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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/mdt/mdt_open.c
  33  *
  34  * Lustre Metadata Target (mdt) open/close file handling
  35  *
  36  * Author: Huang Hua <huanghua@clusterfs.com>
  37  */
  38
  39 #define DEBUG_SUBSYSTEM S_MDS
  40
  41 #include <lustre_acl.h>
  42 #include <lustre_mds.h>
  43 #include <lustre_swab.h>
  44 #include "mdt_internal.h"
  45 #include <lustre_nodemap.h>
  46
  47 /* we do nothing because we do not have refcount now */
  48 static void mdt_mfd_get(void *mfdp)
  49 {
  50 }
  51
  52 static struct portals_handle_ops mfd_handle_ops = {
  53         .hop_addref = mdt_mfd_get,
  54         .hop_free   = NULL,
  55 };
  56
  57 /* Create a new mdt_file_data struct, initialize it,
  58  * and insert it to global hash table */
  59 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
  60 {
  61         struct mdt_file_data *mfd;
  62         ENTRY;
  63
  64         OBD_ALLOC_PTR(mfd);
  65         if (mfd != NULL) {
  66                 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  67                 mfd->mfd_handle.h_owner = med;
  68                 INIT_LIST_HEAD(&mfd->mfd_list);
  69                 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
  70         }
  71
  72         RETURN(mfd);
  73 }
  74
  75 /*
  76  * Find the mfd pointed to by handle in global hash table.
  77  * In case of replay the handle is obsoleted
  78  * but mfd can be found in mfd list by that handle.
  79  * Callers need to be holding med_open_lock.
  80  */
  81 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
  82                                      const struct lustre_handle *handle,
  83                                      bool is_replay_or_resent)
  84 {
  85         struct mdt_file_data   *mfd;
  86         ENTRY;
  87
  88         LASSERT(handle != NULL);
  89         mfd = class_handle2object(handle->cookie, med);
  90         /* during dw/setattr replay the mfd can be found by old handle */
  91         if (mfd == NULL && is_replay_or_resent) {
  92                 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  93                         if (mfd->mfd_old_handle.cookie == handle->cookie)
  94                                 RETURN(mfd);
  95                 }
  96                 mfd = NULL;
  97         }
  98
  99         RETURN(mfd);
 100 }
 101
 102 /* free mfd */
 103 void mdt_mfd_free(struct mdt_file_data *mfd)
 104 {
 105         LASSERT(list_empty(&mfd->mfd_list));
 106         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 107 }
 108
 109 static int mdt_create_data(struct mdt_thread_info *info,
 110                            struct mdt_object *p, struct mdt_object *o)
 111 {
 112         struct md_op_spec     *spec = &info->mti_spec;
 113         struct md_attr        *ma   = &info->mti_attr;
 114         int                    rc   = 0;
 115         ENTRY;
 116
 117         if (!md_should_create(spec->sp_cr_flags))
 118                 RETURN(0);
 119
 120         ma->ma_need = MA_INODE | MA_LOV;
 121         ma->ma_valid = 0;
 122         mutex_lock(&o->mot_lov_mutex);
 123         if (!o->mot_lov_created) {
 124                 rc = mdo_create_data(info->mti_env,
 125                                      p ? mdt_object_child(p) : NULL,
 126                                      mdt_object_child(o), spec, ma);
 127                 if (rc == 0)
 128                         rc = mdt_attr_get_complex(info, o, ma);
 129
 130                 if (rc == 0 && ma->ma_valid & MA_LOV)
 131                         o->mot_lov_created = 1;
 132         }
 133
 134         mutex_unlock(&o->mot_lov_mutex);
 135         RETURN(rc);
 136 }
 137
 138 int mdt_write_read(struct mdt_object *o)
 139 {
 140         int rc = 0;
 141         ENTRY;
 142         spin_lock(&o->mot_write_lock);
 143         rc = o->mot_write_count;
 144         spin_unlock(&o->mot_write_lock);
 145         RETURN(rc);
 146 }
 147
 148 int mdt_write_get(struct mdt_object *o)
 149 {
 150         int rc = 0;
 151         ENTRY;
 152         spin_lock(&o->mot_write_lock);
 153         if (o->mot_write_count < 0)
 154                 rc = -ETXTBSY;
 155         else
 156                 o->mot_write_count++;
 157         spin_unlock(&o->mot_write_lock);
 158
 159         RETURN(rc);
 160 }
 161
 162 void mdt_write_put(struct mdt_object *o)
 163 {
 164         ENTRY;
 165         spin_lock(&o->mot_write_lock);
 166         o->mot_write_count--;
 167         spin_unlock(&o->mot_write_lock);
 168         EXIT;
 169 }
 170
 171 static int mdt_write_deny(struct mdt_object *o)
 172 {
 173         int rc = 0;
 174         ENTRY;
 175         spin_lock(&o->mot_write_lock);
 176         if (o->mot_write_count > 0)
 177                 rc = -ETXTBSY;
 178         else
 179                 o->mot_write_count--;
 180         spin_unlock(&o->mot_write_lock);
 181         RETURN(rc);
 182 }
 183
 184 static void mdt_write_allow(struct mdt_object *o)
 185 {
 186         ENTRY;
 187         spin_lock(&o->mot_write_lock);
 188         o->mot_write_count++;
 189         spin_unlock(&o->mot_write_lock);
 190         EXIT;
 191 }
 192
 193 /* there can be no real transaction so prepare the fake one */
 194 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 195 {
 196         struct mdt_device      *mdt = info->mti_mdt;
 197         struct ptlrpc_request  *req = mdt_info_req(info);
 198         struct tg_export_data  *ted;
 199         struct lsd_client_data *lcd;
 200         ENTRY;
 201
 202         if (mdt_rdonly(req->rq_export))
 203                 RETURN_EXIT;
 204
 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 %llu 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 = %llu, last_committed = %llu\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 = %llu < last_transno = %llu\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_DENTRY, DFID " Change mfd mode %#llo -> %#llo.\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 = %#llx 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 = %#llx\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 = %#llx\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 %#llx 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_tail(&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         /* compatibility check for 2.10 clients when it tries to open mirrored
 526          * files. 2.10 clients don't verify overlapping components so they
 527          * would read and write mirrored files just as if they were normal
 528          * PFL files, which will cause the problem that sycned mirrors actually
 529          * contain different data.
 530          * Older clients are not a concern here because they don't even
 531          * understand PFL layout. */
 532         if (isreg && !exp_connect_flr(exp) && ma->ma_valid & MA_LOV &&
 533             mdt_lmm_is_flr(ma->ma_lmm)) {
 534                 /* LU-10286: for simplicity clients who don't understand
 535                  * mirrored layout(with connect flag OBD_CONNECT2_FLR) won't
 536                  * be able to open mirrored files */
 537                 RETURN(-EOPNOTSUPP);
 538         }
 539
 540         /* LU-2275, simulate broken behaviour (esp. prevalent in
 541          * pre-2.4 servers where a very strange reply is sent on error
 542          * that looks like it was actually almost successful and a
 543          * failure at the same time.) */
 544         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 545                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 546                                                DISP_LOOKUP_NEG |
 547                                                DISP_LOOKUP_POS);
 548
 549                 if (flags & MDS_OPEN_LOCK)
 550                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 551
 552                 RETURN(-ENOENT);
 553         }
 554
 555 #ifdef CONFIG_FS_POSIX_ACL
 556         if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 557                 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
 558                 if (IS_ERR(nodemap))
 559                         RETURN(PTR_ERR(nodemap));
 560
 561                 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
 562                 nodemap_putref(nodemap);
 563                 if (rc)
 564                         RETURN(rc);
 565         }
 566 #endif
 567
 568         /*
 569          * If we are following a symlink, don't open; and do not return open
 570          * handle for special nodes as client required.
 571          */
 572         if (islnk || (!isreg && !isdir &&
 573             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 574                 lustre_msg_set_transno(req->rq_repmsg, 0);
 575                 RETURN(0);
 576         }
 577
 578         /*
 579          * We need to return the existing object's fid back, so it is done here,
 580          * after preparing the reply.
