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, Whamcloud, Inc.
  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         cfs_mutex_lock(&o->mot_lov_mutex);
 122         if (!(o->mot_flags & MOF_LOV_CREATED)) {
 123                 rc = mdo_create_data(info->mti_env,
 124                                      p ? mdt_object_child(p) : NULL,
 125                                      mdt_object_child(o), spec, ma);
 126                 if (rc == 0)
 127                         rc = mdt_attr_get_complex(info, o, ma);
 128
 129                 if (rc == 0 && ma->ma_valid & MA_LOV)
 130                         o->mot_flags |= MOF_LOV_CREATED;
 131         }
 132         cfs_mutex_unlock(&o->mot_lov_mutex);
 133         RETURN(rc);
 134 }
 135
 136 static int mdt_ioepoch_opened(struct mdt_object *mo)
 137 {
 138         return mo->mot_ioepoch_count;
 139 }
 140
 141 int mdt_object_is_som_enabled(struct mdt_object *mo)
 142 {
 143         return !mo->mot_ioepoch;
 144 }
 145
 146 /**
 147  * Re-enable Size-on-MDS.
 148  * Call under ->mot_ioepoch_mutex.
 149  */
 150 static void mdt_object_som_enable(struct mdt_object *mo, __u64 ioepoch)
 151 {
 152         if (ioepoch == mo->mot_ioepoch) {
 153                 LASSERT(!mdt_ioepoch_opened(mo));
 154                 mo->mot_ioepoch = 0;
 155                 mo->mot_flags = 0;
 156         }
 157 }
 158
 159 /**
 160  * Open the IOEpoch. It is allowed if @writecount is not negative.
 161  * The epoch and writecount handling is performed under the mot_ioepoch_mutex.
 162  */
 163 int mdt_ioepoch_open(struct mdt_thread_info *info, struct mdt_object *o,
 164                      int created)
 165 {
 166         struct mdt_device *mdt = info->mti_mdt;
 167         int cancel = 0;
 168         int rc = 0;
 169         ENTRY;
 170
 171         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 172             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
 173                 RETURN(0);
 174
 175         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 176         if (mdt_ioepoch_opened(o)) {
 177                 /* Epoch continues even if there is no writers yet. */
 178                 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
 179                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 180         } else {
 181                 /* XXX: ->mdt_ioepoch is not initialized at the mount */
 182                 cfs_spin_lock(&mdt->mdt_ioepoch_lock);
 183                 if (mdt->mdt_ioepoch < info->mti_replayepoch)
 184                         mdt->mdt_ioepoch = info->mti_replayepoch;
 185
 186                 if (info->mti_replayepoch)
 187                         o->mot_ioepoch = info->mti_replayepoch;
 188                 else if (++mdt->mdt_ioepoch == IOEPOCH_INVAL)
 189                         o->mot_ioepoch = ++mdt->mdt_ioepoch;
 190                 else
 191                         o->mot_ioepoch = mdt->mdt_ioepoch;
 192
 193                 cfs_spin_unlock(&mdt->mdt_ioepoch_lock);
 194
 195                 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
 196                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 197                 if (created)
 198                         o->mot_flags |= MOF_SOM_CREATED;
 199                 cancel = 1;
 200         }
 201         o->mot_ioepoch_count++;
 202         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 203
 204         /* Cancel Size-on-MDS attributes cached on clients for the open case.
 205          * In the truncate case, see mdt_reint_setattr(). */
 206         if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
 207                 struct mdt_lock_handle  *lh = &info->mti_lh[MDT_LH_CHILD];
 208                 mdt_lock_reg_init(lh, LCK_EX);
 209                 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
 210                                      MDT_LOCAL_LOCK);
 211                 if (rc == 0)
 212                         mdt_object_unlock(info, o, lh, 1);
 213         }
 214         RETURN(rc);
 215 }
 216
 217 /**
 218  * Update SOM on-disk attributes.
 219  * If enabling, write update inodes and lustre-ea with the proper IOEpoch,
 220  * mountid and attributes. If disabling, zero IOEpoch id in lustre-ea.
 221  * Call under ->mot_ioepoch_mutex.
 222  */
 223 static int mdt_som_attr_set(struct mdt_thread_info *info,
 224                             struct mdt_object *obj, __u64 ioepoch, int enable)
 225 {
 226         struct lustre_mdt_attrs *lma;
 227         struct md_attr          *ma = &info->mti_attr;
 228         struct lu_buf           *buf = &info->mti_buf;
 229         struct md_object        *next = mdt_object_child(obj);
 230         struct mdt_device       *mdt = info->mti_mdt;
 231         struct lu_attr          *la = &ma->ma_attr;
 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         lma = (struct lustre_mdt_attrs *) info->mti_xattr_buf;
 240         CLASSERT(sizeof(info->mti_xattr_buf) >= sizeof(*lma));
 241
 242         buf->lb_buf = lma;
 243         buf->lb_len = sizeof(info->mti_xattr_buf);
 244         rc = mo_xattr_get(info->mti_env, next, buf, XATTR_NAME_LMA);
 245         if (rc > 0) {
 246                 lustre_lma_swab(lma);
 247         } else if (rc == -ENODATA) {
 248                 memset(lma, 0, sizeof(*lma));
 249         } else {
 250                 RETURN(rc);
 251         }
 252
 253         /* Copy FID */
 254         memcpy(&lma->lma_self_fid, mdt_object_fid(obj), sizeof(lma->lma_self_fid));
 255
 256         /* Copy SOM data */
 257         lma->lma_ioepoch     = ioepoch;
 258         lma->lma_som_size    = la->la_valid & LA_SIZE ? la->la_size : 0;
 259         lma->lma_som_blocks  = la->la_valid & LA_BLOCKS ?  la->la_blocks : 0;
 260         lma->lma_som_mountid = mdt->mdt_lut.lut_obd->u.obt.obt_mount_count;
 261         if (enable)
 262                 lma->lma_compat |= LMAC_SOM;
 263         else
 264                 lma->lma_compat &= ~LMAC_SOM;
 265
 266         rc = mo_xattr_set(info->mti_env, next, buf, XATTR_NAME_LMA, 0);
 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         cfs_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         cfs_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         cfs_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         cfs_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         cfs_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         cfs_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         cfs_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         cfs_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         cfs_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         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 517         rc = o->mot_writecount;
 518         cfs_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         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 527         if (o->mot_writecount < 0)
 528                 rc = -ETXTBSY;
 529         else
 530                 o->mot_writecount++;
 531         cfs_mutex_unlock(&o->mot_ioepoch_mutex);
 532         RETURN(rc);
 533 }
 534
 535 void mdt_write_put(struct mdt_object *o)
 536 {
 537         ENTRY;
 538         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 539         o->mot_writecount--;
 540         cfs_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         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 549         if (o->mot_writecount > 0)
 550                 rc = -ETXTBSY;
 551         else
 552                 o->mot_writecount--;
 553         cfs_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         cfs_mutex_lock(&o->mot_ioepoch_mutex);
 561         o->mot_writecount++;
 562         cfs_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         cfs_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         cfs_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         cfs_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                 cfs_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         cfs_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                 rc = mdt_write_deny(o);
 713         }
 714         if (rc)
 715                 RETURN(rc);
 716
 717         rc = mo_open(info->mti_env, mdt_object_child(o),
 718                      created ? flags | MDS_OPEN_CREATED : flags);
 719         if (rc)
 720                 GOTO(err_out, rc);
 721
 722         mfd = mdt_mfd_new();
 723         if (mfd != NULL) {
 724                 /*
 725                  * Keep a reference on this object for this open, and is
 726                  * released by mdt_mfd_close().
