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         __u32 dom_stripe = 0;
 813         unsigned int dom_only = 0;
 814         unsigned int dom_lock = 0;
 815
 816         ENTRY;
 817
 818         *ibits = 0;
 819         mdt_lock_handle_init(lhc);
 820
 821         if (req_is_replay(mdt_info_req(info)))
 822                 RETURN(0);
 823
 824         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
 825                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
 826                     md_should_create(open_flags))
 827                         create_layout = true;
 828                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
 829                     ma->ma_need & MA_LOV)
 830                         try_layout = true;
 831
 832                 /* DoM files can take IO lock at OPEN when it makes sense,
 833                  * check if file has DoM stripe and ask for lock if client
 834                  * no lock on that resource yet.
 835                  */
 836                 if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
 837                         dom_stripe = mdt_lmm_dom_entry_check(ma->ma_lmm,
 838                                                              &dom_only);
 839                 /* If only DOM stripe is being used then we can expect IO
 840                  * to it after OPEN and will return corresponding DOM ibit
 841                  * using default strategy from mdt_opts.mo_dom_lock.
 842                  * Otherwise trylock mode is used always and DOM ibit will
 843                  * be returned optionally.
 844                  */
 845                 if (dom_stripe &&
 846                     !mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
 847                         dom_lock = !dom_only ? TRYLOCK_DOM_ON_OPEN :
 848                                    info->mti_mdt->mdt_opts.mo_dom_lock;
 849         }
 850
 851         if (acq_lease) {
 852                 /* lease open, acquire write mode of open sem */
 853                 down_write(&obj->mot_open_sem);
 854
 855                 /* Lease exists and ask for new lease */
 856                 if (atomic_read(&obj->mot_lease_count) > 0) {
 857                         /* only exclusive open is supported, so lease
 858                          * are conflicted to each other */
 859                         GOTO(out, rc = -EBUSY);
 860                 }
 861
 862                 /* Lease must be with open lock */
 863                 if (!(open_flags & MDS_OPEN_LOCK)) {
 864                         CERROR("%s: Request lease for file:"DFID ", but open lock "
 865                                "is missed, open_flags = %#llo : rc = %d\n",
 866                                mdt_obd_name(info->mti_mdt),
 867                                PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
 868                         GOTO(out, rc = -EPROTO);
 869                 }
 870
 871                 /* XXX: only exclusive open is supported. */
 872                 lm = LCK_EX;
 873                 *ibits = MDS_INODELOCK_OPEN;
 874
 875                 /* never grant LCK_EX layout lock to client */
 876                 try_layout = false;
 877         } else { /* normal open */
 878                 /* normal open holds read mode of open sem */
 879                 down_read(&obj->mot_open_sem);
 880
 881                 if (open_flags & MDS_OPEN_LOCK) {
 882                         if (open_flags & MDS_FMODE_WRITE)
 883                                 lm = LCK_CW;
 884                         else if (open_flags & MDS_FMODE_EXEC)
 885                                 lm = LCK_PR;
 886                         else
 887                                 lm = LCK_CR;
 888
 889                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
 890                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
 891                         if (open_flags & MDS_FMODE_WRITE)
 892                                 lm = LCK_CW;
 893                         else
 894                                 lm = LCK_CR;
 895
 896                         /* revoke lease */
 897                         *ibits = MDS_INODELOCK_OPEN;
 898                         try_layout = false;
 899
 900                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 901                 } else if (dom_lock) {
 902                         lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR;
 903                         trybits |= MDS_INODELOCK_DOM | MDS_INODELOCK_LAYOUT;
 904                 }
 905
 906                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
 907                         PFID(mdt_object_fid(obj)),
 908                         atomic_read(&obj->mot_lease_count), lm);
 909         }
 910
 911         mdt_lock_reg_init(lhc, lm);
 912
 913         /* Return lookup lock to validate inode at the client side.
 914          * This is pretty important otherwise MDT will return layout
 915          * lock for each open.
 916          * However this is a double-edged sword because changing
 917          * permission will revoke a huge number of LOOKUP locks.
 918          */
 919         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
 920                 if (!(*ibits & MDS_INODELOCK_LOOKUP))
 921                         trybits |= MDS_INODELOCK_LOOKUP;
 922                 trybits |= MDS_INODELOCK_LAYOUT;
 923         }
 924
 925         if (*ibits | trybits)
 926                 rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, false);
 927
 928         CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
 929                ", open_flags = %#llo, try_layout = %d : rc = %d\n",
 930                mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
 931                *ibits, trybits, open_flags, try_layout, rc);
 932
 933         /* will change layout, revoke layout locks by enqueuing EX lock. */
 934         if (rc == 0 && create_layout) {
 935                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
 936
 937                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
 938                         ", open_flags = %#llo\n",
 939                         PFID(mdt_object_fid(obj)), open_flags);
 940
 941                 /* We cannot enqueue another lock for the same resource we
 942                  * already have a lock for, due to mechanics of waiting list
 943                  * iterating in ldlm, see LU-3601.
 944                  * As such we'll drop the open lock we just got above here,
 945                  * it's ok not to have this open lock as it's main purpose is to
 946                  * flush unused cached client open handles. */
 947                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 948                         mdt_object_unlock(info, obj, lhc, 1);
 949
 950                 LASSERT(!try_layout);
 951                 mdt_lock_handle_init(ll);
 952                 mdt_lock_reg_init(ll, LCK_EX);
 953                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
 954
 955                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
 956         }
 957
 958         /* Check if there is any other open handles after acquiring
 959          * open lock. At this point, caching open handles have been revoked
 960          * by open lock.
 961          * XXX: Now only exclusive open is supported. Need to check the
 962          * type of open for generic lease support. */
 963         if (rc == 0 && acq_lease) {
 964                 struct ptlrpc_request *req = mdt_info_req(info);
 965                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 966                 struct mdt_file_data *mfd;
 967                 bool is_replay_or_resent;
 968                 int open_count = 0;
 969
 970                 /* For lease: application can open a file and then apply lease,
 971                  * @handle contains original open handle in that case.
 972                  * In recovery, open REQ will be replayed and the lease REQ may
 973                  * be resent that means the open handle is already stale, so we
 974                  * need to fix it up here by finding new handle. */
 975                 is_replay_or_resent = req_is_replay(req) ||
 976                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
 977
 978                 /* if the request is _not_ a replay request, rr_open_handle
 979                  * may be used to hold an open file handle which is issuing the
 980                  * lease request, so that this openhandle doesn't count. */
 981                 mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 982                                           is_replay_or_resent);
 983                 if (mfd != NULL)
 984                         ++open_count;
 985
 986                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
 987                         PFID(mdt_object_fid(obj)),
 988                         atomic_read(&obj->mot_open_count), open_count);
 989
 990                 if (atomic_read(&obj->mot_open_count) > open_count)
 991                         GOTO(out, rc = -EBUSY);
 992         }
 993         GOTO(out, rc);
 994
 995 out:
 996         RETURN(rc);
 997 }
 998
 999 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1000                                    struct mdt_object *obj,
1001                                    struct mdt_lock_handle *lhc,
1002                                    __u64 ibits, int rc)
1003 {
1004         __u64 open_flags = info->mti_spec.sp_cr_flags;
1005         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1006         ENTRY;
1007
1008         if (req_is_replay(mdt_info_req(info)))
1009                 RETURN_EXIT;
1010
1011         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1012          * return to client side. */
1013         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1014                 mdt_object_unlock(info, obj, ll, 1);
1015
1016         ll = &info->mti_lh[MDT_LH_LAYOUT];
1017         /* Release local layout lock, layout was created */
1018         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1019                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1020                 mdt_object_unlock(info, obj, ll, 1);
1021         }
1022
1023         if (open_flags & MDS_OPEN_LEASE)
1024                 up_write(&obj->mot_open_sem);
1025         else
1026                 up_read(&obj->mot_open_sem);
1027
1028         /* Cross-ref case, the lock should be returned to the client */
1029         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1030                 RETURN_EXIT;
1031
1032         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1033             !(ibits & MDS_INODELOCK_DOM)) {
1034                 /* for the open request, the lock will only return to client
1035                  * if open or layout lock is granted. */
1036                 rc = 1;
1037         }
1038
1039         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1040                 struct ldlm_reply       *ldlm_rep;
1041
1042                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1043                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1044                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1045                         mdt_object_unlock(info, obj, lhc, 1);
1046         }
1047         RETURN_EXIT;
1048 }
1049
1050 /**
1051  * Check release is permitted for the current HSM flags.
