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, 2016, 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 "mdt_internal.h"
  44 #include <lustre_nodemap.h>
  45
  46 /* we do nothing because we do not have refcount now */
  47 static void mdt_mfd_get(void *mfdp)
  48 {
  49 }
  50
  51 static struct portals_handle_ops mfd_handle_ops = {
  52         .hop_addref = mdt_mfd_get,
  53         .hop_free   = NULL,
  54 };
  55
  56 /* Create a new mdt_file_data struct, initialize it,
  57  * and insert it to global hash table */
  58 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
  59 {
  60         struct mdt_file_data *mfd;
  61         ENTRY;
  62
  63         OBD_ALLOC_PTR(mfd);
  64         if (mfd != NULL) {
  65                 INIT_LIST_HEAD(&mfd->mfd_handle.h_link);
  66                 mfd->mfd_handle.h_owner = med;
  67                 INIT_LIST_HEAD(&mfd->mfd_list);
  68                 class_handle_hash(&mfd->mfd_handle, &mfd_handle_ops);
  69         }
  70
  71         RETURN(mfd);
  72 }
  73
  74 /*
  75  * Find the mfd pointed to by handle in global hash table.
  76  * In case of replay the handle is obsoleted
  77  * but mfd can be found in mfd list by that handle.
  78  * Callers need to be holding med_open_lock.
  79  */
  80 struct mdt_file_data *mdt_handle2mfd(struct mdt_export_data *med,
  81                                      const struct lustre_handle *handle,
  82                                      bool is_replay_or_resent)
  83 {
  84         struct mdt_file_data   *mfd;
  85         ENTRY;
  86
  87         LASSERT(handle != NULL);
  88         mfd = class_handle2object(handle->cookie, med);
  89         /* during dw/setattr replay the mfd can be found by old handle */
  90         if (mfd == NULL && is_replay_or_resent) {
  91                 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  92                         if (mfd->mfd_old_handle.cookie == handle->cookie)
  93                                 RETURN(mfd);
  94                 }
  95                 mfd = NULL;
  96         }
  97
  98         RETURN(mfd);
  99 }
 100
 101 /* free mfd */
 102 void mdt_mfd_free(struct mdt_file_data *mfd)
 103 {
 104         LASSERT(list_empty(&mfd->mfd_list));
 105         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_handle);
 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
 200         ENTRY;
 201         /* transaction has occurred already */
 202         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 203                 RETURN_EXIT;
 204
 205         if (tgt_is_multimodrpcs_client(req->rq_export)) {
 206                 struct thandle         *th;
 207
 208                 /* generate an empty transaction to get a transno
 209                  * and reply data */
 210                 th = dt_trans_create(info->mti_env, mdt->mdt_bottom);
 211                 if (!IS_ERR(th)) {
 212                         rc = dt_trans_start(info->mti_env, mdt->mdt_bottom, th);
 213                         dt_trans_stop(info->mti_env, mdt->mdt_bottom, th);
 214                 }
 215                 RETURN_EXIT;
 216         }
 217
 218         spin_lock(&mdt->mdt_lut.lut_translock);
 219         if (rc != 0) {
 220                 if (info->mti_transno != 0) {
 221                         struct obd_export *exp = req->rq_export;
 222
 223                         CERROR("%s: replay trans %llu NID %s: rc = %d\n",
 224                                mdt_obd_name(mdt), info->mti_transno,
 225                                libcfs_nid2str(exp->exp_connection->c_peer.nid),
 226                                rc);
 227                         spin_unlock(&mdt->mdt_lut.lut_translock);
 228                         RETURN_EXIT;
 229                 }
 230         } else if (info->mti_transno == 0) {
 231                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
 232         } else {
 233                 /* should be replay */
 234                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 235                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 236         }
 237         spin_unlock(&mdt->mdt_lut.lut_translock);
 238
 239         CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
 240                info->mti_transno,
 241                req->rq_export->exp_obd->obd_last_committed);
 242
 243         req->rq_transno = info->mti_transno;
 244         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 245
 246         /* update lcd in memory only for resent cases */
 247         ted = &req->rq_export->exp_target_data;
 248         LASSERT(ted);
 249         mutex_lock(&ted->ted_lcd_lock);
 250         lcd = ted->ted_lcd;
 251         if (info->mti_transno < lcd->lcd_last_transno &&
 252             info->mti_transno != 0) {
 253                 /* This should happen during replay. Do not update
 254                  * last rcvd info if replay req transno < last transno,
 255                  * otherwise the following resend(after replay) can not
 256                  * be checked correctly by xid */
 257                 mutex_unlock(&ted->ted_lcd_lock);
 258                 CDEBUG(D_HA, "%s: transno = %llu < last_transno = %llu\n",
 259                        mdt_obd_name(mdt), info->mti_transno,
 260                        lcd->lcd_last_transno);
 261                 RETURN_EXIT;
 262         }
 263
 264         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
 265                 if (info->mti_transno != 0)
 266                         lcd->lcd_last_close_transno = info->mti_transno;
 267                 lcd->lcd_last_close_xid = req->rq_xid;
 268                 lcd->lcd_last_close_result = rc;
 269         } else {
 270                 /* VBR: save versions in last_rcvd for reconstruct. */
 271                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 272                 if (pre_versions) {
 273                         lcd->lcd_pre_versions[0] = pre_versions[0];
 274                         lcd->lcd_pre_versions[1] = pre_versions[1];
 275                         lcd->lcd_pre_versions[2] = pre_versions[2];
 276                         lcd->lcd_pre_versions[3] = pre_versions[3];
 277                 }
 278                 if (info->mti_transno != 0)
 279                         lcd->lcd_last_transno = info->mti_transno;
 280
 281                 lcd->lcd_last_xid = req->rq_xid;
 282                 lcd->lcd_last_result = rc;
 283                 lcd->lcd_last_data = info->mti_opdata;
 284         }
 285         mutex_unlock(&ted->ted_lcd_lock);
 286
 287         EXIT;
 288 }
 289
 290 void mdt_mfd_set_mode(struct mdt_file_data *mfd, __u64 mode)
 291 {
 292         LASSERT(mfd != NULL);
 293
 294         CDEBUG(D_HA, DFID " Change mfd mode %#llo -> %#llo.\n",
 295                PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_mode, mode);
 296
 297         mfd->mfd_mode = mode;
 298 }
 299
 300 /**
 301  * prep ma_lmm/ma_lmv for md_attr from reply
 302  */
 303 static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
 304                                      struct mdt_object *obj,
 305                                      struct md_attr *ma)
 306 {
 307         LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
 308         if (S_ISDIR(obj->mot_header.loh_attr)) {
 309                 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
 310                                                     &RMF_MDT_MD);
 311                 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
 312                                                        &RMF_MDT_MD,
 313                                                        RCL_SERVER);
 314                 if (ma->ma_lmv_size > 0)
 315                         ma->ma_need |= MA_LMV;
 316         } else {
 317                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
 318                                                     &RMF_MDT_MD);
 319                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 320                                                        &RMF_MDT_MD,
 321                                                        RCL_SERVER);
 322                 if (ma->ma_lmm_size > 0)
 323                         ma->ma_need |= MA_LOV;
 324         }
 325 }
 326
 327 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 328                         struct mdt_object *o, __u64 flags, int created,
 329                         struct ldlm_reply *rep)
 330 {
 331         struct ptlrpc_request   *req = mdt_info_req(info);
 332         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 333         struct mdt_file_data    *mfd;
 334         struct md_attr          *ma  = &info->mti_attr;
 335         struct lu_attr          *la  = &ma->ma_attr;
 336         struct mdt_body         *repbody;
 337         int                      rc = 0, isdir, isreg;
 338         ENTRY;
 339
 340         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 341
 342         isreg = S_ISREG(la->la_mode);
 343         isdir = S_ISDIR(la->la_mode);
 344         if (isreg && !(ma->ma_valid & MA_LOV) && !(flags & MDS_OPEN_RELEASE)) {
 345                 /*
 346                  * No EA, check whether it is will set regEA and dirEA since in
 347                  * above attr get, these size might be zero, so reset it, to
 348                  * retrieve the MD after create obj.
