlock(&ted->ted_lcd_lock);
 652
 653         EXIT;
 654 }
 655
 656 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
 657 {
 658         LASSERT(mfd != NULL);
 659
 660         CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n",
 661                PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
 662
 663         mfd->mfd_mode = mode;
 664 }
 665
 666 /**
 667  * prep ma_lmm/ma_lmv for md_attr from reply
 668  */
 669 void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
 670                               struct mdt_object *obj,
 671                               struct md_attr *ma)
 672 {
 673         LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
 674         if (S_ISDIR(obj->mot_header.loh_attr)) {
 675                 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
 676                                                     &RMF_MDT_MD);
 677                 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
 678                                                        &RMF_MDT_MD,
 679                                                        RCL_SERVER);
 680                 if (ma->ma_lmv_size > 0)
 681                         ma->ma_need |= MA_LMV;
 682         } else {
 683                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
 684                                                     &RMF_MDT_MD);
 685                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 686                                                        &RMF_MDT_MD,
 687                                                        RCL_SERVER);
 688                 if (ma->ma_lmm_size > 0)
 689                         ma->ma_need |= MA_LOV;
 690         }
 691 }
 692
 693 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 694                         struct mdt_object *o, __u64 flags, int created,
 695                         struct ldlm_reply *rep)
 696 {
 697         struct ptlrpc_request   *req = mdt_info_req(info);
 698         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 699         struct mdt_file_data    *mfd;
 700         struct md_attr          *ma  = &info->mti_attr;
 701         struct lu_attr          *la  = &ma->ma_attr;
 702         struct mdt_body         *repbody;
 703         int                      rc = 0, isdir, isreg;
 704         ENTRY;
 705
 706         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 707
 708         isreg = S_ISREG(la->la_mode);
 709         isdir = S_ISDIR(la->la_mode);
 710         if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
 711                 /*
 712                  * No EA, check whether it is will set regEA and dirEA since in
 713                  * above attr get, these size might be zero, so reset it, to
 714                  * retrieve the MD after create obj.
 715                  */
 716                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 717                                                        &RMF_MDT_MD,
 718                                                        RCL_SERVER);
 719                 /* in replay case, p == NULL */
 720                 rc = mdt_create_data(info, p, o);
 721                 if (rc)
 722                         RETURN(rc);
 723
 724                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
 725                         mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
 726         }
 727
 728         CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
 729                ma->ma_valid, ma->ma_lmm_size);
 730
 731         if (ma->ma_valid & MA_LOV) {
 732                 LASSERT(ma->ma_lmm_size != 0);
 733                 repbody->eadatasize = ma->ma_lmm_size;
 734                 if (isdir)
 735                         repbody->valid |= OBD_MD_FLDIREA;
 736                 else
 737                         repbody->valid |= OBD_MD_FLEASIZE;
 738         }
 739
 740         if (ma->ma_valid & MA_LMV) {
 741                 LASSERT(ma->ma_lmv_size != 0);
 742                 repbody->eadatasize = ma->ma_lmv_size;
 743                 LASSERT(isdir);
 744                 repbody->valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
 745         }
 746
 747         if (flags & FMODE_WRITE) {
 748                 rc = mdt_write_get(o);
 749                 if (rc == 0) {
 750                         mdt_ioepoch_open(info, o, created);
 751                         repbody->ioepoch = o->mot_ioepoch;
 752                 }
 753         } else if (flags & MDS_FMODE_EXEC) {
 754                 rc = mdt_write_deny(o);
 755         }
 756         if (rc)
 757                 RETURN(rc);
 758
 759         rc = mo_open(info->mti_env, mdt_object_child(o),
 760                      created ? flags | MDS_OPEN_CREATED : flags);
 761         if (rc)
 762                 GOTO(err_out, rc);
 763
 764         mfd = mdt_mfd_new(med);
 765         if (mfd == NULL)
 766                 GOTO(err_out, rc = -ENOMEM);
 767
 768         /*
 769          * Keep a reference on this object for this open, and is
 770          * released by mdt_mfd_close().
 771          */
 772         mdt_object_get(info->mti_env, o);
 773         mfd->mfd_object = o;
 774         mfd->mfd_xid = req->rq_xid;
 775
 776         /*
 777          * @flags is always not zero. At least it should be FMODE_READ,
 778          * FMODE_WRITE or MDS_FMODE_EXEC.
 779          */
 780         LASSERT(flags != 0);
 781
 782         /* Open handling. */
 783         mdt_mfd_set_mode(mfd, flags);
 784
 785         atomic_inc(&o->mot_open_count);
 786         if (flags & MDS_OPEN_LEASE)
 787                 atomic_inc(&o->mot_lease_count);
 788
 789         /* replay handle */
 790         if (req_is_replay(req)) {
 791                 struct mdt_file_data *old_mfd;
 792                 /* Check wheather old cookie already exist in
 793                  * the list, becasue when do recovery, client
 794                  * might be disconnected from server, and
 795                  * restart replay, so there maybe some orphan
 796                  * mfd here, we should remove them */
 797                 LASSERT(info->mti_rr.rr_handle != NULL);
 798                 spin_lock(&med->med_open_lock);
 799                 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
 800                 if (old_mfd != NULL) {
 801                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 802                                "cookie = "LPX64"\n", mfd,
 803                                PFID(mdt_object_fid(mfd->mfd_object)),
 804                                info->mti_rr.rr_handle->cookie);
 805                         class_handle_unhash(&old_mfd->mfd_handle);
 806                         cfs_list_del_init(&old_mfd->mfd_list);
 807                         spin_unlock(&med->med_open_lock);
 808                         /* no attr update for that close */
 809                         la->la_valid = 0;
 810                         ma->ma_valid |= MA_FLAGS;
 811                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 812                         mdt_mfd_close(info, old_mfd);
 813                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 814                         ma->ma_valid &= ~MA_FLAGS;
 815                 } else {
 816                         spin_unlock(&med->med_open_lock);
 817                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 818                                "cookie = "LPX64"\n",
 819                                PFID(mdt_object_fid(mfd->mfd_object)),
 820                                info->mti_rr.rr_handle->cookie);
 821                 }
 822
 823                 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
 824                        info->mti_rr.rr_handle->cookie);
 825
 826                 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
 827         }
 828
 829         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 830
 831         if (req->rq_export->exp_disconnected) {
 832                 spin_lock(&med->med_open_lock);
 833                 class_handle_unhash(&mfd->mfd_handle);
 834                 cfs_list_del_init(&mfd->mfd_list);
 835                 spin_unlock(&med->med_open_lock);
 836                 mdt_mfd_close(info, mfd);
 837         } else {
 838                 spin_lock(&med->med_open_lock);
 839                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
 840                 spin_unlock(&med->med_open_lock);
 841         }
 842
 843         mdt_empty_transno(info, rc);
 844
 845         RETURN(rc);
 846
 847 err_out:
 848         if (flags & FMODE_WRITE)
 849                 /* XXX We also need to close io epoch here.
