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 static const char mfd_open_handle_owner[] = "mdt";
  48
  49 /* Create a new mdt_file_data struct, initialize it,
  50  * and insert it to global hash table */
  51 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
  52 {
  53         struct mdt_file_data *mfd;
  54         ENTRY;
  55
  56         OBD_ALLOC_PTR(mfd);
  57         if (mfd != NULL) {
  58                 refcount_set(&mfd->mfd_open_handle.h_ref, 1);
  59                 INIT_HLIST_NODE(&mfd->mfd_open_handle.h_link);
  60                 mfd->mfd_owner = med;
  61                 INIT_LIST_HEAD(&mfd->mfd_list);
  62                 class_handle_hash(&mfd->mfd_open_handle, mfd_open_handle_owner);
  63         }
  64
  65         RETURN(mfd);
  66 }
  67
  68 /*
  69  * Find the mfd pointed to by handle in global hash table.
  70  * In case of replay the handle is obsoleted
  71  * but mfd can be found in mfd list by that handle.
  72  * Callers need to be holding med_open_lock.
  73  */
  74 struct mdt_file_data *mdt_open_handle2mfd(struct mdt_export_data *med,
  75                                         const struct lustre_handle *open_handle,
  76                                         bool is_replay_or_resent)
  77 {
  78         struct mdt_file_data   *mfd;
  79         ENTRY;
  80
  81         LASSERT(open_handle != NULL);
  82         mfd = class_handle2object(open_handle->cookie, mfd_open_handle_owner);
  83         if (mfd)
  84                 refcount_dec(&mfd->mfd_open_handle.h_ref);
  85
  86         /* during dw/setattr replay the mfd can be found by old handle */
  87         if ((!mfd || mfd->mfd_owner != med) && is_replay_or_resent) {
  88                 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  89                         if (mfd->mfd_open_handle_old.cookie ==
  90                             open_handle->cookie)
  91                                 RETURN(mfd);
  92                 }
  93                 mfd = NULL;
  94         }
  95
  96         RETURN(mfd);
  97 }
  98
  99 /* free mfd */
 100 void mdt_mfd_free(struct mdt_file_data *mfd)
 101 {
 102         LASSERT(refcount_read(&mfd->mfd_open_handle.h_ref) == 1);
 103         LASSERT(list_empty(&mfd->mfd_list));
 104         OBD_FREE_PRE(mfd, sizeof(*mfd), "rcu");
 105         kfree_rcu(mfd, mfd_open_handle.h_rcu);
 106 }
 107
 108 static int mdt_create_data(struct mdt_thread_info *info,
 109                            struct mdt_object *p, struct mdt_object *o)
 110 {
 111         struct md_op_spec     *spec = &info->mti_spec;
 112         struct md_attr        *ma   = &info->mti_attr;
 113         int                    rc   = 0;
 114         ENTRY;
 115
 116         if (!md_should_create(spec->sp_cr_flags))
 117                 RETURN(0);
 118
 119         ma->ma_need = MA_INODE | MA_LOV;
 120         ma->ma_valid = 0;
 121         mutex_lock(&o->mot_lov_mutex);
 122         if (!o->mot_lov_created) {
 123                 rc = mdo_create_data(info->mti_env,
 124                                      p ? mdt_object_child(p) : NULL,
 125                                      mdt_object_child(o), spec, ma);
 126                 if (rc == 0)
 127                         rc = mdt_attr_get_complex(info, o, ma);
 128
 129                 if (rc == 0 && ma->ma_valid & MA_LOV)
 130                         o->mot_lov_created = 1;
 131         }
 132
 133         mutex_unlock(&o->mot_lov_mutex);
 134         RETURN(rc);
 135 }
 136
 137 int mdt_write_read(struct mdt_object *o)
 138 {
 139         int rc = 0;
 140         ENTRY;
 141         spin_lock(&o->mot_write_lock);
 142         rc = o->mot_write_count;
 143         spin_unlock(&o->mot_write_lock);
 144         RETURN(rc);
 145 }
 146
 147 int mdt_write_get(struct mdt_object *o)
 148 {
 149         int rc = 0;
 150         ENTRY;
 151         spin_lock(&o->mot_write_lock);
 152         if (o->mot_write_count < 0)
 153                 rc = -ETXTBSY;
 154         else
 155                 o->mot_write_count++;
 156         spin_unlock(&o->mot_write_lock);
 157
 158         RETURN(rc);
 159 }
 160
 161 void mdt_write_put(struct mdt_object *o)
 162 {
 163         ENTRY;
 164         spin_lock(&o->mot_write_lock);
 165         o->mot_write_count--;
 166         spin_unlock(&o->mot_write_lock);
 167         EXIT;
 168 }
 169
 170 static int mdt_write_deny(struct mdt_object *o)
 171 {
 172         int rc = 0;
 173         ENTRY;
 174         spin_lock(&o->mot_write_lock);
 175         if (o->mot_write_count > 0)
 176                 rc = -ETXTBSY;
 177         else
 178                 o->mot_write_count--;
 179         spin_unlock(&o->mot_write_lock);
 180         RETURN(rc);
 181 }
 182
 183 static void mdt_write_allow(struct mdt_object *o)
 184 {
 185         ENTRY;
 186         spin_lock(&o->mot_write_lock);
 187         o->mot_write_count++;
 188         spin_unlock(&o->mot_write_lock);
 189         EXIT;
 190 }
 191
 192 /* there can be no real transaction so prepare the fake one */
 193 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 194 {
 195         struct mdt_device      *mdt = info->mti_mdt;
 196         struct ptlrpc_request  *req = mdt_info_req(info);
 197         struct tg_export_data  *ted;
 198         struct lsd_client_data *lcd;
 199         ENTRY;
 200
 201         if (mdt_rdonly(req->rq_export))
 202                 RETURN_EXIT;
 203
 204         /* transaction has occurred already */
 205         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 206                 RETURN_EXIT;
 207
 208         if (tgt_is_multimodrpcs_client(req->rq_export)) {
 209                 struct thandle         *th;
 210
 211                 /* generate an empty transaction to get a transno
 212                  * and reply data */
 213                 th = dt_trans_create(info->mti_env, mdt->mdt_bottom);
 214                 if (!IS_ERR(th)) {
 215                         rc = dt_trans_start(info->mti_env, mdt->mdt_bottom, th);
 216                         dt_trans_stop(info->mti_env, mdt->mdt_bottom, th);
 217                 }
 218                 RETURN_EXIT;
 219         }
 220
 221         spin_lock(&mdt->mdt_lut.lut_translock);
 222         if (rc != 0) {
 223                 if (info->mti_transno != 0) {
 224                         struct obd_export *exp = req->rq_export;
 225
 226                         CERROR("%s: replay trans %llu NID %s: rc = %d\n",
 227                                mdt_obd_name(mdt), info->mti_transno,
 228                                obd_export_nid2str(exp), rc);
 229                         spin_unlock(&mdt->mdt_lut.lut_translock);
 230                         RETURN_EXIT;
 231                 }
 232         } else if (info->mti_transno == 0) {
 233                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
 234         } else {
 235                 /* should be replay */
 236                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 237                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 238         }
 239         spin_unlock(&mdt->mdt_lut.lut_translock);
 240
 241         CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
 242                info->mti_transno,
 243                req->rq_export->exp_obd->obd_last_committed);
 244
 245         req->rq_transno = info->mti_transno;
 246         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 247
 248         /* update lcd in memory only for resent cases */
 249         ted = &req->rq_export->exp_target_data;
 250         LASSERT(ted);
 251         mutex_lock(&ted->ted_lcd_lock);
 252         lcd = ted->ted_lcd;
 253         if (info->mti_transno < lcd->lcd_last_transno &&
 254             info->mti_transno != 0) {
 255                 /* This should happen during replay. Do not update
 256                  * last rcvd info if replay req transno < last transno,
 257                  * otherwise the following resend(after replay) can not
 258                  * be checked correctly by xid */
 259                 mutex_unlock(&ted->ted_lcd_lock);
 260                 CDEBUG(D_HA, "%s: transno = %llu < last_transno = %llu\n",
 261                        mdt_obd_name(mdt), info->mti_transno,
 262                        lcd->lcd_last_transno);
 263                 RETURN_EXIT;
 264         }
 265
 266         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
 267                 if (info->mti_transno != 0)
 268                         lcd->lcd_last_close_transno = info->mti_transno;
 269                 lcd->lcd_last_close_xid = req->rq_xid;
 270                 lcd->lcd_last_close_result = rc;
 271         } else {
 272                 /* VBR: save versions in last_rcvd for reconstruct. */
 273                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 274                 if (pre_versions) {
 275                         lcd->lcd_pre_versions[0] = pre_versions[0];
 276                         lcd->lcd_pre_versions[1] = pre_versions[1];
 277                         lcd->lcd_pre_versions[2] = pre_versions[2];
 278                         lcd->lcd_pre_versions[3] = pre_versions[3];
 279                 }
 280                 if (info->mti_transno != 0)
 281                         lcd->lcd_last_transno = info->mti_transno;
 282
 283                 lcd->lcd_last_xid = req->rq_xid;
 284                 lcd->lcd_last_result = rc;
 285                 lcd->lcd_last_data = info->mti_opdata;
 286         }
 287         mutex_unlock(&ted->ted_lcd_lock);
 288
 289         EXIT;
 290 }
 291
 292 void mdt_mfd_set_mode(struct mdt_file_data *mfd, u64 open_flags)
 293 {
 294         LASSERT(mfd != NULL);
 295
 296         CDEBUG(D_DENTRY, DFID " Change mfd open_flags %#llo -> %#llo.\n",
 297                PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_open_flags,
 298                open_flags);
 299
 300         mfd->mfd_open_flags = open_flags;
 301 }
 302
 303 /**
 304  * prep ma_lmm/ma_lmv for md_attr from reply
 305  */
 306 static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
 307                                      struct mdt_object *obj,
 308                                      struct md_attr *ma)
 309 {
 310         if (ma->ma_lmv || ma->ma_lmm) {
 311                 CDEBUG(D_INFO, DFID " %s already set.\n",
 312                        PFID(mdt_object_fid(obj)),
 313                        ma->ma_lmv ? (ma->ma_lmm ? "ma_lmv and ma_lmm"
 314                                                 : "ma_lmv")
 315                                   : "ma_lmm");
 316                 return;
 317         }
 318
 319         if (S_ISDIR(obj->mot_header.loh_attr)) {
 320                 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
 321                                                     &RMF_MDT_MD);
 322                 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
 323                                                        &RMF_MDT_MD,
 324                                                        RCL_SERVER);
 325                 if (ma->ma_lmv_size > 0)
 326                         ma->ma_need |= MA_LMV;
 327         } else {
 328                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
 329                                                     &RMF_MDT_MD);
 330                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 331                                                        &RMF_MDT_MD,
 332                                                        RCL_SERVER);
 333                 if (ma->ma_lmm_size > 0)
 334                         ma->ma_need |= MA_LOV;
 335         }
 336 }
 337
 338 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 339                         struct mdt_object *o, u64 open_flags, int created,
 340                         struct ldlm_reply *rep)
 341 {
 342         struct ptlrpc_request *req = mdt_info_req(info);
 343         struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 344         struct mdt_file_data *mfd;
 345         struct md_attr *ma  = &info->mti_attr;
 346         struct lu_attr *la  = &ma->ma_attr;
 347         struct mdt_body *repbody;
 348         bool isdir, isreg;
 349         int rc = 0;
 350
 351         ENTRY;
 352         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 353
 354         isreg = S_ISREG(la->la_mode);
 355         isdir = S_ISDIR(la->la_mode);
 356         if (isreg && !(ma->ma_valid & MA_LOV) &&
 357             !(open_flags & MDS_OPEN_RELEASE)) {
 358                 /*
 359                  * No EA, check whether it is will set regEA and dirEA since in
 360                  * above attr get, these size might be zero, so reset it, to
 361                  * retrieve the MD after create obj.
