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 (CFS_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 (CFS_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 636                                  OBD_FAIL_MDS_LDLM_REPLY_NET | CFS_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_PR;
 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         if (req_is_replay(mdt_info_req(info)))
 805                 RETURN(0);
 806
 807         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
 808                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
 809                     md_should_create(open_flags))
 810                         create_layout = true;
 811                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
 812                     ma->ma_need & MA_LOV)
 813                         try_layout = true;
 814
 815                 /* DoM files can take IO lock at OPEN when it makes sense,
 816                  * check if file has DoM stripe and ask for lock if client
 817                  * no lock on that resource yet.
 818                  */
 819                 if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
 820                         dom_stripe = mdt_lmm_dom_entry_check(ma->ma_lmm,
 821                                                              &dom_only);
 822                 /* If only DOM stripe is being used then we can expect IO
 823                  * to it after OPEN and will return corresponding DOM ibit
 824                  * using default strategy from mdt_opts.mo_dom_lock.
 825                  * Otherwise trylock mode is used always and DOM ibit will
 826                  * be returned optionally.
 827                  */
 828                 if (dom_stripe &&
 829                     !mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
 830                         dom_lock = !dom_only ? TRYLOCK_DOM_ON_OPEN :
 831                                    info->mti_mdt->mdt_opts.mo_dom_lock;
 832         }
 833
 834         if (acq_lease) {
 835                 /* lease open, acquire write mode of open sem */
 836                 down_write(&obj->mot_open_sem);
 837
 838                 /* Lease exists and ask for new lease */
 839                 if (atomic_read(&obj->mot_lease_count) > 0) {
 840                         /* only exclusive open is supported, so lease
 841                          * are conflicted to each other */
 842                         GOTO(out, rc = -EBUSY);
 843                 }
 844
 845                 /* Lease must be with open lock */
 846                 if (!(open_flags & MDS_OPEN_LOCK)) {
 847                         CERROR("%s: Request lease for file:"DFID ", but open lock "
 848                                "is missed, open_flags = %#llo : rc = %d\n",
 849                                mdt_obd_name(info->mti_mdt),
 850                                PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
 851                         GOTO(out, rc = -EPROTO);
 852                 }
 853
 854                 /* should conflict with new opens for write/execute */
 855                 lm = LCK_PW;
 856                 *ibits = MDS_INODELOCK_OPEN;
 857
 858                 /* never grant LCK_EX layout lock to client */
 859                 try_layout = false;
 860         } else { /* normal open */
 861                 /* normal open holds read mode of open sem */
 862                 down_read(&obj->mot_open_sem);
 863
 864                 if (open_flags & MDS_OPEN_LOCK) {
 865                         if (open_flags & MDS_FMODE_WRITE)
 866                                 lm = LCK_CW;
 867                         else if (open_flags & MDS_FMODE_EXEC)
 868                                 lm = LCK_PR;
 869                         else
 870                                 lm = LCK_CR;
 871
 872                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
 873                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
 874                         if (open_flags & MDS_FMODE_WRITE)
 875                                 lm = LCK_CW;
 876                         else
 877                                 lm = LCK_CR;
 878
 879                         /* revoke lease */
 880                         *ibits = MDS_INODELOCK_OPEN;
 881                         try_layout = false;
 882
 883                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 884                 } else if (dom_lock) {
 885                         lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR;
 886                         trybits |= MDS_INODELOCK_DOM | MDS_INODELOCK_LAYOUT;
 887                 }
 888
 889                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
 890                         PFID(mdt_object_fid(obj)),
 891                         atomic_read(&obj->mot_lease_count), lm);
 892         }
 893
 894         /* Return lookup lock to validate inode at the client side.
 895          * This is pretty important otherwise MDT will return layout
 896          * lock for each open.
 897          * Since trylocks are opportunistic now and we also have
 898          * ability to selectively cancel lock bits from clients on conflict,
 899          * we would also return other useful bits: PERM and UPDATE - both
 900          * necessary to ensure subsequent stats from the client would not
 901          * need to talk to us again. Historically speaking both permissions
 902          * and other inode data rarely changes after creation anyway, so
 903          * there should not be any downsides from doing it for normal
 904          * operations now.
 905          */
 906         if (!CFS_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
 907                 if (!(*ibits & MDS_INODELOCK_LOOKUP))
 908                         trybits |= MDS_INODELOCK_LOOKUP;
 909                 trybits |= MDS_INODELOCK_LAYOUT;
 910                 /* PERM|UPDATE ibits in CR mode don't make sense, since
 911                  * all modifications lock these ibits in PW mode.
 912                  * Don't lock UPDATE|PERM ibits upon PCC attach, otherwise
 913                  * PCC will be detached when these ibits are revoked.
 914                  */
 915                 if (lm != LCK_CR && !(open_flags & MDS_OPEN_PCC))
 916                         trybits |= MDS_INODELOCK_UPDATE | MDS_INODELOCK_PERM;
 917         }
 918
 919         if (*ibits | trybits)
 920                 rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, lm);
 921
 922         CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
 923                ", open_flags = %#llo, try_layout = %d : rc = %d\n",
 924                mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
 925                *ibits, trybits, open_flags, try_layout, rc);
 926
 927         /* will change layout, revoke layout locks by enqueuing EX lock. */
 928         if (rc == 0 && create_layout) {
 929                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
 930
 931                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
 932                         ", open_flags = %#llo\n",
 933                         PFID(mdt_object_fid(obj)), open_flags);
 934
 935                 /* We cannot enqueue another lock for the same resource we
 936                  * already have a lock for, due to mechanics of waiting list
 937                  * iterating in ldlm, see LU-3601.
 938                  * As such we'll drop the open lock we just got above here,
 939                  * it's ok not to have this open lock as it's main purpose is to
 940                  * flush unused cached client open handles. */
 941                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 942                         mdt_object_unlock(info, obj, lhc, 1);
 943
 944                 LASSERT(!try_layout);
 945                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
 946                                      LCK_EX);
 947
 948                 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
 949         }
 950
 951         /* Check if there is any other open handles after acquiring
 952          * open lock. At this point, caching open handles have been revoked
 953          * by open lock.
 954          * XXX: Now only exclusive open is supported. Need to check the
 955          * type of open for generic lease support. */
 956         if (rc == 0 && acq_lease) {
 957                 struct ptlrpc_request *req = mdt_info_req(info);
 958                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 959                 struct mdt_file_data *mfd;
 960                 bool is_replay_or_resent;
 961                 int open_count = 0;
 962
 963                 /* For lease: application can open a file and then apply lease,
 964                  * @handle contains original open handle in that case.
 965                  * In recovery, open REQ will be replayed and the lease REQ may
 966                  * be resent that means the open handle is already stale, so we
 967                  * need to fix it up here by finding new handle. */
 968                 is_replay_or_resent = req_is_replay(req) ||
 969                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
 970
 971                 /* if the request is _not_ a replay request, rr_open_handle
 972                  * may be used to hold an open file handle which is issuing the
 973                  * lease request, so that this openhandle doesn't count. */
 974                 mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 975                                           is_replay_or_resent);
 976                 if (mfd != NULL)
 977                         ++open_count;
 978
 979                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
 980                         PFID(mdt_object_fid(obj)),
 981                         atomic_read(&obj->mot_open_count), open_count);
 982
 983                 if (atomic_read(&obj->mot_open_count) > open_count) {
 984                         /* fail if anyone *else* has opened file for write */
 985                         if (mdt_write_read(obj) > 1)
 986                                 GOTO(out, rc = -EBUSY);
 987                 }
 988         }
 989         GOTO(out, rc);
 990
 991 out:
 992         RETURN(rc);
 993 }
 994
 995 static void mdt_object_open_unlock(struct mdt_thread_info *info,
 996                                    struct mdt_object *obj,
 997                                    struct mdt_lock_handle *lhc,
 998                                    __u64 ibits, int rc)
 999 {
1000         __u64 open_flags = info->mti_spec.sp_cr_flags;
1001         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1002
1003         ENTRY;
1004         if (req_is_replay(mdt_info_req(info)))
1005                 RETURN_EXIT;
1006
1007         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1008          * return to client side. */
1009         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1010                 mdt_object_unlock(info, obj, ll, 1);
1011
1012         ll = &info->mti_lh[MDT_LH_LAYOUT];
1013         /* Release local layout lock, layout was created */
1014         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1015                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1016                 mdt_object_unlock(info, obj, ll, 1);
1017         }
1018
1019         if (open_flags & MDS_OPEN_LEASE)
1020                 up_write(&obj->mot_open_sem);
1021         else
1022                 up_read(&obj->mot_open_sem);
1023
1024         /* Cross-ref case, the lock should be returned to the client */
1025         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1026                 RETURN_EXIT;
1027
1028         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1029             !(ibits & MDS_INODELOCK_DOM)) {
1030                 /* for the open request, the lock will only return to client
1031                  * if open or layout lock is granted. */
1032                 rc = 1;
1033         }
1034
1035         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1036                 struct ldlm_reply       *ldlm_rep;
1037
1038                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1039                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1040                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1041                         mdt_object_unlock(info, obj, lhc, 1);
1042         }
1043         RETURN_EXIT;
1044 }
1045
1046 /**
1047  * Check release is permitted for the current HSM flags.