 581          */
 582         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 583                 RETURN(-EEXIST);
 584
 585         /* This can't be done earlier, we need to return reply body */
 586         if (isdir) {
 587                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 588                         /* We are trying to create or write an existing dir. */
 589                         RETURN(-EISDIR);
 590                 }
 591         } else if (flags & MDS_OPEN_DIRECTORY)
 592                 RETURN(-ENOTDIR);
 593
 594         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 595                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 596                 RETURN(-EAGAIN);
 597         }
 598
 599         mfd = NULL;
 600         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 601                 spin_lock(&med->med_open_lock);
 602                 list_for_each(t, &med->med_open_head) {
 603                         mfd = list_entry(t, struct mdt_file_data, mfd_list);
 604                         if (mfd->mfd_xid == req->rq_xid)
 605                                 break;
 606                         mfd = NULL;
 607                 }
 608                 spin_unlock(&med->med_open_lock);
 609
 610                 if (mfd != NULL) {
 611                         repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
 612                         /* set repbody->ea_size for resent case */
 613                         if (ma->ma_valid & MA_LOV) {
 614                                 LASSERT(ma->ma_lmm_size != 0);
 615                                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 616                                 if (isdir)
 617                                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 618                                 else
 619                                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 620                         }
 621                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 622                         RETURN(0);
 623                 }
 624         }
 625
 626         rc = mdt_mfd_open(info, p, o, flags, created, rep);
 627         if (!rc)
 628                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 629
 630         RETURN(rc);
 631 }
 632
 633 void mdt_reconstruct_open(struct mdt_thread_info *info,
 634                           struct mdt_lock_handle *lhc)
 635 {
 636         const struct lu_env *env = info->mti_env;
 637         struct mdt_device       *mdt  = info->mti_mdt;
 638         struct req_capsule      *pill = info->mti_pill;
 639         struct ptlrpc_request   *req  = mdt_info_req(info);
 640         struct md_attr          *ma   = &info->mti_attr;
 641         struct mdt_reint_record *rr   = &info->mti_rr;
 642         __u64                   flags = info->mti_spec.sp_cr_flags;
 643         struct ldlm_reply       *ldlm_rep;
 644         struct mdt_object       *parent;
 645         struct mdt_object       *child;
 646         struct mdt_body         *repbody;
 647         int                      rc;
 648         __u64                    opdata;
 649         ENTRY;
 650
 651         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 652         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 653         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 654
 655         ma->ma_need = MA_INODE | MA_HSM;
 656         ma->ma_valid = 0;
 657
 658         opdata = mdt_req_from_lrd(req, info->mti_reply_data);
 659         mdt_set_disposition(info, ldlm_rep, opdata);
 660
 661         CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n",
 662                ldlm_rep->lock_policy_res1, req->rq_status);
 663
 664         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
 665             req->rq_status != 0)
 666                 /* We did not create successfully, return error to client. */
 667                 GOTO(out, rc = req->rq_status);
 668
 669         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
 670                 struct obd_export *exp = req->rq_export;
 671                 /*
 672                  * We failed after creation, but we do not know in which step
 673                  * we failed. So try to check the child object.
 674                  */
 675                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
 676                 if (IS_ERR(parent)) {
 677                         rc = PTR_ERR(parent);
 678                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
 679                                       " Evicting client %s with export %s.\n",
 680                                       PFID(rr->rr_fid1), rc,
 681                                       obd_uuid2str(&exp->exp_client_uuid),
 682                                       obd_export_nid2str(exp));
 683                         mdt_export_evict(exp);
 684                         RETURN_EXIT;
 685                 }
 686
 687                 child = mdt_object_find(env, mdt, rr->rr_fid2);
 688                 if (IS_ERR(child)) {
 689                         rc = PTR_ERR(child);
 690                         LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
 691                                       "evicting client %s with export %s\n",
 692                                       PFID(rr->rr_fid2), rc,
 693                                       obd_uuid2str(&exp->exp_client_uuid),
 694                                       obd_export_nid2str(exp));
 695                         mdt_object_put(env, parent);
 696                         mdt_export_evict(exp);
 697                         RETURN_EXIT;
 698                 }
 699
 700                 if (unlikely(mdt_object_remote(child))) {
 701                         /* the child object was created on remote server */
 702                         if (!mdt_is_dne_client(exp)) {
 703                                 /* Return -EIO for old client */
 704                                 mdt_object_put(env, parent);
 705                                 mdt_object_put(env, child);
 706                                 GOTO(out, rc = -EIO);
 707                         }
 708                         repbody->mbo_fid1 = *rr->rr_fid2;
 709                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 710                         rc = 0;
 711                 } else {
 712                         if (mdt_object_exists(child)) {
 713                                 mdt_prep_ma_buf_from_rep(info, child, ma);
 714                                 rc = mdt_attr_get_complex(info, child, ma);
 715                                 if (rc == 0)
 716                                         rc = mdt_finish_open(info, parent,
 717                                                              child, flags,
 718                                                              1, ldlm_rep);
 719                         } else {
 720                                 /* the child does not exist, we should do
 721                                  * regular open */
 722                                 mdt_object_put(env, parent);
 723                                 mdt_object_put(env, child);
 724                                 GOTO(regular_open, 0);
 725                         }
 726                 }
 727                 mdt_object_put(env, parent);
 728                 mdt_object_put(env, child);
 729                 GOTO(out, rc);
 730         } else {
 731 regular_open:
 732                 /* We did not try to create, so we are a pure open */
 733                 rc = mdt_reint_open(info, lhc);
 734         }
 735
 736         EXIT;
 737 out:
 738         req->rq_status = rc;
 739         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 740         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
 741 }
 742
 743 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
 744 {
 745         __u64                    flags = info->mti_spec.sp_cr_flags;
 746         struct mdt_reint_record *rr = &info->mti_rr;
 747         struct md_attr          *ma = &info->mti_attr;
 748         struct mdt_object       *o;
 749         int                      rc;
 750         ENTRY;
 751
 752         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
 753         if (IS_ERR(o))
 754                 RETURN(rc = PTR_ERR(o));
 755
 756         if (unlikely(mdt_object_remote(o))) {
 757                 /* the child object was created on remote server */
 758                 struct mdt_body *repbody;
 759
 760                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 761                                                 DISP_LOOKUP_EXECD |
 762                                                 DISP_LOOKUP_POS));
 763                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 764                 repbody->mbo_fid1 = *rr->rr_fid2;
 765                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 766                 rc = 0;
 767         } else {
 768                 if (mdt_object_exists(o)) {
 769                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 770                                                         DISP_LOOKUP_EXECD |
 771                                                         DISP_LOOKUP_POS));
 772                         mdt_prep_ma_buf_from_rep(info, o, ma);
 773                         rc = mdt_attr_get_complex(info, o, ma);
 774                         if (rc == 0)
 775                                 rc = mdt_finish_open(info, NULL, o, flags, 0,
 776                                                      rep);
 777                 } else {
 778                         rc = -ENOENT;
 779                 }
 780         }
 781
 782         mdt_object_put(info->mti_env, o);
 783         RETURN(rc);
 784 }
 785
 786 /* lock object for open */
 787 static int mdt_object_open_lock(struct mdt_thread_info *info,
 788                                 struct mdt_object *obj,
 789                                 struct mdt_lock_handle *lhc,
 790                                 __u64 *ibits)
 791 {
 792         struct md_attr *ma = &info->mti_attr;
 793         __u64 open_flags = info->mti_spec.sp_cr_flags;
 794         __u64 trybits = 0;
 795         enum ldlm_mode lm = LCK_CR;
 796         bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
 797         bool try_layout = false;
 798         bool create_layout = false;
 799         int rc = 0;
 800         int dom_stripes = LMM_NO_DOM;
 801         bool dom_lock = false;
 802
 803         ENTRY;
 804
 805         *ibits = 0;
 806         mdt_lock_handle_init(lhc);
 807
 808         if (req_is_replay(mdt_info_req(info)))
 809                 RETURN(0);
 810
 811         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
 812                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
 813                     md_should_create(open_flags))
 814                         create_layout = true;
 815                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
 816                     ma->ma_need & MA_LOV)
 817                         try_layout = true;
 818
 819                 /* DoM files can have just MDT stripe or combined MDT + OST
 820                  * stripes.
 821                  * - In the first case the open for read/write will do IO to
 822                  *   the MDT stripe and it makes sense to take IO lock in
 823                  *   advance along with OPEN even if it is blocking lock.
 824                  * - In the second case it is just size of MDT stripe and it
 825                  *   is quite unlikely that client will write into it, though
 826                  *   it may read it. So IO lock will be taken optionally if it
 827                  *   is non-blocking one.