 727                  */
 728                 mdt_object_get(info->mti_env, o);
 729
 730                 /*
 731                  * @flags is always not zero. At least it should be FMODE_READ,
 732                  * FMODE_WRITE or MDS_FMODE_EXEC.
 733                  */
 734                 LASSERT(flags != 0);
 735
 736                 /* Open handling. */
 737                 mdt_mfd_set_mode(mfd, flags);
 738
 739                 mfd->mfd_object = o;
 740                 mfd->mfd_xid = req->rq_xid;
 741
 742                 /* replay handle */
 743                 if (req_is_replay(req)) {
 744                         struct mdt_file_data *old_mfd;
 745                         /* Check wheather old cookie already exist in
 746                          * the list, becasue when do recovery, client
 747                          * might be disconnected from server, and
 748                          * restart replay, so there maybe some orphan
 749                          * mfd here, we should remove them */
 750                         LASSERT(info->mti_rr.rr_handle != NULL);
 751                         old_mfd = mdt_handle2mfd(info, info->mti_rr.rr_handle);
 752                         if (old_mfd) {
 753                                 CDEBUG(D_HA, "del orph mfd %p fid=("DFID") "
 754                                        "cookie=" LPX64"\n", mfd,
 755                                        PFID(mdt_object_fid(mfd->mfd_object)),
 756                                        info->mti_rr.rr_handle->cookie);
 757                                 cfs_spin_lock(&med->med_open_lock);
 758                                 class_handle_unhash(&old_mfd->mfd_handle);
 759                                 cfs_list_del_init(&old_mfd->mfd_list);
 760                                 cfs_spin_unlock(&med->med_open_lock);
 761                                 /* no attr update for that close */
 762                                 la->la_valid = 0;
 763                                 ma->ma_valid |= MA_FLAGS;
 764                                 ma->ma_attr_flags |= MDS_RECOV_OPEN;
 765                                 mdt_mfd_close(info, old_mfd);
 766                                 ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 767                                 ma->ma_valid &= ~MA_FLAGS;
 768                         }
 769                         CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
 770                                info->mti_rr.rr_handle->cookie);
 771                         mfd->mfd_old_handle.cookie =
 772                                                 info->mti_rr.rr_handle->cookie;
 773                 }
 774                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 775
 776                 if (req->rq_export->exp_disconnected) {
 777                         cfs_spin_lock(&med->med_open_lock);
 778                         class_handle_unhash(&mfd->mfd_handle);
 779                         cfs_list_del_init(&mfd->mfd_list);
 780                         cfs_spin_unlock(&med->med_open_lock);
 781                         mdt_mfd_close(info, mfd);
 782                 } else {
 783                         cfs_spin_lock(&med->med_open_lock);
 784                         cfs_list_add(&mfd->mfd_list, &med->med_open_head);
 785                         cfs_spin_unlock(&med->med_open_lock);
 786                 }
 787
 788                 mdt_empty_transno(info, rc);
 789         } else {
 790                 GOTO(err_out, rc = -ENOMEM);
 791         }
 792
 793         RETURN(rc);
 794
 795 err_out:
 796         if (flags & FMODE_WRITE)
 797                         /* XXX We also need to close io epoch here.