1048  */
1049 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1050 {
1051         if (!(ma->ma_valid & MA_HSM))
1052                 return false;
1053
1054         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1055                 return false;
1056
1057         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1058                 return false;
1059
1060         return true;
1061 }
1062
1063 static int mdt_refetch_lovea(struct mdt_thread_info *info,
1064                              struct mdt_object *o, struct md_attr *ma,
1065                              u64 ibits)
1066 {
1067         struct mdt_body *repbody;
1068         int rc;
1069
1070         if ((ibits & MDS_INODELOCK_LAYOUT) == 0)
1071                 return 0;
1072         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
1073                 return 0;
1074
1075         if ((ma->ma_valid & MA_LOV) == 0)
1076                 return 0;
1077
1078         ma->ma_valid &= ~MA_LOV;
1079         info->mti_big_lmm_used = 0;
1080         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1081         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1082                                                RCL_SERVER);
1083         rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1084         if (rc)
1085                 return rc;
1086
1087         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1088         repbody->mbo_eadatasize = ma->ma_lmm_size;
1089         return 0;
1090 }
1091
1092 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1093                                 struct ldlm_reply *rep,
1094                                 struct mdt_lock_handle *lhc)
1095 {
1096         const struct lu_env *env = info->mti_env;
1097         struct mdt_device *mdt = info->mti_mdt;
1098         u64 open_flags = info->mti_spec.sp_cr_flags;
1099         struct mdt_reint_record *rr = &info->mti_rr;
1100         struct md_attr *ma = &info->mti_attr;
1101         struct mdt_object *parent = NULL;
1102         struct mdt_object *o;
1103         bool object_locked = false;
1104         u64 ibits = 0;
1105         int rc;
1106
1107         ENTRY;
1108         if (md_should_create(open_flags)) {
1109                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1110                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1111                         if (IS_ERR(parent)) {
1112                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1113                                        " for anonymous created %ld, try to"
1114                                        " use server-side parent.\n",
1115                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1116                                 parent = NULL;
1117                         }
1118                 }
1119                 if (parent == NULL)
1120                         ma->ma_need |= MA_PFID;
1121         }
1122
1123         o = mdt_object_find(env, mdt, rr->rr_fid2);
1124         if (IS_ERR(o))
1125                 GOTO(out_parent_put, rc = PTR_ERR(o));
1126
1127         if (mdt_object_remote(o)) {
1128                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1129                        mdt_obd_name(info->mti_mdt),
1130                        PFID(rr->rr_fid2));
1131                 GOTO(out, rc = -EREMOTE);
1132         } else if (!mdt_object_exists(o)) {
1133                 mdt_set_disposition(info, rep,
1134                                     DISP_IT_EXECD |
1135                                     DISP_LOOKUP_EXECD |
1136                                     DISP_LOOKUP_NEG);
1137                 GOTO(out, rc = -ENOENT);
1138         }
1139
1140         /* do not check enc for directory: always allow open */
1141         if (!S_ISDIR(lu_object_attr(&o->mot_obj))) {
1142                 rc = mdt_check_enc(info, o);
1143                 if (rc)
1144                         GOTO(out, rc);
1145         }
1146
1147         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1148
1149         mdt_prep_ma_buf_from_rep(info, o, ma, open_flags);
1150         if (open_flags & MDS_OPEN_RELEASE)
1151                 ma->ma_need |= MA_HSM;
1152         rc = mdt_attr_get_complex(info, o, ma);
1153         if (rc)
1154                 GOTO(out, rc);
1155
1156         /* We should not change file's existing LOV EA */
1157         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1158             open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1159                 GOTO(out, rc = -EEXIST);
1160
1161         /* If a release request, check file open flags are fine and ask for an
1162          * exclusive open access. */
1163         if (open_flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1164                 GOTO(out, rc = -EPERM);
1165
1166         rc = mdt_check_resent_lock(info, o, lhc);
1167         if (rc < 0) {
1168                 GOTO(out, rc);
1169         } else if (rc > 0) {
1170                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1171                 object_locked = true;
1172                 if (rc)
1173                         GOTO(out_unlock, rc);
1174         }
1175
1176         if (ma->ma_valid & MA_PFID) {
1177                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1178                 if (IS_ERR(parent)) {
1179                         CDEBUG(D_INODE, "Fail to find parent "DFID
1180                                " for anonymous created %ld, try to"
1181                                " use system default.\n",
1182                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1183                         parent = NULL;
1184                 }
1185         }
1186
1187         tgt_open_obj_set(info->mti_env, mdt_obj2dt(o));
1188         rc = mdt_finish_open(info, parent, o, open_flags, rep);
1189         if (rc)
1190                 GOTO(out_unlock, rc);
1191
1192         mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1193         if (open_flags & MDS_OPEN_LOCK)
1194                 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1195         if (open_flags & MDS_OPEN_LEASE)
1196                 mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1197
1198         /*
1199          * if layout lock is granted, then we should re-fetch LOVEA
1200          * which was originally taken w/o the lock
1201          */
1202         rc = mdt_refetch_lovea(info, o, ma, ibits);
1203
1204         GOTO(out_unlock, rc);
1205
1206 out_unlock:
1207         if (object_locked)
1208                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1209 out:
1210         mdt_object_put(env, o);
1211         if (rc == 0)
1212                 mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
1213 out_parent_put:
1214         if (parent != NULL)
1215                 mdt_object_put(env, parent);
1216         return rc;
1217 }
1218
1219 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1220 static int mdt_cross_open(struct mdt_thread_info *info,
1221                           const struct lu_fid *parent_fid,
1222                           const struct lu_fid *fid,
1223                           struct ldlm_reply *rep, u64 open_flags)
1224 {
1225         struct md_attr *ma = &info->mti_attr;
1226         struct mdt_object *o;
1227         int rc;
1228
1229         ENTRY;
1230         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1231         if (IS_ERR(o))
1232                 RETURN(rc = PTR_ERR(o));
1233
1234         rc = mdt_check_enc(info, o);
1235         if (rc)
1236                 GOTO(out, rc);
1237
1238         if (mdt_object_remote(o)) {
1239                 /* Something is wrong here, the object is on another MDS! */
1240                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1241                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1242                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1243                                 &o->mot_obj,
1244                                 "Object isn't on this server! FLD error?");
1245                 rc = -EFAULT;
1246         } else {
1247                 if (mdt_object_exists(o)) {
1248                         int mask;
1249
1250                         /* Do permission check for cross-open after converting
1251                          * MDS_OPEN_* flags to MAY_* permission mask.
1252                          */
1253                         mask = mds_accmode(open_flags);
1254
1255                         rc = mo_permission(info->mti_env, NULL,
1256                                            mdt_object_child(o), NULL, mask);
1257                         if (rc)
1258                                 goto out;
1259
1260                         mdt_prep_ma_buf_from_rep(info, o, ma, open_flags);
1261                         rc = mdt_attr_get_complex(info, o, ma);
1262                         if (rc != 0)
1263                                 GOTO(out, rc);
1264
1265                         rc = mdt_pack_secctx_in_reply(info, o);
1266                         if (unlikely(rc))
1267                                 GOTO(out, rc);
1268
1269                         rc = mdt_pack_encctx_in_reply(info, o);
1270                         if (unlikely(rc))
1271                                 GOTO(out, rc);
1272
1273                         rc = mdt_finish_open(info, NULL, o, open_flags, rep);
1274                 } else {
1275                         /*
1276                          * Something is wrong here. lookup was positive but
1277                          * there is no object!
1278                          */
1279                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1280                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1281                         rc = -EFAULT;
1282                 }
1283         }
1284 out:
1285         mdt_object_put(info->mti_env, o);
1286
1287         RETURN(rc);
1288 }
1289
1290 /*
1291  * find root object and take its xattr lock if it's on remote MDT, later create
1292  * may use fs default striping (which is stored in root xattr).
