lustre/mdt/mdt_open.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/mdt/mdt_open.c
  33  *
  34  * Lustre Metadata Target (mdt) open/close file handling
  35  *
  36  * Author: Huang Hua <huanghua@clusterfs.com>
  37  */
  38
  39 #define DEBUG_SUBSYSTEM S_MDS
  40
  41 #include <lustre_acl.h>
  42 #include <lustre_mds.h>
  43 #include <lustre_swab.h>
  44 #include "mdt_internal.h"
  45 #include <lustre_nodemap.h>
  46
  47 /* we do nothing because we do not have refcount now */
  48 static void mdt_mfd_get(void *mfdp)
  49 {
  50 }
  51
  52 static struct portals_handle_ops mfd_open_handle_ops = {
  53         .hop_addref = mdt_mfd_get,
  54         .hop_free   = NULL,
  55 };
  56
  57 /* Create a new mdt_file_data struct, initialize it,
  58  * and insert it to global hash table */
  59 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
  60 {
  61         struct mdt_file_data *mfd;
  62         ENTRY;
  63
  64         OBD_ALLOC_PTR(mfd);
  65         if (mfd != NULL) {
  66                 INIT_LIST_HEAD_RCU(&mfd->mfd_open_handle.h_link);
  67                 mfd->mfd_open_handle.h_owner = med;
  68                 INIT_LIST_HEAD(&mfd->mfd_list);
  69                 class_handle_hash(&mfd->mfd_open_handle, &mfd_open_handle_ops);
  70         }
  71
  72         RETURN(mfd);
  73 }
  74
  75 /*
  76  * Find the mfd pointed to by handle in global hash table.
  77  * In case of replay the handle is obsoleted
  78  * but mfd can be found in mfd list by that handle.
  79  * Callers need to be holding med_open_lock.
  80  */
  81 struct mdt_file_data *mdt_open_handle2mfd(struct mdt_export_data *med,
  82                                         const struct lustre_handle *open_handle,
  83                                         bool is_replay_or_resent)
  84 {
  85         struct mdt_file_data   *mfd;
  86         ENTRY;
  87
  88         LASSERT(open_handle != NULL);
  89         mfd = class_handle2object(open_handle->cookie, med);
  90         /* during dw/setattr replay the mfd can be found by old handle */
  91         if (mfd == NULL && is_replay_or_resent) {
  92                 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  93                         if (mfd->mfd_open_handle_old.cookie ==
  94                             open_handle->cookie)
  95                                 RETURN(mfd);
  96                 }
  97                 mfd = NULL;
  98         }
  99
 100         RETURN(mfd);
 101 }
 102
 103 /* free mfd */
 104 void mdt_mfd_free(struct mdt_file_data *mfd)
 105 {
 106         LASSERT(list_empty(&mfd->mfd_list));
 107         OBD_FREE_RCU(mfd, sizeof *mfd, &mfd->mfd_open_handle);
 108 }
 109
 110 static int mdt_create_data(struct mdt_thread_info *info,
 111                            struct mdt_object *p, struct mdt_object *o)
 112 {
 113         struct md_op_spec     *spec = &info->mti_spec;
 114         struct md_attr        *ma   = &info->mti_attr;
 115         int                    rc   = 0;
 116         ENTRY;
 117
 118         if (!md_should_create(spec->sp_cr_flags))
 119                 RETURN(0);
 120
 121         ma->ma_need = MA_INODE | MA_LOV;
 122         ma->ma_valid = 0;
 123         mutex_lock(&o->mot_lov_mutex);
 124         if (!o->mot_lov_created) {
 125                 rc = mdo_create_data(info->mti_env,
 126                                      p ? mdt_object_child(p) : NULL,
 127                                      mdt_object_child(o), spec, ma);
 128                 if (rc == 0)
 129                         rc = mdt_attr_get_complex(info, o, ma);
 130
 131                 if (rc == 0 && ma->ma_valid & MA_LOV)
 132                         o->mot_lov_created = 1;
 133         }
 134
 135         mutex_unlock(&o->mot_lov_mutex);
 136         RETURN(rc);
 137 }
 138
 139 int mdt_write_read(struct mdt_object *o)
 140 {
 141         int rc = 0;
 142         ENTRY;
 143         spin_lock(&o->mot_write_lock);
 144         rc = o->mot_write_count;
 145         spin_unlock(&o->mot_write_lock);
 146         RETURN(rc);
 147 }
 148
 149 int mdt_write_get(struct mdt_object *o)
 150 {
 151         int rc = 0;
 152         ENTRY;
 153         spin_lock(&o->mot_write_lock);
 154         if (o->mot_write_count < 0)
 155                 rc = -ETXTBSY;
 156         else
 157                 o->mot_write_count++;
 158         spin_unlock(&o->mot_write_lock);
 159
 160         RETURN(rc);
 161 }
 162
 163 void mdt_write_put(struct mdt_object *o)
 164 {
 165         ENTRY;
 166         spin_lock(&o->mot_write_lock);
 167         o->mot_write_count--;
 168         spin_unlock(&o->mot_write_lock);
 169         EXIT;
 170 }
 171
 172 static int mdt_write_deny(struct mdt_object *o)
 173 {
 174         int rc = 0;
 175         ENTRY;
 176         spin_lock(&o->mot_write_lock);
 177         if (o->mot_write_count > 0)
 178                 rc = -ETXTBSY;
 179         else
 180                 o->mot_write_count--;
 181         spin_unlock(&o->mot_write_lock);
 182         RETURN(rc);
 183 }
 184
 185 static void mdt_write_allow(struct mdt_object *o)
 186 {
 187         ENTRY;
 188         spin_lock(&o->mot_write_lock);
 189         o->mot_write_count++;
 190         spin_unlock(&o->mot_write_lock);
 191         EXIT;
 192 }
 193
 194 /* there can be no real transaction so prepare the fake one */
 195 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 196 {
 197         struct mdt_device      *mdt = info->mti_mdt;
 198         struct ptlrpc_request  *req = mdt_info_req(info);
 199         struct tg_export_data  *ted;
 200         struct lsd_client_data *lcd;
 201         ENTRY;
 202
 203         if (mdt_rdonly(req->rq_export))
 204                 RETURN_EXIT;
 205
 206         /* transaction has occurred already */
 207         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 208                 RETURN_EXIT;
 209
 210         if (tgt_is_multimodrpcs_client(req->rq_export)) {
 211                 struct thandle         *th;
 212
 213                 /* generate an empty transaction to get a transno
 214                  * and reply data */
 215                 th = dt_trans_create(info->mti_env, mdt->mdt_bottom);
 216                 if (!IS_ERR(th)) {
 217                         rc = dt_trans_start(info->mti_env, mdt->mdt_bottom, th);
 218                         dt_trans_stop(info->mti_env, mdt->mdt_bottom, th);
 219                 }
 220                 RETURN_EXIT;
 221         }
 222
 223         spin_lock(&mdt->mdt_lut.lut_translock);
 224         if (rc != 0) {
 225                 if (info->mti_transno != 0) {
 226                         struct obd_export *exp = req->rq_export;
 227
 228                         CERROR("%s: replay trans %llu NID %s: rc = %d\n",
 229                                mdt_obd_name(mdt), info->mti_transno,
 230                                obd_export_nid2str(exp), rc);
 231                         spin_unlock(&mdt->mdt_lut.lut_translock);
 232                         RETURN_EXIT;
 233                 }
 234         } else if (info->mti_transno == 0) {
 235                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
 236         } else {
 237                 /* should be replay */
 238                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 239                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 240         }
 241         spin_unlock(&mdt->mdt_lut.lut_translock);
 242
 243         CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
 244                info->mti_transno,
 245                req->rq_export->exp_obd->obd_last_committed);
 246
 247         req->rq_transno = info->mti_transno;
 248         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 249
 250         /* update lcd in memory only for resent cases */
 251         ted = &req->rq_export->exp_target_data;
 252         LASSERT(ted);
 253         mutex_lock(&ted->ted_lcd_lock);
 254         lcd = ted->ted_lcd;
 255         if (info->mti_transno < lcd->lcd_last_transno &&
 256             info->mti_transno != 0) {
 257                 /* This should happen during replay. Do not update
 258                  * last rcvd info if replay req transno < last transno,
 259                  * otherwise the following resend(after replay) can not
 260                  * be checked correctly by xid */
 261                 mutex_unlock(&ted->ted_lcd_lock);
 262                 CDEBUG(D_HA, "%s: transno = %llu < last_transno = %llu\n",
 263                        mdt_obd_name(mdt), info->mti_transno,
 264                        lcd->lcd_last_transno);
 265                 RETURN_EXIT;
 266         }
 267
 268         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
 269                 if (info->mti_transno != 0)
 270                         lcd->lcd_last_close_transno = info->mti_transno;
 271                 lcd->lcd_last_close_xid = req->rq_xid;
 272                 lcd->lcd_last_close_result = rc;
 273         } else {
 274                 /* VBR: save versions in last_rcvd for reconstruct. */
 275                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 276                 if (pre_versions) {
 277                         lcd->lcd_pre_versions[0] = pre_versions[0];
 278                         lcd->lcd_pre_versions[1] = pre_versions[1];
 279                         lcd->lcd_pre_versions[2] = pre_versions[2];
 280                         lcd->lcd_pre_versions[3] = pre_versions[3];
 281                 }
 282                 if (info->mti_transno != 0)
 283                         lcd->lcd_last_transno = info->mti_transno;
 284
 285                 lcd->lcd_last_xid = req->rq_xid;
 286                 lcd->lcd_last_result = rc;
 287                 lcd->lcd_last_data = info->mti_opdata;
 288         }
 289         mutex_unlock(&ted->ted_lcd_lock);
 290
 291         EXIT;
 292 }
 293
 294 void mdt_mfd_set_mode(struct mdt_file_data *mfd, u64 open_flags)
 295 {
 296         LASSERT(mfd != NULL);
 297
 298         CDEBUG(D_DENTRY, DFID " Change mfd open_flags %#llo -> %#llo.\n",
 299                PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_open_flags,
 300                open_flags);
 301
 302         mfd->mfd_open_flags = open_flags;
 303 }
 304
 305 /**
 306  * prep ma_lmm/ma_lmv for md_attr from reply
 307  */
 308 static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
 309                                      struct mdt_object *obj,
 310                                      struct md_attr *ma)
 311 {
 312         LASSERT(ma->ma_lmv == NULL && ma->ma_lmm == NULL);
 313         if (S_ISDIR(obj->mot_header.loh_attr)) {
 314                 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
 315                                                     &RMF_MDT_MD);
 316                 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
 317                                                        &RMF_MDT_MD,
 318                                                        RCL_SERVER);
 319                 if (ma->ma_lmv_size > 0)
 320                         ma->ma_need |= MA_LMV;
 321         } else {
 322                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
 323                                                     &RMF_MDT_MD);
 324                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 325                                                        &RMF_MDT_MD,
 326                                                        RCL_SERVER);
 327                 if (ma->ma_lmm_size > 0)
 328                         ma->ma_need |= MA_LOV;
 329         }
 330 }
 331
 332 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 333                         struct mdt_object *o, u64 open_flags, int created,
 334                         struct ldlm_reply *rep)
 335 {
 336         struct ptlrpc_request *req = mdt_info_req(info);
 337         struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 338         struct mdt_file_data *mfd;
 339         struct md_attr *ma  = &info->mti_attr;
 340         struct lu_attr *la  = &ma->ma_attr;
 341         struct mdt_body *repbody;
 342         bool isdir, isreg;
 343         int rc = 0;
 344
 345         ENTRY;
 346         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 347
 348         isreg = S_ISREG(la->la_mode);
 349         isdir = S_ISDIR(la->la_mode);
 350         if (isreg && !(ma->ma_valid & MA_LOV) &&
 351             !(open_flags & MDS_OPEN_RELEASE)) {
 352                 /*
 353                  * No EA, check whether it is will set regEA and dirEA since in
 354                  * above attr get, these size might be zero, so reset it, to
 355                  * retrieve the MD after create obj.