 798                          * See LU-1220 - green */
 799                 mdt_write_put(o);
 800         else if (flags & FMODE_EXEC)
 801                 mdt_write_allow(o);
 802         return rc;
 803 }
 804
 805 int mdt_finish_open(struct mdt_thread_info *info,
 806                     struct mdt_object *p, struct mdt_object *o,
 807                     __u64 flags, int created, struct ldlm_reply *rep)
 808 {
 809         struct ptlrpc_request   *req = mdt_info_req(info);
 810         struct obd_export       *exp = req->rq_export;
 811         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 812         struct md_attr          *ma  = &info->mti_attr;
 813         struct lu_attr          *la  = &ma->ma_attr;
 814         struct mdt_file_data    *mfd;
 815         struct mdt_body         *repbody;
 816         int                      rc = 0;
 817         int                      isreg, isdir, islnk;
 818         cfs_list_t              *t;
 819         ENTRY;
 820
 821         LASSERT(ma->ma_valid & MA_INODE);
 822
 823         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 824
 825         isreg = S_ISREG(la->la_mode);
 826         isdir = S_ISDIR(la->la_mode);
 827         islnk = S_ISLNK(la->la_mode);
 828         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 829
 830         if (exp_connect_rmtclient(exp)) {
 831                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 832
 833                 rc = mdt_pack_remote_perm(info, o, buf);
 834                 if (rc) {
 835                         repbody->valid &= ~OBD_MD_FLRMTPERM;
 836                         repbody->aclsize = 0;
 837                 } else {
 838                         repbody->valid |= OBD_MD_FLRMTPERM;
 839                         repbody->aclsize = sizeof(struct mdt_remote_perm);
 840                 }
 841         }
 842 #ifdef CONFIG_FS_POSIX_ACL
 843         else if (exp->exp_connect_flags & OBD_CONNECT_ACL) {
 844                 const struct lu_env *env = info->mti_env;
 845                 struct md_object *next = mdt_object_child(o);
 846                 struct lu_buf *buf = &info->mti_buf;
 847
 848                 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 849                 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
 850                                                    RCL_SERVER);
 851                 if (buf->lb_len > 0) {
 852                         rc = mo_xattr_get(env, next, buf,
 853                                           XATTR_NAME_ACL_ACCESS);
 854                         if (rc < 0) {
 855                                 if (rc == -ENODATA) {
 856                                         repbody->aclsize = 0;
 857                                         repbody->valid |= OBD_MD_FLACL;
 858                                         rc = 0;
 859                                 } else if (rc == -EOPNOTSUPP) {
 860                                         rc = 0;
 861                                 } else {
 862                                         CERROR("got acl size: %d\n", rc);
 863                                 }
 864                         } else {
 865                                 repbody->aclsize = rc;
 866                                 repbody->valid |= OBD_MD_FLACL;
 867                                 rc = 0;
 868                         }
 869                 }
 870         }
 871 #endif
 872
 873         if (info->mti_mdt->mdt_opts.mo_mds_capa &&
 874             exp->exp_connect_flags & OBD_CONNECT_MDS_CAPA) {
 875                 struct lustre_capa *capa;
 876
 877                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
 878                 LASSERT(capa);
 879                 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
 880                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 881                 if (rc)
 882                         RETURN(rc);
 883                 repbody->valid |= OBD_MD_FLMDSCAPA;
 884         }
 885         if (info->mti_mdt->mdt_opts.mo_oss_capa &&
 886             exp->exp_connect_flags & OBD_CONNECT_OSS_CAPA &&
 887             S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) {
 888                 struct lustre_capa *capa;
 889
 890                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
 891                 LASSERT(capa);
 892                 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
 893                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 894                 if (rc)
 895                         RETURN(rc);
 896                 repbody->valid |= OBD_MD_FLOSSCAPA;
 897         }
 898
 899         /*
 900          * If we are following a symlink, don't open; and do not return open
 901          * handle for special nodes as client required.
 902          */
 903         if (islnk || (!isreg && !isdir &&
 904             (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH))) {
 905                 lustre_msg_set_transno(req->rq_repmsg, 0);
 906                 RETURN(0);
 907         }
 908
 909         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 910
 911         /*
 912          * We need to return the existing object's fid back, so it is done here,
 913          * after preparing the reply.
 914          */
 915         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 916                 RETURN(-EEXIST);
 917
 918         /* This can't be done earlier, we need to return reply body */
 919         if (isdir) {
 920                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 921                         /* We are trying to create or write an existing dir. */
 922                         RETURN(-EISDIR);
 923                 }
 924         } else if (flags & MDS_OPEN_DIRECTORY)
 925                 RETURN(-ENOTDIR);
 926
 927         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 928                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 929                 RETURN(-EAGAIN);
 930         }
 931
 932         mfd = NULL;
 933         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 934                 cfs_spin_lock(&med->med_open_lock);
 935                 cfs_list_for_each(t, &med->med_open_head) {
 936                         mfd = cfs_list_entry(t, struct mdt_file_data, mfd_list);
 937                         if (mfd->mfd_xid == req->rq_xid) {
 938                                 break;
 939                         }
 940                         mfd = NULL;
 941                 }
 942                 cfs_spin_unlock(&med->med_open_lock);
 943
 944                 if (mfd != NULL) {
 945                         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 946                         /*set repbody->ea_size for resent case*/
 947                         if (ma->ma_valid & MA_LOV) {
 948                                 LASSERT(ma->ma_lmm_size != 0);
 949                                 repbody->eadatasize = ma->ma_lmm_size;
 950                                 if (isdir)
 951                                         repbody->valid |= OBD_MD_FLDIREA;
 952                                 else
 953                                         repbody->valid |= OBD_MD_FLEASIZE;
 954                         }
 955                         RETURN(0);
 956                 }
 957         }
 958
 959         rc = mdt_mfd_open(info, p, o, flags, created);
 960         RETURN(rc);
 961 }
 962
 963 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
 964                              struct lsd_client_data *lcd);
 965
 966 void mdt_reconstruct_open(struct mdt_thread_info *info,
 967                           struct mdt_lock_handle *lhc)
 968 {
 969         const struct lu_env *env = info->mti_env;
 970         struct mdt_device       *mdt  = info->mti_mdt;
 971         struct req_capsule      *pill = info->mti_pill;
 972         struct ptlrpc_request   *req  = mdt_info_req(info);
 973         struct tg_export_data   *ted  = &req->rq_export->exp_target_data;
 974         struct lsd_client_data  *lcd  = ted->ted_lcd;
 975         struct md_attr          *ma   = &info->mti_attr;
 976         struct mdt_reint_record *rr   = &info->mti_rr;
 977         __u32                   flags = info->mti_spec.sp_cr_flags;
 978         struct ldlm_reply       *ldlm_rep;
 979         struct mdt_object       *parent;
 980         struct mdt_object       *child;
 981         struct mdt_body         *repbody;
 982         int                      rc;
 983         ENTRY;
 984
 985         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 986         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 987         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 988
 989         ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
 990         ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
 991                                                RCL_SERVER);
 992         ma->ma_need = MA_INODE;
 993         if (ma->ma_lmm_size > 0)
 994                 ma->ma_need |= MA_LOV;
 995
 996         ma->ma_valid = 0;
 997
 998         mdt_req_from_lcd(req, lcd);
 999         mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
1000
1001         CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
1002                ldlm_rep->lock_policy_res1, req->rq_status);
1003
1004         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
1005             req->rq_status != 0)
1006                 /* We did not create successfully, return error to client. */
1007                 GOTO(out, rc = req->rq_status);
1008
1009         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
1010                 struct obd_export *exp = req->rq_export;
1011                 /*
1012                  * We failed after creation, but we do not know in which step
1013                  * we failed. So try to check the child object.