 349                  */
 350                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 351                                                        &RMF_MDT_MD,
 352                                                        RCL_SERVER);
 353                 /* in replay case, p == NULL */
 354                 rc = mdt_create_data(info, p, o);
 355                 if (rc)
 356                         RETURN(rc);
 357
 358                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
 359                         mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
 360         }
 361
 362         CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n",
 363                ma->ma_valid, ma->ma_lmm_size);
 364
 365         if (ma->ma_valid & MA_LOV) {
 366                 LASSERT(ma->ma_lmm_size != 0);
 367                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 368                 if (isdir)
 369                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 370                 else
 371                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 372         }
 373
 374         if (ma->ma_valid & MA_LMV) {
 375                 LASSERT(ma->ma_lmv_size != 0);
 376                 repbody->mbo_eadatasize = ma->ma_lmv_size;
 377                 LASSERT(isdir);
 378                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
 379         }
 380
 381         if (flags & FMODE_WRITE)
 382                 rc = mdt_write_get(o);
 383         else if (flags & MDS_FMODE_EXEC)
 384                 rc = mdt_write_deny(o);
 385
 386         if (rc)
 387                 RETURN(rc);
 388
 389         rc = mo_open(info->mti_env, mdt_object_child(o),
 390                      created ? flags | MDS_OPEN_CREATED : flags);
 391         if (rc != 0) {
 392                 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
 393                  * client does not know which suppgid should be sent to the MDS,
 394                  * or some other(s) changed the target file's GID after this RPC
 395                  * sent to the MDS with the suppgid as the original GID, then we
 396                  * give the client another chance to send the right suppgid. */
 397                 if (rc == -EACCES &&
 398                     allow_client_chgrp(info, lu_ucred(info->mti_env)))
 399                         mdt_set_disposition(info, rep, DISP_OPEN_DENY);
 400
 401                 GOTO(err_out, rc);
 402         }
 403
 404         mfd = mdt_mfd_new(med);
 405         if (mfd == NULL)
 406                 GOTO(err_out, rc = -ENOMEM);
 407
 408         /*
 409          * Keep a reference on this object for this open, and is
 410          * released by mdt_mfd_close().
 411          */
 412         mdt_object_get(info->mti_env, o);
 413         mfd->mfd_object = o;
 414         mfd->mfd_xid = req->rq_xid;
 415
 416         /*
 417          * @flags is always not zero. At least it should be FMODE_READ,
 418          * FMODE_WRITE or MDS_FMODE_EXEC.
 419          */
 420         LASSERT(flags != 0);
 421
 422         /* Open handling. */
 423         mdt_mfd_set_mode(mfd, flags);
 424
 425         atomic_inc(&o->mot_open_count);
 426         if (flags & MDS_OPEN_LEASE)
 427                 atomic_inc(&o->mot_lease_count);
 428
 429         /* replay handle */
 430         if (req_is_replay(req)) {
 431                 struct mdt_file_data *old_mfd;
 432                 /* Check wheather old cookie already exist in
 433                  * the list, becasue when do recovery, client
 434                  * might be disconnected from server, and
 435                  * restart replay, so there maybe some orphan
 436                  * mfd here, we should remove them */
 437                 LASSERT(info->mti_rr.rr_handle != NULL);
 438                 spin_lock(&med->med_open_lock);
 439                 old_mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle, true);
 440                 if (old_mfd != NULL) {
 441                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 442                                "cookie = %#llx\n", mfd,
 443                                PFID(mdt_object_fid(mfd->mfd_object)),
 444                                info->mti_rr.rr_handle->cookie);
 445                         class_handle_unhash(&old_mfd->mfd_handle);
 446                         list_del_init(&old_mfd->mfd_list);
 447                         spin_unlock(&med->med_open_lock);
 448                         /* no attr update for that close */
 449                         la->la_valid = 0;
 450                         ma->ma_valid |= MA_FLAGS;
 451                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 452                         mdt_mfd_close(info, old_mfd);
 453                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 454                         ma->ma_valid &= ~MA_FLAGS;
 455                 } else {
 456                         spin_unlock(&med->med_open_lock);
 457                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 458                                "cookie = %#llx\n",
 459                                PFID(mdt_object_fid(mfd->mfd_object)),
 460                                info->mti_rr.rr_handle->cookie);
 461                 }
 462
 463                 CDEBUG(D_HA, "Store old cookie %#llx in new mfd\n",
 464                        info->mti_rr.rr_handle->cookie);
 465
 466                 mfd->mfd_old_handle.cookie = info->mti_rr.rr_handle->cookie;
 467         }
 468
 469         repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
 470
 471         if (req->rq_export->exp_disconnected) {
 472                 spin_lock(&med->med_open_lock);
 473                 class_handle_unhash(&mfd->mfd_handle);
 474                 list_del_init(&mfd->mfd_list);
 475                 spin_unlock(&med->med_open_lock);
 476                 mdt_mfd_close(info, mfd);
 477         } else {
 478                 spin_lock(&med->med_open_lock);
 479                 list_add(&mfd->mfd_list, &med->med_open_head);
 480                 spin_unlock(&med->med_open_lock);
 481         }
 482
 483         mdt_empty_transno(info, rc);
 484
 485         RETURN(rc);
 486
 487 err_out:
 488         if (flags & FMODE_WRITE)
 489                 mdt_write_put(o);
 490         else if (flags & MDS_FMODE_EXEC)
 491                 mdt_write_allow(o);
 492
 493         return rc;
 494 }
 495
 496 static int mdt_finish_open(struct mdt_thread_info *info,
 497                            struct mdt_object *p, struct mdt_object *o,
 498                            __u64 flags, int created, struct ldlm_reply *rep)
 499 {
 500         struct ptlrpc_request   *req = mdt_info_req(info);
 501         struct obd_export       *exp = req->rq_export;
 502         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 503         struct md_attr          *ma  = &info->mti_attr;
 504         struct lu_attr          *la  = &ma->ma_attr;
 505         struct mdt_file_data    *mfd;
 506         struct mdt_body         *repbody;
 507         int                      rc = 0;
 508         int                      isreg, isdir, islnk;
 509         struct list_head        *t;
 510         ENTRY;
 511
 512         LASSERT(ma->ma_valid & MA_INODE);
 513
 514         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 515
 516         isreg = S_ISREG(la->la_mode);
 517         isdir = S_ISDIR(la->la_mode);
 518         islnk = S_ISLNK(la->la_mode);
 519         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 520
 521         /* LU-2275, simulate broken behaviour (esp. prevalent in
 522          * pre-2.4 servers where a very strange reply is sent on error
 523          * that looks like it was actually almost successful and a
 524          * failure at the same time.) */
 525         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 526                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 527                                                DISP_LOOKUP_NEG |
 528                                                DISP_LOOKUP_POS);
 529
 530                 if (flags & MDS_OPEN_LOCK)
 531                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 532
 533                 RETURN(-ENOENT);
 534         }
 535
 536 #ifdef CONFIG_FS_POSIX_ACL
 537         if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 538                 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
 539                 if (IS_ERR(nodemap))
 540                         RETURN(PTR_ERR(nodemap));
 541
 542                 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
 543                 nodemap_putref(nodemap);
 544                 if (rc)
 545                         RETURN(rc);
 546         }
 547 #endif
 548
 549         /*
 550          * If we are following a symlink, don't open; and do not return open
 551          * handle for special nodes as client required.
 552          */
 553         if (islnk || (!isreg && !isdir &&
 554             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 555                 lustre_msg_set_transno(req->rq_repmsg, 0);
 556                 RETURN(0);
 557         }
 558
 559         /*
 560          * We need to return the existing object's fid back, so it is done here,
 561          * after preparing the reply.
 562          */
 563         if (!created && (flags & MDS_OPEN_EXCL) && (flags & MDS_OPEN_CREAT))
 564                 RETURN(-EEXIST);
 565
 566         /* This can't be done earlier, we need to return reply body */
 567         if (isdir) {
 568                 if (flags & (MDS_OPEN_CREAT | FMODE_WRITE)) {
 569                         /* We are trying to create or write an existing dir. */
 570                         RETURN(-EISDIR);
 571                 }
 572         } else if (flags & MDS_OPEN_DIRECTORY)
 573                 RETURN(-ENOTDIR);
 574
 575         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 576                                  OBD_FAIL_LDLM_REPLY | OBD_FAIL_ONCE)) {
 577                 RETURN(-EAGAIN);
 578         }
 579
 580         mfd = NULL;
 581         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 582                 spin_lock(&med->med_open_lock);
 583                 list_for_each(t, &med->med_open_head) {
 584                         mfd = list_entry(t, struct mdt_file_data, mfd_list);
 585                         if (mfd->mfd_xid == req->rq_xid)
 586                                 break;
 587                         mfd = NULL;
 588                 }
 589                 spin_unlock(&med->med_open_lock);
 590
 591                 if (mfd != NULL) {
 592                         repbody->mbo_handle.cookie = mfd->mfd_handle.h_cookie;
 593                         /* set repbody->ea_size for resent case */
 594                         if (ma->ma_valid & MA_LOV) {
 595                                 LASSERT(ma->ma_lmm_size != 0);
 596                                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 597                                 if (isdir)
 598                                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 599                                 else
 600                                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 601                         }
 602                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 603                         RETURN(0);
 604                 }
 605         }
 606
 607         rc = mdt_mfd_open(info, p, o, flags, created, rep);
 608         if (!rc)
 609                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 610
 611         RETURN(rc);
 612 }
 613
 614 void mdt_reconstruct_open(struct mdt_thread_info *info,
 615                           struct mdt_lock_handle *lhc)
 616 {
 617         const struct lu_env *env = info->mti_env;
 618         struct mdt_device       *mdt  = info->mti_mdt;
 619         struct req_capsule      *pill = info->mti_pill;
 620         struct ptlrpc_request   *req  = mdt_info_req(info);
 621         struct md_attr          *ma   = &info->mti_attr;
 622         struct mdt_reint_record *rr   = &info->mti_rr;
 623         __u64                   flags = info->mti_spec.sp_cr_flags;
 624         struct ldlm_reply       *ldlm_rep;
 625         struct mdt_object       *parent;
 626         struct mdt_object       *child;
 627         struct mdt_body         *repbody;
 628         int                      rc;
 629         __u64                    opdata;
 630         ENTRY;
 631
 632         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 633         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 634         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 635
 636         ma->ma_need = MA_INODE | MA_HSM;
 637         ma->ma_valid = 0;
 638
 639         opdata = mdt_req_from_lrd(req, info->mti_reply_data);
 640         mdt_set_disposition(info, ldlm_rep, opdata);
 641
 642         CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n",
 643                ldlm_rep->lock_policy_res1, req->rq_status);
 644
 645         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
 646             req->rq_status != 0)
 647                 /* We did not create successfully, return error to client. */
 648                 GOTO(out, rc = req->rq_status);
 649
 650         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
 651                 struct obd_export *exp = req->rq_export;
 652                 /*
 653                  * We failed after creation, but we do not know in which step
 654                  * we failed. So try to check the child object.