 850                  * See LU-1220 - green */
 851                 mdt_write_put(o);
 852         else if (flags & MDS_FMODE_EXEC)
 853                 mdt_write_allow(o);
 854
 855         return rc;
 856 }
 857
 858 int mdt_finish_open(struct mdt_thread_info *info,
 859                     struct mdt_object *p, struct mdt_object *o,
 860                     __u64 flags, int created, struct ldlm_reply *rep)
 861 {
 862         struct ptlrpc_request   *req = mdt_info_req(info);
 863         struct obd_export       *exp = req->rq_export;
 864         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 865         struct md_attr          *ma  = &info->mti_attr;
 866         struct lu_attr          *la  = &ma->ma_attr;
 867         struct mdt_file_data    *mfd;
 868         struct mdt_body         *repbody;
 869         int                      rc = 0;
 870         int                      isreg, isdir, islnk;
 871         cfs_list_t              *t;
 872         ENTRY;
 873
 874         LASSERT(ma->ma_valid & MA_INODE);
 875
 876         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 877
 878         isreg = S_ISREG(la->la_mode);
 879         isdir = S_ISDIR(la->la_mode);
 880         islnk = S_ISLNK(la->la_mode);
 881         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 882
 883         /* LU-2275, simulate broken behaviour (esp. prevalent in
 884          * pre-2.4 servers where a very strange reply is sent on error
 885          * that looks like it was actually almost succesful and a failure at the
 886          * same time */
 887         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 888                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 889                                                DISP_LOOKUP_NEG |
 890                                                DISP_LOOKUP_POS);
 891
 892                 if (flags & MDS_OPEN_LOCK)
 893                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 894
 895                 RETURN(-ENOENT);
 896         }
 897
 898         if (exp_connect_rmtclient(exp)) {
 899                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 900
 901                 rc = mdt_pack_remote_perm(info, o, buf);
 902                 if (rc) {
 903                         repbody->valid &= ~OBD_MD_FLRMTPERM;
 904                         repbody->aclsize = 0;
 905                 } else {
 906                         repbody->valid |= OBD_MD_FLRMTPERM;
 907                         repbody->aclsize = sizeof(struct mdt_remote_perm);
 908                 }
 909         }
 910 #ifdef CONFIG_FS_POSIX_ACL
 911         else if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 912                 const struct lu_env *env = info->mti_env;
 913                 struct md_object *next = mdt_object_child(o);
 914                 struct lu_buf *buf = &info->mti_buf;
 915
 916                 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 917                 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
 918                                                    RCL_SERVER);
 919                 if (buf->lb_len > 0) {
 920                         rc = mo_xattr_get(env, next, buf,
 921                                           XATTR_NAME_ACL_ACCESS);
 922                         if (rc < 0) {
 923                                 if (rc == -ENODATA) {
 924                                         repbody->aclsize = 0;
 925                                         repbody->valid |= OBD_MD_FLACL;
 926                                         rc = 0;
 927                                 } else if (rc == -EOPNOTSUPP) {
 928                                         rc = 0;
 929                                 } else {
 930                                         CERROR("got acl size: %d\n", rc);
 931                                 }
 932                         } else {
 933                                 repbody->aclsize = rc;
 934                                 repbody->valid |= OBD_MD_FLACL;
 935                                 rc = 0;
 936                         }
 937                 }
 938         }
 939 #endif
 940
 941         if (info->mti_mdt->mdt_lut.lut_mds_capa &&
 942             exp_connect_flags(exp) & OBD_CONNECT_MDS_CAPA) {
 943                 struct lustre_capa *capa;
 944
 945                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
 946                 LASSERT(capa);
 947                 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
 948                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 949                 if (rc)
 950                         RETURN(rc);
 951                 repbody->valid |= OBD_MD_FLMDSCAPA;
 952         }
 953         if (info->mti_mdt->mdt_lut.lut_oss_capa &&
 954             exp_connect_flags(exp) & OBD_CONNECT_OSS_CAPA &&
 955             S_ISREG(lu_object_attr(&o->mot_obj))) {
 956                 struct lustre_capa *capa;
 957
 958                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
 959                 LASSERT(capa);
 960                 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
 961                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 962                 if (rc)
 963                         RETURN(rc);
 964                 repbody->valid |= OBD_MD_FLOSSCAPA;
 965         }
 966
 967         /*
 968          * If we are following a symlink, don't open; and do not return open
 969          * handle for special nodes as client required.
 970          */
 971         if (islnk || (!isreg && !isdir &&
 972             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 973                 lustre_msg_set_transno(req->rq_repmsg, 0);
 974                 RETURN(0);
 975         }
 976
 977         /*
 978          * We need to return the existing object's fid back, so it is done here,
 979          * after preparing the reply.
 980          */
 981         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 982                 RETURN(-EEXIST);
 983
 984         /* This can't be done earlier, we need to return reply body */
 985         if (isdir) {
 986                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 987                         /* We are trying to create or write an existing dir. */
 988                         RETURN(-EISDIR);
 989                 }
 990         } else if (flags & MDS_OPEN_DIRECTORY)
 991                 RETURN(-ENOTDIR);
 992
 993         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 994                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 995                 RETURN(-EAGAIN);
 996         }
 997
 998         mfd = NULL;
 999         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1000                 spin_lock(&med->med_open_lock);
1001                 cfs_list_for_each(t, &med->med_open_head) {
1002                         mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
1003                         if (mfd->mfd_xid == req->rq_xid)
1004                                 break;
1005                         mfd = NULL;
1006                 }
1007                 spin_unlock(&med->med_open_lock);
1008
1009                 if (mfd != NULL) {
1010                         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
1011                         /*set repbody->ea_size for resent case*/
1012                         if (ma->ma_valid & MA_LOV) {
1013                                 LASSERT(ma->ma_lmm_size != 0);
1014                                 repbody->eadatasize = ma->ma_lmm_size;
1015                                 if (isdir)
1016                                         repbody->valid |= OBD_MD_FLDIREA;
1017                                 else
1018                                         repbody->valid |= OBD_MD_FLEASIZE;
1019                         }
1020                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
1021                         RETURN(0);
1022                 }
1023         }
1024
1025         rc = mdt_mfd_open(info, p, o, flags, created, rep);
1026         if (!rc)
1027                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
1028
1029         RETURN(rc);
1030 }
1031
1032 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
1033                              struct lsd_client_data *lcd);
1034
1035 void mdt_reconstruct_open(struct mdt_thread_info *info,
1036                           struct mdt_lock_handle *lhc)
1037 {
1038         const struct lu_env *env = info->mti_env;
1039         struct mdt_device       *mdt  = info->mti_mdt;
1040         struct req_capsule      *pill = info->mti_pill;
1041         struct ptlrpc_request   *req  = mdt_info_req(info);
1042         struct tg_export_data   *ted  = &req->rq_export->exp_target_data;
1043         struct lsd_client_data  *lcd  = ted->ted_lcd;
1044         struct md_attr          *ma   = &info->mti_attr;
1045         struct mdt_reint_record *rr   = &info->mti_rr;
1046         __u64                   flags = info->mti_spec.sp_cr_flags;
1047         struct ldlm_reply       *ldlm_rep;
1048         struct mdt_object       *parent;
1049         struct mdt_object       *child;
1050         struct mdt_body         *repbody;
1051         int                      rc;
1052         ENTRY;
1053
1054         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1055         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
1056         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
1057
1058         ma->ma_need = MA_INODE | MA_HSM;
1059         ma->ma_valid = 0;
1060
1061         mdt_req_from_lcd(req, lcd);
1062         mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1063
1064         CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1065                ldlm_rep->lock_policy_res1, req->rq_status);
1066
1067         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1068             req->rq_status != 0)
1069                 /* We did not create successfully, return error to client. */
1070                 GOTO(out, rc = req->rq_status);
1071
1072         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1073                 struct obd_export *exp = req->rq_export;
1074                 /*
1075                  * We failed after creation, but we do not know in which step
1076                  * we failed. So try to check the child object.
