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_LIST_HEAD_RCU(&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         LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
 311         if (S_ISDIR(obj->mot_header.loh_attr)) {
 312                 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
 313                                                     &RMF_MDT_MD);
 314                 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
 315                                                        &RMF_MDT_MD,
 316                                                        RCL_SERVER);
 317                 if (ma->ma_lmv_size > 0)
 318                         ma->ma_need |= MA_LMV;
 319         } else {
 320                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
 321                                                     &RMF_MDT_MD);
 322                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 323                                                        &RMF_MDT_MD,
 324                                                        RCL_SERVER);
 325                 if (ma->ma_lmm_size > 0)
 326                         ma->ma_need |= MA_LOV;
 327         }
 328 }
 329
 330 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 331                         struct mdt_object *o, u64 open_flags, int created,
 332                         struct ldlm_reply *rep)
 333 {
 334         struct ptlrpc_request *req = mdt_info_req(info);
 335         struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 336         struct mdt_file_data *mfd;
 337         struct md_attr *ma  = &info->mti_attr;
 338         struct lu_attr *la  = &ma->ma_attr;
 339         struct mdt_body *repbody;
 340         bool isdir, isreg;
 341         int rc = 0;
 342
 343         ENTRY;
 344         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 345
 346         isreg = S_ISREG(la->la_mode);
 347         isdir = S_ISDIR(la->la_mode);
 348         if (isreg && !(ma->ma_valid & MA_LOV) &&
 349             !(open_flags & MDS_OPEN_RELEASE)) {
 350                 /*
 351                  * No EA, check whether it is will set regEA and dirEA since in
 352                  * above attr get, these size might be zero, so reset it, to
 353                  * retrieve the MD after create obj.
 354                  */
 355                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 356                                                        &RMF_MDT_MD,
 357                                                        RCL_SERVER);
 358                 /* in replay case, p == NULL */
 359                 rc = mdt_create_data(info, p, o);
 360                 if (rc)
 361                         RETURN(rc);
 362
 363                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
 364                         mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
 365         }
 366
 367         CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n",
 368                ma->ma_valid, ma->ma_lmm_size);
 369
 370         if (ma->ma_valid & MA_LOV) {
 371                 LASSERT(ma->ma_lmm_size != 0);
 372                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 373                 if (isdir)
 374                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 375                 else
 376                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 377         }
 378
 379         if (ma->ma_valid & MA_LMV) {
 380                 LASSERT(ma->ma_lmv_size != 0);
 381                 repbody->mbo_eadatasize = ma->ma_lmv_size;
 382                 LASSERT(isdir);
 383                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
 384         }
 385
 386         if (open_flags & MDS_FMODE_WRITE)
 387                 rc = mdt_write_get(o);
 388         else if (open_flags & MDS_FMODE_EXEC)
 389                 rc = mdt_write_deny(o);
 390
 391         if (rc)
 392                 RETURN(rc);
 393
 394         rc = mo_open(info->mti_env, mdt_object_child(o),
 395                      created ? open_flags | MDS_OPEN_CREATED : open_flags,
 396                      &info->mti_spec);
 397         if (rc != 0) {
 398                 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
 399                  * client does not know which suppgid should be sent to the MDS,
 400                  * or some other(s) changed the target file's GID after this RPC
 401                  * sent to the MDS with the suppgid as the original GID, then we
 402                  * give the client another chance to send the right suppgid. */
 403                 if (rc == -EACCES &&
 404                     allow_client_chgrp(info, lu_ucred(info->mti_env)))
 405                         mdt_set_disposition(info, rep, DISP_OPEN_DENY);
 406
 407                 GOTO(err_out, rc);
 408         }
 409
 410         mfd = mdt_mfd_new(med);
 411         if (mfd == NULL)
 412                 GOTO(err_out, rc = -ENOMEM);
 413
 414         /*
 415          * Keep a reference on this object for this open, and is
 416          * released by mdt_mfd_close().
 417          */
 418         mdt_object_get(info->mti_env, o);
 419         mfd->mfd_object = o;
 420         mfd->mfd_xid = req->rq_xid;
 421
 422         /*
 423          * @open_flags is always not zero. At least it should be FMODE_READ,
 424          * FMODE_WRITE or MDS_FMODE_EXEC.
 425          */
 426         LASSERT(open_flags != 0);
 427
 428         /* Open handling. */
 429         mdt_mfd_set_mode(mfd, open_flags);
 430
 431         atomic_inc(&o->mot_open_count);
 432         if (open_flags & MDS_OPEN_LEASE)
 433                 atomic_inc(&o->mot_lease_count);
 434
 435         /* replay handle */
 436         if (req_is_replay(req)) {
 437                 struct mdt_file_data *old_mfd;
 438                 /* Check wheather old cookie already exist in
 439                  * the list, becasue when do recovery, client
 440                  * might be disconnected from server, and
 441                  * restart replay, so there maybe some orphan
 442                  * mfd here, we should remove them */
 443                 LASSERT(info->mti_rr.rr_open_handle != NULL);
 444                 spin_lock(&med->med_open_lock);
 445                 old_mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 446                                               true);
 447                 if (old_mfd != NULL) {
 448                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 449                                "cookie = %#llx\n", mfd,
 450                                PFID(mdt_object_fid(mfd->mfd_object)),
 451                                info->mti_rr.rr_open_handle->cookie);
 452                         class_handle_unhash(&old_mfd->mfd_open_handle);
 453                         list_del_init(&old_mfd->mfd_list);
 454                         spin_unlock(&med->med_open_lock);
 455                         /* no attr update for that close */
 456                         la->la_valid = 0;
 457                         ma->ma_valid |= MA_FLAGS;
 458                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 459                         mdt_mfd_close(info, old_mfd);
 460                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 461                         ma->ma_valid &= ~MA_FLAGS;
 462                 } else {
 463                         spin_unlock(&med->med_open_lock);
 464                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 465                                "cookie = %#llx\n",
 466                                PFID(mdt_object_fid(mfd->mfd_object)),
 467                                info->mti_rr.rr_open_handle->cookie);
 468                 }
 469
 470                 CDEBUG(D_HA, "Store old cookie %#llx in new mfd\n",
 471                        info->mti_rr.rr_open_handle->cookie);
 472
 473                 mfd->mfd_open_handle_old = *info->mti_rr.rr_open_handle;
 474         }
 475
 476         repbody->mbo_open_handle.cookie = mfd->mfd_open_handle.h_cookie;
 477
 478         if (req->rq_export->exp_disconnected) {
 479                 spin_lock(&med->med_open_lock);
 480                 class_handle_unhash(&mfd->mfd_open_handle);
 481                 list_del_init(&mfd->mfd_list);
 482                 spin_unlock(&med->med_open_lock);
 483                 mdt_mfd_close(info, mfd);
 484         } else {
 485                 spin_lock(&med->med_open_lock);
 486                 list_add_tail(&mfd->mfd_list, &med->med_open_head);
 487                 spin_unlock(&med->med_open_lock);
 488         }
 489
 490         mdt_empty_transno(info, rc);
 491
 492         RETURN(rc);
 493
 494 err_out:
 495         if (open_flags & MDS_FMODE_WRITE)
 496                 mdt_write_put(o);
 497         else if (open_flags & MDS_FMODE_EXEC)
 498                 mdt_write_allow(o);
 499
 500         return rc;
 501 }
 502
 503 static int mdt_finish_open(struct mdt_thread_info *info,
 504                            struct mdt_object *p, struct mdt_object *o,
 505                            u64 open_flags, int created,
 506                            struct ldlm_reply *rep)
 507 {
 508         struct ptlrpc_request   *req = mdt_info_req(info);
 509         struct obd_export       *exp = req->rq_export;
 510         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 511         struct md_attr          *ma  = &info->mti_attr;
 512         struct lu_attr          *la  = &ma->ma_attr;
 513         struct mdt_file_data    *mfd;
 514         struct mdt_body         *repbody;
 515         int                      rc = 0;
 516         int                      isreg, isdir, islnk;
 517         struct list_head        *t;
 518         ENTRY;
 519
 520         LASSERT(ma->ma_valid & MA_INODE);
 521
 522         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 523
 524         isreg = S_ISREG(la->la_mode);
 525         isdir = S_ISDIR(la->la_mode);
 526         islnk = S_ISLNK(la->la_mode);
 527         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 528
 529         /* compatibility check for 2.10 clients when it tries to open mirrored
 530          * files. 2.10 clients don't verify overlapping components so they
 531          * would read and write mirrored files just as if they were normal
 532          * PFL files, which will cause the problem that sycned mirrors actually
 533          * contain different data.
 534          * Older clients are not a concern here because they don't even
 535          * understand PFL layout. */
 536         if (isreg && !exp_connect_flr(exp) && ma->ma_valid & MA_LOV &&
 537             mdt_lmm_is_flr(ma->ma_lmm)) {
 538                 /* LU-10286: for simplicity clients who don't understand
 539                  * mirrored layout(with connect flag OBD_CONNECT2_FLR) won't
 540                  * be able to open mirrored files */
 541                 RETURN(-EOPNOTSUPP);
 542         }
 543
 544         /* Overstriped files can crash older clients */
 545         if (isreg && !exp_connect_overstriping(exp) &&
 546             mdt_lmm_is_overstriping(ma->ma_lmm))
 547                 RETURN(-EOPNOTSUPP);
 548
 549         /* LU-2275, simulate broken behaviour (esp. prevalent in
 550          * pre-2.4 servers where a very strange reply is sent on error
 551          * that looks like it was actually almost successful and a
 552          * failure at the same time.) */
 553         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 554                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 555                                                DISP_LOOKUP_NEG |
 556                                                DISP_LOOKUP_POS);
 557
 558                 if (open_flags & MDS_OPEN_LOCK)
 559                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 560
 561                 RETURN(-ENOENT);
 562         }
 563
 564 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
 565         if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 566                 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
 567                 if (IS_ERR(nodemap))
 568                         RETURN(PTR_ERR(nodemap));
 569
 570                 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
 571                 nodemap_putref(nodemap);
 572                 if (rc)
 573                         RETURN(rc);
 574         }
 575 #endif
 576
 577         /*
 578          * If we are following a symlink, don't open; and do not return open
 579          * handle for special nodes as client required.