 362                  */
 363                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 364                                                        &RMF_MDT_MD,
 365                                                        RCL_SERVER);
 366                 /* in replay case, p == NULL */
 367                 rc = mdt_create_data(info, p, o);
 368                 if (rc)
 369                         RETURN(rc);
 370
 371                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
 372                         mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
 373         }
 374
 375         CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n",
 376                ma->ma_valid, ma->ma_lmm_size);
 377
 378         if (ma->ma_valid & MA_LOV) {
 379                 LASSERT(ma->ma_lmm_size != 0);
 380                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 381                 if (isdir)
 382                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 383                 else
 384                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 385         }
 386
 387         if (ma->ma_valid & MA_LMV) {
 388                 LASSERT(ma->ma_lmv_size != 0);
 389                 repbody->mbo_eadatasize = ma->ma_lmv_size;
 390                 LASSERT(isdir);
 391                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
 392         }
 393
 394         if (open_flags & MDS_FMODE_WRITE)
 395                 rc = mdt_write_get(o);
 396         else if (open_flags & MDS_FMODE_EXEC)
 397                 rc = mdt_write_deny(o);
 398
 399         if (rc)
 400                 RETURN(rc);
 401
 402         rc = mo_open(info->mti_env, mdt_object_child(o),
 403                      created ? open_flags | MDS_OPEN_CREATED : open_flags,
 404                      &info->mti_spec);
 405         if (rc != 0) {
 406                 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
 407                  * client does not know which suppgid should be sent to the MDS,
 408                  * or some other(s) changed the target file's GID after this RPC
 409                  * sent to the MDS with the suppgid as the original GID, then we
 410                  * give the client another chance to send the right suppgid. */
 411                 if (rc == -EACCES &&
 412                     allow_client_chgrp(info, lu_ucred(info->mti_env)))
 413                         mdt_set_disposition(info, rep, DISP_OPEN_DENY);
 414
 415                 GOTO(err_out, rc);
 416         }
 417
 418         mfd = mdt_mfd_new(med);
 419         if (mfd == NULL)
 420                 GOTO(err_out, rc = -ENOMEM);
 421
 422         /*
 423          * Keep a reference on this object for this open, and is
 424          * released by mdt_mfd_close().
 425          */
 426         mdt_object_get(info->mti_env, o);
 427         mfd->mfd_object = o;
 428         mfd->mfd_xid = req->rq_xid;
 429
 430         /*
 431          * @open_flags is always not zero. At least it should be FMODE_READ,
 432          * FMODE_WRITE or MDS_FMODE_EXEC.
 433          */
 434         LASSERT(open_flags != 0);
 435
 436         /* Open handling. */
 437         mdt_mfd_set_mode(mfd, open_flags);
 438
 439         atomic_inc(&o->mot_open_count);
 440         if (open_flags & MDS_OPEN_LEASE)
 441                 atomic_inc(&o->mot_lease_count);
 442
 443         /* replay handle */
 444         if (req_is_replay(req)) {
 445                 struct mdt_file_data *old_mfd;
 446                 /* Check wheather old cookie already exist in
 447                  * the list, becasue when do recovery, client
 448                  * might be disconnected from server, and
 449                  * restart replay, so there maybe some orphan
 450                  * mfd here, we should remove them */
 451                 LASSERT(info->mti_rr.rr_open_handle != NULL);
 452                 spin_lock(&med->med_open_lock);
 453                 old_mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 454                                               true);
 455                 if (old_mfd != NULL) {
 456                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 457                                "cookie = %#llx\n", mfd,
 458                                PFID(mdt_object_fid(mfd->mfd_object)),
 459                                info->mti_rr.rr_open_handle->cookie);
 460                         class_handle_unhash(&old_mfd->mfd_open_handle);
 461                         list_del_init(&old_mfd->mfd_list);
 462                         spin_unlock(&med->med_open_lock);
 463                         /* no attr update for that close */
 464                         la->la_valid = 0;
 465                         ma->ma_valid |= MA_FLAGS;
 466                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 467                         mdt_mfd_close(info, old_mfd);
 468                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 469                         ma->ma_valid &= ~MA_FLAGS;
 470                 } else {
 471                         spin_unlock(&med->med_open_lock);
 472                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 473                                "cookie = %#llx\n",
 474                                PFID(mdt_object_fid(mfd->mfd_object)),
 475                                info->mti_rr.rr_open_handle->cookie);
 476                 }
 477
 478                 CDEBUG(D_HA, "Store old cookie %#llx in new mfd\n",
 479                        info->mti_rr.rr_open_handle->cookie);
 480
 481                 mfd->mfd_open_handle_old = *info->mti_rr.rr_open_handle;
 482         }
 483
 484         repbody->mbo_open_handle.cookie = mfd->mfd_open_handle.h_cookie;
 485
 486         if (req->rq_export->exp_disconnected) {
 487                 spin_lock(&med->med_open_lock);
 488                 class_handle_unhash(&mfd->mfd_open_handle);
 489                 list_del_init(&mfd->mfd_list);
 490                 spin_unlock(&med->med_open_lock);
 491                 mdt_mfd_close(info, mfd);
 492         } else {
 493                 spin_lock(&med->med_open_lock);
 494                 list_add_tail(&mfd->mfd_list, &med->med_open_head);
 495                 spin_unlock(&med->med_open_lock);
 496         }
 497
 498         mdt_empty_transno(info, rc);
 499
 500         RETURN(rc);
 501
 502 err_out:
 503         if (open_flags & MDS_FMODE_WRITE)
 504                 mdt_write_put(o);
 505         else if (open_flags & MDS_FMODE_EXEC)
 506                 mdt_write_allow(o);
 507
 508         return rc;
 509 }
 510
 511 static int mdt_finish_open(struct mdt_thread_info *info,
 512                            struct mdt_object *p, struct mdt_object *o,
 513                            u64 open_flags, int created,
 514                            struct ldlm_reply *rep)
 515 {
 516         struct ptlrpc_request   *req = mdt_info_req(info);
 517         struct obd_export       *exp = req->rq_export;
 518         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 519         struct md_attr          *ma  = &info->mti_attr;
 520         struct lu_attr          *la  = &ma->ma_attr;
 521         struct mdt_file_data    *mfd;
 522         struct mdt_body         *repbody;
 523         int                      rc = 0;
 524         int                      isreg, isdir, islnk;
 525         struct list_head        *t;
 526         ENTRY;
 527
 528         LASSERT(ma->ma_valid & MA_INODE);
 529
 530         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 531
 532         isreg = S_ISREG(la->la_mode);
 533         isdir = S_ISDIR(la->la_mode);
 534         islnk = S_ISLNK(la->la_mode);
 535         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 536
 537         /* compatibility check for 2.10 clients when it tries to open mirrored
 538          * files. 2.10 clients don't verify overlapping components so they
 539          * would read and write mirrored files just as if they were normal
 540          * PFL files, which will cause the problem that sycned mirrors actually
 541          * contain different data.
 542          * Older clients are not a concern here because they don't even
 543          * understand PFL layout. */
 544         if (isreg && !exp_connect_flr(exp) && ma->ma_valid & MA_LOV &&
 545             mdt_lmm_is_flr(ma->ma_lmm)) {
 546                 /* LU-10286: for simplicity clients who don't understand
 547                  * mirrored layout(with connect flag OBD_CONNECT2_FLR) won't
 548                  * be able to open mirrored files */
 549                 RETURN(-EOPNOTSUPP);
 550         }
 551
 552         /* Overstriped files can crash older clients */
 553         if (isreg && !exp_connect_overstriping(exp) &&
 554             mdt_lmm_is_overstriping(ma->ma_lmm))
 555                 RETURN(-EOPNOTSUPP);
 556
 557         /* LU-2275, simulate broken behaviour (esp. prevalent in
 558          * pre-2.4 servers where a very strange reply is sent on error
 559          * that looks like it was actually almost successful and a
 560          * failure at the same time.) */
 561         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 562                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 563                                                DISP_LOOKUP_NEG |
 564                                                DISP_LOOKUP_POS);
 565
 566                 if (open_flags & MDS_OPEN_LOCK)
 567                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 568
 569                 RETURN(-ENOENT);
 570         }
 571
 572 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
 573         if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 574                 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
 575                 if (IS_ERR(nodemap))
 576                         RETURN(PTR_ERR(nodemap));
 577
 578                 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
 579                 nodemap_putref(nodemap);
 580                 if (rc)
 581                         RETURN(rc);
 582         }
 583 #endif
 584
 585         /*
 586          * If we are following a symlink, don't open; and do not return open
 587          * handle for special nodes as client required.
 588          */
 589         if (islnk || (!isreg && !isdir &&
 590             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 591                 lustre_msg_set_transno(req->rq_repmsg, 0);
 592                 RETURN(0);
 593         }
 594
 595         /*
 596          * We need to return the existing object's fid back, so it is done here,
 597          * after preparing the reply.