1077                  */
1078                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1079                 if (IS_ERR(parent)) {
1080                         rc = PTR_ERR(parent);
1081                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
1082                                       " Evicting client %s with export %s.\n",
1083                                       PFID(rr->rr_fid1), rc,
1084                                       obd_uuid2str(&exp->exp_client_uuid),
1085                                       obd_export_nid2str(exp));
1086                         mdt_export_evict(exp);
1087                         RETURN_EXIT;
1088                 }
1089                 child = mdt_object_find(env, mdt, rr->rr_fid2);
1090                 if (IS_ERR(child)) {
1091                         rc = PTR_ERR(child);
1092                         LCONSOLE_WARN("Child "DFID" lookup error %d."
1093                                       " Evicting client %s with export %s.\n",
1094                                       PFID(mdt_object_fid(child)), rc,
1095                                       obd_uuid2str(&exp->exp_client_uuid),
1096                                       obd_export_nid2str(exp));
1097                         mdt_object_put(env, parent);
1098                         mdt_export_evict(exp);
1099                         RETURN_EXIT;
1100                 }
1101
1102                 if (unlikely(mdt_object_remote(child))) {
1103                         /* the child object was created on remote server */
1104                         if (!mdt_is_dne_client(exp)) {
1105                                 /* Return -EIO for old client */
1106                                 mdt_object_put(env, parent);
1107                                 mdt_object_put(env, child);
1108                                 GOTO(out, rc = -EIO);
1109                         }
1110                         repbody->fid1 = *rr->rr_fid2;
1111                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1112                         rc = 0;
1113                 } else {
1114                         if (mdt_object_exists(child)) {
1115                                 mdt_set_capainfo(info, 1, rr->rr_fid2,
1116                                                  BYPASS_CAPA);
1117                                 mdt_prep_ma_buf_from_rep(info, child, ma);
1118                                 rc = mdt_attr_get_complex(info, child, ma);
1119                                 if (rc == 0)
1120                                         rc = mdt_finish_open(info, parent,
1121                                                              child, flags,
1122                                                              1, ldlm_rep);
1123                         } else {
1124                                 /* the child does not exist, we should do
1125                                  * regular open */
1126                                 mdt_object_put(env, parent);
1127                                 mdt_object_put(env, child);
1128                                 GOTO(regular_open, 0);
1129                         }
1130                 }
1131                 mdt_object_put(env, parent);
1132                 mdt_object_put(env, child);
1133                 GOTO(out, rc);
1134         } else {
1135 regular_open:
1136                 /* We did not try to create, so we are a pure open */
1137                 rc = mdt_reint_open(info, lhc);
1138         }
1139
1140         EXIT;
1141 out:
1142         req->rq_status = rc;
1143         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1144         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1145 }
1146
1147 int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
1148 {
1149         __u64                    flags = info->mti_spec.sp_cr_flags;
1150         struct mdt_reint_record *rr = &info->mti_rr;
1151         struct md_attr          *ma = &info->mti_attr;
1152         struct mdt_object       *o;
1153         int                      rc;
1154         ENTRY;
1155
1156         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1157         if (IS_ERR(o))
1158                 RETURN(rc = PTR_ERR(o));
1159
1160         if (unlikely(mdt_object_remote(o))) {
1161                 /* the child object was created on remote server */
1162                 struct mdt_body *repbody;
1163
1164                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1165                                                 DISP_LOOKUP_EXECD |
1166                                                 DISP_LOOKUP_POS));
1167                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1168                 repbody->fid1 = *rr->rr_fid2;
1169                 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1170                 rc = 0;
1171         } else {
1172                 if (mdt_object_exists(o)) {
1173                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1174                                                         DISP_LOOKUP_EXECD |
1175                                                         DISP_LOOKUP_POS));
1176                         mdt_prep_ma_buf_from_rep(info, o, ma);
1177                         rc = mdt_attr_get_complex(info, o, ma);
1178                         if (rc == 0)
1179                                 rc = mdt_finish_open(info, NULL, o, flags, 0,
1180                                                      rep);
1181                 } else {
1182                         rc = -ENOENT;
1183                 }
1184         }
1185
1186         mdt_object_put(info->mti_env, o);
1187         RETURN(rc);
1188 }
1189
1190 /* lock object for open */
1191 static int mdt_object_open_lock(struct mdt_thread_info *info,
1192                                 struct mdt_object *obj,
1193                                 struct mdt_lock_handle *lhc,
1194                                 __u64 *ibits)
1195 {
1196         struct md_attr  *ma = &info->mti_attr;
1197         __u64            open_flags = info->mti_spec.sp_cr_flags;
1198         ldlm_mode_t      lm = LCK_CR;
1199         bool             acq_lease = !!(open_flags & MDS_OPEN_LEASE);
1200         bool             try_layout = false;
1201         bool             create_layout = false;
1202         int              rc = 0;
1203         ENTRY;
1204
1205         *ibits = 0;
1206         mdt_lock_handle_init(lhc);
1207
1208         if (req_is_replay(mdt_info_req(info)))
1209                 RETURN(0);
1210
1211         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
1212                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
1213                     md_should_create(open_flags))
1214                         create_layout = true;
1215                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
1216                     ma->ma_need & MA_LOV)
1217                         try_layout = true;
1218         }
1219
1220         if (acq_lease) {
1221                 /* lease open, acquire write mode of open sem */
1222                 down_write(&obj->mot_open_sem);
1223
1224                 /* Lease exists and ask for new lease */
1225                 if (atomic_read(&obj->mot_lease_count) > 0) {
1226                         /* only exclusive open is supported, so lease
1227                          * are conflicted to each other */
1228                         GOTO(out, rc = -EBUSY);
1229                 }
1230
1231                 /* Lease must be with open lock */
1232                 if (!(open_flags & MDS_OPEN_LOCK)) {
1233                         CERROR("Request lease for file:"DFID ", but open lock "
1234                                 "is missed, open_flags = "LPO64".\n",
1235                                 PFID(mdt_object_fid(obj)), open_flags);
1236                         GOTO(out, rc = -EPROTO);
1237                 }
1238
1239                 /* XXX: only exclusive open is supported. */
1240                 lm = LCK_EX;
1241                 *ibits = MDS_INODELOCK_OPEN;
1242
1243                 /* never grant LCK_EX layout lock to client */
1244                 try_layout = false;
1245         } else { /* normal open */
1246                 /* normal open holds read mode of open sem */
1247                 down_read(&obj->mot_open_sem);
1248
1249                 if (open_flags & MDS_OPEN_LOCK) {
1250                         if (open_flags & FMODE_WRITE)
1251                                 lm = LCK_CW;
1252                         else if (open_flags & MDS_FMODE_EXEC)
1253                                 lm = LCK_PR;
1254                         else
1255                                 lm = LCK_CR;
1256
1257                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
1258                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
1259                         if (open_flags & FMODE_WRITE)
1260                                 lm = LCK_CW;
1261                         else
1262                                 lm = LCK_CR;
1263
1264                         /* revoke lease */
1265                         *ibits = MDS_INODELOCK_OPEN;
1266                         try_layout = false;
1267
1268                         lhc = &info->mti_lh[MDT_LH_LOCAL];
1269                 }
1270                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
1271                         PFID(mdt_object_fid(obj)),
1272                         atomic_read(&obj->mot_open_count), lm);
1273         }
1274
1275         mdt_lock_reg_init(lhc, lm);
1276
1277         /* one problem to return layout lock on open is that it may result
1278          * in too many layout locks cached on the client side. */
1279         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
1280                 /* return lookup lock to validate inode at the client side,
1281                  * this is pretty important otherwise mdt will return layout
1282                  * lock for each open.
1283                  * However this is a double-edged sword because changing
1284                  * permission will revoke huge # of LOOKUP locks. */
1285                 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
1286                 if (!mdt_object_lock_try(info, obj, lhc, *ibits,
1287                                          MDT_CROSS_LOCK)) {
1288                         *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
1289                         if (*ibits != 0)
1290                                 rc = mdt_object_lock(info, obj, lhc, *ibits,
1291                                                 MDT_CROSS_LOCK);
1292                 }
1293         } else if (*ibits != 0) {
1294                 rc = mdt_object_lock(info, obj, lhc, *ibits, MDT_CROSS_LOCK);
1295         }
1296
1297         CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
1298                 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
1299                 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
1300
1301         /* will change layout, revoke layout locks by enqueuing EX lock. */
1302         if (rc == 0 && create_layout) {
1303                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
1304
1305                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
1306                         ", open_flags = "LPO64"\n",
1307                         PFID(mdt_object_fid(obj)), open_flags);
1308
1309                 /* We cannot enqueue another lock for the same resource we
1310                  * already have a lock for, due to mechanics of waiting list
1311                  * iterating in ldlm, see LU-3601.