 356                  */
 357                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 358                                                        &RMF_MDT_MD,
 359                                                        RCL_SERVER);
 360                 /* in replay case, p == NULL */
 361                 rc = mdt_create_data(info, p, o);
 362                 if (rc)
 363                         RETURN(rc);
 364
 365                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
 366                         mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
 367         }
 368
 369         CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n",
 370                ma->ma_valid, ma->ma_lmm_size);
 371
 372         if (ma->ma_valid & MA_LOV) {
 373                 LASSERT(ma->ma_lmm_size != 0);
 374                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 375                 if (isdir)
 376                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 377                 else
 378                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 379         }
 380
 381         if (ma->ma_valid & MA_LMV) {
 382                 LASSERT(ma->ma_lmv_size != 0);
 383                 repbody->mbo_eadatasize = ma->ma_lmv_size;
 384                 LASSERT(isdir);
 385                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
 386         }
 387
 388         if (open_flags & MDS_FMODE_WRITE)
 389                 rc = mdt_write_get(o);
 390         else if (open_flags & MDS_FMODE_EXEC)
 391                 rc = mdt_write_deny(o);
 392
 393         if (rc)
 394                 RETURN(rc);
 395
 396         rc = mo_open(info->mti_env, mdt_object_child(o),
 397                      created ? open_flags | MDS_OPEN_CREATED : open_flags);
 398         if (rc != 0) {
 399                 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
 400                  * client does not know which suppgid should be sent to the MDS,
 401                  * or some other(s) changed the target file's GID after this RPC
 402                  * sent to the MDS with the suppgid as the original GID, then we
 403                  * give the client another chance to send the right suppgid. */
 404                 if (rc == -EACCES &&
 405                     allow_client_chgrp(info, lu_ucred(info->mti_env)))
 406                         mdt_set_disposition(info, rep, DISP_OPEN_DENY);
 407
 408                 GOTO(err_out, rc);
 409         }
 410
 411         mfd = mdt_mfd_new(med);
 412         if (mfd == NULL)
 413                 GOTO(err_out, rc = -ENOMEM);
 414
 415         /*
 416          * Keep a reference on this object for this open, and is
 417          * released by mdt_mfd_close().
 418          */
 419         mdt_object_get(info->mti_env, o);
 420         mfd->mfd_object = o;
 421         mfd->mfd_xid = req->rq_xid;
 422
 423         /*
 424          * @open_flags is always not zero. At least it should be FMODE_READ,
 425          * FMODE_WRITE or MDS_FMODE_EXEC.
 426          */
 427         LASSERT(open_flags != 0);
 428
 429         /* Open handling. */
 430         mdt_mfd_set_mode(mfd, open_flags);
 431
 432         atomic_inc(&o->mot_open_count);
 433         if (open_flags & MDS_OPEN_LEASE)
 434                 atomic_inc(&o->mot_lease_count);
 435
 436         /* replay handle */
 437         if (req_is_replay(req)) {
 438                 struct mdt_file_data *old_mfd;
 439                 /* Check wheather old cookie already exist in
 440                  * the list, becasue when do recovery, client
 441                  * might be disconnected from server, and
 442                  * restart replay, so there maybe some orphan
 443                  * mfd here, we should remove them */
 444                 LASSERT(info->mti_rr.rr_open_handle != NULL);
 445                 spin_lock(&med->med_open_lock);
 446                 old_mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 447                                               true);
 448                 if (old_mfd != NULL) {
 449                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 450                                "cookie = %#llx\n", mfd,
 451                                PFID(mdt_object_fid(mfd->mfd_object)),
 452                                info->mti_rr.rr_open_handle->cookie);
 453                         class_handle_unhash(&old_mfd->mfd_open_handle);
 454                         list_del_init(&old_mfd->mfd_list);
 455                         spin_unlock(&med->med_open_lock);
 456                         /* no attr update for that close */
 457                         la->la_valid = 0;
 458                         ma->ma_valid |= MA_FLAGS;
 459                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 460                         mdt_mfd_close(info, old_mfd);
 461                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 462                         ma->ma_valid &= ~MA_FLAGS;
 463                 } else {
 464                         spin_unlock(&med->med_open_lock);
 465                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 466                                "cookie = %#llx\n",
 467                                PFID(mdt_object_fid(mfd->mfd_object)),
 468                                info->mti_rr.rr_open_handle->cookie);
 469                 }
 470
 471                 CDEBUG(D_HA, "Store old cookie %#llx in new mfd\n",
 472                        info->mti_rr.rr_open_handle->cookie);
 473
 474                 mfd->mfd_open_handle_old = *info->mti_rr.rr_open_handle;
 475         }
 476
 477         repbody->mbo_open_handle.cookie = mfd->mfd_open_handle.h_cookie;
 478
 479         if (req->rq_export->exp_disconnected) {
 480                 spin_lock(&med->med_open_lock);
 481                 class_handle_unhash(&mfd->mfd_open_handle);
 482                 list_del_init(&mfd->mfd_list);
 483                 spin_unlock(&med->med_open_lock);
 484                 mdt_mfd_close(info, mfd);
 485         } else {
 486                 spin_lock(&med->med_open_lock);
 487                 list_add_tail(&mfd->mfd_list, &med->med_open_head);
 488                 spin_unlock(&med->med_open_lock);
 489         }
 490
 491         mdt_empty_transno(info, rc);
 492
 493         RETURN(rc);
 494
 495 err_out:
 496         if (open_flags & MDS_FMODE_WRITE)
 497                 mdt_write_put(o);
 498         else if (open_flags & MDS_FMODE_EXEC)
 499                 mdt_write_allow(o);
 500
 501         return rc;
 502 }
 503
 504 static int mdt_finish_open(struct mdt_thread_info *info,
 505                            struct mdt_object *p, struct mdt_object *o,
 506                            u64 open_flags, int created,
 507                            struct ldlm_reply *rep)
 508 {
 509         struct ptlrpc_request   *req = mdt_info_req(info);
 510         struct obd_export       *exp = req->rq_export;
 511         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 512         struct md_attr          *ma  = &info->mti_attr;
 513         struct lu_attr          *la  = &ma->ma_attr;
 514         struct mdt_file_data    *mfd;
 515         struct mdt_body         *repbody;
 516         int                      rc = 0;
 517         int                      isreg, isdir, islnk;
 518         struct list_head        *t;
 519         ENTRY;
 520
 521         LASSERT(ma->ma_valid & MA_INODE);
 522
 523         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 524
 525         isreg = S_ISREG(la->la_mode);
 526         isdir = S_ISDIR(la->la_mode);
 527         islnk = S_ISLNK(la->la_mode);
 528         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 529
 530         /* compatibility check for 2.10 clients when it tries to open mirrored
 531          * files. 2.10 clients don't verify overlapping components so they
 532          * would read and write mirrored files just as if they were normal
 533          * PFL files, which will cause the problem that sycned mirrors actually
 534          * contain different data.
 535          * Older clients are not a concern here because they don't even
 536          * understand PFL layout. */
 537         if (isreg && !exp_connect_flr(exp) && ma->ma_valid & MA_LOV &&
 538             mdt_lmm_is_flr(ma->ma_lmm)) {
 539                 /* LU-10286: for simplicity clients who don't understand
 540                  * mirrored layout(with connect flag OBD_CONNECT2_FLR) won't
 541                  * be able to open mirrored files */
 542                 RETURN(-EOPNOTSUPP);
 543         }
 544
 545         /* LU-2275, simulate broken behaviour (esp. prevalent in
 546          * pre-2.4 servers where a very strange reply is sent on error
 547          * that looks like it was actually almost successful and a
 548          * failure at the same time.) */
 549         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 550                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 551                                                DISP_LOOKUP_NEG |
 552                                                DISP_LOOKUP_POS);
 553
 554                 if (open_flags & MDS_OPEN_LOCK)
 555                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 556
 557                 RETURN(-ENOENT);
 558         }
 559
 560 #ifdef CONFIG_FS_POSIX_ACL
 561         if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 562                 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
 563                 if (IS_ERR(nodemap))
 564                         RETURN(PTR_ERR(nodemap));
 565
 566                 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
 567                 nodemap_putref(nodemap);
 568                 if (rc)
 569                         RETURN(rc);
 570         }
 571 #endif
 572
 573         /*
 574          * If we are following a symlink, don't open; and do not return open
 575          * handle for special nodes as client required.
 576          */
 577         if (islnk || (!isreg && !isdir &&
 578             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 579                 lustre_msg_set_transno(req->rq_repmsg, 0);
 580                 RETURN(0);
 581         }
 582
 583         /*
 584          * We need to return the existing object's fid back, so it is done here,
 585          * after preparing the reply.