 598          */
 599         if (!created && (open_flags & MDS_OPEN_EXCL) &&
 600             (open_flags & MDS_OPEN_CREAT))
 601                 RETURN(-EEXIST);
 602
 603         /* This can't be done earlier, we need to return reply body */
 604         if (isdir) {
 605                 if (open_flags & (MDS_OPEN_CREAT | MDS_FMODE_WRITE)) {
 606                         /* We are trying to create or write an existing dir. */
 607                         RETURN(-EISDIR);
 608                 }
 609         } else if (open_flags & MDS_OPEN_DIRECTORY)
 610                 RETURN(-ENOTDIR);
 611
 612         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 613                                  OBD_FAIL_MDS_LDLM_REPLY_NET | OBD_FAIL_ONCE))
 614                 RETURN(-EAGAIN);
 615
 616         mfd = NULL;
 617         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 618                 spin_lock(&med->med_open_lock);
 619                 list_for_each(t, &med->med_open_head) {
 620                         mfd = list_entry(t, struct mdt_file_data, mfd_list);
 621                         if (mfd->mfd_xid == req->rq_xid)
 622                                 break;
 623                         mfd = NULL;
 624                 }
 625                 spin_unlock(&med->med_open_lock);
 626
 627                 if (mfd != NULL) {
 628                         repbody->mbo_open_handle.cookie =
 629                                 mfd->mfd_open_handle.h_cookie;
 630                         /* set repbody->ea_size for resent case */
 631                         if (ma->ma_valid & MA_LOV) {
 632                                 LASSERT(ma->ma_lmm_size != 0);
 633                                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 634                                 if (isdir)
 635                                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 636                                 else
 637                                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 638                         }
 639                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 640                         RETURN(0);
 641                 }
 642         }
 643
 644         rc = mdt_mfd_open(info, p, o, open_flags, created, rep);
 645         if (!rc)
 646                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 647
 648         RETURN(rc);
 649 }
 650
 651 void mdt_reconstruct_open(struct mdt_thread_info *info,
 652                           struct mdt_lock_handle *lhc)
 653 {
 654         const struct lu_env *env = info->mti_env;
 655         struct mdt_device *mdt = info->mti_mdt;
 656         struct req_capsule *pill = info->mti_pill;
 657         struct ptlrpc_request *req = mdt_info_req(info);
 658         struct md_attr *ma = &info->mti_attr;
 659         struct mdt_reint_record *rr = &info->mti_rr;
 660         u64 open_flags = info->mti_spec.sp_cr_flags;
 661         struct ldlm_reply *ldlm_rep;
 662         struct mdt_object *parent;
 663         struct mdt_object *child;
 664         struct mdt_body *repbody;
 665         u64 opdata;
 666         int rc;
 667         ENTRY;
 668
 669         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 670         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 671         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 672
 673         ma->ma_need = MA_INODE | MA_HSM;
 674         ma->ma_valid = 0;
 675
 676         opdata = mdt_req_from_lrd(req, info->mti_reply_data);
 677         mdt_set_disposition(info, ldlm_rep, opdata);
 678
 679         CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n",
 680                ldlm_rep->lock_policy_res1, req->rq_status);
 681
 682         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
 683             req->rq_status != 0)
 684                 /* We did not create successfully, return error to client. */
 685                 GOTO(out, rc = req->rq_status);
 686
 687         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
 688                 struct obd_export *exp = req->rq_export;
 689                 /*
 690                  * We failed after creation, but we do not know in which step
 691                  * we failed. So try to check the child object.
 692                  */
 693                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
 694                 if (IS_ERR(parent)) {
 695                         rc = PTR_ERR(parent);
 696                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
 697                                       " Evicting client %s with export %s.\n",
 698                                       PFID(rr->rr_fid1), rc,
 699                                       obd_uuid2str(&exp->exp_client_uuid),
 700                                       obd_export_nid2str(exp));
 701                         mdt_export_evict(exp);
 702                         RETURN_EXIT;
 703                 }
 704
 705                 child = mdt_object_find(env, mdt, rr->rr_fid2);
 706                 if (IS_ERR(child)) {
 707                         rc = PTR_ERR(child);
 708                         LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
 709                                       "evicting client %s with export %s\n",
 710                                       PFID(rr->rr_fid2), rc,
 711                                       obd_uuid2str(&exp->exp_client_uuid),
 712                                       obd_export_nid2str(exp));
 713                         mdt_object_put(env, parent);
 714                         mdt_export_evict(exp);
 715                         RETURN_EXIT;
 716                 }
 717
 718                 if (unlikely(mdt_object_remote(child))) {
 719                         mdt_object_put(env, parent);
 720                         mdt_object_put(env, child);
 721                         /* the child object was created on remote server */
 722                         if (!mdt_is_dne_client(exp))
 723                                 /* Return -EIO for old client */
 724                                 GOTO(out, rc = -EIO);
 725                         repbody->mbo_fid1 = *rr->rr_fid2;
 726                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 727                         GOTO(out, rc = 0);
 728                 }
 729                 if (mdt_object_exists(child)) {
 730                         mdt_prep_ma_buf_from_rep(info, child, ma);
 731                         rc = mdt_attr_get_complex(info, child, ma);
 732                         if (!rc)
 733                                 rc = mdt_finish_open(info, parent, child,
 734                                                      open_flags, 1, ldlm_rep);
 735                         mdt_object_put(env, parent);
 736                         mdt_object_put(env, child);
 737                         if (!rc)
 738                                 mdt_pack_size2body(info, rr->rr_fid2,
 739                                                    &lhc->mlh_reg_lh);
 740                         GOTO(out, rc);
 741                 }
 742                 /* the child does not exist, we should do regular open */
 743                 mdt_object_put(env, parent);
 744                 mdt_object_put(env, child);
 745         }
 746         /* We did not try to create, so we are a pure open */
 747         rc = mdt_reint_open(info, lhc);
 748         EXIT;
 749 out:
 750         req->rq_status = rc;
 751         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 752         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
 753 }
 754
 755 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
 756 {
 757         u64 open_flags = info->mti_spec.sp_cr_flags;
 758         struct mdt_reint_record *rr = &info->mti_rr;
 759         struct md_attr *ma = &info->mti_attr;
 760         struct mdt_object *o;
 761         int rc;
 762
 763         ENTRY;
 764         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
 765         if (IS_ERR(o))
 766                 RETURN(rc = PTR_ERR(o));
 767
 768         if (unlikely(mdt_object_remote(o))) {
 769                 /* the child object was created on remote server */
 770                 struct mdt_body *repbody;
 771
 772                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 773                                                 DISP_LOOKUP_EXECD |
 774                                                 DISP_LOOKUP_POS));
 775                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 776                 repbody->mbo_fid1 = *rr->rr_fid2;
 777                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 778                 rc = 0;
 779         } else {
 780                 if (mdt_object_exists(o)) {
 781                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 782                                                         DISP_LOOKUP_EXECD |
 783                                                         DISP_LOOKUP_POS));
 784                         mdt_prep_ma_buf_from_rep(info, o, ma);
 785                         rc = mdt_attr_get_complex(info, o, ma);
 786                         if (rc == 0)
 787                                 rc = mdt_finish_open(info, NULL, o, open_flags,
 788                                                      0, rep);
 789                 } else {
 790                         rc = -ENOENT;
 791                 }
 792         }
 793
 794         mdt_object_put(info->mti_env, o);
 795         RETURN(rc);
 796 }
 797
 798 /* lock object for open */
 799 static int mdt_object_open_lock(struct mdt_thread_info *info,
 800                                 struct mdt_object *obj,
 801                                 struct mdt_lock_handle *lhc,
 802                                 __u64 *ibits)
 803 {
 804         struct md_attr *ma = &info->mti_attr;
 805         __u64 open_flags = info->mti_spec.sp_cr_flags;
 806         __u64 trybits = 0;
 807         enum ldlm_mode lm = LCK_CR;
 808         bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
 809         bool try_layout = false;
 810         bool create_layout = false;
 811         int rc = 0;
 812         int dom_stripes = LMM_NO_DOM;
 813         bool dom_lock = false;
 814
 815         ENTRY;
 816
 817         *ibits = 0;
 818         mdt_lock_handle_init(lhc);
 819
 820         if (req_is_replay(mdt_info_req(info)))
 821                 RETURN(0);
 822
 823         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
 824                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
 825                     md_should_create(open_flags))
 826                         create_layout = true;
 827                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
 828                     ma->ma_need & MA_LOV)
 829                         try_layout = true;
 830
 831                 /* DoM files can have just MDT stripe or combined MDT + OST
 832                  * stripes.
 833                  * - In the first case the open for read/write will do IO to
 834                  *   the MDT stripe and it makes sense to take IO lock in
 835                  *   advance along with OPEN even if it is blocking lock.
 836                  * - In the second case it is just size of MDT stripe and it
 837                  *   is quite unlikely that client will write into it, though
 838                  *   it may read it. So IO lock will be taken optionally if it
 839                  *   is non-blocking one.
 840                  */
 841                 if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
 842                         dom_stripes = mdt_lmm_dom_entry(ma->ma_lmm);
 843
 844                 if (dom_stripes == LMM_DOM_ONLY &&
 845                     info->mti_mdt->mdt_opts.mo_dom_lock > 0 &&
 846                     !mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
 847                         dom_lock = true;
 848         }
 849
 850         if (acq_lease) {
 851                 /* lease open, acquire write mode of open sem */
 852                 down_write(&obj->mot_open_sem);
 853
 854                 /* Lease exists and ask for new lease */
 855                 if (atomic_read(&obj->mot_lease_count) > 0) {
 856                         /* only exclusive open is supported, so lease
 857                          * are conflicted to each other */
 858                         GOTO(out, rc = -EBUSY);
 859                 }
 860
 861                 /* Lease must be with open lock */
 862                 if (!(open_flags & MDS_OPEN_LOCK)) {
 863                         CERROR("%s: Request lease for file:"DFID ", but open lock "
 864                                "is missed, open_flags = %#llo : rc = %d\n",
 865                                mdt_obd_name(info->mti_mdt),
 866                                PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
 867                         GOTO(out, rc = -EPROTO);
 868                 }
 869
 870                 /* XXX: only exclusive open is supported. */
 871                 lm = LCK_EX;
 872                 *ibits = MDS_INODELOCK_OPEN;
 873
 874                 /* never grant LCK_EX layout lock to client */
 875                 try_layout = false;
 876         } else { /* normal open */
 877                 /* normal open holds read mode of open sem */
 878                 down_read(&obj->mot_open_sem);
 879
 880                 if (open_flags & MDS_OPEN_LOCK) {
 881                         if (open_flags & MDS_FMODE_WRITE)
 882                                 lm = LCK_CW;
 883                         else if (open_flags & MDS_FMODE_EXEC)
 884                                 lm = LCK_PR;
 885                         else
 886                                 lm = LCK_CR;
 887
 888                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
 889                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
 890                         if (open_flags & MDS_FMODE_WRITE)
 891                                 lm = LCK_CW;
 892                         else
 893                                 lm = LCK_CR;
 894
 895                         /* revoke lease */
 896                         *ibits = MDS_INODELOCK_OPEN;
 897                         try_layout = false;
 898
 899                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 900                 } else if (dom_lock) {
 901                         lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR;
 902                         if (info->mti_mdt->mdt_opts.mo_dom_lock ==
 903                             TRYLOCK_DOM_ON_OPEN) {
 904                                 trybits |= MDS_INODELOCK_DOM |
 905                                            MDS_INODELOCK_LAYOUT;
 906                         } else {
 907                                 /* mo_dom_lock == ALWAYS_DOM_LOCK_ON_OPEN */
 908                                 *ibits = MDS_INODELOCK_DOM;
 909                                 if (info->mti_mdt->mdt_opts.mo_dom_read_open) {
 910                                         trybits |= MDS_INODELOCK_LAYOUT;
 911                                 }
 912                         }
 913                 }
 914
 915                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
 916                         PFID(mdt_object_fid(obj)),
 917                         atomic_read(&obj->mot_lease_count), lm);
 918         }
 919
 920         mdt_lock_reg_init(lhc, lm);
 921
 922         /* Return lookup lock to validate inode at the client side.
 923          * This is pretty important otherwise MDT will return layout
 924          * lock for each open.
 925          * However this is a double-edged sword because changing
 926          * permission will revoke a huge number of LOOKUP locks.
