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, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/mdt/mdt_open.c
  37  *
  38  * Lustre Metadata Target (mdt) open/close file handling
  39  *
  40  * Author: Huang Hua <huanghua@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_MDS
  44
  45 #include <lustre_acl.h>
  46 #include <lustre_mds.h>
  47 #include "mdt_internal.h"
  48 #include <lustre_nodemap.h>
  49
  50 /* we do nothing because we do not have refcount now */
  51 static void mdt_mfd_get(void *mfdp)
  52 {
  53 }
  54
  55 static struct portals_handle_ops mfd_handle_ops = {
  56         .hop_addref = mdt_mfd_get,
  57         .hop_free   = NULL,
  58 };
  59
  60 /* Create a new mdt_file_data struct, initialize it,
  61  * and insert it to global hash table */
  62 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
  63 {
  64         struct mdt_file_data *mfd;
  65         ENTRY;
  66
  67         OBD_ALLOC_PTR(mfd);
  68         if (mfd != NULL) {
  69                 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  70                 mfd->mfd_handle.h_owner = med;
  71                 INIT_LIST_HEAD(&mfd->mfd_list);
  72                 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
  73         }
  74
  75         RETURN(mfd);
  76 }
  77
  78 /*
  79  * Find the mfd pointed to by handle in global hash table.
  80  * In case of replay the handle is obsoleted
  81  * but mfd can be found in mfd list by that handle.
  82  * Callers need to be holding med_open_lock.
  83  */
  84 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
  85                                      const struct lustre_handle *handle,
  86                                      bool is_replay_or_resent)
  87 {
  88         struct mdt_file_data   *mfd;
  89         ENTRY;
  90
  91         LASSERT(handle != NULL);
  92         mfd = class_handle2object(handle->cookie, med);
  93         /* during dw/setattr replay the mfd can be found by old handle */
  94         if (mfd == NULL && is_replay_or_resent) {
  95                 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  96                         if (mfd->mfd_old_handle.cookie == handle->cookie)
  97                                 RETURN(mfd);
  98                 }
  99                 mfd = NULL;
 100         }
 101
 102         RETURN(mfd);
 103 }
 104
 105 /* free mfd */
 106 void mdt_mfd_free(struct mdt_file_data *mfd)
 107 {
 108         LASSERT(list_empty(&mfd->mfd_list));
 109         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 110 }
 111
 112 static int mdt_create_data(struct mdt_thread_info *info,
 113                            struct mdt_object *p, struct mdt_object *o)
 114 {
 115         struct md_op_spec     *spec = &info->mti_spec;
 116         struct md_attr        *ma   = &info->mti_attr;
 117         int                    rc   = 0;
 118         ENTRY;
 119
 120         if (!md_should_create(spec->sp_cr_flags))
 121                 RETURN(0);
 122
 123         ma->ma_need = MA_INODE | MA_LOV;
 124         ma->ma_valid = 0;
 125         mutex_lock(&o->mot_lov_mutex);
 126         if (!(o->mot_flags & MOF_LOV_CREATED)) {
 127                 rc = mdo_create_data(info->mti_env,
 128                                      p ? mdt_object_child(p) : NULL,
 129                                      mdt_object_child(o), spec, ma);
 130                 if (rc == 0)
 131                         rc = mdt_attr_get_complex(info, o, ma);
 132
 133                 if (rc == 0 && ma->ma_valid & MA_LOV)
 134                         o->mot_flags |= MOF_LOV_CREATED;
 135         }
 136
 137         mutex_unlock(&o->mot_lov_mutex);
 138         RETURN(rc);
 139 }
 140
 141 int mdt_write_read(struct mdt_object *o)
 142 {
 143         int rc = 0;
 144         ENTRY;
 145         spin_lock(&o->mot_write_lock);
 146         rc = o->mot_write_count;
 147         spin_unlock(&o->mot_write_lock);
 148         RETURN(rc);
 149 }
 150
 151 int mdt_write_get(struct mdt_object *o)
 152 {
 153         int rc = 0;
 154         ENTRY;
 155         spin_lock(&o->mot_write_lock);
 156         if (o->mot_write_count < 0)
 157                 rc = -ETXTBSY;
 158         else
 159                 o->mot_write_count++;
 160         spin_unlock(&o->mot_write_lock);
 161
 162         RETURN(rc);
 163 }
 164
 165 void mdt_write_put(struct mdt_object *o)
 166 {
 167         ENTRY;
 168         spin_lock(&o->mot_write_lock);
 169         o->mot_write_count--;
 170         spin_unlock(&o->mot_write_lock);
 171         EXIT;
 172 }
 173
 174 static int mdt_write_deny(struct mdt_object *o)
 175 {
 176         int rc = 0;
 177         ENTRY;
 178         spin_lock(&o->mot_write_lock);
 179         if (o->mot_write_count > 0)
 180                 rc = -ETXTBSY;
 181         else
 182                 o->mot_write_count--;
 183         spin_unlock(&o->mot_write_lock);
 184         RETURN(rc);
 185 }
 186
 187 static void mdt_write_allow(struct mdt_object *o)
 188 {
 189         ENTRY;
 190         spin_lock(&o->mot_write_lock);
 191         o->mot_write_count++;
 192         spin_unlock(&o->mot_write_lock);
 193         EXIT;
 194 }
 195
 196 /* there can be no real transaction so prepare the fake one */
 197 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 198 {
 199         struct mdt_device      *mdt = info->mti_mdt;
 200         struct ptlrpc_request  *req = mdt_info_req(info);
 201         struct tg_export_data  *ted;
 202         struct lsd_client_data *lcd;
 203
 204         ENTRY;
 205         /* transaction has occurred already */
 206         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 207                 RETURN_EXIT;
 208
 209         if (tgt_is_multimodrpcs_client(req->rq_export)) {
 210                 struct thandle         *th;
 211
 212                 /* generate an empty transaction to get a transno
 213                  * and reply data */
 214                 th = dt_trans_create(info->mti_env, mdt->mdt_bottom);
 215                 if (!IS_ERR(th)) {
 216                         rc = dt_trans_start(info->mti_env, mdt->mdt_bottom, th);
 217                         dt_trans_stop(info->mti_env, mdt->mdt_bottom, th);
 218                 }
 219                 RETURN_EXIT;
 220         }
 221
 222         spin_lock(&mdt->mdt_lut.lut_translock);
 223         if (rc != 0) {
 224                 if (info->mti_transno != 0) {
 225                         struct obd_export *exp = req->rq_export;
 226
 227                         CERROR("%s: replay trans "LPU64" NID %s: rc = %d\n",
 228                                mdt_obd_name(mdt), info->mti_transno,
 229                                libcfs_nid2str(exp->exp_connection->c_peer.nid),
 230                                rc);
 231                         spin_unlock(&mdt->mdt_lut.lut_translock);
 232                         RETURN_EXIT;
 233                 }
 234         } else if (info->mti_transno == 0) {
 235                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
 236         } else {
 237                 /* should be replay */
 238                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 239                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 240         }
 241         spin_unlock(&mdt->mdt_lut.lut_translock);
 242
 243         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 244                info->mti_transno,
 245                req->rq_export->exp_obd->obd_last_committed);
 246
 247         req->rq_transno = info->mti_transno;
 248         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 249
 250         /* update lcd in memory only for resent cases */
 251         ted = &req->rq_export->exp_target_data;
 252         LASSERT(ted);
 253         mutex_lock(&ted->ted_lcd_lock);
 254         lcd = ted->ted_lcd;
 255         if (info->mti_transno < lcd->lcd_last_transno &&
 256             info->mti_transno != 0) {
 257                 /* This should happen during replay. Do not update
 258                  * last rcvd info if replay req transno < last transno,
 259                  * otherwise the following resend(after replay) can not
 260                  * be checked correctly by xid */
 261                 mutex_unlock(&ted->ted_lcd_lock);
 262                 CDEBUG(D_HA, "%s: transno = "LPU64" < last_transno = "LPU64"\n",
 263                        mdt_obd_name(mdt), info->mti_transno,
 264                        lcd->lcd_last_transno);
 265                 RETURN_EXIT;
 266         }
 267
 268         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
 269                 if (info->mti_transno != 0)
 270                         lcd->lcd_last_close_transno = info->mti_transno;
 271                 lcd->lcd_last_close_xid = req->rq_xid;
 272                 lcd->lcd_last_close_result = rc;
 273         } else {
 274                 /* VBR: save versions in last_rcvd for reconstruct. */
 275                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 276                 if (pre_versions) {
 277                         lcd->lcd_pre_versions[0] = pre_versions[0];
 278                         lcd->lcd_pre_versions[1] = pre_versions[1];
 279                         lcd->lcd_pre_versions[2] = pre_versions[2];
 280                         lcd->lcd_pre_versions[3] = pre_versions[3];
 281                 }
 282                 if (info->mti_transno != 0)
 283                         lcd->lcd_last_transno = info->mti_transno;
 284
 285                 lcd->lcd_last_xid = req->rq_xid;
 286                 lcd->lcd_last_result = rc;
 287                 lcd->lcd_last_data = info->mti_opdata;
 288         }
 289         mutex_unlock(&ted->ted_lcd_lock);
 290
 291         EXIT;
 292 }
 293
 294 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
 295 {
 296         LASSERT(mfd != NULL);
 297
 298         CDEBUG(D_HA, DFID " Change mfd mode "LPO64" -> "LPO64".\n",
 299                PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
 300
 301         mfd->mfd_mode = mode;
 302 }
 303
 304 /**
 305  * prep ma_lmm/ma_lmv for md_attr from reply
 306  */
 307 static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
 308                                      struct mdt_object *obj,
 309                                      struct md_attr *ma)
 310 {
 311         LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
 312         if (S_ISDIR(obj->mot_header.loh_attr)) {
 313                 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
 314                                                     &RMF_MDT_MD);
 315                 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
 316                                                        &RMF_MDT_MD,
 317                                                        RCL_SERVER);
 318                 if (ma->ma_lmv_size > 0)
 319                         ma->ma_need |= MA_LMV;
 320         } else {
 321                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
 322                                                     &RMF_MDT_MD);
 323                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 324                                                        &RMF_MDT_MD,
 325                                                        RCL_SERVER);
 326                 if (ma->ma_lmm_size > 0)
 327                         ma->ma_need |= MA_LOV;
 328         }
 329 }
 330
 331 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 332                         struct mdt_object *o, __u64 flags, int created,
 333                         struct ldlm_reply *rep)
 334 {
 335         struct ptlrpc_request   *req = mdt_info_req(info);
 336         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 337         struct mdt_file_data    *mfd;
 338         struct md_attr          *ma  = &info->mti_attr;
 339         struct lu_attr          *la  = &ma->ma_attr;
 340         struct mdt_body         *repbody;
 341         int                      rc = 0, isdir, isreg;
 342         ENTRY;
 343
 344         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 345
 346         isreg = S_ISREG(la->la_mode);
 347         isdir = S_ISDIR(la->la_mode);
 348         if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
 349                 /*
 350                  * No EA, check whether it is will set regEA and dirEA since in
 351                  * above attr get, these size might be zero, so reset it, to
 352                  * retrieve the MD after create obj.