 580          */
 581         if (islnk || (!isreg && !isdir &&
 582             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 583                 lustre_msg_set_transno(req->rq_repmsg, 0);
 584                 RETURN(0);
 585         }
 586
 587         /*
 588          * We need to return the existing object's fid back, so it is done here,
 589          * after preparing the reply.
 590          */
 591         if (!created && (open_flags & MDS_OPEN_EXCL) &&
 592             (open_flags & MDS_OPEN_CREAT))
 593                 RETURN(-EEXIST);
 594
 595         /* This can't be done earlier, we need to return reply body */
 596         if (isdir) {
 597                 if (open_flags & (MDS_OPEN_CREAT | MDS_FMODE_WRITE)) {
 598                         /* We are trying to create or write an existing dir. */
 599                         RETURN(-EISDIR);
 600                 }
 601         } else if (open_flags & MDS_OPEN_DIRECTORY)
 602                 RETURN(-ENOTDIR);
 603
 604         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 605                                  OBD_FAIL_MDS_LDLM_REPLY_NET | OBD_FAIL_ONCE))
 606                 RETURN(-EAGAIN);
 607
 608         mfd = NULL;
 609         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 610                 spin_lock(&med->med_open_lock);
 611                 list_for_each(t, &med->med_open_head) {
 612                         mfd = list_entry(t, struct mdt_file_data, mfd_list);
 613                         if (mfd->mfd_xid == req->rq_xid)
 614                                 break;
 615                         mfd = NULL;
 616                 }
 617                 spin_unlock(&med->med_open_lock);
 618
 619                 if (mfd != NULL) {
 620                         repbody->mbo_open_handle.cookie =
 621                                 mfd->mfd_open_handle.h_cookie;
 622                         /* set repbody->ea_size for resent case */
 623                         if (ma->ma_valid & MA_LOV) {
 624                                 LASSERT(ma->ma_lmm_size != 0);
 625                                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 626                                 if (isdir)
 627                                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 628                                 else
 629                                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 630                         }
 631                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 632                         RETURN(0);
 633                 }
 634         }
 635
 636         rc = mdt_mfd_open(info, p, o, open_flags, created, rep);
 637         if (!rc)
 638                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 639
 640         RETURN(rc);
 641 }
 642
 643 void mdt_reconstruct_open(struct mdt_thread_info *info,
 644                           struct mdt_lock_handle *lhc)
 645 {
 646         const struct lu_env *env = info->mti_env;
 647         struct mdt_device *mdt = info->mti_mdt;
 648         struct req_capsule *pill = info->mti_pill;
 649         struct ptlrpc_request *req = mdt_info_req(info);
 650         struct md_attr *ma = &info->mti_attr;
 651         struct mdt_reint_record *rr = &info->mti_rr;
 652         u64 open_flags = info->mti_spec.sp_cr_flags;
 653         struct ldlm_reply *ldlm_rep;
 654         struct mdt_object *parent;
 655         struct mdt_object *child;
 656         struct mdt_body *repbody;
 657         u64 opdata;
 658         int rc;
 659         ENTRY;
 660
 661         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 662         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 663         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 664
 665         ma->ma_need = MA_INODE | MA_HSM;
 666         ma->ma_valid = 0;
 667
 668         opdata = mdt_req_from_lrd(req, info->mti_reply_data);
 669         mdt_set_disposition(info, ldlm_rep, opdata);
 670
 671         CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n",
 672                ldlm_rep->lock_policy_res1, req->rq_status);
 673
 674         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
 675             req->rq_status != 0)
 676                 /* We did not create successfully, return error to client. */
 677                 GOTO(out, rc = req->rq_status);
 678
 679         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
 680                 struct obd_export *exp = req->rq_export;
 681                 /*
 682                  * We failed after creation, but we do not know in which step
 683                  * we failed. So try to check the child object.
 684                  */
 685                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
 686                 if (IS_ERR(parent)) {
 687                         rc = PTR_ERR(parent);
 688                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
 689                                       " Evicting client %s with export %s.\n",
 690                                       PFID(rr->rr_fid1), rc,
 691                                       obd_uuid2str(&exp->exp_client_uuid),
 692                                       obd_export_nid2str(exp));
 693                         mdt_export_evict(exp);
 694                         RETURN_EXIT;
 695                 }
 696
 697                 child = mdt_object_find(env, mdt, rr->rr_fid2);
 698                 if (IS_ERR(child)) {
 699                         rc = PTR_ERR(child);
 700                         LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
 701                                       "evicting client %s with export %s\n",
 702                                       PFID(rr->rr_fid2), rc,
 703                                       obd_uuid2str(&exp->exp_client_uuid),
 704                                       obd_export_nid2str(exp));
 705                         mdt_object_put(env, parent);
 706                         mdt_export_evict(exp);
 707                         RETURN_EXIT;
 708                 }
 709
 710                 if (unlikely(mdt_object_remote(child))) {
 711                         mdt_object_put(env, parent);
 712                         mdt_object_put(env, child);
 713                         /* the child object was created on remote server */
 714                         if (!mdt_is_dne_client(exp))
 715                                 /* Return -EIO for old client */
 716                                 GOTO(out, rc = -EIO);
 717                         repbody->mbo_fid1 = *rr->rr_fid2;
 718                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 719                         GOTO(out, rc = 0);
 720                 }
 721                 if (mdt_object_exists(child)) {
 722                         mdt_prep_ma_buf_from_rep(info, child, ma);
 723                         rc = mdt_attr_get_complex(info, child, ma);
 724                         if (!rc)
 725                                 rc = mdt_finish_open(info, parent, child,
 726                                                      open_flags, 1, ldlm_rep);
 727                         mdt_object_put(env, parent);
 728                         mdt_object_put(env, child);
 729                         if (!rc)
 730                                 mdt_pack_size2body(info, rr->rr_fid2,
 731                                                    &lhc->mlh_reg_lh);
 732                         GOTO(out, rc);
 733                 }
 734                 /* the child does not exist, we should do regular open */
 735                 mdt_object_put(env, parent);
 736                 mdt_object_put(env, child);
 737         }
 738         /* We did not try to create, so we are a pure open */
 739         rc = mdt_reint_open(info, lhc);
 740         EXIT;
 741 out:
 742         req->rq_status = rc;
 743         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 744         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
 745 }
 746
 747 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
 748 {
 749         u64 open_flags = info->mti_spec.sp_cr_flags;
 750         struct mdt_reint_record *rr = &info->mti_rr;
 751         struct md_attr *ma = &info->mti_attr;
 752         struct mdt_object *o;
 753         int rc;
 754
 755         ENTRY;
 756         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
 757         if (IS_ERR(o))
 758                 RETURN(rc = PTR_ERR(o));
 759
 760         if (unlikely(mdt_object_remote(o))) {
 761                 /* the child object was created on remote server */
 762                 struct mdt_body *repbody;
 763
 764                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 765                                                 DISP_LOOKUP_EXECD |
 766                                                 DISP_LOOKUP_POS));
 767                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 768                 repbody->mbo_fid1 = *rr->rr_fid2;
 769                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 770                 rc = 0;
 771         } else {
 772                 if (mdt_object_exists(o)) {
 773                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 774                                                         DISP_LOOKUP_EXECD |
 775                                                         DISP_LOOKUP_POS));
 776                         mdt_prep_ma_buf_from_rep(info, o, ma);
 777                         rc = mdt_attr_get_complex(info, o, ma);
 778                         if (rc == 0)
 779                                 rc = mdt_finish_open(info, NULL, o, open_flags,
 780                                                      0, rep);
 781                 } else {
 782                         rc = -ENOENT;
 783                 }
 784         }
 785
 786         mdt_object_put(info->mti_env, o);
 787         RETURN(rc);
 788 }
 789
 790 /* lock object for open */
 791 static int mdt_object_open_lock(struct mdt_thread_info *info,
 792                                 struct mdt_object *obj,
 793                                 struct mdt_lock_handle *lhc,
 794                                 __u64 *ibits)
 795 {
 796         struct md_attr *ma = &info->mti_attr;
 797         __u64 open_flags = info->mti_spec.sp_cr_flags;
 798         __u64 trybits = 0;
 799         enum ldlm_mode lm = LCK_CR;
 800         bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
 801         bool try_layout = false;
 802         bool create_layout = false;
 803         int rc = 0;
 804         int dom_stripes = LMM_NO_DOM;
 805         bool dom_lock = false;
 806
 807         ENTRY;
 808
 809         *ibits = 0;
 810         mdt_lock_handle_init(lhc);
 811
 812         if (req_is_replay(mdt_info_req(info)))
 813                 RETURN(0);
 814
 815         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
 816                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
 817                     md_should_create(open_flags))
 818                         create_layout = true;
 819                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
 820                     ma->ma_need & MA_LOV)
 821                         try_layout = true;
 822
 823                 /* DoM files can have just MDT stripe or combined MDT + OST
 824                  * stripes.