 655                  */
 656                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
 657                 if (IS_ERR(parent)) {
 658                         rc = PTR_ERR(parent);
 659                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
 660                                       " Evicting client %s with export %s.\n",
 661                                       PFID(rr->rr_fid1), rc,
 662                                       obd_uuid2str(&exp->exp_client_uuid),
 663                                       obd_export_nid2str(exp));
 664                         mdt_export_evict(exp);
 665                         RETURN_EXIT;
 666                 }
 667
 668                 child = mdt_object_find(env, mdt, rr->rr_fid2);
 669                 if (IS_ERR(child)) {
 670                         rc = PTR_ERR(child);
 671                         LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
 672                                       "evicting client %s with export %s\n",
 673                                       PFID(rr->rr_fid2), rc,
 674                                       obd_uuid2str(&exp->exp_client_uuid),
 675                                       obd_export_nid2str(exp));
 676                         mdt_object_put(env, parent);
 677                         mdt_export_evict(exp);
 678                         RETURN_EXIT;
 679                 }
 680
 681                 if (unlikely(mdt_object_remote(child))) {
 682                         /* the child object was created on remote server */
 683                         if (!mdt_is_dne_client(exp)) {
 684                                 /* Return -EIO for old client */
 685                                 mdt_object_put(env, parent);
 686                                 mdt_object_put(env, child);
 687                                 GOTO(out, rc = -EIO);
 688                         }
 689                         repbody->mbo_fid1 = *rr->rr_fid2;
 690                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 691                         rc = 0;
 692                 } else {
 693                         if (mdt_object_exists(child)) {
 694                                 mdt_prep_ma_buf_from_rep(info, child, ma);
 695                                 rc = mdt_attr_get_complex(info, child, ma);
 696                                 if (rc == 0)
 697                                         rc = mdt_finish_open(info, parent,
 698                                                              child, flags,
 699                                                              1, ldlm_rep);
 700                         } else {
 701                                 /* the child does not exist, we should do
 702                                  * regular open */
 703                                 mdt_object_put(env, parent);
 704                                 mdt_object_put(env, child);
 705                                 GOTO(regular_open, 0);
 706                         }
 707                 }
 708                 mdt_object_put(env, parent);
 709                 mdt_object_put(env, child);
 710                 GOTO(out, rc);
 711         } else {
 712 regular_open:
 713                 /* We did not try to create, so we are a pure open */
 714                 rc = mdt_reint_open(info, lhc);
 715         }
 716
 717         EXIT;
 718 out:
 719         req->rq_status = rc;
 720         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 721         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
 722 }
 723
 724 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
 725 {
 726         __u64                    flags = info->mti_spec.sp_cr_flags;
 727         struct mdt_reint_record *rr = &info->mti_rr;
 728         struct md_attr          *ma = &info->mti_attr;
 729         struct mdt_object       *o;
 730         int                      rc;
 731         ENTRY;
 732
 733         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
 734         if (IS_ERR(o))
 735                 RETURN(rc = PTR_ERR(o));
 736
 737         if (unlikely(mdt_object_remote(o))) {
 738                 /* the child object was created on remote server */
 739                 struct mdt_body *repbody;
 740
 741                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 742                                                 DISP_LOOKUP_EXECD |
 743                                                 DISP_LOOKUP_POS));
 744                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 745                 repbody->mbo_fid1 = *rr->rr_fid2;
 746                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 747                 rc = 0;
 748         } else {
 749                 if (mdt_object_exists(o)) {
 750                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 751                                                         DISP_LOOKUP_EXECD |
 752                                                         DISP_LOOKUP_POS));
 753                         mdt_prep_ma_buf_from_rep(info, o, ma);
 754                         rc = mdt_attr_get_complex(info, o, ma);
 755                         if (rc == 0)
 756                                 rc = mdt_finish_open(info, NULL, o, flags, 0,
 757                                                      rep);
 758                 } else {
 759                         rc = -ENOENT;
 760                 }
 761         }
 762
 763         mdt_object_put(info->mti_env, o);
 764         RETURN(rc);
 765 }
 766
 767 /* lock object for open */
 768 static int mdt_object_open_lock(struct mdt_thread_info *info,
 769                                 struct mdt_object *obj,
 770                                 struct mdt_lock_handle *lhc,
 771                                 __u64 *ibits)
 772 {
 773         struct md_attr *ma = &info->mti_attr;
 774         __u64 open_flags = info->mti_spec.sp_cr_flags;
 775         enum ldlm_mode lm = LCK_CR;
 776         bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
 777         bool try_layout = false;
 778         bool create_layout = false;
 779         int rc = 0;
 780         ENTRY;
 781
 782         *ibits = 0;
 783         mdt_lock_handle_init(lhc);
 784
 785         if (req_is_replay(mdt_info_req(info)))
 786                 RETURN(0);
 787
 788         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
 789                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
 790                     md_should_create(open_flags))
 791                         create_layout = true;
 792                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
 793                     ma->ma_need & MA_LOV)
 794                         try_layout = true;
 795         }
 796
 797         if (acq_lease) {
 798                 /* lease open, acquire write mode of open sem */
 799                 down_write(&obj->mot_open_sem);
 800
 801                 /* Lease exists and ask for new lease */
 802                 if (atomic_read(&obj->mot_lease_count) > 0) {
 803                         /* only exclusive open is supported, so lease
 804                          * are conflicted to each other */
 805                         GOTO(out, rc = -EBUSY);
 806                 }
 807
 808                 /* Lease must be with open lock */
 809                 if (!(open_flags & MDS_OPEN_LOCK)) {
 810                         CERROR("%s: Request lease for file:"DFID ", but open lock "
 811                                "is missed, open_flags = %#llo : rc = %d\n",
 812                                mdt_obd_name(info->mti_mdt),
 813                                PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
 814                         GOTO(out, rc = -EPROTO);
 815                 }
 816
 817                 /* XXX: only exclusive open is supported. */
 818                 lm = LCK_EX;
 819                 *ibits = MDS_INODELOCK_OPEN;
 820
 821                 /* never grant LCK_EX layout lock to client */
 822                 try_layout = false;
 823         } else { /* normal open */
 824                 /* normal open holds read mode of open sem */
 825                 down_read(&obj->mot_open_sem);
 826
 827                 if (open_flags & MDS_OPEN_LOCK) {
 828                         if (open_flags & FMODE_WRITE)
 829                                 lm = LCK_CW;
 830                         else if (open_flags & MDS_FMODE_EXEC)
 831                                 lm = LCK_PR;
 832                         else
 833                                 lm = LCK_CR;
 834
 835                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
 836                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
 837                         if (open_flags & FMODE_WRITE)
 838                                 lm = LCK_CW;
 839                         else
 840                                 lm = LCK_CR;
 841
 842                         /* revoke lease */
 843                         *ibits = MDS_INODELOCK_OPEN;
 844                         try_layout = false;
 845
 846                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 847                 }
 848                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
 849                         PFID(mdt_object_fid(obj)),
 850                         atomic_read(&obj->mot_open_count), lm);
 851         }
 852
 853         mdt_lock_reg_init(lhc, lm);
 854
 855         /* one problem to return layout lock on open is that it may result
 856          * in too many layout locks cached on the client side. */
 857         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
 858                 /* return lookup lock to validate inode at the client side,
 859                  * this is pretty important otherwise mdt will return layout
 860                  * lock for each open.
 861                  * However this is a double-edged sword because changing
 862                  * permission will revoke huge # of LOOKUP locks. */
 863                 *ibits |= MDS_INODELOCK_LAYOUT | MDS_INODELOCK_LOOKUP;
 864                 if (!mdt_object_lock_try(info, obj, lhc, *ibits)) {
 865                         *ibits &= ~(MDS_INODELOCK_LAYOUT|MDS_INODELOCK_LOOKUP);
 866                         if (*ibits != 0)
 867                                 rc = mdt_object_lock(info, obj, lhc, *ibits);
 868                 }
 869         } else if (*ibits != 0) {
 870                 rc = mdt_object_lock(info, obj, lhc, *ibits);
 871         }
 872
 873         CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx"
 874                ", open_flags = %#llo, try_layout = %d : rc = %d\n",
 875                mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
 876                *ibits, open_flags, try_layout, rc);
 877
 878         /* will change layout, revoke layout locks by enqueuing EX lock. */
 879         if (rc == 0 && create_layout) {
 880                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
 881
 882                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
 883                         ", open_flags = %#llo\n",
 884                         PFID(mdt_object_fid(obj)), open_flags);
 885
 886                 /* We cannot enqueue another lock for the same resource we
 887                  * already have a lock for, due to mechanics of waiting list
 888                  * iterating in ldlm, see LU-3601.