1293  */
1294 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1295                                struct mdt_device *mdt)
1296 {
1297         struct mdt_object *md_root = mdt->mdt_md_root;
1298         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1299         __u64 ibits = MDS_INODELOCK_XATTR;
1300         int rc;
1301
1302         if (md_root == NULL) {
1303                 lu_root_fid(&info->mti_tmp_fid1);
1304                 md_root = mdt_object_find(info->mti_env, mdt,
1305                                           &info->mti_tmp_fid1);
1306                 if (IS_ERR(md_root))
1307                         return PTR_ERR(md_root);
1308
1309                 spin_lock(&mdt->mdt_lock);
1310                 if (mdt->mdt_md_root != NULL) {
1311                         spin_unlock(&mdt->mdt_lock);
1312
1313                         LASSERTF(mdt->mdt_md_root == md_root,
1314                                  "Different root object ("
1315                                  DFID") instances, %p, %p\n",
1316                                  PFID(&info->mti_tmp_fid1),
1317                                  mdt->mdt_md_root, md_root);
1318                         LASSERT(atomic_read(
1319                                 &md_root->mot_obj.lo_header->loh_ref) > 1);
1320
1321                         mdt_object_put(info->mti_env, md_root);
1322                 } else {
1323                         mdt->mdt_md_root = md_root;
1324                         spin_unlock(&mdt->mdt_lock);
1325                 }
1326         }
1327
1328         if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1329                 return 0;
1330
1331         mdt_lock_reg_init(lh, LCK_PR);
1332         rc = mdt_object_lock_internal(info, md_root, mdt_object_fid(md_root),
1333                                       lh, &ibits, 0, true);
1334         if (rc < 0)
1335                 return rc;
1336
1337         md_root->mot_cache_attr = 1;
1338
1339         /* don't cancel this lock, so that we know the cached xattr is valid. */
1340         ldlm_lock_decref(&lh->mlh_rreg_lh, LCK_PR);
1341         lh->mlh_rreg_lh.cookie = 0ull;
1342
1343         return 0;
1344 }
1345
1346 static inline enum ldlm_mode mdt_open_lock_mode(struct mdt_thread_info *info,
1347                                                 struct mdt_object *p,
1348                                                 struct lu_name *name,
1349                                                 u64 open_flags)
1350 {
1351         int result;
1352         struct lu_fid fid;
1353
1354         /* We don't need to take the DLM lock for a volatile */
1355         if (open_flags & MDS_OPEN_VOLATILE)
1356                 return LCK_NL;
1357
1358         if (!(open_flags & MDS_OPEN_CREAT))
1359                 return LCK_PR;
1360
1361         result = mdo_lookup(info->mti_env, mdt_object_child(p), name, &fid,
1362                             &info->mti_spec);
1363
1364         /* If the file exists we only need a read lock on the parent */
1365         return (result == 0) ? LCK_PR : LCK_PW;
1366 }
1367
1368 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1369 {
1370         struct mdt_device *mdt = info->mti_mdt;
1371         struct ptlrpc_request *req = mdt_info_req(info);
1372         struct mdt_object *parent;
1373         struct mdt_object *child;
1374         struct mdt_lock_handle *lh = NULL;
1375         struct ldlm_reply *ldlm_rep;
1376         struct mdt_body *repbody;
1377         struct lu_fid *child_fid = &info->mti_tmp_fid1;
1378         struct lu_ucred *uc = mdt_ucred(info);
1379         struct md_attr *ma = &info->mti_attr;
1380         u64 open_flags = info->mti_spec.sp_cr_flags;
1381         u64 ibits = 0;
1382         struct mdt_reint_record *rr = &info->mti_rr;
1383         int result, rc;
1384         int created = 0;
1385         int object_locked = 0;
1386         enum ldlm_mode lock_mode;
1387         u32 msg_flags;
1388         ktime_t kstart = ktime_get();
1389
1390         ENTRY;
1391         CFS_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, CFS_FAIL_ONCE,
1392                                (obd_timeout + 1) / 4);
1393
1394         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1395
1396         ma->ma_need = MA_INODE;
1397         ma->ma_valid = 0;
1398
1399         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1400         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1401
1402         if (unlikely(open_flags & MDS_OPEN_JOIN_FILE)) {
1403                 CERROR("file join is not supported anymore.\n");
1404                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1405         }
1406         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1407
1408         if ((open_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1409             info->mti_spec.u.sp_ea.eadata == NULL)
1410                 GOTO(out, result = err_serious(-EINVAL));
1411
1412         if (open_flags & MDS_FMODE_WRITE &&
1413             exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1414                 GOTO(out, result = -EROFS);
1415
1416         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1417                "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1418                PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2),
1419                open_flags, ma->ma_attr.la_mode, msg_flags);
1420
1421         /* Prevent by-fid operation if parent fid is .lustre/fid.
1422          * Also, we want rbac roles to have precedence over any other
1423          * permission or capability checks
1424          */
1425         if (lu_fid_eq(rr->rr_fid1, &LU_OBF_FID) && !uc->uc_rbac_byfid_ops)
1426                 GOTO(out, result = -EPERM);
1427
1428         if (info->mti_cross_ref) {
1429                 /* This is cross-ref open */
1430                 mdt_set_disposition(info, ldlm_rep,
1431                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1432                              DISP_LOOKUP_POS));
1433                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1434                                         ldlm_rep, open_flags);
1435                 GOTO(out, result);
1436         } else if (req_is_replay(req)) {
1437                 result = mdt_open_by_fid(info, ldlm_rep, lhc);
1438
1439                 if (result != -ENOENT)
1440                         GOTO(out, result);
1441
1442                 /* We didn't find the correct object, so we need to re-create it
1443                  * via a regular replay. */
1444                 if (!(open_flags & MDS_OPEN_CREAT)) {
1445                         DEBUG_REQ(D_ERROR, req,
1446                                   "OPEN & CREAT not in open replay/by_fid");
1447                         GOTO(out, result = -EFAULT);
1448                 }
1449                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1450         } else if (open_flags & MDS_OPEN_BY_FID) {
1451                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1452                 if (result < 0)
1453                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1454                                PFID(rr->rr_fid2), result);
1455                 GOTO(out, result);
1456         }
1457
1458         if (CFS_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1459                 GOTO(out, result = err_serious(-ENOMEM));
1460
1461         mdt_set_disposition(info, ldlm_rep,
1462                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1463
1464         if (!lu_name_is_valid(&rr->rr_name))
1465                 GOTO(out, result = -EPROTO);
1466
1467         result = mdt_lock_root_xattr(info, mdt);
1468         if (result < 0)
1469                 GOTO(out, result);
1470
1471         parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1472         if (IS_ERR(parent))
1473                 GOTO(out, result = PTR_ERR(parent));
1474
1475         /* get and check version of parent */
1476         result = mdt_version_get_check(info, parent, 0);
1477         if (result) {
1478                 mdt_object_put(info->mti_env, parent);
1479                 GOTO(out, result);
1480         }
1481
1482         if (!uc->uc_rbac_fscrypt_admin &&
1483             parent->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD &&
1484             open_flags & MDS_OPEN_CREAT)
1485                 GOTO(out_parent, result = -EPERM);
1486
1487         result = mdt_check_enc(info, parent);
1488         if (result)
1489                 GOTO(out_parent, result);
1490
1491         fid_zero(child_fid);
1492         result = -ENOENT;
1493         lock_mode = mdt_open_lock_mode(info, parent, &rr->rr_name, open_flags);
1494
1495         CFS_RACE(OBD_FAIL_MDS_REINT_OPEN);
1496 again_pw:
1497         if (lock_mode != LCK_NL) {
1498                 lh = &info->mti_lh[MDT_LH_PARENT];
1499                 result = mdt_parent_lock(info, parent, lh, &rr->rr_name,
1500                                          lock_mode);
1501                 if (result != 0)
1502                         GOTO(out_parent, result);
1503
1504                 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1505                                     &rr->rr_name, child_fid, &info->mti_spec);
1506         }
1507
1508         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1509                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1510                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1511                  PFID(child_fid));
1512
1513         if (result != 0 && result != -ENOENT)
1514                 GOTO(out_parent_unlock, result);
1515
1516         CFS_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1517
1518         if (result == -ENOENT) {
1519                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1520                 if (!(open_flags & MDS_OPEN_CREAT))
1521                         GOTO(out_parent_unlock, result);
1522                 if (mdt_rdonly(req->rq_export))
1523                         GOTO(out_parent_unlock, result = -EROFS);
1524
1525                 LASSERT(equi(lh == NULL, lock_mode == LCK_NL));
1526
1527                 if (lock_mode == LCK_PR) {
1528                         /* unlink vs create race: get write lock and restart */
1529                         mdt_object_unlock(info, parent, lh, 1);
1530                         mdt_clear_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1531                         lock_mode = LCK_PW;
1532                         goto again_pw;
1533                 }
1534
1535                 *child_fid = *info->mti_rr.rr_fid2;
1536                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1537                          PFID(child_fid));
1538                 /* In the function below, .hs_keycmp resolves to
1539                  * lu_obj_hop_keycmp() */
1540                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1541         } else {
1542                 /*
1543                  * Check for O_EXCL is moved to the mdt_finish_open(),
1544                  * we need to return FID back in that case.