 586          */
 587         if (!created && (open_flags & MDS_OPEN_EXCL) &&
 588             (open_flags & MDS_OPEN_CREAT))
 589                 RETURN(-EEXIST);
 590
 591         /* This can't be done earlier, we need to return reply body */
 592         if (isdir) {
 593                 if (open_flags & (MDS_OPEN_CREAT | MDS_FMODE_WRITE)) {
 594                         /* We are trying to create or write an existing dir. */
 595                         RETURN(-EISDIR);
 596                 }
 597         } else if (open_flags & MDS_OPEN_DIRECTORY)
 598                 RETURN(-ENOTDIR);
 599
 600         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 601                                  OBD_FAIL_MDS_LDLM_REPLY_NET | OBD_FAIL_ONCE))
 602                 RETURN(-EAGAIN);
 603
 604         mfd = NULL;
 605         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 606                 spin_lock(&med->med_open_lock);
 607                 list_for_each(t, &med->med_open_head) {
 608                         mfd = list_entry(t, struct mdt_file_data, mfd_list);
 609                         if (mfd->mfd_xid == req->rq_xid)
 610                                 break;
 611                         mfd = NULL;
 612                 }
 613                 spin_unlock(&med->med_open_lock);
 614
 615                 if (mfd != NULL) {
 616                         repbody->mbo_open_handle.cookie =
 617                                 mfd->mfd_open_handle.h_cookie;
 618                         /* set repbody->ea_size for resent case */
 619                         if (ma->ma_valid & MA_LOV) {
 620                                 LASSERT(ma->ma_lmm_size != 0);
 621                                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 622                                 if (isdir)
 623                                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 624                                 else
 625                                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 626                         }
 627                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 628                         RETURN(0);
 629                 }
 630         }
 631
 632         rc = mdt_mfd_open(info, p, o, open_flags, created, rep);
 633         if (!rc)
 634                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 635
 636         RETURN(rc);
 637 }
 638
 639 void mdt_reconstruct_open(struct mdt_thread_info *info,
 640                           struct mdt_lock_handle *lhc)
 641 {
 642         const struct lu_env *env = info->mti_env;
 643         struct mdt_device *mdt = info->mti_mdt;
 644         struct req_capsule *pill = info->mti_pill;
 645         struct ptlrpc_request *req = mdt_info_req(info);
 646         struct md_attr *ma = &info->mti_attr;
 647         struct mdt_reint_record *rr = &info->mti_rr;
 648         u64 open_flags = info->mti_spec.sp_cr_flags;
 649         struct ldlm_reply *ldlm_rep;
 650         struct mdt_object *parent;
 651         struct mdt_object *child;
 652         struct mdt_body *repbody;
 653         u64 opdata;
 654         int rc;
 655         ENTRY;
 656
 657         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 658         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 659         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 660
 661         ma->ma_need = MA_INODE | MA_HSM;
 662         ma->ma_valid = 0;
 663
 664         opdata = mdt_req_from_lrd(req, info->mti_reply_data);
 665         mdt_set_disposition(info, ldlm_rep, opdata);
 666
 667         CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n",
 668                ldlm_rep->lock_policy_res1, req->rq_status);
 669
 670         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
 671             req->rq_status != 0)
 672                 /* We did not create successfully, return error to client. */
 673                 GOTO(out, rc = req->rq_status);
 674
 675         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
 676                 struct obd_export *exp = req->rq_export;
 677                 /*
 678                  * We failed after creation, but we do not know in which step
 679                  * we failed. So try to check the child object.
 680                  */
 681                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
 682                 if (IS_ERR(parent)) {
 683                         rc = PTR_ERR(parent);
 684                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
 685                                       " Evicting client %s with export %s.\n",
 686                                       PFID(rr->rr_fid1), rc,
 687                                       obd_uuid2str(&exp->exp_client_uuid),
 688                                       obd_export_nid2str(exp));
 689                         mdt_export_evict(exp);
 690                         RETURN_EXIT;
 691                 }
 692
 693                 child = mdt_object_find(env, mdt, rr->rr_fid2);
 694                 if (IS_ERR(child)) {
 695                         rc = PTR_ERR(child);
 696                         LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
 697                                       "evicting client %s with export %s\n",
 698                                       PFID(rr->rr_fid2), rc,
 699                                       obd_uuid2str(&exp->exp_client_uuid),
 700                                       obd_export_nid2str(exp));
 701                         mdt_object_put(env, parent);
 702                         mdt_export_evict(exp);
 703                         RETURN_EXIT;
 704                 }
 705
 706                 if (unlikely(mdt_object_remote(child))) {
 707                         mdt_object_put(env, parent);
 708                         mdt_object_put(env, child);
 709                         /* the child object was created on remote server */
 710                         if (!mdt_is_dne_client(exp))
 711                                 /* Return -EIO for old client */
 712                                 GOTO(out, rc = -EIO);
 713                         repbody->mbo_fid1 = *rr->rr_fid2;
 714                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 715                         GOTO(out, rc = 0);
 716                 }
 717                 if (mdt_object_exists(child)) {
 718                         mdt_prep_ma_buf_from_rep(info, child, ma);
 719                         rc = mdt_attr_get_complex(info, child, ma);
 720                         if (!rc)
 721                                 rc = mdt_finish_open(info, parent, child,
 722                                                      open_flags, 1, ldlm_rep);
 723                         mdt_object_put(env, parent);
 724                         mdt_object_put(env, child);
 725                         if (!rc)
 726                                 mdt_pack_size2body(info, rr->rr_fid2,
 727                                                    &lhc->mlh_reg_lh);
 728                         GOTO(out, rc);
 729                 }
 730                 /* the child does not exist, we should do regular open */
 731                 mdt_object_put(env, parent);
 732                 mdt_object_put(env, child);
 733         }
 734         /* We did not try to create, so we are a pure open */
 735         rc = mdt_reint_open(info, lhc);
 736         EXIT;
 737 out:
 738         req->rq_status = rc;
 739         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 740         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
 741 }
 742
 743 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
 744 {
 745         u64 open_flags = info->mti_spec.sp_cr_flags;
 746         struct mdt_reint_record *rr = &info->mti_rr;
 747         struct md_attr *ma = &info->mti_attr;
 748         struct mdt_object *o;
 749         int rc;
 750
 751         ENTRY;
 752         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
 753         if (IS_ERR(o))
 754                 RETURN(rc = PTR_ERR(o));
 755
 756         if (unlikely(mdt_object_remote(o))) {
 757                 /* the child object was created on remote server */
 758                 struct mdt_body *repbody;
 759
 760                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 761                                                 DISP_LOOKUP_EXECD |
 762                                                 DISP_LOOKUP_POS));
 763                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 764                 repbody->mbo_fid1 = *rr->rr_fid2;
 765                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 766                 rc = 0;
 767         } else {
 768                 if (mdt_object_exists(o)) {
 769                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 770                                                         DISP_LOOKUP_EXECD |
 771                                                         DISP_LOOKUP_POS));
 772                         mdt_prep_ma_buf_from_rep(info, o, ma);
 773                         rc = mdt_attr_get_complex(info, o, ma);
 774                         if (rc == 0)
 775                                 rc = mdt_finish_open(info, NULL, o, open_flags,
 776                                                      0, rep);
 777                 } else {
 778                         rc = -ENOENT;
 779                 }
 780         }
 781
 782         mdt_object_put(info->mti_env, o);
 783         RETURN(rc);
 784 }
 785
 786 /* lock object for open */
 787 static int mdt_object_open_lock(struct mdt_thread_info *info,
 788                                 struct mdt_object *obj,
 789                                 struct mdt_lock_handle *lhc,
 790                                 __u64 *ibits)
 791 {
 792         struct md_attr *ma = &info->mti_attr;
 793         __u64 open_flags = info->mti_spec.sp_cr_flags;
 794         __u64 trybits = 0;
 795         enum ldlm_mode lm = LCK_CR;
 796         bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
 797         bool try_layout = false;
 798         bool create_layout = false;
 799         int rc = 0;
 800         int dom_stripes = LMM_NO_DOM;
 801         bool dom_lock = false;
 802
 803         ENTRY;
 804
 805         *ibits = 0;
 806         mdt_lock_handle_init(lhc);
 807
 808         if (req_is_replay(mdt_info_req(info)))
 809                 RETURN(0);
 810
 811         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
 812                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
 813                     md_should_create(open_flags))
 814                         create_layout = true;
 815                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
 816                     ma->ma_need & MA_LOV)
 817                         try_layout = true;
 818
 819                 /* DoM files can have just MDT stripe or combined MDT + OST
 820                  * stripes.
 821                  * - In the first case the open for read/write will do IO to
 822                  *   the MDT stripe and it makes sense to take IO lock in
 823                  *   advance along with OPEN even if it is blocking lock.
 824                  * - In the second case it is just size of MDT stripe and it
 825                  *   is quite unlikely that client will write into it, though
 826                  *   it may read it. So IO lock will be taken optionally if it
 827                  *   is non-blocking one.
 828                  */
 829                 if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
 830                         dom_stripes = mdt_lmm_dom_entry(ma->ma_lmm);
 831
 832                 if (dom_stripes == LMM_DOM_ONLY &&
 833                     info->mti_mdt->mdt_opts.mo_dom_lock > 0 &&
 834                     !mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
 835                         dom_lock = true;
 836         }
 837
 838         if (acq_lease) {
 839                 /* lease open, acquire write mode of open sem */
 840                 down_write(&obj->mot_open_sem);
 841
 842                 /* Lease exists and ask for new lease */
 843                 if (atomic_read(&obj->mot_lease_count) > 0) {
 844                         /* only exclusive open is supported, so lease
 845                          * are conflicted to each other */
 846                         GOTO(out, rc = -EBUSY);
 847                 }
 848
 849                 /* Lease must be with open lock */
 850                 if (!(open_flags & MDS_OPEN_LOCK)) {
 851                         CERROR("%s: Request lease for file:"DFID ", but open lock "
 852                                "is missed, open_flags = %#llo : rc = %d\n",
 853                                mdt_obd_name(info->mti_mdt),
 854                                PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
 855                         GOTO(out, rc = -EPROTO);
 856                 }
 857
 858                 /* XXX: only exclusive open is supported. */
 859                 lm = LCK_EX;
 860                 *ibits = MDS_INODELOCK_OPEN;
 861
 862                 /* never grant LCK_EX layout lock to client */
 863                 try_layout = false;
 864         } else { /* normal open */
 865                 /* normal open holds read mode of open sem */
 866                 down_read(&obj->mot_open_sem);
 867
 868                 if (open_flags & MDS_OPEN_LOCK) {
 869                         if (open_flags & MDS_FMODE_WRITE)
 870                                 lm = LCK_CW;
 871                         else if (open_flags & MDS_FMODE_EXEC)
 872                                 lm = LCK_PR;
 873                         else
 874                                 lm = LCK_CR;
 875
 876                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
 877                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
 878                         if (open_flags & MDS_FMODE_WRITE)
 879                                 lm = LCK_CW;
 880                         else
 881                                 lm = LCK_CR;
 882
 883                         /* revoke lease */
 884                         *ibits = MDS_INODELOCK_OPEN;
 885                         try_layout = false;
 886
 887                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 888                 } else if (dom_lock) {
 889                         lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR;
 890                         if (info->mti_mdt->mdt_opts.mo_dom_lock ==
 891                             TRYLOCK_DOM_ON_OPEN) {
 892                                 trybits |= MDS_INODELOCK_DOM |
 893                                            MDS_INODELOCK_LAYOUT;
 894                         } else {
 895                                 /* mo_dom_lock == ALWAYS_DOM_LOCK_ON_OPEN */
 896                                 *ibits = MDS_INODELOCK_DOM;
 897                                 if (info->mti_mdt->mdt_opts.mo_dom_read_open) {
 898                                         trybits |= MDS_INODELOCK_LAYOUT;
 899                                 }
 900                         }
 901                 }
 902
 903                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
 904                         PFID(mdt_object_fid(obj)),
 905                         atomic_read(&obj->mot_lease_count), lm);
 906         }
 907
 908         mdt_lock_reg_init(lhc, lm);
 909
 910         /* Return lookup lock to validate inode at the client side.
