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  *
  31  * lustre/mdt/mdt_open.c
  32  *
  33  * Lustre Metadata Target (mdt) open/close file handling
  34  *
  35  * Author: Huang Hua <huanghua@clusterfs.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_MDS
  39
  40 #include <lustre_acl.h>
  41 #include <lustre_mds.h>
  42 #include <lustre_swab.h>
  43 #include "mdt_internal.h"
  44 #include <lustre_nodemap.h>
  45
  46 static const char mfd_open_handle_owner[] = "mdt";
  47
  48 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep,
  49                            struct mdt_lock_handle *lhc);
  50
  51 /* Create a new mdt_file_data struct, initialize it,
  52  * and insert it to global hash table */
  53 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
  54 {
  55         struct mdt_file_data *mfd;
  56
  57         ENTRY;
  58         OBD_ALLOC_PTR(mfd);
  59         if (mfd != NULL) {
  60                 refcount_set(&mfd->mfd_open_handle.h_ref, 1);
  61                 INIT_HLIST_NODE(&mfd->mfd_open_handle.h_link);
  62                 mfd->mfd_owner = med;
  63                 INIT_LIST_HEAD(&mfd->mfd_list);
  64                 class_handle_hash(&mfd->mfd_open_handle, mfd_open_handle_owner);
  65         }
  66
  67         RETURN(mfd);
  68 }
  69
  70 /*
  71  * Find the mfd pointed to by handle in global hash table.
  72  * In case of replay the handle is obsoleted
  73  * but mfd can be found in mfd list by that handle.
  74  * Callers need to be holding med_open_lock.
  75  */
  76 struct mdt_file_data *mdt_open_handle2mfd(struct mdt_export_data *med,
  77                                         const struct lustre_handle *open_handle,
  78                                         bool is_replay_or_resent)
  79 {
  80         struct mdt_file_data   *mfd;
  81
  82         ENTRY;
  83         LASSERT(open_handle != NULL);
  84         mfd = class_handle2object(open_handle->cookie, mfd_open_handle_owner);
  85         if (mfd)
  86                 refcount_dec(&mfd->mfd_open_handle.h_ref);
  87
  88         /* during dw/setattr replay the mfd can be found by old handle */
  89         if ((!mfd || mfd->mfd_owner != med) && is_replay_or_resent) {
  90                 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  91                         if (mfd->mfd_open_handle_old.cookie ==
  92                             open_handle->cookie)
  93                                 RETURN(mfd);
  94                 }
  95                 mfd = NULL;
  96         }
  97
  98         RETURN(mfd);
  99 }
 100
 101 /* free mfd */
 102 void mdt_mfd_free(struct mdt_file_data *mfd)
 103 {
 104         LASSERT(refcount_read(&mfd->mfd_open_handle.h_ref) == 1);
 105         LASSERT(list_empty(&mfd->mfd_list));
 106         OBD_FREE_PRE(mfd, sizeof(*mfd), "kfree_rcu");
 107         kfree_rcu(mfd, mfd_open_handle.h_rcu);
 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
 117         ENTRY;
 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
 143         ENTRY;
 144         spin_lock(&o->mot_write_lock);
 145         rc = o->mot_write_count;
 146         spin_unlock(&o->mot_write_lock);
 147         RETURN(rc);
 148 }
 149
 150 int mdt_write_get(struct mdt_object *o)
 151 {
 152         int rc = 0;
 153
 154         ENTRY;
 155         spin_lock(&o->mot_write_lock);
 156         if (o->mot_write_count < 0)
 157                 rc = -ETXTBSY;
 158         else
 159                 o->mot_write_count++;
 160         spin_unlock(&o->mot_write_lock);
 161
 162         RETURN(rc);
 163 }
 164
 165 void mdt_write_put(struct mdt_object *o)
 166 {
 167         ENTRY;
 168         spin_lock(&o->mot_write_lock);
 169         o->mot_write_count--;
 170         spin_unlock(&o->mot_write_lock);
 171         EXIT;
 172 }
 173
 174 static int mdt_write_deny(struct mdt_object *o)
 175 {
 176         int rc = 0;
 177
 178         ENTRY;
 179         spin_lock(&o->mot_write_lock);
 180         if (o->mot_write_count > 0)
 181                 rc = -ETXTBSY;
 182         else
 183                 o->mot_write_count--;
 184         spin_unlock(&o->mot_write_lock);
 185         RETURN(rc);
 186 }
 187
 188 static void mdt_write_allow(struct mdt_object *o)
 189 {
 190         ENTRY;
 191         spin_lock(&o->mot_write_lock);
 192         o->mot_write_count++;
 193         spin_unlock(&o->mot_write_lock);
 194         EXIT;
 195 }
 196
 197 /* there can be no real transaction so prepare the fake one */
 198 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 199 {
 200         struct mdt_device *mdt = info->mti_mdt;
 201         struct ptlrpc_request *req = mdt_info_req(info);
 202         struct tg_export_data *ted;
 203         struct lsd_client_data *lcd;
 204
 205         ENTRY;
 206         if (mdt_rdonly(req->rq_export))
 207                 RETURN_EXIT;
 208
 209         /* transaction has occurred already */
 210         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 211                 RETURN_EXIT;
 212
 213         if (tgt_is_multimodrpcs_client(req->rq_export)) {
 214                 struct thandle *th;
 215
 216                 /* generate an empty transaction to get a transno
 217                  * and reply data */
 218                 th = dt_trans_create(info->mti_env, mdt->mdt_bottom);
 219                 if (!IS_ERR(th)) {
 220                         rc = dt_trans_start(info->mti_env, mdt->mdt_bottom, th);
 221                         dt_trans_stop(info->mti_env, mdt->mdt_bottom, th);
 222                 }
 223                 RETURN_EXIT;
 224         }
 225
 226         spin_lock(&mdt->mdt_lut.lut_translock);
 227         if (rc != 0) {
 228                 if (info->mti_transno != 0) {
 229                         struct obd_export *exp = req->rq_export;
 230
 231                         CERROR("%s: replay trans %llu NID %s: rc = %d\n",
 232                                mdt_obd_name(mdt), info->mti_transno,
 233                                obd_export_nid2str(exp), rc);
 234                         spin_unlock(&mdt->mdt_lut.lut_translock);
 235                         RETURN_EXIT;
 236                 }
 237         } else if (info->mti_transno == 0) {
 238                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
 239         } else {
 240                 /* should be replay */
 241                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 242                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 243         }
 244         spin_unlock(&mdt->mdt_lut.lut_translock);
 245
 246         CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
 247                info->mti_transno,
 248                req->rq_export->exp_obd->obd_last_committed);
 249
 250         req->rq_transno = info->mti_transno;
 251         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 252
 253         /* update lcd in memory only for resent cases */
 254         ted = &req->rq_export->exp_target_data;
 255         LASSERT(ted);
 256         mutex_lock(&ted->ted_lcd_lock);
 257         lcd = ted->ted_lcd;
 258         if (info->mti_transno < lcd->lcd_last_transno &&
 259             info->mti_transno != 0) {
 260                 /* This should happen during replay. Do not update
 261                  * last rcvd info if replay req transno < last transno,
 262                  * otherwise the following resend(after replay) can not
 263                  * be checked correctly by xid */
 264                 mutex_unlock(&ted->ted_lcd_lock);
 265                 CDEBUG(D_HA, "%s: transno = %llu < last_transno = %llu\n",
 266                        mdt_obd_name(mdt), info->mti_transno,
 267                        lcd->lcd_last_transno);
 268                 RETURN_EXIT;
 269         }
 270
 271         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
 272                 if (info->mti_transno != 0)
 273                         lcd->lcd_last_close_transno = info->mti_transno;
 274                 lcd->lcd_last_close_xid = req->rq_xid;
 275                 lcd->lcd_last_close_result = rc;
 276         } else {
 277                 /* VBR: save versions in last_rcvd for reconstruct. */
 278                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 279
 280                 if (pre_versions) {
 281                         lcd->lcd_pre_versions[0] = pre_versions[0];
 282                         lcd->lcd_pre_versions[1] = pre_versions[1];
 283                         lcd->lcd_pre_versions[2] = pre_versions[2];
 284                         lcd->lcd_pre_versions[3] = pre_versions[3];
 285                 }
 286                 if (info->mti_transno != 0)
 287                         lcd->lcd_last_transno = info->mti_transno;
 288
 289                 lcd->lcd_last_xid = req->rq_xid;
 290                 lcd->lcd_last_result = rc;
 291                 lcd->lcd_last_data = info->mti_opdata;
 292         }
 293         mutex_unlock(&ted->ted_lcd_lock);
 294
 295         EXIT;
 296 }
 297
 298 void mdt_mfd_set_mode(struct mdt_file_data *mfd, u64 open_flags)
 299 {
 300         LASSERT(mfd != NULL);
 301
 302         CDEBUG(D_DENTRY, DFID " Change mfd open_flags %#llo -> %#llo.\n",
 303                PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_open_flags,
 304                open_flags);
 305
 306         mfd->mfd_open_flags = open_flags;
 307 }
 308
 309 /**
 310  * prep ma_lmm/ma_lmv for md_attr from reply
 311  */
 312 static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
 313                                      struct mdt_object *obj,
 314                                      struct md_attr *ma, __u64 open_flags)
 315 {
 316         struct req_capsule *pill;
 317
 318         if (ma->ma_lmv || ma->ma_lmm) {
 319                 CDEBUG(D_INFO, DFID " %s already set.\n",
 320                        PFID(mdt_object_fid(obj)),
 321                        ma->ma_lmv ? (ma->ma_lmm ? "ma_lmv and ma_lmm"
 322                                                 : "ma_lmv")
 323                                   : "ma_lmm");
 324                 return;
 325         }
 326
 327         pill = info->mti_pill;
 328         if (S_ISDIR(obj->mot_header.loh_attr)) {
 329                 ma->ma_lmv = req_capsule_server_get(pill, &RMF_MDT_MD);
 330                 ma->ma_lmv_size = req_capsule_get_size(pill, &RMF_MDT_MD,
 331                                                        RCL_SERVER);
 332                 if (ma->ma_lmv_size > 0)
 333                         ma->ma_need |= MA_LMV;
 334
 335                 if (open_flags & MDS_OPEN_DEFAULT_LMV) {
 336                         ma->ma_default_lmv = req_capsule_server_get(pill,
 337                                                         &RMF_DEFAULT_MDT_MD);
 338                         ma->ma_default_lmv_size = req_capsule_get_size(pill,
 339                                                         &RMF_DEFAULT_MDT_MD,
 340                                                         RCL_SERVER);
 341                         if (ma->ma_default_lmv_size > 0)
 342                                 ma->ma_need |= MA_LMV_DEF;
 343                 }
 344         } else {
 345                 ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
 346                 ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
 347                                                        RCL_SERVER);
 348                 if (ma->ma_lmm_size > 0)
 349                         ma->ma_need |= MA_LOV;
 350         }
 351 }
 352
 353 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 354                         struct mdt_object *o, u64 open_flags, int created,
 355                         struct ldlm_reply *rep)
 356 {
 357         struct ptlrpc_request *req = mdt_info_req(info);
 358         struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 359         struct mdt_file_data *mfd;
 360         struct md_attr *ma  = &info->mti_attr;
 361         struct lu_attr *la  = &ma->ma_attr;
 362         struct mdt_body *repbody;
 363         bool isdir, isreg;
 364         int rc = 0;
 365
 366         ENTRY;
 367         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 368
 369         isreg = S_ISREG(la->la_mode);
 370         isdir = S_ISDIR(la->la_mode);
 371         if (isreg && !(ma->ma_valid & MA_LOV) &&
 372             !(open_flags & MDS_OPEN_RELEASE)) {
 373                 /*
 374                  * No EA, check whether it is will set regEA and dirEA since in
 375                  * above attr get, these size might be zero, so reset it, to
 376                  * retrieve the MD after create obj.