 828                  */
 829                 if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
 830                         dom_stripes = mdt_lmm_dom_entry(ma->ma_lmm);
 831
 832                 if (dom_stripes == LMM_DOM_ONLY &&
 833                     info->mti_mdt->mdt_opts.mo_dom_lock != 0 &&
 834                     !mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
 835                         dom_lock = true;
 836         }
 837
 838         if (acq_lease) {
 839                 /* lease open, acquire write mode of open sem */
 840                 down_write(&obj->mot_open_sem);
 841
 842                 /* Lease exists and ask for new lease */
 843                 if (atomic_read(&obj->mot_lease_count) > 0) {
 844                         /* only exclusive open is supported, so lease
 845                          * are conflicted to each other */
 846                         GOTO(out, rc = -EBUSY);
 847                 }
 848
 849                 /* Lease must be with open lock */
 850                 if (!(open_flags & MDS_OPEN_LOCK)) {
 851                         CERROR("%s: Request lease for file:"DFID ", but open lock "
 852                                "is missed, open_flags = %#llo : rc = %d\n",
 853                                mdt_obd_name(info->mti_mdt),
 854                                PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
 855                         GOTO(out, rc = -EPROTO);
 856                 }
 857
 858                 /* XXX: only exclusive open is supported. */
 859                 lm = LCK_EX;
 860                 *ibits = MDS_INODELOCK_OPEN;
 861
 862                 /* never grant LCK_EX layout lock to client */
 863                 try_layout = false;
 864         } else { /* normal open */
 865                 /* normal open holds read mode of open sem */
 866                 down_read(&obj->mot_open_sem);
 867
 868                 if (open_flags & MDS_OPEN_LOCK) {
 869                         if (open_flags & FMODE_WRITE)
 870                                 lm = LCK_CW;
 871                         else if (open_flags & MDS_FMODE_EXEC)
 872                                 lm = LCK_PR;
 873                         else
 874                                 lm = LCK_CR;
 875
 876                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
 877                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
 878                         if (open_flags & FMODE_WRITE)
 879                                 lm = LCK_CW;
 880                         else
 881                                 lm = LCK_CR;
 882
 883                         /* revoke lease */
 884                         *ibits = MDS_INODELOCK_OPEN;
 885                         try_layout = false;
 886
 887                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 888                 } else if (dom_lock) {
 889                         lm = (open_flags & FMODE_WRITE) ? LCK_PW : LCK_PR;
 890                         *ibits = MDS_INODELOCK_DOM;
 891                 }
 892
 893                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
 894                         PFID(mdt_object_fid(obj)),
 895                         atomic_read(&obj->mot_open_count), lm);
 896         }
 897
 898         mdt_lock_reg_init(lhc, lm);
 899
 900         /* one problem to return layout lock on open is that it may result
 901          * in too many layout locks cached on the client side. */
 902         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
 903                 trybits |= MDS_INODELOCK_LAYOUT;
 904         }
 905
 906         if (*ibits | trybits)
 907                 rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, false);
 908
 909         CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
 910                ", open_flags = %#llo, try_layout = %d : rc = %d\n",
 911                mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
 912                *ibits, trybits, open_flags, try_layout, rc);
 913
 914         /* will change layout, revoke layout locks by enqueuing EX lock. */
 915         if (rc == 0 && create_layout) {
 916                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
 917
 918                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
 919                         ", open_flags = %#llo\n",
 920                         PFID(mdt_object_fid(obj)), open_flags);
 921
 922                 /* We cannot enqueue another lock for the same resource we
 923                  * already have a lock for, due to mechanics of waiting list
 924                  * iterating in ldlm, see LU-3601.
 925                  * As such we'll drop the open lock we just got above here,
 926                  * it's ok not to have this open lock as it's main purpose is to
 927                  * flush unused cached client open handles. */
 928                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 929                         mdt_object_unlock(info, obj, lhc, 1);
 930
 931                 LASSERT(!try_layout);
 932                 mdt_lock_handle_init(ll);
 933                 mdt_lock_reg_init(ll, LCK_EX);
 934                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
 935
 936                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
 937         }
 938
 939         /* Check if there is any other open handles after acquiring
 940          * open lock. At this point, caching open handles have been revoked
 941          * by open lock.
 942          * XXX: Now only exclusive open is supported. Need to check the
 943          * type of open for generic lease support. */
 944         if (rc == 0 && acq_lease) {
 945                 struct ptlrpc_request *req = mdt_info_req(info);
 946                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 947                 struct mdt_file_data *mfd;
 948                 bool is_replay_or_resent;
 949                 int open_count = 0;
 950
 951                 /* For lease: application can open a file and then apply lease,
 952                  * @handle contains original open handle in that case.
 953                  * In recovery, open REQ will be replayed and the lease REQ may
 954                  * be resent that means the open handle is already stale, so we
 955                  * need to fix it up here by finding new handle. */
 956                 is_replay_or_resent = req_is_replay(req) ||
 957                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
 958
 959                 /* if the request is _not_ a replay request, rr_handle
 960                  * may be used to hold an openhandle which is issuing the
 961                  * lease request, so that this openhandle doesn't count. */
 962                 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
 963                                      is_replay_or_resent);
 964                 if (mfd != NULL)
 965                         ++open_count;
 966
 967                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
 968                         PFID(mdt_object_fid(obj)),
 969                         atomic_read(&obj->mot_open_count), open_count);
 970
 971                 if (atomic_read(&obj->mot_open_count) > open_count)
 972                         GOTO(out, rc = -EBUSY);
 973         }
 974         GOTO(out, rc);
 975
 976 out:
 977         RETURN(rc);
 978 }
 979
 980 static void mdt_object_open_unlock(struct mdt_thread_info *info,
 981                                    struct mdt_object *obj,
 982                                    struct mdt_lock_handle *lhc,
 983                                    __u64 ibits, int rc)
 984 {
 985         __u64 open_flags = info->mti_spec.sp_cr_flags;
 986         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
 987         ENTRY;
 988
 989         if (req_is_replay(mdt_info_req(info)))
 990                 RETURN_EXIT;
 991
 992         /* Release local lock - the lock put in MDT_LH_LOCAL will never
 993          * return to client side. */
 994         if (lustre_handle_is_used(&ll->mlh_reg_lh))
 995                 mdt_object_unlock(info, obj, ll, 1);
 996
 997         ll = &info->mti_lh[MDT_LH_LAYOUT];
 998         /* Release local layout lock, layout was created */
 999         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1000                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1001                 mdt_object_unlock(info, obj, ll, 1);
1002         }
1003
1004         if (open_flags & MDS_OPEN_LEASE)
1005                 up_write(&obj->mot_open_sem);
1006         else
1007                 up_read(&obj->mot_open_sem);
1008
1009         /* Cross-ref case, the lock should be returned to the client */
1010         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1011                 RETURN_EXIT;
1012
1013         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1014             !(ibits & MDS_INODELOCK_DOM)) {
1015                 /* for the open request, the lock will only return to client
1016                  * if open or layout lock is granted. */
1017                 rc = 1;
1018         }
1019
1020         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1021                 struct ldlm_reply       *ldlm_rep;
1022
1023                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1024                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1025                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1026                         mdt_object_unlock(info, obj, lhc, 1);
1027         }
1028         RETURN_EXIT;
1029 }
1030
1031 /**
1032  * Check release is permitted for the current HSM flags.