1545                  */
1546                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1547                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1548         }
1549         if (IS_ERR(child))
1550                 GOTO(out_parent_unlock, result = PTR_ERR(child));
1551
1552         /** check version of child  */
1553         rc = mdt_version_get_check(info, child, 1);
1554         if (rc)
1555                 GOTO(out_child, result = rc);
1556
1557         tgt_open_obj_set(info->mti_env, mdt_obj2dt(child));
1558
1559         if (result == -ENOENT) {
1560                 /* Create under OBF and .lustre is not permitted */
1561                 if (!fid_is_md_operative(rr->rr_fid1) &&
1562                     (open_flags & MDS_OPEN_VOLATILE) == 0)
1563                         GOTO(out_child, result = -EPERM);
1564
1565                 /* save versions in reply */
1566                 mdt_version_get_save(info, parent, 0);
1567                 mdt_version_get_save(info, child, 1);
1568
1569                 /* version of child will be changed */
1570                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1571
1572                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1573                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1574
1575                 /* Don't do lookup sanity check. We know name doesn't exist. */
1576                 info->mti_spec.sp_cr_lookup = 0;
1577                 info->mti_spec.sp_feat = &dt_directory_features;
1578
1579                 /* set jobid xattr name from sysfs parameter */
1580                 strncpy(info->mti_spec.sp_cr_job_xattr, mdt->mdt_job_xattr,
1581                         XATTR_JOB_MAX_LEN);
1582
1583                 result = mdo_create(info->mti_env, mdt_object_child(parent),
1584                                     &rr->rr_name, mdt_object_child(child),
1585                                     &info->mti_spec, &info->mti_attr);
1586                 if (result == -ERESTART) {
1587                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1588                         GOTO(out_child, result);
1589                 } else {
1590                         mdt_prep_ma_buf_from_rep(info, child, ma, open_flags);
1591                         /* XXX: we should call this once, see few lines below */
1592                         if (result == 0)
1593                                 result = mdt_attr_get_complex(info, child, ma);
1594
1595                         if (result != 0)
1596                                 GOTO(out_child, result);
1597                 }
1598                 created = 1;
1599                 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1600                                  ktime_us_delta(ktime_get(), kstart));
1601         } else {
1602                 /*
1603                  * The object is on remote node, return its FID for remote open.
1604                  */
1605                 if (mdt_object_remote(child)) {
1606                         /*
1607                          * Check if this lock already was sent to client and
1608                          * this is resent case. For resent case do not take lock
1609                          * again, use what is already granted.
1610                          */
1611                         LASSERT(lhc != NULL);
1612
1613                         rc = mdt_check_resent_lock(info, child, lhc);
1614                         if (rc < 0)
1615                                 GOTO(out_child, result = rc);
1616                         else if (rc > 0)
1617                                 rc = mdt_object_lookup_lock(info, NULL, child,
1618                                                             lhc, LCK_PR);
1619                         repbody->mbo_fid1 = *mdt_object_fid(child);
1620                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1621                         if (rc != 0)
1622                                 result = rc;
1623                         else
1624                                 result = -MDT_EREMOTE_OPEN;
1625                         GOTO(out_child, result);
1626                 } else if (mdt_object_exists(child)) {
1627                         /* Check early for MDS_OPEN_DIRECTORY/O_DIRECTORY to
1628                          * avoid opening regular files from lfs getstripe
1629                          * since doing so breaks the leases used by lfs
1630                          * mirror. See LU-13693. */
1631                         if (open_flags & MDS_OPEN_DIRECTORY &&
1632                             S_ISREG(lu_object_attr(&child->mot_obj)))
1633                                 GOTO(out_child, result = -ENOTDIR);
1634
1635                         /* We have to get attr & LOV EA & HSM for this
1636                          * object. */
1637                         mdt_prep_ma_buf_from_rep(info, child, ma, open_flags);
1638                         ma->ma_need |= MA_HSM;
1639                         result = mdt_attr_get_complex(info, child, ma);
1640                         if (result != 0)
1641                                 GOTO(out_child, result);
1642                 } else {
1643                         /* Object does not exist. Likely FS corruption. */
1644                         CERROR("%s: name '"DNAME"' present, but FID "
1645                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1646                                PNAME(&rr->rr_name), PFID(child_fid));
1647                         GOTO(out_child, result = -EIO);
1648                 }
1649         }
1650
1651         repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1652         repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1653
1654         rc = mdt_pack_secctx_in_reply(info, child);
1655         if (unlikely(rc))
1656                 GOTO(out_child, result = rc);
1657
1658         rc = mdt_pack_encctx_in_reply(info, child);
1659         if (unlikely(rc))
1660                 GOTO(out_child, result = rc);
1661
1662         rc = mdt_check_resent_lock(info, child, lhc);
1663         if (rc < 0) {
1664                 GOTO(out_child, result = rc);
1665         } else if (rc == 0) {
1666                 /* the open lock might already be gotten in
1667                  * ldlm_handle_enqueue() */
1668                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1669                 if (open_flags & MDS_OPEN_LOCK)
1670                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1671         } else {
1672                 /* get openlock if this isn't replay and client requested it */
1673                 if (!req_is_replay(req)) {
1674                         CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_DELAY_OPEN, cfs_fail_val);
1675                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1676                         object_locked = 1;
1677                         if (rc != 0)
1678                                 GOTO(out_child_unlock, result = rc);
1679                         else if (open_flags & MDS_OPEN_LOCK)
1680                                 mdt_set_disposition(info, ldlm_rep,
1681                                                     DISP_OPEN_LOCK);
1682                 }
1683         }
1684         /* Try to open it now. */
1685         rc = mdt_finish_open(info, parent, child, open_flags, ldlm_rep);
1686         if (rc) {
1687                 result = rc;
1688                 /* openlock will be released if mdt_finish_open() failed */
1689                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1690
1691                 if (created && (open_flags & MDS_OPEN_VOLATILE)) {
1692                         CERROR("%s: cannot open volatile file "DFID", orphan "
1693                                "file will be left in PENDING directory until "
1694                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1695                                PFID(mdt_object_fid(child)), rc);
1696                         GOTO(out_child_unlock, result);
1697                 }
1698
1699                 if (created) {
1700                         ma->ma_need = 0;
1701                         ma->ma_valid = 0;
1702                         rc = mdo_unlink(info->mti_env,
1703                                         mdt_object_child(parent),
1704                                         mdt_object_child(child),
1705                                         &rr->rr_name,
1706                                         &info->mti_attr, 0);
1707                         if (rc != 0)
1708                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1709                                        mdt_obd_name(info->mti_mdt),
1710                                        PFID(mdt_object_fid(child)), rc);
1711                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1712                 }
1713                 GOTO(out_child_unlock, result);
1714         }
1715
1716         /*
1717          * if layout lock is granted, then we should re-fetch LOVEA
1718          * which was originally taken w/o the lock.
1719          */
1720         result = mdt_refetch_lovea(info, child, ma, ibits);
1721
1722         mdt_counter_incr(req, LPROC_MDT_OPEN,
1723                          ktime_us_delta(ktime_get(), kstart));
1724
1725         GOTO(out_child_unlock, result);
1726
1727 out_child_unlock:
1728         if (object_locked)
1729                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1730 out_child:
1731         mdt_object_put(info->mti_env, child);
1732         if (result == 0)
1733                 mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh);
1734 out_parent_unlock:
1735         if (lh != NULL)
1736                 mdt_object_unlock(info, parent, lh, result || !created);
1737
1738 out_parent:
1739         mdt_object_put(info->mti_env, parent);
1740 out:
1741         if (result)
1742                 lustre_msg_set_transno(req->rq_repmsg, 0);
1743         return result;
1744 }
1745
1746 /**
1747  * Create an orphan object use local root.