 377                  */
 378                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 379                                                        &RMF_MDT_MD,
 380                                                        RCL_SERVER);
 381                 /* in replay case, p == NULL */
 382                 rc = mdt_create_data(info, p, o);
 383                 if (rc)
 384                         RETURN(rc);
 385
 386                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
 387                         mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
 388         }
 389
 390         CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n",
 391                ma->ma_valid, ma->ma_lmm_size);
 392
 393         if (ma->ma_valid & MA_LOV) {
 394                 LASSERT(ma->ma_lmm_size != 0);
 395                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 396                 if (isdir)
 397                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 398                 else
 399                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 400         }
 401
 402         if (ma->ma_valid & MA_LMV) {
 403                 LASSERT(ma->ma_lmv_size != 0);
 404                 repbody->mbo_eadatasize = ma->ma_lmv_size;
 405                 LASSERT(isdir);
 406                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
 407         }
 408
 409         if (ma->ma_valid & MA_LMV_DEF) {
 410                 LASSERT(ma->ma_default_lmv_size != 0);
 411                 LASSERT(isdir);
 412                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_DEFAULT_MEA;
 413         }
 414
 415         if (open_flags & MDS_FMODE_WRITE)
 416                 rc = mdt_write_get(o);
 417         else if (open_flags & MDS_FMODE_EXEC)
 418                 rc = mdt_write_deny(o);
 419
 420         if (rc)
 421                 RETURN(rc);
 422
 423         rc = mo_open(info->mti_env, mdt_object_child(o),
 424                      created ? open_flags | MDS_OPEN_CREATED : open_flags,
 425                      &info->mti_spec);
 426         if (rc != 0) {
 427                 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
 428                  * client does not know which suppgid should be sent to the MDS,
 429                  * or some other(s) changed the target file's GID after this RPC
 430                  * sent to the MDS with the suppgid as the original GID, then we
 431                  * give the client another chance to send the right suppgid. */
 432                 if (rc == -EACCES &&
 433                     allow_client_chgrp(info, lu_ucred(info->mti_env)))
 434                         mdt_set_disposition(info, rep, DISP_OPEN_DENY);
 435
 436                 GOTO(err_out, rc);
 437         }
 438
 439         mfd = mdt_mfd_new(med);
 440         if (mfd == NULL)
 441                 GOTO(err_out, rc = -ENOMEM);
 442
 443         /*
 444          * Keep a reference on this object for this open, and is
 445          * released by mdt_mfd_close().
 446          */
 447         mdt_object_get(info->mti_env, o);
 448         mfd->mfd_object = o;
 449         mfd->mfd_xid = req->rq_xid;
 450
 451         /*
 452          * @open_flags is always not zero. At least it should be FMODE_READ,
 453          * FMODE_WRITE or MDS_FMODE_EXEC.
 454          */
 455         LASSERT(open_flags != 0);
 456
 457         /* Open handling. */
 458         mdt_mfd_set_mode(mfd, open_flags);
 459
 460         atomic_inc(&o->mot_open_count);
 461         if (open_flags & MDS_OPEN_LEASE)
 462                 atomic_inc(&o->mot_lease_count);
 463
 464         /* replay handle */
 465         if (req_is_replay(req)) {
 466                 struct mdt_file_data *old_mfd;
 467                 /* Check wheather old cookie already exist in
 468                  * the list, becasue when do recovery, client
 469                  * might be disconnected from server, and
 470                  * restart replay, so there maybe some orphan
 471                  * mfd here, we should remove them */
 472                 LASSERT(info->mti_rr.rr_open_handle != NULL);
 473                 spin_lock(&med->med_open_lock);
 474                 old_mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 475                                               true);
 476                 if (old_mfd != NULL) {
 477                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 478                                "cookie = %#llx\n", mfd,
 479                                PFID(mdt_object_fid(mfd->mfd_object)),
 480                                info->mti_rr.rr_open_handle->cookie);
 481                         class_handle_unhash(&old_mfd->mfd_open_handle);
 482                         list_del_init(&old_mfd->mfd_list);
 483                         spin_unlock(&med->med_open_lock);
 484                         /* no attr update for that close */
 485                         la->la_valid = 0;
 486                         ma->ma_valid |= MA_FLAGS;
 487                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 488                         mdt_mfd_close(info, old_mfd);
 489                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 490                         ma->ma_valid &= ~MA_FLAGS;
 491                 } else {
 492                         spin_unlock(&med->med_open_lock);
 493                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 494                                "cookie = %#llx\n",
 495                                PFID(mdt_object_fid(mfd->mfd_object)),
 496                                info->mti_rr.rr_open_handle->cookie);
 497                 }
 498
 499                 CDEBUG(D_HA, "Store old cookie %#llx in new mfd\n",
 500                        info->mti_rr.rr_open_handle->cookie);
 501
 502                 mfd->mfd_open_handle_old = *info->mti_rr.rr_open_handle;
 503         }
 504
 505         repbody->mbo_open_handle.cookie = mfd->mfd_open_handle.h_cookie;
 506
 507         if (req->rq_export->exp_disconnected) {
 508                 spin_lock(&med->med_open_lock);
 509                 class_handle_unhash(&mfd->mfd_open_handle);
 510                 list_del_init(&mfd->mfd_list);
 511                 spin_unlock(&med->med_open_lock);
 512                 mdt_mfd_close(info, mfd);
 513         } else {
 514                 spin_lock(&med->med_open_lock);
 515                 list_add_tail(&mfd->mfd_list, &med->med_open_head);
 516                 spin_unlock(&med->med_open_lock);
 517         }
 518
 519         mdt_empty_transno(info, rc);
 520
 521         RETURN(rc);
 522
 523 err_out:
 524         if (open_flags & MDS_FMODE_WRITE)
 525                 mdt_write_put(o);
 526         else if (open_flags & MDS_FMODE_EXEC)
 527                 mdt_write_allow(o);
 528
 529         return rc;
 530 }
 531
 532 static int mdt_finish_open(struct mdt_thread_info *info,
 533                            struct mdt_object *p, struct mdt_object *o,
 534                            u64 open_flags,
 535                            struct ldlm_reply *rep)
 536 {
 537         struct ptlrpc_request *req = mdt_info_req(info);
 538         struct obd_export *exp = req->rq_export;
 539         struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 540         struct md_attr *ma  = &info->mti_attr;
 541         struct lu_attr *la  = &ma->ma_attr;
 542         struct mdt_file_data *mfd;
 543         struct mdt_body *repbody;
 544         int created;
 545         int rc = 0;
 546         int isreg, isdir, islnk;
 547         struct list_head *t;
 548
 549         ENTRY;
 550         LASSERT(ma->ma_valid & MA_INODE);
 551         created = mdt_get_disposition(rep, DISP_OPEN_CREATE);
 552
 553         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 554
 555         isreg = S_ISREG(la->la_mode);
 556         isdir = S_ISDIR(la->la_mode);
 557         islnk = S_ISLNK(la->la_mode);
 558         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 559
 560         /* compatibility check for 2.10 clients when it tries to open mirrored
 561          * files. 2.10 clients don't verify overlapping components so they
 562          * would read and write mirrored files just as if they were normal
 563          * PFL files, which will cause the problem that sycned mirrors actually
 564          * contain different data.
 565          * Older clients are not a concern here because they don't even
 566          * understand PFL layout. */
 567         if (isreg && !exp_connect_flr(exp) && ma->ma_valid & MA_LOV &&
 568             mdt_lmm_is_flr(ma->ma_lmm)) {
 569                 /* LU-10286: for simplicity clients who don't understand
 570                  * mirrored layout(with connect flag OBD_CONNECT2_FLR) won't
 571                  * be able to open mirrored files */
 572                 RETURN(-EOPNOTSUPP);
 573         }
 574
 575         /* Overstriped files can crash older clients */
 576         if (isreg && !exp_connect_overstriping(exp) &&
 577             mdt_lmm_is_overstriping(ma->ma_lmm))
 578                 RETURN(-EOPNOTSUPP);
 579
 580         /* LU-2275, simulate broken behaviour (esp. prevalent in
 581          * pre-2.4 servers where a very strange reply is sent on error
 582          * that looks like it was actually almost successful and a
 583          * failure at the same time.) */
 584         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 585                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 586                                                DISP_LOOKUP_NEG |
 587                                                DISP_LOOKUP_POS);
 588
 589                 if (open_flags & MDS_OPEN_LOCK)
 590                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 591
 592                 RETURN(-ENOENT);
 593         }
 594
 595 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
 596         if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 597                 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
 598                 if (IS_ERR(nodemap))
 599                         RETURN(PTR_ERR(nodemap));
 600
 601                 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
 602                 nodemap_putref(nodemap);
 603                 if (rc)
 604                         RETURN(rc);
 605         }
 606 #endif
 607
 608         /*
 609          * If we are following a symlink, don't open; and do not return open
 610          * handle for special nodes as client required.
 611          */
 612         if (islnk || (!isreg && !isdir &&
 613             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 614                 lustre_msg_set_transno(req->rq_repmsg, 0);
 615                 RETURN(0);
 616         }
 617
 618         /*
 619          * We need to return the existing object's fid back, so it is done here,
 620          * after preparing the reply.