 825                  * - In the first case the open for read/write will do IO to
 826                  *   the MDT stripe and it makes sense to take IO lock in
 827                  *   advance along with OPEN even if it is blocking lock.
 828                  * - In the second case it is just size of MDT stripe and it
 829                  *   is quite unlikely that client will write into it, though
 830                  *   it may read it. So IO lock will be taken optionally if it
 831                  *   is non-blocking one.
 832                  */
 833                 if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
 834                         dom_stripes = mdt_lmm_dom_entry(ma->ma_lmm);
 835
 836                 if (dom_stripes == LMM_DOM_ONLY &&
 837                     info->mti_mdt->mdt_opts.mo_dom_lock > 0 &&
 838                     !mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
 839                         dom_lock = true;
 840         }
 841
 842         if (acq_lease) {
 843                 /* lease open, acquire write mode of open sem */
 844                 down_write(&obj->mot_open_sem);
 845
 846                 /* Lease exists and ask for new lease */
 847                 if (atomic_read(&obj->mot_lease_count) > 0) {
 848                         /* only exclusive open is supported, so lease
 849                          * are conflicted to each other */
 850                         GOTO(out, rc = -EBUSY);
 851                 }
 852
 853                 /* Lease must be with open lock */
 854                 if (!(open_flags & MDS_OPEN_LOCK)) {
 855                         CERROR("%s: Request lease for file:"DFID ", but open lock "
 856                                "is missed, open_flags = %#llo : rc = %d\n",
 857                                mdt_obd_name(info->mti_mdt),
 858                                PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
 859                         GOTO(out, rc = -EPROTO);
 860                 }
 861
 862                 /* XXX: only exclusive open is supported. */
 863                 lm = LCK_EX;
 864                 *ibits = MDS_INODELOCK_OPEN;
 865
 866                 /* never grant LCK_EX layout lock to client */
 867                 try_layout = false;
 868         } else { /* normal open */
 869                 /* normal open holds read mode of open sem */
 870                 down_read(&obj->mot_open_sem);
 871
 872                 if (open_flags & MDS_OPEN_LOCK) {
 873                         if (open_flags & MDS_FMODE_WRITE)
 874                                 lm = LCK_CW;
 875                         else if (open_flags & MDS_FMODE_EXEC)
 876                                 lm = LCK_PR;
 877                         else
 878                                 lm = LCK_CR;
 879
 880                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
 881                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
 882                         if (open_flags & MDS_FMODE_WRITE)
 883                                 lm = LCK_CW;
 884                         else
 885                                 lm = LCK_CR;
 886
 887                         /* revoke lease */
 888                         *ibits = MDS_INODELOCK_OPEN;
 889                         try_layout = false;
 890
 891                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 892                 } else if (dom_lock) {
 893                         lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR;
 894                         if (info->mti_mdt->mdt_opts.mo_dom_lock ==
 895                             TRYLOCK_DOM_ON_OPEN) {
 896                                 trybits |= MDS_INODELOCK_DOM |
 897                                            MDS_INODELOCK_LAYOUT;
 898                         } else {
 899                                 /* mo_dom_lock == ALWAYS_DOM_LOCK_ON_OPEN */
 900                                 *ibits = MDS_INODELOCK_DOM;
 901                                 if (info->mti_mdt->mdt_opts.mo_dom_read_open) {
 902                                         trybits |= MDS_INODELOCK_LAYOUT;
 903                                 }
 904                         }
 905                 }
 906
 907                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
 908                         PFID(mdt_object_fid(obj)),
 909                         atomic_read(&obj->mot_lease_count), lm);
 910         }
 911
 912         mdt_lock_reg_init(lhc, lm);
 913
 914         /* Return lookup lock to validate inode at the client side.
 915          * This is pretty important otherwise MDT will return layout
 916          * lock for each open.
 917          * However this is a double-edged sword because changing
 918          * permission will revoke a huge number of LOOKUP locks.
 919          */
 920         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
 921                 if (!(*ibits & MDS_INODELOCK_LOOKUP))
 922                         trybits |= MDS_INODELOCK_LOOKUP;
 923                 trybits |= MDS_INODELOCK_LAYOUT;
 924         }
 925
 926         if (*ibits | trybits)
 927                 rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, false);
 928
 929         CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
 930                ", open_flags = %#llo, try_layout = %d : rc = %d\n",
 931                mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
 932                *ibits, trybits, open_flags, try_layout, rc);
 933
 934         /* will change layout, revoke layout locks by enqueuing EX lock. */
 935         if (rc == 0 && create_layout) {
 936                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
 937
 938                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
 939                         ", open_flags = %#llo\n",
 940                         PFID(mdt_object_fid(obj)), open_flags);
 941
 942                 /* We cannot enqueue another lock for the same resource we
 943                  * already have a lock for, due to mechanics of waiting list
 944                  * iterating in ldlm, see LU-3601.
 945                  * As such we'll drop the open lock we just got above here,
 946                  * it's ok not to have this open lock as it's main purpose is to
 947                  * flush unused cached client open handles. */
 948                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 949                         mdt_object_unlock(info, obj, lhc, 1);
 950
 951                 LASSERT(!try_layout);
 952                 mdt_lock_handle_init(ll);
 953                 mdt_lock_reg_init(ll, LCK_EX);
 954                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
 955
 956                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
 957         }
 958
 959         /* Check if there is any other open handles after acquiring
 960          * open lock. At this point, caching open handles have been revoked
 961          * by open lock.
 962          * XXX: Now only exclusive open is supported. Need to check the
 963          * type of open for generic lease support. */
 964         if (rc == 0 && acq_lease) {
 965                 struct ptlrpc_request *req = mdt_info_req(info);
 966                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 967                 struct mdt_file_data *mfd;
 968                 bool is_replay_or_resent;
 969                 int open_count = 0;
 970
 971                 /* For lease: application can open a file and then apply lease,
 972                  * @handle contains original open handle in that case.
 973                  * In recovery, open REQ will be replayed and the lease REQ may
 974                  * be resent that means the open handle is already stale, so we
 975                  * need to fix it up here by finding new handle. */
 976                 is_replay_or_resent = req_is_replay(req) ||
 977                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
 978
 979                 /* if the request is _not_ a replay request, rr_open_handle
 980                  * may be used to hold an open file handle which is issuing the
 981                  * lease request, so that this openhandle doesn't count. */
 982                 mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 983                                           is_replay_or_resent);
 984                 if (mfd != NULL)
 985                         ++open_count;
 986
 987                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
 988                         PFID(mdt_object_fid(obj)),
 989                         atomic_read(&obj->mot_open_count), open_count);
 990
 991                 if (atomic_read(&obj->mot_open_count) > open_count)
 992                         GOTO(out, rc = -EBUSY);
 993         }
 994         GOTO(out, rc);
 995
 996 out:
 997         RETURN(rc);
 998 }
 999
1000 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1001                                    struct mdt_object *obj,
1002                                    struct mdt_lock_handle *lhc,
1003                                    __u64 ibits, int rc)
1004 {
1005         __u64 open_flags = info->mti_spec.sp_cr_flags;
1006         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1007         ENTRY;
1008
1009         if (req_is_replay(mdt_info_req(info)))
1010                 RETURN_EXIT;
1011
1012         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1013          * return to client side. */
1014         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1015                 mdt_object_unlock(info, obj, ll, 1);
1016
1017         ll = &info->mti_lh[MDT_LH_LAYOUT];
1018         /* Release local layout lock, layout was created */
1019         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1020                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1021                 mdt_object_unlock(info, obj, ll, 1);
1022         }
1023
1024         if (open_flags & MDS_OPEN_LEASE)
1025                 up_write(&obj->mot_open_sem);
1026         else
1027                 up_read(&obj->mot_open_sem);
1028
1029         /* Cross-ref case, the lock should be returned to the client */
1030         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1031                 RETURN_EXIT;
1032
1033         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1034             !(ibits & MDS_INODELOCK_DOM)) {
1035                 /* for the open request, the lock will only return to client
1036                  * if open or layout lock is granted. */
1037                 rc = 1;
1038         }
1039
1040         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1041                 struct ldlm_reply       *ldlm_rep;
1042
1043                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1044                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1045                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1046                         mdt_object_unlock(info, obj, lhc, 1);
1047         }
1048         RETURN_EXIT;
1049 }
1050
1051 /**
1052  * Check release is permitted for the current HSM flags.
