lustre/mdt/mdt_open.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lustre/mdt/mdt_open.c
  32  *
  33  * Lustre Metadata Target (mdt) open/close file handling
  34  *
  35  * Author: Huang Hua <huanghua@clusterfs.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_MDS
  39
  40 #include <lustre_acl.h>
  41 #include <lustre_mds.h>
  42 #include <lustre_swab.h>
  43 #include "mdt_internal.h"
  44 #include <lustre_nodemap.h>
  45
  46 static const char mfd_open_handle_owner[] = "mdt";
  47
  48 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep,
  49                            struct mdt_lock_handle *lhc);
  50
  51 /* Create a new mdt_file_data struct, initialize it,
  52  * and insert it to global hash table */
  53 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
  54 {
  55         struct mdt_file_data *mfd;
  56
  57         ENTRY;
  58         OBD_ALLOC_PTR(mfd);
  59         if (mfd != NULL) {
  60                 refcount_set(&mfd->mfd_open_handle.h_ref, 1);
  61                 INIT_HLIST_NODE(&mfd->mfd_open_handle.h_link);
  62                 mfd->mfd_owner = med;
  63                 INIT_LIST_HEAD(&mfd->mfd_list);
  64                 class_handle_hash(&mfd->mfd_open_handle, mfd_open_handle_owner);
  65         }
  66
  67         RETURN(mfd);
  68 }
  69
  70 /*
  71  * Find the mfd pointed to by handle in global hash table.
  72  * In case of replay the handle is obsoleted
  73  * but mfd can be found in mfd list by that handle.
  74  * Callers need to be holding med_open_lock.
  75  */
  76 struct mdt_file_data *mdt_open_handle2mfd(struct mdt_export_data *med,
  77                                         const struct lustre_handle *open_handle,
  78                                         bool is_replay_or_resent)
  79 {
  80         struct mdt_file_data   *mfd;
  81
  82         ENTRY;
  83         LASSERT(open_handle != NULL);
  84         mfd = class_handle2object(open_handle->cookie, mfd_open_handle_owner);
  85         if (mfd)
  86                 refcount_dec(&mfd->mfd_open_handle.h_ref);
  87
  88         /* during dw/setattr replay the mfd can be found by old handle */
  89         if ((!mfd || mfd->mfd_owner != med) && is_replay_or_resent) {
  90                 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  91                         if (mfd->mfd_open_handle_old.cookie ==
  92                             open_handle->cookie)
  93                                 RETURN(mfd);
  94                 }
  95                 mfd = NULL;
  96         }
  97
  98         RETURN(mfd);
  99 }
 100
 101 /* free mfd */
 102 void mdt_mfd_free(struct mdt_file_data *mfd)
 103 {
 104         LASSERT(refcount_read(&mfd->mfd_open_handle.h_ref) == 1);
 105         LASSERT(list_empty(&mfd->mfd_list));
 106         OBD_FREE_PRE(mfd, sizeof(*mfd), "kfree_rcu");
 107         kfree_rcu(mfd, mfd_open_handle.h_rcu);
 108 }
 109
 110 static int mdt_create_data(struct mdt_thread_info *info,
 111                            struct mdt_object *p, struct mdt_object *o)
 112 {
 113         struct md_op_spec *spec = &info->mti_spec;
 114         struct md_attr *ma = &info->mti_attr;
 115         int rc = 0;
 116
 117         ENTRY;
 118         if (!md_should_create(spec->sp_cr_flags))
 119                 RETURN(0);
 120
 121         ma->ma_need = MA_INODE | MA_LOV;
 122         ma->ma_valid = 0;
 123         mutex_lock(&o->mot_lov_mutex);
 124         if (!o->mot_lov_created) {
 125                 rc = mdo_create_data(info->mti_env,
 126                                      p ? mdt_object_child(p) : NULL,
 127                                      mdt_object_child(o), spec, ma);
 128                 if (rc == 0)
 129                         rc = mdt_attr_get_complex(info, o, ma);
 130
 131                 if (rc == 0 && ma->ma_valid & MA_LOV)
 132                         o->mot_lov_created = 1;
 133         }
 134
 135         mutex_unlock(&o->mot_lov_mutex);
 136         RETURN(rc);
 137 }
 138
 139 int mdt_write_read(struct mdt_object *o)
 140 {
 141         int rc = 0;
 142
 143         ENTRY;