 621          */
 622         if (!created && (open_flags & MDS_OPEN_EXCL) &&
 623             (open_flags & MDS_OPEN_CREAT))
 624                 RETURN(-EEXIST);
 625
 626         /* This can't be done earlier, we need to return reply body */
 627         if (isdir) {
 628                 if (open_flags & (MDS_OPEN_CREAT | MDS_FMODE_WRITE)) {
 629                         /* We are trying to create or write an existing dir. */
 630                         RETURN(-EISDIR);
 631                 }
 632         } else if (open_flags & MDS_OPEN_DIRECTORY)
 633                 RETURN(-ENOTDIR);
 634
 635         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 636                                  OBD_FAIL_MDS_LDLM_REPLY_NET | OBD_FAIL_ONCE))
 637                 RETURN(-EAGAIN);
 638
 639         mfd = NULL;
 640         if (info->mti_rr.rr_flags & MRF_OPEN_RESEND) {
 641                 spin_lock(&med->med_open_lock);
 642                 list_for_each(t, &med->med_open_head) {
 643                         mfd = list_entry(t, struct mdt_file_data, mfd_list);
 644                         if (mfd->mfd_xid == req->rq_xid) {
 645                                 repbody->mbo_open_handle.cookie =
 646                                                 mfd->mfd_open_handle.h_cookie;
 647                                 break;
 648                         }
 649                         mfd = NULL;
 650                 }
 651                 spin_unlock(&med->med_open_lock);
 652
 653                 if (mfd != NULL) {
 654                         /* set repbody->ea_size for resent case */
 655                         if (ma->ma_valid & MA_LOV) {
 656                                 LASSERT(ma->ma_lmm_size != 0);
 657                                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 658                                 if (isdir)
 659                                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 660                                 else
 661                                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 662                         }
 663                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 664                         RETURN(0);
 665                 }
 666                 /* if we have a real resend (not a resend afrer failover), it
 667                  * means close is already happend, so lets return error
 668                  */
 669                 RETURN(-ESTALE);
 670         }
 671
 672         rc = mdt_mfd_open(info, p, o, open_flags, created, rep);
 673         if (!rc)
 674                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 675
 676         RETURN(rc);
 677 }
 678
 679 void mdt_reconstruct_open(struct mdt_thread_info *info,
 680                           struct mdt_lock_handle *lhc)
 681 {
 682         struct req_capsule *pill = info->mti_pill;
 683         struct ptlrpc_request *req = mdt_info_req(info);
 684         struct mdt_reint_record *rr = &info->mti_rr;
 685         struct md_attr *ma = &info->mti_attr;
 686         struct ldlm_reply *ldlm_rep;
 687         u64 opdata;
 688         int rc;
 689
 690         ENTRY;
 691         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 692         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 693
 694         ma->ma_need = MA_INODE | MA_HSM;
 695         ma->ma_valid = 0;
 696         opdata = mdt_req_from_lrd(req, info->mti_reply_data);
 697         mdt_set_disposition(info, ldlm_rep, opdata);
 698
 699         CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n",
 700                ldlm_rep->lock_policy_res1, req->rq_status);
 701         if (req->rq_status)
 702                 /* We did not create successfully, return error to client. */
 703                 GOTO(out, rc = req->rq_status);
 704
 705         /* tg_reply_data is just memory only  structure, so any non zero fid
 706          * means a real resend not a resend after recovery which need to be
 707          * handled as regular open
 708          */
 709         if (likely(!fid_is_zero(&info->mti_reply_data->trd_object))) {
 710                 rr->rr_fid2 = &info->mti_reply_data->trd_object;
 711                 rr->rr_flags |= MRF_OPEN_RESEND;
 712                 rc = mdt_open_by_fid(info, ldlm_rep, lhc);
 713                 if (rc)
 714                         lustre_msg_set_transno(req->rq_repmsg, 0);
 715         } else {
 716                 /* We did not try to create, so we are a pure open */
 717                 rc = mdt_reint_open(info, lhc);
 718         }
 719         EXIT;
 720 out:
 721         req->rq_status = rc;
 722         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 723         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
 724 }
 725
 726 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep,
 727                            struct mdt_lock_handle *lhc)
 728 {
 729         u64 open_flags = info->mti_spec.sp_cr_flags;
 730         struct mdt_reint_record *rr = &info->mti_rr;
 731         struct md_attr *ma = &info->mti_attr;
 732         struct mdt_object *o;
 733         int rc;
 734
 735         ENTRY;
 736         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
 737         if (IS_ERR(o))
 738                 RETURN(rc = PTR_ERR(o));
 739
 740         rc = mdt_check_enc(info, o);
 741         if (rc)
 742                 GOTO(out, rc);
 743
 744         if (unlikely(mdt_object_remote(o))) {
 745                 /* the child object was created on remote server */
 746                 struct mdt_body *repbody;
 747
 748                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 749                                                 DISP_LOOKUP_EXECD |
 750                                                 DISP_LOOKUP_POS));
 751                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 752                 repbody->mbo_fid1 = *rr->rr_fid2;
 753                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 754                 rc = 0;
 755         } else {
 756                 if (mdt_object_exists(o)) {
 757                         tgt_open_obj_set(info->mti_env, mdt_obj2dt(o));
 758                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 759                                                         DISP_LOOKUP_EXECD |
 760                                                         DISP_LOOKUP_POS));
 761                         if ((open_flags & MDS_OPEN_EXCL) &&
 762                             (open_flags & MDS_OPEN_CREAT))
 763                                 mdt_set_disposition(info, rep,
 764                                                     DISP_OPEN_CREATE);
 765
 766                         mdt_prep_ma_buf_from_rep(info, o, ma, open_flags);
 767                         rc = mdt_attr_get_complex(info, o, ma);
 768                         if (rc)
 769                                 GOTO(out, rc);
 770                         rc = mdt_finish_open(info, NULL, o, open_flags, rep);
 771                         if (rc)
 772                                 GOTO(out, rc);
 773                         mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
 774                 } else {
 775                         rc = -ENOENT;
 776                 }
 777         }
 778
 779 out:
 780         mdt_object_put(info->mti_env, o);
 781         RETURN(rc);
 782 }
 783
 784 /* lock object for open */
 785 static int mdt_object_open_lock(struct mdt_thread_info *info,
 786                                 struct mdt_object *obj,
 787                                 struct mdt_lock_handle *lhc,
 788                                 __u64 *ibits)
 789 {
 790         struct md_attr *ma = &info->mti_attr;
 791         __u64 open_flags = info->mti_spec.sp_cr_flags;
 792         __u64 trybits = 0;
 793         enum ldlm_mode lm = LCK_CR;
 794         bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
 795         bool try_layout = false;
 796         bool create_layout = false;
 797         int rc = 0;
 798         __u32 dom_stripe = 0;
 799         unsigned int dom_only = 0;
 800         unsigned int dom_lock = 0;
 801
 802         ENTRY;
 803         *ibits = 0;
 804         mdt_lock_handle_init(lhc);
 805
 806         if (req_is_replay(mdt_info_req(info)))
 807                 RETURN(0);
 808
 809         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
 810                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
 811                     md_should_create(open_flags))
 812                         create_layout = true;
 813                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
 814                     ma->ma_need & MA_LOV)
 815                         try_layout = true;
 816
 817                 /* DoM files can take IO lock at OPEN when it makes sense,
 818                  * check if file has DoM stripe and ask for lock if client
 819                  * no lock on that resource yet.
 820                  */
 821                 if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
 822                         dom_stripe = mdt_lmm_dom_entry_check(ma->ma_lmm,
 823                                                              &dom_only);
 824                 /* If only DOM stripe is being used then we can expect IO
 825                  * to it after OPEN and will return corresponding DOM ibit
 826                  * using default strategy from mdt_opts.mo_dom_lock.
 827                  * Otherwise trylock mode is used always and DOM ibit will
 828                  * be returned optionally.
 829                  */
 830                 if (dom_stripe &&
 831                     !mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
 832                         dom_lock = !dom_only ? TRYLOCK_DOM_ON_OPEN :
 833                                    info->mti_mdt->mdt_opts.mo_dom_lock;
 834         }
 835
 836         if (acq_lease) {
 837                 /* lease open, acquire write mode of open sem */
 838                 down_write(&obj->mot_open_sem);
 839
 840                 /* Lease exists and ask for new lease */
 841                 if (atomic_read(&obj->mot_lease_count) > 0) {
 842                         /* only exclusive open is supported, so lease
 843                          * are conflicted to each other */
 844                         GOTO(out, rc = -EBUSY);
 845                 }
 846
 847                 /* Lease must be with open lock */
 848                 if (!(open_flags & MDS_OPEN_LOCK)) {
 849                         CERROR("%s: Request lease for file:"DFID ", but open lock "
 850                                "is missed, open_flags = %#llo : rc = %d\n",
 851                                mdt_obd_name(info->mti_mdt),
 852                                PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
 853                         GOTO(out, rc = -EPROTO);
 854                 }
 855
 856                 /* should conflict with new opens for write/execute */
 857                 lm = LCK_PW;
 858                 *ibits = MDS_INODELOCK_OPEN;
 859
 860                 /* never grant LCK_EX layout lock to client */
 861                 try_layout = false;
 862         } else { /* normal open */
 863                 /* normal open holds read mode of open sem */
 864                 down_read(&obj->mot_open_sem);
 865
 866                 if (open_flags & MDS_OPEN_LOCK) {
 867                         if (open_flags & MDS_FMODE_WRITE)
 868                                 lm = LCK_CW;
 869                         else if (open_flags & MDS_FMODE_EXEC)
 870                                 lm = LCK_PR;
 871                         else
 872                                 lm = LCK_CR;
 873
 874                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
 875                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
 876                         if (open_flags & MDS_FMODE_WRITE)
 877                                 lm = LCK_CW;
 878                         else
 879                                 lm = LCK_CR;
 880
 881                         /* revoke lease */
 882                         *ibits = MDS_INODELOCK_OPEN;
 883                         try_layout = false;
 884
 885                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 886                 } else if (dom_lock) {
 887                         lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR;
 888                         trybits |= MDS_INODELOCK_DOM | MDS_INODELOCK_LAYOUT;
 889                 }
 890
 891                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
 892                         PFID(mdt_object_fid(obj)),
 893                         atomic_read(&obj->mot_lease_count), lm);
 894         }
 895
 896         mdt_lock_reg_init(lhc, lm);
 897
 898         /* Return lookup lock to validate inode at the client side.
 899          * This is pretty important otherwise MDT will return layout
 900          * lock for each open.
 901          * However this is a double-edged sword because changing
 902          * permission will revoke a huge number of LOOKUP locks.
 903          */
 904         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
 905                 if (!(*ibits & MDS_INODELOCK_LOOKUP))
 906                         trybits |= MDS_INODELOCK_LOOKUP;
 907                 trybits |= MDS_INODELOCK_LAYOUT;
 908         }
 909
 910         if (*ibits | trybits)
 911                 rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, false);
 912
 913         CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
 914                ", open_flags = %#llo, try_layout = %d : rc = %d\n",
 915                mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
 916                *ibits, trybits, open_flags, try_layout, rc);
 917
 918         /* will change layout, revoke layout locks by enqueuing EX lock. */
 919         if (rc == 0 && create_layout) {
 920                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
 921
 922                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
 923                         ", open_flags = %#llo\n",
 924                         PFID(mdt_object_fid(obj)), open_flags);
 925
 926                 /* We cannot enqueue another lock for the same resource we
 927                  * already have a lock for, due to mechanics of waiting list
 928                  * iterating in ldlm, see LU-3601.
 929                  * As such we'll drop the open lock we just got above here,
 930                  * it's ok not to have this open lock as it's main purpose is to
 931                  * flush unused cached client open handles. */
 932                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 933                         mdt_object_unlock(info, obj, lhc, 1);
 934
 935                 LASSERT(!try_layout);
 936                 mdt_lock_handle_init(ll);
 937                 mdt_lock_reg_init(ll, LCK_EX);
 938                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
 939
 940                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
 941         }
 942
 943         /* Check if there is any other open handles after acquiring
 944          * open lock. At this point, caching open handles have been revoked
 945          * by open lock.
 946          * XXX: Now only exclusive open is supported. Need to check the
 947          * type of open for generic lease support. */
 948         if (rc == 0 && acq_lease) {
 949                 struct ptlrpc_request *req = mdt_info_req(info);
 950                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 951                 struct mdt_file_data *mfd;
 952                 bool is_replay_or_resent;
 953                 int open_count = 0;
 954
 955                 /* For lease: application can open a file and then apply lease,
 956                  * @handle contains original open handle in that case.