1033  */
1034 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1035 {
1036         if (!(ma->ma_valid & MA_HSM))
1037                 return false;
1038
1039         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1040                 return false;
1041
1042         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1043                 return false;
1044
1045         return true;
1046 }
1047
1048 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1049                                 struct ldlm_reply *rep,
1050                                 struct mdt_lock_handle *lhc)
1051 {
1052         const struct lu_env     *env   = info->mti_env;
1053         struct mdt_device       *mdt   = info->mti_mdt;
1054         __u64                    flags = info->mti_spec.sp_cr_flags;
1055         struct mdt_reint_record *rr    = &info->mti_rr;
1056         struct md_attr          *ma    = &info->mti_attr;
1057         struct mdt_object       *parent= NULL;
1058         struct mdt_object       *o;
1059         int                      rc;
1060         bool                     object_locked = false;
1061         __u64                    ibits = 0;
1062         ENTRY;
1063
1064         if (md_should_create(flags)) {
1065                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1066                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1067                         if (IS_ERR(parent)) {
1068                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1069                                        " for anonymous created %ld, try to"
1070                                        " use server-side parent.\n",
1071                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1072                                 parent = NULL;
1073                         }
1074                 }
1075                 if (parent == NULL)
1076                         ma->ma_need |= MA_PFID;
1077         }
1078
1079         o = mdt_object_find(env, mdt, rr->rr_fid2);
1080         if (IS_ERR(o))
1081                 GOTO(out_parent_put, rc = PTR_ERR(o));
1082
1083         if (mdt_object_remote(o)) {
1084                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1085                        mdt_obd_name(info->mti_mdt),
1086                        PFID(rr->rr_fid2));
1087                 GOTO(out, rc = -EREMOTE);
1088         } else if (!mdt_object_exists(o)) {
1089                 mdt_set_disposition(info, rep,
1090                                     DISP_IT_EXECD |
1091                                     DISP_LOOKUP_EXECD |
1092                                     DISP_LOOKUP_NEG);
1093                 GOTO(out, rc = -ENOENT);
1094         }
1095
1096         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1097
1098         mdt_prep_ma_buf_from_rep(info, o, ma);
1099         if (flags & MDS_OPEN_RELEASE)
1100                 ma->ma_need |= MA_HSM;
1101         rc = mdt_attr_get_complex(info, o, ma);
1102         if (rc)
1103                 GOTO(out, rc);
1104
1105         /* We should not change file's existing LOV EA */
1106         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1107             flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1108                 GOTO(out, rc = -EEXIST);
1109
1110         /* If a release request, check file flags are fine and ask for an
1111          * exclusive open access. */
1112         if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1113                 GOTO(out, rc = -EPERM);
1114
1115         rc = mdt_check_resent_lock(info, o, lhc);
1116         if (rc < 0) {
1117                 GOTO(out, rc);
1118         } else if (rc > 0) {
1119                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1120                 object_locked = true;
1121                 if (rc)
1122                         GOTO(out_unlock, rc);
1123         }
1124
1125         if (ma->ma_valid & MA_PFID) {
1126                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1127                 if (IS_ERR(parent)) {
1128                         CDEBUG(D_INODE, "Fail to find parent "DFID
1129                                " for anonymous created %ld, try to"
1130                                " use system default.\n",
1131                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1132                         parent = NULL;
1133                 }
1134         }
1135
1136         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1137         if (!rc) {
1138                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1139                 if (flags & MDS_OPEN_LOCK)
1140                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1141                 if (flags & MDS_OPEN_LEASE)
1142                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1143         }
1144         GOTO(out_unlock, rc);
1145
1146 out_unlock:
1147         if (object_locked)
1148                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1149 out:
1150         mdt_object_put(env, o);
1151         if (rc == 0)
1152                 mdt_pack_size2body(info, rr->rr_fid2,
1153                                    ibits & MDS_INODELOCK_DOM);
1154 out_parent_put:
1155         if (parent != NULL)
1156                 mdt_object_put(env, parent);
1157         return rc;
1158 }
1159
1160 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1161 static int mdt_cross_open(struct mdt_thread_info *info,
1162                           const struct lu_fid *parent_fid,
1163                           const struct lu_fid *fid,
1164                           struct ldlm_reply *rep, __u32 flags)
1165 {
1166         struct md_attr    *ma = &info->mti_attr;
1167         struct mdt_object *o;
1168         int                rc;
1169         ENTRY;
1170
1171         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1172         if (IS_ERR(o))
1173                 RETURN(rc = PTR_ERR(o));
1174
1175         if (mdt_object_remote(o)) {
1176                 /* Something is wrong here, the object is on another MDS! */
1177                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1178                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1179                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1180                                 &o->mot_obj,
1181                                 "Object isn't on this server! FLD error?");
1182                 rc = -EFAULT;
1183         } else {
1184                 if (mdt_object_exists(o)) {
1185                         /* Do permission check for cross-open. */
1186                         rc = mo_permission(info->mti_env, NULL,
1187                                            mdt_object_child(o),
1188                                            NULL, flags | MDS_OPEN_CROSS);
1189                         if (rc)
1190                                 goto out;
1191
1192                         mdt_prep_ma_buf_from_rep(info, o, ma);
1193                         rc = mdt_attr_get_complex(info, o, ma);
1194                         if (rc != 0)
1195                                 GOTO(out, rc);
1196
1197                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1198                 } else {
1199                         /*
1200                          * Something is wrong here. lookup was positive but
1201                          * there is no object!
1202                          */
1203                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1204                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1205                         rc = -EFAULT;
1206                 }
1207         }
1208 out:
1209         mdt_object_put(info->mti_env, o);
1210         RETURN(rc);
1211 }
1212
1213 /*
1214  * find root object and take its xattr lock if it's on remote MDT, later create
1215  * may use fs default striping (which is stored in root xattr).
1216  */
1217 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1218                                struct mdt_device *mdt)
1219 {
1220         struct mdt_object *md_root = mdt->mdt_md_root;
1221         struct lustre_handle lhroot;
1222         int rc;
1223
1224         if (md_root == NULL) {
1225                 lu_root_fid(&info->mti_tmp_fid1);
1226                 md_root = mdt_object_find(info->mti_env, mdt,
1227                                           &info->mti_tmp_fid1);
1228                 if (IS_ERR(md_root))
1229                         return PTR_ERR(md_root);
1230
1231                 spin_lock(&mdt->mdt_lock);
1232                 if (mdt->mdt_md_root != NULL) {
1233                         spin_unlock(&mdt->mdt_lock);
1234
1235                         LASSERTF(mdt->mdt_md_root == md_root,
1236                                  "Different root object ("
1237                                  DFID") instances, %p, %p\n",
1238                                  PFID(&info->mti_tmp_fid1),
1239                                  mdt->mdt_md_root, md_root);
1240                         LASSERT(atomic_read(
1241                                 &md_root->mot_obj.lo_header->loh_ref) > 1);
1242
1243                         mdt_object_put(info->mti_env, md_root);
1244                 } else {
1245                         mdt->mdt_md_root = md_root;
1246                         spin_unlock(&mdt->mdt_lock);
1247                 }
1248         }
1249
1250         if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1251                 return 0;
1252
1253         rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root),
1254                                     &lhroot, LCK_PR, MDS_INODELOCK_XATTR,
1255                                     true);
1256         if (rc < 0)
1257                 return rc;
1258
1259         md_root->mot_cache_attr = 1;
1260
1261         /* don't cancel this lock, so that we know the cached xattr is valid. */
1262         ldlm_lock_decref(&lhroot, LCK_PR);
1263
1264         return 0;
1265 }
1266
1267 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1268 {
1269         struct mdt_device       *mdt = info->mti_mdt;
1270         struct ptlrpc_request   *req = mdt_info_req(info);
1271         struct mdt_object       *parent;
1272         struct mdt_object       *child;
1273         struct mdt_lock_handle  *lh;
1274         struct ldlm_reply       *ldlm_rep;
1275         struct mdt_body         *repbody;
1276         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1277         struct md_attr          *ma = &info->mti_attr;
1278         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1279         __u64                    ibits = 0;
1280         struct mdt_reint_record *rr = &info->mti_rr;
1281         int                      result, rc;
1282         int                      created = 0;
1283         int                      object_locked = 0;
1284         __u32                    msg_flags;
1285         ENTRY;
1286
1287         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1288                                (obd_timeout + 1) / 4);
1289
1290         mdt_counter_incr(req, LPROC_MDT_OPEN);
1291         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1292
1293         ma->ma_need = MA_INODE;
1294         ma->ma_valid = 0;
1295
1296         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1297         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1298
1299         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1300                 CERROR("file join is not supported anymore.\n");
1301                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1302         }
1303         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1304
1305         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1306             info->mti_spec.u.sp_ea.eadata == NULL)
1307                 GOTO(out, result = err_serious(-EINVAL));
1308
1309         if (create_flags & FMODE_WRITE &&
1310             exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1311                 GOTO(out, result = -EROFS);
1312
1313         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1314                "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1315                PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1316                PFID(rr->rr_fid2), create_flags,
1317                ma->ma_attr.la_mode, msg_flags);
1318
1319         if (info->mti_cross_ref) {
1320                 /* This is cross-ref open */
1321                 mdt_set_disposition(info, ldlm_rep,
1322                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1323                              DISP_LOOKUP_POS));
1324                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1325                                         ldlm_rep, create_flags);
1326                 GOTO(out, result);
1327         } else if (req_is_replay(req)) {
1328                 result = mdt_open_by_fid(info, ldlm_rep);
1329
1330                 if (result != -ENOENT)
1331                         GOTO(out, result);
1332
1333                 /* We didn't find the correct object, so we need to re-create it
1334                  * via a regular replay. */
1335                 if (!(create_flags & MDS_OPEN_CREAT)) {
1336                         DEBUG_REQ(D_ERROR, req,
1337                                   "OPEN & CREAT not in open replay/by_fid.");
1338                         GOTO(out, result = -EFAULT);
1339                 }
1340                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1341         } else if (create_flags & (MDS_OPEN_BY_FID | MDS_OPEN_LOCK)) {
1342                 /*
1343                  * MDS_OPEN_LOCK is checked for backward compatibility with 2.1
1344                  * client.