 927          */
 928         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
 929                 if (!(*ibits & MDS_INODELOCK_LOOKUP))
 930                         trybits |= MDS_INODELOCK_LOOKUP;
 931                 trybits |= MDS_INODELOCK_LAYOUT;
 932         }
 933
 934         if (*ibits | trybits)
 935                 rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, false);
 936
 937         CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
 938                ", open_flags = %#llo, try_layout = %d : rc = %d\n",
 939                mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
 940                *ibits, trybits, open_flags, try_layout, rc);
 941
 942         /* will change layout, revoke layout locks by enqueuing EX lock. */
 943         if (rc == 0 && create_layout) {
 944                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
 945
 946                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
 947                         ", open_flags = %#llo\n",
 948                         PFID(mdt_object_fid(obj)), open_flags);
 949
 950                 /* We cannot enqueue another lock for the same resource we
 951                  * already have a lock for, due to mechanics of waiting list
 952                  * iterating in ldlm, see LU-3601.
 953                  * As such we'll drop the open lock we just got above here,
 954                  * it's ok not to have this open lock as it's main purpose is to
 955                  * flush unused cached client open handles. */
 956                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 957                         mdt_object_unlock(info, obj, lhc, 1);
 958
 959                 LASSERT(!try_layout);
 960                 mdt_lock_handle_init(ll);
 961                 mdt_lock_reg_init(ll, LCK_EX);
 962                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
 963
 964                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
 965         }
 966
 967         /* Check if there is any other open handles after acquiring
 968          * open lock. At this point, caching open handles have been revoked
 969          * by open lock.
 970          * XXX: Now only exclusive open is supported. Need to check the
 971          * type of open for generic lease support. */
 972         if (rc == 0 && acq_lease) {
 973                 struct ptlrpc_request *req = mdt_info_req(info);
 974                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 975                 struct mdt_file_data *mfd;
 976                 bool is_replay_or_resent;
 977                 int open_count = 0;
 978
 979                 /* For lease: application can open a file and then apply lease,
 980                  * @handle contains original open handle in that case.
 981                  * In recovery, open REQ will be replayed and the lease REQ may
 982                  * be resent that means the open handle is already stale, so we
 983                  * need to fix it up here by finding new handle. */
 984                 is_replay_or_resent = req_is_replay(req) ||
 985                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
 986
 987                 /* if the request is _not_ a replay request, rr_open_handle
 988                  * may be used to hold an open file handle which is issuing the
 989                  * lease request, so that this openhandle doesn't count. */
 990                 mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 991                                           is_replay_or_resent);
 992                 if (mfd != NULL)
 993                         ++open_count;
 994
 995                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
 996                         PFID(mdt_object_fid(obj)),
 997                         atomic_read(&obj->mot_open_count), open_count);
 998
 999                 if (atomic_read(&obj->mot_open_count) > open_count)
1000                         GOTO(out, rc = -EBUSY);
1001         }
1002         GOTO(out, rc);
1003
1004 out:
1005         RETURN(rc);
1006 }
1007
1008 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1009                                    struct mdt_object *obj,
1010                                    struct mdt_lock_handle *lhc,
1011                                    __u64 ibits, int rc)
1012 {
1013         __u64 open_flags = info->mti_spec.sp_cr_flags;
1014         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1015         ENTRY;
1016
1017         if (req_is_replay(mdt_info_req(info)))
1018                 RETURN_EXIT;
1019
1020         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1021          * return to client side. */
1022         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1023                 mdt_object_unlock(info, obj, ll, 1);
1024
1025         ll = &info->mti_lh[MDT_LH_LAYOUT];
1026         /* Release local layout lock, layout was created */
1027         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1028                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1029                 mdt_object_unlock(info, obj, ll, 1);
1030         }
1031
1032         if (open_flags & MDS_OPEN_LEASE)
1033                 up_write(&obj->mot_open_sem);
1034         else
1035                 up_read(&obj->mot_open_sem);
1036
1037         /* Cross-ref case, the lock should be returned to the client */
1038         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1039                 RETURN_EXIT;
1040
1041         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1042             !(ibits & MDS_INODELOCK_DOM)) {
1043                 /* for the open request, the lock will only return to client
1044                  * if open or layout lock is granted. */
1045                 rc = 1;
1046         }
1047
1048         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1049                 struct ldlm_reply       *ldlm_rep;
1050
1051                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1052                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1053                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1054                         mdt_object_unlock(info, obj, lhc, 1);
1055         }
1056         RETURN_EXIT;
1057 }
1058
1059 /**
1060  * Check release is permitted for the current HSM flags.
1061  */
1062 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1063 {
1064         if (!(ma->ma_valid & MA_HSM))
1065                 return false;
1066
1067         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1068                 return false;
1069
1070         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1071                 return false;
1072
1073         return true;
1074 }
1075
1076 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1077                                 struct ldlm_reply *rep,
1078                                 struct mdt_lock_handle *lhc)
1079 {
1080         const struct lu_env *env = info->mti_env;
1081         struct mdt_device *mdt = info->mti_mdt;
1082         u64 open_flags = info->mti_spec.sp_cr_flags;
1083         struct mdt_reint_record *rr = &info->mti_rr;
1084         struct md_attr *ma = &info->mti_attr;
1085         struct mdt_object *parent = NULL;
1086         struct mdt_object *o;
1087         bool object_locked = false;
1088         u64 ibits = 0;
1089         int rc;
1090
1091         ENTRY;
1092         if (md_should_create(open_flags)) {
1093                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1094                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1095                         if (IS_ERR(parent)) {
1096                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1097                                        " for anonymous created %ld, try to"
1098                                        " use server-side parent.\n",
1099                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1100                                 parent = NULL;
1101                         }
1102                 }
1103                 if (parent == NULL)
1104                         ma->ma_need |= MA_PFID;
1105         }
1106
1107         o = mdt_object_find(env, mdt, rr->rr_fid2);
1108         if (IS_ERR(o))
1109                 GOTO(out_parent_put, rc = PTR_ERR(o));
1110
1111         if (mdt_object_remote(o)) {
1112                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1113                        mdt_obd_name(info->mti_mdt),
1114                        PFID(rr->rr_fid2));
1115                 GOTO(out, rc = -EREMOTE);
1116         } else if (!mdt_object_exists(o)) {
1117                 mdt_set_disposition(info, rep,
1118                                     DISP_IT_EXECD |
1119                                     DISP_LOOKUP_EXECD |
1120                                     DISP_LOOKUP_NEG);
1121                 GOTO(out, rc = -ENOENT);
1122         }
1123
1124         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1125
1126         mdt_prep_ma_buf_from_rep(info, o, ma);
1127         if (open_flags & MDS_OPEN_RELEASE)
1128                 ma->ma_need |= MA_HSM;
1129         rc = mdt_attr_get_complex(info, o, ma);
1130         if (rc)
1131                 GOTO(out, rc);
1132
1133         /* We should not change file's existing LOV EA */
1134         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1135             open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1136                 GOTO(out, rc = -EEXIST);
1137
1138         /* If a release request, check file open flags are fine and ask for an
1139          * exclusive open access. */
1140         if (open_flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1141                 GOTO(out, rc = -EPERM);
1142
1143         rc = mdt_check_resent_lock(info, o, lhc);
1144         if (rc < 0) {
1145                 GOTO(out, rc);
1146         } else if (rc > 0) {
1147                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1148                 object_locked = true;
1149                 if (rc)
1150                         GOTO(out_unlock, rc);
1151         }
1152
1153         if (ma->ma_valid & MA_PFID) {
1154                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1155                 if (IS_ERR(parent)) {
1156                         CDEBUG(D_INODE, "Fail to find parent "DFID
1157                                " for anonymous created %ld, try to"
1158                                " use system default.\n",
1159                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1160                         parent = NULL;
1161                 }
1162         }
1163
1164         rc = mdt_finish_open(info, parent, o, open_flags, 0, rep);
1165         if (!rc) {
1166                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1167                 if (open_flags & MDS_OPEN_LOCK)
1168                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1169                 if (open_flags & MDS_OPEN_LEASE)
1170                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1171         }
1172         GOTO(out_unlock, rc);
1173
1174 out_unlock:
1175         if (object_locked)
1176                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1177 out:
1178         mdt_object_put(env, o);
1179         if (rc == 0)
1180                 mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
1181 out_parent_put:
1182         if (parent != NULL)
1183                 mdt_object_put(env, parent);
1184         return rc;
1185 }
1186
1187 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1188 static int mdt_cross_open(struct mdt_thread_info *info,
1189                           const struct lu_fid *parent_fid,
1190                           const struct lu_fid *fid,
1191                           struct ldlm_reply *rep, u64 open_flags)
1192 {
1193         struct md_attr *ma = &info->mti_attr;
1194         struct mdt_object *o;
1195         int rc;
1196         ENTRY;
1197
1198         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1199         if (IS_ERR(o))
1200                 RETURN(rc = PTR_ERR(o));
1201
1202         if (mdt_object_remote(o)) {
1203                 /* Something is wrong here, the object is on another MDS! */
1204                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1205                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1206                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1207                                 &o->mot_obj,
1208                                 "Object isn't on this server! FLD error?");
1209                 rc = -EFAULT;
1210         } else {
1211                 if (mdt_object_exists(o)) {
1212                         int mask;
1213
1214                         /* Do permission check for cross-open after converting
1215                          * MDS_OPEN_* flags to MAY_* permission mask.
1216                          */
1217                         mask = mds_accmode(open_flags);
1218
1219                         rc = mo_permission(info->mti_env, NULL,
1220                                            mdt_object_child(o), NULL, mask);
1221                         if (rc)
1222                                 goto out;
1223
1224                         mdt_prep_ma_buf_from_rep(info, o, ma);
1225                         rc = mdt_attr_get_complex(info, o, ma);
1226                         if (rc != 0)
1227                                 GOTO(out, rc);
1228
1229                         rc = mdt_pack_secctx_in_reply(info, o);
1230                         if (unlikely(rc))
1231                                 GOTO(out, rc);
1232
1233                         rc = mdt_finish_open(info, NULL, o, open_flags, 0, rep);
1234                 } else {
1235                         /*
1236                          * Something is wrong here. lookup was positive but
1237                          * there is no object!
1238                          */
1239                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1240                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1241                         rc = -EFAULT;
1242                 }
1243         }
1244 out:
1245         mdt_object_put(info->mti_env, o);
1246
1247         RETURN(rc);
1248 }
1249
1250 /*
1251  * find root object and take its xattr lock if it's on remote MDT, later create
1252  * may use fs default striping (which is stored in root xattr).