 353                  */
 354                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 355                                                        &RMF_MDT_MD,
 356                                                        RCL_SERVER);
 357                 /* in replay case, p == NULL */
 358                 rc = mdt_create_data(info, p, o);
 359                 if (rc)
 360                         RETURN(rc);
 361
 362                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
 363                         mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
 364         }
 365
 366         CDEBUG(D_INODE, "after open, ma_valid bit = "LPX64" lmm_size = %d\n",
 367                ma->ma_valid, ma->ma_lmm_size);
 368
 369         if (ma->ma_valid & MA_LOV) {
 370                 LASSERT(ma->ma_lmm_size != 0);
 371                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 372                 if (isdir)
 373                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 374                 else
 375                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 376         }
 377
 378         if (ma->ma_valid & MA_LMV) {
 379                 LASSERT(ma->ma_lmv_size != 0);
 380                 repbody->mbo_eadatasize = ma->ma_lmv_size;
 381                 LASSERT(isdir);
 382                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
 383         }
 384
 385         if (flags & FMODE_WRITE)
 386                 rc = mdt_write_get(o);
 387         else if (flags & MDS_FMODE_EXEC)
 388                 rc = mdt_write_deny(o);
 389
 390         if (rc)
 391                 RETURN(rc);
 392
 393         rc = mo_open(info->mti_env, mdt_object_child(o),
 394                      created ? flags | MDS_OPEN_CREATED : flags);
 395         if (rc != 0) {
 396                 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
 397                  * client does not know which suppgid should be sent to the MDS,
 398                  * or some other(s) changed the target file's GID after this RPC
 399                  * sent to the MDS with the suppgid as the original GID, then we
 400                  * give the client another chance to send the right suppgid. */
 401                 if (rc == -EACCES &&
 402                     allow_client_chgrp(info, lu_ucred(info->mti_env)))
 403                         mdt_set_disposition(info, rep, DISP_OPEN_DENY);
 404
 405                 GOTO(err_out, rc);
 406         }
 407
 408         mfd = mdt_mfd_new(med);
 409         if (mfd == NULL)
 410                 GOTO(err_out, rc = -ENOMEM);
 411
 412         /*
 413          * Keep a reference on this object for this open, and is
 414          * released by mdt_mfd_close().
 415          */
 416         mdt_object_get(info->mti_env, o);
 417         mfd->mfd_object = o;
 418         mfd->mfd_xid = req->rq_xid;
 419
 420         /*
 421          * @flags is always not zero. At least it should be FMODE_READ,
 422          * FMODE_WRITE or MDS_FMODE_EXEC.
 423          */
 424         LASSERT(flags != 0);
 425
 426         /* Open handling. */
 427         mdt_mfd_set_mode(mfd, flags);
 428
 429         atomic_inc(&o->mot_open_count);
 430         if (flags & MDS_OPEN_LEASE)
 431                 atomic_inc(&o->mot_lease_count);
 432
 433         /* replay handle */
 434         if (req_is_replay(req)) {
 435                 struct mdt_file_data *old_mfd;
 436                 /* Check wheather old cookie already exist in
 437                  * the list, becasue when do recovery, client
 438                  * might be disconnected from server, and
 439                  * restart replay, so there maybe some orphan
 440                  * mfd here, we should remove them */
 441                 LASSERT(info->mti_rr.rr_handle != NULL);
 442                 spin_lock(&med->med_open_lock);
 443                 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
 444                 if (old_mfd != NULL) {
 445                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 446                                "cookie = "LPX64"\n", mfd,
 447                                PFID(mdt_object_fid(mfd->mfd_object)),
 448                                info->mti_rr.rr_handle->cookie);
 449                         class_handle_unhash(&old_mfd->mfd_handle);
 450                         list_del_init(&old_mfd->mfd_list);
 451                         spin_unlock(&med->med_open_lock);
 452                         /* no attr update for that close */
 453                         la->la_valid = 0;
 454                         ma->ma_valid |= MA_FLAGS;
 455                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 456                         mdt_mfd_close(info, old_mfd);
 457                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 458                         ma->ma_valid &= ~MA_FLAGS;
 459                 } else {
 460                         spin_unlock(&med->med_open_lock);
 461                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 462                                "cookie = "LPX64"\n",
 463                                PFID(mdt_object_fid(mfd->mfd_object)),
 464                                info->mti_rr.rr_handle->cookie);
 465                 }
 466
 467                 CDEBUG(D_HA, "Store old cookie "LPX64" in new mfd\n",
 468                        info->mti_rr.rr_handle->cookie);
 469
 470                 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
 471         }
 472
 473         repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
 474
 475         if (req->rq_export->exp_disconnected) {
 476                 spin_lock(&med->med_open_lock);
 477                 class_handle_unhash(&mfd->mfd_handle);
 478                 list_del_init(&mfd->mfd_list);
 479                 spin_unlock(&med->med_open_lock);
 480                 mdt_mfd_close(info, mfd);
 481         } else {
 482                 spin_lock(&med->med_open_lock);
 483                 list_add(&mfd->mfd_list, &med->med_open_head);
 484                 spin_unlock(&med->med_open_lock);
 485         }
 486
 487         mdt_empty_transno(info, rc);
 488
 489         RETURN(rc);
 490
 491 err_out:
 492         if (flags & FMODE_WRITE)
 493                 mdt_write_put(o);
 494         else if (flags & MDS_FMODE_EXEC)
 495                 mdt_write_allow(o);
 496
 497         return rc;
 498 }
 499
 500 static int mdt_finish_open(struct mdt_thread_info *info,
 501                            struct mdt_object *p, struct mdt_object *o,
 502                            __u64 flags, int created, struct ldlm_reply *rep)
 503 {
 504         struct ptlrpc_request   *req = mdt_info_req(info);
 505         struct obd_export       *exp = req->rq_export;
 506         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 507         struct md_attr          *ma  = &info->mti_attr;
 508         struct lu_attr          *la  = &ma->ma_attr;
 509         struct mdt_file_data    *mfd;
 510         struct mdt_body         *repbody;
 511         int                      rc = 0;
 512         int                      isreg, isdir, islnk;
 513         struct list_head        *t;
 514         ENTRY;
 515
 516         LASSERT(ma->ma_valid & MA_INODE);
 517
 518         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 519
 520         isreg = S_ISREG(la->la_mode);
 521         isdir = S_ISDIR(la->la_mode);
 522         islnk = S_ISLNK(la->la_mode);
 523         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 524
 525         /* LU-2275, simulate broken behaviour (esp. prevalent in
 526          * pre-2.4 servers where a very strange reply is sent on error
 527          * that looks like it was actually almost successful and a
 528          * failure at the same time.) */
 529         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 530                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 531                                                DISP_LOOKUP_NEG |
 532                                                DISP_LOOKUP_POS);
 533
 534                 if (flags & MDS_OPEN_LOCK)
 535                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 536
 537                 RETURN(-ENOENT);
 538         }
 539
 540         if (exp_connect_rmtclient(exp)) {
 541                 void *buf = req_capsule_server_get(info->mti_pill, &RMF_ACL);
 542
 543                 rc = mdt_pack_remote_perm(info, o, buf);
 544                 if (rc) {
 545                         repbody->mbo_valid &= ~OBD_MD_FLRMTPERM;
 546                         repbody->mbo_aclsize = 0;
 547                 } else {
 548                         repbody->mbo_valid |= OBD_MD_FLRMTPERM;
 549                         repbody->mbo_aclsize = sizeof(struct mdt_remote_perm);
 550                 }
 551         }
 552 #ifdef CONFIG_FS_POSIX_ACL
 553         else if (exp_connect_flags(exp) & OBD_CONNECT_ACL)
 554                 rc = mdt_pack_acl2body(info, repbody, o,
 555                                        exp->exp_target_data.ted_nodemap);
 556 #endif
 557
 558         /*
 559          * If we are following a symlink, don't open; and do not return open
 560          * handle for special nodes as client required.