1014                  */
1015                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
1016                 if (IS_ERR(parent)) {
1017                         rc = PTR_ERR(parent);
1018                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
1019                                       " Evicting client %s with export %s.\n",
1020                                       PFID(rr->rr_fid1), rc,
1021                                       obd_uuid2str(&exp->exp_client_uuid),
1022                                       obd_export_nid2str(exp));
1023                         mdt_export_evict(exp);
1024                         RETURN_EXIT;
1025                 }
1026                 child = mdt_object_find(env, mdt, rr->rr_fid2);
1027                 if (IS_ERR(child)) {
1028                         rc = PTR_ERR(child);
1029                         LCONSOLE_WARN("Child "DFID" lookup error %d."
1030                                       " Evicting client %s with export %s.\n",
1031                                       PFID(mdt_object_fid(child)), rc,
1032                                       obd_uuid2str(&exp->exp_client_uuid),
1033                                       obd_export_nid2str(exp));
1034                         mdt_object_put(env, parent);
1035                         mdt_export_evict(exp);
1036                         RETURN_EXIT;
1037                 }
1038                 rc = mdt_object_exists(child);
1039                 if (rc > 0) {
1040
1041                         mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA);
1042                         rc = mdt_attr_get_complex(info, child, ma);
1043                         if (rc == 0)
1044                               rc = mdt_finish_open(info, parent, child,
1045                                                    flags, 1, ldlm_rep);
1046                 } else if (rc < 0) {
1047                         /* the child object was created on remote server */
1048                         repbody->fid1 = *rr->rr_fid2;
1049                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1050                         rc = 0;
1051                 } else if (rc == 0) {
1052                         /* the child does not exist, we should do regular open */
1053                         mdt_object_put(env, parent);
1054                         mdt_object_put(env, child);
1055                         GOTO(regular_open, 0);
1056                 }
1057                 mdt_object_put(env, parent);
1058                 mdt_object_put(env, child);
1059                 GOTO(out, rc);
1060         } else {
1061 regular_open:
1062                 /* We did not try to create, so we are a pure open */
1063                 rc = mdt_reint_open(info, lhc);
1064         }
1065
1066         EXIT;
1067 out:
1068         req->rq_status = rc;
1069         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
1070         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
1071 }
1072
1073 int mdt_open_by_fid(struct mdt_thread_info* info,
1074                     struct ldlm_reply *rep)
1075 {
1076         __u32                    flags = info->mti_spec.sp_cr_flags;
1077         struct mdt_reint_record *rr = &info->mti_rr;
1078         struct md_attr          *ma = &info->mti_attr;
1079         struct mdt_object       *o;
1080         int                      rc;
1081         ENTRY;
1082
1083         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1084         if (IS_ERR(o))
1085                 RETURN(rc = PTR_ERR(o));
1086
1087         rc = mdt_object_exists(o);
1088         if (rc > 0) {
1089                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1090                                                 DISP_LOOKUP_EXECD |
1091                                                 DISP_LOOKUP_POS));
1092
1093                 rc = mdt_attr_get_complex(info, o, ma);
1094                 if (rc == 0)
1095                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1096         } else if (rc == 0) {
1097                 rc = -ENOENT;
1098         } else  {
1099                 /* the child object was created on remote server */
1100                 struct mdt_body *repbody;
1101                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1102                 repbody->fid1 = *rr->rr_fid2;
1103                 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1104                 rc = 0;
1105         }
1106
1107         mdt_object_put(info->mti_env, o);
1108         RETURN(rc);
1109 }
1110
1111 int mdt_open_by_fid_lock(struct mdt_thread_info *info, struct ldlm_reply *rep,
1112                          struct mdt_lock_handle *lhc)
1113 {
1114         const struct lu_env     *env   = info->mti_env;
1115         struct mdt_device       *mdt   = info->mti_mdt;
1116         __u32                    flags = info->mti_spec.sp_cr_flags;
1117         struct mdt_reint_record *rr    = &info->mti_rr;
1118         struct md_attr          *ma    = &info->mti_attr;
1119         struct mdt_object       *parent= NULL;
1120         struct mdt_object       *o;
1121         int                      rc;
1122         ldlm_mode_t              lm;
1123         ENTRY;
1124
1125         if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1126                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1127                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1128                         if (IS_ERR(parent)) {
1129                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1130                                        " for anonymous created %ld, try to"
1131                                        " use server-side parent.\n",
1132                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1133                                 parent = NULL;
1134                         }
1135                 }
1136                 if (parent == NULL)
1137                         ma->ma_need |= MA_PFID;
1138         }
1139
1140         o = mdt_object_find(env, mdt, rr->rr_fid2);
1141         if (IS_ERR(o))
1142                 RETURN(rc = PTR_ERR(o));
1143
1144         rc = mdt_object_exists(o);
1145         if (rc == 0) {
1146                 mdt_set_disposition(info, rep, (DISP_LOOKUP_EXECD |
1147                                     DISP_LOOKUP_NEG));
1148                 GOTO(out, rc = -ENOENT);
1149         } else if (rc < 0) {
1150                 CERROR("NFS remote open shouldn't happen.\n");
1151                 GOTO(out, rc);
1152         }
1153         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
1154                                         DISP_LOOKUP_EXECD |
1155                                         DISP_LOOKUP_POS));
1156
1157         if (flags & FMODE_WRITE)
1158                 lm = LCK_CW;
1159         else if (flags & MDS_FMODE_EXEC)
1160                 lm = LCK_PR;
1161         else
1162                 lm = LCK_CR;
1163
