lustre/mdt/mdt_open.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  linux/mdt/mdt_open.c
   5  *  Lustre Metadata Target (mdt) open/close file handling
   6  *
   7  *  Copyright (C) 2002-2006 Cluster File Systems, Inc.
   8  *   Author: Huang Hua <huanghua@clusterfs.com>
   9  *
  10  *   This file is part of the Lustre file system, http://www.lustre.org
  11  *   Lustre is a trademark of Cluster File Systems, Inc.
  12  *
  13  *   You may have signed or agreed to another license before downloading
  14  *   this software.  If so, you are bound by the terms and conditions
  15  *   of that agreement, and the following does not apply to you.  See the
  16  *   LICENSE file included with this distribution for more information.
  17  *
  18  *   If you did not agree to a different license, then this copy of Lustre
  19  *   is open source software; you can redistribute it and/or modify it
  20  *   under the terms of version 2 of the GNU General Public License as
  21  *   published by the Free Software Foundation.
  22  *
  23  *   In either case, Lustre is distributed in the hope that it will be
  24  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  25  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  26  *   license text for more details.
  27  */
  28
  29 #ifndef EXPORT_SYMTAB
  30 # define EXPORT_SYMTAB
  31 #endif
  32 #define DEBUG_SUBSYSTEM S_MDS
  33
  34 #include <linux/lustre_acl.h>
  35 #include <lustre_mds.h>
  36 #include "mdt_internal.h"
  37
  38 /* we do nothing because we do not have refcount now */
  39 static void mdt_mfd_get(void *mfdp)
  40 {
  41 }
  42
  43 /* Create a new mdt_file_data struct, initialize it,
  44  * and insert it to global hash table */
  45 struct mdt_file_data *mdt_mfd_new(void)
  46 {
  47         struct mdt_file_data *mfd;
  48         ENTRY;
  49
  50         OBD_ALLOC_PTR(mfd);
  51         if (mfd != NULL) {
  52                 CFS_INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  53                 CFS_INIT_LIST_HEAD(&mfd->mfd_list);
  54                 class_handle_hash(&mfd->mfd_handle, mdt_mfd_get);
  55         }
  56         RETURN(mfd);
  57 }
  58
  59 /*
  60  * Find the mfd pointed to by handle in global hash table.
  61  * In case of replay the handle is obsoleted
  62  * but mfd can be found in mfd list by that handle
  63  */
  64 struct mdt_file_data *mdt_handle2mfd(struct mdt_thread_info *info,
  65                                      const struct lustre_handle *handle)
  66 {
  67         struct ptlrpc_request *req = mdt_info_req(info);
  68         struct mdt_file_data  *mfd;
  69         ENTRY;
  70
  71         LASSERT(handle != NULL);
  72         mfd = class_handle2object(handle->cookie);
  73         /* during dw/setattr replay the mfd can be found by old handle */
  74         if (mfd == NULL &&
  75             lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
  76                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
  77                 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  78                         if (mfd->mfd_old_handle.cookie == handle->cookie)
  79                                 RETURN (mfd);
  80                 }
  81                 mfd = NULL;
  82         }
  83         RETURN (mfd);
  84 }
  85
  86 /* free mfd */
  87 void mdt_mfd_free(struct mdt_file_data *mfd)
  88 {
  89         LASSERT(list_empty(&mfd->mfd_list));
  90         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
  91 }
  92
  93 static int mdt_create_data(struct mdt_thread_info *info,
  94                            struct mdt_object *p, struct mdt_object *o)
  95 {
  96         struct md_op_spec     *spec = &info->mti_spec;
  97         struct md_attr        *ma = &info->mti_attr;
  98         int rc;
  99         ENTRY;
 100
 101         if (!md_should_create(spec->sp_cr_flags))
 102                 RETURN(0);
 103
 104         ma->ma_need = MA_INODE | MA_LOV;
 105         ma->ma_valid = 0;
 106         rc = mdo_create_data(info->mti_env,
 107                              p ? mdt_object_child(p) : NULL,
 108                              mdt_object_child(o), spec, ma);
 109         RETURN(rc);
 110 }
 111
 112 static int mdt_epoch_opened(struct mdt_object *mo)
 113 {
 114         return mo->mot_epochcount;
 115 }
 116
 117 int mdt_sizeonmds_enabled(struct mdt_object *mo)
 118 {
 119         return !mo->mot_ioepoch;
 120 }
 121
 122 /* Re-enable Size-on-MDS. */
 123 void mdt_sizeonmds_enable(struct mdt_thread_info *info,
 124                           struct mdt_object *mo)
 125 {
 126        spin_lock(&info->mti_mdt->mdt_ioepoch_lock);
 127        if (info->mti_epoch->ioepoch == mo->mot_ioepoch) {
 128                 LASSERT(!mdt_epoch_opened(mo));
 129                 mo->mot_ioepoch = 0;
 130                 mo->mot_flags = 0;
 131        }
 132        spin_unlock(&info->mti_mdt->mdt_ioepoch_lock);
 133 }
 134
 135 /* Open the epoch. Epoch open is allowed if @writecount is not negative.
 136  * The epoch and writecount handling is performed under the mdt_ioepoch_lock. */
 137 int mdt_epoch_open(struct mdt_thread_info *info, struct mdt_object *o)
 138 {
 139         struct mdt_device *mdt = info->mti_mdt;
 140         int cancel = 0;
 141         int rc = 0;
 142         ENTRY;
 143
 144         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 145             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
 146                 RETURN(0);
 147
 148         spin_lock(&mdt->mdt_ioepoch_lock);
 149         if (mdt_epoch_opened(o)) {
 150                 /* Epoch continues even if there is no writers yet. */
 151                 CDEBUG(D_INODE, "continue epoch "LPU64" for "DFID"\n",
 152                        o->mot_ioepoch, PFID(mdt_object_fid(o)));
 153         } else {
 154                 if (info->mti_replayepoch > mdt->mdt_ioepoch)
 155                         mdt->mdt_ioepoch = info->mti_replayepoch;
 156                 else
 157                         mdt->mdt_ioepoch++;
 158                 o->mot_ioepoch = info->mti_replayepoch ?
 159                         info->mti_replayepoch : mdt->mdt_ioepoch;
 160                 CDEBUG(D_INODE, "starting epoch "LPU64" for "DFID"\n",
 161                        mdt->mdt_ioepoch, PFID(mdt_object_fid(o)));
 162                 cancel = 1;
 163         }
 164         o->mot_epochcount++;
 165         spin_unlock(&mdt->mdt_ioepoch_lock);
 166
 167         /* Cancel Size-on-MDS attributes on clients if not truncate.