 561          */
 562         if (islnk || (!isreg && !isdir &&
 563             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 564                 lustre_msg_set_transno(req->rq_repmsg, 0);
 565                 RETURN(0);
 566         }
 567
 568         /*
 569          * We need to return the existing object's fid back, so it is done here,
 570          * after preparing the reply.
 571          */
 572         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 573                 RETURN(-EEXIST);
 574
 575         /* This can't be done earlier, we need to return reply body */
 576         if (isdir) {
 577                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 578                         /* We are trying to create or write an existing dir. */
 579                         RETURN(-EISDIR);
 580                 }
 581         } else if (flags & MDS_OPEN_DIRECTORY)
 582                 RETURN(-ENOTDIR);
 583
 584         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 585                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 586                 RETURN(-EAGAIN);
 587         }
 588
 589         mfd = NULL;
 590         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 591                 spin_lock(&med->med_open_lock);
 592                 list_for_each(t, &med->med_open_head) {
 593                         mfd = list_entry(t, struct mdt_file_data, mfd_list);
 594                         if (mfd->mfd_xid == req->rq_xid)
 595                                 break;
 596                         mfd = NULL;
 597                 }
 598                 spin_unlock(&med->med_open_lock);
 599
 600                 if (mfd != NULL) {
 601                         repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
 602                         /* set repbody->ea_size for resent case */
 603                         if (ma->ma_valid & MA_LOV) {
 604                                 LASSERT(ma->ma_lmm_size != 0);
 605                                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 606                                 if (isdir)
 607                                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 608                                 else
 609                                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 610                         }
 611                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 612                         RETURN(0);
 613                 }
 614         }
 615
 616         rc = mdt_mfd_open(info, p, o, flags, created, rep);
 617         if (!rc)
 618                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 619
 620         RETURN(rc);
 621 }
 622
 623 void mdt_reconstruct_open(struct mdt_thread_info *info,
 624                           struct mdt_lock_handle *lhc)
 625 {
 626         const struct lu_env *env = info->mti_env;
 627         struct mdt_device       *mdt  = info->mti_mdt;
 628         struct req_capsule      *pill = info->mti_pill;
 629         struct ptlrpc_request   *req  = mdt_info_req(info);
 630         struct md_attr          *ma   = &info->mti_attr;
 631         struct mdt_reint_record *rr   = &info->mti_rr;
 632         __u64                   flags = info->mti_spec.sp_cr_flags;
 633         struct ldlm_reply       *ldlm_rep;
 634         struct mdt_object       *parent;
 635         struct mdt_object       *child;
 636         struct mdt_body         *repbody;
 637         int                      rc;
 638         __u64                    opdata;
 639         ENTRY;
 640
 641         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 642         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 643         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 644
 645         ma->ma_need = MA_INODE | MA_HSM;
 646         ma->ma_valid = 0;
 647
 648         opdata = mdt_req_from_lrd(req, info->mti_reply_data);
 649         mdt_set_disposition(info, ldlm_rep, opdata);
 650
 651         CDEBUG(D_INODE, "This is reconstruct open: disp="LPX64", result=%d\n",
 652                ldlm_rep->lock_policy_res1, req->rq_status);
 653
 654         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
 655             req->rq_status != 0)
 656                 /* We did not create successfully, return error to client. */
 657                 GOTO(out, rc = req->rq_status);
 658
 659         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
 660                 struct obd_export *exp = req->rq_export;
 661                 /*
 662                  * We failed after creation, but we do not know in which step
 663                  * we failed. So try to check the child object.
 664                  */
 665                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
 666                 if (IS_ERR(parent)) {
 667                         rc = PTR_ERR(parent);
 668                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
 669                                       " Evicting client %s with export %s.\n",
 670                                       PFID(rr->rr_fid1), rc,
 671                                       obd_uuid2str(&exp->exp_client_uuid),
 672                                       obd_export_nid2str(exp));
 673                         mdt_export_evict(exp);
 674                         RETURN_EXIT;
 675                 }
 676
 677                 child = mdt_object_find(env, mdt, rr->rr_fid2);
 678                 if (IS_ERR(child)) {
 679                         rc = PTR_ERR(child);
 680                         LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
 681                                       "evicting client %s with export %s\n",
 682                                       PFID(rr->rr_fid2), rc,
 683                                       obd_uuid2str(&exp->exp_client_uuid),
 684                                       obd_export_nid2str(exp));
 685                         mdt_object_put(env, parent);
 686                         mdt_export_evict(exp);
 687                         RETURN_EXIT;
 688                 }
 689
 690                 if (unlikely(mdt_object_remote(child))) {
 691                         /* the child object was created on remote server */
 692                         if (!mdt_is_dne_client(exp)) {
 693                                 /* Return -EIO for old client */
 694                                 mdt_object_put(env, parent);
 695                                 mdt_object_put(env, child);
 696                                 GOTO(out, rc = -EIO);
 697                         }
 698                         repbody->mbo_fid1 = *rr->rr_fid2;
 699                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 700                         rc = 0;
 701                 } else {
 702                         if (mdt_object_exists(child)) {
 703                                 mdt_prep_ma_buf_from_rep(info, child, ma);
 704                                 rc = mdt_attr_get_complex(info, child, ma);
 705                                 if (rc == 0)
 706                                         rc = mdt_finish_open(info, parent,
 707                                                              child, flags,
 708                                                              1, ldlm_rep);
 709                         } else {
 710                                 /* the child does not exist, we should do
 711                                  * regular open */
 712                                 mdt_object_put(env, parent);
 713                                 mdt_object_put(env, child);
 714                                 GOTO(regular_open, 0);
 715                         }
 716                 }
 717                 mdt_object_put(env, parent);
 718                 mdt_object_put(env, child);
 719                 GOTO(out, rc);
 720         } else {
 721 regular_open:
 722                 /* We did not try to create, so we are a pure open */
 723                 rc = mdt_reint_open(info, lhc);
 724         }
 725
 726         EXIT;
 727 out:
 728         req->rq_status = rc;
 729         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 730         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
 731 }
 732
 733 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
 734 {
 735         __u64                    flags = info->mti_spec.sp_cr_flags;
 736         struct mdt_reint_record *rr = &info->mti_rr;
 737         struct md_attr          *ma = &info->mti_attr;
 738         struct mdt_object       *o;
 739         int                      rc;
 740         ENTRY;
 741
 742         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
 743         if (IS_ERR(o))
 744                 RETURN(rc = PTR_ERR(o));
 745
 746         if (unlikely(mdt_object_remote(o))) {
 747                 /* the child object was created on remote server */
 748                 struct mdt_body *repbody;
 749
 750                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 751                                                 DISP_LOOKUP_EXECD |
 752                                                 DISP_LOOKUP_POS));
 753                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 754                 repbody->mbo_fid1 = *rr->rr_fid2;
 755                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 756                 rc = 0;
 757         } else {
 758                 if (mdt_object_exists(o)) {
 759                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 760                                                         DISP_LOOKUP_EXECD |
 761                                                         DISP_LOOKUP_POS));
 762                         mdt_prep_ma_buf_from_rep(info, o, ma);
 763                         rc = mdt_attr_get_complex(info, o, ma);
 764                         if (rc == 0)
 765                                 rc = mdt_finish_open(info, NULL, o, flags, 0,
 766                                                      rep);
 767                 } else {
 768                         rc = -ENOENT;
 769                 }
 770         }
 771
 772         mdt_object_put(info->mti_env, o);
 773         RETURN(rc);
 774 }
 775
 776 /* lock object for open */
 777 static int mdt_object_open_lock(struct mdt_thread_info *info,
 778                                 struct mdt_object *obj,
 779                                 struct mdt_lock_handle *lhc,
 780                                 __u64 *ibits)