1052  */
1053 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1054 {
1055         if (!(ma->ma_valid & MA_HSM))
1056                 return false;
1057
1058         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1059                 return false;
1060
1061         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1062                 return false;
1063
1064         return true;
1065 }
1066
1067 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1068                                 struct ldlm_reply *rep,
1069                                 struct mdt_lock_handle *lhc)
1070 {
1071         const struct lu_env *env = info->mti_env;
1072         struct mdt_device *mdt = info->mti_mdt;
1073         u64 open_flags = info->mti_spec.sp_cr_flags;
1074         struct mdt_reint_record *rr = &info->mti_rr;
1075         struct md_attr *ma = &info->mti_attr;
1076         struct mdt_object *parent = NULL;
1077         struct mdt_object *o;
1078         bool object_locked = false;
1079         u64 ibits = 0;
1080         int rc;
1081
1082         ENTRY;
1083         if (md_should_create(open_flags)) {
1084                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1085                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1086                         if (IS_ERR(parent)) {
1087                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1088                                        " for anonymous created %ld, try to"
1089                                        " use server-side parent.\n",
1090                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1091                                 parent = NULL;
1092                         }
1093                 }
1094                 if (parent == NULL)
1095                         ma->ma_need |= MA_PFID;
1096         }
1097
1098         o = mdt_object_find(env, mdt, rr->rr_fid2);
1099         if (IS_ERR(o))
1100                 GOTO(out_parent_put, rc = PTR_ERR(o));
1101
1102         if (mdt_object_remote(o)) {
1103                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1104                        mdt_obd_name(info->mti_mdt),
1105                        PFID(rr->rr_fid2));
1106                 GOTO(out, rc = -EREMOTE);
1107         } else if (!mdt_object_exists(o)) {
1108                 mdt_set_disposition(info, rep,
1109                                     DISP_IT_EXECD |
1110                                     DISP_LOOKUP_EXECD |
1111                                     DISP_LOOKUP_NEG);
1112                 GOTO(out, rc = -ENOENT);
1113         }
1114
1115         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1116
1117         mdt_prep_ma_buf_from_rep(info, o, ma);
1118         if (open_flags & MDS_OPEN_RELEASE)
1119                 ma->ma_need |= MA_HSM;
1120         rc = mdt_attr_get_complex(info, o, ma);
1121         if (rc)
1122                 GOTO(out, rc);
1123
1124         /* We should not change file's existing LOV EA */
1125         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1126             open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1127                 GOTO(out, rc = -EEXIST);
1128
1129         /* If a release request, check file open flags are fine and ask for an
1130          * exclusive open access. */
1131         if (open_flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1132                 GOTO(out, rc = -EPERM);
1133
1134         rc = mdt_check_resent_lock(info, o, lhc);
1135         if (rc < 0) {
1136                 GOTO(out, rc);
1137         } else if (rc > 0) {
1138                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1139                 object_locked = true;
1140                 if (rc)
1141                         GOTO(out_unlock, rc);
1142         }
1143
1144         if (ma->ma_valid & MA_PFID) {
1145                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1146                 if (IS_ERR(parent)) {
1147                         CDEBUG(D_INODE, "Fail to find parent "DFID
1148                                " for anonymous created %ld, try to"
1149                                " use system default.\n",
1150                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1151                         parent = NULL;
1152                 }
1153         }
1154
1155         rc = mdt_finish_open(info, parent, o, open_flags, 0, rep);
1156         if (!rc) {
1157                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1158                 if (open_flags & MDS_OPEN_LOCK)
1159                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1160                 if (open_flags & MDS_OPEN_LEASE)
1161                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1162         }
1163         GOTO(out_unlock, rc);
1164
1165 out_unlock:
1166         if (object_locked)
1167                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1168 out:
1169         mdt_object_put(env, o);
1170         if (rc == 0)
1171                 mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
1172 out_parent_put:
1173         if (parent != NULL)
1174                 mdt_object_put(env, parent);
1175         return rc;
1176 }
1177
1178 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1179 static int mdt_cross_open(struct mdt_thread_info *info,
1180                           const struct lu_fid *parent_fid,
1181                           const struct lu_fid *fid,
1182                           struct ldlm_reply *rep, u64 open_flags)
1183 {
1184         struct md_attr *ma = &info->mti_attr;
1185         struct mdt_object *o;
1186         int rc;
1187         ENTRY;
1188
1189         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1190         if (IS_ERR(o))
1191                 RETURN(rc = PTR_ERR(o));
1192
1193         if (mdt_object_remote(o)) {
1194                 /* Something is wrong here, the object is on another MDS! */
1195                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1196                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1197                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1198                                 &o->mot_obj,
1199                                 "Object isn't on this server! FLD error?");
1200                 rc = -EFAULT;
1201         } else {
1202                 if (mdt_object_exists(o)) {
1203                         int mask;
1204
1205                         /* Do permission check for cross-open after converting
1206                          * MDS_OPEN_* flags to MAY_* permission mask.
1207                          */
1208                         mask = mds_accmode(open_flags);
1209
1210                         rc = mo_permission(info->mti_env, NULL,
1211                                            mdt_object_child(o), NULL, mask);
1212                         if (rc)
1213                                 goto out;
1214
1215                         mdt_prep_ma_buf_from_rep(info, o, ma);
1216                         rc = mdt_attr_get_complex(info, o, ma);
1217                         if (rc != 0)
1218                                 GOTO(out, rc);
1219
1220                         rc = mdt_pack_secctx_in_reply(info, o);
1221                         if (unlikely(rc))
1222                                 GOTO(out, rc);
1223
1224                         rc = mdt_pack_encctx_in_reply(info, o);
1225                         if (unlikely(rc))
1226                                 GOTO(out, rc);
1227
1228                         rc = mdt_finish_open(info, NULL, o, open_flags, 0, rep);
1229                 } else {
1230                         /*
1231                          * Something is wrong here. lookup was positive but
1232                          * there is no object!
1233                          */
1234                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1235                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1236                         rc = -EFAULT;
1237                 }
1238         }
1239 out:
1240         mdt_object_put(info->mti_env, o);
1241
1242         RETURN(rc);
1243 }
1244
1245 /*
1246  * find root object and take its xattr lock if it's on remote MDT, later create
1247  * may use fs default striping (which is stored in root xattr).