1253  */
1254 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1255                                struct mdt_device *mdt)
1256 {
1257         struct mdt_object *md_root = mdt->mdt_md_root;
1258         struct lustre_handle lhroot;
1259         int rc;
1260
1261         if (md_root == NULL) {
1262                 lu_root_fid(&info->mti_tmp_fid1);
1263                 md_root = mdt_object_find(info->mti_env, mdt,
1264                                           &info->mti_tmp_fid1);
1265                 if (IS_ERR(md_root))
1266                         return PTR_ERR(md_root);
1267
1268                 spin_lock(&mdt->mdt_lock);
1269                 if (mdt->mdt_md_root != NULL) {
1270                         spin_unlock(&mdt->mdt_lock);
1271
1272                         LASSERTF(mdt->mdt_md_root == md_root,
1273                                  "Different root object ("
1274                                  DFID") instances, %p, %p\n",
1275                                  PFID(&info->mti_tmp_fid1),
1276                                  mdt->mdt_md_root, md_root);
1277                         LASSERT(atomic_read(
1278                                 &md_root->mot_obj.lo_header->loh_ref) > 1);
1279
1280                         mdt_object_put(info->mti_env, md_root);
1281                 } else {
1282                         mdt->mdt_md_root = md_root;
1283                         spin_unlock(&mdt->mdt_lock);
1284                 }
1285         }
1286
1287         if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1288                 return 0;
1289
1290         rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root),
1291                                     &lhroot, LCK_PR, MDS_INODELOCK_XATTR,
1292                                     true);
1293         if (rc < 0)
1294                 return rc;
1295
1296         md_root->mot_cache_attr = 1;
1297
1298         /* don't cancel this lock, so that we know the cached xattr is valid. */
1299         ldlm_lock_decref(&lhroot, LCK_PR);
1300
1301         return 0;
1302 }
1303
1304 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1305 {
1306         struct mdt_device *mdt = info->mti_mdt;
1307         struct ptlrpc_request *req = mdt_info_req(info);
1308         struct mdt_object *parent;
1309         struct mdt_object *child;
1310         struct mdt_lock_handle *lh;
1311         struct ldlm_reply *ldlm_rep;
1312         struct mdt_body *repbody;
1313         struct lu_fid *child_fid = &info->mti_tmp_fid1;
1314         struct md_attr *ma = &info->mti_attr;
1315         u64 open_flags = info->mti_spec.sp_cr_flags;
1316         u64 ibits = 0;
1317         struct mdt_reint_record *rr = &info->mti_rr;
1318         int result, rc;
1319         int created = 0;
1320         int object_locked = 0;
1321         u32 msg_flags;
1322
1323         ENTRY;
1324         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1325                                (obd_timeout + 1) / 4);
1326
1327         mdt_counter_incr(req, LPROC_MDT_OPEN);
1328         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1329
1330         ma->ma_need = MA_INODE;
1331         ma->ma_valid = 0;
1332
1333         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1334         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1335
1336         if (unlikely(open_flags & MDS_OPEN_JOIN_FILE)) {
1337                 CERROR("file join is not supported anymore.\n");
1338                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1339         }
1340         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1341
1342         if ((open_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1343             info->mti_spec.u.sp_ea.eadata == NULL)
1344                 GOTO(out, result = err_serious(-EINVAL));
1345
1346         if (open_flags & MDS_FMODE_WRITE &&
1347             exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1348                 GOTO(out, result = -EROFS);
1349
1350         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1351                "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1352                PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2),
1353                open_flags, ma->ma_attr.la_mode, msg_flags);
1354
1355         if (info->mti_cross_ref) {
1356                 /* This is cross-ref open */
1357                 mdt_set_disposition(info, ldlm_rep,
1358                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1359                              DISP_LOOKUP_POS));
1360                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1361                                         ldlm_rep, open_flags);
1362                 GOTO(out, result);
1363         } else if (req_is_replay(req)) {
1364                 result = mdt_open_by_fid(info, ldlm_rep);
1365
1366                 if (result != -ENOENT)
1367                         GOTO(out, result);
1368
1369                 /* We didn't find the correct object, so we need to re-create it
1370                  * via a regular replay. */
1371                 if (!(open_flags & MDS_OPEN_CREAT)) {
1372                         DEBUG_REQ(D_ERROR, req,
1373                                   "OPEN & CREAT not in open replay/by_fid");
1374                         GOTO(out, result = -EFAULT);
1375                 }
1376                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1377         } else if (open_flags & MDS_OPEN_BY_FID) {
1378                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1379                 if (result < 0)
1380                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1381                                PFID(rr->rr_fid2), result);
1382                 GOTO(out, result);
1383         }
1384
1385         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1386                 GOTO(out, result = err_serious(-ENOMEM));
1387
1388         mdt_set_disposition(info, ldlm_rep,
1389                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1390
1391         if (!lu_name_is_valid(&rr->rr_name))
1392                 GOTO(out, result = -EPROTO);
1393
1394         result = mdt_lock_root_xattr(info, mdt);
1395         if (result < 0)
1396                 GOTO(out, result);
1397
1398 again:
1399         lh = &info->mti_lh[MDT_LH_PARENT];
1400         mdt_lock_pdo_init(lh, (open_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1401                           &rr->rr_name);
1402
1403         parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1404         if (IS_ERR(parent))
1405                 GOTO(out, result = PTR_ERR(parent));
1406
1407         result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1408         if (result != 0) {
1409                 mdt_object_put(info->mti_env, parent);
1410                 GOTO(out, result);
1411         }
1412
1413         /* get and check version of parent */
1414         result = mdt_version_get_check(info, parent, 0);
1415         if (result)
1416                 GOTO(out_parent, result);
1417
1418         fid_zero(child_fid);
1419
1420         result = -ENOENT;
1421         if ((open_flags & MDS_OPEN_VOLATILE) == 0)
1422                 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1423                                     &rr->rr_name, child_fid, &info->mti_spec);
1424
1425         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1426                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1427                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1428                  PFID(child_fid));
1429
1430         if (result != 0 && result != -ENOENT && result != -ESTALE)
1431                 GOTO(out_parent, result);
1432
1433         if (result == -ENOENT || result == -ESTALE) {
1434                 /* If the object is dead, let's check if the object
1435                  * is being migrated to a new object */
1436                 if (result == -ESTALE) {
1437                         struct lu_buf lmv_buf;
1438
1439                         lmv_buf.lb_buf = info->mti_xattr_buf;
1440                         lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
1441                         rc = mo_xattr_get(info->mti_env,
1442                                           mdt_object_child(parent),
1443                                           &lmv_buf, XATTR_NAME_LMV);
1444                         if (rc > 0) {
1445                                 struct lmv_mds_md_v1 *lmv;
1446
1447                                 lmv = lmv_buf.lb_buf;
1448                                 if (le32_to_cpu(lmv->lmv_hash_type) &
1449                                                 LMV_HASH_FLAG_MIGRATION) {
1450                                         /* Get the new parent FID and retry */
1451                                         mdt_object_unlock_put(info, parent,
1452                                                               lh, 1);
1453                                         mdt_lock_handle_init(lh);
1454                                         fid_le_to_cpu(
1455                                                 (struct lu_fid *)rr->rr_fid1,
1456                                                 &lmv->lmv_stripe_fids[1]);
1457                                         goto again;
1458                                 }
1459                         }
1460                 }
1461
1462                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1463                 if (result == -ESTALE) {
1464                         /*
1465                          * -ESTALE means the parent is a dead(unlinked) dir, so
1466                          * it should return -ENOENT to in accordance with the
1467                          * original mds implementaion.
1468                          */
1469                         GOTO(out_parent, result = -ENOENT);
1470                 }
1471
1472                 if (!(open_flags & MDS_OPEN_CREAT))
1473                         GOTO(out_parent, result);
1474                 if (mdt_rdonly(req->rq_export))
1475                         GOTO(out_parent, result = -EROFS);
1476                 *child_fid = *info->mti_rr.rr_fid2;
1477                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1478                          PFID(child_fid));
1479                 /* In the function below, .hs_keycmp resolves to
1480                  * lu_obj_hop_keycmp() */
1481                 /* coverity[overrun-buffer-val] */
1482                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1483         } else {
1484                 /*
1485                  * Check for O_EXCL is moved to the mdt_finish_open(),
1486                  * we need to return FID back in that case.
1487                  */
1488                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1489                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1490         }
1491         if (IS_ERR(child))
1492                 GOTO(out_parent, result = PTR_ERR(child));
1493
1494         /** check version of child  */
1495         rc = mdt_version_get_check(info, child, 1);
1496         if (rc)
1497                 GOTO(out_child, result = rc);
1498
1499         if (result == -ENOENT) {
1500                 /* Create under OBF and .lustre is not permitted */
1501                 if (!fid_is_md_operative(rr->rr_fid1) &&
1502                     (open_flags & MDS_OPEN_VOLATILE) == 0)
1503                         GOTO(out_child, result = -EPERM);
1504
1505                 /* save versions in reply */
1506                 mdt_version_get_save(info, parent, 0);
1507                 mdt_version_get_save(info, child, 1);
1508
1509                 /* version of child will be changed */
1510                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1511
1512                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1513                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1514
1515                 /* Let lower layers know what is lock mode on directory. */
1516                 info->mti_spec.sp_cr_mode =
1517                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1518
1519                 /* Don't do lookup sanity check. We know name doesn't exist. */
1520                 info->mti_spec.sp_cr_lookup = 0;
1521                 info->mti_spec.sp_feat = &dt_directory_features;
1522
1523                 result = mdo_create(info->mti_env, mdt_object_child(parent),
1524                                     &rr->rr_name, mdt_object_child(child),
1525                                     &info->mti_spec, &info->mti_attr);
1526                 if (result == -ERESTART) {
1527                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1528                         GOTO(out_child, result);
1529                 } else {
1530                         mdt_prep_ma_buf_from_rep(info, child, ma);
1531                         /* XXX: we should call this once, see few lines below */
1532                         if (result == 0)
1533                                 result = mdt_attr_get_complex(info, child, ma);
1534
1535                         if (result != 0)
1536                                 GOTO(out_child, result);
1537                 }
1538                 created = 1;
1539                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
1540         } else {
1541                 /*
1542                  * The object is on remote node, return its FID for remote open.
1543                  */
1544                 if (mdt_object_remote(child)) {
1545                         /*
1546                          * Check if this lock already was sent to client and
1547                          * this is resent case. For resent case do not take lock
1548                          * again, use what is already granted.