 781 {
 782         struct md_attr *ma = &info->mti_attr;
 783         __u64 open_flags = info->mti_spec.sp_cr_flags;
 784         enum ldlm_mode lm = LCK_CR;
 785         bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
 786         bool try_layout = false;
 787         bool create_layout = false;
 788         int rc = 0;
 789         ENTRY;
 790
 791         *ibits = 0;
 792         mdt_lock_handle_init(lhc);
 793
 794         if (req_is_replay(mdt_info_req(info)))
 795                 RETURN(0);
 796
 797         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
 798                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
 799                     md_should_create(open_flags))
 800                         create_layout = true;
 801                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
 802                     ma->ma_need & MA_LOV)
 803                         try_layout = true;
 804         }
 805
 806         if (acq_lease) {
 807                 /* lease open, acquire write mode of open sem */
 808                 down_write(&obj->mot_open_sem);
 809
 810                 /* Lease exists and ask for new lease */
 811                 if (atomic_read(&obj->mot_lease_count) > 0) {
 812                         /* only exclusive open is supported, so lease
 813                          * are conflicted to each other */
 814                         GOTO(out, rc = -EBUSY);
 815                 }
 816
 817                 /* Lease must be with open lock */
 818                 if (!(open_flags & MDS_OPEN_LOCK)) {
 819                         CERROR("Request lease for file:"DFID ", but open lock "
 820                                 "is missed, open_flags = "LPO64".\n",
 821                                 PFID(mdt_object_fid(obj)), open_flags);
 822                         GOTO(out, rc = -EPROTO);
 823                 }
 824
 825                 /* XXX: only exclusive open is supported. */
 826                 lm = LCK_EX;
 827                 *ibits = MDS_INODELOCK_OPEN;
 828
 829                 /* never grant LCK_EX layout lock to client */
 830                 try_layout = false;
 831         } else { /* normal open */
 832                 /* normal open holds read mode of open sem */
 833                 down_read(&obj->mot_open_sem);
 834
 835                 if (open_flags & MDS_OPEN_LOCK) {
 836                         if (open_flags & FMODE_WRITE)
 837                                 lm = LCK_CW;
 838                         else if (open_flags & MDS_FMODE_EXEC)
 839                                 lm = LCK_PR;
 840                         else
 841                                 lm = LCK_CR;
 842
 843                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
 844                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
 845                         if (open_flags & FMODE_WRITE)
 846                                 lm = LCK_CW;
 847                         else
 848                                 lm = LCK_CR;
 849
 850                         /* revoke lease */
 851                         *ibits = MDS_INODELOCK_OPEN;
 852                         try_layout = false;
 853
 854                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 855                 }
 856                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
 857                         PFID(mdt_object_fid(obj)),
 858                         atomic_read(&obj->mot_open_count), lm);
 859         }
 860
 861         mdt_lock_reg_init(lhc, lm);
 862
 863         /* one problem to return layout lock on open is that it may result
 864          * in too many layout locks cached on the client side. */
 865         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
 866                 /* return lookup lock to validate inode at the client side,
 867                  * this is pretty important otherwise mdt will return layout
 868                  * lock for each open.
 869                  * However this is a double-edged sword because changing
 870                  * permission will revoke huge # of LOOKUP locks. */
 871                 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
 872                 if (!mdt_object_lock_try(info, obj, lhc, *ibits)) {
 873                         *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
 874                         if (*ibits != 0)
 875                                 rc = mdt_object_lock(info, obj, lhc, *ibits);
 876                 }
 877         } else if (*ibits != 0) {
 878                 rc = mdt_object_lock(info, obj, lhc, *ibits);
 879         }
 880
 881         CDEBUG(D_INODE, "Requested bits lock:"DFID ", ibits = "LPX64
 882                 ", open_flags = "LPO64", try_layout = %d, rc = %d\n",
 883                 PFID(mdt_object_fid(obj)), *ibits, open_flags, try_layout, rc);
 884
 885         /* will change layout, revoke layout locks by enqueuing EX lock. */
 886         if (rc == 0 && create_layout) {
 887                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
 888
 889                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
 890                         ", open_flags = "LPO64"\n",
 891                         PFID(mdt_object_fid(obj)), open_flags);
 892
 893                 /* We cannot enqueue another lock for the same resource we
 894                  * already have a lock for, due to mechanics of waiting list
 895                  * iterating in ldlm, see LU-3601.
 896                  * As such we'll drop the open lock we just got above here,
 897                  * it's ok not to have this open lock as it's main purpose is to
 898                  * flush unused cached client open handles. */
 899                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 900                         mdt_object_unlock(info, obj, lhc, 1);
 901
 902                 LASSERT(!try_layout);
 903                 mdt_lock_handle_init(ll);
 904                 mdt_lock_reg_init(ll, LCK_EX);
 905                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
 906
 907                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
 908         }
 909
 910         /* Check if there is any other open handles after acquiring
 911          * open lock. At this point, caching open handles have been revoked
 912          * by open lock.
 913          * XXX: Now only exclusive open is supported. Need to check the
 914          * type of open for generic lease support. */
 915         if (rc == 0 && acq_lease) {
 916                 struct ptlrpc_request *req = mdt_info_req(info);
 917                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 918                 struct mdt_file_data *mfd;
 919                 bool is_replay_or_resent;
 920                 int open_count = 0;
 921
 922                 /* For lease: application can open a file and then apply lease,
 923                  * @handle contains original open handle in that case.
 924                  * In recovery, open REQ will be replayed and the lease REQ may
 925                  * be resent that means the open handle is already stale, so we
 926                  * need to fix it up here by finding new handle. */
 927                 is_replay_or_resent = req_is_replay(req) ||
 928                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
 929
 930                 /* if the request is _not_ a replay request, rr_handle
 931                  * may be used to hold an openhandle which is issuing the
 932                  * lease request, so that this openhandle doesn't count. */
 933                 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
 934                                      is_replay_or_resent);
 935                 if (mfd != NULL)
 936                         ++open_count;
 937
 938                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
 939                         PFID(mdt_object_fid(obj)),
 940                         atomic_read(&obj->mot_open_count), open_count);
 941
 942                 if (atomic_read(&obj->mot_open_count) > open_count)
 943                         GOTO(out, rc = -EBUSY);
 944         }
 945         GOTO(out, rc);
 946
 947 out:
 948         RETURN(rc);
 949 }
 950
 951 static void mdt_object_open_unlock(struct mdt_thread_info *info,
 952                                    struct mdt_object *obj,
 953                                    struct mdt_lock_handle *lhc,
 954                                    __u64 ibits, int rc)
 955 {
 956         __u64 open_flags = info->mti_spec.sp_cr_flags;
 957         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
 958         ENTRY;
 959
 960         if (req_is_replay(mdt_info_req(info)))
 961                 RETURN_EXIT;
 962
 963         /* Release local lock - the lock put in MDT_LH_LOCAL will never
 964          * return to client side. */
 965         if (lustre_handle_is_used(&ll->mlh_reg_lh))
 966                 mdt_object_unlock(info, obj, ll, 1);
 967
 968         ll = &info->mti_lh[MDT_LH_LAYOUT];
 969         /* Release local layout lock, layout was created */
 970         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
 971                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
 972                 mdt_object_unlock(info, obj, ll, 1);
 973         }
 974
 975         if (open_flags & MDS_OPEN_LEASE)
 976                 up_write(&obj->mot_open_sem);
 977         else
 978                 up_read(&obj->mot_open_sem);
 979
 980         /* Cross-ref case, the lock should be returned to the client */
 981         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
 982                 RETURN_EXIT;
 983
 984         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
 985                 /* for the open request, the lock will only return to client
 986                  * if open or layout lock is granted. */
 987                 rc = 1;
 988         }
 989
 990         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
 991                 struct ldlm_reply       *ldlm_rep;
 992
 993                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
 994                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
 995                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 996                         mdt_object_unlock(info, obj, lhc, 1);
 997         }
 998         RETURN_EXIT;
 999 }
1000
1001 /**
1002  * Check release is permitted for the current HSM flags.
