lustre/mdt/mdt_open.c

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