1164         mdt_lock_handle_init(lhc);
1165         mdt_lock_reg_init(lhc, lm);
1166         rc = mdt_object_lock(info, o, lhc,
1167                              MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1168                              MDT_CROSS_LOCK);
1169         if (rc)
1170                 GOTO(out, rc);
1171
1172         rc = mdt_attr_get_complex(info, o, ma);
1173         if (rc)
1174                 GOTO(out, rc);
1175
1176         if (ma->ma_valid & MA_PFID) {
1177                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1178                 if (IS_ERR(parent)) {
1179                         CDEBUG(D_INODE, "Fail to find parent "DFID
1180                                " for anonymous created %ld, try to"
1181                                " use system default.\n",
1182                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1183                         parent = NULL;
1184                 }
1185         }
1186
1187         if (flags & MDS_OPEN_LOCK)
1188                 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1189         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1190
1191         if (!(flags & MDS_OPEN_LOCK) || rc)
1192                 mdt_object_unlock(info, o, lhc, 1);
1193
1194         GOTO(out, rc);
1195 out:
1196         mdt_object_put(env, o);
1197         if (parent != NULL)
1198                 mdt_object_put(env, parent);
1199         return rc;
1200 }
1201
1202 int mdt_pin(struct mdt_thread_info* info)
1203 {
1204         ENTRY;
1205         RETURN(err_serious(-EOPNOTSUPP));
1206 }
1207
1208 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1209 int mdt_cross_open(struct mdt_thread_info* info,
1210                    const struct lu_fid *fid,
1211                    struct ldlm_reply *rep, __u32 flags)
1212 {
1213         struct md_attr    *ma = &info->mti_attr;
1214         struct mdt_object *o;
1215         int                rc;
1216         ENTRY;
1217
1218         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1219         if (IS_ERR(o))
1220                 RETURN(rc = PTR_ERR(o));
1221
1222         rc = mdt_object_exists(o);
1223         if (rc > 0) {
1224                 /* Do permission check for cross-open. */
1225                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o),
1226                                    NULL, flags | MDS_OPEN_CROSS);
1227                 if (rc)
1228                         goto out;
1229
1230                 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
1231                 rc = mdt_attr_get_complex(info, o, ma);
1232                 if (rc == 0)
1233                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1234         } else if (rc == 0) {
1235                 /*
1236                  * Something is wrong here. lookup was positive but there is
1237                  * no object!
1238                  */
1239                 CERROR("Cross-ref object doesn't exist!\n");
1240                 rc = -EFAULT;
1241         } else  {
1242                 /* Something is wrong here, the object is on another MDS! */
1243                 CERROR("The object isn't on this server! FLD error?\n");
1244                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1245                                 &o->mot_obj.mo_lu,
1246                                 "Object isn't on this server! FLD error?\n");
1247
1248                 rc = -EFAULT;
1249         }
1250
1251 out:
1252         mdt_object_put(info->mti_env, o);
1253         RETURN(rc);
1254 }
1255
1256 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1257 {
1258         struct mdt_device       *mdt = info->mti_mdt;
1259         struct ptlrpc_request   *req = mdt_info_req(info);
1260         struct mdt_object       *parent;
1261         struct mdt_object       *child;
1262         struct mdt_lock_handle  *lh;
1263         struct ldlm_reply       *ldlm_rep;
1264         struct mdt_body         *repbody;
1265         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1266         struct md_attr          *ma = &info->mti_attr;
1267         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1268         struct mdt_reint_record *rr = &info->mti_rr;
1269         struct lu_name          *lname;
1270         int                      result, rc;
1271         int                      created = 0;
1272         __u32                    msg_flags;
1273         ENTRY;
1274
1275         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1276                                (obd_timeout + 1) / 4);
1277
1278         mdt_counter_incr(req, LPROC_MDT_OPEN);
1279         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1280
1281         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1282         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1283                                                RCL_SERVER);
1284         ma->ma_need = MA_INODE;
1285         if (ma->ma_lmm_size > 0)
1286                 ma->ma_need |= MA_LOV;
1287
1288         ma->ma_valid = 0;
1289
1290         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1291         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1292
1293         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1294                 CERROR("file join is not supported anymore.\n");
1295                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1296         }
1297         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1298
1299         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1300             info->mti_spec.u.sp_ea.eadata == NULL)
1301                 GOTO(out, result = err_serious(-EINVAL));
1302
1303         CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
1304                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1305                PFID(rr->rr_fid1), rr->rr_name,
1306                PFID(rr->rr_fid2), create_flags,
1307                ma->ma_attr.la_mode, msg_flags);
1308
1309         if (req_is_replay(req) ||
1310             (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1311                 /* This is a replay request or from liblustre with ea. */
1312                 result = mdt_open_by_fid(info, ldlm_rep);
1313
1314                 if (result != -ENOENT) {
1315                         if (req->rq_export->exp_libclient &&
1316                             create_flags & MDS_OPEN_HAS_EA)
1317                                 GOTO(out, result = 0);
1318                         GOTO(out, result);
1319                 }
1320                 /* We didn't find the correct object, so we need to re-create it
1321                  * via a regular replay. */
1322                 if (!(create_flags & MDS_OPEN_CREAT)) {
1323                         DEBUG_REQ(D_ERROR, req,
1324                                   "OPEN & CREAT not in open replay/by_fid.");
1325                         GOTO(out, result = -EFAULT);
1326                 }