1312                  * As such we'll drop the open lock we just got above here,
1313                  * it's ok not to have this open lock as it's main purpose is to
1314                  * flush unused cached client open handles. */
1315                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1316                         mdt_object_unlock(info, obj, lhc, 1);
1317
1318                 LASSERT(!try_layout);
1319                 mdt_lock_handle_init(ll);
1320                 mdt_lock_reg_init(ll, LCK_EX);
1321                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
1322                                         MDT_LOCAL_LOCK);
1323
1324                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
1325         }
1326
1327         /* Check if there is any other open handles after acquiring
1328          * open lock. At this point, caching open handles have been revoked
1329          * by open lock.
1330          * XXX: Now only exclusive open is supported. Need to check the
1331          * type of open for generic lease support. */
1332         if (rc == 0 && acq_lease) {
1333                 struct ptlrpc_request *req = mdt_info_req(info);
1334                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
1335                 struct mdt_file_data *mfd;
1336                 bool is_replay_or_resent;
1337                 int open_count = 0;
1338
1339                 /* For lease: application can open a file and then apply lease,
1340                  * @handle contains original open handle in that case.
1341                  * In recovery, open REQ will be replayed and the lease REQ may
1342                  * be resent that means the open handle is already stale, so we
1343                  * need to fix it up here by finding new handle. */
1344                 is_replay_or_resent = req_is_replay(req) ||
1345                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
1346
1347                 /* if the request is _not_ a replay request, rr_handle
1348                  * may be used to hold an openhandle which is issuing the
1349                  * lease request, so that this openhandle doesn't count. */
1350                 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
1351                                      is_replay_or_resent);
1352                 if (mfd != NULL)
1353                         ++open_count;
1354
1355                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
1356                         PFID(mdt_object_fid(obj)),
1357                         atomic_read(&obj->mot_open_count), open_count);
1358
1359                 if (atomic_read(&obj->mot_open_count) > open_count)
1360                         GOTO(out, rc = -EBUSY);
1361         }
1362         GOTO(out, rc);
1363
1364 out:
1365         RETURN(rc);
1366 }
1367
1368 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1369                                    struct mdt_object *obj,
1370                                    struct mdt_lock_handle *lhc,
1371                                    __u64 ibits, int rc)
1372 {
1373         __u64 open_flags = info->mti_spec.sp_cr_flags;
1374         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1375         ENTRY;
1376
1377         if (req_is_replay(mdt_info_req(info)))
1378                 RETURN_EXIT;
1379
1380         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1381          * return to client side. */
1382         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1383                 mdt_object_unlock(info, obj, ll, 1);
1384
1385         ll = &info->mti_lh[MDT_LH_LAYOUT];
1386         /* Release local layout lock, layout was created */
1387         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1388                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1389                 mdt_object_unlock(info, obj, ll, 1);
1390         }
1391
1392         if (open_flags & MDS_OPEN_LEASE)
1393                 up_write(&obj->mot_open_sem);
1394         else
1395                 up_read(&obj->mot_open_sem);
1396
1397         /* Cross-ref case, the lock should be returned to the client */
1398         if (ibits == 0 || rc == -EREMOTE)
1399                 RETURN_EXIT;
1400
1401         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
1402                 /* for the open request, the lock will only return to client
1403                  * if open or layout lock is granted. */
1404                 rc = 1;
1405         }
1406
1407         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1408                 struct ldlm_reply       *ldlm_rep;
1409
1410                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1411                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1412                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1413                         mdt_object_unlock(info, obj, lhc, 1);
1414         }
1415         RETURN_EXIT;
1416 }
1417
1418 /**
1419  * Check release is permitted for the current HSM flags.
1420  */
1421 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1422 {
1423         if (!(ma->ma_valid & MA_HSM))
1424                 return false;
1425
1426         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1427                 return false;
1428
1429         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1430                 return false;
1431
1432         return true;
1433 }
1434
1435 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1436                          struct mdt_lock_handle *lhc)
1437 {
1438         const struct lu_env     *env   = info->mti_env;
1439         struct mdt_device       *mdt   = info->mti_mdt;
1440         __u64                    flags = info->mti_spec.sp_cr_flags;
1441         struct mdt_reint_record *rr    = &info->mti_rr;
1442         struct md_attr          *ma    = &info->mti_attr;
1443         struct mdt_object       *parent= NULL;
1444         struct mdt_object       *o;
1445         int                      rc;
1446         __u64                    ibits = 0;
1447         ENTRY;
1448
1449         if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1450                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1451                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1452                         if (IS_ERR(parent)) {
1453                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1454                                        " for anonymous created %ld, try to"
1455                                        " use server-side parent.\n",
1456                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1457                                 parent = NULL;
1458                         }
1459                 }
1460                 if (parent == NULL)
1461                         ma->ma_need |= MA_PFID;
1462         }
1463
1464         o = mdt_object_find(env, mdt, rr->rr_fid2);
1465         if (IS_ERR(o))
1466                 RETURN(rc = PTR_ERR(o));
1467
1468         if (mdt_object_remote(o)) {
1469                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1470                        mdt_obd_name(info->mti_mdt),
1471                        PFID(rr->rr_fid2));
1472                 GOTO(out, rc = -EREMOTE);
1473         } else if (!mdt_object_exists(o)) {
1474                 mdt_set_disposition(info, rep,
1475                                     DISP_IT_EXECD |
1476                                     DISP_LOOKUP_EXECD |
1477                                     DISP_LOOKUP_NEG);
1478                 GOTO(out, rc = -ENOENT);
1479         }
1480
1481         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1482
1483         mdt_prep_ma_buf_from_rep(info, o, ma);
1484         if (flags & MDS_OPEN_RELEASE)
1485                 ma->ma_need |= MA_HSM;
1486         rc = mdt_attr_get_complex(info, o, ma);
1487         if (rc)
1488                 GOTO(out, rc);
1489
1490         /* If a release request, check file flags are fine and ask for an
1491          * exclusive open access. */
1492         if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1493                 GOTO(out, rc = -EPERM);
1494
1495         rc = mdt_object_open_lock(info, o, lhc, &ibits);
1496         if (rc)
1497                 GOTO(out_unlock, rc);
1498
1499         if (ma->ma_valid & MA_PFID) {
1500                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1501                 if (IS_ERR(parent)) {
1502                         CDEBUG(D_INODE, "Fail to find parent "DFID
1503                                " for anonymous created %ld, try to"
1504                                " use system default.\n",
1505                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1506                         parent = NULL;
1507                 }
1508         }
1509
1510         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1511         if (!rc) {
1512                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1513                 if (flags & MDS_OPEN_LOCK)
1514                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1515                 if (flags & MDS_OPEN_LEASE)
1516                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1517         }
1518         GOTO(out_unlock, rc);
1519
1520 out_unlock:
1521         mdt_object_open_unlock(info, o, lhc, ibits, rc);
1522 out:
1523         mdt_object_put(env, o);
1524         if (parent != NULL)
1525                 mdt_object_put(env, parent);
1526         return rc;
1527 }
1528
1529 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1530 static int mdt_cross_open(struct mdt_thread_info *info,
1531                           const struct lu_fid *parent_fid,
1532                           const struct lu_fid *fid,
1533                           struct ldlm_reply *rep, __u32 flags)
1534 {
1535         struct md_attr    *ma = &info->mti_attr;
1536         struct mdt_object *o;
1537         int                rc;
1538         ENTRY;
1539
1540         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1541         if (IS_ERR(o))
1542                 RETURN(rc = PTR_ERR(o));
1543
1544         if (mdt_object_remote(o)) {
1545                 /* Something is wrong here, the object is on another MDS! */
1546                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1547                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1548                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1549                                 &o->mot_obj,
1550                                 "Object isn't on this server! FLD error?\n");
1551                 rc = -EFAULT;
1552         } else {
1553                 if (mdt_object_exists(o)) {
1554                         /* Do permission check for cross-open. */
1555                         rc = mo_permission(info->mti_env, NULL,
1556                                            mdt_object_child(o),
1557                                            NULL, flags | MDS_OPEN_CROSS);
1558                         if (rc)
1559                                 goto out;
1560
1561                         mdt_prep_ma_buf_from_rep(info, o, ma);
1562                         mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1563                         rc = mdt_attr_get_complex(info, o, ma);
1564                         if (rc != 0)
1565                                 GOTO(out, rc);
1566
1567                         /* Do not create lov object if the fid is opened
1568                          * under OBF */
1569                         if (S_ISREG(ma->ma_attr.la_mode) &&
1570                             !(ma->ma_valid & MA_LOV) && (flags & FMODE_WRITE) &&
1571                             fid_is_obf(parent_fid))
1572                                 GOTO(out, rc = -EPERM);
1573
1574                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1575                 } else {
1576                         /*
1577                          * Something is wrong here. lookup was positive but
1578                          * there is no object!