 957                  * In recovery, open REQ will be replayed and the lease REQ may
 958                  * be resent that means the open handle is already stale, so we
 959                  * need to fix it up here by finding new handle. */
 960                 is_replay_or_resent = req_is_replay(req) ||
 961                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
 962
 963                 /* if the request is _not_ a replay request, rr_open_handle
 964                  * may be used to hold an open file handle which is issuing the
 965                  * lease request, so that this openhandle doesn't count. */
 966                 mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 967                                           is_replay_or_resent);
 968                 if (mfd != NULL)
 969                         ++open_count;
 970
 971                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
 972                         PFID(mdt_object_fid(obj)),
 973                         atomic_read(&obj->mot_open_count), open_count);
 974
 975                 if (atomic_read(&obj->mot_open_count) > open_count) {
 976                         /* fail if anyone *else* has opened file for write */
 977                         if (mdt_write_read(obj) > 1)
 978                                 GOTO(out, rc = -EBUSY);
 979                 }
 980         }
 981         GOTO(out, rc);
 982
 983 out:
 984         RETURN(rc);
 985 }
 986
 987 static void mdt_object_open_unlock(struct mdt_thread_info *info,
 988                                    struct mdt_object *obj,
 989                                    struct mdt_lock_handle *lhc,
 990                                    __u64 ibits, int rc)
 991 {
 992         __u64 open_flags = info->mti_spec.sp_cr_flags;
 993         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
 994
 995         ENTRY;
 996         if (req_is_replay(mdt_info_req(info)))
 997                 RETURN_EXIT;
 998
 999         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1000          * return to client side. */
1001         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1002                 mdt_object_unlock(info, obj, ll, 1);
1003
1004         ll = &info->mti_lh[MDT_LH_LAYOUT];
1005         /* Release local layout lock, layout was created */
1006         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1007                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1008                 mdt_object_unlock(info, obj, ll, 1);
1009         }
1010
1011         if (open_flags & MDS_OPEN_LEASE)
1012                 up_write(&obj->mot_open_sem);
1013         else
1014                 up_read(&obj->mot_open_sem);
1015
1016         /* Cross-ref case, the lock should be returned to the client */
1017         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1018                 RETURN_EXIT;
1019
1020         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1021             !(ibits & MDS_INODELOCK_DOM)) {
1022                 /* for the open request, the lock will only return to client
1023                  * if open or layout lock is granted. */
1024                 rc = 1;
1025         }
1026
1027         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1028                 struct ldlm_reply       *ldlm_rep;
1029
1030                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1031                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1032                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1033                         mdt_object_unlock(info, obj, lhc, 1);
1034         }
1035         RETURN_EXIT;
1036 }
1037
1038 /**
1039  * Check release is permitted for the current HSM flags.
1040  */
1041 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1042 {
1043         if (!(ma->ma_valid & MA_HSM))
1044                 return false;
1045
1046         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1047                 return false;
1048
1049         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1050                 return false;
1051
1052         return true;
1053 }
1054
1055 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1056                                 struct ldlm_reply *rep,
1057                                 struct mdt_lock_handle *lhc)
1058 {
1059         const struct lu_env *env = info->mti_env;
1060         struct mdt_device *mdt = info->mti_mdt;
1061         u64 open_flags = info->mti_spec.sp_cr_flags;
1062         struct mdt_reint_record *rr = &info->mti_rr;
1063         struct md_attr *ma = &info->mti_attr;
1064         struct mdt_object *parent = NULL;
1065         struct mdt_object *o;
1066         bool object_locked = false;
1067         u64 ibits = 0;
1068         int rc;
1069
1070         ENTRY;
1071         if (md_should_create(open_flags)) {
1072                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1073                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1074                         if (IS_ERR(parent)) {
1075                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1076                                        " for anonymous created %ld, try to"
1077                                        " use server-side parent.\n",
1078                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1079                                 parent = NULL;
1080                         }
1081                 }
1082                 if (parent == NULL)
1083                         ma->ma_need |= MA_PFID;
1084         }
1085
1086         o = mdt_object_find(env, mdt, rr->rr_fid2);
1087         if (IS_ERR(o))
1088                 GOTO(out_parent_put, rc = PTR_ERR(o));
1089
1090         if (mdt_object_remote(o)) {
1091                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1092                        mdt_obd_name(info->mti_mdt),
1093                        PFID(rr->rr_fid2));
1094                 GOTO(out, rc = -EREMOTE);
1095         } else if (!mdt_object_exists(o)) {
1096                 mdt_set_disposition(info, rep,
1097                                     DISP_IT_EXECD |
1098                                     DISP_LOOKUP_EXECD |
1099                                     DISP_LOOKUP_NEG);
1100                 GOTO(out, rc = -ENOENT);
1101         }
1102
1103         /* do not check enc for directory: always allow open */
1104         if (!S_ISDIR(lu_object_attr(&o->mot_obj))) {
1105                 rc = mdt_check_enc(info, o);
1106                 if (rc)
1107                         GOTO(out, rc);
1108         }
1109
1110         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1111
1112         mdt_prep_ma_buf_from_rep(info, o, ma, open_flags);
1113         if (open_flags & MDS_OPEN_RELEASE)
1114                 ma->ma_need |= MA_HSM;
1115         rc = mdt_attr_get_complex(info, o, ma);
1116         if (rc)
1117                 GOTO(out, rc);
1118
1119         /* We should not change file's existing LOV EA */
1120         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1121             open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1122                 GOTO(out, rc = -EEXIST);
1123
1124         /* If a release request, check file open flags are fine and ask for an
1125          * exclusive open access. */
1126         if (open_flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1127                 GOTO(out, rc = -EPERM);
1128
1129         rc = mdt_check_resent_lock(info, o, lhc);
1130         if (rc < 0) {
1131                 GOTO(out, rc);
1132         } else if (rc > 0) {
1133                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1134                 object_locked = true;
1135                 if (rc)
1136                         GOTO(out_unlock, rc);
1137         }
1138
1139         if (ma->ma_valid & MA_PFID) {
1140                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1141                 if (IS_ERR(parent)) {
1142                         CDEBUG(D_INODE, "Fail to find parent "DFID
1143                                " for anonymous created %ld, try to"
1144                                " use system default.\n",
1145                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1146                         parent = NULL;
1147                 }
1148         }
1149
1150         tgt_open_obj_set(info->mti_env, mdt_obj2dt(o));
1151         rc = mdt_finish_open(info, parent, o, open_flags, rep);
1152         if (!rc) {
1153                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1154                 if (open_flags & MDS_OPEN_LOCK)
1155                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1156                 if (open_flags & MDS_OPEN_LEASE)
1157                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1158         }
1159         GOTO(out_unlock, rc);
1160
1161 out_unlock:
1162         if (object_locked)
1163                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1164 out:
1165         mdt_object_put(env, o);
1166         if (rc == 0)
1167                 mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
1168 out_parent_put:
1169         if (parent != NULL)
1170                 mdt_object_put(env, parent);
1171         return rc;
1172 }
1173
1174 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1175 static int mdt_cross_open(struct mdt_thread_info *info,
1176                           const struct lu_fid *parent_fid,
1177                           const struct lu_fid *fid,
1178                           struct ldlm_reply *rep, u64 open_flags)
1179 {
1180         struct md_attr *ma = &info->mti_attr;
1181         struct mdt_object *o;
1182         int rc;
1183
1184         ENTRY;
1185         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1186         if (IS_ERR(o))
1187                 RETURN(rc = PTR_ERR(o));
1188
1189         rc = mdt_check_enc(info, o);
1190         if (rc)
1191                 GOTO(out, rc);
1192
1193         if (mdt_object_remote(o)) {
1194                 /* Something is wrong here, the object is on another MDS! */
1195                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1196                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1197                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1198                                 &o->mot_obj,
1199                                 "Object isn't on this server! FLD error?");
1200                 rc = -EFAULT;
1201         } else {
1202                 if (mdt_object_exists(o)) {
1203                         int mask;
1204
1205                         /* Do permission check for cross-open after converting
1206                          * MDS_OPEN_* flags to MAY_* permission mask.
1207                          */
1208                         mask = mds_accmode(open_flags);
1209
1210                         rc = mo_permission(info->mti_env, NULL,
1211                                            mdt_object_child(o), NULL, mask);
1212                         if (rc)
1213                                 goto out;
1214
1215                         mdt_prep_ma_buf_from_rep(info, o, ma, open_flags);
1216                         rc = mdt_attr_get_complex(info, o, ma);
1217                         if (rc != 0)
1218                                 GOTO(out, rc);
1219
1220                         rc = mdt_pack_secctx_in_reply(info, o);
1221                         if (unlikely(rc))
1222                                 GOTO(out, rc);
1223
1224                         rc = mdt_pack_encctx_in_reply(info, o);
1225                         if (unlikely(rc))
1226                                 GOTO(out, rc);
1227
1228                         rc = mdt_finish_open(info, NULL, o, open_flags, rep);
1229                 } else {
1230                         /*
1231                          * Something is wrong here. lookup was positive but
1232                          * there is no object!
1233                          */
1234                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1235                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1236                         rc = -EFAULT;
1237                 }
1238         }
1239 out:
1240         mdt_object_put(info->mti_env, o);
1241
1242         RETURN(rc);
1243 }
1244
1245 /*
1246  * find root object and take its xattr lock if it's on remote MDT, later create
1247  * may use fs default striping (which is stored in root xattr).
1248  */
1249 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1250                                struct mdt_device *mdt)
1251 {
1252         struct mdt_object *md_root = mdt->mdt_md_root;
1253         struct lustre_handle lhroot;
1254         int rc;
1255
1256         if (md_root == NULL) {
1257                 lu_root_fid(&info->mti_tmp_fid1);
1258                 md_root = mdt_object_find(info->mti_env, mdt,
1259                                           &info->mti_tmp_fid1);
1260                 if (IS_ERR(md_root))
1261                         return PTR_ERR(md_root);
1262
1263                 spin_lock(&mdt->mdt_lock);
1264                 if (mdt->mdt_md_root != NULL) {
1265                         spin_unlock(&mdt->mdt_lock);
1266
1267                         LASSERTF(mdt->mdt_md_root == md_root,
1268                                  "Different root object ("
1269                                  DFID") instances, %p, %p\n",
1270                                  PFID(&info->mti_tmp_fid1),
1271                                  mdt->mdt_md_root, md_root);
1272                         LASSERT(atomic_read(
1273                                 &md_root->mot_obj.lo_header->loh_ref) > 1);
1274
1275                         mdt_object_put(info->mti_env, md_root);
1276                 } else {
1277                         mdt->mdt_md_root = md_root;
1278                         spin_unlock(&mdt->mdt_lock);
1279                 }
1280         }
1281
1282         if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1283                 return 0;
1284
1285         rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root),
1286                                     &lhroot, LCK_PR, MDS_INODELOCK_XATTR,
1287                                     true);
1288         if (rc < 0)
1289                 return rc;
1290
1291         md_root->mot_cache_attr = 1;
1292
1293         /* don't cancel this lock, so that we know the cached xattr is valid. */
1294         ldlm_lock_decref(&lhroot, LCK_PR);
1295
1296         return 0;
1297 }
1298
1299 static inline enum ldlm_mode mdt_open_lock_mode(struct mdt_thread_info *info,
1300                                                 struct mdt_object *p,
1301                                                 struct lu_name *name,
1302                                                 u64 open_flags)
1303 {
1304         int result;
1305         struct lu_fid fid;
1306
1307         /* We don't need to take the DLM lock for a volatile */
1308         if (open_flags & MDS_OPEN_VOLATILE)
1309                 return LCK_NL;
1310
1311         if (!(open_flags & MDS_OPEN_CREAT))
1312                 return LCK_PR;
1313
1314         result = mdo_lookup(info->mti_env, mdt_object_child(p), name, &fid,
1315                             &info->mti_spec);
1316
1317         /* If the file exists we only need a read lock on the parent */
1318         return (result == 0) ? LCK_PR : LCK_PW;
1319 }
1320
1321 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1322 {
1323         struct mdt_device *mdt = info->mti_mdt;
1324         struct ptlrpc_request *req = mdt_info_req(info);
1325         struct mdt_object *parent;
1326         struct mdt_object *child;
1327         struct mdt_lock_handle *lh = NULL;
1328         struct ldlm_reply *ldlm_rep;
1329         struct mdt_body *repbody;
1330         struct lu_fid *child_fid = &info->mti_tmp_fid1;
1331         struct lu_ucred *uc = mdt_ucred(info);
1332         struct md_attr *ma = &info->mti_attr;
1333         u64 open_flags = info->mti_spec.sp_cr_flags;
1334         u64 ibits = 0;
1335         struct mdt_reint_record *rr = &info->mti_rr;
1336         int result, rc;
1337         int created = 0;
1338         int object_locked = 0;
1339         enum ldlm_mode lock_mode;
1340         u32 msg_flags;
1341         ktime_t kstart = ktime_get();
1342
1343         ENTRY;
1344         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1345                                (obd_timeout + 1) / 4);
1346
1347         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1348
1349         ma->ma_need = MA_INODE;
1350         ma->ma_valid = 0;
1351
1352         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1353         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1354
1355         if (unlikely(open_flags & MDS_OPEN_JOIN_FILE)) {
1356                 CERROR("file join is not supported anymore.\n");
1357                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1358         }
1359         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1360
1361         if ((open_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1362             info->mti_spec.u.sp_ea.eadata == NULL)
1363                 GOTO(out, result = err_serious(-EINVAL));
1364
1365         if (open_flags & MDS_FMODE_WRITE &&
1366             exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1367                 GOTO(out, result = -EROFS);
1368
1369         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1370                "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1371                PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2),
1372                open_flags, ma->ma_attr.la_mode, msg_flags);
1373
1374         /* Prevent by-fid operation if parent fid is .lustre/fid.