 168          * In the later case, mdt_reint_setattr will do it. */
 169         if (cancel && (info->mti_rr.rr_fid1 != NULL)) {
 170                 struct mdt_lock_handle  *lh = &info->mti_lh[MDT_LH_CHILD];
 171                 mdt_lock_reg_init(lh, LCK_EX);
 172                 rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_UPDATE,
 173                                      MDT_LOCAL_LOCK);
 174                 if (rc == 0)
 175                         mdt_object_unlock(info, o, lh, 1);
 176         }
 177         RETURN(rc);
 178 }
 179
 180 /* Update the on-disk attributes if needed and re-enable Size-on-MDS caching. */
 181 static int mdt_sizeonmds_update(struct mdt_thread_info *info,
 182                                 struct mdt_object *o)
 183 {
 184         ENTRY;
 185
 186         CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
 187                o->mot_ioepoch, PFID(mdt_object_fid(o)), o->mot_epochcount);
 188
 189         if (info->mti_attr.ma_attr.la_valid & LA_SIZE) {
 190                 /* Do Size-on-MDS attribute update.
 191                  * Size-on-MDS is re-enabled inside. */
 192                 /* XXX: since we have opened the file, it is unnecessary
 193                  * to check permission when close it. Between the "open"
 194                  * and "close", maybe someone has changed the file mode
 195                  * or flags, or the file created mode do not permit wirte,
 196                  * and so on. Just set MDS_PERM_BYPASS for all the cases. */
 197                 info->mti_attr.ma_attr_flags |= MDS_PERM_BYPASS;
 198                 info->mti_attr.ma_attr.la_valid &= LA_SIZE | LA_BLOCKS |
 199                                                 LA_ATIME | LA_MTIME | LA_CTIME;
 200                 RETURN(mdt_attr_set(info, o, 0));
 201         } else
 202                 mdt_sizeonmds_enable(info, o);
 203         RETURN(0);
 204 }
 205
 206 /* Epoch closes.
 207  * Returns 1 if epoch does not close.
 208  * Returns 0 if epoch closes.
 209  * Returns -EAGAIN if epoch closes but an Size-on-MDS Update is still needed
 210  * from the client. */
 211 static int mdt_epoch_close(struct mdt_thread_info *info, struct mdt_object *o)
 212 {
 213         int eviction = (mdt_info_req(info) == NULL ? 1 : 0);
 214         struct lu_attr *la = &info->mti_attr.ma_attr;
 215         int achange = 0;
 216         int opened;
 217         int rc = 1;
 218         ENTRY;
 219
 220         if (!(mdt_conn_flags(info) & OBD_CONNECT_SOM) ||
 221             !S_ISREG(lu_object_attr(&o->mot_obj.mo_lu)))
 222                 RETURN(0);
 223
 224         spin_lock(&info->mti_mdt->mdt_ioepoch_lock);
 225
 226         /* Epoch closes only if client tells about it or eviction occures. */
 227         if (eviction || (info->mti_epoch->flags & MF_EPOCH_CLOSE)) {
 228                 LASSERT(o->mot_epochcount);
 229                 o->mot_epochcount--;
 230
 231                 CDEBUG(D_INODE, "Closing epoch "LPU64" on "DFID". Count %d\n",
 232                        o->mot_ioepoch, PFID(mdt_object_fid(o)),
 233                        o->mot_epochcount);
 234
 235                 if (!eviction)
 236                         achange = (info->mti_epoch->flags & MF_SOM_CHANGE);
 237
 238                 rc = 0;
 239                 if (!eviction && !mdt_epoch_opened(o)) {
 240                         /* Epoch ends. Is an Size-on-MDS update needed? */
 241                         if (o->mot_flags & MF_SOM_CHANGE) {
 242                                 /* Some previous writer changed the attribute.
 243                                  * Do not believe to the current Size-on-MDS
 244                                  * update, re-ask client. */
 245                                 rc = -EAGAIN;
 246                         } else if (!(la->la_valid & LA_SIZE) && achange) {
 247                                 /* Attributes were changed by the last writer
 248                                  * only but no Size-on-MDS update is received.*/
 249                                 rc = -EAGAIN;
 250                         }
 251                 }
 252
 253                 if (achange || eviction)
 254                         o->mot_flags |= MF_SOM_CHANGE;
 255         }
 256
 257         opened = mdt_epoch_opened(o);
 258         spin_unlock(&info->mti_mdt->mdt_ioepoch_lock);
 259
 260         /* If eviction occurred, do nothing. */
 261         if ((rc == 0) && !opened && !eviction) {
 262                 /* Epoch ends and wanted Size-on-MDS update is obtained. */
 263                 rc = mdt_sizeonmds_update(info, o);
 264                 /* Avoid the following setattrs of these attributes, e.g.