1579                          */
1580                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1581                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1582                         rc = -EFAULT;
1583                 }
1584         }
1585 out:
1586         mdt_object_put(info->mti_env, o);
1587         RETURN(rc);
1588 }
1589
1590 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1591 {
1592         struct mdt_device       *mdt = info->mti_mdt;
1593         struct ptlrpc_request   *req = mdt_info_req(info);
1594         struct mdt_object       *parent;
1595         struct mdt_object       *child;
1596         struct mdt_lock_handle  *lh;
1597         struct ldlm_reply       *ldlm_rep;
1598         struct mdt_body         *repbody;
1599         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1600         struct md_attr          *ma = &info->mti_attr;
1601         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1602         __u64                    ibits = 0;
1603         struct mdt_reint_record *rr = &info->mti_rr;
1604         int                      result, rc;
1605         int                      created = 0;
1606         __u32                    msg_flags;
1607         ENTRY;
1608
1609         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1610                                (obd_timeout + 1) / 4);
1611
1612         mdt_counter_incr(req, LPROC_MDT_OPEN);
1613         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1614
1615         ma->ma_need = MA_INODE;
1616         ma->ma_valid = 0;
1617
1618         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1619         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1620
1621         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1622                 CERROR("file join is not supported anymore.\n");
1623                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1624         }
1625         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1626
1627         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1628             info->mti_spec.u.sp_ea.eadata == NULL)
1629                 GOTO(out, result = err_serious(-EINVAL));
1630
1631         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1632                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1633                PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1634                PFID(rr->rr_fid2), create_flags,
1635                ma->ma_attr.la_mode, msg_flags);
1636
1637         if (info->mti_cross_ref) {
1638                 /* This is cross-ref open */
1639                 mdt_set_disposition(info, ldlm_rep,
1640                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1641                              DISP_LOOKUP_POS));
1642                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1643                                         ldlm_rep, create_flags);
1644                 GOTO(out, result);
1645         } else if (req_is_replay(req) ||
1646             (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1647                 /* This is a replay request or from liblustre with ea. */
1648                 result = mdt_open_by_fid(info, ldlm_rep);
1649
1650                 if (result != -ENOENT) {
1651                         if (req->rq_export->exp_libclient &&
1652                             create_flags & MDS_OPEN_HAS_EA)
1653                                 GOTO(out, result = 0);
1654                         GOTO(out, result);
1655                 }
1656                 /* We didn't find the correct object, so we need to re-create it
1657                  * via a regular replay. */
1658                 if (!(create_flags & MDS_OPEN_CREAT)) {
1659                         DEBUG_REQ(D_ERROR, req,
1660                                   "OPEN & CREAT not in open replay/by_fid.");
1661                         GOTO(out, result = -EFAULT);
1662                 }
1663                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1664         } else if ((!lu_name_is_valid(&rr->rr_name) &&
1665                     (create_flags & MDS_OPEN_LOCK)) ||
1666                    (create_flags & MDS_OPEN_BY_FID)) {
1667                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1668                 /* If result is 0 then open by FID has found the file
1669                  * and there is nothing left for us to do here.  More
1670                  * generally if it is anything other than -ENOENT or
1671                  * -EREMOTE then we return that now.  If -ENOENT and
1672                  * MDS_OPEN_CREAT is set then we must create the file
1673                  * below.  If -EREMOTE then we need to return a LOOKUP
1674                  * lock to the client, which we do below.  Hence this
1675                  * odd looking condition.  See LU-2523. */
1676                 if (!(result == -ENOENT && (create_flags & MDS_OPEN_CREAT)) &&
1677                     result != -EREMOTE)
1678                         GOTO(out, result);
1679
1680                 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1681         }
1682
1683         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1684                 GOTO(out, result = err_serious(-ENOMEM));
1685
1686         mdt_set_disposition(info, ldlm_rep,
1687                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1688
1689         if (!lu_name_is_valid(&rr->rr_name))
1690                 GOTO(out, result = -EPROTO);
1691
1692         lh = &info->mti_lh[MDT_LH_PARENT];
1693         mdt_lock_pdo_init(lh,
1694                           (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1695                           &rr->rr_name);
1696
1697         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1698                                       MDS_INODELOCK_UPDATE);
1699         if (IS_ERR(parent))
1700                 GOTO(out, result = PTR_ERR(parent));
1701
1702         /* get and check version of parent */
1703         result = mdt_version_get_check(info, parent, 0);
1704         if (result)
1705                 GOTO(out_parent, result);
1706
1707         fid_zero(child_fid);
1708
1709         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1710                             &rr->rr_name, child_fid, &info->mti_spec);
1711
1712         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1713                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1714                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1715                  PFID(child_fid));
1716
1717         if (result != 0 && result != -ENOENT && result != -ESTALE)
1718                 GOTO(out_parent, result);
1719
1720         if (result == -ENOENT || result == -ESTALE) {
1721                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1722                 if (result == -ESTALE) {
1723                         /*
1724                          * -ESTALE means the parent is a dead(unlinked) dir, so
1725                          * it should return -ENOENT to in accordance with the
1726                          * original mds implementaion.
1727                          */
1728                         GOTO(out_parent, result = -ENOENT);
1729                 }
1730                 if (!(create_flags & MDS_OPEN_CREAT))
1731                         GOTO(out_parent, result);
1732                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1733                         GOTO(out_parent, result = -EROFS);
1734                 *child_fid = *info->mti_rr.rr_fid2;
1735                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1736                          PFID(child_fid));
1737                 /* In the function below, .hs_keycmp resolves to
1738                  * lu_obj_hop_keycmp() */
1739                 /* coverity[overrun-buffer-val] */
1740                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1741         } else {
1742                 /*
1743                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1744                  * return FID back in that case.
1745                  */
1746                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1747                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1748         }
1749         if (IS_ERR(child))
1750                 GOTO(out_parent, result = PTR_ERR(child));
1751
1752         /** check version of child  */
1753         rc = mdt_version_get_check(info, child, 1);
1754         if (rc)
1755                 GOTO(out_child, result = rc);
1756
1757         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1758         if (result == -ENOENT) {
1759                 /* Create under OBF and .lustre is not permitted */
1760                 if (fid_is_obf(rr->rr_fid1) || fid_is_dot_lustre(rr->rr_fid1))
1761                         GOTO(out_child, result = -EPERM);
1762
1763                 /* save versions in reply */
1764                 mdt_version_get_save(info, parent, 0);
1765                 mdt_version_get_save(info, child, 1);
1766
1767                 /* version of child will be changed */
1768                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1769
1770                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1771                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1772
1773                 /* Let lower layers know what is lock mode on directory. */
1774                 info->mti_spec.sp_cr_mode =
1775                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1776
1777                 /*
1778                  * Do not perform lookup sanity check. We know that name does
1779                  * not exist.