1748  */
1749 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1750                                           struct mdt_device *mdt,
1751                                           const struct lu_fid *fid,
1752                                           struct md_attr *attr, fmode_t fmode)
1753 {
1754         const struct lu_env *env = info->mti_env;
1755         struct md_op_spec *spec = &info->mti_spec;
1756         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1757         struct mdt_object *obj = NULL;
1758         struct mdt_object *local_root;
1759         static const struct lu_name lname = {
1760                 .ln_name = "i_am_nobody",
1761                 .ln_namelen = sizeof("i_am_nobody") - 1,
1762         };
1763         struct lu_ucred *uc;
1764         kernel_cap_t uc_cap_save;
1765         int rc;
1766
1767         ENTRY;
1768         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1769         if (rc != 0)
1770                 RETURN(ERR_PTR(rc));
1771
1772         local_root = mdt_object_find(env, mdt, local_root_fid);
1773         if (IS_ERR(local_root))
1774                 RETURN(local_root);
1775
1776         obj = mdt_object_new(env, mdt, fid);
1777         if (IS_ERR(obj))
1778                 GOTO(out, rc = PTR_ERR(obj));
1779
1780         spec->sp_cr_lookup = 0;
1781         spec->sp_feat = &dt_directory_features;
1782         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1783         if (attr->ma_valid & MA_LOV) {
1784                 spec->u.sp_ea.eadata = attr->ma_lmm;
1785                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1786                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1787         } else {
1788                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1789         }
1790
1791         uc = lu_ucred(env);
1792         uc_cap_save = uc->uc_cap;
1793         cap_raise(uc->uc_cap, CAP_DAC_OVERRIDE);
1794         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1795                         mdt_object_child(obj), spec, attr);
1796         uc->uc_cap = uc_cap_save;
1797         if (rc < 0) {
1798                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1799                        mdt_obd_name(mdt), PFID(fid), rc);
1800                 GOTO(out, rc);
1801         }
1802
1803         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED, spec);
1804         if (rc < 0)
1805                 CERROR("%s: cannot open volatile file "DFID", orphan "
1806                        "file will be left in PENDING directory until "
1807                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1808                        PFID(fid), rc);
1809         GOTO(out, rc);
1810
1811 out:
1812         if (rc < 0) {
1813                 if (!IS_ERR(obj))
1814                         mdt_object_put(env, obj);
1815                 obj = ERR_PTR(rc);
1816         }
1817         mdt_object_put(env, local_root);
1818         return obj;
1819 }
1820
1821 /* XXX Look into layout in MDT layer. */
1822 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1823 {
1824         struct lov_comp_md_v1 *comp_v1;
1825         struct lov_mds_md *v1;
1826         __u32 off;
1827         int i;
1828
1829         if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1830                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1831
1832                 if (comp_v1->lcm_entry_count == 0)
1833                         return -EINVAL;
1834
1835                 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1836                         off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1837                         v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1838                         if (v1->lmm_magic == cpu_to_le32(LOV_MAGIC_FOREIGN))
1839                                 continue;
1840
1841                         v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1842                 }
1843         } else {
1844                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1845         }
1846         return 0;
1847 }
1848
1849 static inline int mdt_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1850 {
1851         struct lov_comp_md_v1 *comp_v1;
1852
1853         if (le32_to_cpu(lmm->lmm_magic == LOV_MAGIC_COMP_V1)) {
1854                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1855                 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1856         } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1857                    le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1858                 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1859         } else {
1860                 return -EINVAL;
1861         }
1862         return 0;
1863 }
1864
1865 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1866                            struct md_attr *ma)
1867 {
1868         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1869         struct lu_ucred *uc = mdt_ucred(info);
1870         struct close_data *data;
1871         struct ldlm_lock *lease;
1872         struct mdt_object *orphan;
1873         struct md_attr *orp_ma;
1874         struct lu_buf *buf;
1875         kernel_cap_t cap;
1876         bool lease_broken;
1877         int rc;
1878         int rc2;
1879
1880         ENTRY;
1881         if (mdt_rdonly(info->mti_exp))
1882                 RETURN(-EROFS);
1883
1884         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1885         if (data == NULL)
1886                 RETURN(-EPROTO);
1887
1888         lease = ldlm_handle2lock(&data->cd_handle);
1889         if (lease == NULL)
1890                 RETURN(-ESTALE);
1891
1892         /* try to hold open_sem so that nobody else can open the file */
1893         if (!down_write_trylock(&o->mot_open_sem)) {
1894                 ldlm_lock_cancel(lease);
1895                 GOTO(out_reprocess, rc = -EBUSY);
1896         }
1897
1898         /* Check if the lease open lease has already canceled */
1899         lock_res_and_lock(lease);
1900         lease_broken = ldlm_is_cancel(lease);
1901         unlock_res_and_lock(lease);
1902
1903         LDLM_DEBUG(lease, DFID " lease broken? %d",
1904                    PFID(mdt_object_fid(o)), lease_broken);
1905
1906         /* Cancel server side lease. Client side counterpart should
1907          * have been cancelled. It's okay to cancel it now as we've
1908          * held mot_open_sem. */
1909         ldlm_lock_cancel(lease);
1910
1911         if (lease_broken) /* don't perform release task */
1912                 GOTO(out_unlock, rc = -ESTALE);
1913
1914         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1915                 GOTO(out_unlock, rc = -EINVAL);
1916
1917         /* ma_need was set before but it seems fine to change it in order to
1918          * avoid modifying the one from RPC */
1919         ma->ma_need = MA_HSM;
1920         rc = mdt_attr_get_complex(info, o, ma);
1921         if (rc != 0)
1922                 GOTO(out_unlock, rc);
1923
1924         if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
1925                 if (ma->ma_valid & MA_HSM) {
1926                         if (ma->ma_hsm.mh_flags & HS_RELEASED)
1927                                 GOTO(out_unlock, rc = -EALREADY);
1928
1929                         if (ma->ma_hsm.mh_arch_id != data->cd_archive_id)
1930                                 CDEBUG(D_CACHE,
1931                                        DFID" archive id diff: %llu:%u\n",
1932                                        PFID(mdt_object_fid(o)),
1933                                        ma->ma_hsm.mh_arch_id,
1934                                        data->cd_archive_id);
1935
1936                         if (!(ma->ma_hsm.mh_flags & HS_DIRTY) &&
1937                             ma->ma_hsm.mh_arch_ver == data->cd_data_version) {
1938                                 CDEBUG(D_CACHE,
1939                                        DFID" data version matches: packed=%llu "
1940                                        "and on-disk=%llu\n",
1941                                        PFID(mdt_object_fid(o)),
1942                                        data->cd_data_version,
1943                                        ma->ma_hsm.mh_arch_ver);
1944                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1945                         }
1946
1947                         if (ma->ma_hsm.mh_flags & HS_DIRTY)
1948                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1949                 } else {
1950                         /* Set up HSM attribte for PCC archived object */
1951                         BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1952                                      sizeof(info->mti_xattr_buf));
1953                         buf = &info->mti_buf;
1954                         buf->lb_buf = info->mti_xattr_buf;
1955                         buf->lb_len = sizeof(struct hsm_attrs);
1956                         memset(&ma->ma_hsm, 0, sizeof(ma->ma_hsm));
1957                         ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1958                         ma->ma_hsm.mh_arch_id = data->cd_archive_id;
1959                         ma->ma_hsm.mh_arch_ver = data->cd_data_version;
1960                         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1961
1962                         rc = mo_xattr_set(info->mti_env, mdt_object_child(o),
1963                                           buf, XATTR_NAME_HSM, 0);
1964                         if (rc)
1965                                 GOTO(out_unlock, rc);
1966                 }
1967         } else {
1968                 if (!mdt_hsm_release_allow(ma))
1969                         GOTO(out_unlock, rc = -EPERM);
1970
1971                 /* already released? */
1972                 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1973                         GOTO(out_unlock, rc = 0);
1974
1975                 /* Compare on-disk and packed data_version */
1976                 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1977                         CDEBUG(D_HSM, DFID" data_version mismatches: "
1978                                "packed=%llu and on-disk=%llu\n",
1979                                PFID(mdt_object_fid(o)),
1980                                data->cd_data_version,