1248  */
1249 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1250                                struct mdt_device *mdt)
1251 {
1252         struct mdt_object *md_root = mdt->mdt_md_root;
1253         struct lustre_handle lhroot;
1254         int rc;
1255
1256         if (md_root == NULL) {
1257                 lu_root_fid(&info->mti_tmp_fid1);
1258                 md_root = mdt_object_find(info->mti_env, mdt,
1259                                           &info->mti_tmp_fid1);
1260                 if (IS_ERR(md_root))
1261                         return PTR_ERR(md_root);
1262
1263                 spin_lock(&mdt->mdt_lock);
1264                 if (mdt->mdt_md_root != NULL) {
1265                         spin_unlock(&mdt->mdt_lock);
1266
1267                         LASSERTF(mdt->mdt_md_root == md_root,
1268                                  "Different root object ("
1269                                  DFID") instances, %p, %p\n",
1270                                  PFID(&info->mti_tmp_fid1),
1271                                  mdt->mdt_md_root, md_root);
1272                         LASSERT(atomic_read(
1273                                 &md_root->mot_obj.lo_header->loh_ref) > 1);
1274
1275                         mdt_object_put(info->mti_env, md_root);
1276                 } else {
1277                         mdt->mdt_md_root = md_root;
1278                         spin_unlock(&mdt->mdt_lock);
1279                 }
1280         }
1281
1282         if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1283                 return 0;
1284
1285         rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root),
1286                                     &lhroot, LCK_PR, MDS_INODELOCK_XATTR,
1287                                     true);
1288         if (rc < 0)
1289                 return rc;
1290
1291         md_root->mot_cache_attr = 1;
1292
1293         /* don't cancel this lock, so that we know the cached xattr is valid. */
1294         ldlm_lock_decref(&lhroot, LCK_PR);
1295
1296         return 0;
1297 }
1298
1299 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1300 {
1301         struct mdt_device *mdt = info->mti_mdt;
1302         struct ptlrpc_request *req = mdt_info_req(info);
1303         struct mdt_object *parent;
1304         struct mdt_object *child;
1305         struct mdt_lock_handle *lh;
1306         struct ldlm_reply *ldlm_rep;
1307         struct mdt_body *repbody;
1308         struct lu_fid *child_fid = &info->mti_tmp_fid1;
1309         struct md_attr *ma = &info->mti_attr;
1310         u64 open_flags = info->mti_spec.sp_cr_flags;
1311         u64 ibits = 0;
1312         struct mdt_reint_record *rr = &info->mti_rr;
1313         int result, rc;
1314         int created = 0;
1315         int object_locked = 0;
1316         enum ldlm_mode lock_mode = LCK_PR;
1317         u32 msg_flags;
1318         ktime_t kstart = ktime_get();
1319
1320         ENTRY;
1321         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1322                                (obd_timeout + 1) / 4);
1323
1324         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1325
1326         ma->ma_need = MA_INODE;
1327         ma->ma_valid = 0;
1328
1329         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1330         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1331
1332         if (unlikely(open_flags & MDS_OPEN_JOIN_FILE)) {
1333                 CERROR("file join is not supported anymore.\n");
1334                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1335         }
1336         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1337
1338         if ((open_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1339             info->mti_spec.u.sp_ea.eadata == NULL)
1340                 GOTO(out, result = err_serious(-EINVAL));
1341
1342         if (open_flags & MDS_FMODE_WRITE &&
1343             exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1344                 GOTO(out, result = -EROFS);
1345
1346         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1347                "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1348                PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2),
1349                open_flags, ma->ma_attr.la_mode, msg_flags);
1350
1351         if (info->mti_cross_ref) {
1352                 /* This is cross-ref open */
1353                 mdt_set_disposition(info, ldlm_rep,
1354                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1355                              DISP_LOOKUP_POS));
1356                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1357                                         ldlm_rep, open_flags);
1358                 GOTO(out, result);
1359         } else if (req_is_replay(req)) {
1360                 result = mdt_open_by_fid(info, ldlm_rep);
1361
1362                 if (result != -ENOENT)
1363                         GOTO(out, result);
1364
1365                 /* We didn't find the correct object, so we need to re-create it
1366                  * via a regular replay. */
1367                 if (!(open_flags & MDS_OPEN_CREAT)) {
1368                         DEBUG_REQ(D_ERROR, req,
1369                                   "OPEN & CREAT not in open replay/by_fid");
1370                         GOTO(out, result = -EFAULT);
1371                 }
1372                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1373         } else if (open_flags & MDS_OPEN_BY_FID) {
1374                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1375                 if (result < 0)
1376                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1377                                PFID(rr->rr_fid2), result);
1378                 GOTO(out, result);
1379         }
1380
1381         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1382                 GOTO(out, result = err_serious(-ENOMEM));
1383
1384         mdt_set_disposition(info, ldlm_rep,
1385                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1386
1387         if (!lu_name_is_valid(&rr->rr_name))
1388                 GOTO(out, result = -EPROTO);
1389
1390         result = mdt_lock_root_xattr(info, mdt);
1391         if (result < 0)
1392                 GOTO(out, result);
1393
1394         parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1395         if (IS_ERR(parent))
1396                 GOTO(out, result = PTR_ERR(parent));
1397
1398         /* get and check version of parent */
1399         result = mdt_version_get_check(info, parent, 0);
1400         if (result) {
1401                 mdt_object_put(info->mti_env, parent);
1402                 GOTO(out, result);
1403         }
1404
1405         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
1406 again_pw:
1407         lh = &info->mti_lh[MDT_LH_PARENT];
1408         mdt_lock_pdo_init(lh, lock_mode, &rr->rr_name);
1409
1410         result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1411         if (result != 0) {
1412                 mdt_object_put(info->mti_env, parent);
1413                 GOTO(out, result);
1414         }
1415         fid_zero(child_fid);
1416
1417         result = -ENOENT;
1418         if ((open_flags & MDS_OPEN_VOLATILE) == 0)
1419                 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1420                                     &rr->rr_name, child_fid, &info->mti_spec);
1421
1422         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1423                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1424                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1425                  PFID(child_fid));
1426
1427         if (result != 0 && result != -ENOENT)
1428                 GOTO(out_parent, result);
1429
1430         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1431
1432         if (result == -ENOENT) {
1433                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1434                 if (!(open_flags & MDS_OPEN_CREAT))
1435                         GOTO(out_parent, result);
1436                 if (mdt_rdonly(req->rq_export))
1437                         GOTO(out_parent, result = -EROFS);
1438
1439                 if (lock_mode == LCK_PR) {
1440                         /* first pass: get write lock and restart */
1441                         mdt_object_unlock(info, parent, lh, 1);
1442                         mdt_clear_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1443                         mdt_lock_handle_init(lh);
1444                         lock_mode = LCK_PW;
1445                         goto again_pw;
1446                 }
1447
1448                 *child_fid = *info->mti_rr.rr_fid2;
1449                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1450                          PFID(child_fid));
1451                 /* In the function below, .hs_keycmp resolves to
1452                  * lu_obj_hop_keycmp() */
1453                 /* coverity[overrun-buffer-val] */
1454                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1455         } else {
1456                 /*
1457                  * Check for O_EXCL is moved to the mdt_finish_open(),
1458                  * we need to return FID back in that case.
1459                  */
1460                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1461                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1462         }
1463         if (IS_ERR(child))
1464                 GOTO(out_parent, result = PTR_ERR(child));
1465
1466         /** check version of child  */
1467         rc = mdt_version_get_check(info, child, 1);
1468         if (rc)
1469                 GOTO(out_child, result = rc);
1470
1471         if (result == -ENOENT) {
1472                 /* Create under OBF and .lustre is not permitted */
1473                 if (!fid_is_md_operative(rr->rr_fid1) &&
1474                     (open_flags & MDS_OPEN_VOLATILE) == 0)
1475                         GOTO(out_child, result = -EPERM);
1476
1477                 /* save versions in reply */
1478                 mdt_version_get_save(info, parent, 0);
1479                 mdt_version_get_save(info, child, 1);
1480
1481                 /* version of child will be changed */
1482                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1483
1484                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1485                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1486
1487                 /* Let lower layers know what is lock mode on directory. */
1488                 info->mti_spec.sp_cr_mode =
1489                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1490
1491                 /* Don't do lookup sanity check. We know name doesn't exist. */
1492                 info->mti_spec.sp_cr_lookup = 0;
1493                 info->mti_spec.sp_feat = &dt_directory_features;
1494
1495                 result = mdo_create(info->mti_env, mdt_object_child(parent),
1496                                     &rr->rr_name, mdt_object_child(child),
1497                                     &info->mti_spec, &info->mti_attr);
1498                 if (result == -ERESTART) {
1499                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1500                         GOTO(out_child, result);
1501                 } else {
1502                         mdt_prep_ma_buf_from_rep(info, child, ma);
1503                         /* XXX: we should call this once, see few lines below */
1504                         if (result == 0)
1505                                 result = mdt_attr_get_complex(info, child, ma);
1506
1507                         if (result != 0)
1508                                 GOTO(out_child, result);
1509                 }
1510                 created = 1;
1511                 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1512                                  ktime_us_delta(ktime_get(), kstart));
1513         } else {
1514                 /*
1515                  * The object is on remote node, return its FID for remote open.