 911          * This is pretty important otherwise MDT will return layout
 912          * lock for each open.
 913          * However this is a double-edged sword because changing
 914          * permission will revoke a huge number of LOOKUP locks.
 915          */
 916         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
 917                 if (!(*ibits & MDS_INODELOCK_LOOKUP))
 918                         trybits |= MDS_INODELOCK_LOOKUP;
 919                 trybits |= MDS_INODELOCK_LAYOUT;
 920         }
 921
 922         if (*ibits | trybits)
 923                 rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, false);
 924
 925         CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
 926                ", open_flags = %#llo, try_layout = %d : rc = %d\n",
 927                mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
 928                *ibits, trybits, open_flags, try_layout, rc);
 929
 930         /* will change layout, revoke layout locks by enqueuing EX lock. */
 931         if (rc == 0 && create_layout) {
 932                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
 933
 934                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
 935                         ", open_flags = %#llo\n",
 936                         PFID(mdt_object_fid(obj)), open_flags);
 937
 938                 /* We cannot enqueue another lock for the same resource we
 939                  * already have a lock for, due to mechanics of waiting list
 940                  * iterating in ldlm, see LU-3601.
 941                  * As such we'll drop the open lock we just got above here,
 942                  * it's ok not to have this open lock as it's main purpose is to
 943                  * flush unused cached client open handles. */
 944                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 945                         mdt_object_unlock(info, obj, lhc, 1);
 946
 947                 LASSERT(!try_layout);
 948                 mdt_lock_handle_init(ll);
 949                 mdt_lock_reg_init(ll, LCK_EX);
 950                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
 951
 952                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
 953         }
 954
 955         /* Check if there is any other open handles after acquiring
 956          * open lock. At this point, caching open handles have been revoked
 957          * by open lock.
 958          * XXX: Now only exclusive open is supported. Need to check the
 959          * type of open for generic lease support. */
 960         if (rc == 0 && acq_lease) {
 961                 struct ptlrpc_request *req = mdt_info_req(info);
 962                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 963                 struct mdt_file_data *mfd;
 964                 bool is_replay_or_resent;
 965                 int open_count = 0;
 966
 967                 /* For lease: application can open a file and then apply lease,
 968                  * @handle contains original open handle in that case.
 969                  * In recovery, open REQ will be replayed and the lease REQ may
 970                  * be resent that means the open handle is already stale, so we
 971                  * need to fix it up here by finding new handle. */
 972                 is_replay_or_resent = req_is_replay(req) ||
 973                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
 974
 975                 /* if the request is _not_ a replay request, rr_open_handle
 976                  * may be used to hold an open file handle which is issuing the
 977                  * lease request, so that this openhandle doesn't count. */
 978                 mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 979                                           is_replay_or_resent);
 980                 if (mfd != NULL)
 981                         ++open_count;
 982
 983                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
 984                         PFID(mdt_object_fid(obj)),
 985                         atomic_read(&obj->mot_open_count), open_count);
 986
 987                 if (atomic_read(&obj->mot_open_count) > open_count)
 988                         GOTO(out, rc = -EBUSY);
 989         }
 990         GOTO(out, rc);
 991
 992 out:
 993         RETURN(rc);
 994 }
 995
 996 static void mdt_object_open_unlock(struct mdt_thread_info *info,
 997                                    struct mdt_object *obj,
 998                                    struct mdt_lock_handle *lhc,
 999                                    __u64 ibits, int rc)
1000 {
1001         __u64 open_flags = info->mti_spec.sp_cr_flags;
1002         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1003         ENTRY;
1004
1005         if (req_is_replay(mdt_info_req(info)))
1006                 RETURN_EXIT;
1007
1008         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1009          * return to client side. */
1010         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1011                 mdt_object_unlock(info, obj, ll, 1);
1012
1013         ll = &info->mti_lh[MDT_LH_LAYOUT];
1014         /* Release local layout lock, layout was created */
1015         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1016                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1017                 mdt_object_unlock(info, obj, ll, 1);
1018         }
1019
1020         if (open_flags & MDS_OPEN_LEASE)
1021                 up_write(&obj->mot_open_sem);
1022         else
1023                 up_read(&obj->mot_open_sem);
1024
1025         /* Cross-ref case, the lock should be returned to the client */
1026         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1027                 RETURN_EXIT;
1028
1029         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1030             !(ibits & MDS_INODELOCK_DOM)) {
1031                 /* for the open request, the lock will only return to client
1032                  * if open or layout lock is granted. */
1033                 rc = 1;
1034         }
1035
1036         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1037                 struct ldlm_reply       *ldlm_rep;
1038
1039                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1040                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1041                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1042                         mdt_object_unlock(info, obj, lhc, 1);
1043         }
1044         RETURN_EXIT;
1045 }
1046
1047 /**
1048  * Check release is permitted for the current HSM flags.
1049  */
1050 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1051 {
1052         if (!(ma->ma_valid & MA_HSM))
1053                 return false;
1054
1055         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1056                 return false;
1057
1058         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1059                 return false;
1060
1061         return true;
1062 }
1063
1064 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1065                                 struct ldlm_reply *rep,
1066                                 struct mdt_lock_handle *lhc)
1067 {
1068         const struct lu_env *env = info->mti_env;
1069         struct mdt_device *mdt = info->mti_mdt;
1070         u64 open_flags = info->mti_spec.sp_cr_flags;
1071         struct mdt_reint_record *rr = &info->mti_rr;
1072         struct md_attr *ma = &info->mti_attr;
1073         struct mdt_object *parent = NULL;
1074         struct mdt_object *o;
1075         bool object_locked = false;
1076         u64 ibits = 0;
1077         int rc;
1078
1079         ENTRY;
1080         if (md_should_create(open_flags)) {
1081                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1082                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1083                         if (IS_ERR(parent)) {
1084                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1085                                        " for anonymous created %ld, try to"
1086                                        " use server-side parent.\n",
1087                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1088                                 parent = NULL;
1089                         }
1090                 }
1091                 if (parent == NULL)
1092                         ma->ma_need |= MA_PFID;
1093         }
1094
1095         o = mdt_object_find(env, mdt, rr->rr_fid2);
1096         if (IS_ERR(o))
1097                 GOTO(out_parent_put, rc = PTR_ERR(o));
1098
1099         if (mdt_object_remote(o)) {
1100                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1101                        mdt_obd_name(info->mti_mdt),
1102                        PFID(rr->rr_fid2));
1103                 GOTO(out, rc = -EREMOTE);
1104         } else if (!mdt_object_exists(o)) {
1105                 mdt_set_disposition(info, rep,
1106                                     DISP_IT_EXECD |
1107                                     DISP_LOOKUP_EXECD |
1108                                     DISP_LOOKUP_NEG);
1109                 GOTO(out, rc = -ENOENT);
1110         }
1111
1112         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1113
1114         mdt_prep_ma_buf_from_rep(info, o, ma);
1115         if (open_flags & MDS_OPEN_RELEASE)
1116                 ma->ma_need |= MA_HSM;
1117         rc = mdt_attr_get_complex(info, o, ma);
1118         if (rc)
1119                 GOTO(out, rc);
1120
1121         /* We should not change file's existing LOV EA */
1122         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1123             open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1124                 GOTO(out, rc = -EEXIST);
1125
1126         /* If a release request, check file open flags are fine and ask for an
1127          * exclusive open access. */
1128         if (open_flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1129                 GOTO(out, rc = -EPERM);
1130
1131         rc = mdt_check_resent_lock(info, o, lhc);
1132         if (rc < 0) {
1133                 GOTO(out, rc);
1134         } else if (rc > 0) {
1135                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1136                 object_locked = true;
1137                 if (rc)
1138                         GOTO(out_unlock, rc);
1139         }
1140
1141         if (ma->ma_valid & MA_PFID) {
1142                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1143                 if (IS_ERR(parent)) {
1144                         CDEBUG(D_INODE, "Fail to find parent "DFID
1145                                " for anonymous created %ld, try to"
1146                                " use system default.\n",
1147                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1148                         parent = NULL;
1149                 }
1150         }
1151
1152         rc = mdt_finish_open(info, parent, o, open_flags, 0, rep);
1153         if (!rc) {
1154                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1155                 if (open_flags & MDS_OPEN_LOCK)
1156                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1157                 if (open_flags & MDS_OPEN_LEASE)
1158                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1159         }
1160         GOTO(out_unlock, rc);
1161
1162 out_unlock:
1163         if (object_locked)
1164                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1165 out:
1166         mdt_object_put(env, o);
1167         if (rc == 0)
1168                 mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
1169 out_parent_put:
1170         if (parent != NULL)
1171                 mdt_object_put(env, parent);
1172         return rc;
1173 }
1174
1175 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1176 static int mdt_cross_open(struct mdt_thread_info *info,
1177                           const struct lu_fid *parent_fid,
1178                           const struct lu_fid *fid,
1179                           struct ldlm_reply *rep, u64 open_flags)
1180 {
1181         struct md_attr *ma = &info->mti_attr;
1182         struct mdt_object *o;
1183         int rc;
1184         ENTRY;
1185
1186         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1187         if (IS_ERR(o))
1188                 RETURN(rc = PTR_ERR(o));
1189
1190         if (mdt_object_remote(o)) {
1191                 /* Something is wrong here, the object is on another MDS! */
1192                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1193                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1194                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1195                                 &o->mot_obj,
1196                                 "Object isn't on this server! FLD error?");
1197                 rc = -EFAULT;
1198         } else {
1199                 if (mdt_object_exists(o)) {
1200                         int mask;
1201
1202                         /* Do permission check for cross-open after converting
1203                          * MDS_OPEN_* flags to MAY_* permission mask.