 265                  * for atime update. */
 266                 info->mti_attr.ma_valid = 0;
 267         }
 268         RETURN(rc);
 269 }
 270
 271 int mdt_write_read(struct mdt_device *mdt, struct mdt_object *o)
 272 {
 273         int rc = 0;
 274         ENTRY;
 275         spin_lock(&mdt->mdt_ioepoch_lock);
 276         rc = o->mot_writecount;
 277         spin_unlock(&mdt->mdt_ioepoch_lock);
 278         RETURN(rc);
 279 }
 280
 281 int mdt_write_get(struct mdt_device *mdt, struct mdt_object *o)
 282 {
 283         int rc = 0;
 284         ENTRY;
 285         spin_lock(&mdt->mdt_ioepoch_lock);
 286         if (o->mot_writecount < 0)
 287                 rc = -ETXTBSY;
 288         else
 289                 o->mot_writecount++;
 290         spin_unlock(&mdt->mdt_ioepoch_lock);
 291         RETURN(rc);
 292 }
 293
 294 static void mdt_write_put(struct mdt_device *mdt, struct mdt_object *o)
 295 {
 296         ENTRY;
 297         spin_lock(&mdt->mdt_ioepoch_lock);
 298         o->mot_writecount--;
 299         spin_unlock(&mdt->mdt_ioepoch_lock);
 300         EXIT;
 301 }
 302
 303 static int mdt_write_deny(struct mdt_device *mdt, struct mdt_object *o)
 304 {
 305         int rc = 0;
 306         ENTRY;
 307         spin_lock(&mdt->mdt_ioepoch_lock);
 308         if (o->mot_writecount > 0)
 309                 rc = -ETXTBSY;
 310         else
 311                 o->mot_writecount--;
 312         spin_unlock(&mdt->mdt_ioepoch_lock);
 313         RETURN(rc);
 314 }
 315
 316 static void mdt_write_allow(struct mdt_device *mdt, struct mdt_object *o)
 317 {
 318         ENTRY;
 319         spin_lock(&mdt->mdt_ioepoch_lock);
 320         o->mot_writecount++;
 321         spin_unlock(&mdt->mdt_ioepoch_lock);
 322         EXIT;
 323 }
 324
 325 /* there can be no real transaction so prepare the fake one */
 326 static void mdt_empty_transno(struct mdt_thread_info* info)
 327 {
 328         struct mdt_device *mdt = info->mti_mdt;
 329         struct ptlrpc_request *req = mdt_info_req(info);
 330
 331         ENTRY;
 332         /* transaction is occured already */
 333         if (lustre_msg_get_transno(req->rq_repmsg) != 0) {
 334                 EXIT;
 335                 return;
 336         }
 337
 338         spin_lock(&mdt->mdt_transno_lock);
 339         if (info->mti_transno == 0) {
 340                 info->mti_transno = ++ mdt->mdt_last_transno;
 341         } else {
 342                 /* should be replay */
 343                 if (info->mti_transno > mdt->mdt_last_transno)
 344                         mdt->mdt_last_transno = info->mti_transno;
 345         }
 346         spin_unlock(&mdt->mdt_transno_lock);
 347
 348         CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
 349                         info->mti_transno,
 350                         req->rq_export->exp_obd->obd_last_committed);
 351
 352         req->rq_transno = info->mti_transno;
 353         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 354         lustre_msg_set_last_xid(req->rq_repmsg, req->rq_xid);
 355         EXIT;
 356 }
 357
 358 void mdt_mfd_set_mode(struct mdt_file_data *mfd, int mode)
 359 {
 360         LASSERT(mfd != NULL);
 361
 362         CDEBUG(D_HA, "Change mfd %p mode 0x%x->0x%x\n",
 363                mfd, (unsigned int)mfd->mfd_mode, (unsigned int)mode);
 364
 365         mfd->mfd_mode = mode;
 366 }
 367
 368 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 369                         struct mdt_object *o, int flags, int created)
 370 {
 371         struct ptlrpc_request   *req = mdt_info_req(info);
 372         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 373         struct mdt_file_data    *mfd;
 374         struct md_attr          *ma  = &info->mti_attr;
 375         struct lu_attr          *la  = &ma->ma_attr;
 376         struct mdt_body         *repbody;
 377         int                      rc = 0, isdir, isreg;
 378         ENTRY;
 379
 380         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 381
 382         isreg = S_ISREG(la->la_mode);
 383         isdir = S_ISDIR(la->la_mode);
 384         if ((isreg && !(ma->ma_valid & MA_LOV))) {
 385                 /*
 386                  * No EA, check whether it is will set regEA and dirEA since in
 387                  * above attr get, these size might be zero, so reset it, to
 388                  * retrieve the MD after create obj.
 389                  */
 390                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 391                                                        &RMF_MDT_MD,
 392                                                        RCL_SERVER);
 393                 /* in replay case, p == NULL */
 394                 rc = mdt_create_data(info, p, o);
 395                 if (rc)
 396                         RETURN(rc);
 397         }
 398
 399         CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
 400                ma->ma_valid, ma->ma_lmm_size);
 401
 402         if (ma->ma_valid & MA_LOV) {
 403                 LASSERT(ma->ma_lmm_size != 0);
 404                 repbody->eadatasize = ma->ma_lmm_size;
 405                 if (isdir)
 406                         repbody->valid |= OBD_MD_FLDIREA;
 407                 else
 408                         repbody->valid |= OBD_MD_FLEASIZE;
 409         }
 410
 411         if (flags & FMODE_WRITE) {
 412                 rc = mdt_write_get(info->mti_mdt, o);
 413                 if (rc == 0) {
 414                         mdt_epoch_open(info, o);
 415                         repbody->ioepoch = o->mot_ioepoch;
 416                 }
 417         } else if (flags & MDS_FMODE_EXEC) {
 418                 rc = mdt_write_deny(info->mti_mdt, o);
 419         }
 420         if (rc)
 421                 RETURN(rc);
 422
 423         rc = mo_open(info->mti_env, mdt_object_child(o),
 424                      created ? flags | MDS_OPEN_CREATED : flags);
 425         if (rc)
 426                 RETURN(rc);
 427
 428         mfd = mdt_mfd_new();
 429         if (mfd != NULL) {
 430                 /*
 431                  * Keep a reference on this object for this open, and is
 432                  * released by mdt_mfd_close().
 433                  */
 434                 mdt_object_get(info->mti_env, o);
 435
 436                 /*
 437                  * @flags is always not zero. At least it should be FMODE_READ,
 438                  * FMODE_WRITE or FMODE_EXEC.