1516                  */
1517                 if (mdt_object_remote(child)) {
1518                         /*
1519                          * Check if this lock already was sent to client and
1520                          * this is resent case. For resent case do not take lock
1521                          * again, use what is already granted.
1522                          */
1523                         LASSERT(lhc != NULL);
1524
1525                         rc = mdt_check_resent_lock(info, child, lhc);
1526                         if (rc < 0) {
1527                                 GOTO(out_child, result = rc);
1528                         } else if (rc > 0) {
1529                                 mdt_lock_handle_init(lhc);
1530                                 mdt_lock_reg_init(lhc, LCK_PR);
1531
1532                                 rc = mdt_object_lock(info, child, lhc,
1533                                                      MDS_INODELOCK_LOOKUP);
1534                         }
1535                         repbody->mbo_fid1 = *mdt_object_fid(child);
1536                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1537                         if (rc != 0)
1538                                 result = rc;
1539                         else
1540                                 result = -MDT_EREMOTE_OPEN;
1541                         GOTO(out_child, result);
1542                 } else if (mdt_object_exists(child)) {
1543                         /* Check early for MDS_OPEN_DIRECTORY/O_DIRECTORY to
1544                          * avoid opening regular files from lfs getstripe
1545                          * since doing so breaks the leases used by lfs
1546                          * mirror. See LU-13693. */
1547                         if (open_flags & MDS_OPEN_DIRECTORY &&
1548                             S_ISREG(lu_object_attr(&child->mot_obj)))
1549                                 GOTO(out_child, result = -ENOTDIR);
1550
1551                         /* We have to get attr & LOV EA & HSM for this
1552                          * object. */
1553                         mdt_prep_ma_buf_from_rep(info, child, ma);
1554                         ma->ma_need |= MA_HSM;
1555                         result = mdt_attr_get_complex(info, child, ma);
1556                         if (result != 0)
1557                                 GOTO(out_child, result);
1558                 } else {
1559                         /* Object does not exist. Likely FS corruption. */
1560                         CERROR("%s: name '"DNAME"' present, but FID "
1561                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1562                                PNAME(&rr->rr_name), PFID(child_fid));
1563                         GOTO(out_child, result = -EIO);
1564                 }
1565         }
1566
1567         repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1568         repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1569
1570         rc = mdt_pack_secctx_in_reply(info, child);
1571         if (unlikely(rc))
1572                 GOTO(out_child, result = rc);
1573
1574         rc = mdt_pack_encctx_in_reply(info, child);
1575         if (unlikely(rc))
1576                 GOTO(out_child, result = rc);
1577
1578         rc = mdt_check_resent_lock(info, child, lhc);
1579         if (rc < 0) {
1580                 GOTO(out_child, result = rc);
1581         } else if (rc == 0) {
1582                 /* the open lock might already be gotten in
1583                  * ldlm_handle_enqueue() */
1584                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1585                 if (open_flags & MDS_OPEN_LOCK)
1586                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1587         } else {
1588                 /* get openlock if this isn't replay and client requested it */
1589                 if (!req_is_replay(req)) {
1590                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1591                         object_locked = 1;
1592                         if (rc != 0)
1593                                 GOTO(out_child_unlock, result = rc);
1594                         else if (open_flags & MDS_OPEN_LOCK)
1595                                 mdt_set_disposition(info, ldlm_rep,
1596                                                     DISP_OPEN_LOCK);
1597                 }
1598         }
1599         /* Try to open it now. */
1600         rc = mdt_finish_open(info, parent, child, open_flags,
1601                              created, ldlm_rep);
1602         if (rc) {
1603                 result = rc;
1604                 /* openlock will be released if mdt_finish_open() failed */
1605                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1606
1607                 if (created && (open_flags & MDS_OPEN_VOLATILE)) {
1608                         CERROR("%s: cannot open volatile file "DFID", orphan "
1609                                "file will be left in PENDING directory until "
1610                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1611                                PFID(mdt_object_fid(child)), rc);
1612                         GOTO(out_child_unlock, result);
1613                 }
1614
1615                 if (created) {
1616                         ma->ma_need = 0;
1617                         ma->ma_valid = 0;
1618                         rc = mdo_unlink(info->mti_env,
1619                                         mdt_object_child(parent),
1620                                         mdt_object_child(child),
1621                                         &rr->rr_name,
1622                                         &info->mti_attr, 0);
1623                         if (rc != 0)
1624                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1625                                        mdt_obd_name(info->mti_mdt),
1626                                        PFID(mdt_object_fid(child)), rc);
1627                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1628                 }
1629         }
1630
1631         mdt_counter_incr(req, LPROC_MDT_OPEN,
1632                          ktime_us_delta(ktime_get(), kstart));
1633
1634         EXIT;
1635 out_child_unlock:
1636         if (object_locked)
1637                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1638 out_child:
1639         mdt_object_put(info->mti_env, child);
1640         if (result == 0)
1641                 mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh);
1642 out_parent:
1643         mdt_object_unlock_put(info, parent, lh, result || !created);
1644 out:
1645         if (result)
1646                 lustre_msg_set_transno(req->rq_repmsg, 0);
1647         return result;
1648 }
1649
1650 /**
1651  * Create an orphan object use local root.