1204                          */
1205                         mask = mds_accmode(open_flags);
1206
1207                         rc = mo_permission(info->mti_env, NULL,
1208                                            mdt_object_child(o), NULL, mask);
1209                         if (rc)
1210                                 goto out;
1211
1212                         mdt_prep_ma_buf_from_rep(info, o, ma);
1213                         rc = mdt_attr_get_complex(info, o, ma);
1214                         if (rc != 0)
1215                                 GOTO(out, rc);
1216
1217                         mdt_pack_secctx_in_reply(info, o);
1218
1219                         rc = mdt_finish_open(info, NULL, o, open_flags, 0, rep);
1220                 } else {
1221                         /*
1222                          * Something is wrong here. lookup was positive but
1223                          * there is no object!
1224                          */
1225                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1226                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1227                         rc = -EFAULT;
1228                 }
1229         }
1230 out:
1231         mdt_object_put(info->mti_env, o);
1232
1233         RETURN(rc);
1234 }
1235
1236 /*
1237  * find root object and take its xattr lock if it's on remote MDT, later create
1238  * may use fs default striping (which is stored in root xattr).
1239  */
1240 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1241                                struct mdt_device *mdt)
1242 {
1243         struct mdt_object *md_root = mdt->mdt_md_root;
1244         struct lustre_handle lhroot;
1245         int rc;
1246
1247         if (md_root == NULL) {
1248                 lu_root_fid(&info->mti_tmp_fid1);
1249                 md_root = mdt_object_find(info->mti_env, mdt,
1250                                           &info->mti_tmp_fid1);
1251                 if (IS_ERR(md_root))
1252                         return PTR_ERR(md_root);
1253
1254                 spin_lock(&mdt->mdt_lock);
1255                 if (mdt->mdt_md_root != NULL) {
1256                         spin_unlock(&mdt->mdt_lock);
1257
1258                         LASSERTF(mdt->mdt_md_root == md_root,
1259                                  "Different root object ("
1260                                  DFID") instances, %p, %p\n",
1261                                  PFID(&info->mti_tmp_fid1),
1262                                  mdt->mdt_md_root, md_root);
1263                         LASSERT(atomic_read(
1264                                 &md_root->mot_obj.lo_header->loh_ref) > 1);
1265
1266                         mdt_object_put(info->mti_env, md_root);
1267                 } else {
1268                         mdt->mdt_md_root = md_root;
1269                         spin_unlock(&mdt->mdt_lock);
1270                 }
1271         }
1272
1273         if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1274                 return 0;
1275
1276         rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root),
1277                                     &lhroot, LCK_PR, MDS_INODELOCK_XATTR,
1278                                     true);
1279         if (rc < 0)
1280                 return rc;
1281
1282         md_root->mot_cache_attr = 1;
1283
1284         /* don't cancel this lock, so that we know the cached xattr is valid. */
1285         ldlm_lock_decref(&lhroot, LCK_PR);
1286
1287         return 0;
1288 }
1289
1290 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1291 {
1292         struct mdt_device *mdt = info->mti_mdt;
1293         struct ptlrpc_request *req = mdt_info_req(info);
1294         struct mdt_object *parent;
1295         struct mdt_object *child;
1296         struct mdt_lock_handle *lh;
1297         struct ldlm_reply *ldlm_rep;
1298         struct mdt_body *repbody;
1299         struct lu_fid *child_fid = &info->mti_tmp_fid1;
1300         struct md_attr *ma = &info->mti_attr;
1301         u64 open_flags = info->mti_spec.sp_cr_flags;
1302         u64 ibits = 0;
1303         struct mdt_reint_record *rr = &info->mti_rr;
1304         int result, rc;
1305         int created = 0;
1306         int object_locked = 0;
1307         u32 msg_flags;
1308
1309         ENTRY;
1310         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1311                                (obd_timeout + 1) / 4);
1312
1313         mdt_counter_incr(req, LPROC_MDT_OPEN);
1314         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1315
1316         ma->ma_need = MA_INODE;
1317         ma->ma_valid = 0;
1318
1319         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1320         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1321
1322         if (unlikely(open_flags & MDS_OPEN_JOIN_FILE)) {
1323                 CERROR("file join is not supported anymore.\n");
1324                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1325         }
1326         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1327
1328         if ((open_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1329             info->mti_spec.u.sp_ea.eadata == NULL)
1330                 GOTO(out, result = err_serious(-EINVAL));
1331
1332         if (open_flags & MDS_FMODE_WRITE &&
1333             exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1334                 GOTO(out, result = -EROFS);
1335
1336         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1337                "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1338                PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2),
1339                open_flags, ma->ma_attr.la_mode, msg_flags);
1340
1341         if (info->mti_cross_ref) {
1342                 /* This is cross-ref open */
1343                 mdt_set_disposition(info, ldlm_rep,
1344                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1345                              DISP_LOOKUP_POS));
1346                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1347                                         ldlm_rep, open_flags);
1348                 GOTO(out, result);
1349         } else if (req_is_replay(req)) {
1350                 result = mdt_open_by_fid(info, ldlm_rep);
1351
1352                 if (result != -ENOENT)
1353                         GOTO(out, result);
1354
1355                 /* We didn't find the correct object, so we need to re-create it
1356                  * via a regular replay. */
1357                 if (!(open_flags & MDS_OPEN_CREAT)) {
1358                         DEBUG_REQ(D_ERROR, req,
1359                                   "OPEN & CREAT not in open replay/by_fid.");
1360                         GOTO(out, result = -EFAULT);
1361                 }
1362                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1363         } else if (open_flags & (MDS_OPEN_BY_FID | MDS_OPEN_LOCK)) {
1364                 /*
1365                  * MDS_OPEN_LOCK is checked for backward compatibility with 2.1
1366                  * client.
1367                  */
1368                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1369                 if (result < 0)
1370                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1371                                PFID(rr->rr_fid2), result);
1372                 GOTO(out, result);
1373         }
1374
1375         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1376                 GOTO(out, result = err_serious(-ENOMEM));
1377
1378         mdt_set_disposition(info, ldlm_rep,
1379                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1380
1381         if (!lu_name_is_valid(&rr->rr_name))
1382                 GOTO(out, result = -EPROTO);
1383
1384         result = mdt_lock_root_xattr(info, mdt);
1385         if (result < 0)
1386                 GOTO(out, result);
1387
1388 again:
1389         lh = &info->mti_lh[MDT_LH_PARENT];
1390         mdt_lock_pdo_init(lh, (open_flags & MDS_OPEN_CREAT) ? LCK_PW : LCK_PR,
1391                           &rr->rr_name);
1392
1393         parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1394         if (IS_ERR(parent))
1395                 GOTO(out, result = PTR_ERR(parent));
1396
1397         result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1398         if (result != 0) {
1399                 mdt_object_put(info->mti_env, parent);
1400                 GOTO(out, result);
1401         }
1402
1403         /* get and check version of parent */
1404         result = mdt_version_get_check(info, parent, 0);
1405         if (result)
1406                 GOTO(out_parent, result);
1407
1408         fid_zero(child_fid);
1409
1410         result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1411                             &rr->rr_name, child_fid, &info->mti_spec);
1412
1413         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1414                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1415                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1416                  PFID(child_fid));
1417
1418         if (result != 0 && result != -ENOENT && result != -ESTALE)
1419                 GOTO(out_parent, result);
1420
1421         if (result == -ENOENT || result == -ESTALE) {
1422                 /* If the object is dead, let's check if the object
1423                  * is being migrated to a new object */
1424                 if (result == -ESTALE) {
1425                         struct lu_buf lmv_buf;
1426
1427                         lmv_buf.lb_buf = info->mti_xattr_buf;
1428                         lmv_buf.lb_len = sizeof(info->mti_xattr_buf);
1429                         rc = mo_xattr_get(info->mti_env,
1430                                           mdt_object_child(parent),
1431                                           &lmv_buf, XATTR_NAME_LMV);
1432                         if (rc > 0) {
1433                                 struct lmv_mds_md_v1 *lmv;
1434
1435                                 lmv = lmv_buf.lb_buf;
1436                                 if (le32_to_cpu(lmv->lmv_hash_type) &
1437                                                 LMV_HASH_FLAG_MIGRATION) {
1438                                         /* Get the new parent FID and retry */
1439                                         mdt_object_unlock_put(info, parent,
1440                                                               lh, 1);
1441                                         mdt_lock_handle_init(lh);
1442                                         fid_le_to_cpu(
1443                                                 (struct lu_fid *)rr->rr_fid1,
1444                                                 &lmv->lmv_stripe_fids[1]);
1445                                         goto again;
1446                                 }
1447                         }
1448                 }
1449
1450                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1451                 if (result == -ESTALE) {
1452                         /*
1453                          * -ESTALE means the parent is a dead(unlinked) dir, so
1454                          * it should return -ENOENT to in accordance with the
1455                          * original mds implementaion.
1456                          */
1457                         GOTO(out_parent, result = -ENOENT);
1458                 }
1459
1460                 if (!(open_flags & MDS_OPEN_CREAT))
1461                         GOTO(out_parent, result);
1462                 if (mdt_rdonly(req->rq_export))
1463                         GOTO(out_parent, result = -EROFS);
1464                 *child_fid = *info->mti_rr.rr_fid2;
1465                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1466                          PFID(child_fid));
1467                 /* In the function below, .hs_keycmp resolves to
1468                  * lu_obj_hop_keycmp() */
1469                 /* coverity[overrun-buffer-val] */
1470                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1471         } else {
1472                 /*
1473                  * Check for O_EXCL is moved to the mdt_finish_open(),
1474                  * we need to return FID back in that case.
1475                  */
1476                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1477                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1478         }
1479         if (IS_ERR(child))
1480                 GOTO(out_parent, result = PTR_ERR(child));
1481
1482         /** check version of child  */
1483         rc = mdt_version_get_check(info, child, 1);
1484         if (rc)
1485                 GOTO(out_child, result = rc);
1486
1487         if (result == -ENOENT) {
1488                 /* Create under OBF and .lustre is not permitted */
1489                 if (!fid_is_md_operative(rr->rr_fid1))
1490                         GOTO(out_child, result = -EPERM);
1491
1492                 /* save versions in reply */
1493                 mdt_version_get_save(info, parent, 0);
1494                 mdt_version_get_save(info, child, 1);
1495
1496                 /* version of child will be changed */
1497                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1498
1499                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1500                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1501
1502                 /* Let lower layers know what is lock mode on directory. */
1503                 info->mti_spec.sp_cr_mode =
1504                         mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
1505
1506                 /* Don't do lookup sanity check. We know name doesn't exist. */
1507                 info->mti_spec.sp_cr_lookup = 0;
1508                 info->mti_spec.sp_feat = &dt_directory_features;
1509
1510                 result = mdo_create(info->mti_env, mdt_object_child(parent),
1511                                     &rr->rr_name, mdt_object_child(child),
1512                                     &info->mti_spec, &info->mti_attr);
1513                 if (result == -ERESTART) {
1514                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1515                         GOTO(out_child, result);
1516                 } else {
1517                         mdt_prep_ma_buf_from_rep(info, child, ma);
1518                         /* XXX: we should call this once, see few lines below */
1519                         if (result == 0)
1520                                 result = mdt_attr_get_complex(info, child, ma);
1521
1522                         if (result != 0)
1523                                 GOTO(out_child, result);
1524                 }
1525                 created = 1;
1526                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
1527         } else {
1528                 /*
1529                  * The object is on remote node, return its FID for remote open.