1780                  */
1781                 info->mti_spec.sp_cr_lookup = 0;
1782                 info->mti_spec.sp_feat = &dt_directory_features;
1783
1784                 result = mdo_create(info->mti_env,
1785                                     mdt_object_child(parent),
1786                                     &rr->rr_name,
1787                                     mdt_object_child(child),
1788                                     &info->mti_spec,
1789                                     &info->mti_attr);
1790                 if (result == -ERESTART) {
1791                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1792                         GOTO(out_child, result);
1793                 } else {
1794                         mdt_prep_ma_buf_from_rep(info, child, ma);
1795                         /* XXX: we should call this once, see few lines below */
1796                         if (result == 0)
1797                                 result = mdt_attr_get_complex(info, child, ma);
1798
1799                         if (result != 0)
1800                                 GOTO(out_child, result);
1801                 }
1802                 created = 1;
1803         } else {
1804                 /*
1805                  * The object is on remote node, return its FID for remote open.
1806                  */
1807                 if (mdt_object_remote(child)) {
1808                         /*
1809                          * Check if this lock already was sent to client and
1810                          * this is resent case. For resent case do not take lock
1811                          * again, use what is already granted.
1812                          */
1813                         LASSERT(lhc != NULL);
1814
1815                         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1816                                 struct ldlm_lock *lock;
1817
1818                                 LASSERT(msg_flags & MSG_RESENT);
1819
1820                                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1821                                 if (!lock) {
1822                                         CERROR("Invalid lock handle "LPX64"\n",
1823                                                lhc->mlh_reg_lh.cookie);
1824                                         LBUG();
1825                                 }
1826                                 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1827                                                         &lock->l_resource->lr_name));
1828                                 LDLM_LOCK_PUT(lock);
1829                                 rc = 0;
1830                         } else {
1831                                 mdt_lock_handle_init(lhc);
1832                                 mdt_lock_reg_init(lhc, LCK_PR);
1833
1834                                 rc = mdt_object_lock(info, child, lhc,
1835                                                      MDS_INODELOCK_LOOKUP,
1836                                                      MDT_CROSS_LOCK);
1837                         }
1838                         repbody->fid1 = *mdt_object_fid(child);
1839                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1840                         if (rc != 0)
1841                                 result = rc;
1842                         else
1843                                 result = -EREMOTE;
1844                         GOTO(out_child, result);
1845                 } else if (mdt_object_exists(child)) {
1846                         /* We have to get attr & LOV EA & HSM for this
1847                          * object. */
1848                         mdt_prep_ma_buf_from_rep(info, child, ma);
1849                         ma->ma_need |= MA_HSM;
1850                         result = mdt_attr_get_complex(info, child, ma);
1851                 } else {
1852                         /* Object does not exist. Likely FS corruption. */
1853                         CERROR("%s: name '"DNAME"' present, but FID "
1854                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1855                                PNAME(&rr->rr_name), PFID(child_fid));
1856                         GOTO(out_child, result = -EIO);
1857                 }
1858         }
1859
1860         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1861                 /* the open lock might already be gotten in
1862                  * mdt_intent_fixup_resent */
1863                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1864                 if (create_flags & MDS_OPEN_LOCK)
1865                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1866         } else {
1867                 /* get openlock if this isn't replay and client requested it */
1868                 if (!req_is_replay(req)) {
1869                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1870                         if (rc != 0)
1871                                 GOTO(out_child_unlock, result = rc);
1872                         else if (create_flags & MDS_OPEN_LOCK)
1873                                 mdt_set_disposition(info, ldlm_rep,
1874                                                     DISP_OPEN_LOCK);
1875                 }
1876         }
1877         /* Try to open it now. */
1878         rc = mdt_finish_open(info, parent, child, create_flags,
1879                              created, ldlm_rep);
1880         if (rc) {
1881                 result = rc;
1882                 /* openlock will be released if mdt_finish_open failed */
1883                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1884
1885                 if (created && create_flags & MDS_OPEN_VOLATILE) {
1886                         CERROR("%s: cannot open volatile file "DFID", orphan "
1887                                "file will be left in PENDING directory until "
1888                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1889                                PFID(mdt_object_fid(child)), rc);
1890                         GOTO(out_child_unlock, result);
1891                 }
1892
1893                 if (created) {
1894                         ma->ma_need = 0;
1895                         ma->ma_valid = 0;
1896                         ma->ma_cookie_size = 0;
1897                         rc = mdo_unlink(info->mti_env,
1898                                         mdt_object_child(parent),
1899                                         mdt_object_child(child),
1900                                         &rr->rr_name,
1901                                         &info->mti_attr, 0);
1902                         if (rc != 0)
1903                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1904                                        mdt_obd_name(info->mti_mdt),
1905                                        PFID(mdt_object_fid(child)), rc);
1906                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1907                 }
1908         }
1909         EXIT;
1910 out_child_unlock:
1911         mdt_object_open_unlock(info, child, lhc, ibits, result);
1912 out_child:
1913         mdt_object_put(info->mti_env, child);
1914 out_parent:
1915         mdt_object_unlock_put(info, parent, lh, result || !created);
1916 out:
1917         if (result)
1918                 lustre_msg_set_transno(req->rq_repmsg, 0);
1919         return result;
1920 }
1921
1922 /**
1923  * Create an orphan object use local root.