1345                  */
1346                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1347                 if (result < 0)
1348                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1349                                PFID(rr->rr_fid2), result);
1350                 GOTO(out, result);
1351         }
1352
1353         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1354                 GOTO(out, result = err_serious(-ENOMEM));
1355
1356         mdt_set_disposition(info, ldlm_rep,
1357                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1358
1359         if (!lu_name_is_valid(&rr->rr_name))
1360                 GOTO(out, result = -EPROTO);
1361
1362         result = mdt_lock_root_xattr(info, mdt);
1363         if (result < 0)
1364                 GOTO(out, result);
1365
1366 again:
1367         lh = &info->mti_lh[MDT_LH_PARENT];
1368         mdt_lock_pdo_init(lh,
1369                           (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1370                           &rr->rr_name);
1371
1372         parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1373         if (IS_ERR(parent))
1374                 GOTO(out, result = PTR_ERR(parent));
1375
1376         result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1377         if (result != 0) {
1378                 mdt_object_put(info->mti_env, parent);
1379                 GOTO(out, result);
1380         }
1381
1382         /* get and check version of parent */
1383         result = mdt_version_get_check(info, parent, 0);
1384         if (result)
1385                 GOTO(out_parent, result);
1386
1387         fid_zero(child_fid);
1388
1389         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1390                             &rr->rr_name, child_fid, &info->mti_spec);
1391
1392         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1393                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1394                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1395                  PFID(child_fid));
1396
1397         if (result != 0 && result != -ENOENT && result != -ESTALE)
1398                 GOTO(out_parent, result);
1399
1400         if (result == -ENOENT || result == -ESTALE) {
1401                 /* If the object is dead, let's check if the object
1402                  * is being migrated to a new object */
1403                 if (result == -ESTALE) {
1404                         struct lu_buf lmv_buf;
1405
1406                         lmv_buf.lb_buf = info->mti_xattr_buf;
1407                         lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
1408                         rc = mo_xattr_get(info->mti_env,
1409                                           mdt_object_child(parent),
1410                                           &lmv_buf, XATTR_NAME_LMV);
1411                         if (rc > 0) {
1412                                 struct lmv_mds_md_v1 *lmv;
1413
1414                                 lmv = lmv_buf.lb_buf;
1415                                 if (le32_to_cpu(lmv->lmv_hash_type) &
1416                                                 LMV_HASH_FLAG_MIGRATION) {
1417                                         /* Get the new parent FID and retry */
1418                                         mdt_object_unlock_put(info, parent,
1419                                                               lh, 1);
1420                                         mdt_lock_handle_init(lh);
1421                                         fid_le_to_cpu(
1422                                                 (struct lu_fid *)rr->rr_fid1,
1423                                                 &lmv->lmv_stripe_fids[1]);
1424                                         goto again;
1425                                 }
1426                         }
1427                 }
1428
1429                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1430                 if (result == -ESTALE) {
1431                         /*
1432                          * -ESTALE means the parent is a dead(unlinked) dir, so
1433                          * it should return -ENOENT to in accordance with the
1434                          * original mds implementaion.
1435                          */
1436                         GOTO(out_parent, result = -ENOENT);
1437                 }
1438
1439                 if (!(create_flags & MDS_OPEN_CREAT))
1440                         GOTO(out_parent, result);
1441                 if (mdt_rdonly(req->rq_export))
1442                         GOTO(out_parent, result = -EROFS);
1443                 *child_fid = *info->mti_rr.rr_fid2;
1444                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1445                          PFID(child_fid));
1446                 /* In the function below, .hs_keycmp resolves to
1447                  * lu_obj_hop_keycmp() */
1448                 /* coverity[overrun-buffer-val] */
1449                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1450         } else {
1451                 /*
1452                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1453                  * return FID back in that case.
1454                  */
1455                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1456                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1457         }
1458         if (IS_ERR(child))
1459                 GOTO(out_parent, result = PTR_ERR(child));
1460
1461         /** check version of child  */
1462         rc = mdt_version_get_check(info, child, 1);
1463         if (rc)
1464                 GOTO(out_child, result = rc);
1465
1466         if (result == -ENOENT) {
1467                 /* Create under OBF and .lustre is not permitted */
1468                 if (!fid_is_md_operative(rr->rr_fid1))
1469                         GOTO(out_child, result = -EPERM);
1470
1471                 /* save versions in reply */
1472                 mdt_version_get_save(info, parent, 0);
1473                 mdt_version_get_save(info, child, 1);
1474
1475                 /* version of child will be changed */
1476                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1477
1478                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1479                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1480
1481                 /* Let lower layers know what is lock mode on directory. */
1482                 info->mti_spec.sp_cr_mode =
1483                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1484
1485                 /* Don't do lookup sanity check. We know name doesn't exist. */
1486                 info->mti_spec.sp_cr_lookup = 0;
1487                 info->mti_spec.sp_feat = &dt_directory_features;
1488
1489                 result = mdo_create(info->mti_env, mdt_object_child(parent),
1490                                     &rr->rr_name, mdt_object_child(child),
1491                                     &info->mti_spec, &info->mti_attr);
1492                 if (result == -ERESTART) {
1493                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1494                         GOTO(out_child, result);
1495                 } else {
1496                         mdt_prep_ma_buf_from_rep(info, child, ma);
1497                         /* XXX: we should call this once, see few lines below */
1498                         if (result == 0)
1499                                 result = mdt_attr_get_complex(info, child, ma);
1500
1501                         if (result != 0)
1502                                 GOTO(out_child, result);
1503                 }
1504                 created = 1;
1505                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
1506         } else {
1507                 /*
1508                  * The object is on remote node, return its FID for remote open.
1509                  */
1510                 if (mdt_object_remote(child)) {
1511                         /*
1512                          * Check if this lock already was sent to client and
1513                          * this is resent case. For resent case do not take lock
1514                          * again, use what is already granted.
1515                          */
1516                         LASSERT(lhc != NULL);
1517
1518                         rc = mdt_check_resent_lock(info, child, lhc);
1519                         if (rc < 0) {
1520                                 GOTO(out_child, result = rc);
1521                         } else if (rc > 0) {
1522                                 mdt_lock_handle_init(lhc);
1523                                 mdt_lock_reg_init(lhc, LCK_PR);
1524
1525                                 rc = mdt_object_lock(info, child, lhc,
1526                                                      MDS_INODELOCK_LOOKUP);
1527                         }
1528                         repbody->mbo_fid1 = *mdt_object_fid(child);
1529                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1530                         if (rc != 0)
1531                                 result = rc;
1532                         else
1533                                 result = -MDT_EREMOTE_OPEN;
1534                         GOTO(out_child, result);
1535                 } else if (mdt_object_exists(child)) {
1536                         /* We have to get attr & LOV EA & HSM for this
1537                          * object. */
1538                         mdt_prep_ma_buf_from_rep(info, child, ma);
1539                         ma->ma_need |= MA_HSM;
1540                         result = mdt_attr_get_complex(info, child, ma);
1541                         if (result != 0)
1542                                 GOTO(out_child, result);
1543                 } else {
1544                         /* Object does not exist. Likely FS corruption. */
1545                         CERROR("%s: name '"DNAME"' present, but FID "
1546                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1547                                PNAME(&rr->rr_name), PFID(child_fid));
1548                         GOTO(out_child, result = -EIO);
1549                 }
1550         }
1551
1552         rc = mdt_check_resent_lock(info, child, lhc);
1553         if (rc < 0) {
1554                 GOTO(out_child, result = rc);
1555         } else if (rc == 0) {
1556                 /* the open lock might already be gotten in
1557                  * ldlm_handle_enqueue() */
1558                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1559                 if (create_flags & MDS_OPEN_LOCK)
1560                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1561         } else {
1562                 /* get openlock if this isn't replay and client requested it */
1563                 if (!req_is_replay(req)) {
1564                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1565                         object_locked = 1;
1566                         if (rc != 0)
1567                                 GOTO(out_child_unlock, result = rc);
1568                         else if (create_flags & MDS_OPEN_LOCK)
1569                                 mdt_set_disposition(info, ldlm_rep,
1570                                                     DISP_OPEN_LOCK);
1571                 }
1572         }
1573         /* Try to open it now. */
1574         rc = mdt_finish_open(info, parent, child, create_flags,
1575                              created, ldlm_rep);
1576         if (rc) {
1577                 result = rc;
1578                 /* openlock will be released if mdt_finish_open failed */
1579                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1580
1581                 if (created && create_flags & MDS_OPEN_VOLATILE) {
1582                         CERROR("%s: cannot open volatile file "DFID", orphan "
1583                                "file will be left in PENDING directory until "
1584                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1585                                PFID(mdt_object_fid(child)), rc);
1586                         GOTO(out_child_unlock, result);
1587                 }
1588
1589                 if (created) {
1590                         ma->ma_need = 0;
1591                         ma->ma_valid = 0;
1592                         rc = mdo_unlink(info->mti_env,
1593                                         mdt_object_child(parent),
1594                                         mdt_object_child(child),
1595                                         &rr->rr_name,
1596                                         &info->mti_attr, 0);
1597                         if (rc != 0)
1598                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1599                                        mdt_obd_name(info->mti_mdt),
1600                                        PFID(mdt_object_fid(child)), rc);
1601                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1602                 }
1603         }
1604         EXIT;
1605 out_child_unlock:
1606         if (object_locked)
1607                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1608 out_child:
1609         mdt_object_put(info->mti_env, child);
1610         if (result == 0)
1611                 mdt_pack_size2body(info, child_fid, ibits & MDS_INODELOCK_DOM);
1612 out_parent:
1613         mdt_object_unlock_put(info, parent, lh, result || !created);
1614 out:
1615         if (result)
1616                 lustre_msg_set_transno(req->rq_repmsg, 0);
1617         return result;
1618 }
1619
1620 /**
1621  * Create an orphan object use local root.