1530                  */
1531                 if (mdt_object_remote(child)) {
1532                         /*
1533                          * Check if this lock already was sent to client and
1534                          * this is resent case. For resent case do not take lock
1535                          * again, use what is already granted.
1536                          */
1537                         LASSERT(lhc != NULL);
1538
1539                         rc = mdt_check_resent_lock(info, child, lhc);
1540                         if (rc < 0) {
1541                                 GOTO(out_child, result = rc);
1542                         } else if (rc > 0) {
1543                                 mdt_lock_handle_init(lhc);
1544                                 mdt_lock_reg_init(lhc, LCK_PR);
1545
1546                                 rc = mdt_object_lock(info, child, lhc,
1547                                                      MDS_INODELOCK_LOOKUP);
1548                         }
1549                         repbody->mbo_fid1 = *mdt_object_fid(child);
1550                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1551                         if (rc != 0)
1552                                 result = rc;
1553                         else
1554                                 result = -MDT_EREMOTE_OPEN;
1555                         GOTO(out_child, result);
1556                 } else if (mdt_object_exists(child)) {
1557                         /* We have to get attr & LOV EA & HSM for this
1558                          * object. */
1559                         mdt_prep_ma_buf_from_rep(info, child, ma);
1560                         ma->ma_need |= MA_HSM;
1561                         result = mdt_attr_get_complex(info, child, ma);
1562                         if (result != 0)
1563                                 GOTO(out_child, result);
1564                 } else {
1565                         /* Object does not exist. Likely FS corruption. */
1566                         CERROR("%s: name '"DNAME"' present, but FID "
1567                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1568                                PNAME(&rr->rr_name), PFID(child_fid));
1569                         GOTO(out_child, result = -EIO);
1570                 }
1571         }
1572
1573         mdt_pack_secctx_in_reply(info, child);
1574
1575         rc = mdt_check_resent_lock(info, child, lhc);
1576         if (rc < 0) {
1577                 GOTO(out_child, result = rc);
1578         } else if (rc == 0) {
1579                 /* the open lock might already be gotten in
1580                  * ldlm_handle_enqueue() */
1581                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1582                 if (open_flags & MDS_OPEN_LOCK)
1583                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1584         } else {
1585                 /* get openlock if this isn't replay and client requested it */
1586                 if (!req_is_replay(req)) {
1587                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1588                         object_locked = 1;
1589                         if (rc != 0)
1590                                 GOTO(out_child_unlock, result = rc);
1591                         else if (open_flags & MDS_OPEN_LOCK)
1592                                 mdt_set_disposition(info, ldlm_rep,
1593                                                     DISP_OPEN_LOCK);
1594                 }
1595         }
1596         /* Try to open it now. */
1597         rc = mdt_finish_open(info, parent, child, open_flags,
1598                              created, ldlm_rep);
1599         if (rc) {
1600                 result = rc;
1601                 /* openlock will be released if mdt_finish_open() failed */
1602                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1603
1604                 if (created && (open_flags & MDS_OPEN_VOLATILE)) {
1605                         CERROR("%s: cannot open volatile file "DFID", orphan "
1606                                "file will be left in PENDING directory until "
1607                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1608                                PFID(mdt_object_fid(child)), rc);
1609                         GOTO(out_child_unlock, result);
1610                 }
1611
1612                 if (created) {
1613                         ma->ma_need = 0;
1614                         ma->ma_valid = 0;
1615                         rc = mdo_unlink(info->mti_env,
1616                                         mdt_object_child(parent),
1617                                         mdt_object_child(child),
1618                                         &rr->rr_name,
1619                                         &info->mti_attr, 0);
1620                         if (rc != 0)
1621                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1622                                        mdt_obd_name(info->mti_mdt),
1623                                        PFID(mdt_object_fid(child)), rc);
1624                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1625                 }
1626         }
1627         EXIT;
1628 out_child_unlock:
1629         if (object_locked)
1630                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1631 out_child:
1632         mdt_object_put(info->mti_env, child);
1633         if (result == 0)
1634                 mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh);
1635 out_parent:
1636         mdt_object_unlock_put(info, parent, lh, result || !created);
1637 out:
1638         if (result)
1639                 lustre_msg_set_transno(req->rq_repmsg, 0);
1640         return result;
1641 }
1642
1643 /**
1644  * Create an orphan object use local root.
1645  */
1646 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1647                                           struct mdt_device *mdt,
1648                                           const struct lu_fid *fid,
1649                                           struct md_attr *attr, fmode_t fmode)
1650 {
1651         const struct lu_env *env = info->mti_env;
1652         struct md_op_spec *spec = &info->mti_spec;
1653         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1654         struct mdt_object *obj = NULL;
1655         struct mdt_object *local_root;
1656         static const struct lu_name lname = {
1657                 .ln_name = "i_am_nobody",
1658                 .ln_namelen = sizeof("i_am_nobody") - 1,
1659         };
1660         struct lu_ucred *uc;
1661         cfs_cap_t uc_cap_save;
1662         int rc;
1663         ENTRY;
1664
1665         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1666         if (rc != 0)
1667                 RETURN(ERR_PTR(rc));
1668
1669         local_root = mdt_object_find(env, mdt, local_root_fid);
1670         if (IS_ERR(local_root))
1671                 RETURN(local_root);
1672
1673         obj = mdt_object_new(env, mdt, fid);
1674         if (IS_ERR(obj))
1675                 GOTO(out, rc = PTR_ERR(obj));
1676
1677         spec->sp_cr_lookup = 0;
1678         spec->sp_feat = &dt_directory_features;
1679         spec->sp_cr_mode = MDL_MINMODE; /* no lock */
1680         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1681         if (attr->ma_valid & MA_LOV) {
1682                 spec->u.sp_ea.eadata = attr->ma_lmm;
1683                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1684                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1685         } else {
1686                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1687         }
1688
1689         uc = lu_ucred(env);
1690         uc_cap_save = uc->uc_cap;
1691         uc->uc_cap |= 1 << CFS_CAP_DAC_OVERRIDE;
1692         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1693                         mdt_object_child(obj), spec, attr);
1694         uc->uc_cap = uc_cap_save;
1695         if (rc < 0) {
1696                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1697                        mdt_obd_name(mdt), PFID(fid), rc);
1698                 GOTO(out, rc);
1699         }
1700
1701         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED);
1702         if (rc < 0)
1703                 CERROR("%s: cannot open volatile file "DFID", orphan "
1704                        "file will be left in PENDING directory until "
1705                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1706                        PFID(fid), rc);
1707         GOTO(out, rc);
1708
1709 out:
1710         if (rc < 0) {
1711                 if (!IS_ERR(obj))
1712                         mdt_object_put(env, obj);
1713                 obj = ERR_PTR(rc);
1714         }
1715         mdt_object_put(env, local_root);
1716         return obj;
1717 }
1718
1719 /* XXX Look into layout in MDT layer. */
1720 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1721 {
1722         struct lov_comp_md_v1   *comp_v1;
1723         struct lov_mds_md       *v1;
1724         __u32   off;
1725         int     i;
1726
1727         if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1728                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1729
1730                 if (comp_v1->lcm_entry_count == 0)
1731                         return -EINVAL;
1732
1733                 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1734                         off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1735                         v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1736                         v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1737                 }
1738         } else {
1739                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1740         }
1741         return 0;
1742 }
1743
1744 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1745                            struct md_attr *ma)
1746 {
1747         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1748         struct lu_ucred        *uc = mdt_ucred(info);
1749         struct close_data      *data;
1750         struct ldlm_lock       *lease;
1751         struct mdt_object      *orphan;
1752         struct md_attr         *orp_ma;
1753         struct lu_buf          *buf;
1754         cfs_cap_t               cap;
1755         bool                    lease_broken;
1756         int                     rc;
1757         int                     rc2;
1758         ENTRY;
1759
1760         if (mdt_rdonly(info->mti_exp))
1761                 RETURN(-EROFS);
1762
1763         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1764         if (data == NULL)
1765                 RETURN(-EPROTO);
1766
1767         lease = ldlm_handle2lock(&data->cd_handle);
1768         if (lease == NULL)
1769                 RETURN(-ESTALE);
1770
1771         /* try to hold open_sem so that nobody else can open the file */
1772         if (!down_write_trylock(&o->mot_open_sem)) {
1773                 ldlm_lock_cancel(lease);
1774                 GOTO(out_reprocess, rc = -EBUSY);
1775         }
1776
1777         /* Check if the lease open lease has already canceled */
1778         lock_res_and_lock(lease);
1779         lease_broken = ldlm_is_cancel(lease);
1780         unlock_res_and_lock(lease);
1781
1782         LDLM_DEBUG(lease, DFID " lease broken? %d",
1783                    PFID(mdt_object_fid(o)), lease_broken);
1784
1785         /* Cancel server side lease. Client side counterpart should
1786          * have been cancelled. It's okay to cancel it now as we've
1787          * held mot_open_sem. */
1788         ldlm_lock_cancel(lease);
1789
1790         if (lease_broken) /* don't perform release task */
1791                 GOTO(out_unlock, rc = -ESTALE);
1792
1793         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1794                 GOTO(out_unlock, rc = -EINVAL);
1795
1796         /* ma_need was set before but it seems fine to change it in order to
1797          * avoid modifying the one from RPC */
1798         ma->ma_need = MA_HSM;
1799         rc = mdt_attr_get_complex(info, o, ma);
1800         if (rc != 0)
1801                 GOTO(out_unlock, rc);
1802
1803         if (!mdt_hsm_release_allow(ma))
1804                 GOTO(out_unlock, rc = -EPERM);
1805
1806         /* already released? */
1807         if (ma->ma_hsm.mh_flags & HS_RELEASED)
1808                 GOTO(out_unlock, rc = 0);
1809
1810         /* Compare on-disk and packed data_version */
1811         if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1812                 CDEBUG(D_HSM, DFID" data_version mismatches: packed=%llu"
1813                        " and on-disk=%llu\n", PFID(mdt_object_fid(o)),
1814                        data->cd_data_version, ma->ma_hsm.mh_arch_ver);