1053  */
1054 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1055 {
1056         if (!(ma->ma_valid & MA_HSM))
1057                 return false;
1058
1059         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1060                 return false;
1061
1062         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1063                 return false;
1064
1065         return true;
1066 }
1067
1068 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1069                                 struct ldlm_reply *rep,
1070                                 struct mdt_lock_handle *lhc)
1071 {
1072         const struct lu_env *env = info->mti_env;
1073         struct mdt_device *mdt = info->mti_mdt;
1074         u64 open_flags = info->mti_spec.sp_cr_flags;
1075         struct mdt_reint_record *rr = &info->mti_rr;
1076         struct md_attr *ma = &info->mti_attr;
1077         struct mdt_object *parent = NULL;
1078         struct mdt_object *o;
1079         bool object_locked = false;
1080         u64 ibits = 0;
1081         int rc;
1082
1083         ENTRY;
1084         if (md_should_create(open_flags)) {
1085                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1086                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1087                         if (IS_ERR(parent)) {
1088                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1089                                        " for anonymous created %ld, try to"
1090                                        " use server-side parent.\n",
1091                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1092                                 parent = NULL;
1093                         }
1094                 }
1095                 if (parent == NULL)
1096                         ma->ma_need |= MA_PFID;
1097         }
1098
1099         o = mdt_object_find(env, mdt, rr->rr_fid2);
1100         if (IS_ERR(o))
1101                 GOTO(out_parent_put, rc = PTR_ERR(o));
1102
1103         if (mdt_object_remote(o)) {
1104                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1105                        mdt_obd_name(info->mti_mdt),
1106                        PFID(rr->rr_fid2));
1107                 GOTO(out, rc = -EREMOTE);
1108         } else if (!mdt_object_exists(o)) {
1109                 mdt_set_disposition(info, rep,
1110                                     DISP_IT_EXECD |
1111                                     DISP_LOOKUP_EXECD |
1112                                     DISP_LOOKUP_NEG);
1113                 GOTO(out, rc = -ENOENT);
1114         }
1115
1116         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1117
1118         mdt_prep_ma_buf_from_rep(info, o, ma);
1119         if (open_flags & MDS_OPEN_RELEASE)
1120                 ma->ma_need |= MA_HSM;
1121         rc = mdt_attr_get_complex(info, o, ma);
1122         if (rc)
1123                 GOTO(out, rc);
1124
1125         /* We should not change file's existing LOV EA */
1126         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1127             open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1128                 GOTO(out, rc = -EEXIST);
1129
1130         /* If a release request, check file open flags are fine and ask for an
1131          * exclusive open access. */
1132         if (open_flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1133                 GOTO(out, rc = -EPERM);
1134
1135         rc = mdt_check_resent_lock(info, o, lhc);
1136         if (rc < 0) {
1137                 GOTO(out, rc);
1138         } else if (rc > 0) {
1139                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1140                 object_locked = true;
1141                 if (rc)
1142                         GOTO(out_unlock, rc);
1143         }
1144
1145         if (ma->ma_valid & MA_PFID) {
1146                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1147                 if (IS_ERR(parent)) {
1148                         CDEBUG(D_INODE, "Fail to find parent "DFID
1149                                " for anonymous created %ld, try to"
1150                                " use system default.\n",
1151                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1152                         parent = NULL;
1153                 }
1154         }
1155
1156         rc = mdt_finish_open(info, parent, o, open_flags, 0, rep);
1157         if (!rc) {
1158                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1159                 if (open_flags & MDS_OPEN_LOCK)
1160                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1161                 if (open_flags & MDS_OPEN_LEASE)
1162                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1163         }
1164         GOTO(out_unlock, rc);
1165
1166 out_unlock:
1167         if (object_locked)
1168                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1169 out:
1170         mdt_object_put(env, o);
1171         if (rc == 0)
1172                 mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
1173 out_parent_put:
1174         if (parent != NULL)
1175                 mdt_object_put(env, parent);
1176         return rc;
1177 }
1178
1179 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1180 static int mdt_cross_open(struct mdt_thread_info *info,
1181                           const struct lu_fid *parent_fid,
1182                           const struct lu_fid *fid,
1183                           struct ldlm_reply *rep, u64 open_flags)
1184 {
1185         struct md_attr *ma = &info->mti_attr;
1186         struct mdt_object *o;
1187         int rc;
1188         ENTRY;
1189
1190         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1191         if (IS_ERR(o))
1192                 RETURN(rc = PTR_ERR(o));
1193
1194         if (mdt_object_remote(o)) {
1195                 /* Something is wrong here, the object is on another MDS! */
1196                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1197                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1198                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1199                                 &o->mot_obj,
1200                                 "Object isn't on this server! FLD error?");
1201                 rc = -EFAULT;
1202         } else {
1203                 if (mdt_object_exists(o)) {
1204                         int mask;
1205
1206                         /* Do permission check for cross-open after converting
1207                          * MDS_OPEN_* flags to MAY_* permission mask.
1208                          */
1209                         mask = mds_accmode(open_flags);
1210
1211                         rc = mo_permission(info->mti_env, NULL,
1212                                            mdt_object_child(o), NULL, mask);
1213                         if (rc)
1214                                 goto out;
1215
1216                         mdt_prep_ma_buf_from_rep(info, o, ma);
1217                         rc = mdt_attr_get_complex(info, o, ma);
1218                         if (rc != 0)
1219                                 GOTO(out, rc);
1220
1221                         mdt_pack_secctx_in_reply(info, o);
1222
1223                         rc = mdt_finish_open(info, NULL, o, open_flags, 0, rep);
1224                 } else {
1225                         /*
1226                          * Something is wrong here. lookup was positive but
1227                          * there is no object!
1228                          */
1229                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1230                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1231                         rc = -EFAULT;
1232                 }
1233         }
1234 out:
1235         mdt_object_put(info->mti_env, o);
1236
1237         RETURN(rc);
1238 }
1239
1240 /*
1241  * find root object and take its xattr lock if it's on remote MDT, later create
1242  * may use fs default striping (which is stored in root xattr).
1243  */
1244 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1245                                struct mdt_device *mdt)
1246 {
1247         struct mdt_object *md_root = mdt->mdt_md_root;
1248         struct lustre_handle lhroot;
1249         int rc;
1250
1251         if (md_root == NULL) {
1252                 lu_root_fid(&info->mti_tmp_fid1);
1253                 md_root = mdt_object_find(info->mti_env, mdt,
1254                                           &info->mti_tmp_fid1);
1255                 if (IS_ERR(md_root))
1256                         return PTR_ERR(md_root);
1257
1258                 spin_lock(&mdt->mdt_lock);
1259                 if (mdt->mdt_md_root != NULL) {
1260                         spin_unlock(&mdt->mdt_lock);
1261
1262                         LASSERTF(mdt->mdt_md_root == md_root,
1263                                  "Different root object ("
1264                                  DFID") instances, %p, %p\n",
1265                                  PFID(&info->mti_tmp_fid1),
1266                                  mdt->mdt_md_root, md_root);
1267                         LASSERT(atomic_read(
1268                                 &md_root->mot_obj.lo_header->loh_ref) > 1);
1269
1270                         mdt_object_put(info->mti_env, md_root);
1271                 } else {
1272                         mdt->mdt_md_root = md_root;
1273                         spin_unlock(&mdt->mdt_lock);
1274                 }
1275         }
1276
1277         if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1278                 return 0;
1279
1280         rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root),
1281                                     &lhroot, LCK_PR, MDS_INODELOCK_XATTR,
1282                                     true);
1283         if (rc < 0)
1284                 return rc;
1285
1286         md_root->mot_cache_attr = 1;
1287
1288         /* don't cancel this lock, so that we know the cached xattr is valid. */
1289         ldlm_lock_decref(&lhroot, LCK_PR);
1290
1291         return 0;
1292 }
1293
1294 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1295 {
1296         struct mdt_device *mdt = info->mti_mdt;
1297         struct ptlrpc_request *req = mdt_info_req(info);
1298         struct mdt_object *parent;
1299         struct mdt_object *child;
1300         struct mdt_lock_handle *lh;
1301         struct ldlm_reply *ldlm_rep;
1302         struct mdt_body *repbody;
1303         struct lu_fid *child_fid = &info->mti_tmp_fid1;
1304         struct md_attr *ma = &info->mti_attr;
1305         u64 open_flags = info->mti_spec.sp_cr_flags;
1306         u64 ibits = 0;
1307         struct mdt_reint_record *rr = &info->mti_rr;
1308         int result, rc;
1309         int created = 0;
1310         int object_locked = 0;
1311         u32 msg_flags;
1312
1313         ENTRY;
1314         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1315                                (obd_timeout + 1) / 4);
1316
1317         mdt_counter_incr(req, LPROC_MDT_OPEN);
1318         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1319
1320         ma->ma_need = MA_INODE;
1321         ma->ma_valid = 0;
1322
1323         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1324         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1325
1326         if (unlikely(open_flags & MDS_OPEN_JOIN_FILE)) {
1327                 CERROR("file join is not supported anymore.\n");
1328                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1329         }
1330         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1331
1332         if ((open_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1333             info->mti_spec.u.sp_ea.eadata == NULL)
1334                 GOTO(out, result = err_serious(-EINVAL));
1335
1336         if (open_flags & MDS_FMODE_WRITE &&
1337             exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1338                 GOTO(out, result = -EROFS);
1339
1340         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1341                "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1342                PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2),
1343                open_flags, ma->ma_attr.la_mode, msg_flags);