 889                  * As such we'll drop the open lock we just got above here,
 890                  * it's ok not to have this open lock as it's main purpose is to
 891                  * flush unused cached client open handles. */
 892                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 893                         mdt_object_unlock(info, obj, lhc, 1);
 894
 895                 LASSERT(!try_layout);
 896                 mdt_lock_handle_init(ll);
 897                 mdt_lock_reg_init(ll, LCK_EX);
 898                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
 899
 900                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
 901         }
 902
 903         /* Check if there is any other open handles after acquiring
 904          * open lock. At this point, caching open handles have been revoked
 905          * by open lock.
 906          * XXX: Now only exclusive open is supported. Need to check the
 907          * type of open for generic lease support. */
 908         if (rc == 0 && acq_lease) {
 909                 struct ptlrpc_request *req = mdt_info_req(info);
 910                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 911                 struct mdt_file_data *mfd;
 912                 bool is_replay_or_resent;
 913                 int open_count = 0;
 914
 915                 /* For lease: application can open a file and then apply lease,
 916                  * @handle contains original open handle in that case.
 917                  * In recovery, open REQ will be replayed and the lease REQ may
 918                  * be resent that means the open handle is already stale, so we
 919                  * need to fix it up here by finding new handle. */
 920                 is_replay_or_resent = req_is_replay(req) ||
 921                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
 922
 923                 /* if the request is _not_ a replay request, rr_handle
 924                  * may be used to hold an openhandle which is issuing the
 925                  * lease request, so that this openhandle doesn't count. */
 926                 mfd = mdt_handle2mfd(med, info->mti_rr.rr_handle,
 927                                      is_replay_or_resent);
 928                 if (mfd != NULL)
 929                         ++open_count;
 930
 931                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
 932                         PFID(mdt_object_fid(obj)),
 933                         atomic_read(&obj->mot_open_count), open_count);
 934
 935                 if (atomic_read(&obj->mot_open_count) > open_count)
 936                         GOTO(out, rc = -EBUSY);
 937         }
 938         GOTO(out, rc);
 939
 940 out:
 941         RETURN(rc);
 942 }
 943
 944 static void mdt_object_open_unlock(struct mdt_thread_info *info,
 945                                    struct mdt_object *obj,
 946                                    struct mdt_lock_handle *lhc,
 947                                    __u64 ibits, int rc)
 948 {
 949         __u64 open_flags = info->mti_spec.sp_cr_flags;
 950         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
 951         ENTRY;
 952
 953         if (req_is_replay(mdt_info_req(info)))
 954                 RETURN_EXIT;
 955
 956         /* Release local lock - the lock put in MDT_LH_LOCAL will never
 957          * return to client side. */
 958         if (lustre_handle_is_used(&ll->mlh_reg_lh))
 959                 mdt_object_unlock(info, obj, ll, 1);
 960
 961         ll = &info->mti_lh[MDT_LH_LAYOUT];
 962         /* Release local layout lock, layout was created */
 963         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
 964                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
 965                 mdt_object_unlock(info, obj, ll, 1);
 966         }
 967
 968         if (open_flags & MDS_OPEN_LEASE)
 969                 up_write(&obj->mot_open_sem);
 970         else
 971                 up_read(&obj->mot_open_sem);
 972
 973         /* Cross-ref case, the lock should be returned to the client */
 974         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
 975                 RETURN_EXIT;
 976
 977         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT)) {
 978                 /* for the open request, the lock will only return to client
 979                  * if open or layout lock is granted. */
 980                 rc = 1;
 981         }
 982
 983         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
 984                 struct ldlm_reply       *ldlm_rep;
 985
 986                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
 987                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
 988                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 989                         mdt_object_unlock(info, obj, lhc, 1);
 990         }
 991         RETURN_EXIT;
 992 }
 993
 994 /**
 995  * Check release is permitted for the current HSM flags.
 996  */
 997 static bool mdt_hsm_release_allow(const struct md_attr *ma)
 998 {
 999         if (!(ma->ma_valid & MA_HSM))
1000                 return false;
1001
1002         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1003                 return false;
1004
1005         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1006                 return false;
1007
1008         return true;
1009 }
1010
1011 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1012                                 struct ldlm_reply *rep,
1013                                 struct mdt_lock_handle *lhc)
1014 {
1015         const struct lu_env     *env   = info->mti_env;
1016         struct mdt_device       *mdt   = info->mti_mdt;
1017         __u64                    flags = info->mti_spec.sp_cr_flags;
1018         struct mdt_reint_record *rr    = &info->mti_rr;
1019         struct md_attr          *ma    = &info->mti_attr;
1020         struct mdt_object       *parent= NULL;
1021         struct mdt_object       *o;
1022         int                      rc;
1023         bool                     object_locked = false;
1024         __u64                    ibits = 0;
1025         ENTRY;
1026
1027         if (md_should_create(flags) && !(flags & MDS_OPEN_HAS_EA)) {
1028                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1029                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1030                         if (IS_ERR(parent)) {
1031                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1032                                        " for anonymous created %ld, try to"
1033                                        " use server-side parent.\n",
1034                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1035                                 parent = NULL;
1036                         }
1037                 }
1038                 if (parent == NULL)
1039                         ma->ma_need |= MA_PFID;
1040         }
1041
1042         o = mdt_object_find(env, mdt, rr->rr_fid2);
1043         if (IS_ERR(o))
1044                 GOTO(out_parent_put, rc = PTR_ERR(o));
1045
1046         if (mdt_object_remote(o)) {
1047                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1048                        mdt_obd_name(info->mti_mdt),
1049                        PFID(rr->rr_fid2));
1050                 GOTO(out, rc = -EREMOTE);
1051         } else if (!mdt_object_exists(o)) {
1052                 mdt_set_disposition(info, rep,
1053                                     DISP_IT_EXECD |
1054                                     DISP_LOOKUP_EXECD |
1055                                     DISP_LOOKUP_NEG);
1056                 GOTO(out, rc = -ENOENT);
1057         }
1058
1059         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1060
1061         mdt_prep_ma_buf_from_rep(info, o, ma);
1062         if (flags & MDS_OPEN_RELEASE)
1063                 ma->ma_need |= MA_HSM;
1064         rc = mdt_attr_get_complex(info, o, ma);
1065         if (rc)
1066                 GOTO(out, rc);
1067
1068         /* We should not change file's existing LOV EA */
1069         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1070             flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1071                 GOTO(out, rc = -EEXIST);
1072
1073         /* If a release request, check file flags are fine and ask for an
1074          * exclusive open access. */
1075         if (flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1076                 GOTO(out, rc = -EPERM);
1077
1078         rc = mdt_check_resent_lock(info, o, lhc);
1079         if (rc < 0) {
1080                 GOTO(out, rc);
1081         } else if (rc > 0) {
1082                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1083                 object_locked = true;
1084                 if (rc)
1085                         GOTO(out_unlock, rc);
1086         }
1087
1088         if (ma->ma_valid & MA_PFID) {
1089                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1090                 if (IS_ERR(parent)) {
1091                         CDEBUG(D_INODE, "Fail to find parent "DFID
1092                                " for anonymous created %ld, try to"
1093                                " use system default.\n",
1094                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1095                         parent = NULL;
1096                 }
1097         }
1098
1099         rc = mdt_finish_open(info, parent, o, flags, 0, rep);
1100         if (!rc) {
1101                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1102                 if (flags & MDS_OPEN_LOCK)
1103                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1104                 if (flags & MDS_OPEN_LEASE)
1105                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1106         }
1107         GOTO(out_unlock, rc);
1108
1109 out_unlock:
1110         if (object_locked)
1111                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1112 out:
1113         mdt_object_put(env, o);
1114 out_parent_put:
1115         if (parent != NULL)
1116                 mdt_object_put(env, parent);
1117         return rc;
1118 }
1119
1120 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1121 static int mdt_cross_open(struct mdt_thread_info *info,
1122                           const struct lu_fid *parent_fid,
1123                           const struct lu_fid *fid,
1124                           struct ldlm_reply *rep, __u32 flags)
1125 {
1126         struct md_attr    *ma = &info->mti_attr;
1127         struct mdt_object *o;
1128         int                rc;
1129         ENTRY;
1130
1131         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1132         if (IS_ERR(o))
1133                 RETURN(rc = PTR_ERR(o));
1134
1135         if (mdt_object_remote(o)) {
1136                 /* Something is wrong here, the object is on another MDS! */
1137                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1138                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1139                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1140                                 &o->mot_obj,
1141                                 "Object isn't on this server! FLD error?");
1142                 rc = -EFAULT;
1143         } else {
1144                 if (mdt_object_exists(o)) {
1145                         /* Do permission check for cross-open. */
1146                         rc = mo_permission(info->mti_env, NULL,
1147                                            mdt_object_child(o),
1148                                            NULL, flags | MDS_OPEN_CROSS);
1149                         if (rc)
1150                                 goto out;
1151
1152                         mdt_prep_ma_buf_from_rep(info, o, ma);
1153                         rc = mdt_attr_get_complex(info, o, ma);
1154                         if (rc != 0)
1155                                 GOTO(out, rc);
1156
1157                         rc = mdt_finish_open(info, NULL, o, flags, 0, rep);
1158                 } else {
1159                         /*
1160                          * Something is wrong here. lookup was positive but
1161                          * there is no object!