1327                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1328         } else if ((rr->rr_namelen == 0 && !info->mti_cross_ref &&
1329                     create_flags & MDS_OPEN_LOCK) ||
1330                    (create_flags & MDS_OPEN_BY_FID)) {
1331                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1332                 if (result != -ENOENT && !(create_flags & MDS_OPEN_CREAT))
1333                         GOTO(out, result);
1334                 if (unlikely(rr->rr_namelen == 0))
1335                         GOTO(out, result = -EINVAL);
1336                 CDEBUG(D_INFO, "No object(2), continue as regular open.\n");
1337         }
1338
1339         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1340                 GOTO(out, result = err_serious(-ENOMEM));
1341
1342         mdt_set_disposition(info, ldlm_rep,
1343                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1344
1345         if (info->mti_cross_ref) {
1346                 /* This is cross-ref open */
1347                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1348                 result = mdt_cross_open(info, rr->rr_fid1, ldlm_rep,
1349                                         create_flags);
1350                 GOTO(out, result);
1351         }
1352
1353         lh = &info->mti_lh[MDT_LH_PARENT];
1354         mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
1355                           LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
1356
1357         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
1358                                       MDS_INODELOCK_UPDATE);
1359         if (IS_ERR(parent))
1360                 GOTO(out, result = PTR_ERR(parent));
1361
1362         /* get and check version of parent */
1363         result = mdt_version_get_check(info, parent, 0);
1364         if (result)
1365                 GOTO(out_parent, result);
1366
1367         fid_zero(child_fid);
1368
1369         lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
1370         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1371                             lname, child_fid, &info->mti_spec);
1372         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1373                  "looking for "DFID"/%s, result fid="DFID"\n",
1374                  PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
1375
1376         if (result != 0 && result != -ENOENT && result != -ESTALE)
1377                 GOTO(out_parent, result);
1378
1379         if (result == -ENOENT || result == -ESTALE) {
1380                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1381                 if (result == -ESTALE) {
1382                         /*
1383                          * -ESTALE means the parent is a dead(unlinked) dir, so
1384                          * it should return -ENOENT to in accordance with the
1385                          * original mds implementaion.
1386                          */
1387                         GOTO(out_parent, result = -ENOENT);
1388                 }
1389                 if (!(create_flags & MDS_OPEN_CREAT))
1390                         GOTO(out_parent, result);
1391                 *child_fid = *info->mti_rr.rr_fid2;
1392                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1393                          PFID(child_fid));
1394                 /* In the function below, .hs_keycmp resolves to
1395                  * lu_obj_hop_keycmp() */
1396                 /* coverity[overrun-buffer-val] */
1397                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1398         } else {
1399                 /*
1400                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1401                  * return FID back in that case.
1402                  */
1403                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1404                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1405         }
1406         if (IS_ERR(child))
1407                 GOTO(out_parent, result = PTR_ERR(child));
1408
1409         /** check version of child  */
1410         rc = mdt_version_get_check(info, child, 1);
1411         if (rc)
1412                 GOTO(out_child, result = rc);
1413
1414         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
1415         if (result == -ENOENT) {
1416                 if (mdt_object_obf(parent))
1417                         GOTO(out_child, result = -EPERM);
1418
1419                 /* save versions in reply */
1420                 mdt_version_get_save(info, parent, 0);
1421                 mdt_version_get_save(info, child, 1);
1422
1423                 /* version of child will be changed */
1424                 info->mti_mos = child;
1425
1426                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1427                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1428
1429                 /* Let lower layers know what is lock mode on directory. */
1430                 info->mti_spec.sp_cr_mode =
1431                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1432
1433                 /*
1434                  * Do not perform lookup sanity check. We know that name does
1435                  * not exist.
1436                  */
1437                 info->mti_spec.sp_cr_lookup = 0;
1438                 info->mti_spec.sp_feat = &dt_directory_features;
1439
1440                 result = mdo_create(info->mti_env,
1441                                     mdt_object_child(parent),
1442                                     lname,
1443                                     mdt_object_child(child),
1444                                     &info->mti_spec,
1445                                     &info->mti_attr);
1446                 if (result == -ERESTART) {
1447                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1448                         GOTO(out_child, result);
1449                 } else {
1450
1451                         /* XXX: we should call this once, see few lines below */
1452                         if (result == 0)
1453                                 result = mdt_attr_get_complex(info, child, ma);
1454
1455                         if (result != 0)
1456                                 GOTO(out_child, result);
1457                 }
1458                 created = 1;
1459         } else {
1460                 /* We have to get attr & lov ea for this object */
1461                 result = mdt_attr_get_complex(info, child, ma);
1462                 /*
1463                  * The object is on remote node, return its FID for remote open.
1464                  */
1465                 if (result == -EREMOTE) {
1466                         /*
1467                          * Check if this lock already was sent to client and
1468                          * this is resent case. For resent case do not take lock
1469                          * again, use what is already granted.