1549                          */
1550                         LASSERT(lhc != NULL);
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                                 mdt_lock_handle_init(lhc);
1557                                 mdt_lock_reg_init(lhc, LCK_PR);
1558
1559                                 rc = mdt_object_lock(info, child, lhc,
1560                                                      MDS_INODELOCK_LOOKUP);
1561                         }
1562                         repbody->mbo_fid1 = *mdt_object_fid(child);
1563                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1564                         if (rc != 0)
1565                                 result = rc;
1566                         else
1567                                 result = -MDT_EREMOTE_OPEN;
1568                         GOTO(out_child, result);
1569                 } else if (mdt_object_exists(child)) {
1570                         /* We have to get attr & LOV EA & HSM for this
1571                          * object. */
1572                         mdt_prep_ma_buf_from_rep(info, child, ma);
1573                         ma->ma_need |= MA_HSM;
1574                         result = mdt_attr_get_complex(info, child, ma);
1575                         if (result != 0)
1576                                 GOTO(out_child, result);
1577                 } else {
1578                         /* Object does not exist. Likely FS corruption. */
1579                         CERROR("%s: name '"DNAME"' present, but FID "
1580                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1581                                PNAME(&rr->rr_name), PFID(child_fid));
1582                         GOTO(out_child, result = -EIO);
1583                 }
1584         }
1585
1586         repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1587         repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1588
1589         rc = mdt_pack_secctx_in_reply(info, child);
1590         if (unlikely(rc))
1591                 GOTO(out_child, result = rc);
1592
1593         rc = mdt_check_resent_lock(info, child, lhc);
1594         if (rc < 0) {
1595                 GOTO(out_child, result = rc);
1596         } else if (rc == 0) {
1597                 /* the open lock might already be gotten in
1598                  * ldlm_handle_enqueue() */
1599                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1600                 if (open_flags & MDS_OPEN_LOCK)
1601                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1602         } else {
1603                 /* get openlock if this isn't replay and client requested it */
1604                 if (!req_is_replay(req)) {
1605                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1606                         object_locked = 1;
1607                         if (rc != 0)
1608                                 GOTO(out_child_unlock, result = rc);
1609                         else if (open_flags & MDS_OPEN_LOCK)
1610                                 mdt_set_disposition(info, ldlm_rep,
1611                                                     DISP_OPEN_LOCK);
1612                 }
1613         }
1614         /* Try to open it now. */
1615         rc = mdt_finish_open(info, parent, child, open_flags,
1616                              created, ldlm_rep);
1617         if (rc) {
1618                 result = rc;
1619                 /* openlock will be released if mdt_finish_open() failed */
1620                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1621
1622                 if (created && (open_flags & MDS_OPEN_VOLATILE)) {
1623                         CERROR("%s: cannot open volatile file "DFID", orphan "
1624                                "file will be left in PENDING directory until "
1625                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1626                                PFID(mdt_object_fid(child)), rc);
1627                         GOTO(out_child_unlock, result);
1628                 }
1629
1630                 if (created) {
1631                         ma->ma_need = 0;
1632                         ma->ma_valid = 0;
1633                         rc = mdo_unlink(info->mti_env,
1634                                         mdt_object_child(parent),
1635                                         mdt_object_child(child),
1636                                         &rr->rr_name,
1637                                         &info->mti_attr, 0);
1638                         if (rc != 0)
1639                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1640                                        mdt_obd_name(info->mti_mdt),
1641                                        PFID(mdt_object_fid(child)), rc);
1642                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1643                 }
1644         }
1645         EXIT;
1646 out_child_unlock:
1647         if (object_locked)
1648                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1649 out_child:
1650         mdt_object_put(info->mti_env, child);
1651         if (result == 0)
1652                 mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh);
1653 out_parent:
1654         mdt_object_unlock_put(info, parent, lh, result || !created);
1655 out:
1656         if (result)
1657                 lustre_msg_set_transno(req->rq_repmsg, 0);
1658         return result;
1659 }
1660
1661 /**
1662  * Create an orphan object use local root.
1663  */
1664 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1665                                           struct mdt_device *mdt,
1666                                           const struct lu_fid *fid,
1667                                           struct md_attr *attr, fmode_t fmode)
1668 {
1669         const struct lu_env *env = info->mti_env;
1670         struct md_op_spec *spec = &info->mti_spec;
1671         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1672         struct mdt_object *obj = NULL;
1673         struct mdt_object *local_root;
1674         static const struct lu_name lname = {
1675                 .ln_name = "i_am_nobody",
1676                 .ln_namelen = sizeof("i_am_nobody") - 1,
1677         };
1678         struct lu_ucred *uc;
1679         cfs_cap_t uc_cap_save;
1680         int rc;
1681         ENTRY;
1682
1683         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1684         if (rc != 0)
1685                 RETURN(ERR_PTR(rc));
1686
1687         local_root = mdt_object_find(env, mdt, local_root_fid);
1688         if (IS_ERR(local_root))
1689                 RETURN(local_root);
1690
1691         obj = mdt_object_new(env, mdt, fid);
1692         if (IS_ERR(obj))
1693                 GOTO(out, rc = PTR_ERR(obj));
1694
1695         spec->sp_cr_lookup = 0;
1696         spec->sp_feat = &dt_directory_features;
1697         spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1698         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1699         if (attr->ma_valid & MA_LOV) {
1700                 spec->u.sp_ea.eadata = attr->ma_lmm;
1701                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1702                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1703         } else {
1704                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1705         }
1706
1707         uc = lu_ucred(env);
1708         uc_cap_save = uc->uc_cap;
1709         uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1710         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1711                         mdt_object_child(obj), spec, attr);
1712         uc->uc_cap = uc_cap_save;
1713         if (rc < 0) {
1714                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1715                        mdt_obd_name(mdt), PFID(fid), rc);
1716                 GOTO(out, rc);
1717         }
1718
1719         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED, spec);
1720         if (rc < 0)
1721                 CERROR("%s: cannot open volatile file "DFID", orphan "
1722                        "file will be left in PENDING directory until "
1723                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1724                        PFID(fid), rc);
1725         GOTO(out, rc);
1726
1727 out:
1728         if (rc < 0) {
1729                 if (!IS_ERR(obj))
1730                         mdt_object_put(env, obj);
1731                 obj = ERR_PTR(rc);
1732         }
1733         mdt_object_put(env, local_root);
1734         return obj;
1735 }
1736
1737 /* XXX Look into layout in MDT layer. */
1738 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1739 {
1740         struct lov_comp_md_v1   *comp_v1;
1741         struct lov_mds_md       *v1;
1742         __u32   off;
1743         int     i;
1744
1745         if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1746                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1747
1748                 if (comp_v1->lcm_entry_count == 0)
1749                         return -EINVAL;
1750
1751                 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1752                         off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1753                         v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1754                         v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1755                 }
1756         } else {
1757                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1758         }
1759         return 0;
1760 }
1761
1762 static inline int mdt_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1763 {
1764         struct lov_comp_md_v1 *comp_v1;
1765
1766         if (le32_to_cpu(lmm->lmm_magic == LOV_MAGIC_COMP_V1)) {
1767                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1768                 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1769         } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1770                    le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1771                 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1772         } else {
1773                 return -EINVAL;
1774         }
1775         return 0;
1776 }
1777
1778 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1779                            struct md_attr *ma)
1780 {
1781         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1782         struct lu_ucred        *uc = mdt_ucred(info);
1783         struct close_data      *data;
1784         struct ldlm_lock       *lease;
1785         struct mdt_object      *orphan;
1786         struct md_attr         *orp_ma;
1787         struct lu_buf          *buf;
1788         cfs_cap_t               cap;
1789         bool                    lease_broken;
1790         int                     rc;
1791         int                     rc2;
1792         ENTRY;
1793
1794         if (mdt_rdonly(info->mti_exp))
1795                 RETURN(-EROFS);
1796
1797         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1798         if (data == NULL)
1799                 RETURN(-EPROTO);
1800
1801         lease = ldlm_handle2lock(&data->cd_handle);
1802         if (lease == NULL)
1803                 RETURN(-ESTALE);
1804
1805         /* try to hold open_sem so that nobody else can open the file */
1806         if (!down_write_trylock(&o->mot_open_sem)) {
1807                 ldlm_lock_cancel(lease);
1808                 GOTO(out_reprocess, rc = -EBUSY);
1809         }
1810
1811         /* Check if the lease open lease has already canceled */
1812         lock_res_and_lock(lease);
1813         lease_broken = ldlm_is_cancel(lease);
1814         unlock_res_and_lock(lease);
1815
1816         LDLM_DEBUG(lease, DFID " lease broken? %d",
1817                    PFID(mdt_object_fid(o)), lease_broken);
1818
1819         /* Cancel server side lease. Client side counterpart should
1820          * have been cancelled. It's okay to cancel it now as we've
1821          * held mot_open_sem. */
1822         ldlm_lock_cancel(lease);
1823
1824         if (lease_broken) /* don't perform release task */
1825                 GOTO(out_unlock, rc = -ESTALE);
1826
1827         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1828                 GOTO(out_unlock, rc = -EINVAL);
1829
1830         /* ma_need was set before but it seems fine to change it in order to
1831          * avoid modifying the one from RPC */
1832         ma->ma_need = MA_HSM;
1833         rc = mdt_attr_get_complex(info, o, ma);
1834         if (rc != 0)
1835                 GOTO(out_unlock, rc);
1836
1837         if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
1838                 if (ma->ma_valid & MA_HSM) {
1839                         if (ma->ma_hsm.mh_flags & HS_RELEASED)
1840                                 GOTO(out_unlock, rc = -EALREADY);
1841
1842                         if (ma->ma_hsm.mh_arch_id != data->cd_archive_id)
1843                                 CDEBUG(D_CACHE,
1844                                        DFID" archive id diff: %llu:%u\n",
1845                                        PFID(mdt_object_fid(o)),
1846                                        ma->ma_hsm.mh_arch_id,
1847                                        data->cd_archive_id);
1848
1849                         if (!(ma->ma_hsm.mh_flags & HS_DIRTY) &&
1850                             ma->ma_hsm.mh_arch_ver == data->cd_data_version) {
1851                                 CDEBUG(D_CACHE,
1852                                        DFID" data version matches: packed=%llu "
1853                                        "and on-disk=%llu\n",
1854                                        PFID(mdt_object_fid(o)),
1855                                        data->cd_data_version,
1856                                        ma->ma_hsm.mh_arch_ver);
1857                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1858                         }
1859
1860                         if (ma->ma_hsm.mh_flags & HS_DIRTY)
1861                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1862                 } else {
1863                         /* Set up HSM attribte for PCC archived object */