1162                          */
1163                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1164                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1165                         rc = -EFAULT;
1166                 }
1167         }
1168 out:
1169         mdt_object_put(info->mti_env, o);
1170         RETURN(rc);
1171 }
1172
1173 /*
1174  * find root object and take its xattr lock if it's on remote MDT, later create
1175  * may use fs default striping (which is stored in root xattr).
1176  */
1177 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1178                                struct mdt_device *mdt)
1179 {
1180         struct mdt_object *md_root = mdt->mdt_md_root;
1181         struct lustre_handle lhroot;
1182         int rc;
1183
1184         if (md_root == NULL) {
1185                 lu_root_fid(&info->mti_tmp_fid1);
1186                 md_root = mdt_object_find(info->mti_env, mdt,
1187                                           &info->mti_tmp_fid1);
1188                 if (IS_ERR(md_root))
1189                         return PTR_ERR(md_root);
1190
1191                 spin_lock(&mdt->mdt_lock);
1192                 if (mdt->mdt_md_root != NULL) {
1193                         spin_unlock(&mdt->mdt_lock);
1194
1195                         LASSERTF(mdt->mdt_md_root == md_root,
1196                                  "Different root object ("
1197                                  DFID") instances, %p, %p\n",
1198                                  PFID(&info->mti_tmp_fid1),
1199                                  mdt->mdt_md_root, md_root);
1200                         LASSERT(atomic_read(
1201                                 &md_root->mot_obj.lo_header->loh_ref) > 1);
1202
1203                         mdt_object_put(info->mti_env, md_root);
1204                 } else {
1205                         mdt->mdt_md_root = md_root;
1206                         spin_unlock(&mdt->mdt_lock);
1207                 }
1208         }
1209
1210         if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1211                 return 0;
1212
1213         rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root),
1214                                     &lhroot, LCK_PR, MDS_INODELOCK_XATTR,
1215                                     false, true);
1216         if (rc < 0)
1217                 return rc;
1218
1219         md_root->mot_cache_attr = 1;
1220
1221         /* don't cancel this lock, so that we know the cached xattr is valid. */
1222         ldlm_lock_decref(&lhroot, LCK_PR);
1223
1224         return 0;
1225 }
1226
1227 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1228 {
1229         struct mdt_device       *mdt = info->mti_mdt;
1230         struct ptlrpc_request   *req = mdt_info_req(info);
1231         struct mdt_object       *parent;
1232         struct mdt_object       *child;
1233         struct mdt_lock_handle  *lh;
1234         struct ldlm_reply       *ldlm_rep;
1235         struct mdt_body         *repbody;
1236         struct lu_fid           *child_fid = &info->mti_tmp_fid1;
1237         struct md_attr          *ma = &info->mti_attr;
1238         __u64                    create_flags = info->mti_spec.sp_cr_flags;
1239         __u64                    ibits = 0;
1240         struct mdt_reint_record *rr = &info->mti_rr;
1241         int                      result, rc;
1242         int                      created = 0;
1243         int                      object_locked = 0;
1244         __u32                    msg_flags;
1245         ENTRY;
1246
1247         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1248                                (obd_timeout + 1) / 4);
1249
1250         mdt_counter_incr(req, LPROC_MDT_OPEN);
1251         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1252
1253         ma->ma_need = MA_INODE;
1254         ma->ma_valid = 0;
1255
1256         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1257         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1258
1259         if (unlikely(create_flags & MDS_OPEN_JOIN_FILE)) {
1260                 CERROR("file join is not supported anymore.\n");
1261                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1262         }
1263         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1264
1265         if ((create_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1266             info->mti_spec.u.sp_ea.eadata == NULL)
1267                 GOTO(out, result = err_serious(-EINVAL));
1268
1269         if (create_flags & FMODE_WRITE &&
1270             exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1271                 GOTO(out, result = -EROFS);
1272
1273         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1274                "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1275                PFID(rr->rr_fid1), PNAME(&rr->rr_name),
1276                PFID(rr->rr_fid2), create_flags,
1277                ma->ma_attr.la_mode, msg_flags);
1278
1279         if (info->mti_cross_ref) {
1280                 /* This is cross-ref open */
1281                 mdt_set_disposition(info, ldlm_rep,
1282                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1283                              DISP_LOOKUP_POS));
1284                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1285                                         ldlm_rep, create_flags);
1286                 GOTO(out, result);
1287         } else if (req_is_replay(req) ||
1288             (req->rq_export->exp_libclient && create_flags & MDS_OPEN_HAS_EA)) {
1289                 /* This is a replay request or from liblustre with ea. */
1290                 result = mdt_open_by_fid(info, ldlm_rep);
1291
1292                 if (result != -ENOENT) {
1293                         if (req->rq_export->exp_libclient &&
1294                             create_flags & MDS_OPEN_HAS_EA)
1295                                 GOTO(out, result = 0);
1296                         GOTO(out, result);
1297                 }
1298                 /* We didn't find the correct object, so we need to re-create it
1299                  * via a regular replay. */
1300                 if (!(create_flags & MDS_OPEN_CREAT)) {
1301                         DEBUG_REQ(D_ERROR, req,
1302                                   "OPEN & CREAT not in open replay/by_fid.");
1303                         GOTO(out, result = -EFAULT);
1304                 }
1305                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1306         } else if (create_flags & (MDS_OPEN_BY_FID | MDS_OPEN_LOCK)) {
1307                 /*
1308                  * MDS_OPEN_LOCK is checked for backward compatibility with 2.1
1309                  * client.
1310                  */
1311                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1312                 if (result < 0)
1313                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1314                                PFID(rr->rr_fid2), result);
1315                 GOTO(out, result);
1316         }
1317
1318         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1319                 GOTO(out, result = err_serious(-ENOMEM));
1320
1321         mdt_set_disposition(info, ldlm_rep,
1322                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1323
1324         if (!lu_name_is_valid(&rr->rr_name))
1325                 GOTO(out, result = -EPROTO);
1326
1327         result = mdt_lock_root_xattr(info, mdt);
1328         if (result < 0)
1329                 GOTO(out, result);
1330
1331 again:
1332         lh = &info->mti_lh[MDT_LH_PARENT];
1333         mdt_lock_pdo_init(lh,
1334                           (create_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1335                           &rr->rr_name);
1336
1337         parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1338         if (IS_ERR(parent))
1339                 GOTO(out, result = PTR_ERR(parent));
1340
1341         result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1342         if (result != 0) {
1343                 mdt_object_put(info->mti_env, parent);
1344                 GOTO(out, result);
1345         }
1346
1347         /* get and check version of parent */
1348         result = mdt_version_get_check(info, parent, 0);
1349         if (result)
1350                 GOTO(out_parent, result);
1351
1352         fid_zero(child_fid);
1353
1354         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1355                             &rr->rr_name, child_fid, &info->mti_spec);
1356
1357         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1358                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1359                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1360                  PFID(child_fid));
1361
1362         if (result != 0 && result != -ENOENT && result != -ESTALE)
1363                 GOTO(out_parent, result);
1364
1365         if (result == -ENOENT || result == -ESTALE) {
1366                 /* If the object is dead, let's check if the object
1367                  * is being migrated to a new object */
1368                 if (result == -ESTALE) {
1369                         struct lu_buf lmv_buf;
1370
1371                         lmv_buf.lb_buf = info->mti_xattr_buf;
1372                         lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
1373                         rc = mo_xattr_get(info->mti_env,
1374                                           mdt_object_child(parent),
1375                                           &lmv_buf, XATTR_NAME_LMV);
1376                         if (rc > 0) {
1377                                 struct lmv_mds_md_v1 *lmv;
1378
1379                                 lmv = lmv_buf.lb_buf;
1380                                 if (le32_to_cpu(lmv->lmv_hash_type) &
1381                                                 LMV_HASH_FLAG_MIGRATION) {
1382                                         /* Get the new parent FID and retry */
1383                                         mdt_object_unlock_put(info, parent,
1384                                                               lh, 1);
1385                                         mdt_lock_handle_init(lh);
1386                                         fid_le_to_cpu(
1387                                                 (struct lu_fid *)rr->rr_fid1,
1388                                                 &lmv->lmv_stripe_fids[1]);
1389                                         goto again;
1390                                 }
1391                         }
1392                 }
1393
1394                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1395                 if (result == -ESTALE) {
1396                         /*
1397                          * -ESTALE means the parent is a dead(unlinked) dir, so
1398                          * it should return -ENOENT to in accordance with the
1399                          * original mds implementaion.