1344
1345         if (info->mti_cross_ref) {
1346                 /* This is cross-ref open */
1347                 mdt_set_disposition(info, ldlm_rep,
1348                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1349                              DISP_LOOKUP_POS));
1350                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1351                                         ldlm_rep, open_flags);
1352                 GOTO(out, result);
1353         } else if (req_is_replay(req)) {
1354                 result = mdt_open_by_fid(info, ldlm_rep);
1355
1356                 if (result != -ENOENT)
1357                         GOTO(out, result);
1358
1359                 /* We didn't find the correct object, so we need to re-create it
1360                  * via a regular replay. */
1361                 if (!(open_flags & MDS_OPEN_CREAT)) {
1362                         DEBUG_REQ(D_ERROR, req,
1363                                   "OPEN & CREAT not in open replay/by_fid");
1364                         GOTO(out, result = -EFAULT);
1365                 }
1366                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1367         } else if (open_flags & MDS_OPEN_BY_FID) {
1368                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1369                 if (result < 0)
1370                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1371                                PFID(rr->rr_fid2), result);
1372                 GOTO(out, result);
1373         }
1374
1375         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1376                 GOTO(out, result = err_serious(-ENOMEM));
1377
1378         mdt_set_disposition(info, ldlm_rep,
1379                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1380
1381         if (!lu_name_is_valid(&rr->rr_name))
1382                 GOTO(out, result = -EPROTO);
1383
1384         result = mdt_lock_root_xattr(info, mdt);
1385         if (result < 0)
1386                 GOTO(out, result);
1387
1388 again:
1389         lh = &info->mti_lh[MDT_LH_PARENT];
1390         mdt_lock_pdo_init(lh, (open_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1391                           &rr->rr_name);
1392
1393         parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1394         if (IS_ERR(parent))
1395                 GOTO(out, result = PTR_ERR(parent));
1396
1397         result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1398         if (result != 0) {
1399                 mdt_object_put(info->mti_env, parent);
1400                 GOTO(out, result);
1401         }
1402
1403         /* get and check version of parent */
1404         result = mdt_version_get_check(info, parent, 0);
1405         if (result)
1406                 GOTO(out_parent, result);
1407
1408         fid_zero(child_fid);
1409
1410         result = -ENOENT;
1411         if ((open_flags & MDS_OPEN_VOLATILE) == 0)
1412                 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1413                                     &rr->rr_name, child_fid, &info->mti_spec);
1414
1415         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1416                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1417                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1418                  PFID(child_fid));
1419
1420         if (result != 0 && result != -ENOENT && result != -ESTALE)
1421                 GOTO(out_parent, result);
1422
1423         if (result == -ENOENT || result == -ESTALE) {
1424                 /* If the object is dead, let's check if the object
1425                  * is being migrated to a new object */
1426                 if (result == -ESTALE) {
1427                         struct lu_buf lmv_buf;
1428
1429                         lmv_buf.lb_buf = info->mti_xattr_buf;
1430                         lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
1431                         rc = mo_xattr_get(info->mti_env,
1432                                           mdt_object_child(parent),
1433                                           &lmv_buf, XATTR_NAME_LMV);
1434                         if (rc > 0) {
1435                                 struct lmv_mds_md_v1 *lmv;
1436
1437                                 lmv = lmv_buf.lb_buf;
1438                                 if (le32_to_cpu(lmv->lmv_hash_type) &
1439                                                 LMV_HASH_FLAG_MIGRATION) {
1440                                         /* Get the new parent FID and retry */
1441                                         mdt_object_unlock_put(info, parent,
1442                                                               lh, 1);
1443                                         mdt_lock_handle_init(lh);
1444                                         fid_le_to_cpu(
1445                                                 (struct lu_fid *)rr->rr_fid1,
1446                                                 &lmv->lmv_stripe_fids[1]);
1447                                         goto again;
1448                                 }
1449                         }
1450                 }
1451
1452                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1453                 if (result == -ESTALE) {
1454                         /*
1455                          * -ESTALE means the parent is a dead(unlinked) dir, so
1456                          * it should return -ENOENT to in accordance with the
1457                          * original mds implementaion.
1458                          */
1459                         GOTO(out_parent, result = -ENOENT);
1460                 }
1461
1462                 if (!(open_flags & MDS_OPEN_CREAT))
1463                         GOTO(out_parent, result);
1464                 if (mdt_rdonly(req->rq_export))
1465                         GOTO(out_parent, result = -EROFS);
1466                 *child_fid = *info->mti_rr.rr_fid2;
1467                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1468                          PFID(child_fid));
1469                 /* In the function below, .hs_keycmp resolves to
1470                  * lu_obj_hop_keycmp() */
1471                 /* coverity[overrun-buffer-val] */
1472                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1473         } else {
1474                 /*
1475                  * Check for O_EXCL is moved to the mdt_finish_open(),
1476                  * we need to return FID back in that case.
1477                  */
1478                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1479                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1480         }
1481         if (IS_ERR(child))
1482                 GOTO(out_parent, result = PTR_ERR(child));
1483
1484         /** check version of child  */
1485         rc = mdt_version_get_check(info, child, 1);
1486         if (rc)
1487                 GOTO(out_child, result = rc);
1488
1489         if (result == -ENOENT) {
1490                 /* Create under OBF and .lustre is not permitted */
1491                 if (!fid_is_md_operative(rr->rr_fid1) &&
1492                     (open_flags & MDS_OPEN_VOLATILE) == 0)
1493                         GOTO(out_child, result = -EPERM);
1494
1495                 /* save versions in reply */
1496                 mdt_version_get_save(info, parent, 0);
1497                 mdt_version_get_save(info, child, 1);
1498
1499                 /* version of child will be changed */
1500                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1501
1502                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1503                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1504
1505                 /* Let lower layers know what is lock mode on directory. */
1506                 info->mti_spec.sp_cr_mode =
1507                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1508
1509                 /* Don't do lookup sanity check. We know name doesn't exist. */
1510                 info->mti_spec.sp_cr_lookup = 0;
1511                 info->mti_spec.sp_feat = &dt_directory_features;
1512
1513                 result = mdo_create(info->mti_env, mdt_object_child(parent),
1514                                     &rr->rr_name, mdt_object_child(child),
1515                                     &info->mti_spec, &info->mti_attr);
1516                 if (result == -ERESTART) {
1517                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1518                         GOTO(out_child, result);
1519                 } else {
1520                         mdt_prep_ma_buf_from_rep(info, child, ma);
1521                         /* XXX: we should call this once, see few lines below */
1522                         if (result == 0)
1523                                 result = mdt_attr_get_complex(info, child, ma);
1524
1525                         if (result != 0)
1526                                 GOTO(out_child, result);
1527                 }
1528                 created = 1;
1529                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
1530         } else {
1531                 /*
1532                  * The object is on remote node, return its FID for remote open.
1533                  */
1534                 if (mdt_object_remote(child)) {
1535                         /*
1536                          * Check if this lock already was sent to client and
1537                          * this is resent case. For resent case do not take lock
1538                          * again, use what is already granted.
1539                          */
1540                         LASSERT(lhc != NULL);
1541
1542                         rc = mdt_check_resent_lock(info, child, lhc);
1543                         if (rc < 0) {
1544                                 GOTO(out_child, result = rc);
1545                         } else if (rc > 0) {
1546                                 mdt_lock_handle_init(lhc);
1547                                 mdt_lock_reg_init(lhc, LCK_PR);
1548
1549                                 rc = mdt_object_lock(info, child, lhc,
1550                                                      MDS_INODELOCK_LOOKUP);
1551                         }
1552                         repbody->mbo_fid1 = *mdt_object_fid(child);
1553                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1554                         if (rc != 0)
1555                                 result = rc;
1556                         else
1557                                 result = -MDT_EREMOTE_OPEN;
1558                         GOTO(out_child, result);
1559                 } else if (mdt_object_exists(child)) {
1560                         /* We have to get attr & LOV EA & HSM for this
1561                          * object. */
1562                         mdt_prep_ma_buf_from_rep(info, child, ma);
1563                         ma->ma_need |= MA_HSM;
1564                         result = mdt_attr_get_complex(info, child, ma);
1565                         if (result != 0)
1566                                 GOTO(out_child, result);
1567                 } else {
1568                         /* Object does not exist. Likely FS corruption. */
1569                         CERROR("%s: name '"DNAME"' present, but FID "
1570                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1571                                PNAME(&rr->rr_name), PFID(child_fid));
1572                         GOTO(out_child, result = -EIO);
1573                 }
1574         }
1575
1576         mdt_pack_secctx_in_reply(info, child);
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         EXIT;
1631 out_child_unlock:
1632         if (object_locked)
1633                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1634 out_child:
1635         mdt_object_put(info->mti_env, child);
1636         if (result == 0)
1637                 mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh);
1638 out_parent:
1639         mdt_object_unlock_put(info, parent, lh, result || !created);
1640 out:
1641         if (result)
1642                 lustre_msg_set_transno(req->rq_repmsg, 0);
1643         return result;
1644 }
1645
1646 /**
1647  * Create an orphan object use local root.