1864                         BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1865                                      sizeof(info->mti_xattr_buf));
1866                         buf = &info->mti_buf;
1867                         buf->lb_buf = info->mti_xattr_buf;
1868                         buf->lb_len = sizeof(struct hsm_attrs);
1869                         memset(&ma->ma_hsm, 0, sizeof(ma->ma_hsm));
1870                         ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1871                         ma->ma_hsm.mh_arch_id = data->cd_archive_id;
1872                         ma->ma_hsm.mh_arch_ver = data->cd_data_version;
1873                         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1874
1875                         rc = mo_xattr_set(info->mti_env, mdt_object_child(o),
1876                                           buf, XATTR_NAME_HSM, 0);
1877                         if (rc)
1878                                 GOTO(out_unlock, rc);
1879                 }
1880         } else {
1881                 if (!mdt_hsm_release_allow(ma))
1882                         GOTO(out_unlock, rc = -EPERM);
1883
1884                 /* already released? */
1885                 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1886                         GOTO(out_unlock, rc = 0);
1887
1888                 /* Compare on-disk and packed data_version */
1889                 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1890                         CDEBUG(D_HSM, DFID" data_version mismatches: "
1891                                "packed=%llu and on-disk=%llu\n",
1892                                PFID(mdt_object_fid(o)),
1893                                data->cd_data_version,
1894                                ma->ma_hsm.mh_arch_ver);
1895                         GOTO(out_unlock, rc = -EPERM);
1896                 }
1897         }
1898
1899         ma->ma_valid = MA_INODE;
1900         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1901         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1902         if (rc < 0)
1903                 GOTO(out_unlock, rc);
1904
1905         mutex_lock(&o->mot_som_mutex);
1906         rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size,
1907                            ma->ma_attr.la_blocks);
1908         mutex_unlock(&o->mot_som_mutex);
1909         if (rc2 < 0)
1910                 CDEBUG(D_INODE,
1911                        "%s: File "DFID" SOM update failed: rc = %d\n",
1912                        mdt_obd_name(info->mti_mdt),
1913                        PFID(mdt_object_fid(o)), rc2);
1914
1915
1916         ma->ma_need = MA_INODE | MA_LOV;
1917         rc = mdt_attr_get_complex(info, o, ma);
1918         if (rc < 0)
1919                 GOTO(out_unlock, rc);
1920
1921         if (!(ma->ma_valid & MA_LOV)) {
1922                 /* Even empty file are released */
1923                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1924                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
1925                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1926                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1927                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1928         } else {
1929                 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
1930                  * ma_lmm as lov_user_md, then it will be confused by union of
1931                  * layout_gen and stripe_offset. */
1932                 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
1933                     LOV_MAGIC_MAGIC)
1934                         ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1935                 else
1936                         GOTO(out_unlock, rc = -EINVAL);
1937         }
1938
1939         /* Set file as released. */
1940         rc = mdt_hsm_set_released(ma->ma_lmm);
1941         if (rc)
1942                 GOTO(out_unlock, rc);
1943
1944         orp_ma = &info->mti_u.hsm.attr;
1945         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1946         /* We use root ownership to bypass potential quota
1947          * restrictions on the user and group of the file. */
1948         orp_ma->ma_attr.la_uid = 0;
1949         orp_ma->ma_attr.la_gid = 0;
1950         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1951         orp_ma->ma_lmm = ma->ma_lmm;
1952         orp_ma->ma_lmm_size = ma->ma_lmm_size;
1953         orp_ma->ma_valid = MA_INODE | MA_LOV;
1954         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1955                                  MDS_FMODE_WRITE);
1956         if (IS_ERR(orphan)) {
1957                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1958                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1959                        PTR_ERR(orphan));
1960                 GOTO(out_unlock, rc = PTR_ERR(orphan));
1961         }
1962
1963         /* Set up HSM attribute for orphan object */
1964         BUILD_BUG_ON(sizeof(struct hsm_attrs) > sizeof(info->mti_xattr_buf));
1965         buf = &info->mti_buf;
1966         buf->lb_buf = info->mti_xattr_buf;
1967         buf->lb_len = sizeof(struct hsm_attrs);
1968         ma->ma_hsm.mh_flags |= HS_RELEASED;
1969         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1970         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
1971
1972         mdt_lock_reg_init(lh, LCK_EX);
1973         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
1974                              MDS_INODELOCK_XATTR);
1975         if (rc != 0)
1976                 GOTO(out_close, rc);
1977
1978         /* The orphan has root ownership so we need to raise
1979          * CAP_FOWNER to set the HSM attributes. */
1980         cap = uc->uc_cap;
1981         uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
1982         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
1983                           XATTR_NAME_HSM, 0);
1984         uc->uc_cap = cap;
1985         if (rc != 0)
1986                 GOTO(out_layout_lock, rc);
1987
1988         /* Swap layout with orphan objects. */
1989         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
1990                              mdt_object_child(orphan),
1991                              SWAP_LAYOUTS_MDS_HSM);
1992
1993         if (!rc && ma->ma_attr_flags & MDS_PCC_ATTACH) {
1994                 ma->ma_need = MA_LOV;
1995                 rc = mdt_attr_get_complex(info, o, ma);
1996         }
1997
1998         EXIT;
1999
2000 out_layout_lock:
2001         /* Release exclusive LL */
2002         mdt_object_unlock(info, o, lh, 1);
2003 out_close:
2004         /* Close orphan object anyway */
2005         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2006                        MDS_FMODE_WRITE);
2007         if (rc2 < 0)
2008                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2009                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2010         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2011                         "object closed");
2012         mdt_object_put(info->mti_env, orphan);
2013
2014 out_unlock:
2015         up_write(&o->mot_open_sem);
2016
2017         /* already released */
2018         if (rc == 0) {
2019                 struct mdt_body *repbody;
2020
2021                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2022                 LASSERT(repbody != NULL);
2023                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2024                 if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
2025                         LASSERT(ma->ma_valid & MA_LOV);
2026                         rc = mdt_get_lmm_gen(ma->ma_lmm,
2027                                              &repbody->mbo_layout_gen);
2028                         if (!rc)
2029                                 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
2030                 }
2031         }
2032
2033 out_reprocess:
2034         ldlm_reprocess_all(lease->l_resource, lease);
2035         LDLM_LOCK_PUT(lease);
2036
2037         ma->ma_valid = 0;
2038         ma->ma_need = 0;
2039
2040         return rc;
2041 }
2042
2043 int mdt_close_handle_layouts(struct mdt_thread_info *info,
2044                              struct mdt_object *o, struct md_attr *ma)
2045 {
2046         struct mdt_lock_handle  *lh1 = &info->mti_lh[MDT_LH_NEW];
2047         struct mdt_lock_handle  *lh2 = &info->mti_lh[MDT_LH_OLD];
2048         struct close_data       *data;
2049         struct ldlm_lock        *lease;
2050         struct mdt_object       *o1 = o, *o2;
2051         bool                     lease_broken;
2052         bool                     swap_objects;
2053         int                      rc;
2054         ENTRY;
2055
2056         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2057                 RETURN(-EROFS);
2058
2059         if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
2060                 RETURN(-EINVAL);
2061
2062         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2063         if (data == NULL)
2064                 RETURN(-EPROTO);
2065
2066         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2067                 RETURN(-EINVAL);
2068
2069         rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
2070         if (unlikely(rc == 0))
2071                 RETURN(-EINVAL);
2072
2073         /* Exchange o1 and o2, to enforce locking order */
2074         swap_objects = (rc < 0);
2075
2076         lease = ldlm_handle2lock(&data->cd_handle);
2077         if (lease == NULL)
2078                 RETURN(-ESTALE);
2079
2080         o2 = mdt_object_find(info->mti_env, info->mti_mdt, &data->cd_fid);
2081         if (IS_ERR(o2))
2082                 GOTO(out_lease, rc = PTR_ERR(o2));
2083
2084         if (!S_ISREG(lu_object_attr(&o2->mot_obj))) {
2085                 swap_objects = false; /* not swapped yet */
2086                 GOTO(out_obj, rc = -EINVAL);
2087         }
2088
2089         if (swap_objects)
2090                 swap(o1, o2);
2091
2092         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
2093                            MAY_WRITE);
2094         if (rc < 0)
2095                 GOTO(out_obj, rc);
2096
2097         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
2098                            MAY_WRITE);
2099         if (rc < 0)
2100                 GOTO(out_obj, rc);
2101
2102         /* try to hold open_sem so that nobody else can open the file */
2103         if (!down_write_trylock(&o->mot_open_sem)) {
2104                 ldlm_lock_cancel(lease);
2105                 GOTO(out_obj, rc = -EBUSY);
2106         }
2107
2108         /* Check if the lease open lease has already canceled */
2109         lock_res_and_lock(lease);
2110         lease_broken = ldlm_is_cancel(lease);
2111         unlock_res_and_lock(lease);
2112
2113         LDLM_DEBUG(lease, DFID " lease broken? %d",
2114                    PFID(mdt_object_fid(o)), lease_broken);
2115
2116         /* Cancel server side lease. Client side counterpart should
2117          * have been cancelled. It's okay to cancel it now as we've
2118          * held mot_open_sem. */
2119         ldlm_lock_cancel(lease);
2120
2121         if (lease_broken)
2122                 GOTO(out_unlock_sem, rc = -ESTALE);
2123
2124         mdt_lock_reg_init(lh1, LCK_EX);
2125         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
2126                              MDS_INODELOCK_XATTR);
2127         if (rc < 0)
2128                 GOTO(out_unlock_sem, rc);
2129
2130         mdt_lock_reg_init(lh2, LCK_EX);
2131         rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2132                              MDS_INODELOCK_XATTR);
2133         if (rc < 0)
2134                 GOTO(out_unlock1, rc);
2135
2136         /* Swap layout with orphan object */
2137         if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2138                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2139                                      mdt_object_child(o2), 0);
2140         } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2141                    ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2142                 struct lu_buf *buf = &info->mti_buf;
2143                 struct md_rejig_data mrd;
2144
2145                 mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2146                 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)
2147                         mrd.mrd_mirror_id = data->cd_mirror_id;
2148
2149                 buf->lb_len = sizeof(mrd);
2150                 buf->lb_buf = &mrd;
2151                 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2152                                   XATTR_LUSTRE_LOV,
2153                                   ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2154                                   LU_XATTR_SPLIT : LU_XATTR_MERGE);
2155                 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2156                         int rc2;
2157
2158                         mutex_lock(&o->mot_som_mutex);
2159                         rc2 = mdt_set_som(info, o, SOM_FL_STRICT,
2160                                           ma->ma_attr.la_size,
2161                                           ma->ma_attr.la_blocks);
2162                         mutex_unlock(&o->mot_som_mutex);
2163                         if (rc2 < 0)
2164                                 CERROR(DFID": Setting i_blocks error: %d, "
2165                                        "i_blocks will be reported wrongly and "
2166                                        "can only be fixed in next resync\n",
2167                                        PFID(mdt_object_fid(o)), rc2);
2168                 }
2169         }
2170         if (rc < 0)
2171                 GOTO(out_unlock2, rc);
2172
2173         EXIT;
2174
2175 out_unlock2:
2176         /* Release exclusive LL */
2177         mdt_object_unlock(info, o2, lh2, 1);
2178
2179 out_unlock1:
2180         mdt_object_unlock(info, o1, lh1, 1);
2181
2182 out_unlock_sem:
2183         up_write(&o->mot_open_sem);
2184
2185         /* already swapped */