1622  */
1623 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1624                                           struct mdt_device *mdt,
1625                                           const struct lu_fid *fid,
1626                                           struct md_attr *attr, fmode_t fmode)
1627 {
1628         const struct lu_env *env = info->mti_env;
1629         struct md_op_spec *spec = &info->mti_spec;
1630         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1631         struct mdt_object *obj = NULL;
1632         struct mdt_object *local_root;
1633         static const struct lu_name lname = {
1634                 .ln_name = "i_am_nobody",
1635                 .ln_namelen = sizeof("i_am_nobody") - 1,
1636         };
1637         struct lu_ucred *uc;
1638         cfs_cap_t uc_cap_save;
1639         int rc;
1640         ENTRY;
1641
1642         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1643         if (rc != 0)
1644                 RETURN(ERR_PTR(rc));
1645
1646         local_root = mdt_object_find(env, mdt, local_root_fid);
1647         if (IS_ERR(local_root))
1648                 RETURN(local_root);
1649
1650         obj = mdt_object_new(env, mdt, fid);
1651         if (IS_ERR(obj))
1652                 GOTO(out, rc = PTR_ERR(obj));
1653
1654         spec->sp_cr_lookup = 0;
1655         spec->sp_feat = &dt_directory_features;
1656         spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1657         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1658         if (attr->ma_valid & MA_LOV) {
1659                 spec->u.sp_ea.eadata = attr->ma_lmm;
1660                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1661                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1662         } else {
1663                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1664         }
1665
1666         uc = lu_ucred(env);
1667         uc_cap_save = uc->uc_cap;
1668         uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1669         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1670                         mdt_object_child(obj), spec, attr);
1671         uc->uc_cap = uc_cap_save;
1672         if (rc < 0) {
1673                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1674                        mdt_obd_name(mdt), PFID(fid), rc);
1675                 GOTO(out, rc);
1676         }
1677
1678         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1679         if (rc < 0)
1680                 CERROR("%s: cannot open volatile file "DFID", orphan "
1681                        "file will be left in PENDING directory until "
1682                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1683                        PFID(fid), rc);
1684         GOTO(out, rc);
1685
1686 out:
1687         if (rc < 0) {
1688                 if (!IS_ERR(obj))
1689                         mdt_object_put(env, obj);
1690                 obj = ERR_PTR(rc);
1691         }
1692         mdt_object_put(env, local_root);
1693         return obj;
1694 }
1695
1696 /* XXX Look into layout in MDT layer. */
1697 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1698 {
1699         struct lov_comp_md_v1   *comp_v1;
1700         struct lov_mds_md       *v1;
1701         __u32   off;
1702         int     i;
1703
1704         if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1705                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1706
1707                 if (comp_v1->lcm_entry_count == 0)
1708                         return -EINVAL;
1709
1710                 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1711                         off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1712                         v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1713                         v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1714                 }
1715         } else {
1716                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1717         }
1718         return 0;
1719 }
1720
1721 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1722                            struct md_attr *ma)
1723 {
1724         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1725         struct lu_ucred        *uc = mdt_ucred(info);
1726         struct close_data      *data;
1727         struct ldlm_lock       *lease;
1728         struct mdt_object      *orphan;
1729         struct md_attr         *orp_ma;
1730         struct lu_buf          *buf;
1731         cfs_cap_t               cap;
1732         bool                    lease_broken;
1733         int                     rc;
1734         int                     rc2;
1735         ENTRY;
1736
1737         if (mdt_rdonly(info->mti_exp))
1738                 RETURN(-EROFS);
1739
1740         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1741         if (data == NULL)
1742                 RETURN(-EPROTO);
1743
1744         lease = ldlm_handle2lock(&data->cd_handle);
1745         if (lease == NULL)
1746                 RETURN(-ESTALE);
1747
1748         /* try to hold open_sem so that nobody else can open the file */
1749         if (!down_write_trylock(&o->mot_open_sem)) {
1750                 ldlm_lock_cancel(lease);
1751                 GOTO(out_reprocess, rc = -EBUSY);
1752         }
1753
1754         /* Check if the lease open lease has already canceled */
1755         lock_res_and_lock(lease);
1756         lease_broken = ldlm_is_cancel(lease);
1757         unlock_res_and_lock(lease);
1758
1759         LDLM_DEBUG(lease, DFID " lease broken? %d",
1760                    PFID(mdt_object_fid(o)), lease_broken);
1761
1762         /* Cancel server side lease. Client side counterpart should
1763          * have been cancelled. It's okay to cancel it now as we've
1764          * held mot_open_sem. */
1765         ldlm_lock_cancel(lease);
1766
1767         if (lease_broken) /* don't perform release task */
1768                 GOTO(out_unlock, rc = -ESTALE);
1769
1770         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1771                 GOTO(out_unlock, rc = -EINVAL);
1772
1773         /* ma_need was set before but it seems fine to change it in order to
1774          * avoid modifying the one from RPC */
1775         ma->ma_need = MA_HSM;
1776         rc = mdt_attr_get_complex(info, o, ma);
1777         if (rc != 0)
1778                 GOTO(out_unlock, rc);
1779
1780         if (!mdt_hsm_release_allow(ma))
1781                 GOTO(out_unlock, rc = -EPERM);
1782
1783         /* already released? */
1784         if (ma->ma_hsm.mh_flags & HS_RELEASED)
1785                 GOTO(out_unlock, rc = 0);
1786
1787         /* Compare on-disk and packed data_version */
1788         if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1789                 CDEBUG(D_HSM, DFID" data_version mismatches: packed=%llu"
1790                        " and on-disk=%llu\n", PFID(mdt_object_fid(o)),
1791                        data->cd_data_version, ma->ma_hsm.mh_arch_ver);
1792                 GOTO(out_unlock, rc = -EPERM);
1793         }
1794
1795         ma->ma_valid = MA_INODE;
1796         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1797         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1798         if (rc < 0)
1799                 GOTO(out_unlock, rc);
1800
1801         ma->ma_need = MA_INODE | MA_LOV;
1802         rc = mdt_attr_get_complex(info, o, ma);
1803         if (rc < 0)
1804                 GOTO(out_unlock, rc);
1805
1806         if (!(ma->ma_valid & MA_LOV)) {
1807                 /* Even empty file are released */
1808                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1809                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
1810                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1811                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1812                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1813         } else {
1814                 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
1815                  * ma_lmm as lov_user_md, then it will be confused by union of
1816                  * layout_gen and stripe_offset. */
1817                 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
1818                     LOV_MAGIC_MAGIC)
1819                         ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1820                 else
1821                         GOTO(out_unlock, rc = -EINVAL);
1822         }
1823
1824         /* Set file as released. */
1825         rc = mdt_hsm_set_released(ma->ma_lmm);
1826         if (rc)
1827                 GOTO(out_unlock, rc);
1828
1829         orp_ma = &info->mti_u.hsm.attr;
1830         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1831         /* We use root ownership to bypass potential quota
1832          * restrictions on the user and group of the file. */
1833         orp_ma->ma_attr.la_uid = 0;
1834         orp_ma->ma_attr.la_gid = 0;
1835         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1836         orp_ma->ma_lmm = ma->ma_lmm;
1837         orp_ma->ma_lmm_size = ma->ma_lmm_size;
1838         orp_ma->ma_valid = MA_INODE | MA_LOV;
1839         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1840                                  FMODE_WRITE);
1841         if (IS_ERR(orphan)) {
1842                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1843                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1844                        PTR_ERR(orphan));
1845                 GOTO(out_unlock, rc = PTR_ERR(orphan));
1846         }
1847
1848         /* Set up HSM attribute for orphan object */
1849         CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
1850         buf = &info->mti_buf;
1851         buf->lb_buf = info->mti_xattr_buf;
1852         buf->lb_len = sizeof(struct hsm_attrs);
1853         ma->ma_hsm.mh_flags |= HS_RELEASED;
1854         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1855         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
1856
1857         mdt_lock_reg_init(lh, LCK_EX);
1858         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
1859                              MDS_INODELOCK_XATTR);
1860         if (rc != 0)