1400                          */
1401                         GOTO(out_parent, result = -ENOENT);
1402                 }
1403
1404                 if (!(create_flags & MDS_OPEN_CREAT))
1405                         GOTO(out_parent, result);
1406                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1407                         GOTO(out_parent, result = -EROFS);
1408                 *child_fid = *info->mti_rr.rr_fid2;
1409                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1410                          PFID(child_fid));
1411                 /* In the function below, .hs_keycmp resolves to
1412                  * lu_obj_hop_keycmp() */
1413                 /* coverity[overrun-buffer-val] */
1414                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1415         } else {
1416                 /*
1417                  * Check for O_EXCL is moved to the mdt_finish_open(), we need to
1418                  * return FID back in that case.
1419                  */
1420                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1421                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1422         }
1423         if (IS_ERR(child))
1424                 GOTO(out_parent, result = PTR_ERR(child));
1425
1426         /** check version of child  */
1427         rc = mdt_version_get_check(info, child, 1);
1428         if (rc)
1429                 GOTO(out_child, result = rc);
1430
1431         if (result == -ENOENT) {
1432                 /* Create under OBF and .lustre is not permitted */
1433                 if (!fid_is_md_operative(rr->rr_fid1))
1434                         GOTO(out_child, result = -EPERM);
1435
1436                 /* save versions in reply */
1437                 mdt_version_get_save(info, parent, 0);
1438                 mdt_version_get_save(info, child, 1);
1439
1440                 /* version of child will be changed */
1441                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1442
1443                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1444                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1445
1446                 /* Let lower layers know what is lock mode on directory. */
1447                 info->mti_spec.sp_cr_mode =
1448                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1449
1450                 /*
1451                  * Do not perform lookup sanity check. We know that name does
1452                  * not exist.
1453                  */
1454                 info->mti_spec.sp_cr_lookup = 0;
1455                 info->mti_spec.sp_feat = &dt_directory_features;
1456
1457                 result = mdo_create(info->mti_env,
1458                                     mdt_object_child(parent),
1459                                     &rr->rr_name,
1460                                     mdt_object_child(child),
1461                                     &info->mti_spec,
1462                                     &info->mti_attr);
1463                 if (result == -ERESTART) {
1464                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1465                         GOTO(out_child, result);
1466                 } else {
1467                         mdt_prep_ma_buf_from_rep(info, child, ma);
1468                         /* XXX: we should call this once, see few lines below */
1469                         if (result == 0)
1470                                 result = mdt_attr_get_complex(info, child, ma);
1471
1472                         if (result != 0)
1473                                 GOTO(out_child, result);
1474                 }
1475                 created = 1;
1476                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
1477         } else {
1478                 /*
1479                  * The object is on remote node, return its FID for remote open.
1480                  */
1481                 if (mdt_object_remote(child)) {
1482                         /*
1483                          * Check if this lock already was sent to client and
1484                          * this is resent case. For resent case do not take lock
1485                          * again, use what is already granted.
1486                          */
1487                         LASSERT(lhc != NULL);
1488
1489                         rc = mdt_check_resent_lock(info, child, lhc);
1490                         if (rc < 0) {
1491                                 GOTO(out_child, result = rc);
1492                         } else if (rc > 0) {
1493                                 mdt_lock_handle_init(lhc);
1494                                 mdt_lock_reg_init(lhc, LCK_PR);
1495
1496                                 rc = mdt_object_lock(info, child, lhc,
1497                                                      MDS_INODELOCK_LOOKUP);
1498                         }
1499                         repbody->mbo_fid1 = *mdt_object_fid(child);
1500                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1501                         if (rc != 0)
1502                                 result = rc;
1503                         else
1504                                 result = -MDT_EREMOTE_OPEN;
1505                         GOTO(out_child, result);
1506                 } else if (mdt_object_exists(child)) {
1507                         /* We have to get attr & LOV EA & HSM for this
1508                          * object. */
1509                         mdt_prep_ma_buf_from_rep(info, child, ma);
1510                         ma->ma_need |= MA_HSM;
1511                         result = mdt_attr_get_complex(info, child, ma);
1512                         if (result != 0)
1513                                 GOTO(out_child, result);
1514                 } else {
1515                         /* Object does not exist. Likely FS corruption. */
1516                         CERROR("%s: name '"DNAME"' present, but FID "
1517                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1518                                PNAME(&rr->rr_name), PFID(child_fid));
1519                         GOTO(out_child, result = -EIO);
1520                 }
1521         }
1522
1523         rc = mdt_check_resent_lock(info, child, lhc);
1524         if (rc < 0) {
1525                 GOTO(out_child, result = rc);
1526         } else if (rc == 0) {
1527                 /* the open lock might already be gotten in
1528                  * ldlm_handle_enqueue() */
1529                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1530                 if (create_flags & MDS_OPEN_LOCK)
1531                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1532         } else {
1533                 /* get openlock if this isn't replay and client requested it */
1534                 if (!req_is_replay(req)) {
1535                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1536                         object_locked = 1;
1537                         if (rc != 0)
1538                                 GOTO(out_child_unlock, result = rc);
1539                         else if (create_flags & MDS_OPEN_LOCK)
1540                                 mdt_set_disposition(info, ldlm_rep,
1541                                                     DISP_OPEN_LOCK);
1542                 }
1543         }
1544         /* Try to open it now. */
1545         rc = mdt_finish_open(info, parent, child, create_flags,
1546                              created, ldlm_rep);
1547         if (rc) {
1548                 result = rc;
1549                 /* openlock will be released if mdt_finish_open failed */
1550                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1551
1552                 if (created && create_flags & MDS_OPEN_VOLATILE) {
1553                         CERROR("%s: cannot open volatile file "DFID", orphan "
1554                                "file will be left in PENDING directory until "
1555                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1556                                PFID(mdt_object_fid(child)), rc);
1557                         GOTO(out_child_unlock, result);
1558                 }
1559
1560                 if (created) {
1561                         ma->ma_need = 0;
1562                         ma->ma_valid = 0;
1563                         rc = mdo_unlink(info->mti_env,
1564                                         mdt_object_child(parent),
1565                                         mdt_object_child(child),
1566                                         &rr->rr_name,
1567                                         &info->mti_attr, 0);
1568                         if (rc != 0)
1569                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1570                                        mdt_obd_name(info->mti_mdt),
1571                                        PFID(mdt_object_fid(child)), rc);
1572                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1573                 }
1574         }
1575         EXIT;
1576 out_child_unlock:
1577         if (object_locked)
1578                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1579 out_child:
1580         mdt_object_put(info->mti_env, child);
1581 out_parent:
1582         mdt_object_unlock_put(info, parent, lh, result || !created);
1583 out:
1584         if (result)
1585                 lustre_msg_set_transno(req->rq_repmsg, 0);
1586         return result;
1587 }
1588
1589 /**
1590  * Create an orphan object use local root.