1470                          */
1471                         LASSERT(lhc != NULL);
1472
1473                         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1474                                 struct ldlm_lock *lock;
1475
1476                                 LASSERT(msg_flags & MSG_RESENT);
1477
1478                                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1479                                 if (!lock) {
1480                                         CERROR("Invalid lock handle "LPX64"\n",
1481                                                lhc->mlh_reg_lh.cookie);
1482                                         LBUG();
1483                                 }
1484                                 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1485                                                         &lock->l_resource->lr_name));
1486                                 LDLM_LOCK_PUT(lock);
1487                                 rc = 0;
1488                         } else {
1489                                 mdt_lock_handle_init(lhc);
1490                                 mdt_lock_reg_init(lhc, LCK_PR);
1491
1492                                 rc = mdt_object_lock(info, child, lhc,
1493                                                      MDS_INODELOCK_LOOKUP,
1494                                                      MDT_CROSS_LOCK);
1495                         }
1496                         repbody->fid1 = *mdt_object_fid(child);
1497                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1498                         if (rc != 0)
1499                                 result = rc;
1500                         GOTO(out_child, result);
1501                 }
1502         }
1503
1504         LASSERT(!lustre_handle_is_used(&lhc->mlh_reg_lh));
1505
1506         /* get openlock if this is not replay and if a client requested it */
1507         if (!req_is_replay(req) && create_flags & MDS_OPEN_LOCK) {
1508                 ldlm_mode_t lm;
1509
1510                 if (create_flags & FMODE_WRITE)
1511                         lm = LCK_CW;
1512                 else if (create_flags & MDS_FMODE_EXEC)
1513                         lm = LCK_PR;
1514                 else
1515                         lm = LCK_CR;
1516                 mdt_lock_handle_init(lhc);
1517                 mdt_lock_reg_init(lhc, lm);
1518                 rc = mdt_object_lock(info, child, lhc,
1519                                      MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN,
1520                                      MDT_CROSS_LOCK);
1521                 if (rc) {
1522                         result = rc;
1523                         GOTO(out_child, result);
1524                 } else {
1525                         result = -EREMOTE;
1526                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1527                 }
1528         }
1529
1530         /* Try to open it now. */
1531         rc = mdt_finish_open(info, parent, child, create_flags,
1532                              created, ldlm_rep);
1533         if (rc) {
1534                 result = rc;
1535                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1536                         /* openlock was acquired and mdt_finish_open failed -
1537                            drop the openlock */
1538                         mdt_object_unlock(info, child, lhc, 1);
1539                 if (created) {
1540                         ma->ma_need = 0;
1541                         ma->ma_valid = 0;
1542                         ma->ma_cookie_size = 0;
1543                         rc = mdo_unlink(info->mti_env,
1544                                         mdt_object_child(parent),
1545                                         mdt_object_child(child),
1546                                         lname,
1547                                         &info->mti_attr);
1548                         if (rc != 0)
1549                                 CERROR("Error in cleanup of open\n");
1550                 }
1551         }
1552         EXIT;
1553 out_child:
1554         mdt_object_put(info->mti_env, child);
1555 out_parent:
1556         mdt_object_unlock_put(info, parent, lh, result || !created);
1557 out:
1558         if (result && result != -EREMOTE)
1559                 lustre_msg_set_transno(req->rq_repmsg, 0);
1560         return result;
1561 }
1562
1563 #define MFD_CLOSED(mode) (((mode) & ~(MDS_FMODE_EPOCH | MDS_FMODE_SOM | \
1564                                       MDS_FMODE_TRUNC)) == MDS_FMODE_CLOSED)
1565
1566 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1567 {
1568         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1569 }
1570
1571 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1572 {
1573         struct mdt_object *o = mfd->mfd_object;
1574         struct md_object *next = mdt_object_child(o);
1575         struct md_attr *ma = &info->mti_attr;
1576         int ret = MDT_IOEPOCH_CLOSED;
1577         int rc = 0;
1578         int mode;
1579         ENTRY;
1580
1581         mode = mfd->mfd_mode;
1582
1583         if ((mode & FMODE_WRITE) || (mode & MDS_FMODE_TRUNC)) {
1584                 mdt_write_put(o);
1585                 ret = mdt_ioepoch_close(info, o);
1586         } else if (mode & MDS_FMODE_EXEC) {
1587                 mdt_write_allow(o);
1588         } else if (mode & MDS_FMODE_EPOCH) {
1589                 ret = mdt_ioepoch_close(info, o);
1590         } else if (mode & MDS_FMODE_SOM) {
1591                 ret = mdt_som_au_close(info, o);
1592         }
1593
1594         /* Update atime on close only. */
1595         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1596             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1597                 /* Set the atime only. */
1598                 ma->ma_valid = MA_INODE;
1599                 ma->ma_attr.la_valid = LA_ATIME;
1600                 rc = mo_attr_set(info->mti_env, next, ma);
1601         }
1602
1603         ma->ma_need |= MA_INODE;
1604         ma->ma_valid &= ~MA_INODE;
1605
1606         if (!MFD_CLOSED(mode))
1607                 rc = mo_close(info->mti_env, next, ma, mode);
1608
1609         if (ret == MDT_IOEPOCH_GETATTR || ret == MDT_IOEPOCH_OPENED) {
1610                 struct mdt_export_data *med;
1611
1612                 /* The IOepoch is still opened or SOM update is needed.
1613                  * Put mfd back into the list. */
1614                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
1615                 mdt_mfd_set_mode(mfd, ret == MDT_IOEPOCH_OPENED ?