1375          * Also, we want rbac roles to have precedence over any other
1376          * permission or capability checks
1377          */
1378         if (lu_fid_eq(rr->rr_fid1, &LU_OBF_FID) && !uc->uc_rbac_byfid_ops)
1379                 GOTO(out, result = -EPERM);
1380
1381         if (info->mti_cross_ref) {
1382                 /* This is cross-ref open */
1383                 mdt_set_disposition(info, ldlm_rep,
1384                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1385                              DISP_LOOKUP_POS));
1386                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1387                                         ldlm_rep, open_flags);
1388                 GOTO(out, result);
1389         } else if (req_is_replay(req)) {
1390                 result = mdt_open_by_fid(info, ldlm_rep, lhc);
1391
1392                 if (result != -ENOENT)
1393                         GOTO(out, result);
1394
1395                 /* We didn't find the correct object, so we need to re-create it
1396                  * via a regular replay. */
1397                 if (!(open_flags & MDS_OPEN_CREAT)) {
1398                         DEBUG_REQ(D_ERROR, req,
1399                                   "OPEN & CREAT not in open replay/by_fid");
1400                         GOTO(out, result = -EFAULT);
1401                 }
1402                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1403         } else if (open_flags & MDS_OPEN_BY_FID) {
1404                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1405                 if (result < 0)
1406                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1407                                PFID(rr->rr_fid2), result);
1408                 GOTO(out, result);
1409         }
1410
1411         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1412                 GOTO(out, result = err_serious(-ENOMEM));
1413
1414         mdt_set_disposition(info, ldlm_rep,
1415                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1416
1417         if (!lu_name_is_valid(&rr->rr_name))
1418                 GOTO(out, result = -EPROTO);
1419
1420         result = mdt_lock_root_xattr(info, mdt);
1421         if (result < 0)
1422                 GOTO(out, result);
1423
1424         parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1425         if (IS_ERR(parent))
1426                 GOTO(out, result = PTR_ERR(parent));
1427
1428         /* get and check version of parent */
1429         result = mdt_version_get_check(info, parent, 0);
1430         if (result) {
1431                 mdt_object_put(info->mti_env, parent);
1432                 GOTO(out, result);
1433         }
1434
1435         if (!uc->uc_rbac_fscrypt_admin &&
1436             parent->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD &&
1437             open_flags & MDS_OPEN_CREAT)
1438                 GOTO(out_parent, result = -EPERM);
1439
1440         result = mdt_check_enc(info, parent);
1441         if (result)
1442                 GOTO(out_parent, result);
1443
1444         fid_zero(child_fid);
1445         result = -ENOENT;
1446         lock_mode = mdt_open_lock_mode(info, parent, &rr->rr_name, open_flags);
1447
1448         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
1449 again_pw:
1450         if (lock_mode != LCK_NL) {
1451                 lh = &info->mti_lh[MDT_LH_PARENT];
1452                 mdt_lock_pdo_init(lh, lock_mode, &rr->rr_name);
1453                 result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1454                 if (result != 0)
1455                         GOTO(out_parent, result);
1456
1457                 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1458                                     &rr->rr_name, child_fid, &info->mti_spec);
1459         }
1460
1461         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1462                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1463                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1464                  PFID(child_fid));
1465
1466         if (result != 0 && result != -ENOENT)
1467                 GOTO(out_parent_unlock, result);
1468
1469         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1470
1471         if (result == -ENOENT) {
1472                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1473                 if (!(open_flags & MDS_OPEN_CREAT))
1474                         GOTO(out_parent_unlock, result);
1475                 if (mdt_rdonly(req->rq_export))
1476                         GOTO(out_parent_unlock, result = -EROFS);
1477
1478                 LASSERT(equi(lh == NULL, lock_mode == LCK_NL));
1479
1480                 if (lock_mode == LCK_PR) {
1481                         /* unlink vs create race: get write lock and restart */
1482                         mdt_object_unlock(info, parent, lh, 1);
1483                         mdt_clear_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1484                         mdt_lock_handle_init(lh);
1485                         lock_mode = LCK_PW;
1486                         goto again_pw;
1487                 }
1488
1489                 *child_fid = *info->mti_rr.rr_fid2;
1490                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1491                          PFID(child_fid));
1492                 /* In the function below, .hs_keycmp resolves to
1493                  * lu_obj_hop_keycmp() */
1494                 /* coverity[overrun-buffer-val] */
1495                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1496         } else {
1497                 /*
1498                  * Check for O_EXCL is moved to the mdt_finish_open(),
1499                  * we need to return FID back in that case.
1500                  */
1501                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1502                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1503         }
1504         if (IS_ERR(child))
1505                 GOTO(out_parent_unlock, result = PTR_ERR(child));
1506
1507         /** check version of child  */
1508         rc = mdt_version_get_check(info, child, 1);
1509         if (rc)
1510                 GOTO(out_child, result = rc);
1511
1512         tgt_open_obj_set(info->mti_env, mdt_obj2dt(child));
1513
1514         if (result == -ENOENT) {
1515                 /* Create under OBF and .lustre is not permitted */
1516                 if (!fid_is_md_operative(rr->rr_fid1) &&
1517                     (open_flags & MDS_OPEN_VOLATILE) == 0)
1518                         GOTO(out_child, result = -EPERM);
1519
1520                 /* save versions in reply */
1521                 mdt_version_get_save(info, parent, 0);
1522                 mdt_version_get_save(info, child, 1);
1523
1524                 /* version of child will be changed */
1525                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1526
1527                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1528                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1529
1530                 /* Don't do lookup sanity check. We know name doesn't exist. */
1531                 info->mti_spec.sp_cr_lookup = 0;
1532                 info->mti_spec.sp_feat = &dt_directory_features;
1533
1534                 result = mdo_create(info->mti_env, mdt_object_child(parent),
1535                                     &rr->rr_name, mdt_object_child(child),
1536                                     &info->mti_spec, &info->mti_attr);
1537                 if (result == -ERESTART) {
1538                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1539                         GOTO(out_child, result);
1540                 } else {
1541                         mdt_prep_ma_buf_from_rep(info, child, ma, open_flags);
1542                         /* XXX: we should call this once, see few lines below */
1543                         if (result == 0)
1544                                 result = mdt_attr_get_complex(info, child, ma);
1545
1546                         if (result != 0)
1547                                 GOTO(out_child, result);
1548                 }
1549                 created = 1;
1550                 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1551                                  ktime_us_delta(ktime_get(), kstart));
1552         } else {
1553                 /*
1554                  * The object is on remote node, return its FID for remote open.
1555                  */
1556                 if (mdt_object_remote(child)) {
1557                         /*
1558                          * Check if this lock already was sent to client and
1559                          * this is resent case. For resent case do not take lock
1560                          * again, use what is already granted.
1561                          */
1562                         LASSERT(lhc != NULL);
1563
1564                         rc = mdt_check_resent_lock(info, child, lhc);
1565                         if (rc < 0) {
1566                                 GOTO(out_child, result = rc);
1567                         } else if (rc > 0) {
1568                                 mdt_lock_handle_init(lhc);
1569                                 mdt_lock_reg_init(lhc, LCK_PR);
1570
1571                                 rc = mdt_object_lock(info, child, lhc,
1572                                                      MDS_INODELOCK_LOOKUP);
1573                         }
1574                         repbody->mbo_fid1 = *mdt_object_fid(child);
1575                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1576                         if (rc != 0)
1577                                 result = rc;
1578                         else
1579                                 result = -MDT_EREMOTE_OPEN;
1580                         GOTO(out_child, result);
1581                 } else if (mdt_object_exists(child)) {
1582                         /* Check early for MDS_OPEN_DIRECTORY/O_DIRECTORY to
1583                          * avoid opening regular files from lfs getstripe
1584                          * since doing so breaks the leases used by lfs
1585                          * mirror. See LU-13693. */
1586                         if (open_flags & MDS_OPEN_DIRECTORY &&
1587                             S_ISREG(lu_object_attr(&child->mot_obj)))
1588                                 GOTO(out_child, result = -ENOTDIR);
1589
1590                         /* We have to get attr & LOV EA & HSM for this
1591                          * object. */
1592                         mdt_prep_ma_buf_from_rep(info, child, ma, open_flags);
1593                         ma->ma_need |= MA_HSM;
1594                         result = mdt_attr_get_complex(info, child, ma);
1595                         if (result != 0)
1596                                 GOTO(out_child, result);
1597                 } else {
1598                         /* Object does not exist. Likely FS corruption. */
1599                         CERROR("%s: name '"DNAME"' present, but FID "
1600                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1601                                PNAME(&rr->rr_name), PFID(child_fid));
1602                         GOTO(out_child, result = -EIO);
1603                 }
1604         }
1605
1606         repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1607         repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1608
1609         rc = mdt_pack_secctx_in_reply(info, child);
1610         if (unlikely(rc))
1611                 GOTO(out_child, result = rc);
1612
1613         rc = mdt_pack_encctx_in_reply(info, child);
1614         if (unlikely(rc))
1615                 GOTO(out_child, result = rc);
1616
1617         rc = mdt_check_resent_lock(info, child, lhc);
1618         if (rc < 0) {
1619                 GOTO(out_child, result = rc);
1620         } else if (rc == 0) {
1621                 /* the open lock might already be gotten in
1622                  * ldlm_handle_enqueue() */
1623                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1624                 if (open_flags & MDS_OPEN_LOCK)
1625                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1626         } else {
1627                 /* get openlock if this isn't replay and client requested it */
1628                 if (!req_is_replay(req)) {
1629                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1630                         object_locked = 1;
1631                         if (rc != 0)
1632                                 GOTO(out_child_unlock, result = rc);
1633                         else if (open_flags & MDS_OPEN_LOCK)
1634                                 mdt_set_disposition(info, ldlm_rep,
1635                                                     DISP_OPEN_LOCK);
1636                 }
1637         }
1638         /* Try to open it now. */
1639         rc = mdt_finish_open(info, parent, child, open_flags, ldlm_rep);
1640         if (rc) {
1641                 result = rc;
1642                 /* openlock will be released if mdt_finish_open() failed */
1643                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1644
1645                 if (created && (open_flags & MDS_OPEN_VOLATILE)) {
1646                         CERROR("%s: cannot open volatile file "DFID", orphan "
1647                                "file will be left in PENDING directory until "
1648                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1649                                PFID(mdt_object_fid(child)), rc);
1650                         GOTO(out_child_unlock, result);
1651                 }
1652
1653                 if (created) {
1654                         ma->ma_need = 0;
1655                         ma->ma_valid = 0;
1656                         rc = mdo_unlink(info->mti_env,
1657                                         mdt_object_child(parent),
1658                                         mdt_object_child(child),
1659                                         &rr->rr_name,
1660                                         &info->mti_attr, 0);
1661                         if (rc != 0)
1662                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1663                                        mdt_obd_name(info->mti_mdt),
1664                                        PFID(mdt_object_fid(child)), rc);
1665                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1666                 }
1667         }
1668
1669         mdt_counter_incr(req, LPROC_MDT_OPEN,
1670                          ktime_us_delta(ktime_get(), kstart));
1671
1672         EXIT;
1673 out_child_unlock:
1674         if (object_locked)
1675                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1676 out_child:
1677         mdt_object_put(info->mti_env, child);
1678         if (result == 0)
1679                 mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh);
1680 out_parent_unlock:
1681         if (lh != NULL)
1682                 mdt_object_unlock(info, parent, lh, result || !created);
1683
1684 out_parent:
1685         mdt_object_put(info->mti_env, parent);
1686 out:
1687         if (result)
1688                 lustre_msg_set_transno(req->rq_repmsg, 0);
1689         return result;
1690 }
1691
1692 /**
1693  * Create an orphan object use local root.