1003  */
1004 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1005 {
1006         if (!(ma->ma_valid & MA_HSM))
1007                 return false;
1008
1009         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1010                 return false;
1011
1012         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1013                 return false;
1014
1015         return true;
1016 }
1017
1018 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1019                                 struct ldlm_reply *rep,
1020                                 struct mdt_lock_handle *lhc)
1021 {
1022         const struct lu_env     *env   = info->mti_env;
1023         struct mdt_device       *mdt   = info->mti_mdt;
1024         __u64                    flags = info->mti_spec.sp_cr_flags;
1025         struct mdt_reint_record *rr    = &info->mti_rr;
1026         struct md_attr          *ma    = &info->mti_attr;
1027         struct mdt_object       *parent= NULL;
1028         struct mdt_object       *o;
1029         int                      rc;
1030         bool                     object_locked = false;
1031         __u64                    ibits = 0;
1032         ENTRY;
1033
1034         if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1035                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1036                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1037                         if (IS_ERR(parent)) {
1038                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1039                                        " for anonymous created %ld, try to"
1040                                        " use server-side parent.\n",
1041                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1042                                 parent = NULL;
1043                         }
1044                 }
1045                 if (parent == NULL)
1046                         ma->ma_need |= MA_PFID;
1047         }
1048
1049         o = mdt_object_find(env, mdt, rr->rr_fid2);
1050         if (IS_ERR(o))
1051                 GOTO(out_parent_put, rc = PTR_ERR(o));
1052
1053         if (mdt_object_remote(o)) {
1054                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1055                        mdt_obd_name(info->mti_mdt),
1056                        PFID(rr->rr_fid2));
1057                 GOTO(out, rc = -EREMOTE);
1058         } else if (!mdt_object_exists(o)) {
1059                 mdt_set_disposition(info, rep,
1060                                     DISP_IT_EXECD |
1061                                     DISP_LOOKUP_EXECD |
1062                                     DISP_LOOKUP_NEG);
1063                 GOTO(out, rc = -ENOENT);
1064         }
1065
1066         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1067
1068         mdt_prep_ma_buf_from_rep(info, o, ma);
1069         if (flags & MDS_OPEN_RELEASE)
1070                 ma->ma_need |= MA_HSM;
1071         rc = mdt_attr_get_complex(info, o, ma);
1072         if (rc)
1073                 GOTO(out, rc);
1074
1075         /* We should not change file's existing LOV EA */
1076         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1077             flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1078                 GOTO(out, rc = -EEXIST);
1079
1080         /* If a release request, check file flags are fine and ask for an
1081          * exclusive open access. */
1082         if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1083                 GOTO(out, rc = -EPERM);
1084
1085         rc = mdt_check_resent_lock(info, o, lhc);
1086         if (rc < 0) {
1087                 GOTO(out, rc);
1088         } else if (rc > 0) {
1089                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1090                 object_locked = true;
1091                 if (rc)
1092                         GOTO(out_unlock, rc);
1093         }
1094
1095         if (ma->ma_valid & MA_PFID) {
1096                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1097                 if (IS_ERR(parent)) {
1098                         CDEBUG(D_INODE, "Fail to find parent "DFID
1099                                " for anonymous created %ld, try to"
1100                                " use system default.\n",
1101                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1102                         parent = NULL;
1103                 }
1104         }
1105
1106         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1107         if (!rc) {
1108                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1109                 if (flags & MDS_OPEN_LOCK)
1110                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1111                 if (flags & MDS_OPEN_LEASE)
1112                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1113         }
1114         GOTO(out_unlock, rc);
1115
1116 out_unlock:
1117         if (object_locked)
1118                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1119 out:
1120         mdt_object_put(env, o);
1121 out_parent_put:
1122         if (parent != NULL)
1123                 mdt_object_put(env, parent);
1124         return rc;
1125 }
1126
1127 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1128 static int mdt_cross_open(struct mdt_thread_info *info,
1129                           const struct lu_fid *parent_fid,
1130                           const struct lu_fid *fid,
1131                           struct ldlm_reply *rep, __u32 flags)
1132 {
1133         struct md_attr    *ma = &info->mti_attr;
1134         struct mdt_object *o;
1135         int                rc;
1136         ENTRY;
1137
1138         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1139         if (IS_ERR(o))
1140                 RETURN(rc = PTR_ERR(o));
1141
1142         if (mdt_object_remote(o)) {
1143                 /* Something is wrong here, the object is on another MDS! */
1144                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1145                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1146                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1147                                 &o->mot_obj,
1148                                 "Object isn't on this server! FLD error?");
1149                 rc = -EFAULT;
1150         } else {
1151                 if (mdt_object_exists(o)) {
1152                         /* Do permission check for cross-open. */
1153                         rc = mo_permission(info->mti_env, NULL,
1154                                            mdt_object_child(o),
1155                                            NULL, flags | MDS_OPEN_CROSS);
1156                         if (rc)
1157                                 goto out;
1158
1159                         mdt_prep_ma_buf_from_rep(info, o, ma);
1160                         rc = mdt_attr_get_complex(info, o, ma);
1161                         if (rc != 0)
1162                                 GOTO(out, rc);
1163
1164                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1165                 } else {
1166                         /*
1167                          * Something is wrong here. lookup was positive but
1168                          * there is no object!
1169                          */
1170                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1171                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1172                         rc = -EFAULT;
1173                 }
1174         }
1175 out:
1176         mdt_object_put(info->mti_env, o);
1177         RETURN(rc);
1178 }
1179
1180 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1181 {
1182         struct mdt_device       *mdt = info->mti_mdt;
1183         struct ptlrpc_request   *req = mdt_info_req(info);
1184         struct mdt_object       *parent;
1185         struct mdt_object       *child;
1186         struct mdt_lock_handle  *lh;
1187         struct ldlm_reply       *ldlm_rep;
1188         struct mdt_body         *repbody;
1189         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1190         struct md_attr          *ma = &info->mti_attr;
1191         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1192         __u64                    ibits = 0;
1193         struct mdt_reint_record *rr = &info->mti_rr;
1194         int                      result, rc;
1195         int                      created = 0;
1196         int                      object_locked = 0;
1197         __u32                    msg_flags;
1198         ENTRY;
1199
1200         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1201                                (obd_timeout + 1) / 4);
1202
1203         mdt_counter_incr(req, LPROC_MDT_OPEN);
1204         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1205
1206         ma->ma_need = MA_INODE;
1207         ma->ma_valid = 0;
1208
1209         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1210         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1211
1212         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1213                 CERROR("file join is not supported anymore.\n");
1214                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1215         }
1216         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1217
1218         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1219             info->mti_spec.u.sp_ea.eadata == NULL)
1220                 GOTO(out, result = err_serious(-EINVAL));
1221
1222         if (create_flags & FMODE_WRITE &&
1223             exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1224                 GOTO(out, result = -EROFS);
1225
1226         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1227                "cr_flag="LPO64" mode=0%06o msg_flag=0x%x\n",
1228                PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1229                PFID(rr->rr_fid2), create_flags,
1230                ma->ma_attr.la_mode, msg_flags);
1231
1232         if (info->mti_cross_ref) {
1233                 /* This is cross-ref open */
1234                 mdt_set_disposition(info, ldlm_rep,
1235                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1236                              DISP_LOOKUP_POS));
1237                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1238                                         ldlm_rep, create_flags);
1239                 GOTO(out, result);
1240         } else if (req_is_replay(req) ||
1241             (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1242                 /* This is a replay request or from liblustre with ea. */
1243                 result = mdt_open_by_fid(info, ldlm_rep);
1244
1245                 if (result != -ENOENT) {
1246                         if (req->rq_export->exp_libclient &&
1247                             create_flags & MDS_OPEN_HAS_EA)
1248                                 GOTO(out, result = 0);
1249                         GOTO(out, result);
1250                 }
1251                 /* We didn't find the correct object, so we need to re-create it
1252                  * via a regular replay. */
1253                 if (!(create_flags & MDS_OPEN_CREAT)) {
1254                         DEBUG_REQ(D_ERROR, req,
1255                                   "OPEN & CREAT not in open replay/by_fid.");
1256                         GOTO(out, result = -EFAULT);
1257                 }
1258                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1259         } else if (create_flags & (MDS_OPEN_BY_FID | MDS_OPEN_LOCK)) {
1260                 /*
1261                  * MDS_OPEN_LOCK is checked for backward compatibility with 2.1
1262                  * client.
1263                  */
1264                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1265                 if (result < 0)
1266                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1267                                PFID(rr->rr_fid2), result);
1268                 GOTO(out, result);
1269         }
1270
1271         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1272                 GOTO(out, result = err_serious(-ENOMEM));
1273
1274         mdt_set_disposition(info, ldlm_rep,
1275                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1276
1277         if (!lu_name_is_valid(&rr->rr_name))
1278                 GOTO(out, result = -EPROTO);
1279
1280 again:
1281         lh = &info->mti_lh[MDT_LH_PARENT];
1282         mdt_lock_pdo_init(lh,
1283                           (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1284                           &rr->rr_name);
1285
1286         parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1287         if (IS_ERR(parent))
1288                 GOTO(out, result = PTR_ERR(parent));
1289
1290         result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1291         if (result != 0) {
1292                 mdt_object_put(info->mti_env, parent);
1293                 GOTO(out, result);
1294         }
1295
1296         /* get and check version of parent */
1297         result = mdt_version_get_check(info, parent, 0);
1298         if (result)
1299                 GOTO(out_parent, result);
1300
1301         fid_zero(child_fid);
1302
1303         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1304                             &rr->rr_name, child_fid, &info->mti_spec);
1305
1306         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1307                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1308                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1309                  PFID(child_fid));
1310
1311         if (result != 0 && result != -ENOENT && result != -ESTALE)
1312                 GOTO(out_parent, result);
1313
1314         if (result == -ENOENT || result == -ESTALE) {
1315                 /* If the object is dead, let's check if the object
1316                  * is being migrated to a new object */
1317                 if (result == -ESTALE) {
1318                         struct lu_buf lmv_buf;
1319
1320                         lmv_buf.lb_buf = info->mti_xattr_buf;
1321                         lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
1322                         rc = mo_xattr_get(info->mti_env,
1323                                           mdt_object_child(parent),
1324                                           &lmv_buf, XATTR_NAME_LMV);
1325                         if (rc > 0) {
1326                                 struct lmv_mds_md_v1 *lmv;
1327
1328                                 lmv = lmv_buf.lb_buf;
1329                                 if (le32_to_cpu(lmv->lmv_hash_type) &
1330                                                 LMV_HASH_FLAG_MIGRATION) {
1331                                         /* Get the new parent FID and retry */
1332                                         mdt_object_unlock_put(info, parent,
1333                                                               lh, 1);
1334                                         mdt_lock_handle_init(lh);
1335                                         fid_le_to_cpu(
1336                                                 (struct lu_fid *)rr->rr_fid1,
1337                                                 &lmv->lmv_stripe_fids[1]);
1338                                         goto again;
1339                                 }
1340                         }
1341                 }
1342
1343                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1344                 if (result == -ESTALE) {
1345                         /*
1346                          * -ESTALE means the parent is a dead(unlinked) dir, so
1347                          * it should return -ENOENT to in accordance with the
1348                          * original mds implementaion.