1652  */
1653 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1654                                           struct mdt_device *mdt,
1655                                           const struct lu_fid *fid,
1656                                           struct md_attr *attr, fmode_t fmode)
1657 {
1658         const struct lu_env *env = info->mti_env;
1659         struct md_op_spec *spec = &info->mti_spec;
1660         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1661         struct mdt_object *obj = NULL;
1662         struct mdt_object *local_root;
1663         static const struct lu_name lname = {
1664                 .ln_name = "i_am_nobody",
1665                 .ln_namelen = sizeof("i_am_nobody") - 1,
1666         };
1667         struct lu_ucred *uc;
1668         cfs_cap_t uc_cap_save;
1669         int rc;
1670         ENTRY;
1671
1672         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1673         if (rc != 0)
1674                 RETURN(ERR_PTR(rc));
1675
1676         local_root = mdt_object_find(env, mdt, local_root_fid);
1677         if (IS_ERR(local_root))
1678                 RETURN(local_root);
1679
1680         obj = mdt_object_new(env, mdt, fid);
1681         if (IS_ERR(obj))
1682                 GOTO(out, rc = PTR_ERR(obj));
1683
1684         spec->sp_cr_lookup = 0;
1685         spec->sp_feat = &dt_directory_features;
1686         spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1687         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1688         if (attr->ma_valid & MA_LOV) {
1689                 spec->u.sp_ea.eadata = attr->ma_lmm;
1690                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1691                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1692         } else {
1693                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1694         }
1695
1696         uc = lu_ucred(env);
1697         uc_cap_save = uc->uc_cap;
1698         uc->uc_cap |= BIT(CAP_DAC_OVERRIDE);
1699         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1700                         mdt_object_child(obj), spec, attr);
1701         uc->uc_cap = uc_cap_save;
1702         if (rc < 0) {
1703                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1704                        mdt_obd_name(mdt), PFID(fid), rc);
1705                 GOTO(out, rc);
1706         }
1707
1708         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED, spec);
1709         if (rc < 0)
1710                 CERROR("%s: cannot open volatile file "DFID", orphan "
1711                        "file will be left in PENDING directory until "
1712                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1713                        PFID(fid), rc);
1714         GOTO(out, rc);
1715
1716 out:
1717         if (rc < 0) {
1718                 if (!IS_ERR(obj))
1719                         mdt_object_put(env, obj);
1720                 obj = ERR_PTR(rc);
1721         }
1722         mdt_object_put(env, local_root);
1723         return obj;
1724 }
1725
1726 /* XXX Look into layout in MDT layer. */
1727 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1728 {
1729         struct lov_comp_md_v1   *comp_v1;
1730         struct lov_mds_md       *v1;
1731         __u32   off;
1732         int     i;
1733
1734         if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1735                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1736
1737                 if (comp_v1->lcm_entry_count == 0)
1738                         return -EINVAL;
1739
1740                 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1741                         off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1742                         v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1743                         v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1744                 }
1745         } else {
1746                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1747         }
1748         return 0;
1749 }
1750
1751 static inline int mdt_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1752 {
1753         struct lov_comp_md_v1 *comp_v1;
1754
1755         if (le32_to_cpu(lmm->lmm_magic == LOV_MAGIC_COMP_V1)) {
1756                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1757                 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1758         } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1759                    le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1760                 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1761         } else {
1762                 return -EINVAL;
1763         }
1764         return 0;
1765 }
1766
1767 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1768                            struct md_attr *ma)
1769 {
1770         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1771         struct lu_ucred        *uc = mdt_ucred(info);
1772         struct close_data      *data;
1773         struct ldlm_lock       *lease;
1774         struct mdt_object      *orphan;
1775         struct md_attr         *orp_ma;
1776         struct lu_buf          *buf;
1777         cfs_cap_t               cap;
1778         bool                    lease_broken;
1779         int                     rc;
1780         int                     rc2;
1781         ENTRY;
1782
1783         if (mdt_rdonly(info->mti_exp))
1784                 RETURN(-EROFS);
1785
1786         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1787         if (data == NULL)
1788                 RETURN(-EPROTO);
1789
1790         lease = ldlm_handle2lock(&data->cd_handle);
1791         if (lease == NULL)
1792                 RETURN(-ESTALE);
1793
1794         /* try to hold open_sem so that nobody else can open the file */
1795         if (!down_write_trylock(&o->mot_open_sem)) {
1796                 ldlm_lock_cancel(lease);
1797                 GOTO(out_reprocess, rc = -EBUSY);
1798         }
1799
1800         /* Check if the lease open lease has already canceled */
1801         lock_res_and_lock(lease);
1802         lease_broken = ldlm_is_cancel(lease);
1803         unlock_res_and_lock(lease);
1804
1805         LDLM_DEBUG(lease, DFID " lease broken? %d",
1806                    PFID(mdt_object_fid(o)), lease_broken);
1807
1808         /* Cancel server side lease. Client side counterpart should
1809          * have been cancelled. It's okay to cancel it now as we've
1810          * held mot_open_sem. */
1811         ldlm_lock_cancel(lease);
1812
1813         if (lease_broken) /* don't perform release task */
1814                 GOTO(out_unlock, rc = -ESTALE);
1815
1816         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1817                 GOTO(out_unlock, rc = -EINVAL);
1818
1819         /* ma_need was set before but it seems fine to change it in order to
1820          * avoid modifying the one from RPC */
1821         ma->ma_need = MA_HSM;
1822         rc = mdt_attr_get_complex(info, o, ma);
1823         if (rc != 0)
1824                 GOTO(out_unlock, rc);
1825
1826         if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
1827                 if (ma->ma_valid & MA_HSM) {
1828                         if (ma->ma_hsm.mh_flags & HS_RELEASED)
1829                                 GOTO(out_unlock, rc = -EALREADY);
1830
1831                         if (ma->ma_hsm.mh_arch_id != data->cd_archive_id)
1832                                 CDEBUG(D_CACHE,
1833                                        DFID" archive id diff: %llu:%u\n",
1834                                        PFID(mdt_object_fid(o)),
1835                                        ma->ma_hsm.mh_arch_id,
1836                                        data->cd_archive_id);
1837
1838                         if (!(ma->ma_hsm.mh_flags & HS_DIRTY) &&
1839                             ma->ma_hsm.mh_arch_ver == data->cd_data_version) {
1840                                 CDEBUG(D_CACHE,
1841                                        DFID" data version matches: packed=%llu "
1842                                        "and on-disk=%llu\n",
1843                                        PFID(mdt_object_fid(o)),
1844                                        data->cd_data_version,
1845                                        ma->ma_hsm.mh_arch_ver);
1846                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1847                         }
1848
1849                         if (ma->ma_hsm.mh_flags & HS_DIRTY)
1850                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1851                 } else {
1852                         /* Set up HSM attribte for PCC archived object */
1853                         BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1854                                      sizeof(info->mti_xattr_buf));
1855                         buf = &info->mti_buf;
1856                         buf->lb_buf = info->mti_xattr_buf;
1857                         buf->lb_len = sizeof(struct hsm_attrs);
1858                         memset(&ma->ma_hsm, 0, sizeof(ma->ma_hsm));
1859                         ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1860                         ma->ma_hsm.mh_arch_id = data->cd_archive_id;
1861                         ma->ma_hsm.mh_arch_ver = data->cd_data_version;
1862                         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1863
1864                         rc = mo_xattr_set(info->mti_env, mdt_object_child(o),
1865                                           buf, XATTR_NAME_HSM, 0);
1866                         if (rc)
1867                                 GOTO(out_unlock, rc);
1868                 }
1869         } else {
1870                 if (!mdt_hsm_release_allow(ma))
1871                         GOTO(out_unlock, rc = -EPERM);