 439                  */
 440                 LASSERT(flags != 0);
 441
 442                 /* Open handling. */
 443                 mdt_mfd_set_mode(mfd, flags);
 444
 445                 mfd->mfd_object = o;
 446                 mfd->mfd_xid = req->rq_xid;
 447
 448                 /* replay handle */
 449                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
 450                         struct mdt_file_data *old_mfd;
 451                         /* Check wheather old cookie already exist in
 452                          * the list, becasue when do recovery, client
 453                          * might be disconnected from server, and
 454                          * restart replay, so there maybe some orphan
 455                          * mfd here, we should remove them */
 456                         LASSERT(info->mti_rr.rr_handle != NULL);
 457                         old_mfd = mdt_handle2mfd(info, info->mti_rr.rr_handle);
 458                         if (old_mfd) {
 459                                 CDEBUG(D_HA, "del orph mfd %p cookie" LPX64"\n",
 460                                        mfd, info->mti_rr.rr_handle->cookie);
 461                                 spin_lock(&med->med_open_lock);
 462                                 class_handle_unhash(&old_mfd->mfd_handle);
 463                                 list_del_init(&old_mfd->mfd_list);
 464                                 spin_unlock(&med->med_open_lock);
 465                                 mdt_mfd_free(old_mfd);
 466                         }
 467                         CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
 468                                info->mti_rr.rr_handle->cookie);
 469                         mfd->mfd_old_handle.cookie =
 470                                                 info->mti_rr.rr_handle->cookie;
 471                 }
 472                 spin_lock(&med->med_open_lock);
 473                 list_add(&mfd->mfd_list, &med->med_open_head);
 474                 spin_unlock(&med->med_open_lock);
 475
 476                 repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 477                 mdt_empty_transno(info);
 478         } else
 479                 rc = -ENOMEM;
 480
 481         RETURN(rc);
 482 }
 483
 484
 485 static int mdt_finish_open(struct mdt_thread_info *info,
 486                            struct mdt_object *p, struct mdt_object *o,
 487                            int flags, int created, struct ldlm_reply *rep)
 488 {
 489         struct ptlrpc_request   *req = mdt_info_req(info);
 490         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 491         struct mdt_device       *mdt = info->mti_mdt;
 492         struct md_attr          *ma  = &info->mti_attr;
 493         struct lu_attr          *la  = &ma->ma_attr;
 494         struct mdt_file_data    *mfd;
 495         struct mdt_body         *repbody;
 496         int                      rc = 0;
 497         int                      isreg, isdir, islnk;
 498         struct list_head        *t;
 499         ENTRY;
 500
 501         LASSERT(ma->ma_valid & MA_INODE);
 502
 503         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 504
 505         isreg = S_ISREG(la->la_mode);
 506         isdir = S_ISDIR(la->la_mode);
 507         islnk = S_ISLNK(la->la_mode);
 508         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 509
 510         if (med->med_rmtclient) {
 511                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 512
 513                 rc = mdt_pack_remote_perm(info, o, buf);
 514                 if (rc) {
 515                         repbody->valid &= ~OBD_MD_FLRMTPERM;
 516                         repbody->aclsize = 0;
 517                 } else {
 518                         repbody->valid |= OBD_MD_FLRMTPERM;
 519                         repbody->aclsize = sizeof(struct mdt_remote_perm);
 520                 }
 521         }
 522 #ifdef CONFIG_FS_POSIX_ACL
 523         else if (req->rq_export->exp_connect_flags & OBD_CONNECT_ACL) {
 524                 const struct lu_env *env = info->mti_env;
 525                 struct md_object *next = mdt_object_child(o);
 526                 struct lu_buf *buf = &info->mti_buf;
 527
 528                 buf->lb_buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 529                 buf->lb_len = req_capsule_get_size(info->mti_pill, &RMF_ACL,
 530                                                    RCL_SERVER);
 531                 if (buf->lb_len > 0) {
 532                         rc = mo_xattr_get(env, next, buf,
 533                                           XATTR_NAME_ACL_ACCESS);
 534                         if (rc < 0) {
 535                                 if (rc == -ENODATA) {
 536                                         repbody->aclsize = 0;
 537                                         repbody->valid |= OBD_MD_FLACL;
 538                                         rc = 0;
 539                                 } else if (rc == -EOPNOTSUPP) {
 540                                         rc = 0;
 541                                 } else {
 542                                         CERROR("got acl size: %d\n", rc);
 543                                 }
 544                         } else {
 545                                 repbody->aclsize = rc;
 546                                 repbody->valid |= OBD_MD_FLACL;
 547                                 rc = 0;
 548                         }
 549                 }
 550         }
 551 #endif
 552
 553         if (mdt->mdt_opts.mo_mds_capa) {
 554                 struct lustre_capa *capa;
 555
 556                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA1);
 557                 LASSERT(capa);
 558                 capa->lc_opc = CAPA_OPC_MDS_DEFAULT;
 559                 capa->lc_uid = 0;
 560                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 561                 if (rc)
 562                         RETURN(rc);
 563                 repbody->valid |= OBD_MD_FLMDSCAPA;
 564         }
 565         if (mdt->mdt_opts.mo_oss_capa &&
 566             S_ISREG(lu_object_attr(&o->mot_obj.mo_lu))) {
 567                 struct lustre_capa *capa;
 568
 569                 capa = req_capsule_server_get(info->mti_pill, &RMF_CAPA2);
 570                 LASSERT(capa);
 571                 capa->lc_opc = CAPA_OPC_OSS_DEFAULT | capa_open_opc(flags);
 572                 capa->lc_uid = 0;
 573                 rc = mo_capa_get(info->mti_env, mdt_object_child(o), capa, 0);
 574                 if (rc)
 575                         RETURN(rc);
 576                 repbody->valid |= OBD_MD_FLOSSCAPA;
 577         }
 578
 579         /*
 580          * If we are following a symlink, don't open; and do not return open
 581          * handle for special nodes as client required.
 582          */
 583         if (islnk || (!isreg && !isdir &&
 584             (req->rq_export->exp_connect_flags & OBD_CONNECT_NODEVOH))) {
 585                 lustre_msg_set_transno(req->rq_repmsg, 0);
 586                 RETURN(0);
 587         }
 588
 589         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 590
 591         /*
 592          * We need to return the existing object's fid back, so it is done here,
 593          * after preparing the reply.