1648  */
1649 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1650                                           struct mdt_device *mdt,
1651                                           const struct lu_fid *fid,
1652                                           struct md_attr *attr, fmode_t fmode)
1653 {
1654         const struct lu_env *env = info->mti_env;
1655         struct md_op_spec *spec = &info->mti_spec;
1656         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1657         struct mdt_object *obj = NULL;
1658         struct mdt_object *local_root;
1659         static const struct lu_name lname = {
1660                 .ln_name = "i_am_nobody",
1661                 .ln_namelen = sizeof("i_am_nobody") - 1,
1662         };
1663         struct lu_ucred *uc;
1664         cfs_cap_t uc_cap_save;
1665         int rc;
1666         ENTRY;
1667
1668         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1669         if (rc != 0)
1670                 RETURN(ERR_PTR(rc));
1671
1672         local_root = mdt_object_find(env, mdt, local_root_fid);
1673         if (IS_ERR(local_root))
1674                 RETURN(local_root);
1675
1676         obj = mdt_object_new(env, mdt, fid);
1677         if (IS_ERR(obj))
1678                 GOTO(out, rc = PTR_ERR(obj));
1679
1680         spec->sp_cr_lookup = 0;
1681         spec->sp_feat = &dt_directory_features;
1682         spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1683         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1684         if (attr->ma_valid & MA_LOV) {
1685                 spec->u.sp_ea.eadata = attr->ma_lmm;
1686                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1687                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1688         } else {
1689                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1690         }
1691
1692         uc = lu_ucred(env);
1693         uc_cap_save = uc->uc_cap;
1694         uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1695         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1696                         mdt_object_child(obj), spec, attr);
1697         uc->uc_cap = uc_cap_save;
1698         if (rc < 0) {
1699                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1700                        mdt_obd_name(mdt), PFID(fid), rc);
1701                 GOTO(out, rc);
1702         }
1703
1704         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED, spec);
1705         if (rc < 0)
1706                 CERROR("%s: cannot open volatile file "DFID", orphan "
1707                        "file will be left in PENDING directory until "
1708                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1709                        PFID(fid), rc);
1710         GOTO(out, rc);
1711
1712 out:
1713         if (rc < 0) {
1714                 if (!IS_ERR(obj))
1715                         mdt_object_put(env, obj);
1716                 obj = ERR_PTR(rc);
1717         }
1718         mdt_object_put(env, local_root);
1719         return obj;
1720 }
1721
1722 /* XXX Look into layout in MDT layer. */
1723 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1724 {
1725         struct lov_comp_md_v1   *comp_v1;
1726         struct lov_mds_md       *v1;
1727         __u32   off;
1728         int     i;
1729
1730         if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1731                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1732
1733                 if (comp_v1->lcm_entry_count == 0)
1734                         return -EINVAL;
1735
1736                 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1737                         off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1738                         v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1739                         v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1740                 }
1741         } else {
1742                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1743         }
1744         return 0;
1745 }
1746
1747 static inline int mdt_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1748 {
1749         struct lov_comp_md_v1 *comp_v1;
1750
1751         if (le32_to_cpu(lmm->lmm_magic == LOV_MAGIC_COMP_V1)) {
1752                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1753                 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1754         } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1755                    le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1756                 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1757         } else {
1758                 return -EINVAL;
1759         }
1760         return 0;
1761 }
1762
1763 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1764                            struct md_attr *ma)
1765 {
1766         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1767         struct lu_ucred        *uc = mdt_ucred(info);
1768         struct close_data      *data;
1769         struct ldlm_lock       *lease;
1770         struct mdt_object      *orphan;
1771         struct md_attr         *orp_ma;
1772         struct lu_buf          *buf;
1773         cfs_cap_t               cap;
1774         bool                    lease_broken;
1775         int                     rc;
1776         int                     rc2;
1777         ENTRY;
1778
1779         if (mdt_rdonly(info->mti_exp))
1780                 RETURN(-EROFS);
1781
1782         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1783         if (data == NULL)
1784                 RETURN(-EPROTO);
1785
1786         lease = ldlm_handle2lock(&data->cd_handle);
1787         if (lease == NULL)
1788                 RETURN(-ESTALE);
1789
1790         /* try to hold open_sem so that nobody else can open the file */
1791         if (!down_write_trylock(&o->mot_open_sem)) {
1792                 ldlm_lock_cancel(lease);
1793                 GOTO(out_reprocess, rc = -EBUSY);
1794         }
1795
1796         /* Check if the lease open lease has already canceled */
1797         lock_res_and_lock(lease);
1798         lease_broken = ldlm_is_cancel(lease);
1799         unlock_res_and_lock(lease);
1800
1801         LDLM_DEBUG(lease, DFID " lease broken? %d",
1802                    PFID(mdt_object_fid(o)), lease_broken);
1803
1804         /* Cancel server side lease. Client side counterpart should
1805          * have been cancelled. It's okay to cancel it now as we've
1806          * held mot_open_sem. */
1807         ldlm_lock_cancel(lease);
1808
1809         if (lease_broken) /* don't perform release task */
1810                 GOTO(out_unlock, rc = -ESTALE);
1811
1812         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1813                 GOTO(out_unlock, rc = -EINVAL);
1814
1815         /* ma_need was set before but it seems fine to change it in order to
1816          * avoid modifying the one from RPC */
1817         ma->ma_need = MA_HSM;
1818         rc = mdt_attr_get_complex(info, o, ma);
1819         if (rc != 0)
1820                 GOTO(out_unlock, rc);
1821
1822         if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
1823                 if (ma->ma_valid & MA_HSM) {
1824                         if (ma->ma_hsm.mh_flags & HS_RELEASED)
1825                                 GOTO(out_unlock, rc = -EALREADY);
1826
1827                         if (ma->ma_hsm.mh_arch_id != data->cd_archive_id)
1828                                 CDEBUG(D_CACHE,
1829                                        DFID" archive id diff: %llu:%u\n",
1830                                        PFID(mdt_object_fid(o)),
1831                                        ma->ma_hsm.mh_arch_id,
1832                                        data->cd_archive_id);
1833
1834                         if (!(ma->ma_hsm.mh_flags & HS_DIRTY) &&
1835                             ma->ma_hsm.mh_arch_ver == data->cd_data_version) {
1836                                 CDEBUG(D_CACHE,
1837                                        DFID" data version matches: packed=%llu "
1838                                        "and on-disk=%llu\n",
1839                                        PFID(mdt_object_fid(o)),
1840                                        data->cd_data_version,
1841                                        ma->ma_hsm.mh_arch_ver);
1842                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1843                         }
1844
1845                         if (ma->ma_hsm.mh_flags & HS_DIRTY)
1846                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1847                 } else {
1848                         /* Set up HSM attribte for PCC archived object */
1849                         BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1850                                      sizeof(info->mti_xattr_buf));
1851                         buf = &info->mti_buf;
1852                         buf->lb_buf = info->mti_xattr_buf;
1853                         buf->lb_len = sizeof(struct hsm_attrs);
1854                         memset(&ma->ma_hsm, 0, sizeof(ma->ma_hsm));
1855                         ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1856                         ma->ma_hsm.mh_arch_id = data->cd_archive_id;
1857                         ma->ma_hsm.mh_arch_ver = data->cd_data_version;
1858                         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1859
1860                         rc = mo_xattr_set(info->mti_env, mdt_object_child(o),
1861                                           buf, XATTR_NAME_HSM, 0);
1862                         if (rc)
1863                                 GOTO(out_unlock, rc);
1864                 }
1865         } else {
1866                 if (!mdt_hsm_release_allow(ma))
1867                         GOTO(out_unlock, rc = -EPERM);
1868
1869                 /* already released? */
1870                 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1871                         GOTO(out_unlock, rc = 0);
1872
1873                 /* Compare on-disk and packed data_version */
1874                 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1875                         CDEBUG(D_HSM, DFID" data_version mismatches: "
1876                                "packed=%llu and on-disk=%llu\n",
1877                                PFID(mdt_object_fid(o)),
1878                                data->cd_data_version,
1879                                ma->ma_hsm.mh_arch_ver);
1880                         GOTO(out_unlock, rc = -EPERM);
1881                 }
1882         }
1883
1884         ma->ma_valid = MA_INODE;
1885         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1886         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1887         if (rc < 0)
1888                 GOTO(out_unlock, rc);
1889
1890         mutex_lock(&o->mot_som_mutex);
1891         rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size,
1892                            ma->ma_attr.la_blocks);
1893         mutex_unlock(&o->mot_som_mutex);
1894         if (rc2 < 0)
1895                 CDEBUG(D_INODE,
1896                        "%s: File "DFID" SOM update failed: rc = %d\n",
1897                        mdt_obd_name(info->mti_mdt),
1898                        PFID(mdt_object_fid(o)), rc2);
1899
1900
1901         ma->ma_need = MA_INODE | MA_LOV;
1902         rc = mdt_attr_get_complex(info, o, ma);
1903         if (rc < 0)
1904                 GOTO(out_unlock, rc);
1905
1906         if (!(ma->ma_valid & MA_LOV)) {
1907                 /* Even empty file are released */
1908                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1909                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
1910                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1911                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1912                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1913         } else {
1914                 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
1915                  * ma_lmm as lov_user_md, then it will be confused by union of
1916                  * layout_gen and stripe_offset. */
1917                 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
1918                     LOV_MAGIC_MAGIC)