1872
1873                 /* already released? */
1874                 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1875                         GOTO(out_unlock, rc = 0);
1876
1877                 /* Compare on-disk and packed data_version */
1878                 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1879                         CDEBUG(D_HSM, DFID" data_version mismatches: "
1880                                "packed=%llu and on-disk=%llu\n",
1881                                PFID(mdt_object_fid(o)),
1882                                data->cd_data_version,
1883                                ma->ma_hsm.mh_arch_ver);
1884                         GOTO(out_unlock, rc = -EPERM);
1885                 }
1886         }
1887
1888         ma->ma_valid = MA_INODE;
1889         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1890         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1891         if (rc < 0)
1892                 GOTO(out_unlock, rc);
1893
1894         mutex_lock(&o->mot_som_mutex);
1895         rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size,
1896                            ma->ma_attr.la_blocks);
1897         mutex_unlock(&o->mot_som_mutex);
1898         if (rc2 < 0)
1899                 CDEBUG(D_INODE,
1900                        "%s: File "DFID" SOM update failed: rc = %d\n",
1901                        mdt_obd_name(info->mti_mdt),
1902                        PFID(mdt_object_fid(o)), rc2);
1903
1904
1905         ma->ma_need = MA_INODE | MA_LOV;
1906         rc = mdt_attr_get_complex(info, o, ma);
1907         if (rc < 0)
1908                 GOTO(out_unlock, rc);
1909
1910         if (!(ma->ma_valid & MA_LOV)) {
1911                 /* Even empty file are released */
1912                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1913                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
1914                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1915                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1916                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1917         } else {
1918                 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
1919                  * ma_lmm as lov_user_md, then it will be confused by union of
1920                  * layout_gen and stripe_offset. */
1921                 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
1922                     LOV_MAGIC_MAGIC)
1923                         ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1924                 else
1925                         GOTO(out_unlock, rc = -EINVAL);
1926         }
1927
1928         /* Set file as released. */
1929         rc = mdt_hsm_set_released(ma->ma_lmm);
1930         if (rc)
1931                 GOTO(out_unlock, rc);
1932
1933         orp_ma = &info->mti_u.hsm.attr;
1934         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1935         /* We use root ownership to bypass potential quota
1936          * restrictions on the user and group of the file. */
1937         orp_ma->ma_attr.la_uid = 0;
1938         orp_ma->ma_attr.la_gid = 0;
1939         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1940         orp_ma->ma_lmm = ma->ma_lmm;
1941         orp_ma->ma_lmm_size = ma->ma_lmm_size;
1942         orp_ma->ma_valid = MA_INODE | MA_LOV;
1943         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1944                                  MDS_FMODE_WRITE);
1945         if (IS_ERR(orphan)) {
1946                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1947                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1948                        PTR_ERR(orphan));
1949                 GOTO(out_unlock, rc = PTR_ERR(orphan));
1950         }
1951
1952         /* Set up HSM attribute for orphan object */
1953         BUILD_BUG_ON(sizeof(struct hsm_attrs) > sizeof(info->mti_xattr_buf));
1954         buf = &info->mti_buf;
1955         buf->lb_buf = info->mti_xattr_buf;
1956         buf->lb_len = sizeof(struct hsm_attrs);
1957         ma->ma_hsm.mh_flags |= HS_RELEASED;
1958         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1959         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
1960
1961         mdt_lock_reg_init(lh, LCK_EX);
1962         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
1963                              MDS_INODELOCK_XATTR);
1964         if (rc != 0)
1965                 GOTO(out_close, rc);
1966
1967         /* The orphan has root ownership so we need to raise
1968          * CAP_FOWNER to set the HSM attributes. */
1969         cap = uc->uc_cap;
1970         uc->uc_cap |= MD_CAP_TO_MASK(CAP_FOWNER);
1971         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
1972                           XATTR_NAME_HSM, 0);
1973         uc->uc_cap = cap;
1974         if (rc != 0)
1975                 GOTO(out_layout_lock, rc);
1976
1977         /* Swap layout with orphan objects. */
1978         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
1979                              mdt_object_child(orphan),
1980                              SWAP_LAYOUTS_MDS_HSM);
1981
1982         if (!rc && ma->ma_attr_flags & MDS_PCC_ATTACH) {
1983                 ma->ma_need = MA_LOV;
1984                 rc = mdt_attr_get_complex(info, o, ma);
1985         }
1986
1987         EXIT;
1988
1989 out_layout_lock:
1990         /* Release exclusive LL */
1991         mdt_object_unlock(info, o, lh, 1);
1992 out_close:
1993         /* Close orphan object anyway */
1994         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
1995                        MDS_FMODE_WRITE);
1996         if (rc2 < 0)
1997                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
1998                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
1999         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2000                         "object closed");
2001         mdt_object_put(info->mti_env, orphan);
2002
2003 out_unlock:
2004         up_write(&o->mot_open_sem);
2005
2006         /* already released */
2007         if (rc == 0) {
2008                 struct mdt_body *repbody;
2009
2010                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2011                 LASSERT(repbody != NULL);
2012                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2013                 if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
2014                         LASSERT(ma->ma_valid & MA_LOV);
2015                         rc = mdt_get_lmm_gen(ma->ma_lmm,
2016                                              &repbody->mbo_layout_gen);
2017                         if (!rc)
2018                                 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
2019                 }
2020         }
2021
2022 out_reprocess:
2023         ldlm_reprocess_all(lease->l_resource, lease);
2024         LDLM_LOCK_PUT(lease);
2025
2026         ma->ma_valid = 0;
2027         ma->ma_need = 0;
2028
2029         return rc;
2030 }
2031
2032 int mdt_close_handle_layouts(struct mdt_thread_info *info,
2033                              struct mdt_object *o, struct md_attr *ma)
2034 {
2035         struct mdt_lock_handle  *lh1 = &info->mti_lh[MDT_LH_NEW];
2036         struct mdt_lock_handle  *lh2 = &info->mti_lh[MDT_LH_OLD];
2037         struct close_data       *data;
2038         struct ldlm_lock        *lease;
2039         struct mdt_object       *o1 = o, *o2 = NULL;
2040         bool                     lease_broken;
2041         bool                     swap_objects = false;
2042         int                      rc;
2043         ENTRY;
2044
2045         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2046                 RETURN(-EROFS);
2047
2048         if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
2049                 RETURN(-EINVAL);
2050
2051         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2052         if (data == NULL)
2053                 RETURN(-EPROTO);
2054
2055         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2056                 RETURN(-EINVAL);
2057
2058         rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
2059         if (rc == 0) {
2060                 /**
2061                  * only MDS_CLOSE_LAYOUT_SPLIT use the same fid to indicate
2062                  * mirror deletion, so we'd zero cd_fid, and keeps o2 be NULL.
2063                  */
2064                 if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT))
2065                         RETURN(-EINVAL);
2066
2067                 /* zero cd_fid to keeps o2 be NULL */
2068                 fid_zero(&data->cd_fid);
2069         } else if (rc < 0) {
2070                 /* Exchange o1 and o2, to enforce locking order */
2071                 swap_objects = true;
2072         }
2073
2074         lease = ldlm_handle2lock(&data->cd_handle);
2075         if (lease == NULL)
2076                 RETURN(-ESTALE);
2077
2078         if (!fid_is_zero(&data->cd_fid)) {
2079                 o2 = mdt_object_find(info->mti_env, info->mti_mdt,
2080                                      &data->cd_fid);
2081                 if (IS_ERR(o2))
2082                         GOTO(out_lease, rc = PTR_ERR(o2));
2083
2084                 if (!mdt_object_exists(o2))
2085                         GOTO(out_obj, rc = -ENOENT);
2086
2087                 if (!S_ISREG(lu_object_attr(&o2->mot_obj)))
2088                         GOTO(out_obj, rc = -EINVAL);
2089
2090                 if (swap_objects)
2091                         swap(o1, o2);
2092         }
2093
2094         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
2095                            MAY_WRITE);
2096         if (rc < 0)
2097                 GOTO(out_obj, rc);
2098