1815                 GOTO(out_unlock, rc = -EPERM);
1816         }
1817
1818         ma->ma_valid = MA_INODE;
1819         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1820         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1821         if (rc < 0)
1822                 GOTO(out_unlock, rc);
1823
1824         mutex_lock(&o->mot_som_mutex);
1825         rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size,
1826                            ma->ma_attr.la_blocks);
1827         mutex_unlock(&o->mot_som_mutex);
1828         if (rc2 < 0)
1829                 CDEBUG(D_INODE,
1830                        "%s: File "DFID" SOM update failed: rc = %d\n",
1831                        mdt_obd_name(info->mti_mdt),
1832                        PFID(mdt_object_fid(o)), rc2);
1833
1834
1835         ma->ma_need = MA_INODE | MA_LOV;
1836         rc = mdt_attr_get_complex(info, o, ma);
1837         if (rc < 0)
1838                 GOTO(out_unlock, rc);
1839
1840         if (!(ma->ma_valid & MA_LOV)) {
1841                 /* Even empty file are released */
1842                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1843                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
1844                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1845                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1846                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1847         } else {
1848                 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
1849                  * ma_lmm as lov_user_md, then it will be confused by union of
1850                  * layout_gen and stripe_offset. */
1851                 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
1852                     LOV_MAGIC_MAGIC)
1853                         ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1854                 else
1855                         GOTO(out_unlock, rc = -EINVAL);
1856         }
1857
1858         /* Set file as released. */
1859         rc = mdt_hsm_set_released(ma->ma_lmm);
1860         if (rc)
1861                 GOTO(out_unlock, rc);
1862
1863         orp_ma = &info->mti_u.hsm.attr;
1864         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1865         /* We use root ownership to bypass potential quota
1866          * restrictions on the user and group of the file. */
1867         orp_ma->ma_attr.la_uid = 0;
1868         orp_ma->ma_attr.la_gid = 0;
1869         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1870         orp_ma->ma_lmm = ma->ma_lmm;
1871         orp_ma->ma_lmm_size = ma->ma_lmm_size;
1872         orp_ma->ma_valid = MA_INODE | MA_LOV;
1873         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1874                                  MDS_FMODE_WRITE);
1875         if (IS_ERR(orphan)) {
1876                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1877                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1878                        PTR_ERR(orphan));
1879                 GOTO(out_unlock, rc = PTR_ERR(orphan));
1880         }
1881
1882         /* Set up HSM attribute for orphan object */
1883         CLASSERT(sizeof(struct hsm_attrs) <= sizeof(info->mti_xattr_buf));
1884         buf = &info->mti_buf;
1885         buf->lb_buf = info->mti_xattr_buf;
1886         buf->lb_len = sizeof(struct hsm_attrs);
1887         ma->ma_hsm.mh_flags |= HS_RELEASED;
1888         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1889         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
1890
1891         mdt_lock_reg_init(lh, LCK_EX);
1892         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
1893                              MDS_INODELOCK_XATTR);
1894         if (rc != 0)
1895                 GOTO(out_close, rc);
1896
1897         /* The orphan has root ownership so we need to raise
1898          * CAP_FOWNER to set the HSM attributes. */
1899         cap = uc->uc_cap;
1900         uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
1901         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
1902                           XATTR_NAME_HSM, 0);
1903         uc->uc_cap = cap;
1904         if (rc != 0)
1905                 GOTO(out_layout_lock, rc);
1906
1907         /* Swap layout with orphan objects. */
1908         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
1909                              mdt_object_child(orphan),
1910                              SWAP_LAYOUTS_MDS_HSM);
1911         EXIT;
1912
1913 out_layout_lock:
1914         /* Release exclusive LL */
1915         mdt_object_unlock(info, o, lh, 1);
1916 out_close:
1917         /* Close orphan object anyway */
1918         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
1919                        MDS_FMODE_WRITE);
1920         if (rc2 < 0)
1921                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
1922                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
1923         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
1924                         "object closed");
1925         mdt_object_put(info->mti_env, orphan);
1926
1927 out_unlock:
1928         up_write(&o->mot_open_sem);
1929
1930         /* already released */
1931         if (rc == 0) {
1932                 struct mdt_body *repbody;
1933
1934                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1935                 LASSERT(repbody != NULL);
1936                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1937         }
1938
1939 out_reprocess:
1940         ldlm_reprocess_all(lease->l_resource);
1941         LDLM_LOCK_PUT(lease);
1942
1943         ma->ma_valid = 0;
1944         ma->ma_need = 0;
1945
1946         return rc;
1947 }
1948
1949 int mdt_close_handle_layouts(struct mdt_thread_info *info,
1950                              struct mdt_object *o, struct md_attr *ma)
1951 {
1952         struct mdt_lock_handle  *lh1 = &info->mti_lh[MDT_LH_NEW];
1953         struct mdt_lock_handle  *lh2 = &info->mti_lh[MDT_LH_OLD];
1954         struct close_data       *data;
1955         struct ldlm_lock        *lease;
1956         struct mdt_object       *o1 = o, *o2;
1957         bool                     lease_broken;
1958         bool                     swap_objects;
1959         int                      rc;
1960         ENTRY;
1961
1962         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
1963                 RETURN(-EROFS);
1964
1965         if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
1966                 RETURN(-EINVAL);
1967
1968         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1969         if (data == NULL)
1970                 RETURN(-EPROTO);
1971
1972         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1973                 RETURN(-EINVAL);
1974
1975         rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
1976         if (unlikely(rc == 0))
1977                 RETURN(-EINVAL);
1978
1979         /* Exchange o1 and o2, to enforce locking order */
1980         swap_objects = (rc < 0);
1981
1982         lease = ldlm_handle2lock(&data->cd_handle);
1983         if (lease == NULL)
1984                 RETURN(-ESTALE);
1985
1986         o2 = mdt_object_find(info->mti_env, info->mti_mdt, &data->cd_fid);
1987         if (IS_ERR(o2))
1988                 GOTO(out_lease, rc = PTR_ERR(o2));
1989
1990         if (!S_ISREG(lu_object_attr(&o2->mot_obj))) {
1991                 swap_objects = false; /* not swapped yet */
1992                 GOTO(out_obj, rc = -EINVAL);
1993         }
1994
1995         if (swap_objects)
1996                 swap(o1, o2);
1997
1998         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
1999                            MAY_WRITE);
2000         if (rc < 0)
2001                 GOTO(out_obj, rc);
2002
2003         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2), NULL,
2004                            MAY_WRITE);
2005         if (rc < 0)
2006                 GOTO(out_obj, rc);
2007
2008         /* try to hold open_sem so that nobody else can open the file */
2009         if (!down_write_trylock(&o->mot_open_sem)) {
2010                 ldlm_lock_cancel(lease);
2011                 GOTO(out_obj, rc = -EBUSY);
2012         }
2013
2014         /* Check if the lease open lease has already canceled */
2015         lock_res_and_lock(lease);
2016         lease_broken = ldlm_is_cancel(lease);
2017         unlock_res_and_lock(lease);
2018
2019         LDLM_DEBUG(lease, DFID " lease broken? %d",
2020                    PFID(mdt_object_fid(o)), lease_broken);
2021
2022         /* Cancel server side lease. Client side counterpart should
2023          * have been cancelled. It's okay to cancel it now as we've
2024          * held mot_open_sem. */
2025         ldlm_lock_cancel(lease);
2026
2027         if (lease_broken)
2028                 GOTO(out_unlock_sem, rc = -ESTALE);
2029
2030         mdt_lock_reg_init(lh1, LCK_EX);
2031         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
2032                              MDS_INODELOCK_XATTR);
2033         if (rc < 0)
2034                 GOTO(out_unlock_sem, rc);
2035
2036         mdt_lock_reg_init(lh2, LCK_EX);
2037         rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2038                              MDS_INODELOCK_XATTR);
2039         if (rc < 0)
2040                 GOTO(out_unlock1, rc);
2041
2042         /* Swap layout with orphan object */
2043         if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2044                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2045                                      mdt_object_child(o2), 0);
2046         } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2047                    ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2048                 struct lu_buf *buf = &info->mti_buf;
2049                 struct md_rejig_data mrd;
2050
2051                 mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2052                 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)
2053                         mrd.mrd_mirror_id = data->cd_mirror_id;
2054
2055                 buf->lb_len = sizeof(mrd);
2056                 buf->lb_buf = &mrd;
2057                 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2058                                   XATTR_LUSTRE_LOV,
2059                                   ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2060                                   LU_XATTR_SPLIT : LU_XATTR_MERGE);
2061                 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2062                         int rc2;
2063
2064                         mutex_lock(&o->mot_som_mutex);
2065                         rc2 = mdt_set_som(info, o, SOM_FL_STRICT,
2066                                           ma->ma_attr.la_size,
2067                                           ma->ma_attr.la_blocks);
2068                         mutex_unlock(&o->mot_som_mutex);
2069                         if (rc2 < 0)
2070                                 CERROR(DFID": Setting i_blocks error: %d, "
2071                                        "i_blocks will be reported wrongly and "
2072                                        "can only be fixed in next resync\n",
2073                                        PFID(mdt_object_fid(o)), rc2);
2074                 }
2075         }
2076         if (rc < 0)
2077                 GOTO(out_unlock2, rc);
2078
2079         EXIT;
2080
2081 out_unlock2:
2082         /* Release exclusive LL */
2083         mdt_object_unlock(info, o2, lh2, 1);
2084
2085 out_unlock1:
2086         mdt_object_unlock(info, o1, lh1, 1);
2087
2088 out_unlock_sem:
2089         up_write(&o->mot_open_sem);
2090
2091         /* already swapped */
2092         if (rc == 0) {
2093                 struct mdt_body *repbody;
2094
2095                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2096                 LASSERT(repbody != NULL);
2097                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2098         }
2099
2100 out_obj:
2101         mdt_object_put(info->mti_env, swap_objects ? o1 : o2);
2102
2103         ldlm_reprocess_all(lease->l_resource);
2104
2105 out_lease:
2106         LDLM_LOCK_PUT(lease);
2107
2108         if (ma != NULL) {
2109                 ma->ma_valid = 0;
2110                 ma->ma_need = 0;
2111         }
2112
2113         return rc;
2114 }
2115
2116 static int mdt_close_resync_done(struct mdt_thread_info *info,
2117                                  struct mdt_object *o, struct md_attr *ma)
2118 {
2119         struct close_data       *data;
2120         struct ldlm_lock        *lease;
2121         struct md_layout_change  layout = { 0 };
2122         __u32                   *resync_ids = NULL;
2123         size_t                   resync_count = 0;