1694  */
1695 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1696                                           struct mdt_device *mdt,
1697                                           const struct lu_fid *fid,
1698                                           struct md_attr *attr, fmode_t fmode)
1699 {
1700         const struct lu_env *env = info->mti_env;
1701         struct md_op_spec *spec = &info->mti_spec;
1702         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1703         struct mdt_object *obj = NULL;
1704         struct mdt_object *local_root;
1705         static const struct lu_name lname = {
1706                 .ln_name = "i_am_nobody",
1707                 .ln_namelen = sizeof("i_am_nobody") - 1,
1708         };
1709         struct lu_ucred *uc;
1710         kernel_cap_t uc_cap_save;
1711         int rc;
1712
1713         ENTRY;
1714         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1715         if (rc != 0)
1716                 RETURN(ERR_PTR(rc));
1717
1718         local_root = mdt_object_find(env, mdt, local_root_fid);
1719         if (IS_ERR(local_root))
1720                 RETURN(local_root);
1721
1722         obj = mdt_object_new(env, mdt, fid);
1723         if (IS_ERR(obj))
1724                 GOTO(out, rc = PTR_ERR(obj));
1725
1726         spec->sp_cr_lookup = 0;
1727         spec->sp_feat = &dt_directory_features;
1728         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1729         if (attr->ma_valid & MA_LOV) {
1730                 spec->u.sp_ea.eadata = attr->ma_lmm;
1731                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1732                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1733         } else {
1734                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1735         }
1736
1737         uc = lu_ucred(env);
1738         uc_cap_save = uc->uc_cap;
1739         cap_raise(uc->uc_cap, CAP_DAC_OVERRIDE);
1740         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1741                         mdt_object_child(obj), spec, attr);
1742         uc->uc_cap = uc_cap_save;
1743         if (rc < 0) {
1744                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1745                        mdt_obd_name(mdt), PFID(fid), rc);
1746                 GOTO(out, rc);
1747         }
1748
1749         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED, spec);
1750         if (rc < 0)
1751                 CERROR("%s: cannot open volatile file "DFID", orphan "
1752                        "file will be left in PENDING directory until "
1753                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1754                        PFID(fid), rc);
1755         GOTO(out, rc);
1756
1757 out:
1758         if (rc < 0) {
1759                 if (!IS_ERR(obj))
1760                         mdt_object_put(env, obj);
1761                 obj = ERR_PTR(rc);
1762         }
1763         mdt_object_put(env, local_root);
1764         return obj;
1765 }
1766
1767 /* XXX Look into layout in MDT layer. */
1768 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1769 {
1770         struct lov_comp_md_v1 *comp_v1;
1771         struct lov_mds_md *v1;
1772         __u32 off;
1773         int i;
1774
1775         if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1776                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1777
1778                 if (comp_v1->lcm_entry_count == 0)
1779                         return -EINVAL;
1780
1781                 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1782                         off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1783                         v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1784                         v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1785                 }
1786         } else {
1787                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1788         }
1789         return 0;
1790 }
1791
1792 static inline int mdt_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1793 {
1794         struct lov_comp_md_v1 *comp_v1;
1795
1796         if (le32_to_cpu(lmm->lmm_magic == LOV_MAGIC_COMP_V1)) {
1797                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1798                 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1799         } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1800                    le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1801                 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1802         } else {
1803                 return -EINVAL;
1804         }
1805         return 0;
1806 }
1807
1808 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1809                            struct md_attr *ma)
1810 {
1811         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1812         struct lu_ucred *uc = mdt_ucred(info);
1813         struct close_data *data;
1814         struct ldlm_lock *lease;
1815         struct mdt_object *orphan;
1816         struct md_attr *orp_ma;
1817         struct lu_buf *buf;
1818         kernel_cap_t cap;
1819         bool lease_broken;
1820         int rc;
1821         int rc2;
1822
1823         ENTRY;
1824         if (mdt_rdonly(info->mti_exp))
1825                 RETURN(-EROFS);
1826
1827         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1828         if (data == NULL)
1829                 RETURN(-EPROTO);
1830
1831         lease = ldlm_handle2lock(&data->cd_handle);
1832         if (lease == NULL)
1833                 RETURN(-ESTALE);
1834
1835         /* try to hold open_sem so that nobody else can open the file */
1836         if (!down_write_trylock(&o->mot_open_sem)) {
1837                 ldlm_lock_cancel(lease);
1838                 GOTO(out_reprocess, rc = -EBUSY);
1839         }
1840
1841         /* Check if the lease open lease has already canceled */
1842         lock_res_and_lock(lease);
1843         lease_broken = ldlm_is_cancel(lease);
1844         unlock_res_and_lock(lease);
1845
1846         LDLM_DEBUG(lease, DFID " lease broken? %d",
1847                    PFID(mdt_object_fid(o)), lease_broken);
1848
1849         /* Cancel server side lease. Client side counterpart should
1850          * have been cancelled. It's okay to cancel it now as we've
1851          * held mot_open_sem. */
1852         ldlm_lock_cancel(lease);
1853
1854         if (lease_broken) /* don't perform release task */
1855                 GOTO(out_unlock, rc = -ESTALE);
1856
1857         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1858                 GOTO(out_unlock, rc = -EINVAL);
1859
1860         /* ma_need was set before but it seems fine to change it in order to
1861          * avoid modifying the one from RPC */
1862         ma->ma_need = MA_HSM;
1863         rc = mdt_attr_get_complex(info, o, ma);
1864         if (rc != 0)
1865                 GOTO(out_unlock, rc);
1866
1867         if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
1868                 if (ma->ma_valid & MA_HSM) {
1869                         if (ma->ma_hsm.mh_flags & HS_RELEASED)
1870                                 GOTO(out_unlock, rc = -EALREADY);
1871
1872                         if (ma->ma_hsm.mh_arch_id != data->cd_archive_id)
1873                                 CDEBUG(D_CACHE,
1874                                        DFID" archive id diff: %llu:%u\n",
1875                                        PFID(mdt_object_fid(o)),
1876                                        ma->ma_hsm.mh_arch_id,
1877                                        data->cd_archive_id);
1878
1879                         if (!(ma->ma_hsm.mh_flags & HS_DIRTY) &&
1880                             ma->ma_hsm.mh_arch_ver == data->cd_data_version) {
1881                                 CDEBUG(D_CACHE,
1882                                        DFID" data version matches: packed=%llu "
1883                                        "and on-disk=%llu\n",
1884                                        PFID(mdt_object_fid(o)),
1885                                        data->cd_data_version,
1886                                        ma->ma_hsm.mh_arch_ver);
1887                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1888                         }
1889
1890                         if (ma->ma_hsm.mh_flags & HS_DIRTY)
1891                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1892                 } else {
1893                         /* Set up HSM attribte for PCC archived object */
1894                         BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1895                                      sizeof(info->mti_xattr_buf));
1896                         buf = &info->mti_buf;
1897                         buf->lb_buf = info->mti_xattr_buf;
1898                         buf->lb_len = sizeof(struct hsm_attrs);
1899                         memset(&ma->ma_hsm, 0, sizeof(ma->ma_hsm));
1900                         ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1901                         ma->ma_hsm.mh_arch_id = data->cd_archive_id;
1902                         ma->ma_hsm.mh_arch_ver = data->cd_data_version;
1903                         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1904
1905                         rc = mo_xattr_set(info->mti_env, mdt_object_child(o),
1906                                           buf, XATTR_NAME_HSM, 0);
1907                         if (rc)
1908                                 GOTO(out_unlock, rc);
1909                 }
1910         } else {
1911                 if (!mdt_hsm_release_allow(ma))
1912                         GOTO(out_unlock, rc = -EPERM);
1913
1914                 /* already released? */
1915                 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1916                         GOTO(out_unlock, rc = 0);
1917
1918                 /* Compare on-disk and packed data_version */
1919                 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1920                         CDEBUG(D_HSM, DFID" data_version mismatches: "
1921                                "packed=%llu and on-disk=%llu\n",
1922                                PFID(mdt_object_fid(o)),
1923                                data->cd_data_version,
1924                                ma->ma_hsm.mh_arch_ver);
1925                         GOTO(out_unlock, rc = -EPERM);
1926                 }
1927         }
1928
1929         ma->ma_valid = MA_INODE;
1930         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1931         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1932         if (rc < 0)
1933                 GOTO(out_unlock, rc);
1934
1935         mutex_lock(&o->mot_som_mutex);
1936         rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size,
1937                            ma->ma_attr.la_blocks);
1938         mutex_unlock(&o->mot_som_mutex);
1939         if (rc2 < 0)
1940                 CDEBUG(D_INODE,
1941                        "%s: File "DFID" SOM update failed: rc = %d\n",
1942                        mdt_obd_name(info->mti_mdt),
1943                        PFID(mdt_object_fid(o)), rc2);
1944
1945
1946         ma->ma_need = MA_INODE | MA_LOV;
1947         rc = mdt_attr_get_complex(info, o, ma);
1948         if (rc < 0)
1949                 GOTO(out_unlock, rc);
1950
1951         if (!(ma->ma_valid & MA_LOV)) {
1952                 /* Even empty file are released */
1953                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1954                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
1955                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1956                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1957                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1958         } else {
1959                 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
1960                  * ma_lmm as lov_user_md, then it will be confused by union of
1961                  * layout_gen and stripe_offset. */
1962                 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
1963                     LOV_MAGIC_MAGIC)
1964                         ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1965                 else
1966                         GOTO(out_unlock, rc = -EINVAL);
1967         }
1968
1969         /* Set file as released. */
1970         rc = mdt_hsm_set_released(ma->ma_lmm);
1971         if (rc)
1972                 GOTO(out_unlock, rc);
1973
1974         orp_ma = &info->mti_u.hsm.attr;
1975         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1976         /* We use root ownership to bypass potential quota
1977          * restrictions on the user and group of the file. */
1978         orp_ma->ma_attr.la_uid = 0;
1979         orp_ma->ma_attr.la_gid = 0;
1980         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1981         orp_ma->ma_lmm = ma->ma_lmm;
1982         orp_ma->ma_lmm_size = ma->ma_lmm_size;
1983         orp_ma->ma_valid = MA_INODE | MA_LOV;
1984         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1985                                  MDS_FMODE_WRITE);
1986         if (IS_ERR(orphan)) {
1987                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1988                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1989                        PTR_ERR(orphan));
1990                 GOTO(out_unlock, rc = PTR_ERR(orphan));
1991         }
1992
1993         /* Set up HSM attribute for orphan object */
1994         BUILD_BUG_ON(sizeof(struct hsm_attrs) > sizeof(info->mti_xattr_buf));
1995         buf = &info->mti_buf;