1919                         ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1920                 else
1921                         GOTO(out_unlock, rc = -EINVAL);
1922         }
1923
1924         /* Set file as released. */
1925         rc = mdt_hsm_set_released(ma->ma_lmm);
1926         if (rc)
1927                 GOTO(out_unlock, rc);
1928
1929         orp_ma = &info->mti_u.hsm.attr;
1930         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1931         /* We use root ownership to bypass potential quota
1932          * restrictions on the user and group of the file. */
1933         orp_ma->ma_attr.la_uid = 0;
1934         orp_ma->ma_attr.la_gid = 0;
1935         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1936         orp_ma->ma_lmm = ma->ma_lmm;
1937         orp_ma->ma_lmm_size = ma->ma_lmm_size;
1938         orp_ma->ma_valid = MA_INODE | MA_LOV;
1939         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1940                                  MDS_FMODE_WRITE);
1941         if (IS_ERR(orphan)) {
1942                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1943                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1944                        PTR_ERR(orphan));
1945                 GOTO(out_unlock, rc = PTR_ERR(orphan));
1946         }
1947
1948         /* Set up HSM attribute for orphan object */
1949         BUILD_BUG_ON(sizeof(struct hsm_attrs) > sizeof(info->mti_xattr_buf));
1950         buf = &info->mti_buf;
1951         buf->lb_buf = info->mti_xattr_buf;
1952         buf->lb_len = sizeof(struct hsm_attrs);
1953         ma->ma_hsm.mh_flags |= HS_RELEASED;
1954         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1955         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
1956
1957         mdt_lock_reg_init(lh, LCK_EX);
1958         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
1959                              MDS_INODELOCK_XATTR);
1960         if (rc != 0)
1961                 GOTO(out_close, rc);
1962
1963         /* The orphan has root ownership so we need to raise
1964          * CAP_FOWNER to set the HSM attributes. */
1965         cap = uc->uc_cap;
1966         uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
1967         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
1968                           XATTR_NAME_HSM, 0);
1969         uc->uc_cap = cap;
1970         if (rc != 0)
1971                 GOTO(out_layout_lock, rc);
1972
1973         /* Swap layout with orphan objects. */
1974         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
1975                              mdt_object_child(orphan),
1976                              SWAP_LAYOUTS_MDS_HSM);
1977
1978         if (!rc && ma->ma_attr_flags & MDS_PCC_ATTACH) {
1979                 ma->ma_need = MA_LOV;
1980                 rc = mdt_attr_get_complex(info, o, ma);
1981         }
1982
1983         EXIT;
1984
1985 out_layout_lock:
1986         /* Release exclusive LL */
1987         mdt_object_unlock(info, o, lh, 1);
1988 out_close:
1989         /* Close orphan object anyway */
1990         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
1991                        MDS_FMODE_WRITE);
1992         if (rc2 < 0)
1993                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
1994                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
1995         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
1996                         "object closed");
1997         mdt_object_put(info->mti_env, orphan);
1998
1999 out_unlock:
2000         up_write(&o->mot_open_sem);
2001
2002         /* already released */
2003         if (rc == 0) {
2004                 struct mdt_body *repbody;
2005
2006                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2007                 LASSERT(repbody != NULL);
2008                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2009                 if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
2010                         LASSERT(ma->ma_valid & MA_LOV);
2011                         rc = mdt_get_lmm_gen(ma->ma_lmm,
2012                                              &repbody->mbo_layout_gen);
2013                         if (!rc)
2014                                 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
2015                 }
2016         }
2017
2018 out_reprocess:
2019         ldlm_reprocess_all(lease->l_resource, lease);
2020         LDLM_LOCK_PUT(lease);
2021
2022         ma->ma_valid = 0;
2023         ma->ma_need = 0;
2024
2025         return rc;
2026 }
2027
2028 int mdt_close_handle_layouts(struct mdt_thread_info *info,
2029                              struct mdt_object *o, struct md_attr *ma)
2030 {
2031         struct mdt_lock_handle  *lh1 = &info->mti_lh[MDT_LH_NEW];
2032         struct mdt_lock_handle  *lh2 = &info->mti_lh[MDT_LH_OLD];
2033         struct close_data       *data;
2034         struct ldlm_lock        *lease;
2035         struct mdt_object       *o1 = o, *o2;
2036         bool                     lease_broken;
2037         bool                     swap_objects;
2038         int                      rc;
2039         ENTRY;
2040
2041         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2042                 RETURN(-EROFS);
2043
2044         if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
2045                 RETURN(-EINVAL);
2046
2047         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2048         if (data == NULL)
2049                 RETURN(-EPROTO);
2050
2051         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2052                 RETURN(-EINVAL);
2053
2054         rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
2055         if (unlikely(rc == 0))
2056                 RETURN(-EINVAL);
2057
2058         /* Exchange o1 and o2, to enforce locking order */
2059         swap_objects = (rc < 0);
2060
2061         lease = ldlm_handle2lock(&data->cd_handle);
2062         if (lease == NULL)
2063                 RETURN(-ESTALE);
2064
2065         o2 = mdt_object_find(info->mti_env, info->mti_mdt, &data->cd_fid);
2066         if (IS_ERR(o2))
2067                 GOTO(out_lease, rc = PTR_ERR(o2));
2068
2069         if (!S_ISREG(lu_object_attr(&o2->mot_obj))) {
2070                 swap_objects = false; /* not swapped yet */
2071                 GOTO(out_obj, rc = -EINVAL);
2072         }
2073
2074         if (swap_objects)
2075                 swap(o1, o2);
2076
2077         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
2078                            MAY_WRITE);
2079         if (rc < 0)
2080                 GOTO(out_obj, rc);
2081
2082         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
2083                            MAY_WRITE);
2084         if (rc < 0)
2085                 GOTO(out_obj, rc);
2086
2087         /* try to hold open_sem so that nobody else can open the file */
2088         if (!down_write_trylock(&o->mot_open_sem)) {
2089                 ldlm_lock_cancel(lease);
2090                 GOTO(out_obj, rc = -EBUSY);
2091         }
2092
2093         /* Check if the lease open lease has already canceled */
2094         lock_res_and_lock(lease);
2095         lease_broken = ldlm_is_cancel(lease);
2096         unlock_res_and_lock(lease);
2097
2098         LDLM_DEBUG(lease, DFID " lease broken? %d",
2099                    PFID(mdt_object_fid(o)), lease_broken);
2100
2101         /* Cancel server side lease. Client side counterpart should
2102          * have been cancelled. It's okay to cancel it now as we've
2103          * held mot_open_sem. */
2104         ldlm_lock_cancel(lease);
2105
2106         if (lease_broken)
2107                 GOTO(out_unlock_sem, rc = -ESTALE);
2108
2109         mdt_lock_reg_init(lh1, LCK_EX);
2110         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
2111                              MDS_INODELOCK_XATTR);
2112         if (rc < 0)
2113                 GOTO(out_unlock_sem, rc);
2114
2115         mdt_lock_reg_init(lh2, LCK_EX);
2116         rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2117                              MDS_INODELOCK_XATTR);
2118         if (rc < 0)
2119                 GOTO(out_unlock1, rc);
2120
2121         /* Swap layout with orphan object */
2122         if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2123                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2124                                      mdt_object_child(o2), 0);
2125         } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2126                    ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2127                 struct lu_buf *buf = &info->mti_buf;
2128                 struct md_rejig_data mrd;
2129
2130                 mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2131                 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)
2132                         mrd.mrd_mirror_id = data->cd_mirror_id;
2133
2134                 buf->lb_len = sizeof(mrd);
2135                 buf->lb_buf = &mrd;
2136                 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2137                                   XATTR_LUSTRE_LOV,
2138                                   ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2139                                   LU_XATTR_SPLIT : LU_XATTR_MERGE);
2140                 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2141                         int rc2;
2142
2143                         mutex_lock(&o->mot_som_mutex);
2144                         rc2 = mdt_set_som(info, o, SOM_FL_STRICT,
2145                                           ma->ma_attr.la_size,
2146                                           ma->ma_attr.la_blocks);
2147                         mutex_unlock(&o->mot_som_mutex);
2148                         if (rc2 < 0)
2149                                 CERROR(DFID": Setting i_blocks error: %d, "
2150                                        "i_blocks will be reported wrongly and "
2151                                        "can only be fixed in next resync\n",
2152                                        PFID(mdt_object_fid(o)), rc2);
2153                 }
2154         }
2155         if (rc < 0)
2156                 GOTO(out_unlock2, rc);
2157
2158         EXIT;
2159
2160 out_unlock2:
2161         /* Release exclusive LL */
2162         mdt_object_unlock(info, o2, lh2, 1);
2163
2164 out_unlock1:
2165         mdt_object_unlock(info, o1, lh1, 1);
2166
2167 out_unlock_sem:
2168         up_write(&o->mot_open_sem);
2169
2170         /* already swapped */
2171         if (rc == 0) {
2172                 struct mdt_body *repbody;
2173
2174                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2175                 LASSERT(repbody != NULL);
2176                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2177         }
2178
2179 out_obj:
2180         mdt_object_put(info->mti_env, swap_objects ? o1 : o2);
2181
2182         ldlm_reprocess_all(lease->l_resource, lease);
2183
2184 out_lease:
2185         LDLM_LOCK_PUT(lease);
2186
2187         if (ma != NULL) {
2188                 ma->ma_valid = 0;
2189                 ma->ma_need = 0;
2190         }
2191
2192         return rc;
2193 }
2194
2195 static int mdt_close_resync_done(struct mdt_thread_info *info,
2196                                  struct mdt_object *o, struct md_attr *ma)
2197 {
2198         struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LOCAL];
2199         struct close_data *data;
2200         struct ldlm_lock *lease;
2201         struct md_layout_change layout = { 0 };
2202         __u32 *resync_ids = NULL;
2203         size_t resync_count = 0;
2204         bool lease_broken;
2205         int rc;
2206
2207         ENTRY;
2208
2209         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2210                 RETURN(-EROFS);
2211
2212         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2213                 RETURN(-EINVAL);
2214
2215         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2216         if (data == NULL)
2217                 RETURN(-EPROTO);
2218
2219         if (ptlrpc_req_need_swab(mdt_info_req(info)))
2220                 lustre_swab_close_data_resync_done(&data->cd_resync);
2221
2222         if (!fid_is_zero(&data->cd_fid))
2223                 RETURN(-EPROTO);
2224
2225         lease = ldlm_handle2lock(&data->cd_handle);