1981                                ma->ma_hsm.mh_arch_ver);
1982                         GOTO(out_unlock, rc = -EPERM);
1983                 }
1984         }
1985
1986         ma->ma_valid = MA_INODE;
1987         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1988         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1989         if (rc < 0)
1990                 GOTO(out_unlock, rc);
1991
1992         mutex_lock(&o->mot_som_mutex);
1993         rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size,
1994                            ma->ma_attr.la_blocks);
1995         mutex_unlock(&o->mot_som_mutex);
1996         if (rc2 < 0)
1997                 CDEBUG(D_INODE,
1998                        "%s: File "DFID" SOM update failed: rc = %d\n",
1999                        mdt_obd_name(info->mti_mdt),
2000                        PFID(mdt_object_fid(o)), rc2);
2001
2002
2003         ma->ma_need = MA_INODE | MA_LOV;
2004         rc = mdt_attr_get_complex(info, o, ma);
2005         if (rc < 0)
2006                 GOTO(out_unlock, rc);
2007
2008         if (!(ma->ma_valid & MA_LOV)) {
2009                 /* Even empty file are released */
2010                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2011                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
2012                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2013                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2014                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2015         } else {
2016                 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
2017                  * ma_lmm as lov_user_md, then it will be confused by union of
2018                  * layout_gen and stripe_offset. */
2019                 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
2020                     LOV_MAGIC_MAGIC)
2021                         ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2022                 else
2023                         GOTO(out_unlock, rc = -EINVAL);
2024         }
2025
2026         /* Set file as released. */
2027         rc = mdt_hsm_set_released(ma->ma_lmm);
2028         if (rc)
2029                 GOTO(out_unlock, rc);
2030
2031         orp_ma = &info->mti_u.hsm.attr;
2032         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2033         /* We use root ownership to bypass potential quota
2034          * restrictions on the user and group of the file. */
2035         orp_ma->ma_attr.la_uid = 0;
2036         orp_ma->ma_attr.la_gid = 0;
2037         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2038         orp_ma->ma_lmm = ma->ma_lmm;
2039         orp_ma->ma_lmm_size = ma->ma_lmm_size;
2040         orp_ma->ma_valid = MA_INODE | MA_LOV;
2041         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2042                                  MDS_FMODE_WRITE);
2043         if (IS_ERR(orphan)) {
2044                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2045                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2046                        PTR_ERR(orphan));
2047                 GOTO(out_unlock, rc = PTR_ERR(orphan));
2048         }
2049
2050         /* Set up HSM attribute for orphan object */
2051         BUILD_BUG_ON(sizeof(struct hsm_attrs) > sizeof(info->mti_xattr_buf));
2052         buf = &info->mti_buf;
2053         buf->lb_buf = info->mti_xattr_buf;
2054         buf->lb_len = sizeof(struct hsm_attrs);
2055         ma->ma_hsm.mh_flags |= HS_RELEASED;
2056         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2057         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2058
2059         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2060                              MDS_INODELOCK_XATTR, LCK_EX);
2061         if (rc != 0)
2062                 GOTO(out_close, rc);
2063
2064         /* The orphan has root ownership so we need to raise
2065          * CAP_FOWNER to set the HSM attributes. */
2066         cap = uc->uc_cap;
2067         cap_raise(uc->uc_cap, CAP_FOWNER);
2068         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2069                           XATTR_NAME_HSM, 0);
2070         uc->uc_cap = cap;
2071         if (rc != 0)
2072                 GOTO(out_layout_lock, rc);
2073
2074         /* Swap layout with orphan objects. */
2075         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2076                              mdt_object_child(orphan),
2077                              SWAP_LAYOUTS_MDS_HSM);
2078
2079         if (!rc && ma->ma_attr_flags & MDS_PCC_ATTACH) {
2080                 ma->ma_need = MA_LOV;
2081                 rc = mdt_attr_get_complex(info, o, ma);
2082         }
2083
2084         EXIT;
2085
2086 out_layout_lock:
2087         /* Release exclusive LL */
2088         mdt_object_unlock(info, o, lh, 1);
2089 out_close:
2090         /* Close orphan object anyway */
2091         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2092                        MDS_FMODE_WRITE);
2093         if (rc2 < 0)
2094                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2095                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2096         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2097                         "object closed");
2098         mdt_object_put(info->mti_env, orphan);
2099
2100 out_unlock:
2101         up_write(&o->mot_open_sem);
2102
2103         /* already released */
2104         if (rc == 0) {
2105                 struct mdt_body *repbody;
2106
2107                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2108                 LASSERT(repbody != NULL);
2109                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2110                 if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
2111                         LASSERT(ma->ma_valid & MA_LOV);
2112                         rc = mdt_get_lmm_gen(ma->ma_lmm,
2113                                              &repbody->mbo_layout_gen);
2114                         if (!rc)
2115                                 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
2116                 }
2117         }
2118
2119 out_reprocess:
2120         ldlm_reprocess_all(lease->l_resource,
2121                            lease->l_policy_data.l_inodebits.bits);
2122         LDLM_LOCK_PUT(lease);
2123
2124         ma->ma_valid = 0;
2125         ma->ma_need = 0;
2126
2127         return rc;
2128 }
2129
2130 static int mdt_close_handle_layouts(struct mdt_thread_info *info,
2131                                     struct mdt_object *o, struct md_attr *ma)
2132 {
2133         struct mdt_lock_handle *lh1 = &info->mti_lh[MDT_LH_NEW];
2134         struct mdt_lock_handle *lh2 = &info->mti_lh[MDT_LH_OLD];
2135         struct close_data *data;
2136         struct ldlm_lock *lease;
2137         struct mdt_object *o1 = o, *o2 = NULL;
2138         bool lease_broken;
2139         bool swap_objects = false;
2140         int rc;
2141
2142         ENTRY;
2143         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2144                 RETURN(-EROFS);
2145
2146         if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
2147                 RETURN(-EINVAL);
2148
2149         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2150         if (data == NULL)
2151                 RETURN(-EPROTO);
2152
2153         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2154                 RETURN(-EINVAL);
2155
2156         rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
2157         if (rc == 0) {
2158                 /**
2159                  * only MDS_CLOSE_LAYOUT_SPLIT use the same fid to indicate
2160                  * mirror deletion, so we'd zero cd_fid, and keeps o2 be NULL.
2161                  */
2162                 if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT))
2163                         RETURN(-EINVAL);
2164
2165                 /* zero cd_fid to keeps o2 be NULL */
2166                 fid_zero(&data->cd_fid);
2167         } else if (rc < 0) {
2168                 /* Exchange o1 and o2, to enforce locking order */
2169                 swap_objects = true;
2170         }
2171
2172         lease = ldlm_handle2lock(&data->cd_handle);
2173         if (lease == NULL)
2174                 RETURN(-ESTALE);
2175
2176         if (!fid_is_zero(&data->cd_fid)) {
2177                 o2 = mdt_object_find(info->mti_env, info->mti_mdt,
2178                                      &data->cd_fid);
2179                 if (IS_ERR(o2))
2180                         GOTO(out_lease, rc = PTR_ERR(o2));
2181
2182                 if (!mdt_object_exists(o2))
2183                         GOTO(out_obj, rc = -ENOENT);
2184
2185                 if (!S_ISREG(lu_object_attr(&o2->mot_obj)))
2186                         GOTO(out_obj, rc = -EINVAL);
2187
2188                 if (swap_objects)
2189                         swap(o1, o2);
2190         }
2191
2192         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
2193                            MAY_WRITE);
2194         if (rc < 0)
2195                 GOTO(out_obj, rc);
2196
2197         if (o2) {
2198                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2),
2199                                    NULL, MAY_WRITE);
2200                 if (rc < 0)
2201                         GOTO(out_obj, rc);
2202         }
2203
2204         /* try to hold open_sem so that nobody else can open the file */
2205         if (!down_write_trylock(&o->mot_open_sem)) {
2206                 ldlm_lock_cancel(lease);
2207                 GOTO(out_obj, rc = -EBUSY);
2208         }
2209
2210         /* Check if the lease open lease has already canceled */
2211         lock_res_and_lock(lease);
2212         lease_broken = ldlm_is_cancel(lease);
2213         unlock_res_and_lock(lease);
2214
2215         LDLM_DEBUG(lease, DFID " lease broken? %d",
2216                    PFID(mdt_object_fid(o)), lease_broken);
2217
2218         /* Cancel server side lease. Client side counterpart should
2219          * have been cancelled. It's okay to cancel it now as we've
2220          * held mot_open_sem. */
2221         ldlm_lock_cancel(lease);
2222
2223         if (lease_broken)
2224                 GOTO(out_unlock_sem, rc = -ESTALE);
2225
2226         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
2227                              MDS_INODELOCK_XATTR, LCK_EX);