1924  */
1925 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1926                                           struct mdt_device *mdt,
1927                                           const struct lu_fid *fid,
1928                                           struct md_attr *attr, fmode_t fmode)
1929 {
1930         const struct lu_env *env = info->mti_env;
1931         struct md_op_spec *spec = &info->mti_spec;
1932         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1933         struct mdt_object *obj = NULL;
1934         struct mdt_object *local_root;
1935         static const struct lu_name lname = {
1936                 .ln_name = "i_am_nobody",
1937                 .ln_namelen = sizeof("i_am_nobody") - 1,
1938         };
1939         struct lu_ucred *uc;
1940         cfs_cap_t uc_cap_save;
1941         int rc;
1942         ENTRY;
1943
1944         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1945         if (rc != 0)
1946                 RETURN(ERR_PTR(rc));
1947
1948         local_root = mdt_object_find(env, mdt, local_root_fid);
1949         if (IS_ERR(local_root))
1950                 RETURN(local_root);
1951
1952         obj = mdt_object_new(env, mdt, fid);
1953         if (IS_ERR(obj))
1954                 GOTO(out, rc = PTR_ERR(obj));
1955
1956         spec->sp_cr_lookup = 0;
1957         spec->sp_feat = &dt_directory_features;
1958         spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1959         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1960         if (attr->ma_valid & MA_LOV) {
1961                 spec->u.sp_ea.eadata = attr->ma_lmm;
1962                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1963                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1964         } else {
1965                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1966         }
1967
1968         uc = lu_ucred(env);
1969         uc_cap_save = uc->uc_cap;
1970         uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1971         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1972                         mdt_object_child(obj), spec, attr);
1973         uc->uc_cap = uc_cap_save;
1974         if (rc < 0) {
1975                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1976                        mdt_obd_name(mdt), PFID(fid), rc);
1977                 GOTO(out, rc);
1978         }
1979
1980         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1981         if (rc < 0)
1982                 CERROR("%s: cannot open volatile file "DFID", orphan "
1983                        "file will be left in PENDING directory until "
1984                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1985                        PFID(fid), rc);
1986         GOTO(out, rc);
1987
1988 out:
1989         if (rc < 0) {
1990                 if (!IS_ERR(obj))
1991                         mdt_object_put(env, obj);
1992                 obj = ERR_PTR(rc);
1993         }
1994         mdt_object_put(env, local_root);
1995         return obj;
1996 }
1997
1998 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1999                            struct md_attr *ma)
2000 {
2001         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
2002         struct close_data      *data;
2003         struct ldlm_lock       *lease;
2004         struct mdt_object      *orphan;
2005         struct md_attr         *orp_ma;
2006         struct lu_buf          *buf;
2007         bool                    lease_broken;
2008         int                     rc;
2009         int                     rc2;
2010         ENTRY;
2011
2012         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2013                 RETURN(-EROFS);
2014
2015         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2016         if (data == NULL)
2017                 RETURN(-EPROTO);
2018
2019         lease = ldlm_handle2lock(&data->cd_handle);
2020         if (lease == NULL)
2021                 RETURN(-ESTALE);
2022
2023         /* try to hold open_sem so that nobody else can open the file */
2024         if (!down_write_trylock(&o->mot_open_sem)) {
2025                 ldlm_lock_cancel(lease);
2026                 GOTO(out_reprocess, rc = -EBUSY);
2027         }
2028
2029         /* Check if the lease open lease has already canceled */
2030         lock_res_and_lock(lease);
2031         lease_broken = ldlm_is_cancel(lease);
2032         unlock_res_and_lock(lease);
2033
2034         LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2035                    PFID(mdt_object_fid(o)), lease_broken);
2036
2037         /* Cancel server side lease. Client side counterpart should
2038          * have been cancelled. It's okay to cancel it now as we've
2039          * held mot_open_sem. */
2040         ldlm_lock_cancel(lease);
2041
2042         if (lease_broken) /* don't perform release task */
2043                 GOTO(out_unlock, rc = -ESTALE);
2044
2045         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2046                 GOTO(out_unlock, rc = -EINVAL);
2047
2048         /* ma_need was set before but it seems fine to change it in order to
2049          * avoid modifying the one from RPC */
2050         ma->ma_need = MA_HSM;
2051         rc = mdt_attr_get_complex(info, o, ma);
2052         if (rc != 0)
2053                 GOTO(out_unlock, rc);
2054
2055         if (!mdt_hsm_release_allow(ma))
2056                 GOTO(out_unlock, rc = -EPERM);
2057
2058         /* already released? */
2059         if (ma->ma_hsm.mh_flags & HS_RELEASED)
2060                 GOTO(out_unlock, rc = 0);
2061
2062         /* Compare on-disk and packed data_version */
2063         if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
2064                 CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
2065                        " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
2066                        data->cd_data_version, ma->ma_hsm.mh_arch_ver);
2067                 GOTO(out_unlock, rc = -EPERM);
2068         }
2069
2070         ma->ma_valid = MA_INODE;
2071         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
2072         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
2073         if (rc < 0)
2074                 GOTO(out_unlock, rc);
2075
2076         ma->ma_need = MA_INODE | MA_LOV;
2077         rc = mdt_attr_get_complex(info, o, ma);
2078         if (rc < 0)
2079                 GOTO(out_unlock, rc);
2080
2081         if (!(ma->ma_valid & MA_LOV)) {
2082                 /* Even empty file are released */
2083                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2084                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2085                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2086                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2087                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2088         } else {
2089                 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
2090                  * ma_lmm as lov_user_md, then it will be confused by union of
2091                  * layout_gen and stripe_offset. */
2092                 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
2093                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
2094                 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
2095                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
2096                 else
2097                         GOTO(out_unlock, rc = -EINVAL);
2098         }
2099
2100         /* Set file as released */
2101         ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
2102
2103         /* Hopefully it's not used in this call path */
2104         orp_ma = &info->mti_u.som.attr;
2105         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2106         orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
2107         orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
2108         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2109         orp_ma->ma_lmm = ma->ma_lmm;
2110         orp_ma->ma_lmm_size = ma->ma_lmm_size;
2111         orp_ma->ma_valid = MA_INODE | MA_LOV;
2112         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2113                                  FMODE_WRITE);
2114         if (IS_ERR(orphan)) {
2115                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2116                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2117                        PTR_ERR(orphan));
2118                 GOTO(out_unlock, rc = PTR_ERR(orphan));
2119         }
2120
2121         /* Set up HSM attribute for orphan object */
2122         CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
2123         buf = &info->mti_buf;
2124         buf->lb_buf = info->mti_xattr_buf;
2125         buf->lb_len = sizeof(struct hsm_attrs);
2126         ma->ma_hsm.mh_flags |= HS_RELEASED;
2127         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2128         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2129
2130         mdt_lock_reg_init(lh, LCK_EX);
2131         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2132                              MDS_INODELOCK_XATTR, MDT_LOCAL_LOCK);
2133         if (rc != 0)
2134                 GOTO(out_close, rc);
2135
2136         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2137                           XATTR_NAME_HSM, 0);
2138
2139         if (rc == 0)
2140                 /* Swap layout with orphan object */
2141                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2142                                      mdt_object_child(orphan),
2143                                      SWAP_LAYOUTS_MDS_HSM);
2144
2145         /* Release exclusive LL */
2146         mdt_object_unlock(info, o, lh, 1);
2147
2148         EXIT;
2149
2150 out_close:
2151         /* Close orphan object anyway */
2152         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2153                        FMODE_WRITE);
2154         if (rc2 < 0)
2155                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2156                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2157         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2158                         "object closed\n");
2159         mdt_object_put(info->mti_env, orphan);
2160
2161 out_unlock:
2162         up_write(&o->mot_open_sem);
2163
2164         if (rc == 0) { /* already released */
2165                 struct mdt_body *repbody;
2166                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2167                 LASSERT(repbody != NULL);
2168                 repbody->valid |= OBD_MD_FLRELEASED;
2169         }
2170
2171 out_reprocess:
2172         ldlm_reprocess_all(lease->l_resource);
2173         LDLM_LOCK_PUT(lease);
2174
2175         ma->ma_valid = 0;
2176         ma->ma_need = 0;
2177
2178         return rc;
2179 }
2180
2181 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
2182                                       MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
2183
2184 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2185 {
2186         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
2187 }
2188
2189 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2190 {
2191         struct mdt_object *o = mfd->mfd_object;
2192         struct md_object *next = mdt_object_child(o);
2193         struct md_attr *ma = &info->mti_attr;
2194         int ret = MDT_IOEPOCH_CLOSED;
2195         int rc = 0;
2196         __u64 mode;
2197         ENTRY;
2198
2199         mode = mfd->mfd_mode;
2200
2201         if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
2202                 rc = mdt_hsm_release(info, o, ma);
2203                 if (rc < 0) {
2204                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2205                                 mdt_obd_name(info->mti_mdt),
2206                                 PFID(mdt_object_fid(o)), rc);
2207                         /* continue to close even error occurred. */
2208                 }
2209         }
2210
2211         if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
2212                 mdt_write_put(o);
2213                 ret = mdt_ioepoch_close(info, o);
2214         } else if (mode & MDS_FMODE_EXEC) {
2215                 mdt_write_allow(o);
2216         } else if (mode & MDS_FMODE_EPOCH) {
2217                 ret = mdt_ioepoch_close(info, o);
2218         } else if (mode & MDS_FMODE_SOM) {
2219                 ret = mdt_som_au_close(info, o);
2220         }
2221
2222         /* Update atime on close only. */
2223         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
2224             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
2225                 /* Set the atime only. */
2226                 ma->ma_valid = MA_INODE;
2227                 ma->ma_attr.la_valid = LA_ATIME;
2228                 rc = mo_attr_set(info->mti_env, next, ma);
2229         }
2230
2231         /* If file data is modified, add the dirty flag. */
2232         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2233                 rc = mdt_add_dirty_flag(info, o, ma);
2234
2235         ma->ma_need |= MA_INODE;
2236         ma->ma_valid &= ~MA_INODE;
2237
2238         if (!MFD_CLOSED(mode))
2239                 rc = mo_close(info->mti_env, next, ma, mode);
2240
2241         if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
2242                 struct mdt_export_data *med;
2243
2244                 /* The IOepoch is still opened or SOM update is needed.