 594          */
 595         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 596                 RETURN(-EEXIST);
 597
 598         /* This can't be done earlier, we need to return reply body */
 599         if (isdir) {
 600                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 601                         /* We are trying to create or write an existing dir. */
 602                         RETURN(-EISDIR);
 603                 }
 604         } else if (flags & MDS_OPEN_DIRECTORY)
 605                 RETURN(-ENOTDIR);
 606
 607         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 608                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 609                 RETURN(-EAGAIN);
 610         }
 611
 612         mfd = NULL;
 613         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 614                 spin_lock(&med->med_open_lock);
 615                 list_for_each(t, &med->med_open_head) {
 616                         mfd = list_entry(t, struct mdt_file_data, mfd_list);
 617                         if (mfd->mfd_xid == req->rq_xid) {
 618                                 break;
 619                         }
 620                         mfd = NULL;
 621                 }
 622                 spin_unlock(&med->med_open_lock);
 623
 624                 if (mfd != NULL) {
 625                         repbody->handle.cookie = mfd->mfd_handle.h_cookie;
 626                         /*set repbody->ea_size for resent case*/
 627                         if (ma->ma_valid & MA_LOV) {
 628                                 LASSERT(ma->ma_lmm_size != 0);
 629                                 repbody->eadatasize = ma->ma_lmm_size;
 630                                 if (isdir)
 631                                         repbody->valid |= OBD_MD_FLDIREA;
 632                                 else
 633                                         repbody->valid |= OBD_MD_FLEASIZE;
 634                         }
 635                         RETURN(0);
 636                 }
 637         }
 638
 639         rc = mdt_mfd_open(info, p, o, flags, created);
 640         RETURN(rc);
 641 }
 642
 643 extern void mdt_req_from_lcd(struct ptlrpc_request *req,
 644                              struct lsd_client_data *lcd);
 645
 646 void mdt_reconstruct_open(struct mdt_thread_info *info,
 647                           struct mdt_lock_handle *lhc)
 648 {
 649         const struct lu_env *env = info->mti_env;
 650         struct mdt_device       *mdt  = info->mti_mdt;
 651         struct req_capsule      *pill = info->mti_pill;
 652         struct ptlrpc_request   *req  = mdt_info_req(info);
 653         struct mdt_export_data  *med  = &req->rq_export->exp_mdt_data;
 654         struct lsd_client_data  *lcd  = med->med_lcd;
 655         struct md_attr          *ma   = &info->mti_attr;
 656         struct mdt_reint_record *rr   = &info->mti_rr;
 657         __u32                   flags = info->mti_spec.sp_cr_flags;
 658         struct ldlm_reply       *ldlm_rep;
 659         struct mdt_object       *parent;
 660         struct mdt_object       *child;
 661         struct mdt_body         *repbody;
 662         int                      rc;
 663         ENTRY;
 664
 665         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 666         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 667         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 668
 669         ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
 670         ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
 671                                                RCL_SERVER);
 672         ma->ma_need = MA_INODE | MA_LOV;
 673         ma->ma_valid = 0;
 674
 675         mdt_req_from_lcd(req, med->med_lcd);
 676         mdt_set_disposition(info, ldlm_rep, lcd->lcd_last_data);
 677
 678         CERROR("This is reconstruct open: disp="LPX64", result=%d\n",
 679                 ldlm_rep->lock_policy_res1, req->rq_status);
 680
 681         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
 682             req->rq_status != 0)
 683                 /* We did not create successfully, return error to client. */
 684                 GOTO(out, rc = req->rq_status);
 685
 686         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
 687                 struct obd_export *exp = req->rq_export;
 688                 /*
 689                  * We failed after creation, but we do not know in which step
 690                  * we failed. So try to check the child object.
 691                  */
 692                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
 693                 if (IS_ERR(parent)) {
 694                         rc = PTR_ERR(parent);
 695                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
 696                                       " Evicting client %s with export %s.\n",
 697                                       PFID(mdt_object_fid(parent)), rc,
 698                                       obd_uuid2str(&exp->exp_client_uuid),
 699                                       obd_export_nid2str(exp));
 700                         mdt_export_evict(exp);
 701                         EXIT;
 702                         return;
 703                 }
 704                 child = mdt_object_find(env, mdt, rr->rr_fid2);
 705                 if (IS_ERR(child)) {
 706                         rc = PTR_ERR(parent);
 707                         LCONSOLE_WARN("Child "DFID" lookup error %d."
 708                                       " Evicting client %s with export %s.\n",
 709                                       PFID(mdt_object_fid(child)), rc,
 710                                       obd_uuid2str(&exp->exp_client_uuid),
 711                                       obd_export_nid2str(exp));
 712                         mdt_export_evict(exp);
 713                         EXIT;
 714                         return;
 715                 }
 716                 rc = mdt_object_exists(child);
 717                 if (rc > 0) {
 718                         struct md_object *next;
 719
 720                         mdt_set_capainfo(info, 1, rr->rr_fid2, BYPASS_CAPA);
 721                         next = mdt_object_child(child);
 722                         rc = mo_attr_get(env, next, ma);
 723                         if (rc == 0)
 724                               rc = mdt_finish_open(info, parent, child,
 725                                                    flags, 1, ldlm_rep);
 726                 } else if (rc < 0) {
 727                         /* the child object was created on remote server */
 728                         repbody->fid1 = *rr->rr_fid2;
 729                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
 730                         rc = 0;
 731                 } else if (rc == 0) {
 732                         /* the child does not exist, we should do regular open */
 733                         mdt_object_put(env, parent);
 734                         mdt_object_put(env, child);
 735                         GOTO(regular_open, 0);
 736                 }
 737                 mdt_object_put(env, parent);
 738                 mdt_object_put(env, child);
 739                 GOTO(out, rc);
 740         } else {
 741 regular_open:
 742                 /* We did not try to create, so we are a pure open */
 743                 rc = mdt_reint_open(info, lhc);
 744         }
 745
 746         EXIT;
 747 out:
 748         req->rq_status = rc;
 749         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 750         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
 751 }
 752
 753 static int mdt_open_by_fid(struct mdt_thread_info* info,
 754                            struct ldlm_reply *rep)
 755 {
 756         const struct lu_env     *env = info->mti_env;
 757         __u32                    flags = info->mti_spec.sp_cr_flags;
 758         struct mdt_reint_record *rr = &info->mti_rr;
 759         struct md_attr          *ma = &info->mti_attr;
 760         struct mdt_object       *o;
 761         int                      rc;
 762         ENTRY;
 763
 764         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
 765         if (IS_ERR(o))
 766                 RETURN(rc = PTR_ERR(o));
 767
 768         rc = mdt_object_exists(o);
 769         if (rc > 0) {
 770                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 771                                                 DISP_LOOKUP_EXECD |
 772                                                 DISP_LOOKUP_POS));
 773
 774                 rc = mo_attr_get(env, mdt_object_child(o), ma);
 775                 if (rc == 0)
 776                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
 777         } else if (rc == 0) {
 778                 rc = -ENOENT;
 779         } else  {
 780                 /* the child object was created on remote server */
 781                 struct mdt_body *repbody;
 782                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 783                 repbody->fid1 = *rr->rr_fid2;
 784                 repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
 785                 rc = 0;
 786         }
 787
 788         mdt_object_put(info->mti_env, o);
 789         RETURN(rc);
 790 }
 791
 792 int mdt_pin(struct mdt_thread_info* info)
 793 {
 794         ENTRY;
 795         RETURN(err_serious(-EOPNOTSUPP));
 796 }
 797
 798 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
 799 static int mdt_cross_open(struct mdt_thread_info* info,
 800                           const struct lu_fid *fid,
 801                           struct ldlm_reply *rep, __u32 flags)
 802 {
 803         struct md_attr    *ma = &info->mti_attr;
 804         struct mdt_object *o;
 805         int                rc;
 806         ENTRY;
 807
 808         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
 809         if (IS_ERR(o))
 810                 RETURN(rc = PTR_ERR(o));
 811
 812         rc = mdt_object_exists(o);
 813         if (rc > 0) {
 814                 /* Do permission check for cross-open. */
 815                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o),
 816                                    NULL, flags | MDS_OPEN_CROSS);
 817                 if (rc)
 818                         goto out;
 819
 820                 mdt_set_capainfo(info, 0, fid, BYPASS_CAPA);
 821                 rc = mo_attr_get(info->mti_env, mdt_object_child(o), ma);
 822                 if (rc == 0)
 823                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
 824         } else if (rc == 0) {
 825                 /*
 826                  * Something is wrong here. lookup was positive but there is
 827                  * no object!