2099         if (o2) {
2100                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2),
2101                                    NULL, MAY_WRITE);
2102                 if (rc < 0)
2103                         GOTO(out_obj, rc);
2104         }
2105
2106         /* try to hold open_sem so that nobody else can open the file */
2107         if (!down_write_trylock(&o->mot_open_sem)) {
2108                 ldlm_lock_cancel(lease);
2109                 GOTO(out_obj, rc = -EBUSY);
2110         }
2111
2112         /* Check if the lease open lease has already canceled */
2113         lock_res_and_lock(lease);
2114         lease_broken = ldlm_is_cancel(lease);
2115         unlock_res_and_lock(lease);
2116
2117         LDLM_DEBUG(lease, DFID " lease broken? %d",
2118                    PFID(mdt_object_fid(o)), lease_broken);
2119
2120         /* Cancel server side lease. Client side counterpart should
2121          * have been cancelled. It's okay to cancel it now as we've
2122          * held mot_open_sem. */
2123         ldlm_lock_cancel(lease);
2124
2125         if (lease_broken)
2126                 GOTO(out_unlock_sem, rc = -ESTALE);
2127
2128         mdt_lock_reg_init(lh1, LCK_EX);
2129         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
2130                              MDS_INODELOCK_XATTR);
2131         if (rc < 0)
2132                 GOTO(out_unlock_sem, rc);
2133
2134         if (o2) {
2135                 mdt_lock_reg_init(lh2, LCK_EX);
2136                 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2137                                      MDS_INODELOCK_XATTR);
2138                 if (rc < 0)
2139                         GOTO(out_unlock1, rc);
2140         }
2141
2142         /* Swap layout with orphan object */
2143         if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2144                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2145                                      mdt_object_child(o2), 0);
2146         } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2147                    ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2148                 struct lu_buf *buf = &info->mti_buf;
2149                 struct md_rejig_data mrd;
2150
2151                 if (o2) {
2152                         mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2153                 } else {
2154                         if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)) {
2155                                 /* paranoid check again */
2156                                 CERROR(DFID
2157                                   ":only mirror split support NULL o2 object\n",
2158                                         PFID(mdt_object_fid(o)));
2159                                 GOTO(out_unlock1, rc = -EINVAL);
2160                         }
2161
2162                         /* set NULL mrd_obj for deleting mirror objects */
2163                         mrd.mrd_obj = NULL;
2164                 }
2165
2166                 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)
2167                         mrd.mrd_mirror_id = data->cd_mirror_id;
2168
2169                 buf->lb_len = sizeof(mrd);
2170                 buf->lb_buf = &mrd;
2171                 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2172                                   XATTR_LUSTRE_LOV,
2173                                   ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2174                                   LU_XATTR_SPLIT : LU_XATTR_MERGE);
2175                 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2176                         int rc2;
2177
2178                         mutex_lock(&o->mot_som_mutex);
2179                         rc2 = mdt_set_som(info, o, SOM_FL_STRICT,
2180                                           ma->ma_attr.la_size,
2181                                           ma->ma_attr.la_blocks);
2182                         mutex_unlock(&o->mot_som_mutex);
2183                         if (rc2 < 0)
2184                                 CERROR(DFID": Setting i_blocks error: %d, "
2185                                        "i_blocks will be reported wrongly and "
2186                                        "can only be fixed in next resync\n",
2187                                        PFID(mdt_object_fid(o)), rc2);
2188                 }
2189         }
2190         if (rc < 0)
2191                 GOTO(out_unlock2, rc);
2192
2193         EXIT;
2194
2195 out_unlock2:
2196         /* Release exclusive LL */
2197         if (o2)
2198                 mdt_object_unlock(info, o2, lh2, 1);
2199
2200 out_unlock1:
2201         mdt_object_unlock(info, o1, lh1, 1);
2202
2203 out_unlock_sem:
2204         up_write(&o->mot_open_sem);
2205
2206         /* already swapped */
2207         if (rc == 0) {
2208                 struct mdt_body *repbody;
2209
2210                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2211                 LASSERT(repbody != NULL);
2212                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2213         }
2214
2215 out_obj:
2216         if (o1 != o)
2217                 /* the 2nd object has been used, and swapped to o1 */
2218                 mdt_object_put(info->mti_env, o1);
2219         else if (o2)
2220                 /* the 2nd object has been used, and not swapped */
2221                 mdt_object_put(info->mti_env, o2);
2222
2223         ldlm_reprocess_all(lease->l_resource, lease);
2224
2225 out_lease:
2226         LDLM_LOCK_PUT(lease);
2227
2228         if (ma != NULL) {
2229                 ma->ma_valid = 0;
2230                 ma->ma_need = 0;
2231         }
2232
2233         return rc;
2234 }
2235
2236 static int mdt_close_resync_done(struct mdt_thread_info *info,
2237                                  struct mdt_object *o, struct md_attr *ma)
2238 {
2239         struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LOCAL];
2240         struct close_data *data;
2241         struct ldlm_lock *lease;
2242         struct md_layout_change layout = { 0 };
2243         __u32 *resync_ids = NULL;
2244         size_t resync_count = 0;
2245         bool lease_broken;
2246         int rc;
2247
2248         ENTRY;
2249
2250         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2251                 RETURN(-EROFS);
2252
2253         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2254                 RETURN(-EINVAL);
2255
2256         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2257         if (data == NULL)
2258                 RETURN(-EPROTO);
2259
2260         if (ptlrpc_req_need_swab(mdt_info_req(info)))
2261                 lustre_swab_close_data_resync_done(&data->cd_resync);
2262
2263         if (!fid_is_zero(&data->cd_fid))
2264                 RETURN(-EPROTO);
2265
2266         lease = ldlm_handle2lock(&data->cd_handle);
2267         if (lease == NULL)
2268                 RETURN(-ESTALE);
2269
2270         /* try to hold open_sem so that nobody else can open the file */
2271         if (!down_write_trylock(&o->mot_open_sem)) {
2272                 ldlm_lock_cancel(lease);
2273                 GOTO(out_reprocess, rc = -EBUSY);
2274         }
2275
2276         /* Check if the lease open lease has already canceled */
2277         lock_res_and_lock(lease);
2278         lease_broken = ldlm_is_cancel(lease);
2279         unlock_res_and_lock(lease);
2280
2281         LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2282                    PFID(mdt_object_fid(o)), lease_broken);
2283
2284         /* Cancel server side lease. Client side counterpart should
2285          * have been cancelled. It's okay to cancel it now as we've
2286          * held mot_open_sem. */
2287         ldlm_lock_cancel(lease);
2288
2289         if (lease_broken) /* don't perform release task */
2290                 GOTO(out_unlock, rc = -ESTALE);
2291
2292         resync_count = data->cd_resync.resync_count;
2293
2294         if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2295                 void *data;
2296
2297                 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2298                                            RCL_CLIENT))
2299                         GOTO(out_unlock, rc = -EPROTO);
2300
2301                 OBD_ALLOC_PTR_ARRAY(resync_ids, resync_count);
2302                 if (!resync_ids)
2303                         GOTO(out_unlock, rc = -ENOMEM);
2304
2305                 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2306                 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2307
2308                 layout.mlc_resync_ids = resync_ids;
2309         } else {
2310                 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2311         }
2312
2313         layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2314         layout.mlc_resync_count = resync_count;
2315         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2316                 layout.mlc_som.lsa_valid = SOM_FL_STRICT;
2317                 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2318                 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2319         }
2320         rc = mdt_layout_change(info, o, lhc, &layout);
2321         if (rc)
2322                 GOTO(out_unlock, rc);
2323
2324         mdt_object_unlock(info, o, lhc, 0);
2325
2326         EXIT;
2327
2328 out_unlock:
2329         up_write(&o->mot_open_sem);
2330
2331         /* already released */
2332         if (rc == 0) {
2333                 struct mdt_body *repbody;
2334
2335                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2336                 LASSERT(repbody != NULL);
2337                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2338         }
2339
2340         if (resync_ids)