1861                 GOTO(out_close, rc);
1862
1863         /* The orphan has root ownership so we need to raise
1864          * CAP_FOWNER to set the HSM attributes. */
1865         cap = uc->uc_cap;
1866         uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
1867         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
1868                           XATTR_NAME_HSM, 0);
1869         uc->uc_cap = cap;
1870         if (rc != 0)
1871                 GOTO(out_layout_lock, rc);
1872
1873         /* Swap layout with orphan objects. */
1874         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
1875                              mdt_object_child(orphan),
1876                              SWAP_LAYOUTS_MDS_HSM);
1877         EXIT;
1878
1879 out_layout_lock:
1880         /* Release exclusive LL */
1881         mdt_object_unlock(info, o, lh, 1);
1882 out_close:
1883         /* Close orphan object anyway */
1884         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
1885                        FMODE_WRITE);
1886         if (rc2 < 0)
1887                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
1888                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
1889         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
1890                         "object closed");
1891         mdt_object_put(info->mti_env, orphan);
1892
1893 out_unlock:
1894         up_write(&o->mot_open_sem);
1895
1896         /* already released */
1897         if (rc == 0) {
1898                 struct mdt_body *repbody;
1899
1900                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1901                 LASSERT(repbody != NULL);
1902                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1903         }
1904
1905 out_reprocess:
1906         ldlm_reprocess_all(lease->l_resource);
1907         LDLM_LOCK_PUT(lease);
1908
1909         ma->ma_valid = 0;
1910         ma->ma_need = 0;
1911
1912         return rc;
1913 }
1914
1915 int mdt_close_handle_layouts(struct mdt_thread_info *info,
1916                              struct mdt_object *o, struct md_attr *ma)
1917 {
1918         struct mdt_lock_handle  *lh1 = &info->mti_lh[MDT_LH_NEW];
1919         struct mdt_lock_handle  *lh2 = &info->mti_lh[MDT_LH_OLD];
1920         struct close_data       *data;
1921         struct ldlm_lock        *lease;
1922         struct mdt_object       *o1 = o, *o2;
1923         bool                     lease_broken;
1924         bool                     swap_objects;
1925         int                      rc;
1926         ENTRY;
1927
1928         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1929                 RETURN(-EROFS);
1930
1931         if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
1932                 RETURN(-EINVAL);
1933
1934         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1935         if (data == NULL)
1936                 RETURN(-EPROTO);
1937
1938         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1939                 RETURN(-EINVAL);
1940
1941         rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
1942         if (unlikely(rc == 0))
1943                 RETURN(-EINVAL);
1944
1945         /* Exchange o1 and o2, to enforce locking order */
1946         swap_objects = (rc < 0);
1947
1948         lease = ldlm_handle2lock(&data->cd_handle);
1949         if (lease == NULL)
1950                 RETURN(-ESTALE);
1951
1952         o2 = mdt_object_find(info->mti_env, info->mti_mdt, &data->cd_fid);
1953         if (IS_ERR(o2))
1954                 GOTO(out_lease, rc = PTR_ERR(o2));
1955
1956         if (!S_ISREG(lu_object_attr(&o2->mot_obj))) {
1957                 swap_objects = false; /* not swapped yet */
1958                 GOTO(out_obj, rc = -EINVAL);
1959         }
1960
1961         if (swap_objects)
1962                 swap(o1, o2);
1963
1964         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1965                            MAY_WRITE);
1966         if (rc < 0)
1967                 GOTO(out_obj, rc);
1968
1969         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1970                            MAY_WRITE);
1971         if (rc < 0)
1972                 GOTO(out_obj, rc);
1973
1974         /* try to hold open_sem so that nobody else can open the file */
1975         if (!down_write_trylock(&o->mot_open_sem)) {
1976                 ldlm_lock_cancel(lease);
1977                 GOTO(out_obj, rc = -EBUSY);
1978         }
1979
1980         /* Check if the lease open lease has already canceled */
1981         lock_res_and_lock(lease);
1982         lease_broken = ldlm_is_cancel(lease);
1983         unlock_res_and_lock(lease);
1984
1985         LDLM_DEBUG(lease, DFID " lease broken? %d",
1986                    PFID(mdt_object_fid(o)), lease_broken);
1987
1988         /* Cancel server side lease. Client side counterpart should
1989          * have been cancelled. It's okay to cancel it now as we've
1990          * held mot_open_sem. */
1991         ldlm_lock_cancel(lease);
1992
1993         if (lease_broken)
1994                 GOTO(out_unlock_sem, rc = -ESTALE);
1995
1996         mdt_lock_reg_init(lh1, LCK_EX);
1997         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1998                              MDS_INODELOCK_XATTR);
1999         if (rc < 0)
2000                 GOTO(out_unlock_sem, rc);
2001
2002         mdt_lock_reg_init(lh2, LCK_EX);
2003         rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2004                              MDS_INODELOCK_XATTR);
2005         if (rc < 0)
2006                 GOTO(out_unlock1, rc);
2007
2008         /* Swap layout with orphan object */
2009         if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2010                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2011                                      mdt_object_child(o2), 0);
2012         } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2013                    ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2014                 struct lu_buf *buf = &info->mti_buf;
2015                 struct md_rejig_data mrd;
2016
2017                 mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2018                 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)
2019                         mrd.mrd_mirror_id = data->cd_mirror_id;
2020
2021                 buf->lb_len = sizeof(mrd);
2022                 buf->lb_buf = &mrd;
2023                 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2024                                   XATTR_LUSTRE_LOV,
2025                                   ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2026                                   LU_XATTR_SPLIT : LU_XATTR_MERGE);
2027                 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2028                         int rc2;
2029
2030                         rc2 = mdt_set_som(info, o, &ma->ma_attr);
2031                         if (rc2 < 0)
2032                                 CERROR(DFID": Setting i_blocks error: %d, "
2033                                        "i_blocks will be reported wrongly and "
2034                                        "can only be fixed in next resync\n",
2035                                        PFID(mdt_object_fid(o)), rc2);
2036                 }
2037         }
2038         if (rc < 0)
2039                 GOTO(out_unlock2, rc);
2040
2041         EXIT;
2042
2043 out_unlock2:
2044         /* Release exclusive LL */
2045         mdt_object_unlock(info, o2, lh2, 1);
2046
2047 out_unlock1:
2048         mdt_object_unlock(info, o1, lh1, 1);
2049
2050 out_unlock_sem:
2051         up_write(&o->mot_open_sem);
2052
2053         /* already swapped */
2054         if (rc == 0) {
2055                 struct mdt_body *repbody;
2056
2057                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2058                 LASSERT(repbody != NULL);
2059                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2060         }
2061
2062 out_obj:
2063         mdt_object_put(info->mti_env, swap_objects ? o1 : o2);
2064
2065         ldlm_reprocess_all(lease->l_resource);
2066
2067 out_lease:
2068         LDLM_LOCK_PUT(lease);
2069
2070         if (ma != NULL) {
2071                 ma->ma_valid = 0;
2072                 ma->ma_need = 0;
2073         }
2074
2075         return rc;
2076 }
2077
2078 static int mdt_close_resync_done(struct mdt_thread_info *info,
2079                                  struct mdt_object *o, struct md_attr *ma)
2080 {
2081         struct close_data       *data;
2082         struct ldlm_lock        *lease;
2083         struct md_layout_change  layout = { 0 };
2084         __u32                   *resync_ids = NULL;
2085         size_t                   resync_count = 0;
2086         bool                     lease_broken;
2087         int                      rc;
2088         ENTRY;
2089
2090         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2091                 RETURN(-EROFS);
2092
2093         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2094                 RETURN(-EINVAL);
2095
2096         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2097         if (data == NULL)
2098                 RETURN(-EPROTO);
2099
2100         if (ptlrpc_req_need_swab(mdt_info_req(info)))
2101                 lustre_swab_close_data_resync_done(&data->cd_resync);
2102
2103         if (!fid_is_zero(&data->cd_fid))
2104                 RETURN(-EPROTO);
2105
2106         lease = ldlm_handle2lock(&data->cd_handle);
2107         if (lease == NULL)
2108                 RETURN(-ESTALE);
2109
2110         /* try to hold open_sem so that nobody else can open the file */
2111         if (!down_write_trylock(&o->mot_open_sem)) {
2112                 ldlm_lock_cancel(lease);
2113                 GOTO(out_reprocess, rc = -EBUSY);
2114         }
2115
2116         /* Check if the lease open lease has already canceled */
2117         lock_res_and_lock(lease);
2118         lease_broken = ldlm_is_cancel(lease);
2119         unlock_res_and_lock(lease);
2120
2121         LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2122                    PFID(mdt_object_fid(o)), lease_broken);
2123
2124         /* Cancel server side lease. Client side counterpart should