1591  */
1592 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1593                                           struct mdt_device *mdt,
1594                                           const struct lu_fid *fid,
1595                                           struct md_attr *attr, fmode_t fmode)
1596 {
1597         const struct lu_env *env = info->mti_env;
1598         struct md_op_spec *spec = &info->mti_spec;
1599         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1600         struct mdt_object *obj = NULL;
1601         struct mdt_object *local_root;
1602         static const struct lu_name lname = {
1603                 .ln_name = "i_am_nobody",
1604                 .ln_namelen = sizeof("i_am_nobody") - 1,
1605         };
1606         struct lu_ucred *uc;
1607         cfs_cap_t uc_cap_save;
1608         int rc;
1609         ENTRY;
1610
1611         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1612         if (rc != 0)
1613                 RETURN(ERR_PTR(rc));
1614
1615         local_root = mdt_object_find(env, mdt, local_root_fid);
1616         if (IS_ERR(local_root))
1617                 RETURN(local_root);
1618
1619         obj = mdt_object_new(env, mdt, fid);
1620         if (IS_ERR(obj))
1621                 GOTO(out, rc = PTR_ERR(obj));
1622
1623         spec->sp_cr_lookup = 0;
1624         spec->sp_feat = &dt_directory_features;
1625         spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1626         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1627         if (attr->ma_valid & MA_LOV) {
1628                 spec->u.sp_ea.eadata = attr->ma_lmm;
1629                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1630                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1631         } else {
1632                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1633         }
1634
1635         uc = lu_ucred(env);
1636         uc_cap_save = uc->uc_cap;
1637         uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1638         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1639                         mdt_object_child(obj), spec, attr);
1640         uc->uc_cap = uc_cap_save;
1641         if (rc < 0) {
1642                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1643                        mdt_obd_name(mdt), PFID(fid), rc);
1644                 GOTO(out, rc);
1645         }
1646
1647         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1648         if (rc < 0)
1649                 CERROR("%s: cannot open volatile file "DFID", orphan "
1650                        "file will be left in PENDING directory until "
1651                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1652                        PFID(fid), rc);
1653         GOTO(out, rc);
1654
1655 out:
1656         if (rc < 0) {
1657                 if (!IS_ERR(obj))
1658                         mdt_object_put(env, obj);
1659                 obj = ERR_PTR(rc);
1660         }
1661         mdt_object_put(env, local_root);
1662         return obj;
1663 }
1664
1665 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1666                            struct md_attr *ma)
1667 {
1668         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1669         struct lu_ucred        *uc = mdt_ucred(info);
1670         struct close_data      *data;
1671         struct ldlm_lock       *lease;
1672         struct mdt_object      *orphan;
1673         struct md_attr         *orp_ma;
1674         struct lu_buf          *buf;
1675         cfs_cap_t               cap;
1676         bool                    lease_broken;
1677         int                     rc;
1678         int                     rc2;
1679         ENTRY;
1680
1681         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1682                 RETURN(-EROFS);
1683
1684         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1685         if (data == NULL)
1686                 RETURN(-EPROTO);
1687
1688         lease = ldlm_handle2lock(&data->cd_handle);
1689         if (lease == NULL)
1690                 RETURN(-ESTALE);
1691
1692         /* try to hold open_sem so that nobody else can open the file */
1693         if (!down_write_trylock(&o->mot_open_sem)) {
1694                 ldlm_lock_cancel(lease);
1695                 GOTO(out_reprocess, rc = -EBUSY);
1696         }
1697
1698         /* Check if the lease open lease has already canceled */
1699         lock_res_and_lock(lease);
1700         lease_broken = ldlm_is_cancel(lease);
1701         unlock_res_and_lock(lease);
1702
1703         LDLM_DEBUG(lease, DFID " lease broken? %d",
1704                    PFID(mdt_object_fid(o)), lease_broken);
1705
1706         /* Cancel server side lease. Client side counterpart should
1707          * have been cancelled. It's okay to cancel it now as we've
1708          * held mot_open_sem. */
1709         ldlm_lock_cancel(lease);
1710
1711         if (lease_broken) /* don't perform release task */
1712                 GOTO(out_unlock, rc = -ESTALE);
1713
1714         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1715                 GOTO(out_unlock, rc = -EINVAL);
1716
1717         /* ma_need was set before but it seems fine to change it in order to
1718          * avoid modifying the one from RPC */
1719         ma->ma_need = MA_HSM;
1720         rc = mdt_attr_get_complex(info, o, ma);
1721         if (rc != 0)
1722                 GOTO(out_unlock, rc);
1723
1724         if (!mdt_hsm_release_allow(ma))
1725                 GOTO(out_unlock, rc = -EPERM);
1726
1727         /* already released? */
1728         if (ma->ma_hsm.mh_flags & HS_RELEASED)
1729                 GOTO(out_unlock, rc = 0);
1730
1731         /* Compare on-disk and packed data_version */
1732         if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1733                 CDEBUG(D_HSM, DFID" data_version mismatches: packed=%llu"
1734                        " and on-disk=%llu\n", PFID(mdt_object_fid(o)),
1735                        data->cd_data_version, ma->ma_hsm.mh_arch_ver);
1736                 GOTO(out_unlock, rc = -EPERM);
1737         }
1738
1739         ma->ma_valid = MA_INODE;
1740         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1741         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1742         if (rc < 0)
1743                 GOTO(out_unlock, rc);
1744
1745         ma->ma_need = MA_INODE | MA_LOV;
1746         rc = mdt_attr_get_complex(info, o, ma);
1747         if (rc < 0)
1748                 GOTO(out_unlock, rc);
1749
1750         if (!(ma->ma_valid & MA_LOV)) {
1751                 /* Even empty file are released */
1752                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1753                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
1754                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1755                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1756                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1757         } else {
1758                 /* Magic must be LOV_MAGIC_Vx_DEF otherwise LOD will interpret
1759                  * ma_lmm as lov_user_md, then it will be confused by union of
1760                  * layout_gen and stripe_offset. */
1761                 if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V1)
1762                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEF);
1763                 else if (le32_to_cpu(ma->ma_lmm->lmm_magic) == LOV_MAGIC_V3)
1764                         ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V3_DEF);
1765                 else
1766                         GOTO(out_unlock, rc = -EINVAL);
1767         }
1768
1769         /* Set file as released */
1770         ma->ma_lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1771
1772         orp_ma = &info->mti_u.hsm.attr;
1773         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1774         /* We use root ownership to bypass potential quota
1775          * restrictions on the user and group of the file. */
1776         orp_ma->ma_attr.la_uid = 0;
1777         orp_ma->ma_attr.la_gid = 0;
1778         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1779         orp_ma->ma_lmm = ma->ma_lmm;
1780         orp_ma->ma_lmm_size = ma->ma_lmm_size;
1781         orp_ma->ma_valid = MA_INODE | MA_LOV;
1782         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1783                                  FMODE_WRITE);
1784         if (IS_ERR(orphan)) {
1785                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1786                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1787                        PTR_ERR(orphan));
1788                 GOTO(out_unlock, rc = PTR_ERR(orphan));
1789         }
1790
1791         /* Set up HSM attribute for orphan object */
1792         CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
1793         buf = &info->mti_buf;
1794         buf->lb_buf = info->mti_xattr_buf;
1795         buf->lb_len = sizeof(struct hsm_attrs);
1796         ma->ma_hsm.mh_flags |= HS_RELEASED;
1797         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1798         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
1799
1800         mdt_lock_reg_init(lh, LCK_EX);
1801         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
1802                              MDS_INODELOCK_XATTR);
1803         if (rc != 0)
1804                 GOTO(out_close, rc);
1805
1806         /* The orphan has root ownership so we need to raise
1807          * CAP_FOWNER to set the HSM attributes. */
1808         cap = uc->uc_cap;
1809         uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
1810         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
1811                           XATTR_NAME_HSM, 0);
1812         uc->uc_cap = cap;
1813         if (rc != 0)
1814                 GOTO(out_layout_lock, rc);
1815
1816         /* Swap layout with orphan objects. */
1817         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
1818                              mdt_object_child(orphan),
1819                              SWAP_LAYOUTS_MDS_HSM);
1820         EXIT;
1821
1822 out_layout_lock:
1823         /* Release exclusive LL */
1824         mdt_object_unlock(info, o, lh, 1);
1825 out_close:
1826         /* Close orphan object anyway */
1827         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
1828                        FMODE_WRITE);
1829         if (rc2 < 0)
1830                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
1831                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
1832         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
1833                         "object closed");
1834         mdt_object_put(info->mti_env, orphan);
1835
1836 out_unlock:
1837         up_write(&o->mot_open_sem);
1838
1839         /* already released */
1840         if (rc == 0) {
1841                 struct mdt_body *repbody;
1842
1843                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1844                 LASSERT(repbody != NULL);
1845                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1846         }
1847
1848 out_reprocess:
1849         ldlm_reprocess_all(lease->l_resource);
1850         LDLM_LOCK_PUT(lease);
1851
1852         ma->ma_valid = 0;
1853         ma->ma_need = 0;
1854
1855         return rc;
1856 }
1857
1858 int mdt_close_swap_layouts(struct mdt_thread_info *info,
1859                            struct mdt_object *o, struct md_attr *ma)
1860 {
1861         struct mdt_lock_handle  *lh1 = &info->mti_lh[MDT_LH_NEW];
1862         struct mdt_lock_handle  *lh2 = &info->mti_lh[MDT_LH_OLD];
1863         struct close_data       *data;
1864         struct ldlm_lock        *lease;
1865         struct mdt_object       *o1 = o, *o2;
1866         bool                     lease_broken;
1867         bool                     swap_objects;