1616                                       MDS_FMODE_EPOCH : MDS_FMODE_SOM);
1617
1618                 LASSERT(mdt_info_req(info));
1619                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
1620                 cfs_spin_lock(&med->med_open_lock);
1621                 cfs_list_add(&mfd->mfd_list, &med->med_open_head);
1622                 class_handle_hash_back(&mfd->mfd_handle);
1623                 cfs_spin_unlock(&med->med_open_lock);
1624
1625                 if (ret == MDT_IOEPOCH_OPENED) {
1626                         ret = 0;
1627                 } else {
1628                         ret = -EAGAIN;
1629                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
1630                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
1631                 }
1632         } else {
1633                 mdt_mfd_free(mfd);
1634                 mdt_object_put(info->mti_env, o);
1635         }
1636
1637         RETURN(rc ? rc : ret);
1638 }
1639
1640 int mdt_close(struct mdt_thread_info *info)
1641 {
1642         struct mdt_export_data *med;
1643         struct mdt_file_data   *mfd;
1644         struct mdt_object      *o;
1645         struct md_attr         *ma = &info->mti_attr;
1646         struct mdt_body        *repbody = NULL;
1647         struct ptlrpc_request  *req = mdt_info_req(info);
1648         int rc, ret = 0;
1649         ENTRY;
1650
1651         mdt_counter_incr(req, LPROC_MDT_CLOSE);
1652         /* Close may come with the Size-on-MDS update. Unpack it. */
1653         rc = mdt_close_unpack(info);
1654         if (rc)
1655                 RETURN(err_serious(rc));
1656
1657         LASSERT(info->mti_ioepoch);
1658
1659         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
1660                              info->mti_mdt->mdt_max_mdsize);
1661         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
1662                              info->mti_mdt->mdt_max_cookiesize);
1663         rc = req_capsule_server_pack(info->mti_pill);
1664         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
1665                 mdt_client_compatibility(info);
1666                 if (rc == 0)
1667                         mdt_fix_reply(info);
1668                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1669         }
1670
1671         /* Continue to close handle even if we can not pack reply */
1672         if (rc == 0) {
1673                 repbody = req_capsule_server_get(info->mti_pill,
1674                                                  &RMF_MDT_BODY);
1675                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
1676                                                     &RMF_MDT_MD);
1677                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
1678                                                        &RMF_MDT_MD,
1679                                                        RCL_SERVER);
1680                 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
1681                                                        &RMF_LOGCOOKIES);
1682                 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
1683                                                           &RMF_LOGCOOKIES,
1684                                                           RCL_SERVER);
1685                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
1686                 repbody->eadatasize = 0;
1687                 repbody->aclsize = 0;
1688         } else {
1689                 rc = err_serious(rc);
1690         }
1691
1692         med = &req->rq_export->exp_mdt_data;
1693         cfs_spin_lock(&med->med_open_lock);
1694         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1695         if (mdt_mfd_closed(mfd)) {
1696                 cfs_spin_unlock(&med->med_open_lock);
1697                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
1698                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
1699                        info->mti_ioepoch->handle.cookie);
1700                 /** not serious error since bug 3633 */
1701                 rc = -ESTALE;
1702         } else {
1703                 class_handle_unhash(&mfd->mfd_handle);
1704                 cfs_list_del_init(&mfd->mfd_list);
1705                 cfs_spin_unlock(&med->med_open_lock);
1706
1707                 /* Do not lose object before last unlink. */
1708                 o = mfd->mfd_object;
1709                 mdt_object_get(info->mti_env, o);
1710                 ret = mdt_mfd_close(info, mfd);
1711                 if (repbody != NULL)
1712                         rc = mdt_handle_last_unlink(info, o, ma);
1713                 mdt_empty_transno(info, rc);
1714                 mdt_object_put(info->mti_env, o);
1715         }
1716         if (repbody != NULL) {
1717                 mdt_client_compatibility(info);
1718                 rc = mdt_fix_reply(info);
1719         }
1720
1721         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
1722                 RETURN(err_serious(-ENOMEM));
1723
1724         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
1725                                  OBD_FAIL_MDS_CLOSE_NET_REP))
1726                 info->mti_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
1727         RETURN(rc ? rc : ret);
1728 }
1729
1730 /**
1731  * DONE_WRITING rpc handler.
1732  *
1733  * As mfd is not kept after replayed CLOSE (see mdt_ioepoch_close_on_replay()),
1734  * only those DONE_WRITING rpc will be replayed which really wrote smth on disk,
1735  * and got a trasid. Waiting for such DONE_WRITING is not reliable, so just
1736  * skip attributes and reconstruct the reply here.
1737  */
1738 int mdt_done_writing(struct mdt_thread_info *info)
1739 {
1740         struct ptlrpc_request   *req = mdt_info_req(info);
1741         struct mdt_body         *repbody = NULL;
1742         struct mdt_export_data  *med;
1743         struct mdt_file_data    *mfd;
1744         int rc;
1745         ENTRY;
1746
1747         rc = req_capsule_server_pack(info->mti_pill);
1748         if (rc)
1749                 RETURN(err_serious(rc));
1750
1751         repbody = req_capsule_server_get(info->mti_pill,
1752                                          &RMF_MDT_BODY);
1753         repbody->eadatasize = 0;
1754         repbody->aclsize = 0;
1755
1756         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
1757         rc = mdt_close_unpack(info);
1758         if (rc)
1759                 RETURN(err_serious(rc));
1760
1761         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL))
1762                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1763
1764         med = &info->mti_exp->exp_mdt_data;
1765         cfs_spin_lock(&med->med_open_lock);
1766         mfd = mdt_handle2mfd(info, &info->mti_ioepoch->handle);
1767         if (mfd == NULL) {
1768                 cfs_spin_unlock(&med->med_open_lock);
1769                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
1770                        ": cookie = "LPX64" ioepoch = "LPU64"\n",
1771                        PFID(info->mti_rr.rr_fid1),
1772                        info->mti_ioepoch->handle.cookie,
1773                        info->mti_ioepoch->ioepoch);
1774                 /* If this is a replay, reconstruct the transno. */
1775                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1776                         rc = info->mti_ioepoch->flags & MF_SOM_AU ?
1777                              -EAGAIN : 0;
1778                         mdt_empty_transno(info, rc);
1779                         RETURN(rc);
1780                 }
1781                 RETURN(-ESTALE);
1782         }
1783
1784         LASSERT(mfd->mfd_mode == MDS_FMODE_EPOCH ||
1785                 mfd->mfd_mode == MDS_FMODE_TRUNC);
1786         class_handle_unhash(&mfd->mfd_handle);
1787         cfs_list_del_init(&mfd->mfd_list);
1788         cfs_spin_unlock(&med->med_open_lock);
1789
1790         /* Set EPOCH CLOSE flag if not set by client. */
1791         info->mti_ioepoch->flags |= MF_EPOCH_CLOSE;
1792         info->mti_attr.ma_valid = 0;
1793
1794         info->mti_attr.ma_lmm_size = info->mti_mdt->mdt_max_mdsize;
1795         OBD_ALLOC_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1796         if (info->mti_attr.ma_lmm == NULL)
1797                 RETURN(-ENOMEM);
1798
1799         rc = mdt_mfd_close(info, mfd);
1800
1801         OBD_FREE_LARGE(info->mti_attr.ma_lmm, info->mti_mdt->mdt_max_mdsize);
1802         mdt_empty_transno(info, rc);
1803         RETURN(rc);
1804 }