1996         buf->lb_buf = info->mti_xattr_buf;
1997         buf->lb_len = sizeof(struct hsm_attrs);
1998         ma->ma_hsm.mh_flags |= HS_RELEASED;
1999         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2000         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2001
2002         mdt_lock_reg_init(lh, LCK_EX);
2003         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2004                              MDS_INODELOCK_XATTR);
2005         if (rc != 0)
2006                 GOTO(out_close, rc);
2007
2008         /* The orphan has root ownership so we need to raise
2009          * CAP_FOWNER to set the HSM attributes. */
2010         cap = uc->uc_cap;
2011         cap_raise(uc->uc_cap, CAP_FOWNER);
2012         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2013                           XATTR_NAME_HSM, 0);
2014         uc->uc_cap = cap;
2015         if (rc != 0)
2016                 GOTO(out_layout_lock, rc);
2017
2018         /* Swap layout with orphan objects. */
2019         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2020                              mdt_object_child(orphan),
2021                              SWAP_LAYOUTS_MDS_HSM);
2022
2023         if (!rc && ma->ma_attr_flags & MDS_PCC_ATTACH) {
2024                 ma->ma_need = MA_LOV;
2025                 rc = mdt_attr_get_complex(info, o, ma);
2026         }
2027
2028         EXIT;
2029
2030 out_layout_lock:
2031         /* Release exclusive LL */
2032         mdt_object_unlock(info, o, lh, 1);
2033 out_close:
2034         /* Close orphan object anyway */
2035         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2036                        MDS_FMODE_WRITE);
2037         if (rc2 < 0)
2038                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2039                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2040         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2041                         "object closed");
2042         mdt_object_put(info->mti_env, orphan);
2043
2044 out_unlock:
2045         up_write(&o->mot_open_sem);
2046
2047         /* already released */
2048         if (rc == 0) {
2049                 struct mdt_body *repbody;
2050
2051                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2052                 LASSERT(repbody != NULL);
2053                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2054                 if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
2055                         LASSERT(ma->ma_valid & MA_LOV);
2056                         rc = mdt_get_lmm_gen(ma->ma_lmm,
2057                                              &repbody->mbo_layout_gen);
2058                         if (!rc)
2059                                 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
2060                 }
2061         }
2062
2063 out_reprocess:
2064         ldlm_reprocess_all(lease->l_resource,
2065                            lease->l_policy_data.l_inodebits.bits);
2066         LDLM_LOCK_PUT(lease);
2067
2068         ma->ma_valid = 0;
2069         ma->ma_need = 0;
2070
2071         return rc;
2072 }
2073
2074 int mdt_close_handle_layouts(struct mdt_thread_info *info,
2075                              struct mdt_object *o, struct md_attr *ma)
2076 {
2077         struct mdt_lock_handle *lh1 = &info->mti_lh[MDT_LH_NEW];
2078         struct mdt_lock_handle *lh2 = &info->mti_lh[MDT_LH_OLD];
2079         struct close_data *data;
2080         struct ldlm_lock *lease;
2081         struct mdt_object *o1 = o, *o2 = NULL;
2082         bool lease_broken;
2083         bool swap_objects = false;
2084         int rc;
2085
2086         ENTRY;
2087         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2088                 RETURN(-EROFS);
2089
2090         if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
2091                 RETURN(-EINVAL);
2092
2093         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2094         if (data == NULL)
2095                 RETURN(-EPROTO);
2096
2097         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2098                 RETURN(-EINVAL);
2099
2100         rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
2101         if (rc == 0) {
2102                 /**
2103                  * only MDS_CLOSE_LAYOUT_SPLIT use the same fid to indicate
2104                  * mirror deletion, so we'd zero cd_fid, and keeps o2 be NULL.
2105                  */
2106                 if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT))
2107                         RETURN(-EINVAL);
2108
2109                 /* zero cd_fid to keeps o2 be NULL */
2110                 fid_zero(&data->cd_fid);
2111         } else if (rc < 0) {
2112                 /* Exchange o1 and o2, to enforce locking order */
2113                 swap_objects = true;
2114         }
2115
2116         lease = ldlm_handle2lock(&data->cd_handle);
2117         if (lease == NULL)
2118                 RETURN(-ESTALE);
2119
2120         if (!fid_is_zero(&data->cd_fid)) {
2121                 o2 = mdt_object_find(info->mti_env, info->mti_mdt,
2122                                      &data->cd_fid);
2123                 if (IS_ERR(o2))
2124                         GOTO(out_lease, rc = PTR_ERR(o2));
2125
2126                 if (!mdt_object_exists(o2))
2127                         GOTO(out_obj, rc = -ENOENT);
2128
2129                 if (!S_ISREG(lu_object_attr(&o2->mot_obj)))
2130                         GOTO(out_obj, rc = -EINVAL);
2131
2132                 if (swap_objects)
2133                         swap(o1, o2);
2134         }
2135
2136         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
2137                            MAY_WRITE);
2138         if (rc < 0)
2139                 GOTO(out_obj, rc);
2140
2141         if (o2) {
2142                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2),
2143                                    NULL, MAY_WRITE);
2144                 if (rc < 0)
2145                         GOTO(out_obj, rc);
2146         }
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_obj, 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",
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)
2168                 GOTO(out_unlock_sem, rc = -ESTALE);
2169
2170         mdt_lock_reg_init(lh1, LCK_EX);
2171         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
2172                              MDS_INODELOCK_XATTR);
2173         if (rc < 0)
2174                 GOTO(out_unlock_sem, rc);
2175
2176         if (o2) {
2177                 mdt_lock_reg_init(lh2, LCK_EX);
2178                 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2179                                      MDS_INODELOCK_XATTR);
2180                 if (rc < 0)
2181                         GOTO(out_unlock1, rc);
2182         }
2183
2184         /* Swap layout with orphan object */
2185         if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2186                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2187                                      mdt_object_child(o2), 0);
2188         } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2189                    ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2190                 struct lu_buf *buf = &info->mti_buf;
2191                 struct md_rejig_data mrd;
2192
2193                 if (o2) {
2194                         mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2195                 } else {
2196                         if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)) {
2197                                 /* paranoid check again */
2198                                 CERROR(DFID
2199                                   ":only mirror split support NULL o2 object\n",
2200                                         PFID(mdt_object_fid(o)));
2201                                 GOTO(out_unlock1, rc = -EINVAL);
2202                         }
2203
2204                         /* set NULL mrd_obj for deleting mirror objects */
2205                         mrd.mrd_obj = NULL;
2206                 }
2207
2208                 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2209                         mrd.mrd_mirror_id = data->cd_mirror_id;
2210                         /* set a small enough blocks in the SoM */
2211                         ma->ma_attr.la_blocks >>= 1;
2212                 }
2213
2214                 buf->lb_len = sizeof(mrd);
2215                 buf->lb_buf = &mrd;
2216                 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2217                                   XATTR_LUSTRE_LOV,
2218                                   ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2219                                   LU_XATTR_SPLIT : LU_XATTR_MERGE);
2220                 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS |
2221                                                        LA_LSIZE | LA_LBLOCKS)) {
2222                         int rc2;
2223                         enum lustre_som_flags lsf;
2224
2225                         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS))
2226                                 lsf = SOM_FL_STRICT;
2227                         else
2228                                 lsf = SOM_FL_LAZY;
2229
2230                         mutex_lock(&o->mot_som_mutex);
2231                         rc2 = mdt_set_som(info, o, lsf,
2232                                           ma->ma_attr.la_size,
2233                                           ma->ma_attr.la_blocks);
2234                         mutex_unlock(&o->mot_som_mutex);
2235                         if (rc2 < 0)
2236                                 CERROR(DFID": Setting i_blocks error: %d, "
2237                                        "i_blocks will be reported wrongly and "
2238                                        "can only be fixed in next resync\n",
2239                                        PFID(mdt_object_fid(o)), rc2);
2240                 }
2241         }
2242         if (rc < 0)
2243                 GOTO(out_unlock2, rc);
2244
2245         EXIT;
2246
2247 out_unlock2:
2248         /* Release exclusive LL */
2249         if (o2)
2250                 mdt_object_unlock(info, o2, lh2, 1);
2251
2252 out_unlock1:
2253         mdt_object_unlock(info, o1, lh1, 1);
2254
2255 out_unlock_sem:
2256         up_write(&o->mot_open_sem);
2257
2258         /* already swapped */
2259         if (rc == 0) {
2260                 struct mdt_body *repbody;
2261
2262                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2263                 LASSERT(repbody != NULL);
2264                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2265         }
2266
2267 out_obj:
2268         if (o1 != o)
2269                 /* the 2nd object has been used, and swapped to o1 */
2270                 mdt_object_put(info->mti_env, o1);
2271         else if (o2)
2272                 /* the 2nd object has been used, and not swapped */
2273                 mdt_object_put(info->mti_env, o2);
2274
2275         ldlm_reprocess_all(lease->l_resource,
2276                            lease->l_policy_data.l_inodebits.bits);
2277
2278 out_lease:
2279         LDLM_LOCK_PUT(lease);
2280
2281         if (ma != NULL) {
2282                 ma->ma_valid = 0;
2283                 ma->ma_need = 0;
2284         }
2285
2286         return rc;
2287 }
2288
2289 static int mdt_close_resync_done(struct mdt_thread_info *info,
2290                                  struct mdt_object *o, struct md_attr *ma)
2291 {
2292         struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LOCAL];
2293         struct close_data *data;
2294         struct ldlm_lock *lease;
2295         struct md_layout_change layout = { 0 };
2296         __u32 *resync_ids = NULL;
2297         size_t resync_count = 0;
2298         bool lease_broken;
2299         int rc;
2300
2301         ENTRY;
2302         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2303                 RETURN(-EROFS);
2304
2305         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2306                 RETURN(-EINVAL);
2307
2308         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2309         if (data == NULL)
2310                 RETURN(-EPROTO);
2311
2312         if (req_capsule_req_need_swab(info->mti_pill))
2313                 lustre_swab_close_data_resync_done(&data->cd_resync);
2314
2315         if (!fid_is_zero(&data->cd_fid))
2316                 RETURN(-EPROTO);
2317
2318         lease = ldlm_handle2lock(&data->cd_handle);
2319         if (lease == NULL)
2320                 RETURN(-ESTALE);
2321
2322         /* try to hold open_sem so that nobody else can open the file */
2323         if (!down_write_trylock(&o->mot_open_sem)) {
2324                 ldlm_lock_cancel(lease);
2325                 GOTO(out_reprocess, rc = -EBUSY);
2326         }
2327
2328         /* Check if the lease open lease has already canceled */
2329         lock_res_and_lock(lease);
2330         lease_broken = ldlm_is_cancel(lease);
2331         unlock_res_and_lock(lease);
2332
2333         LDLM_DEBUG(lease, DFID " lease broken? %d",
2334                    PFID(mdt_object_fid(o)), lease_broken);
2335
2336         /* Cancel server side lease. Client side counterpart should
2337          * have been cancelled. It's okay to cancel it now as we've
2338          * held mot_open_sem. */
2339         ldlm_lock_cancel(lease);
2340
2341         if (lease_broken) /* don't perform release task */
2342                 GOTO(out_unlock, rc = -ESTALE);
2343
2344         resync_count = data->cd_resync.resync_count;