 828                  */
 829                 CERROR("Cross-ref object doesn't exist!\n");
 830                 rc = -EFAULT;
 831         } else  {
 832                 /* Something is wrong here, the object is on another MDS! */
 833                 CERROR("The object isn't on this server! FLD error?\n");
 834                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
 835                                 &o->mot_obj.mo_lu,
 836                                 "Object isn't on this server! FLD error?\n");
 837
 838                 rc = -EFAULT;
 839         }
 840
 841 out:
 842         mdt_object_put(info->mti_env, o);
 843         RETURN(rc);
 844 }
 845
 846 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
 847 {
 848         struct mdt_device       *mdt = info->mti_mdt;
 849         struct ptlrpc_request   *req = mdt_info_req(info);
 850         struct mdt_object       *parent;
 851         struct mdt_object       *child;
 852         struct mdt_lock_handle  *lh;
 853         struct ldlm_reply       *ldlm_rep;
 854         struct mdt_body         *repbody;
 855         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
 856         struct md_attr          *ma = &info->mti_attr;
 857         __u32                    create_flags = info->mti_spec.sp_cr_flags;
 858         struct mdt_reint_record *rr = &info->mti_rr;
 859         struct lu_name          *lname;
 860         int                      result;
 861         int                      created = 0;
 862         ENTRY;
 863
 864         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
 865                                (obd_timeout + 1) / 4);
 866
 867         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 868
 869         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
 870         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
 871                                                RCL_SERVER);
 872         ma->ma_need = MA_INODE | MA_LOV;
 873         ma->ma_valid = 0;
 874
 875         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 876         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
 877
 878         /* TODO: JOIN file */
 879         if (create_flags & MDS_OPEN_JOIN_FILE) {
 880                 CERROR("JOIN file will be supported soon\n");
 881                 GOTO(out, result = err_serious(-EOPNOTSUPP));
 882         }
 883
 884         CDEBUG(D_INODE, "I am going to open "DFID"/(%s->"DFID") "
 885                "cr_flag=0%o mode=0%06o msg_flag=0x%x\n",
 886                PFID(rr->rr_fid1), rr->rr_name,
 887                PFID(rr->rr_fid2), create_flags,
 888                ma->ma_attr.la_mode, lustre_msg_get_flags(req->rq_reqmsg));
 889
 890         if ((lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) ||
 891             (req->rq_export->exp_libclient && create_flags&MDS_OPEN_HAS_EA)) {
 892                 /* This is a replay request or from liblustre with ea. */
 893                 result = mdt_open_by_fid(info, ldlm_rep);
 894
 895                 if (result != -ENOENT) {
 896                         if (req->rq_export->exp_libclient &&
 897                             create_flags&MDS_OPEN_HAS_EA)
 898                                 GOTO(out, result = 0);
 899                         GOTO(out, result);
 900                 }
 901                 /*
 902                  * We didn't find the correct object, so we need to re-create it
 903                  * via a regular replay.
 904                  */
 905                 if (!(create_flags & MDS_OPEN_CREAT)) {
 906                         DEBUG_REQ(D_ERROR, req,"OPEN & CREAT not in open replay.");
 907                         GOTO(out, result = -EFAULT);
 908                 }
 909                 CDEBUG(D_INFO, "Open replay did find object, continue as "
 910                        "regular open\n");
 911         }
 912
 913         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
 914                 GOTO(out, result = err_serious(-ENOMEM));
 915
 916         mdt_set_disposition(info, ldlm_rep,
 917                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
 918
 919         if (info->mti_cross_ref) {
 920                 /* This is cross-ref open */
 921                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
 922                 result = mdt_cross_open(info, rr->rr_fid1, ldlm_rep,
 923                                         create_flags);
 924                 GOTO(out, result);
 925         }
 926
 927         lh = &info->mti_lh[MDT_LH_PARENT];
 928         mdt_lock_pdo_init(lh, (create_flags & MDS_OPEN_CREAT) ?
 929                           LCK_PW : LCK_PR, rr->rr_name, rr->rr_namelen);
 930
 931         parent = mdt_object_find_lock(info, rr->rr_fid1, lh,
 932                                       MDS_INODELOCK_UPDATE);
 933         if (IS_ERR(parent))
 934                 GOTO(out, result = PTR_ERR(parent));
 935
 936         fid_zero(child_fid);
 937
 938         lname = mdt_name(info->mti_env, (char *)rr->rr_name, rr->rr_namelen);
 939
 940         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
 941                             lname, child_fid, &info->mti_spec);
 942         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
 943                  "looking for "DFID"/%s, result fid="DFID"\n",
 944                  PFID(mdt_object_fid(parent)), rr->rr_name, PFID(child_fid));
 945
 946         if (result != 0 && result != -ENOENT && result != -ESTALE)
 947                 GOTO(out_parent, result);
 948
 949         if (result == -ENOENT || result == -ESTALE) {
 950                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
 951                 if (result == -ESTALE) {
 952                         /*
 953                          * -ESTALE means the parent is a dead(unlinked) dir, so
 954                          * it should return -ENOENT to in accordance with the
 955                          * original mds implementaion.
 956                          */
 957                         GOTO(out_parent, result = -ENOENT);
 958                 }
 959                 if (!(create_flags & MDS_OPEN_CREAT))
 960                         GOTO(out_parent, result);
 961                 *child_fid = *info->mti_rr.rr_fid2;
 962                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
 963                          PFID(child_fid));
 964         } else {
 965                 /*
 966                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
 967                  * return FID back in that case.