1349                          */
1350                         GOTO(out_parent, result = -ENOENT);
1351                 }
1352
1353                 if (!(create_flags & MDS_OPEN_CREAT))
1354                         GOTO(out_parent, result);
1355                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1356                         GOTO(out_parent, result = -EROFS);
1357                 *child_fid = *info->mti_rr.rr_fid2;
1358                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1359                          PFID(child_fid));
1360                 /* In the function below, .hs_keycmp resolves to
1361                  * lu_obj_hop_keycmp() */
1362                 /* coverity[overrun-buffer-val] */
1363                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1364         } else {
1365                 /*
1366                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1367                  * return FID back in that case.
1368                  */
1369                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1370                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1371         }
1372         if (IS_ERR(child))
1373                 GOTO(out_parent, result = PTR_ERR(child));
1374
1375         /** check version of child  */
1376         rc = mdt_version_get_check(info, child, 1);
1377         if (rc)
1378                 GOTO(out_child, result = rc);
1379
1380         if (result == -ENOENT) {
1381                 /* Create under OBF and .lustre is not permitted */
1382                 if (!fid_is_md_operative(rr->rr_fid1))
1383                         GOTO(out_child, result = -EPERM);
1384
1385                 /* save versions in reply */
1386                 mdt_version_get_save(info, parent, 0);
1387                 mdt_version_get_save(info, child, 1);
1388
1389                 /* version of child will be changed */
1390                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1391
1392                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1393                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1394
1395                 /* Let lower layers know what is lock mode on directory. */
1396                 info->mti_spec.sp_cr_mode =
1397                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1398
1399                 /*
1400                  * Do not perform lookup sanity check. We know that name does
1401                  * not exist.
1402                  */
1403                 info->mti_spec.sp_cr_lookup = 0;
1404                 info->mti_spec.sp_feat = &dt_directory_features;
1405
1406                 result = mdo_create(info->mti_env,
1407                                     mdt_object_child(parent),
1408                                     &rr->rr_name,
1409                                     mdt_object_child(child),
1410                                     &info->mti_spec,
1411                                     &info->mti_attr);
1412                 if (result == -ERESTART) {
1413                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1414                         GOTO(out_child, result);
1415                 } else {
1416                         mdt_prep_ma_buf_from_rep(info, child, ma);
1417                         /* XXX: we should call this once, see few lines below */
1418                         if (result == 0)
1419                                 result = mdt_attr_get_complex(info, child, ma);
1420
1421                         if (result != 0)
1422                                 GOTO(out_child, result);
1423                 }
1424                 created = 1;
1425         } else {
1426                 /*
1427                  * The object is on remote node, return its FID for remote open.
1428                  */
1429                 if (mdt_object_remote(child)) {
1430                         /*
1431                          * Check if this lock already was sent to client and
1432                          * this is resent case. For resent case do not take lock
1433                          * again, use what is already granted.
1434                          */
1435                         LASSERT(lhc != NULL);
1436
1437                         rc = mdt_check_resent_lock(info, child, lhc);
1438                         if (rc < 0) {
1439                                 GOTO(out_child, result = rc);
1440                         } else if (rc > 0) {
1441                                 mdt_lock_handle_init(lhc);
1442                                 mdt_lock_reg_init(lhc, LCK_PR);
1443
1444                                 rc = mdt_object_lock(info, child, lhc,
1445                                                      MDS_INODELOCK_LOOKUP);
1446                         }
1447                         repbody->mbo_fid1 = *mdt_object_fid(child);
1448                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1449                         if (rc != 0)
1450                                 result = rc;
1451                         else
1452                                 result = -MDT_EREMOTE_OPEN;
1453                         GOTO(out_child, result);
1454                 } else if (mdt_object_exists(child)) {
1455                         /* We have to get attr & LOV EA & HSM for this
1456                          * object. */
1457                         mdt_prep_ma_buf_from_rep(info, child, ma);
1458                         ma->ma_need |= MA_HSM;
1459                         result = mdt_attr_get_complex(info, child, ma);
1460                         if (result != 0)
1461                                 GOTO(out_child, result);
1462                 } else {
1463                         /* Object does not exist. Likely FS corruption. */
1464                         CERROR("%s: name '"DNAME"' present, but FID "
1465                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1466                                PNAME(&rr->rr_name), PFID(child_fid));
1467                         GOTO(out_child, result = -EIO);
1468                 }
1469         }
1470
1471         rc = mdt_check_resent_lock(info, child, lhc);
1472         if (rc < 0) {
1473                 GOTO(out_child, result = rc);
1474         } else if (rc == 0) {
1475                 /* the open lock might already be gotten in
1476                  * ldlm_handle_enqueue() */
1477                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1478                 if (create_flags & MDS_OPEN_LOCK)
1479                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1480         } else {
1481                 /* get openlock if this isn't replay and client requested it */
1482                 if (!req_is_replay(req)) {
1483                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1484                         object_locked = 1;
1485                         if (rc != 0)
1486                                 GOTO(out_child_unlock, result = rc);
1487                         else if (create_flags & MDS_OPEN_LOCK)
1488                                 mdt_set_disposition(info, ldlm_rep,
1489                                                     DISP_OPEN_LOCK);
1490                 }
1491         }
1492         /* Try to open it now. */
1493         rc = mdt_finish_open(info, parent, child, create_flags,
1494                              created, ldlm_rep);
1495         if (rc) {
1496                 result = rc;
1497                 /* openlock will be released if mdt_finish_open failed */
1498                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1499
1500                 if (created && create_flags & MDS_OPEN_VOLATILE) {
1501                         CERROR("%s: cannot open volatile file "DFID", orphan "
1502                                "file will be left in PENDING directory until "
1503                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1504                                PFID(mdt_object_fid(child)), rc);
1505                         GOTO(out_child_unlock, result);
1506                 }
1507
1508                 if (created) {
1509                         ma->ma_need = 0;
1510                         ma->ma_valid = 0;
1511                         rc = mdo_unlink(info->mti_env,
1512                                         mdt_object_child(parent),
1513                                         mdt_object_child(child),
1514                                         &rr->rr_name,
1515                                         &info->mti_attr, 0);
1516                         if (rc != 0)
1517                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1518                                        mdt_obd_name(info->mti_mdt),
1519                                        PFID(mdt_object_fid(child)), rc);
1520                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1521                 }
1522         }
1523         EXIT;
1524 out_child_unlock:
1525         if (object_locked)
1526                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1527 out_child:
1528         mdt_object_put(info->mti_env, child);
1529 out_parent:
1530         mdt_object_unlock_put(info, parent, lh, result || !created);
1531 out:
1532         if (result)
1533                 lustre_msg_set_transno(req->rq_repmsg, 0);
1534         return result;
1535 }
1536
1537 /**
1538  * Create an orphan object use local root.