2245                  * Put mfd back into the list. */
2246                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
2247                 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
2248                                       MDS_FMODE_EPOCH : MDS_FMODE_SOM);
2249
2250                 LASSERT(mdt_info_req(info));
2251                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
2252                 spin_lock(&med->med_open_lock);
2253                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
2254                 class_handle_hash_back(&mfd->mfd_handle);
2255                 spin_unlock(&med->med_open_lock);
2256
2257                 if (ret == MDT_IOEPOCH_OPENED) {
2258                         ret = 0;
2259                 } else {
2260                         ret = -EAGAIN;
2261                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
2262                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
2263                 }
2264         } else {
2265                 /* adjust open and lease count */
2266                 if (mode & MDS_OPEN_LEASE) {
2267                         LASSERT(atomic_read(&o->mot_lease_count) > 0);
2268                         atomic_dec(&o->mot_lease_count);
2269                 }
2270                 LASSERT(atomic_read(&o->mot_open_count) > 0);
2271                 atomic_dec(&o->mot_open_count);
2272
2273                 mdt_mfd_free(mfd);
2274                 mdt_object_put(info->mti_env, o);
2275         }
2276
2277         RETURN(rc ? rc : ret);
2278 }
2279
2280 int mdt_close(struct tgt_session_info *tsi)
2281 {
2282         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2283         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2284         struct mdt_export_data *med;
2285         struct mdt_file_data   *mfd;
2286         struct mdt_object      *o;
2287         struct md_attr         *ma = &info->mti_attr;
2288         struct mdt_body        *repbody = NULL;
2289         int rc, ret = 0;
2290         ENTRY;
2291
2292         mdt_counter_incr(req, LPROC_MDT_CLOSE);
2293         /* Close may come with the Size-on-MDS update. Unpack it. */
2294         rc = mdt_close_unpack(info);
2295         if (rc)
2296                 GOTO(out, rc = err_serious(rc));
2297
2298         LASSERT(info->mti_ioepoch);
2299
2300         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2301                              info->mti_mdt->mdt_max_mdsize);
2302         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2303                              info->mti_mdt->mdt_max_cookiesize);
2304         rc = req_capsule_server_pack(info->mti_pill);
2305         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2306                 mdt_client_compatibility(info);
2307                 if (rc == 0)
2308                         mdt_fix_reply(info);
2309                 mdt_exit_ucred(info);
2310                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2311         }
2312
2313         /* Continue to close handle even if we can not pack reply */
2314         if (rc == 0) {
2315                 repbody = req_capsule_server_get(info->mti_pill,
2316                                                  &RMF_MDT_BODY);
2317                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2318                                                     &RMF_MDT_MD);
2319                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2320                                                        &RMF_MDT_MD,
2321                                                        RCL_SERVER);
2322                 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
2323                                                        &RMF_LOGCOOKIES);
2324                 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
2325                                                           &RMF_LOGCOOKIES,
2326                                                           RCL_SERVER);
2327                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
2328                 repbody->eadatasize = 0;
2329                 repbody->aclsize = 0;
2330         } else {
2331                 rc = err_serious(rc);
2332         }
2333
2334         med = &req->rq_export->exp_mdt_data;
2335         spin_lock(&med->med_open_lock);
2336         mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2337                              req_is_replay(req));
2338         if (mdt_mfd_closed(mfd)) {
2339                 spin_unlock(&med->med_open_lock);
2340                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2341                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
2342                        info->mti_ioepoch->handle.cookie);
2343                 /** not serious error since bug 3633 */
2344                 rc = -ESTALE;
2345         } else {
2346                 class_handle_unhash(&mfd->mfd_handle);
2347                 cfs_list_del_init(&mfd->mfd_list);
2348                 spin_unlock(&med->med_open_lock);
2349
2350                 /* Do not lose object before last unlink. */
2351                 o = mfd->mfd_object;
2352                 mdt_object_get(info->mti_env, o);
2353                 ret = mdt_mfd_close(info, mfd);
2354                 if (repbody != NULL)
2355                         rc = mdt_handle_last_unlink(info, o, ma);
2356                 mdt_empty_transno(info, rc);
2357                 mdt_object_put(info->mti_env, o);
2358         }
2359         if (repbody != NULL) {
2360                 mdt_client_compatibility(info);
2361                 rc = mdt_fix_reply(info);
2362         }
2363
2364         mdt_exit_ucred(info);
2365         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2366                 GOTO(out, rc = err_serious(-ENOMEM));
2367
2368         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2369                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2370                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2371 out:
2372         mdt_thread_info_fini(info);
2373         RETURN(rc ? rc : ret);
2374 }
2375
2376 /**
2377  * DONE_WRITING rpc handler.
2378  *
2379  * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
2380  * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
2381  * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
2382  * skip attributes and reconstruct the reply here.
2383  */
2384 int mdt_done_writing(struct tgt_session_info *tsi)
2385 {
2386         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2387         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2388         struct mdt_body         *repbody = NULL;
2389         struct mdt_export_data  *med;
2390         struct mdt_file_data    *mfd;
2391         int rc;
2392         ENTRY;
2393
2394         rc = req_capsule_server_pack(tsi->tsi_pill);
2395         if (rc)
2396                 GOTO(out, rc = err_serious(rc));
2397
2398         repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_MDT_BODY);
2399         repbody->eadatasize = 0;
2400         repbody->aclsize = 0;
2401
2402         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
2403         rc = mdt_close_unpack(info);
2404         if (rc)
2405                 GOTO(out, rc = err_serious(rc));
2406
2407         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2408                 mdt_exit_ucred(info);
2409                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2410         }
2411
2412         med = &info->mti_exp->exp_mdt_data;
2413         spin_lock(&med->med_open_lock);
2414         mfd = mdt_handle2mfd(med, &info->mti_ioepoch->handle,
2415                              req_is_replay(req));
2416         if (mfd == NULL) {
2417                 spin_unlock(&med->med_open_lock);
2418                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
2419                        ": cookie = "LPX64" ioepoch = "LPU64"\n",
2420                        PFID(info->mti_rr.rr_fid1),
2421                        info->mti_ioepoch->handle.cookie,
2422                        info->mti_ioepoch->ioepoch);
2423                 /* If this is a replay, reconstruct the transno. */
2424                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
2425                         rc = info->mti_ioepoch->flags & MF_SOM_AU ?
2426                              -EAGAIN : 0;
2427                         mdt_empty_transno(info, rc);
2428                 } else
2429                         rc = -ESTALE;
2430                 GOTO(error_ucred, rc);
2431         }
2432
2433         LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
2434                 mfd->mfd_mode == MDS_FMODE_TRUNC);
2435         class_handle_unhash(&mfd->mfd_handle);
2436         cfs_list_del_init(&mfd->mfd_list);
2437         spin_unlock(&med->med_open_lock);
2438
2439         /* Set EPOCH CLOSE flag if not set by client. */
2440         info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
2441         info->mti_attr.ma_valid = 0;
2442
2443         info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
2444         OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2445         if (info->mti_attr.ma_lmm == NULL)
2446                 GOTO(error_ucred, rc = -ENOMEM);
2447
2448         rc = mdt_mfd_close(info, mfd);
2449
2450         OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
2451         mdt_empty_transno(info, rc);
2452 error_ucred:
2453         mdt_exit_ucred(info);
2454 out:
2455         mdt_thread_info_fini(info);
2456         RETURN(rc);
2457 }