2345
2346         if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2347                 void *data;
2348
2349                 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2350                                            RCL_CLIENT))
2351                         GOTO(out_unlock, rc = -EPROTO);
2352
2353                 OBD_ALLOC_PTR_ARRAY(resync_ids, resync_count);
2354                 if (!resync_ids)
2355                         GOTO(out_unlock, rc = -ENOMEM);
2356
2357                 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2358                 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2359
2360                 layout.mlc_resync_ids = resync_ids;
2361         } else {
2362                 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2363         }
2364
2365         layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2366         layout.mlc_resync_count = resync_count;
2367         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2368                 layout.mlc_som.lsa_valid = SOM_FL_STRICT;
2369                 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2370                 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2371         }
2372         rc = mdt_layout_change(info, o, lhc, &layout);
2373         if (rc)
2374                 GOTO(out_unlock, rc);
2375
2376         mdt_object_unlock(info, o, lhc, 0);
2377
2378         EXIT;
2379
2380 out_unlock:
2381         up_write(&o->mot_open_sem);
2382
2383         /* already released */
2384         if (rc == 0) {
2385                 struct mdt_body *repbody;
2386
2387                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2388                 LASSERT(repbody != NULL);
2389                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2390         }
2391
2392         if (resync_ids)
2393                 OBD_FREE_PTR_ARRAY(resync_ids, resync_count);
2394
2395 out_reprocess:
2396         ldlm_reprocess_all(lease->l_resource,
2397                            lease->l_policy_data.l_inodebits.bits);
2398         LDLM_LOCK_PUT(lease);
2399
2400         ma->ma_valid = 0;
2401         ma->ma_need = 0;
2402
2403         return rc;
2404 }
2405
2406 #define MFD_CLOSED(open_flags) ((open_flags) == MDS_FMODE_CLOSED)
2407
2408 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2409 {
2410         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_open_flags));
2411 }
2412
2413 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2414 {
2415         struct mdt_object *o = mfd->mfd_object;
2416         struct md_object *next = mdt_object_child(o);
2417         struct md_attr *ma = &info->mti_attr;
2418         struct lu_fid *ofid = &info->mti_tmp_fid1;
2419         int rc = 0;
2420         int rc2;
2421         u64 open_flags;
2422         u64 intent;
2423
2424         ENTRY;
2425         open_flags = mfd->mfd_open_flags;
2426         intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2427         *ofid = *mdt_object_fid(o);
2428
2429         /* the below message is checked in replay-single.sh test_46 */
2430         CDEBUG(D_INODE, "%s: %sclosing file handle "DFID" with intent: %llx\n",
2431                mdt_obd_name(info->mti_mdt),
2432                ma->ma_valid & MA_FORCE_LOG ? "force " : "", PFID(ofid), intent);
2433
2434         switch (intent) {
2435         case MDS_HSM_RELEASE: {
2436                 rc = mdt_hsm_release(info, o, ma);
2437                 if (rc < 0) {
2438                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2439                                mdt_obd_name(info->mti_mdt),
2440                                PFID(ofid), rc);
2441                         /* continue to close even error occurred. */
2442                 }
2443                 break;
2444         }
2445         case MDS_CLOSE_LAYOUT_MERGE:
2446         case MDS_CLOSE_LAYOUT_SPLIT:
2447         case MDS_CLOSE_LAYOUT_SWAP: {
2448                 rc = mdt_close_handle_layouts(info, o, ma);
2449                 if (rc < 0) {
2450                         CDEBUG(D_INODE,
2451                                "%s: cannot %s layout of "DFID": rc = %d\n",
2452                                mdt_obd_name(info->mti_mdt),
2453                                intent == MDS_CLOSE_LAYOUT_MERGE ? "merge" :
2454                                intent == MDS_CLOSE_LAYOUT_SPLIT ? "split" :
2455                                "swap",
2456                                PFID(ofid), rc);
2457                         /* continue to close even if error occurred. */
2458                 }
2459                 break;
2460         }
2461         case MDS_CLOSE_RESYNC_DONE:
2462                 rc = mdt_close_resync_done(info, o, ma);
2463                 if (rc < 0) {
2464                         CDEBUG(D_INODE,
2465                                "%s: cannot resync layout of "DFID": rc = %d\n",
2466                                mdt_obd_name(info->mti_mdt),
2467                                PFID(ofid), rc);
2468                         /* continue to close even if error occurred. */
2469                 }
2470                 break;
2471         default:
2472                 /* nothing */
2473                 break;
2474         }
2475
2476         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
2477             ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) {
2478                 rc2 = mdt_lsom_update(info, o, false);
2479                 if (rc2 < 0) {
2480                         CDEBUG(D_INODE,
2481                                "%s: File " DFID " LSOM failed: rc = %d\n",
2482                                mdt_obd_name(info->mti_mdt),
2483                                PFID(ofid), rc2);
2484                         if (rc == 0)
2485                                 rc = rc2;
2486                         /* continue to close even if error occurred. */
2487                 }
2488         }
2489
2490         if (open_flags & MDS_FMODE_WRITE)
2491                 mdt_write_put(o);
2492         else if (open_flags & MDS_FMODE_EXEC)
2493                 mdt_write_allow(o);
2494
2495         /* Update atime|mtime|ctime on close. */
2496         if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ ||
2497              open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
2498             (ma->ma_attr.la_valid & LA_ATIME ||
2499              ma->ma_attr.la_valid & LA_MTIME ||
2500              ma->ma_attr.la_valid & LA_CTIME)) {
2501                 ma->ma_valid = MA_INODE;
2502                 ma->ma_attr_flags |= MDS_CLOSE_UPDATE_TIMES;
2503                 ma->ma_attr.la_valid &= (LA_ATIME | LA_MTIME | LA_CTIME);
2504
2505                 if (ma->ma_attr.la_valid & LA_MTIME) {
2506                         if (mdt_attr_get_pfid(info, o, &ma->ma_pfid) == 0)
2507                                 ma->ma_valid |= MA_PFID;
2508                 }
2509
2510                 rc2 = mo_attr_set(info->mti_env, next, ma);
2511                 if (rc2 != 0) {
2512                         CDEBUG(D_INODE,
2513                            "%s: File "DFID" set attr (%#llx) failed: rc = %d\n",
2514                                mdt_obd_name(info->mti_mdt), PFID(ofid),
2515                                ma->ma_attr.la_valid, rc2);
2516                         if (rc == 0)
2517                                 rc = rc2;
2518                 }
2519         }
2520
2521         /* If file data is modified, add the dirty flag. */
2522         if (ma->ma_attr_flags & MDS_DATA_MODIFIED) {
2523                 rc2 = mdt_add_dirty_flag(info, o, ma);
2524                 if (rc2 != 0) {
2525                         CDEBUG(D_INODE,
2526                              "%s: File "DFID" add dirty flag failed: rc = %d\n",
2527                                mdt_obd_name(info->mti_mdt), PFID(ofid), rc2);
2528                         if (rc == 0)
2529                                 rc = rc2;
2530                 }
2531         }
2532
2533         ma->ma_need |= MA_INODE;
2534         ma->ma_valid &= ~MA_INODE;
2535
2536         LASSERT(atomic_read(&o->mot_open_count) > 0);
2537         atomic_dec(&o->mot_open_count);
2538         mdt_handle_last_unlink(info, o, ma);
2539
2540         if (!MFD_CLOSED(open_flags)) {
2541                 rc2 = mo_close(info->mti_env, next, ma, open_flags);
2542                 if (rc2 != 0) {
2543                         CDEBUG(D_INODE,
2544                                "%s: File "DFID" close failed: rc = %d\n",
2545                                mdt_obd_name(info->mti_mdt), PFID(ofid), rc2);
2546                         if (rc == 0)
2547                                 rc = rc2;
2548                 }
2549                 if (mdt_dom_check_for_discard(info, o))
2550                         mdt_dom_discard_data(info, o);
2551         }
2552
2553         /* adjust open and lease count */
2554         if (open_flags & MDS_OPEN_LEASE) {
2555                 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2556                 atomic_dec(&o->mot_lease_count);
2557         }
2558
2559         mdt_mfd_free(mfd);
2560         mdt_object_put(info->mti_env, o);
2561
2562         RETURN(rc);
2563 }
2564
2565 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2566                        struct mdt_body *repbody)
2567 {
2568         struct mdt_export_data *med;
2569         struct mdt_file_data   *mfd;
2570         int ret = 0;
2571         int rc = 0;
2572
2573         ENTRY;
2574         med = &req->rq_export->exp_mdt_data;
2575         spin_lock(&med->med_open_lock);
2576         mfd = mdt_open_handle2mfd(med, &info->mti_open_handle,
2577                                   req_is_replay(req));
2578         if (mdt_mfd_closed(mfd)) {
2579                 spin_unlock(&med->med_open_lock);
2580                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2581                        ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2582                        info->mti_open_handle.cookie);
2583                 /** not serious error since bug 3633 */
2584                 rc = -ESTALE;
2585         } else {
2586                 class_handle_unhash(&mfd->mfd_open_handle);
2587                 list_del_init(&mfd->mfd_list);
2588                 spin_unlock(&med->med_open_lock);
2589                 ret = mdt_mfd_close(info, mfd);
2590         }
2591
2592         RETURN(rc ? rc : ret);
2593 }
2594
2595 int mdt_close(struct tgt_session_info *tsi)
2596 {
2597         struct mdt_thread_info *info = tsi2mdt_info(tsi);
2598         struct ptlrpc_request *req = tgt_ses_req(tsi);
2599         struct md_attr *ma = &info->mti_attr;
2600         struct mdt_body *repbody = NULL;
2601         ktime_t kstart = ktime_get();
2602         int rc;
2603         int rc2;
2604
2605         ENTRY;
2606         /* Close may come with the Size-on-MDS update. Unpack it. */
2607         rc = mdt_close_unpack(info);
2608         if (rc)
2609                 GOTO(out, rc = err_serious(rc));
2610
2611         /* These fields are no longer used and are left for compatibility.
2612          * size is always zero */
2613         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2614                              0);
2615         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2616                              0);
2617         rc = req_capsule_server_pack(info->mti_pill);
2618         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2619                 mdt_client_compatibility(info);
2620                 if (rc == 0)
2621                         mdt_fix_reply(info);
2622                 mdt_exit_ucred(info);
2623                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2624         }
2625
2626         /* Continue to close handle even if we can not pack reply */
2627         if (rc == 0) {
2628                 repbody = req_capsule_server_get(info->mti_pill,
2629                                                  &RMF_MDT_BODY);
2630                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2631                                                     &RMF_MDT_MD);
2632                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2633                                                        &RMF_MDT_MD,
2634                                                        RCL_SERVER);
2635                 ma->ma_need = MA_INODE | MA_LOV;
2636                 repbody->mbo_eadatasize = 0;
2637                 repbody->mbo_aclsize = 0;
2638         } else {
2639                 rc = err_serious(rc);
2640         }
2641
2642         rc2 = mdt_close_internal(info, req, repbody);
2643         if (rc2 != -ESTALE)
2644                 mdt_empty_transno(info, rc2);
2645         if (rc2 != 0 && rc == 0)
2646                 rc = rc2;
2647
2648         if (repbody != NULL) {
2649                 mdt_client_compatibility(info);
2650                 rc2 = mdt_fix_reply(info);
2651                 if (rc2 != 0 && rc == 0)
2652                         rc = rc2;
2653         }
2654
2655         mdt_exit_ucred(info);
2656         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2657                 GOTO(out, rc = err_serious(-ENOMEM));
2658
2659         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2660                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2661                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2662 out:
2663         mdt_thread_info_fini(info);
2664         if (rc == 0)
2665                 mdt_counter_incr(req, LPROC_MDT_CLOSE,
2666                                  ktime_us_delta(ktime_get(), kstart));
2667         RETURN(rc);
2668 }