1539  */
1540 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1541                                           struct mdt_device *mdt,
1542                                           const struct lu_fid *fid,
1543                                           struct md_attr *attr, fmode_t fmode)
1544 {
1545         const struct lu_env *env = info->mti_env;
1546         struct md_op_spec *spec = &info->mti_spec;
1547         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1548         struct mdt_object *obj = NULL;
1549         struct mdt_object *local_root;
1550         static const struct lu_name lname = {
1551                 .ln_name = "i_am_nobody",
1552                 .ln_namelen = sizeof("i_am_nobody") - 1,
1553         };
1554         struct lu_ucred *uc;
1555         cfs_cap_t uc_cap_save;
1556         int rc;
1557         ENTRY;
1558
1559         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1560         if (rc != 0)
1561                 RETURN(ERR_PTR(rc));
1562
1563         local_root = mdt_object_find(env, mdt, local_root_fid);
1564         if (IS_ERR(local_root))
1565                 RETURN(local_root);
1566
1567         obj = mdt_object_new(env, mdt, fid);
1568         if (IS_ERR(obj))
1569                 GOTO(out, rc = PTR_ERR(obj));
1570
1571         spec->sp_cr_lookup = 0;
1572         spec->sp_feat = &dt_directory_features;
1573         spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1574         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1575         if (attr->ma_valid & MA_LOV) {
1576                 spec->u.sp_ea.eadata = attr->ma_lmm;
1577                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1578                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1579         } else {
1580                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1581         }
1582
1583         uc = lu_ucred(env);
1584         uc_cap_save = uc->uc_cap;
1585         uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1586         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1587                         mdt_object_child(obj), spec, attr);
1588         uc->uc_cap = uc_cap_save;
1589         if (rc < 0) {
1590                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1591                        mdt_obd_name(mdt), PFID(fid), rc);
1592                 GOTO(out, rc);
1593         }
1594
1595         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1596         if (rc < 0)
1597                 CERROR("%s: cannot open volatile file "DFID", orphan "
1598                        "file will be left in PENDING directory until "
1599                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1600                        PFID(fid), rc);
1601         GOTO(out, rc);
1602
1603 out:
1604         if (rc < 0) {
1605                 if (!IS_ERR(obj))
1606                         mdt_object_put(env, obj);
1607                 obj = ERR_PTR(rc);
1608         }
1609         mdt_object_put(env, local_root);
1610         return obj;
1611 }
1612
1613 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1614                            struct md_attr *ma)
1615 {
1616         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1617         struct close_data      *data;
1618         struct ldlm_lock       *lease;
1619         struct mdt_object      *orphan;
1620         struct md_attr         *orp_ma;
1621         struct lu_buf          *buf;
1622         bool                    lease_broken;
1623         int                     rc;
1624         int                     rc2;
1625         ENTRY;
1626
1627         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1628                 RETURN(-EROFS);
1629
1630         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1631         if (data == NULL)
1632                 RETURN(-EPROTO);
1633
1634         lease = ldlm_handle2lock(&data->cd_handle);
1635         if (lease == NULL)
1636                 RETURN(-ESTALE);
1637
1638         /* try to hold open_sem so that nobody else can open the file */
1639         if (!down_write_trylock(&o->mot_open_sem)) {
1640                 ldlm_lock_cancel(lease);
1641                 GOTO(out_reprocess, rc = -EBUSY);
1642         }
1643
1644         /* Check if the lease open lease has already canceled */
1645         lock_res_and_lock(lease);
1646         lease_broken = ldlm_is_cancel(lease);
1647         unlock_res_and_lock(lease);
1648
1649         LDLM_DEBUG(lease, DFID " lease broken? %d",
1650                    PFID(mdt_object_fid(o)), lease_broken);
1651
1652         /* Cancel server side lease. Client side counterpart should
1653          * have been cancelled. It's okay to cancel it now as we've
1654          * held mot_open_sem. */
1655         ldlm_lock_cancel(lease);
1656
1657         if (lease_broken) /* don't perform release task */
1658                 GOTO(out_unlock, rc = -ESTALE);
1659
1660         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1661                 GOTO(out_unlock, rc = -EINVAL);
1662
1663         /* ma_need was set before but it seems fine to change it in order to
1664          * avoid modifying the one from RPC */
1665         ma->ma_need = MA_HSM;
1666         rc = mdt_attr_get_complex(info, o, ma);
1667         if (rc != 0)
1668                 GOTO(out_unlock, rc);
1669
1670         if (!mdt_hsm_release_allow(ma))
1671                 GOTO(out_unlock, rc = -EPERM);
1672
1673         /* already released? */
1674         if (ma->ma_hsm.mh_flags & HS_RELEASED)
1675                 GOTO(out_unlock, rc = 0);
1676
1677         /* Compare on-disk and packed data_version */
1678         if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1679                 CDEBUG(D_HSM, DFID" data_version mismatches: packed="LPU64
1680                        " and on-disk="LPU64"\n", PFID(mdt_object_fid(o)),
1681                        data->cd_data_version, ma->ma_hsm.mh_arch_ver);
1682                 GOTO(out_unlock, rc = -EPERM);
1683         }
1684
1685         ma->ma_valid = MA_INODE;
1686         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1687         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1688         if (rc < 0)
1689                 GOTO(out_unlock, rc);
1690
1691         ma->ma_need = MA_INODE | MA_LOV;
1692         rc = mdt_attr_get_complex(info, o, ma);
1693         if (rc < 0)
1694                 GOTO(out_unlock, rc);
1695
1696         if (!(ma->ma_valid & MA_LOV)) {
1697                 /* Even empty file are released */
1698                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1699                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
1700                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1701                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1702                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1703         } else {
1704                 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
1705                  * ma_lmm as lov_user_md, then it will be confused by union of
1706                  * layout_gen and stripe_offset. */
1707                 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
1708                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
1709                 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
1710                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
1711                 else
1712                         GOTO(out_unlock, rc = -EINVAL);
1713         }
1714
1715         /* Set file as released */
1716         ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1717
1718         orp_ma = &info->mti_u.hsm.attr;
1719         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1720         orp_ma->ma_attr.la_uid = ma->ma_attr.la_uid;
1721         orp_ma->ma_attr.la_gid = ma->ma_attr.la_gid;
1722         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1723         orp_ma->ma_lmm = ma->ma_lmm;
1724         orp_ma->ma_lmm_size = ma->ma_lmm_size;
1725         orp_ma->ma_valid = MA_INODE | MA_LOV;
1726         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1727                                  FMODE_WRITE);
1728         if (IS_ERR(orphan)) {
1729                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1730                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1731                        PTR_ERR(orphan));
1732                 GOTO(out_unlock, rc = PTR_ERR(orphan));
1733         }
1734
1735         /* Set up HSM attribute for orphan object */
1736         CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
1737         buf = &info->mti_buf;
1738         buf->lb_buf = info->mti_xattr_buf;
1739         buf->lb_len = sizeof(struct hsm_attrs);
1740         ma->ma_hsm.mh_flags |= HS_RELEASED;
1741         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1742         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
1743
1744         mdt_lock_reg_init(lh, LCK_EX);
1745         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
1746                              MDS_INODELOCK_XATTR);
1747         if (rc != 0)
1748                 GOTO(out_close, rc);
1749
1750         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
1751                           XATTR_NAME_HSM, 0);
1752
1753         if (rc == 0)
1754                 /* Swap layout with orphan object */
1755                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
1756                                      mdt_object_child(orphan),
1757                                      SWAP_LAYOUTS_MDS_HSM);
1758
1759         /* Release exclusive LL */
1760         mdt_object_unlock(info, o, lh, 1);
1761
1762         EXIT;
1763
1764 out_close:
1765         /* Close orphan object anyway */
1766         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
1767                        FMODE_WRITE);
1768         if (rc2 < 0)
1769                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
1770                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
1771         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
1772                         "object closed");
1773         mdt_object_put(info->mti_env, orphan);
1774
1775 out_unlock:
1776         up_write(&o->mot_open_sem);
1777
1778         /* already released */
1779         if (rc == 0) {
1780                 struct mdt_body *repbody;
1781
1782                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1783                 LASSERT(repbody != NULL);
1784                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1785         }
1786
1787 out_reprocess:
1788         ldlm_reprocess_all(lease->l_resource);
1789         LDLM_LOCK_PUT(lease);
1790
1791         ma->ma_valid = 0;
1792         ma->ma_need = 0;
1793
1794         return rc;
1795 }
1796
1797 int mdt_close_swap_layouts(struct mdt_thread_info *info,
1798                            struct mdt_object *o, struct md_attr *ma)
1799 {
1800         struct mdt_lock_handle  *lh1 = &info->mti_lh[MDT_LH_NEW];
1801         struct mdt_lock_handle  *lh2 = &info->mti_lh[MDT_LH_OLD];
1802         struct close_data       *data;
1803         struct ldlm_lock        *lease;
1804         struct mdt_object       *o1 = o, *o2;
1805         bool                     lease_broken;
1806         bool                     swap_objects;
1807         int                      rc;
1808         ENTRY;
1809
1810         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1811                 RETURN(-EROFS);
1812
1813         if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