2186         if (rc == 0) {
2187                 struct mdt_body *repbody;
2188
2189                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2190                 LASSERT(repbody != NULL);
2191                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2192         }
2193
2194 out_obj:
2195         mdt_object_put(info->mti_env, swap_objects ? o1 : o2);
2196
2197         ldlm_reprocess_all(lease->l_resource, lease);
2198
2199 out_lease:
2200         LDLM_LOCK_PUT(lease);
2201
2202         if (ma != NULL) {
2203                 ma->ma_valid = 0;
2204                 ma->ma_need = 0;
2205         }
2206
2207         return rc;
2208 }
2209
2210 static int mdt_close_resync_done(struct mdt_thread_info *info,
2211                                  struct mdt_object *o, struct md_attr *ma)
2212 {
2213         struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LOCAL];
2214         struct close_data *data;
2215         struct ldlm_lock *lease;
2216         struct md_layout_change layout = { 0 };
2217         __u32 *resync_ids = NULL;
2218         size_t resync_count = 0;
2219         bool lease_broken;
2220         int rc;
2221
2222         ENTRY;
2223
2224         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2225                 RETURN(-EROFS);
2226
2227         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2228                 RETURN(-EINVAL);
2229
2230         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2231         if (data == NULL)
2232                 RETURN(-EPROTO);
2233
2234         if (ptlrpc_req_need_swab(mdt_info_req(info)))
2235                 lustre_swab_close_data_resync_done(&data->cd_resync);
2236
2237         if (!fid_is_zero(&data->cd_fid))
2238                 RETURN(-EPROTO);
2239
2240         lease = ldlm_handle2lock(&data->cd_handle);
2241         if (lease == NULL)
2242                 RETURN(-ESTALE);
2243
2244         /* try to hold open_sem so that nobody else can open the file */
2245         if (!down_write_trylock(&o->mot_open_sem)) {
2246                 ldlm_lock_cancel(lease);
2247                 GOTO(out_reprocess, rc = -EBUSY);
2248         }
2249
2250         /* Check if the lease open lease has already canceled */
2251         lock_res_and_lock(lease);
2252         lease_broken = ldlm_is_cancel(lease);
2253         unlock_res_and_lock(lease);
2254
2255         LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2256                    PFID(mdt_object_fid(o)), lease_broken);
2257
2258         /* Cancel server side lease. Client side counterpart should
2259          * have been cancelled. It's okay to cancel it now as we've
2260          * held mot_open_sem. */
2261         ldlm_lock_cancel(lease);
2262
2263         if (lease_broken) /* don't perform release task */
2264                 GOTO(out_unlock, rc = -ESTALE);
2265
2266         resync_count = data->cd_resync.resync_count;
2267
2268         if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2269                 void *data;
2270
2271                 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2272                                            RCL_CLIENT))
2273                         GOTO(out_unlock, rc = -EPROTO);
2274
2275                 OBD_ALLOC(resync_ids, resync_count * sizeof(__u32));
2276                 if (!resync_ids)
2277                         GOTO(out_unlock, rc = -ENOMEM);
2278
2279                 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2280                 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2281
2282                 layout.mlc_resync_ids = resync_ids;
2283         } else {
2284                 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2285         }
2286
2287         layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2288         layout.mlc_resync_count = resync_count;
2289         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2290                 layout.mlc_som.lsa_valid = SOM_FL_STRICT;
2291                 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2292                 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2293         }
2294         rc = mdt_layout_change(info, o, lhc, &layout);
2295         if (rc)
2296                 GOTO(out_unlock, rc);
2297
2298         mdt_object_unlock(info, o, lhc, 0);
2299
2300         EXIT;
2301
2302 out_unlock:
2303         up_write(&o->mot_open_sem);
2304
2305         /* already released */
2306         if (rc == 0) {
2307                 struct mdt_body *repbody;
2308
2309                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2310                 LASSERT(repbody != NULL);
2311                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2312         }
2313
2314         if (resync_ids)
2315                 OBD_FREE(resync_ids, resync_count * sizeof(__u32));
2316
2317 out_reprocess:
2318         ldlm_reprocess_all(lease->l_resource, lease);
2319         LDLM_LOCK_PUT(lease);
2320
2321         ma->ma_valid = 0;
2322         ma->ma_need = 0;
2323
2324         return rc;
2325 }
2326
2327 #define MFD_CLOSED(open_flags) ((open_flags) == MDS_FMODE_CLOSED)
2328
2329 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2330 {
2331         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_open_flags));
2332 }
2333
2334 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2335 {
2336         struct mdt_object *o = mfd->mfd_object;
2337         struct md_object *next = mdt_object_child(o);
2338         struct md_attr *ma = &info->mti_attr;
2339         struct lu_fid *ofid = &info->mti_tmp_fid1;
2340         int rc = 0;
2341         u64 open_flags;
2342         u64 intent;
2343
2344         ENTRY;
2345
2346         open_flags = mfd->mfd_open_flags;
2347         intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2348         *ofid = *mdt_object_fid(o);
2349
2350         /* the below message is checked in replay-single.sh test_46 */
2351         CDEBUG(D_INODE, "%s: %sclosing file handle "DFID" with intent: %llx\n",
2352                mdt_obd_name(info->mti_mdt),
2353                ma->ma_valid & MA_FORCE_LOG ? "force " : "", PFID(ofid), intent);
2354
2355         switch (intent) {
2356         case MDS_HSM_RELEASE: {
2357                 rc = mdt_hsm_release(info, o, ma);
2358                 if (rc < 0) {
2359                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2360                                mdt_obd_name(info->mti_mdt),
2361                                PFID(ofid), rc);
2362                         /* continue to close even error occurred. */
2363                 }
2364                 break;
2365         }
2366         case MDS_CLOSE_LAYOUT_MERGE:
2367         case MDS_CLOSE_LAYOUT_SPLIT:
2368         case MDS_CLOSE_LAYOUT_SWAP: {
2369                 rc = mdt_close_handle_layouts(info, o, ma);
2370                 if (rc < 0) {
2371                         CDEBUG(D_INODE,
2372                                "%s: cannot swap layout of "DFID": rc = %d\n",
2373                                mdt_obd_name(info->mti_mdt),
2374                                PFID(ofid), rc);
2375                         /* continue to close even if error occurred. */
2376                 }
2377                 break;
2378         }
2379         case MDS_CLOSE_RESYNC_DONE:
2380                 rc = mdt_close_resync_done(info, o, ma);
2381                 break;
2382         default:
2383                 /* nothing */
2384                 break;
2385         }
2386
2387         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
2388             ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) {
2389                 int rc2;
2390
2391                 rc2 = mdt_lsom_update(info, o, false);
2392                 if (rc2 < 0)
2393                         CDEBUG(D_INODE,
2394                                "%s: File " DFID " LSOM failed: rc = %d\n",
2395                                mdt_obd_name(info->mti_mdt),
2396                                PFID(ofid), rc2);
2397                         /* continue to close even if error occured. */
2398         }
2399
2400         if (open_flags & MDS_FMODE_WRITE)
2401                 mdt_write_put(o);
2402         else if (open_flags & MDS_FMODE_EXEC)
2403                 mdt_write_allow(o);
2404
2405         /* Update atime|mtime|ctime on close. */
2406         if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ ||
2407              open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
2408             (ma->ma_attr.la_valid & LA_ATIME ||
2409              ma->ma_attr.la_valid & LA_MTIME ||
2410              ma->ma_attr.la_valid & LA_CTIME)) {
2411                 ma->ma_valid = MA_INODE;
2412                 ma->ma_attr_flags |= MDS_CLOSE_UPDATE_TIMES;
2413                 ma->ma_attr.la_valid &= (LA_ATIME | LA_MTIME | LA_CTIME);
2414                 rc = mo_attr_set(info->mti_env, next, ma);
2415         }
2416
2417         /* If file data is modified, add the dirty flag. */
2418         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2419                 rc = mdt_add_dirty_flag(info, o, ma);
2420
2421         ma->ma_need |= MA_INODE;
2422         ma->ma_valid &= ~MA_INODE;
2423
2424         LASSERT(atomic_read(&o->mot_open_count) > 0);
2425         atomic_dec(&o->mot_open_count);
2426         mdt_handle_last_unlink(info, o, ma);
2427
2428         if (!MFD_CLOSED(open_flags)) {
2429                 rc = mo_close(info->mti_env, next, ma, open_flags);
2430                 if (mdt_dom_check_for_discard(info, o))
2431                         mdt_dom_discard_data(info, o);
2432         }
2433
2434         /* adjust open and lease count */
2435         if (open_flags & MDS_OPEN_LEASE) {
2436                 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2437                 atomic_dec(&o->mot_lease_count);
2438         }
2439
2440         mdt_mfd_free(mfd);
2441         mdt_object_put(info->mti_env, o);
2442
2443         RETURN(rc);
2444 }
2445
2446 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2447                        struct mdt_body *repbody)
2448 {
2449         struct mdt_export_data *med;
2450         struct mdt_file_data   *mfd;
2451         int                     ret = 0;
2452         int                     rc = 0;
2453         ENTRY;
2454
2455         med = &req->rq_export->exp_mdt_data;
2456         spin_lock(&med->med_open_lock);
2457         mfd = mdt_open_handle2mfd(med, &info->mti_open_handle,
2458                                   req_is_replay(req));
2459         if (mdt_mfd_closed(mfd)) {
2460                 spin_unlock(&med->med_open_lock);
2461                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2462                        ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2463                        info->mti_open_handle.cookie);
2464                 /** not serious error since bug 3633 */
2465                 rc = -ESTALE;
2466         } else {
2467                 class_handle_unhash(&mfd->mfd_open_handle);
2468                 list_del_init(&mfd->mfd_list);
2469                 spin_unlock(&med->med_open_lock);
2470                 ret = mdt_mfd_close(info, mfd);
2471         }
2472
2473         RETURN(rc ? rc : ret);
2474 }
2475
2476 int mdt_close(struct tgt_session_info *tsi)
2477 {
2478         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2479         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2480         struct md_attr         *ma = &info->mti_attr;
2481         struct mdt_body        *repbody = NULL;
2482         int rc, ret = 0;
2483         ENTRY;
2484
2485         mdt_counter_incr(req, LPROC_MDT_CLOSE);
2486         /* Close may come with the Size-on-MDS update. Unpack it. */
2487         rc = mdt_close_unpack(info);
2488         if (rc)
2489                 GOTO(out, rc = err_serious(rc));
2490
2491         /* These fields are no longer used and are left for compatibility.
2492          * size is always zero */
2493         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2494                              0);
2495         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2496                              0);
2497         rc = req_capsule_server_pack(info->mti_pill);
2498         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2499                 mdt_client_compatibility(info);
2500                 if (rc == 0)
2501                         mdt_fix_reply(info);
2502                 mdt_exit_ucred(info);
2503                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2504         }
2505
2506         /* Continue to close handle even if we can not pack reply */
2507         if (rc == 0) {
2508                 repbody = req_capsule_server_get(info->mti_pill,
2509                                                  &RMF_MDT_BODY);
2510                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2511                                                     &RMF_MDT_MD);
2512                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2513                                                        &RMF_MDT_MD,
2514                                                        RCL_SERVER);
2515                 ma->ma_need = MA_INODE | MA_LOV;
2516                 repbody->mbo_eadatasize = 0;
2517                 repbody->mbo_aclsize = 0;
2518         } else {
2519                 rc = err_serious(rc);
2520         }
2521
2522         rc = mdt_close_internal(info, req, repbody);
2523         if (rc != -ESTALE)
2524                 mdt_empty_transno(info, rc);
2525
2526         if (repbody != NULL) {
2527                 mdt_client_compatibility(info);
2528                 rc = mdt_fix_reply(info);
2529         }
2530
2531         mdt_exit_ucred(info);
2532         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2533                 GOTO(out, rc = err_serious(-ENOMEM));
2534
2535         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2536                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2537                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2538 out:
2539         mdt_thread_info_fini(info);
2540         RETURN(rc ? rc : ret);
2541 }