2125          * have been cancelled. It's okay to cancel it now as we've
2126          * held mot_open_sem. */
2127         ldlm_lock_cancel(lease);
2128
2129         if (lease_broken) /* don't perform release task */
2130                 GOTO(out_unlock, rc = -ESTALE);
2131
2132         resync_count = data->cd_resync.resync_count;
2133
2134         if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2135                 void *data;
2136
2137                 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2138                                            RCL_CLIENT))
2139                         GOTO(out_unlock, rc = -EPROTO);
2140
2141                 OBD_ALLOC(resync_ids, resync_count * sizeof(__u32));
2142                 if (!resync_ids)
2143                         GOTO(out_unlock, rc = -ENOMEM);
2144
2145                 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2146                 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2147
2148                 layout.mlc_resync_ids = resync_ids;
2149         } else {
2150                 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2151         }
2152
2153         layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2154         layout.mlc_resync_count = resync_count;
2155         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2156                 layout.mlc_som.lsa_valid = LSOM_FL_VALID;
2157                 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2158                 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2159         }
2160         rc = mdt_layout_change(info, o, &layout);
2161         if (rc)
2162                 GOTO(out_unlock, rc);
2163
2164         EXIT;
2165
2166 out_unlock:
2167         up_write(&o->mot_open_sem);
2168
2169         /* already released */
2170         if (rc == 0) {
2171                 struct mdt_body *repbody;
2172
2173                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2174                 LASSERT(repbody != NULL);
2175                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2176         }
2177
2178         if (resync_ids)
2179                 OBD_FREE(resync_ids, resync_count * sizeof(__u32));
2180
2181 out_reprocess:
2182         ldlm_reprocess_all(lease->l_resource);
2183         LDLM_LOCK_PUT(lease);
2184
2185         ma->ma_valid = 0;
2186         ma->ma_need = 0;
2187
2188         return rc;
2189 }
2190
2191 #define MFD_CLOSED(mode) ((mode) == MDS_FMODE_CLOSED)
2192 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2193 {
2194         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
2195 }
2196
2197 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2198 {
2199         struct mdt_object *o = mfd->mfd_object;
2200         struct md_object *next = mdt_object_child(o);
2201         struct md_attr *ma = &info->mti_attr;
2202         int rc = 0;
2203         __u64 mode;
2204         __u64 intent;
2205         ENTRY;
2206
2207         mode = mfd->mfd_mode;
2208
2209         intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2210
2211         CDEBUG(D_INODE, "%s: close file "DFID" with intent: %llx\n",
2212                mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(o)), intent);
2213
2214         switch (intent) {
2215         case MDS_HSM_RELEASE: {
2216                 rc = mdt_hsm_release(info, o, ma);
2217                 if (rc < 0) {
2218                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2219                                mdt_obd_name(info->mti_mdt),
2220                                PFID(mdt_object_fid(o)), rc);
2221                         /* continue to close even error occurred. */
2222                 }
2223                 break;
2224         }
2225         case MDS_CLOSE_LAYOUT_MERGE:
2226         case MDS_CLOSE_LAYOUT_SPLIT:
2227         case MDS_CLOSE_LAYOUT_SWAP: {
2228                 rc = mdt_close_handle_layouts(info, o, ma);
2229                 if (rc < 0) {
2230                         CDEBUG(D_INODE,
2231                                "%s: cannot swap layout of "DFID": rc=%d\n",
2232                                mdt_obd_name(info->mti_mdt),
2233                                PFID(mdt_object_fid(o)), rc);
2234                         /* continue to close even if error occurred. */
2235                 }
2236                 break;
2237         }
2238         case MDS_CLOSE_RESYNC_DONE:
2239                 rc = mdt_close_resync_done(info, o, ma);
2240                 break;
2241         default:
2242                 /* nothing */
2243                 break;
2244         }
2245
2246         if (mode & FMODE_WRITE)
2247                 mdt_write_put(o);
2248         else if (mode & MDS_FMODE_EXEC)
2249                 mdt_write_allow(o);
2250
2251         /* Update atime on close only. */
2252         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
2253             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
2254                 /* Set the atime only. */
2255                 ma->ma_valid = MA_INODE;
2256                 ma->ma_attr.la_valid = LA_ATIME;
2257                 rc = mo_attr_set(info->mti_env, next, ma);
2258         }
2259
2260         /* If file data is modified, add the dirty flag. */
2261         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2262                 rc = mdt_add_dirty_flag(info, o, ma);
2263
2264         ma->ma_need |= MA_INODE;
2265         ma->ma_valid &= ~MA_INODE;
2266
2267         LASSERT(atomic_read(&o->mot_open_count) > 0);
2268         atomic_dec(&o->mot_open_count);
2269         mdt_handle_last_unlink(info, o, ma);
2270
2271         if (!MFD_CLOSED(mode)) {
2272                 rc = mo_close(info->mti_env, next, ma, mode);
2273                 mdt_dom_check_and_discard(info, o);
2274         }
2275
2276         /* adjust open and lease count */
2277         if (mode & MDS_OPEN_LEASE) {
2278                 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2279                 atomic_dec(&o->mot_lease_count);
2280         }
2281
2282         mdt_mfd_free(mfd);
2283         mdt_object_put(info->mti_env, o);
2284
2285         RETURN(rc);
2286 }
2287
2288 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2289                        struct mdt_body *repbody)
2290 {
2291         struct mdt_export_data *med;
2292         struct mdt_file_data   *mfd;
2293         int                     ret = 0;
2294         int                     rc = 0;
2295         ENTRY;
2296
2297         med = &req->rq_export->exp_mdt_data;
2298         spin_lock(&med->med_open_lock);
2299         mfd = mdt_handle2mfd(med, &info->mti_close_handle, req_is_replay(req));
2300         if (mdt_mfd_closed(mfd)) {
2301                 spin_unlock(&med->med_open_lock);
2302                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2303                        ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2304                        info->mti_close_handle.cookie);
2305                 /** not serious error since bug 3633 */
2306                 rc = -ESTALE;
2307         } else {
2308                 class_handle_unhash(&mfd->mfd_handle);
2309                 list_del_init(&mfd->mfd_list);
2310                 spin_unlock(&med->med_open_lock);
2311                 ret = mdt_mfd_close(info, mfd);
2312         }
2313
2314         RETURN(rc ? rc : ret);
2315 }
2316
2317 int mdt_close(struct tgt_session_info *tsi)
2318 {
2319         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2320         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2321         struct md_attr         *ma = &info->mti_attr;
2322         struct mdt_body        *repbody = NULL;
2323         int rc, ret = 0;
2324         ENTRY;
2325
2326         mdt_counter_incr(req, LPROC_MDT_CLOSE);
2327         /* Close may come with the Size-on-MDS update. Unpack it. */
2328         rc = mdt_close_unpack(info);
2329         if (rc)
2330                 GOTO(out, rc = err_serious(rc));
2331
2332         /* These fields are no longer used and are left for compatibility.
2333          * size is always zero */
2334         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2335                              0);
2336         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2337                              0);
2338         rc = req_capsule_server_pack(info->mti_pill);
2339         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2340                 mdt_client_compatibility(info);
2341                 if (rc == 0)
2342                         mdt_fix_reply(info);
2343                 mdt_exit_ucred(info);
2344                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2345         }
2346
2347         /* Continue to close handle even if we can not pack reply */
2348         if (rc == 0) {
2349                 repbody = req_capsule_server_get(info->mti_pill,
2350                                                  &RMF_MDT_BODY);
2351                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2352                                                     &RMF_MDT_MD);
2353                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2354                                                        &RMF_MDT_MD,
2355                                                        RCL_SERVER);
2356                 ma->ma_need = MA_INODE | MA_LOV;
2357                 repbody->mbo_eadatasize = 0;
2358                 repbody->mbo_aclsize = 0;
2359         } else {
2360                 rc = err_serious(rc);
2361         }
2362
2363         rc = mdt_close_internal(info, req, repbody);
2364         if (rc != -ESTALE)
2365                 mdt_empty_transno(info, rc);
2366
2367         if (repbody != NULL) {
2368                 mdt_client_compatibility(info);
2369                 rc = mdt_fix_reply(info);
2370         }
2371
2372         mdt_exit_ucred(info);
2373         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2374                 GOTO(out, rc = err_serious(-ENOMEM));
2375
2376         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2377                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2378                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2379 out:
2380         mdt_thread_info_fini(info);
2381         RETURN(rc ? rc : ret);
2382 }