2124         bool                     lease_broken;
2125         int                      rc;
2126         ENTRY;
2127
2128         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2129                 RETURN(-EROFS);
2130
2131         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2132                 RETURN(-EINVAL);
2133
2134         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2135         if (data == NULL)
2136                 RETURN(-EPROTO);
2137
2138         if (ptlrpc_req_need_swab(mdt_info_req(info)))
2139                 lustre_swab_close_data_resync_done(&data->cd_resync);
2140
2141         if (!fid_is_zero(&data->cd_fid))
2142                 RETURN(-EPROTO);
2143
2144         lease = ldlm_handle2lock(&data->cd_handle);
2145         if (lease == NULL)
2146                 RETURN(-ESTALE);
2147
2148         /* try to hold open_sem so that nobody else can open the file */
2149         if (!down_write_trylock(&o->mot_open_sem)) {
2150                 ldlm_lock_cancel(lease);
2151                 GOTO(out_reprocess, rc = -EBUSY);
2152         }
2153
2154         /* Check if the lease open lease has already canceled */
2155         lock_res_and_lock(lease);
2156         lease_broken = ldlm_is_cancel(lease);
2157         unlock_res_and_lock(lease);
2158
2159         LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2160                    PFID(mdt_object_fid(o)), lease_broken);
2161
2162         /* Cancel server side lease. Client side counterpart should
2163          * have been cancelled. It's okay to cancel it now as we've
2164          * held mot_open_sem. */
2165         ldlm_lock_cancel(lease);
2166
2167         if (lease_broken) /* don't perform release task */
2168                 GOTO(out_unlock, rc = -ESTALE);
2169
2170         resync_count = data->cd_resync.resync_count;
2171
2172         if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2173                 void *data;
2174
2175                 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2176                                            RCL_CLIENT))
2177                         GOTO(out_unlock, rc = -EPROTO);
2178
2179                 OBD_ALLOC(resync_ids, resync_count * sizeof(__u32));
2180                 if (!resync_ids)
2181                         GOTO(out_unlock, rc = -ENOMEM);
2182
2183                 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2184                 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2185
2186                 layout.mlc_resync_ids = resync_ids;
2187         } else {
2188                 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2189         }
2190
2191         layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2192         layout.mlc_resync_count = resync_count;
2193         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2194                 layout.mlc_som.lsa_valid = SOM_FL_STRICT;
2195                 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2196                 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2197         }
2198         rc = mdt_layout_change(info, o, &layout);
2199         if (rc)
2200                 GOTO(out_unlock, rc);
2201
2202         EXIT;
2203
2204 out_unlock:
2205         up_write(&o->mot_open_sem);
2206
2207         /* already released */
2208         if (rc == 0) {
2209                 struct mdt_body *repbody;
2210
2211                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2212                 LASSERT(repbody != NULL);
2213                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2214         }
2215
2216         if (resync_ids)
2217                 OBD_FREE(resync_ids, resync_count * sizeof(__u32));
2218
2219 out_reprocess:
2220         ldlm_reprocess_all(lease->l_resource);
2221         LDLM_LOCK_PUT(lease);
2222
2223         ma->ma_valid = 0;
2224         ma->ma_need = 0;
2225
2226         return rc;
2227 }
2228
2229 #define MFD_CLOSED(open_flags) ((open_flags) == MDS_FMODE_CLOSED)
2230
2231 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2232 {
2233         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_open_flags));
2234 }
2235
2236 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2237 {
2238         struct mdt_object *o = mfd->mfd_object;
2239         struct md_object *next = mdt_object_child(o);
2240         struct md_attr *ma = &info->mti_attr;
2241         struct lu_fid *ofid = &info->mti_tmp_fid1;
2242         int rc = 0;
2243         u64 open_flags;
2244         u64 intent;
2245         bool discard = false;
2246
2247         ENTRY;
2248
2249         open_flags = mfd->mfd_open_flags;
2250         intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2251         *ofid = *mdt_object_fid(o);
2252
2253         CDEBUG(D_INODE, "%s: close file "DFID" with intent: %llx\n",
2254                mdt_obd_name(info->mti_mdt), PFID(ofid), intent);
2255
2256         switch (intent) {
2257         case MDS_HSM_RELEASE: {
2258                 rc = mdt_hsm_release(info, o, ma);
2259                 if (rc < 0) {
2260                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2261                                mdt_obd_name(info->mti_mdt),
2262                                PFID(ofid), rc);
2263                         /* continue to close even error occurred. */
2264                 }
2265                 break;
2266         }
2267         case MDS_CLOSE_LAYOUT_MERGE:
2268         case MDS_CLOSE_LAYOUT_SPLIT:
2269         case MDS_CLOSE_LAYOUT_SWAP: {
2270                 rc = mdt_close_handle_layouts(info, o, ma);
2271                 if (rc < 0) {
2272                         CDEBUG(D_INODE,
2273                                "%s: cannot swap layout of "DFID": rc = %d\n",
2274                                mdt_obd_name(info->mti_mdt),
2275                                PFID(ofid), rc);
2276                         /* continue to close even if error occurred. */
2277                 }
2278                 break;
2279         }
2280         case MDS_CLOSE_RESYNC_DONE:
2281                 rc = mdt_close_resync_done(info, o, ma);
2282                 break;
2283         default:
2284                 /* nothing */
2285                 break;
2286         }
2287
2288         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
2289             ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) {
2290                 int rc2;
2291
2292                 rc2 = mdt_lsom_update(info, o, false);
2293                 if (rc2 < 0)
2294                         CDEBUG(D_INODE,
2295                                "%s: File " DFID " LSOM failed: rc = %d\n",
2296                                mdt_obd_name(info->mti_mdt),
2297                                PFID(ofid), rc2);
2298                         /* continue to close even if error occured. */
2299         }
2300
2301         if (open_flags & MDS_FMODE_WRITE)
2302                 mdt_write_put(o);
2303         else if (open_flags & MDS_FMODE_EXEC)
2304                 mdt_write_allow(o);
2305
2306         /* Update atime on close only. */
2307         if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ ||
2308              open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
2309             (ma->ma_attr.la_valid & LA_ATIME)) {
2310                 /* Set the atime only. */
2311                 ma->ma_valid = MA_INODE;
2312                 ma->ma_attr.la_valid = LA_ATIME;
2313                 rc = mo_attr_set(info->mti_env, next, ma);
2314         }
2315
2316         /* If file data is modified, add the dirty flag. */
2317         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2318                 rc = mdt_add_dirty_flag(info, o, ma);
2319
2320         ma->ma_need |= MA_INODE;
2321         ma->ma_valid &= ~MA_INODE;
2322
2323         LASSERT(atomic_read(&o->mot_open_count) > 0);
2324         atomic_dec(&o->mot_open_count);
2325         mdt_handle_last_unlink(info, o, ma);
2326
2327         if (!MFD_CLOSED(open_flags)) {
2328                 rc = mo_close(info->mti_env, next, ma, open_flags);
2329                 discard = mdt_dom_check_for_discard(info, o);
2330         }
2331
2332         /* adjust open and lease count */
2333         if (open_flags & MDS_OPEN_LEASE) {
2334                 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2335                 atomic_dec(&o->mot_lease_count);
2336         }
2337
2338         mdt_mfd_free(mfd);
2339         mdt_object_put(info->mti_env, o);
2340
2341         if (discard)
2342                 mdt_dom_discard_data(info, ofid);
2343
2344         RETURN(rc);
2345 }
2346
2347 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2348                        struct mdt_body *repbody)
2349 {
2350         struct mdt_export_data *med;
2351         struct mdt_file_data   *mfd;
2352         int                     ret = 0;
2353         int                     rc = 0;
2354         ENTRY;
2355
2356         med = &req->rq_export->exp_mdt_data;
2357         spin_lock(&med->med_open_lock);
2358         mfd = mdt_open_handle2mfd(med, &info->mti_open_handle,
2359                                   req_is_replay(req));
2360         if (mdt_mfd_closed(mfd)) {
2361                 spin_unlock(&med->med_open_lock);
2362                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2363                        ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2364                        info->mti_open_handle.cookie);
2365                 /** not serious error since bug 3633 */
2366                 rc = -ESTALE;
2367         } else {
2368                 class_handle_unhash(&mfd->mfd_open_handle);
2369                 list_del_init(&mfd->mfd_list);
2370                 spin_unlock(&med->med_open_lock);
2371                 ret = mdt_mfd_close(info, mfd);
2372         }
2373
2374         RETURN(rc ? rc : ret);
2375 }
2376
2377 int mdt_close(struct tgt_session_info *tsi)
2378 {
2379         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2380         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2381         struct md_attr         *ma = &info->mti_attr;
2382         struct mdt_body        *repbody = NULL;
2383         int rc, ret = 0;
2384         ENTRY;
2385
2386         mdt_counter_incr(req, LPROC_MDT_CLOSE);
2387         /* Close may come with the Size-on-MDS update. Unpack it. */
2388         rc = mdt_close_unpack(info);
2389         if (rc)
2390                 GOTO(out, rc = err_serious(rc));
2391
2392         /* These fields are no longer used and are left for compatibility.
2393          * size is always zero */
2394         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2395                              0);
2396         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2397                              0);
2398         rc = req_capsule_server_pack(info->mti_pill);
2399         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2400                 mdt_client_compatibility(info);
2401                 if (rc == 0)
2402                         mdt_fix_reply(info);
2403                 mdt_exit_ucred(info);
2404                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2405         }
2406
2407         /* Continue to close handle even if we can not pack reply */
2408         if (rc == 0) {
2409                 repbody = req_capsule_server_get(info->mti_pill,
2410                                                  &RMF_MDT_BODY);
2411                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2412                                                     &RMF_MDT_MD);
2413                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2414                                                        &RMF_MDT_MD,
2415                                                        RCL_SERVER);
2416                 ma->ma_need = MA_INODE | MA_LOV;
2417                 repbody->mbo_eadatasize = 0;
2418                 repbody->mbo_aclsize = 0;
2419         } else {
2420                 rc = err_serious(rc);
2421         }
2422
2423         rc = mdt_close_internal(info, req, repbody);
2424         if (rc != -ESTALE)
2425                 mdt_empty_transno(info, rc);
2426
2427         if (repbody != NULL) {
2428                 mdt_client_compatibility(info);
2429                 rc = mdt_fix_reply(info);
2430         }
2431
2432         mdt_exit_ucred(info);
2433         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2434                 GOTO(out, rc = err_serious(-ENOMEM));
2435
2436         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2437                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2438                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2439 out:
2440         mdt_thread_info_fini(info);
2441         RETURN(rc ? rc : ret);
2442 }