2228         if (rc < 0)
2229                 GOTO(out_unlock_sem, rc);
2230
2231         if (o2) {
2232                 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2233                                      MDS_INODELOCK_XATTR, LCK_EX);
2234                 if (rc < 0)
2235                         GOTO(out_unlock1, rc);
2236         }
2237
2238         /* Swap layout with orphan object */
2239         if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2240                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2241                                      mdt_object_child(o2), 0);
2242         } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2243                    ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2244                 struct lu_buf *buf = &info->mti_buf;
2245                 struct md_rejig_data mrd;
2246
2247                 if (o2) {
2248                         mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2249                 } else {
2250                         if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)) {
2251                                 /* paranoid check again */
2252                                 CERROR(DFID
2253                                   ":only mirror split support NULL o2 object\n",
2254                                         PFID(mdt_object_fid(o)));
2255                                 GOTO(out_unlock1, rc = -EINVAL);
2256                         }
2257
2258                         /* set NULL mrd_obj for deleting mirror objects */
2259                         mrd.mrd_obj = NULL;
2260                 }
2261
2262                 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2263                         mrd.mrd_mirror_id = data->cd_mirror_id;
2264                         /* set a small enough blocks in the SoM */
2265                         ma->ma_attr.la_blocks >>= 1;
2266                 }
2267
2268                 buf->lb_len = sizeof(mrd);
2269                 buf->lb_buf = &mrd;
2270                 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2271                                   XATTR_LUSTRE_LOV,
2272                                   ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2273                                   LU_XATTR_SPLIT : LU_XATTR_MERGE);
2274                 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS |
2275                                                        LA_LSIZE | LA_LBLOCKS)) {
2276                         int rc2;
2277                         enum lustre_som_flags lsf;
2278
2279                         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS))
2280                                 lsf = SOM_FL_STRICT;
2281                         else
2282                                 lsf = SOM_FL_LAZY;
2283
2284                         mutex_lock(&o->mot_som_mutex);
2285                         rc2 = mdt_set_som(info, o, lsf,
2286                                           ma->ma_attr.la_size,
2287                                           ma->ma_attr.la_blocks);
2288                         mutex_unlock(&o->mot_som_mutex);
2289                         if (rc2 < 0)
2290                                 CERROR(DFID": Setting i_blocks error: %d, "
2291                                        "i_blocks will be reported wrongly and "
2292                                        "can only be fixed in next resync\n",
2293                                        PFID(mdt_object_fid(o)), rc2);
2294                 }
2295         }
2296         if (rc < 0)
2297                 GOTO(out_unlock2, rc);
2298
2299         EXIT;
2300
2301 out_unlock2:
2302         /* Release exclusive LL */
2303         if (o2)
2304                 mdt_object_unlock(info, o2, lh2, 1);
2305
2306 out_unlock1:
2307         mdt_object_unlock(info, o1, lh1, 1);
2308
2309 out_unlock_sem:
2310         up_write(&o->mot_open_sem);
2311
2312         /* already swapped */
2313         if (rc == 0) {
2314                 struct mdt_body *repbody;
2315
2316                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2317                 LASSERT(repbody != NULL);
2318                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2319         }
2320
2321 out_obj:
2322         if (o1 != o)
2323                 /* the 2nd object has been used, and swapped to o1 */
2324                 mdt_object_put(info->mti_env, o1);
2325         else if (o2)
2326                 /* the 2nd object has been used, and not swapped */
2327                 mdt_object_put(info->mti_env, o2);
2328
2329         ldlm_reprocess_all(lease->l_resource,
2330                            lease->l_policy_data.l_inodebits.bits);
2331
2332 out_lease:
2333         LDLM_LOCK_PUT(lease);
2334
2335         if (ma != NULL) {
2336                 ma->ma_valid = 0;
2337                 ma->ma_need = 0;
2338         }
2339
2340         return rc;
2341 }
2342
2343 static int mdt_close_resync_done(struct mdt_thread_info *info,
2344                                  struct mdt_object *o, struct md_attr *ma)
2345 {
2346         struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LOCAL];
2347         struct close_data *data;
2348         struct ldlm_lock *lease;
2349         struct md_layout_change layout = { 0 };
2350         __u32 *resync_ids = NULL;
2351         size_t resync_count = 0;
2352         bool lease_broken;
2353         int rc;
2354
2355         ENTRY;
2356         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2357                 RETURN(-EROFS);
2358
2359         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2360                 RETURN(-EINVAL);
2361
2362         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2363         if (data == NULL)
2364                 RETURN(-EPROTO);
2365
2366         if (req_capsule_req_need_swab(info->mti_pill))
2367                 lustre_swab_close_data_resync_done(&data->cd_resync);
2368
2369         if (!fid_is_zero(&data->cd_fid))
2370                 RETURN(-EPROTO);
2371
2372         lease = ldlm_handle2lock(&data->cd_handle);
2373         if (lease == NULL)
2374                 RETURN(-ESTALE);
2375
2376         /* try to hold open_sem so that nobody else can open the file */
2377         if (!down_write_trylock(&o->mot_open_sem)) {
2378                 ldlm_lock_cancel(lease);
2379                 GOTO(out_reprocess, rc = -EBUSY);
2380         }
2381
2382         /* Check if the lease open lease has already canceled */
2383         lock_res_and_lock(lease);
2384         lease_broken = ldlm_is_cancel(lease);
2385         unlock_res_and_lock(lease);
2386
2387         LDLM_DEBUG(lease, DFID " lease broken? %d",
2388                    PFID(mdt_object_fid(o)), lease_broken);
2389
2390         /* Cancel server side lease. Client side counterpart should
2391          * have been cancelled. It's okay to cancel it now as we've
2392          * held mot_open_sem. */
2393         ldlm_lock_cancel(lease);
2394
2395         if (lease_broken) /* don't perform release task */
2396                 GOTO(out_unlock, rc = -ESTALE);
2397
2398         resync_count = data->cd_resync.resync_count;
2399
2400         if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2401                 void *data;
2402
2403                 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2404                                            RCL_CLIENT))
2405                         GOTO(out_unlock, rc = -EPROTO);
2406
2407                 OBD_ALLOC_PTR_ARRAY(resync_ids, resync_count);
2408                 if (!resync_ids)
2409                         GOTO(out_unlock, rc = -ENOMEM);
2410
2411                 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2412                 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2413
2414                 layout.mlc_resync_ids = resync_ids;
2415         } else {
2416                 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2417         }
2418
2419         layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2420         layout.mlc_resync_count = resync_count;
2421         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2422                 layout.mlc_som.lsa_valid = SOM_FL_STRICT;
2423                 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2424                 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2425         }
2426         rc = mdt_layout_change(info, o, lhc, &layout);
2427         if (rc)
2428                 GOTO(out_unlock, rc);
2429
2430         mdt_object_unlock(info, o, lhc, 0);
2431
2432         EXIT;
2433
2434 out_unlock:
2435         up_write(&o->mot_open_sem);
2436
2437         /* already released */
2438         if (rc == 0) {
2439                 struct mdt_body *repbody;
2440
2441                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2442                 LASSERT(repbody != NULL);
2443                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2444         }
2445
2446         if (resync_ids)
2447                 OBD_FREE_PTR_ARRAY(resync_ids, resync_count);
2448
2449 out_reprocess:
2450         ldlm_reprocess_all(lease->l_resource,
2451                            lease->l_policy_data.l_inodebits.bits);
2452         LDLM_LOCK_PUT(lease);
2453
2454         ma->ma_valid = 0;
2455         ma->ma_need = 0;
2456
2457         return rc;
2458 }
2459
2460 #define MFD_CLOSED(open_flags) ((open_flags) == MDS_FMODE_CLOSED)
2461
2462 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2463 {
2464         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_open_flags));
2465 }
2466
2467 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2468 {
2469         struct mdt_object *o = mfd->mfd_object;
2470         struct md_object *next = mdt_object_child(o);
2471         struct md_attr *ma = &info->mti_attr;
2472         struct lu_fid *ofid = &info->mti_tmp_fid1;
2473         int rc = 0;
2474         int rc2;
2475         u64 open_flags;
2476         u64 intent;
2477
2478         ENTRY;
2479         open_flags = mfd->mfd_open_flags;
2480         intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2481         *ofid = *mdt_object_fid(o);
2482
2483         /* the below message is checked in replay-single.sh test_46 */
2484         CDEBUG(D_INODE, "%s: %sclosing file handle "DFID" with intent: %llx\n",
2485                mdt_obd_name(info->mti_mdt),