1814                 RETURN(-EINVAL);
1815
1816         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1817         if (data == NULL)
1818                 RETURN(-EPROTO);
1819
1820         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1821                 RETURN(-EINVAL);
1822
1823         rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
1824         if (unlikely(rc == 0))
1825                 RETURN(-EINVAL);
1826
1827         /* Exchange o1 and o2, to enforce locking order */
1828         swap_objects = (rc < 0);
1829
1830         lease = ldlm_handle2lock(&data->cd_handle);
1831         if (lease == NULL)
1832                 RETURN(-ESTALE);
1833
1834         o2 = mdt_object_find(info->mti_env, info->mti_mdt, &data->cd_fid);
1835         if (IS_ERR(o2))
1836                 GOTO(out_lease, rc = PTR_ERR(o2));
1837
1838         if (!S_ISREG(lu_object_attr(&o2->mot_obj))) {
1839                 swap_objects = false; /* not swapped yet */
1840                 GOTO(out_obj, rc = -EINVAL);
1841         }
1842
1843         if (swap_objects)
1844                 swap(o1, o2);
1845
1846         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1847                            MAY_WRITE);
1848         if (rc < 0)
1849                 GOTO(out_obj, rc);
1850
1851         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1852                            MAY_WRITE);
1853         if (rc < 0)
1854                 GOTO(out_obj, rc);
1855
1856         /* try to hold open_sem so that nobody else can open the file */
1857         if (!down_write_trylock(&o->mot_open_sem)) {
1858                 ldlm_lock_cancel(lease);
1859                 GOTO(out_obj, rc = -EBUSY);
1860         }
1861
1862         /* Check if the lease open lease has already canceled */
1863         lock_res_and_lock(lease);
1864         lease_broken = ldlm_is_cancel(lease);
1865         unlock_res_and_lock(lease);
1866
1867         LDLM_DEBUG(lease, DFID " lease broken? %d",
1868                    PFID(mdt_object_fid(o)), lease_broken);
1869
1870         /* Cancel server side lease. Client side counterpart should
1871          * have been cancelled. It's okay to cancel it now as we've
1872          * held mot_open_sem. */
1873         ldlm_lock_cancel(lease);
1874
1875         if (lease_broken)
1876                 GOTO(out_unlock_sem, rc = -ESTALE);
1877
1878         mdt_lock_reg_init(lh1, LCK_EX);
1879         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1880                              MDS_INODELOCK_XATTR);
1881         if (rc < 0)
1882                 GOTO(out_unlock_sem, rc);
1883
1884         mdt_lock_reg_init(lh2, LCK_EX);
1885         rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1886                              MDS_INODELOCK_XATTR);
1887         if (rc < 0)
1888                 GOTO(out_unlock1, rc);
1889
1890         /* Swap layout with orphan object */
1891         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1892                              mdt_object_child(o2), 0);
1893         if (rc < 0)
1894                 GOTO(out_unlock2, rc);
1895
1896         EXIT;
1897
1898 out_unlock2:
1899         /* Release exclusive LL */
1900         mdt_object_unlock(info, o2, lh2, 1);
1901
1902 out_unlock1:
1903         mdt_object_unlock(info, o1, lh1, 1);
1904
1905 out_unlock_sem:
1906         up_write(&o->mot_open_sem);
1907
1908         /* already swapped */
1909         if (rc == 0) {
1910                 struct mdt_body *repbody;
1911
1912                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1913                 LASSERT(repbody != NULL);
1914                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1915         }
1916
1917 out_obj:
1918         mdt_object_put(info->mti_env, swap_objects ? o1 : o2);
1919
1920         ldlm_reprocess_all(lease->l_resource);
1921
1922 out_lease:
1923         LDLM_LOCK_PUT(lease);
1924
1925         if (ma != NULL) {
1926                 ma->ma_valid = 0;
1927                 ma->ma_need = 0;
1928         }
1929
1930         return rc;
1931 }
1932
1933 #define MFD_CLOSED(mode) ((mode) == MDS_FMODE_CLOSED)
1934 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1935 {
1936         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1937 }
1938
1939 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
1940 {
1941         struct mdt_object *o = mfd->mfd_object;
1942         struct md_object *next = mdt_object_child(o);
1943         struct md_attr *ma = &info->mti_attr;
1944         int rc = 0;
1945         __u64 mode;
1946         ENTRY;
1947
1948         mode = mfd->mfd_mode;
1949
1950         if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
1951                 rc = mdt_hsm_release(info, o, ma);
1952                 if (rc < 0) {
1953                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
1954                                mdt_obd_name(info->mti_mdt),
1955                                PFID(mdt_object_fid(o)), rc);
1956                         /* continue to close even error occurred. */
1957                 }
1958         }
1959
1960         if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
1961                 rc = mdt_close_swap_layouts(info, o, ma);
1962                 if (rc < 0) {
1963                         CDEBUG(D_INODE,
1964                                "%s: cannot swap layout of "DFID": rc=%d\n",
1965                                mdt_obd_name(info->mti_mdt),
1966                                PFID(mdt_object_fid(o)), rc);
1967                         /* continue to close even if error occurred. */
1968                 }
1969         }
1970
1971         if (mode & FMODE_WRITE)
1972                 mdt_write_put(o);
1973         else if (mode & MDS_FMODE_EXEC)
1974                 mdt_write_allow(o);
1975
1976         /* Update atime on close only. */
1977         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
1978             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
1979                 /* Set the atime only. */
1980                 ma->ma_valid = MA_INODE;
1981                 ma->ma_attr.la_valid = LA_ATIME;
1982                 rc = mo_attr_set(info->mti_env, next, ma);
1983         }
1984
1985         /* If file data is modified, add the dirty flag. */
1986         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
1987                 rc = mdt_add_dirty_flag(info, o, ma);
1988
1989         ma->ma_need |= MA_INODE;
1990         ma->ma_valid &= ~MA_INODE;
1991
1992         if (!MFD_CLOSED(mode))
1993                 rc = mo_close(info->mti_env, next, ma, mode);
1994
1995         /* adjust open and lease count */
1996         if (mode & MDS_OPEN_LEASE) {
1997                 LASSERT(atomic_read(&o->mot_lease_count) > 0);
1998                 atomic_dec(&o->mot_lease_count);
1999         }
2000
2001         LASSERT(atomic_read(&o->mot_open_count) > 0);
2002         atomic_dec(&o->mot_open_count);
2003         mdt_mfd_free(mfd);
2004         mdt_object_put(info->mti_env, o);
2005
2006         RETURN(rc);
2007 }
2008
2009 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2010                        struct mdt_body *repbody)
2011 {
2012         struct mdt_export_data *med;
2013         struct mdt_file_data   *mfd;
2014         struct mdt_object      *o;
2015         struct md_attr         *ma = &info->mti_attr;
2016         int                     ret = 0;
2017         int                     rc = 0;
2018         ENTRY;
2019
2020         med = &req->rq_export->exp_mdt_data;
2021         spin_lock(&med->med_open_lock);
2022         mfd = mdt_handle2mfd(med, &info->mti_close_handle, req_is_replay(req));
2023         if (mdt_mfd_closed(mfd)) {
2024                 spin_unlock(&med->med_open_lock);
2025                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2026                        ": cookie = "LPX64"\n", PFID(info->mti_rr.rr_fid1),
2027                        info->mti_close_handle.cookie);
2028                 /** not serious error since bug 3633 */
2029                 rc = -ESTALE;
2030         } else {
2031                 class_handle_unhash(&mfd->mfd_handle);
2032                 list_del_init(&mfd->mfd_list);
2033                 spin_unlock(&med->med_open_lock);
2034
2035                 /* Do not lose object before last unlink. */
2036                 o = mfd->mfd_object;
2037                 mdt_object_get(info->mti_env, o);
2038                 ret = mdt_mfd_close(info, mfd);
2039                 if (repbody != NULL)
2040                         rc = mdt_handle_last_unlink(info, o, ma);
2041                 mdt_object_put(info->mti_env, o);
2042         }
2043
2044         RETURN(rc ? rc : ret);
2045 }
2046
2047 int mdt_close(struct tgt_session_info *tsi)
2048 {
2049         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2050         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2051         struct md_attr         *ma = &info->mti_attr;
2052         struct mdt_body        *repbody = NULL;
2053         int rc, ret = 0;
2054         ENTRY;
2055
2056         mdt_counter_incr(req, LPROC_MDT_CLOSE);
2057         /* Close may come with the Size-on-MDS update. Unpack it. */
2058         rc = mdt_close_unpack(info);
2059         if (rc)
2060                 GOTO(out, rc = err_serious(rc));
2061
2062         /* These fields are no longer used and are left for compatibility.
2063          * size is always zero */
2064         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2065                              0);
2066         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2067                              0);
2068         rc = req_capsule_server_pack(info->mti_pill);
2069         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2070                 mdt_client_compatibility(info);
2071                 if (rc == 0)
2072                         mdt_fix_reply(info);
2073                 mdt_exit_ucred(info);
2074                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2075         }
2076
2077         /* Continue to close handle even if we can not pack reply */
2078         if (rc == 0) {
2079                 repbody = req_capsule_server_get(info->mti_pill,
2080                                                  &RMF_MDT_BODY);
2081                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2082                                                     &RMF_MDT_MD);
2083                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2084                                                        &RMF_MDT_MD,
2085                                                        RCL_SERVER);
2086                 ma->ma_need = MA_INODE | MA_LOV;
2087                 repbody->mbo_eadatasize = 0;
2088                 repbody->mbo_aclsize = 0;
2089         } else {
2090                 rc = err_serious(rc);
2091         }
2092
2093         rc = mdt_close_internal(info, req, repbody);
2094         if (rc != -ESTALE)
2095                 mdt_empty_transno(info, rc);
2096
2097         if (repbody != NULL) {
2098                 mdt_client_compatibility(info);
2099                 rc = mdt_fix_reply(info);
2100         }
2101
2102         mdt_exit_ucred(info);
2103         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2104                 GOTO(out, rc = err_serious(-ENOMEM));
2105
2106         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2107                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2108                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2109 out:
2110         mdt_thread_info_fini(info);
2111         RETURN(rc ? rc : ret);
2112 }