 968                  */
 969                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
 970         }
 971
 972         child = mdt_object_find(info->mti_env, mdt, child_fid);
 973         if (IS_ERR(child))
 974                 GOTO(out_parent, result = PTR_ERR(child));
 975
 976         mdt_set_capainfo(info, 1, child_fid, BYPASS_CAPA);
 977         if (result == -ENOENT) {
 978                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
 979                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
 980
 981                 /* Let lower layers know what is lock mode on directory. */
 982                 info->mti_spec.sp_cr_mode =
 983                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
 984
 985                 /*
 986                  * Do not perform lookup sanity check. We know that name does
 987                  * not exist.
 988                  */
 989                 info->mti_spec.sp_cr_lookup = 0;
 990
 991                 result = mdo_create(info->mti_env,
 992                                     mdt_object_child(parent),
 993                                     lname,
 994                                     mdt_object_child(child),
 995                                     &info->mti_spec,
 996                                     &info->mti_attr);
 997                 if (result == -ERESTART) {
 998                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
 999                         GOTO(out_child, result);
1000                 } else {
1001                         if (result != 0)
1002                                 GOTO(out_child, result);
1003                 }
1004                 created = 1;
1005         } else {
1006                 /* We have to get attr & lov ea for this object */
1007                 result = mo_attr_get(info->mti_env, mdt_object_child(child),
1008                                      ma);
1009                 /*
1010                  * The object is on remote node, return its FID for remote open.
1011                  */
1012                 if (result == -EREMOTE) {
1013                         int rc;
1014
1015                         /*
1016                          * Check if this lock already was sent to client and
1017                          * this is resent case. For resent case do not take lock
1018                          * again, use what is already granted.
1019                          */
1020                         LASSERT(lhc != NULL);
1021
1022                         if (lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1023                                 struct ldlm_lock *lock;
1024
1025                                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) &
1026                                         MSG_RESENT);
1027
1028                                 lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
1029                                 if (!lock) {
1030                                         CERROR("Invalid lock handle "LPX64"\n",
1031                                                lhc->mlh_reg_lh.cookie);
1032                                         LBUG();
1033                                 }
1034                                 LASSERT(fid_res_name_eq(mdt_object_fid(child),
1035                                                         &lock->l_resource->lr_name));
1036                                 LDLM_LOCK_PUT(lock);
1037                                 rc = 0;
1038                         } else {
1039                                 mdt_lock_handle_init(lhc);
1040                                 mdt_lock_reg_init(lhc, LCK_PR);
1041
1042                                 rc = mdt_object_lock(info, child, lhc,
1043                                                      MDS_INODELOCK_LOOKUP,
1044                                                      MDT_CROSS_LOCK);
1045                         }
1046                         repbody->fid1 = *mdt_object_fid(child);
1047                         repbody->valid |= (OBD_MD_FLID | OBD_MD_MDS);
1048                         if (rc != 0)
1049                                 result = rc;
1050                         GOTO(out_child, result);
1051                 }
1052         }
1053
1054         /* Try to open it now. */
1055         result = mdt_finish_open(info, parent, child, create_flags,
1056                                  created, ldlm_rep);
1057
1058         if (result != 0 && created) {
1059                 int rc2;
1060                 ma->ma_need = 0;
1061                 ma->ma_valid = 0;
1062                 ma->ma_cookie_size = 0;
1063                 info->mti_no_need_trans = 1;
1064                 rc2 = mdo_unlink(info->mti_env,
1065                                  mdt_object_child(parent),
1066                                  mdt_object_child(child),
1067                                  lname,
1068                                  &info->mti_attr);
1069                 if (rc2 != 0)
1070                         CERROR("Error in cleanup of open\n");
1071         }
1072         EXIT;
1073 out_child:
1074         mdt_object_put(info->mti_env, child);
1075 out_parent:
1076         mdt_object_unlock_put(info, parent, lh, result);
1077 out:
1078         if (result)
1079                 lustre_msg_set_transno(req->rq_repmsg, 0);
1080         return result;
1081 }
1082
1083 #define MFD_CLOSED(mode) (((mode) & ~(FMODE_EPOCH | FMODE_SOM | \
1084                                       FMODE_EPOCHLCK)) == FMODE_CLOSED)
1085
1086 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1087 {
1088         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1089 }
1090
1091 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1092 {
1093         struct mdt_object *o = mfd->mfd_object;
1094         struct md_object *next = mdt_object_child(o);
1095         struct md_attr *ma = &info->mti_attr;
1096         int rc = 0, ret = 0;
1097         int mode;
1098         ENTRY;
1099
1100         mode = mfd->mfd_mode;
1101
1102         if ((mode & FMODE_WRITE) || (mode & FMODE_EPOCHLCK)) {
1103                 mdt_write_put(info->mti_mdt, o);
1104                 ret = mdt_epoch_close(info, o);
1105         } else if (mode & MDS_FMODE_EXEC) {
1106                 mdt_write_allow(info->mti_mdt, o);
1107         } else if (mode & FMODE_EPOCH) {
1108                 ret = mdt_epoch_close(info, o);
1109         }
1110
1111         /* Update atime on close only. */
1112         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1113             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1114                 /* Set the atime only. */
1115                 ma->ma_attr.la_valid = LA_ATIME;
1116                 rc = mo_attr_set(info->mti_env, next, ma);
1117         }
1118
1119         ma->ma_need |= MA_INODE;
1120         ma->ma_valid = 0;
1121
1122         if (!MFD_CLOSED(mode))
1123                 rc = mo_close(info->mti_env, next, ma);
1124         else if (ret == -EAGAIN)
1125                 rc = mo_attr_get(info->mti_env, next, ma);
1126
1127         /* If the object is unlinked, do not try to re-enable SIZEONMDS */
1128         if ((ret == -EAGAIN) && (ma->ma_valid & MA_INODE) &&
1129             (ma->ma_attr.la_nlink == 0)) {
1130                 ret = 0;
1131         }
1132
1133         if ((ret == -EAGAIN) || (ret == 1)) {
1134                 struct mdt_export_data *med;
1135
1136                 /* The epoch has not closed or Size-on-MDS update is needed.