2226         if (lease == NULL)
2227                 RETURN(-ESTALE);
2228
2229         /* try to hold open_sem so that nobody else can open the file */
2230         if (!down_write_trylock(&o->mot_open_sem)) {
2231                 ldlm_lock_cancel(lease);
2232                 GOTO(out_reprocess, rc = -EBUSY);
2233         }
2234
2235         /* Check if the lease open lease has already canceled */
2236         lock_res_and_lock(lease);
2237         lease_broken = ldlm_is_cancel(lease);
2238         unlock_res_and_lock(lease);
2239
2240         LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2241                    PFID(mdt_object_fid(o)), lease_broken);
2242
2243         /* Cancel server side lease. Client side counterpart should
2244          * have been cancelled. It's okay to cancel it now as we've
2245          * held mot_open_sem. */
2246         ldlm_lock_cancel(lease);
2247
2248         if (lease_broken) /* don't perform release task */
2249                 GOTO(out_unlock, rc = -ESTALE);
2250
2251         resync_count = data->cd_resync.resync_count;
2252
2253         if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2254                 void *data;
2255
2256                 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2257                                            RCL_CLIENT))
2258                         GOTO(out_unlock, rc = -EPROTO);
2259
2260                 OBD_ALLOC(resync_ids, resync_count * sizeof(__u32));
2261                 if (!resync_ids)
2262                         GOTO(out_unlock, rc = -ENOMEM);
2263
2264                 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2265                 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2266
2267                 layout.mlc_resync_ids = resync_ids;
2268         } else {
2269                 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2270         }
2271
2272         layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2273         layout.mlc_resync_count = resync_count;
2274         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2275                 layout.mlc_som.lsa_valid = SOM_FL_STRICT;
2276                 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2277                 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2278         }
2279         rc = mdt_layout_change(info, o, lhc, &layout);
2280         if (rc)
2281                 GOTO(out_unlock, rc);
2282
2283         mdt_object_unlock(info, o, lhc, 0);
2284
2285         EXIT;
2286
2287 out_unlock:
2288         up_write(&o->mot_open_sem);
2289
2290         /* already released */
2291         if (rc == 0) {
2292                 struct mdt_body *repbody;
2293
2294                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2295                 LASSERT(repbody != NULL);
2296                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2297         }
2298
2299         if (resync_ids)
2300                 OBD_FREE(resync_ids, resync_count * sizeof(__u32));
2301
2302 out_reprocess:
2303         ldlm_reprocess_all(lease->l_resource, lease);
2304         LDLM_LOCK_PUT(lease);
2305
2306         ma->ma_valid = 0;
2307         ma->ma_need = 0;
2308
2309         return rc;
2310 }
2311
2312 #define MFD_CLOSED(open_flags) ((open_flags) == MDS_FMODE_CLOSED)
2313
2314 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2315 {
2316         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_open_flags));
2317 }
2318
2319 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2320 {
2321         struct mdt_object *o = mfd->mfd_object;
2322         struct md_object *next = mdt_object_child(o);
2323         struct md_attr *ma = &info->mti_attr;
2324         struct lu_fid *ofid = &info->mti_tmp_fid1;
2325         int rc = 0;
2326         u64 open_flags;
2327         u64 intent;
2328
2329         ENTRY;
2330
2331         open_flags = mfd->mfd_open_flags;
2332         intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2333         *ofid = *mdt_object_fid(o);
2334
2335         /* the below message is checked in replay-single.sh test_46 */
2336         CDEBUG(D_INODE, "%s: %sclosing file handle "DFID" with intent: %llx\n",
2337                mdt_obd_name(info->mti_mdt),
2338                ma->ma_valid & MA_FORCE_LOG ? "force " : "", PFID(ofid), intent);
2339
2340         switch (intent) {
2341         case MDS_HSM_RELEASE: {
2342                 rc = mdt_hsm_release(info, o, ma);
2343                 if (rc < 0) {
2344                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2345                                mdt_obd_name(info->mti_mdt),
2346                                PFID(ofid), rc);
2347                         /* continue to close even error occurred. */
2348                 }
2349                 break;
2350         }
2351         case MDS_CLOSE_LAYOUT_MERGE:
2352         case MDS_CLOSE_LAYOUT_SPLIT:
2353         case MDS_CLOSE_LAYOUT_SWAP: {
2354                 rc = mdt_close_handle_layouts(info, o, ma);
2355                 if (rc < 0) {
2356                         CDEBUG(D_INODE,
2357                                "%s: cannot swap layout of "DFID": rc = %d\n",
2358                                mdt_obd_name(info->mti_mdt),
2359                                PFID(ofid), rc);
2360                         /* continue to close even if error occurred. */
2361                 }
2362                 break;
2363         }
2364         case MDS_CLOSE_RESYNC_DONE:
2365                 rc = mdt_close_resync_done(info, o, ma);
2366                 break;
2367         default:
2368                 /* nothing */
2369                 break;
2370         }
2371
2372         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
2373             ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) {
2374                 int rc2;
2375
2376                 rc2 = mdt_lsom_update(info, o, false);
2377                 if (rc2 < 0)
2378                         CDEBUG(D_INODE,
2379                                "%s: File " DFID " LSOM failed: rc = %d\n",
2380                                mdt_obd_name(info->mti_mdt),
2381                                PFID(ofid), rc2);
2382                         /* continue to close even if error occured. */
2383         }
2384
2385         if (open_flags & MDS_FMODE_WRITE)
2386                 mdt_write_put(o);
2387         else if (open_flags & MDS_FMODE_EXEC)
2388                 mdt_write_allow(o);
2389
2390         /* Update atime|mtime|ctime on close. */
2391         if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ ||
2392              open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
2393             (ma->ma_attr.la_valid & LA_ATIME ||
2394              ma->ma_attr.la_valid & LA_MTIME ||
2395              ma->ma_attr.la_valid & LA_CTIME)) {
2396                 ma->ma_valid = MA_INODE;
2397                 ma->ma_attr_flags |= MDS_CLOSE_UPDATE_TIMES;
2398                 ma->ma_attr.la_valid &= (LA_ATIME | LA_MTIME | LA_CTIME);
2399                 rc = mo_attr_set(info->mti_env, next, ma);
2400         }
2401
2402         /* If file data is modified, add the dirty flag. */
2403         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2404                 rc = mdt_add_dirty_flag(info, o, ma);
2405
2406         ma->ma_need |= MA_INODE;
2407         ma->ma_valid &= ~MA_INODE;
2408
2409         LASSERT(atomic_read(&o->mot_open_count) > 0);
2410         atomic_dec(&o->mot_open_count);
2411         mdt_handle_last_unlink(info, o, ma);
2412
2413         if (!MFD_CLOSED(open_flags)) {
2414                 rc = mo_close(info->mti_env, next, ma, open_flags);
2415                 if (mdt_dom_check_for_discard(info, o))
2416                         mdt_dom_discard_data(info, o);
2417         }
2418
2419         /* adjust open and lease count */
2420         if (open_flags & MDS_OPEN_LEASE) {
2421                 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2422                 atomic_dec(&o->mot_lease_count);
2423         }
2424
2425         mdt_mfd_free(mfd);
2426         mdt_object_put(info->mti_env, o);
2427
2428         RETURN(rc);
2429 }
2430
2431 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2432                        struct mdt_body *repbody)
2433 {
2434         struct mdt_export_data *med;
2435         struct mdt_file_data   *mfd;
2436         int                     ret = 0;
2437         int                     rc = 0;
2438         ENTRY;
2439
2440         med = &req->rq_export->exp_mdt_data;
2441         spin_lock(&med->med_open_lock);
2442         mfd = mdt_open_handle2mfd(med, &info->mti_open_handle,
2443                                   req_is_replay(req));
2444         if (mdt_mfd_closed(mfd)) {
2445                 spin_unlock(&med->med_open_lock);
2446                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2447                        ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2448                        info->mti_open_handle.cookie);
2449                 /** not serious error since bug 3633 */
2450                 rc = -ESTALE;
2451         } else {
2452                 class_handle_unhash(&mfd->mfd_open_handle);
2453                 list_del_init(&mfd->mfd_list);
2454                 spin_unlock(&med->med_open_lock);
2455                 ret = mdt_mfd_close(info, mfd);
2456         }
2457
2458         RETURN(rc ? rc : ret);
2459 }
2460
2461 int mdt_close(struct tgt_session_info *tsi)
2462 {
2463         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2464         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2465         struct md_attr         *ma = &info->mti_attr;
2466         struct mdt_body        *repbody = NULL;
2467         int rc, ret = 0;
2468         ENTRY;
2469
2470         mdt_counter_incr(req, LPROC_MDT_CLOSE);
2471         /* Close may come with the Size-on-MDS update. Unpack it. */
2472         rc = mdt_close_unpack(info);
2473         if (rc)
2474                 GOTO(out, rc = err_serious(rc));
2475
2476         /* These fields are no longer used and are left for compatibility.
2477          * size is always zero */
2478         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2479                              0);
2480         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2481                              0);
2482         rc = req_capsule_server_pack(info->mti_pill);
2483         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2484                 mdt_client_compatibility(info);
2485                 if (rc == 0)
2486                         mdt_fix_reply(info);
2487                 mdt_exit_ucred(info);
2488                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2489         }
2490
2491         /* Continue to close handle even if we can not pack reply */
2492         if (rc == 0) {
2493                 repbody = req_capsule_server_get(info->mti_pill,
2494                                                  &RMF_MDT_BODY);
2495                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2496                                                     &RMF_MDT_MD);
2497                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2498                                                        &RMF_MDT_MD,
2499                                                        RCL_SERVER);
2500                 ma->ma_need = MA_INODE | MA_LOV;
2501                 repbody->mbo_eadatasize = 0;
2502                 repbody->mbo_aclsize = 0;
2503         } else {
2504                 rc = err_serious(rc);
2505         }
2506
2507         rc = mdt_close_internal(info, req, repbody);
2508         if (rc != -ESTALE)
2509                 mdt_empty_transno(info, rc);
2510
2511         if (repbody != NULL) {
2512                 mdt_client_compatibility(info);
2513                 rc = mdt_fix_reply(info);
2514         }
2515
2516         mdt_exit_ucred(info);
2517         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2518                 GOTO(out, rc = err_serious(-ENOMEM));
2519
2520         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2521                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2522                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2523 out:
2524         mdt_thread_info_fini(info);
2525         RETURN(rc ? rc : ret);
2526 }