2486                ma->ma_valid & MA_FORCE_LOG ? "force " : "", PFID(ofid), intent);
2487
2488         switch (intent) {
2489         case MDS_HSM_RELEASE: {
2490                 rc = mdt_hsm_release(info, o, ma);
2491                 if (rc < 0) {
2492                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2493                                mdt_obd_name(info->mti_mdt),
2494                                PFID(ofid), rc);
2495                         /* continue to close even error occurred. */
2496                 }
2497                 break;
2498         }
2499         case MDS_CLOSE_LAYOUT_MERGE:
2500         case MDS_CLOSE_LAYOUT_SPLIT:
2501         case MDS_CLOSE_LAYOUT_SWAP: {
2502                 rc = mdt_close_handle_layouts(info, o, ma);
2503                 if (rc < 0) {
2504                         CDEBUG(D_INODE,
2505                                "%s: cannot %s layout of "DFID": rc = %d\n",
2506                                mdt_obd_name(info->mti_mdt),
2507                                intent == MDS_CLOSE_LAYOUT_MERGE ? "merge" :
2508                                intent == MDS_CLOSE_LAYOUT_SPLIT ? "split" :
2509                                "swap",
2510                                PFID(ofid), rc);
2511                         /* continue to close even if error occurred. */
2512                 }
2513                 break;
2514         }
2515         case MDS_CLOSE_RESYNC_DONE:
2516                 rc = mdt_close_resync_done(info, o, ma);
2517                 if (rc < 0) {
2518                         CDEBUG(D_INODE,
2519                                "%s: cannot resync layout of "DFID": rc = %d\n",
2520                                mdt_obd_name(info->mti_mdt),
2521                                PFID(ofid), rc);
2522                         /* continue to close even if error occurred. */
2523                 }
2524                 break;
2525         default:
2526                 /* nothing */
2527                 break;
2528         }
2529
2530         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
2531             ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) {
2532                 rc2 = mdt_lsom_update(info, o, false);
2533                 if (rc2 < 0) {
2534                         CDEBUG(D_INODE,
2535                                "%s: File " DFID " LSOM failed: rc = %d\n",
2536                                mdt_obd_name(info->mti_mdt),
2537                                PFID(ofid), rc2);
2538                         if (rc == 0)
2539                                 rc = rc2;
2540                         /* continue to close even if error occurred. */
2541                 }
2542         }
2543
2544         if (open_flags & MDS_FMODE_WRITE)
2545                 mdt_write_put(o);
2546         else if (open_flags & MDS_FMODE_EXEC)
2547                 mdt_write_allow(o);
2548
2549         /* Update atime|mtime|ctime on close. */
2550         if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ ||
2551              open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
2552             (ma->ma_attr.la_valid & LA_ATIME ||
2553              ma->ma_attr.la_valid & LA_MTIME ||
2554              ma->ma_attr.la_valid & LA_CTIME)) {
2555                 ma->ma_valid = MA_INODE;
2556                 ma->ma_attr_flags |= MDS_CLOSE_UPDATE_TIMES;
2557                 ma->ma_attr.la_valid &= (LA_ATIME | LA_MTIME | LA_CTIME);
2558
2559                 if (ma->ma_attr.la_valid & LA_MTIME) {
2560                         if (mdt_attr_get_pfid(info, o, &ma->ma_pfid) == 0)
2561                                 ma->ma_valid |= MA_PFID;
2562                 }
2563
2564                 rc2 = mo_attr_set(info->mti_env, next, ma);
2565                 if (rc2 != 0) {
2566                         CDEBUG(D_INODE,
2567                            "%s: File "DFID" set attr (%#llx) failed: rc = %d\n",
2568                                mdt_obd_name(info->mti_mdt), PFID(ofid),
2569                                ma->ma_attr.la_valid, rc2);
2570                         if (rc == 0)
2571                                 rc = rc2;
2572                 }
2573         }
2574
2575         /* If file data is modified, add the dirty flag. */
2576         if (ma->ma_attr_flags & MDS_DATA_MODIFIED) {
2577                 rc2 = mdt_add_dirty_flag(info, o, ma);
2578                 if (rc2 != 0) {
2579                         CDEBUG(D_INODE,
2580                              "%s: File "DFID" add dirty flag failed: rc = %d\n",
2581                                mdt_obd_name(info->mti_mdt), PFID(ofid), rc2);
2582                         if (rc == 0)
2583                                 rc = rc2;
2584                 }
2585         }
2586
2587         ma->ma_need |= MA_INODE;
2588         ma->ma_valid &= ~MA_INODE;
2589
2590         LASSERT(atomic_read(&o->mot_open_count) > 0);
2591         atomic_dec(&o->mot_open_count);
2592         mdt_handle_last_unlink(info, o, ma);
2593
2594         if (!MFD_CLOSED(open_flags)) {
2595                 rc2 = mo_close(info->mti_env, next, ma, open_flags);
2596                 if (rc2 != 0) {
2597                         CDEBUG(D_INODE,
2598                                "%s: File "DFID" close failed: rc = %d\n",
2599                                mdt_obd_name(info->mti_mdt), PFID(ofid), rc2);
2600                         if (rc == 0)
2601                                 rc = rc2;
2602                 }
2603                 if (mdt_dom_check_for_discard(info, o))
2604                         mdt_dom_discard_data(info, o);
2605         }
2606
2607         /* adjust open and lease count */
2608         if (open_flags & MDS_OPEN_LEASE) {
2609                 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2610                 atomic_dec(&o->mot_lease_count);
2611         }
2612
2613         mdt_mfd_free(mfd);
2614         mdt_object_put(info->mti_env, o);
2615
2616         RETURN(rc);
2617 }
2618
2619 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2620                        struct mdt_body *repbody)
2621 {
2622         struct mdt_export_data *med;
2623         struct mdt_file_data   *mfd;
2624         int ret = 0;
2625         int rc = 0;
2626
2627         ENTRY;
2628         med = &req->rq_export->exp_mdt_data;
2629         spin_lock(&med->med_open_lock);
2630         mfd = mdt_open_handle2mfd(med, &info->mti_open_handle,
2631                                   req_is_replay(req));
2632         if (mdt_mfd_closed(mfd)) {
2633                 spin_unlock(&med->med_open_lock);
2634                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2635                        ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2636                        info->mti_open_handle.cookie);
2637                 /** not serious error since bug 3633 */
2638                 rc = -ESTALE;
2639         } else {
2640                 class_handle_unhash(&mfd->mfd_open_handle);
2641                 list_del_init(&mfd->mfd_list);
2642                 spin_unlock(&med->med_open_lock);
2643                 ret = mdt_mfd_close(info, mfd);
2644         }
2645
2646         RETURN(rc ? rc : ret);
2647 }
2648
2649 int mdt_close(struct tgt_session_info *tsi)
2650 {
2651         struct mdt_thread_info *info = tsi2mdt_info(tsi);
2652         struct ptlrpc_request *req = tgt_ses_req(tsi);
2653         struct md_attr *ma = &info->mti_attr;
2654         struct mdt_body *repbody = NULL;
2655         ktime_t kstart = ktime_get();
2656         int rc;
2657         int rc2;
2658
2659         ENTRY;
2660         /* Close may come with the Size-on-MDS update. Unpack it. */
2661         rc = mdt_close_unpack(info);
2662         if (rc)
2663                 GOTO(out, rc = err_serious(rc));
2664
2665         /* These fields are no longer used and are left for compatibility.
2666          * size is always zero */
2667         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2668                              0);
2669         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2670                              0);
2671         rc = req_capsule_server_pack(info->mti_pill);
2672         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2673                 mdt_client_compatibility(info);
2674                 if (rc == 0)
2675                         mdt_fix_reply(info);
2676                 mdt_exit_ucred(info);
2677                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2678         }
2679
2680         /* Continue to close handle even if we can not pack reply */
2681         if (rc == 0) {
2682                 repbody = req_capsule_server_get(info->mti_pill,
2683                                                  &RMF_MDT_BODY);
2684                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2685                                                     &RMF_MDT_MD);
2686                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2687                                                        &RMF_MDT_MD,
2688                                                        RCL_SERVER);
2689                 ma->ma_need = MA_INODE | MA_LOV;
2690                 repbody->mbo_eadatasize = 0;
2691                 repbody->mbo_aclsize = 0;
2692         } else {
2693                 rc = err_serious(rc);
2694         }
2695
2696         rc2 = mdt_close_internal(info, req, repbody);
2697         if (rc2 != -ESTALE)
2698                 mdt_empty_transno(info, rc2);
2699         if (rc2 != 0 && rc == 0)
2700                 rc = rc2;
2701
2702         if (repbody != NULL) {
2703                 mdt_client_compatibility(info);
2704                 rc2 = mdt_fix_reply(info);
2705                 if (rc2 != 0 && rc == 0)
2706                         rc = rc2;
2707         }
2708
2709         mdt_exit_ucred(info);
2710         if (CFS_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2711                 GOTO(out, rc = err_serious(-ENOMEM));
2712
2713         if (CFS_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2714                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2715                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2716 out:
2717         mdt_thread_info_fini(info);
2718         if (rc == 0)
2719                 mdt_counter_incr(req, LPROC_MDT_CLOSE,
2720                                  ktime_us_delta(ktime_get(), kstart));
2721         RETURN(rc);
2722 }