2341                 OBD_FREE_PTR_ARRAY(resync_ids, resync_count);
2342
2343 out_reprocess:
2344         ldlm_reprocess_all(lease->l_resource, lease);
2345         LDLM_LOCK_PUT(lease);
2346
2347         ma->ma_valid = 0;
2348         ma->ma_need = 0;
2349
2350         return rc;
2351 }
2352
2353 #define MFD_CLOSED(open_flags) ((open_flags) == MDS_FMODE_CLOSED)
2354
2355 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2356 {
2357         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_open_flags));
2358 }
2359
2360 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2361 {
2362         struct mdt_object *o = mfd->mfd_object;
2363         struct md_object *next = mdt_object_child(o);
2364         struct md_attr *ma = &info->mti_attr;
2365         struct lu_fid *ofid = &info->mti_tmp_fid1;
2366         int rc = 0;
2367         u64 open_flags;
2368         u64 intent;
2369
2370         ENTRY;
2371
2372         open_flags = mfd->mfd_open_flags;
2373         intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2374         *ofid = *mdt_object_fid(o);
2375
2376         /* the below message is checked in replay-single.sh test_46 */
2377         CDEBUG(D_INODE, "%s: %sclosing file handle "DFID" with intent: %llx\n",
2378                mdt_obd_name(info->mti_mdt),
2379                ma->ma_valid & MA_FORCE_LOG ? "force " : "", PFID(ofid), intent);
2380
2381         switch (intent) {
2382         case MDS_HSM_RELEASE: {
2383                 rc = mdt_hsm_release(info, o, ma);
2384                 if (rc < 0) {
2385                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2386                                mdt_obd_name(info->mti_mdt),
2387                                PFID(ofid), rc);
2388                         /* continue to close even error occurred. */
2389                 }
2390                 break;
2391         }
2392         case MDS_CLOSE_LAYOUT_MERGE:
2393         case MDS_CLOSE_LAYOUT_SPLIT:
2394         case MDS_CLOSE_LAYOUT_SWAP: {
2395                 rc = mdt_close_handle_layouts(info, o, ma);
2396                 if (rc < 0) {
2397                         CDEBUG(D_INODE,
2398                                "%s: cannot swap layout of "DFID": rc = %d\n",
2399                                mdt_obd_name(info->mti_mdt),
2400                                PFID(ofid), rc);
2401                         /* continue to close even if error occurred. */
2402                 }
2403                 break;
2404         }
2405         case MDS_CLOSE_RESYNC_DONE:
2406                 rc = mdt_close_resync_done(info, o, ma);
2407                 break;
2408         default:
2409                 /* nothing */
2410                 break;
2411         }
2412
2413         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
2414             ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) {
2415                 int rc2;
2416
2417                 rc2 = mdt_lsom_update(info, o, false);
2418                 if (rc2 < 0)
2419                         CDEBUG(D_INODE,
2420                                "%s: File " DFID " LSOM failed: rc = %d\n",
2421                                mdt_obd_name(info->mti_mdt),
2422                                PFID(ofid), rc2);
2423                         /* continue to close even if error occured. */
2424         }
2425
2426         if (open_flags & MDS_FMODE_WRITE)
2427                 mdt_write_put(o);
2428         else if (open_flags & MDS_FMODE_EXEC)
2429                 mdt_write_allow(o);
2430
2431         /* Update atime|mtime|ctime on close. */
2432         if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ ||
2433              open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
2434             (ma->ma_attr.la_valid & LA_ATIME ||
2435              ma->ma_attr.la_valid & LA_MTIME ||
2436              ma->ma_attr.la_valid & LA_CTIME)) {
2437                 ma->ma_valid = MA_INODE;
2438                 ma->ma_attr_flags |= MDS_CLOSE_UPDATE_TIMES;
2439                 ma->ma_attr.la_valid &= (LA_ATIME | LA_MTIME | LA_CTIME);
2440
2441                 if (ma->ma_attr.la_valid & LA_MTIME) {
2442                         rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
2443                         if (!rc)
2444                                 ma->ma_valid |= MA_PFID;
2445                 }
2446
2447                 rc = mo_attr_set(info->mti_env, next, ma);
2448         }
2449
2450         /* If file data is modified, add the dirty flag. */
2451         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2452                 rc = mdt_add_dirty_flag(info, o, ma);
2453
2454         ma->ma_need |= MA_INODE;
2455         ma->ma_valid &= ~MA_INODE;
2456
2457         LASSERT(atomic_read(&o->mot_open_count) > 0);
2458         atomic_dec(&o->mot_open_count);
2459         mdt_handle_last_unlink(info, o, ma);
2460
2461         if (!MFD_CLOSED(open_flags)) {
2462                 rc = mo_close(info->mti_env, next, ma, open_flags);
2463                 if (mdt_dom_check_for_discard(info, o))
2464                         mdt_dom_discard_data(info, o);
2465         }
2466
2467         /* adjust open and lease count */
2468         if (open_flags & MDS_OPEN_LEASE) {
2469                 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2470                 atomic_dec(&o->mot_lease_count);
2471         }
2472
2473         mdt_mfd_free(mfd);
2474         mdt_object_put(info->mti_env, o);
2475
2476         RETURN(rc);
2477 }
2478
2479 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2480                        struct mdt_body *repbody)
2481 {
2482         struct mdt_export_data *med;
2483         struct mdt_file_data   *mfd;
2484         int                     ret = 0;
2485         int                     rc = 0;
2486         ENTRY;
2487
2488         med = &req->rq_export->exp_mdt_data;
2489         spin_lock(&med->med_open_lock);
2490         mfd = mdt_open_handle2mfd(med, &info->mti_open_handle,
2491                                   req_is_replay(req));
2492         if (mdt_mfd_closed(mfd)) {
2493                 spin_unlock(&med->med_open_lock);
2494                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2495                        ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2496                        info->mti_open_handle.cookie);
2497                 /** not serious error since bug 3633 */
2498                 rc = -ESTALE;
2499         } else {
2500                 class_handle_unhash(&mfd->mfd_open_handle);
2501                 list_del_init(&mfd->mfd_list);
2502                 spin_unlock(&med->med_open_lock);
2503                 ret = mdt_mfd_close(info, mfd);
2504         }
2505
2506         RETURN(rc ? rc : ret);
2507 }
2508
2509 int mdt_close(struct tgt_session_info *tsi)
2510 {
2511         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2512         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2513         struct md_attr         *ma = &info->mti_attr;
2514         struct mdt_body        *repbody = NULL;
2515         ktime_t                 kstart = ktime_get();
2516         int rc, ret = 0;
2517         ENTRY;
2518
2519         /* Close may come with the Size-on-MDS update. Unpack it. */
2520         rc = mdt_close_unpack(info);
2521         if (rc)
2522                 GOTO(out, rc = err_serious(rc));
2523
2524         /* These fields are no longer used and are left for compatibility.
2525          * size is always zero */
2526         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2527                              0);
2528         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2529                              0);
2530         rc = req_capsule_server_pack(info->mti_pill);
2531         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2532                 mdt_client_compatibility(info);
2533                 if (rc == 0)
2534                         mdt_fix_reply(info);
2535                 mdt_exit_ucred(info);
2536                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2537         }
2538
2539         /* Continue to close handle even if we can not pack reply */
2540         if (rc == 0) {
2541                 repbody = req_capsule_server_get(info->mti_pill,
2542                                                  &RMF_MDT_BODY);
2543                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2544                                                     &RMF_MDT_MD);
2545                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2546                                                        &RMF_MDT_MD,
2547                                                        RCL_SERVER);
2548                 ma->ma_need = MA_INODE | MA_LOV;
2549                 repbody->mbo_eadatasize = 0;
2550                 repbody->mbo_aclsize = 0;
2551         } else {
2552                 rc = err_serious(rc);
2553         }
2554
2555         rc = mdt_close_internal(info, req, repbody);
2556         if (rc != -ESTALE)
2557                 mdt_empty_transno(info, rc);
2558
2559         if (repbody != NULL) {
2560                 mdt_client_compatibility(info);
2561                 rc = mdt_fix_reply(info);
2562         }
2563
2564         mdt_exit_ucred(info);
2565         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2566                 GOTO(out, rc = err_serious(-ENOMEM));
2567
2568         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2569                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2570                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2571 out:
2572         mdt_thread_info_fini(info);
2573         if (rc == 0)
2574                 mdt_counter_incr(req, LPROC_MDT_CLOSE,
2575                                  ktime_us_delta(ktime_get(), kstart));
2576         RETURN(rc ? rc : ret);
2577 }