1137                  * Put mfd back into the list. */
1138                 LASSERT(mdt_conn_flags(info) & OBD_CONNECT_SOM);
1139                 mdt_mfd_set_mode(mfd, (ret == 1 ? FMODE_EPOCH : FMODE_SOM));
1140
1141                 LASSERT(mdt_info_req(info));
1142                 med = &mdt_info_req(info)->rq_export->exp_mdt_data;
1143                 spin_lock(&med->med_open_lock);
1144                 list_add(&mfd->mfd_list, &med->med_open_head);
1145                 class_handle_hash_back(&mfd->mfd_handle);
1146                 spin_unlock(&med->med_open_lock);
1147
1148                 if (ret == 1) {
1149                         ret = 0;
1150                 } else {
1151                         CDEBUG(D_INODE, "Size-on-MDS attribute update is "
1152                                "needed on "DFID"\n", PFID(mdt_object_fid(o)));
1153                 }
1154         } else {
1155                 mdt_mfd_free(mfd);
1156                 mdt_object_put(info->mti_env, o);
1157         }
1158
1159         RETURN(rc ? rc : ret);
1160 }
1161
1162 int mdt_close(struct mdt_thread_info *info)
1163 {
1164         struct mdt_export_data *med;
1165         struct mdt_file_data   *mfd;
1166         struct mdt_object      *o;
1167         struct md_attr         *ma = &info->mti_attr;
1168         struct mdt_body        *repbody = NULL;
1169         struct ptlrpc_request  *req = mdt_info_req(info);
1170         int rc, ret = 0;
1171         ENTRY;
1172
1173         /* Close may come with the Size-on-MDS update. Unpack it. */
1174         rc = mdt_close_unpack(info);
1175         if (rc)
1176                 RETURN(err_serious(rc));
1177
1178         LASSERT(info->mti_epoch);
1179
1180         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
1181                              info->mti_mdt->mdt_max_mdsize);
1182         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
1183                              info->mti_mdt->mdt_max_cookiesize);
1184         rc = req_capsule_server_pack(info->mti_pill);
1185         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL))
1186                 RETURN(lustre_msg_get_status(req->rq_repmsg));
1187
1188         /* Continue to close handle even if we can not pack reply */
1189         if (rc == 0) {
1190                 repbody = req_capsule_server_get(info->mti_pill,
1191                                                  &RMF_MDT_BODY);
1192                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
1193                                                     &RMF_MDT_MD);
1194                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
1195                                                        &RMF_MDT_MD,
1196                                                        RCL_SERVER);
1197                 ma->ma_cookie = req_capsule_server_get(info->mti_pill,
1198                                                        &RMF_LOGCOOKIES);
1199                 ma->ma_cookie_size = req_capsule_get_size(info->mti_pill,
1200                                                           &RMF_LOGCOOKIES,
1201                                                           RCL_SERVER);
1202                 ma->ma_need = MA_INODE | MA_LOV | MA_COOKIE;
1203                 repbody->eadatasize = 0;
1204                 repbody->aclsize = 0;
1205         } else
1206                 rc = err_serious(rc);
1207
1208         med = &req->rq_export->exp_mdt_data;
1209         spin_lock(&med->med_open_lock);
1210         mfd = mdt_handle2mfd(info, &info->mti_epoch->handle);
1211         if (mdt_mfd_closed(mfd)) {
1212                 spin_unlock(&med->med_open_lock);
1213                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
1214                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
1215                        info->mti_epoch->handle.cookie);
1216                 rc = err_serious(-ESTALE);
1217         } else {
1218                 class_handle_unhash(&mfd->mfd_handle);
1219                 list_del_init(&mfd->mfd_list);
1220                 spin_unlock(&med->med_open_lock);
1221
1222                 /* Do not lose object before last unlink. */
1223                 o = mfd->mfd_object;
1224                 mdt_object_get(info->mti_env, o);
1225                 ret = mdt_mfd_close(info, mfd);
1226                 if (repbody != NULL)
1227                         rc = mdt_handle_last_unlink(info, o, ma);
1228                 mdt_empty_transno(info);
1229                 mdt_object_put(info->mti_env, o);
1230         }
1231         if (repbody != NULL)
1232                 mdt_shrink_reply(info);
1233
1234         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
1235                 RETURN(err_serious(-ENOMEM));
1236
1237         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
1238                                  OBD_FAIL_MDS_CLOSE_NET_REP))
1239                 info->mti_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
1240         RETURN(rc ? rc : ret);
1241 }
1242
1243 int mdt_done_writing(struct mdt_thread_info *info)
1244 {
1245         struct mdt_body         *repbody = NULL;
1246         struct mdt_export_data  *med;
1247         struct mdt_file_data    *mfd;
1248         int rc;
1249         ENTRY;
1250
1251         rc = req_capsule_server_pack(info->mti_pill);
1252         if (rc)
1253                 RETURN(err_serious(rc));
1254
1255         repbody = req_capsule_server_get(info->mti_pill,
1256                                          &RMF_MDT_BODY);
1257         repbody->eadatasize = 0;
1258         repbody->aclsize = 0;
1259
1260         /* Done Writing may come with the Size-on-MDS update. Unpack it. */
1261         rc = mdt_close_unpack(info);
1262         if (rc)
1263                 RETURN(err_serious(rc));
1264
1265         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL))
1266                 RETURN(lustre_msg_get_status(mdt_info_req(info)->rq_repmsg));
1267
1268         med = &info->mti_exp->exp_mdt_data;
1269         spin_lock(&med->med_open_lock);
1270         mfd = mdt_handle2mfd(info, &info->mti_epoch->handle);
1271         if (mfd == NULL) {
1272                 spin_unlock(&med->med_open_lock);
1273                 CDEBUG(D_INODE, "no handle for done write: fid = "DFID
1274                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
1275                        info->mti_epoch->handle.cookie);
1276                 RETURN(-ESTALE);
1277         }
1278
1279         LASSERT(mfd->mfd_mode == FMODE_EPOCH ||
1280                 mfd->mfd_mode == FMODE_EPOCHLCK);
1281         class_handle_unhash(&mfd->mfd_handle);
1282         list_del_init(&mfd->mfd_list);
1283         spin_unlock(&med->med_open_lock);
1284
1285         /* Set EPOCH CLOSE flag if not set by client. */
1286         info->mti_epoch->flags |= MF_EPOCH_CLOSE;
1287         info->mti_attr.ma_valid = 0;
1288         rc = mdt_mfd_close(info, mfd);
1289         mdt_empty_transno(info);
1290         RETURN(rc);
1291 }
1292