1868         int                      rc;
1869         ENTRY;
1870
1871         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1872                 RETURN(-EROFS);
1873
1874         if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
1875                 RETURN(-EINVAL);
1876
1877         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1878         if (data == NULL)
1879                 RETURN(-EPROTO);
1880
1881         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1882                 RETURN(-EINVAL);
1883
1884         rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
1885         if (unlikely(rc == 0))
1886                 RETURN(-EINVAL);
1887
1888         /* Exchange o1 and o2, to enforce locking order */
1889         swap_objects = (rc < 0);
1890
1891         lease = ldlm_handle2lock(&data->cd_handle);
1892         if (lease == NULL)
1893                 RETURN(-ESTALE);
1894
1895         o2 = mdt_object_find(info->mti_env, info->mti_mdt, &data->cd_fid);
1896         if (IS_ERR(o2))
1897                 GOTO(out_lease, rc = PTR_ERR(o2));
1898
1899         if (!S_ISREG(lu_object_attr(&o2->mot_obj))) {
1900                 swap_objects = false; /* not swapped yet */
1901                 GOTO(out_obj, rc = -EINVAL);
1902         }
1903
1904         if (swap_objects)
1905                 swap(o1, o2);
1906
1907         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1908                            MAY_WRITE);
1909         if (rc < 0)
1910                 GOTO(out_obj, rc);
1911
1912         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
1913                            MAY_WRITE);
1914         if (rc < 0)
1915                 GOTO(out_obj, rc);
1916
1917         /* try to hold open_sem so that nobody else can open the file */
1918         if (!down_write_trylock(&o->mot_open_sem)) {
1919                 ldlm_lock_cancel(lease);
1920                 GOTO(out_obj, rc = -EBUSY);
1921         }
1922
1923         /* Check if the lease open lease has already canceled */
1924         lock_res_and_lock(lease);
1925         lease_broken = ldlm_is_cancel(lease);
1926         unlock_res_and_lock(lease);
1927
1928         LDLM_DEBUG(lease, DFID " lease broken? %d",
1929                    PFID(mdt_object_fid(o)), lease_broken);
1930
1931         /* Cancel server side lease. Client side counterpart should
1932          * have been cancelled. It's okay to cancel it now as we've
1933          * held mot_open_sem. */
1934         ldlm_lock_cancel(lease);
1935
1936         if (lease_broken)
1937                 GOTO(out_unlock_sem, rc = -ESTALE);
1938
1939         mdt_lock_reg_init(lh1, LCK_EX);
1940         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
1941                              MDS_INODELOCK_XATTR);
1942         if (rc < 0)
1943                 GOTO(out_unlock_sem, rc);
1944
1945         mdt_lock_reg_init(lh2, LCK_EX);
1946         rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
1947                              MDS_INODELOCK_XATTR);
1948         if (rc < 0)
1949                 GOTO(out_unlock1, rc);
1950
1951         /* Swap layout with orphan object */
1952         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
1953                              mdt_object_child(o2), 0);
1954         if (rc < 0)
1955                 GOTO(out_unlock2, rc);
1956
1957         EXIT;
1958
1959 out_unlock2:
1960         /* Release exclusive LL */
1961         mdt_object_unlock(info, o2, lh2, 1);
1962
1963 out_unlock1:
1964         mdt_object_unlock(info, o1, lh1, 1);
1965
1966 out_unlock_sem:
1967         up_write(&o->mot_open_sem);
1968
1969         /* already swapped */
1970         if (rc == 0) {
1971                 struct mdt_body *repbody;
1972
1973                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1974                 LASSERT(repbody != NULL);
1975                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1976         }
1977
1978 out_obj:
1979         mdt_object_put(info->mti_env, swap_objects ? o1 : o2);
1980
1981         ldlm_reprocess_all(lease->l_resource);
1982
1983 out_lease:
1984         LDLM_LOCK_PUT(lease);
1985
1986         if (ma != NULL) {
1987                 ma->ma_valid = 0;
1988                 ma->ma_need = 0;
1989         }
1990
1991         return rc;
1992 }
1993
1994 #define MFD_CLOSED(mode) ((mode) == MDS_FMODE_CLOSED)
1995 static int mdt_mfd_closed(struct mdt_file_data *mfd)
1996 {
1997         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_mode));
1998 }
1999
2000 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2001 {
2002         struct mdt_object *o = mfd->mfd_object;
2003         struct md_object *next = mdt_object_child(o);
2004         struct md_attr *ma = &info->mti_attr;
2005         int rc = 0;
2006         __u64 mode;
2007         ENTRY;
2008
2009         mode = mfd->mfd_mode;
2010
2011         if (ma->ma_attr_flags & MDS_HSM_RELEASE) {
2012                 rc = mdt_hsm_release(info, o, ma);
2013                 if (rc < 0) {
2014                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2015                                mdt_obd_name(info->mti_mdt),
2016                                PFID(mdt_object_fid(o)), rc);
2017                         /* continue to close even error occurred. */
2018                 }
2019         }
2020
2021         if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2022                 rc = mdt_close_swap_layouts(info, o, ma);
2023                 if (rc < 0) {
2024                         CDEBUG(D_INODE,
2025                                "%s: cannot swap layout of "DFID": rc=%d\n",
2026                                mdt_obd_name(info->mti_mdt),
2027                                PFID(mdt_object_fid(o)), rc);
2028                         /* continue to close even if error occurred. */
2029                 }
2030         }
2031
2032         if (mode & FMODE_WRITE)
2033                 mdt_write_put(o);
2034         else if (mode & MDS_FMODE_EXEC)
2035                 mdt_write_allow(o);
2036
2037         /* Update atime on close only. */
2038         if ((mode & MDS_FMODE_EXEC || mode & FMODE_READ || mode & FMODE_WRITE)
2039             && (ma->ma_valid & MA_INODE) && (ma->ma_attr.la_valid & LA_ATIME)) {
2040                 /* Set the atime only. */
2041                 ma->ma_valid = MA_INODE;
2042                 ma->ma_attr.la_valid = LA_ATIME;
2043                 rc = mo_attr_set(info->mti_env, next, ma);
2044         }
2045
2046         /* If file data is modified, add the dirty flag. */
2047         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2048                 rc = mdt_add_dirty_flag(info, o, ma);
2049
2050         ma->ma_need |= MA_INODE;
2051         ma->ma_valid &= ~MA_INODE;
2052
2053         if (!MFD_CLOSED(mode))
2054                 rc = mo_close(info->mti_env, next, ma, mode);
2055
2056         /* adjust open and lease count */
2057         if (mode & MDS_OPEN_LEASE) {
2058                 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2059                 atomic_dec(&o->mot_lease_count);
2060         }
2061
2062         LASSERT(atomic_read(&o->mot_open_count) > 0);
2063         atomic_dec(&o->mot_open_count);
2064         mdt_mfd_free(mfd);
2065         mdt_object_put(info->mti_env, o);
2066
2067         RETURN(rc);
2068 }
2069
2070 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2071                        struct mdt_body *repbody)
2072 {
2073         struct mdt_export_data *med;
2074         struct mdt_file_data   *mfd;
2075         struct mdt_object      *o;
2076         struct md_attr         *ma = &info->mti_attr;
2077         int                     ret = 0;
2078         int                     rc = 0;
2079         ENTRY;
2080
2081         med = &req->rq_export->exp_mdt_data;
2082         spin_lock(&med->med_open_lock);
2083         mfd = mdt_handle2mfd(med, &info->mti_close_handle, req_is_replay(req));
2084         if (mdt_mfd_closed(mfd)) {
2085                 spin_unlock(&med->med_open_lock);
2086                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2087                        ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2088                        info->mti_close_handle.cookie);
2089                 /** not serious error since bug 3633 */
2090                 rc = -ESTALE;
2091         } else {
2092                 class_handle_unhash(&mfd->mfd_handle);
2093                 list_del_init(&mfd->mfd_list);
2094                 spin_unlock(&med->med_open_lock);
2095
2096                 /* Do not lose object before last unlink. */
2097                 o = mfd->mfd_object;
2098                 mdt_object_get(info->mti_env, o);
2099                 ret = mdt_mfd_close(info, mfd);
2100                 if (repbody != NULL)
2101                         rc = mdt_handle_last_unlink(info, o, ma);
2102                 mdt_object_put(info->mti_env, o);
2103         }
2104
2105         RETURN(rc ? rc : ret);
2106 }
2107
2108 int mdt_close(struct tgt_session_info *tsi)
2109 {
2110         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2111         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2112         struct md_attr         *ma = &info->mti_attr;
2113         struct mdt_body        *repbody = NULL;
2114         int rc, ret = 0;
2115         ENTRY;
2116
2117         mdt_counter_incr(req, LPROC_MDT_CLOSE);
2118         /* Close may come with the Size-on-MDS update. Unpack it. */
2119         rc = mdt_close_unpack(info);
2120         if (rc)
2121                 GOTO(out, rc = err_serious(rc));
2122
2123         /* These fields are no longer used and are left for compatibility.
2124          * size is always zero */
2125         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2126                              0);
2127         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2128                              0);
2129         rc = req_capsule_server_pack(info->mti_pill);
2130         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2131                 mdt_client_compatibility(info);
2132                 if (rc == 0)
2133                         mdt_fix_reply(info);
2134                 mdt_exit_ucred(info);
2135                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2136         }
2137
2138         /* Continue to close handle even if we can not pack reply */
2139         if (rc == 0) {
2140                 repbody = req_capsule_server_get(info->mti_pill,
2141                                                  &RMF_MDT_BODY);
2142                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2143                                                     &RMF_MDT_MD);
2144                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2145                                                        &RMF_MDT_MD,
2146                                                        RCL_SERVER);
2147                 ma->ma_need = MA_INODE | MA_LOV;
2148                 repbody->mbo_eadatasize = 0;
2149                 repbody->mbo_aclsize = 0;
2150         } else {
2151                 rc = err_serious(rc);
2152         }
2153
2154         rc = mdt_close_internal(info, req, repbody);
2155         if (rc != -ESTALE)
2156                 mdt_empty_transno(info, rc);
2157
2158         if (repbody != NULL) {
2159                 mdt_client_compatibility(info);
2160                 rc = mdt_fix_reply(info);
2161         }
2162
2163         mdt_exit_ucred(info);
2164         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2165                 GOTO(out, rc = err_serious(-ENOMEM));
2166
2167         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2168                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2169                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2170 out:
2171         mdt_thread_info_fini(info);
2172         RETURN(rc ? rc : ret);
2173 }