 144         spin_lock(&o->mot_write_lock);
 145         rc = o->mot_write_count;
 146         spin_unlock(&o->mot_write_lock);
 147         RETURN(rc);
 148 }
 149
 150 int mdt_write_get(struct mdt_object *o)
 151 {
 152         int rc = 0;
 153
 154         ENTRY;
 155         spin_lock(&o->mot_write_lock);
 156         if (o->mot_write_count < 0)
 157                 rc = -ETXTBSY;
 158         else
 159                 o->mot_write_count++;
 160         spin_unlock(&o->mot_write_lock);
 161
 162         RETURN(rc);
 163 }
 164
 165 void mdt_write_put(struct mdt_object *o)
 166 {
 167         ENTRY;
 168         spin_lock(&o->mot_write_lock);
 169         o->mot_write_count--;
 170         spin_unlock(&o->mot_write_lock);
 171         EXIT;
 172 }
 173
 174 static int mdt_write_deny(struct mdt_object *o)
 175 {
 176         int rc = 0;
 177
 178         ENTRY;
 179         spin_lock(&o->mot_write_lock);
 180         if (o->mot_write_count > 0)
 181                 rc = -ETXTBSY;
 182         else
 183                 o->mot_write_count--;
 184         spin_unlock(&o->mot_write_lock);
 185         RETURN(rc);
 186 }
 187
 188 static void mdt_write_allow(struct mdt_object *o)
 189 {
 190         ENTRY;
 191         spin_lock(&o->mot_write_lock);
 192         o->mot_write_count++;
 193         spin_unlock(&o->mot_write_lock);
 194         EXIT;
 195 }
 196
 197 /* there can be no real transaction so prepare the fake one */
 198 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 199 {
 200         struct mdt_device *mdt = info->mti_mdt;
 201         struct ptlrpc_request *req = mdt_info_req(info);
 202         struct tg_export_data *ted;
 203         struct lsd_client_data *lcd;
 204
 205         ENTRY;
 206         if (mdt_rdonly(req->rq_export))
 207                 RETURN_EXIT;
 208
 209         /* transaction has occurred already */
 210         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 211                 RETURN_EXIT;
 212
 213         if (tgt_is_multimodrpcs_client(req->rq_export)) {
 214                 struct thandle *th;
 215
 216                 /* generate an empty transaction to get a transno
 217                  * and reply data */
 218                 th = dt_trans_create(info->mti_env, mdt->mdt_bottom);
 219                 if (!IS_ERR(th)) {
 220                         rc = dt_trans_start(info->mti_env, mdt->mdt_bottom, th);
 221                         dt_trans_stop(info->mti_env, mdt->mdt_bottom, th);
 222                 }
 223                 RETURN_EXIT;
 224         }
 225
 226         spin_lock(&mdt->mdt_lut.lut_translock);
 227         if (rc != 0) {
 228                 if (info->mti_transno != 0) {
 229                         struct obd_export *exp = req->rq_export;
 230
 231                         CERROR("%s: replay trans %llu NID %s: rc = %d\n",
 232                                mdt_obd_name(mdt), info->mti_transno,
 233                                obd_export_nid2str(exp), rc);
 234                         spin_unlock(&mdt->mdt_lut.lut_translock);
 235                         RETURN_EXIT;
 236                 }
 237         } else if (info->mti_transno == 0) {
 238                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
 239         } else {
 240                 /* should be replay */
 241                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 242                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 243         }
 244         spin_unlock(&mdt->mdt_lut.lut_translock);
 245
 246         CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
 247                info->mti_transno,
 248                req->rq_export->exp_obd->obd_last_committed);
 249
 250         req->rq_transno = info->mti_transno;
 251         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 252
 253         /* update lcd in memory only for resent cases */
 254         ted = &req->rq_export->exp_target_data;
 255         LASSERT(ted);
 256         mutex_lock(&ted->ted_lcd_lock);
 257         lcd = ted->ted_lcd;
 258         if (info->mti_transno < lcd->lcd_last_transno &&
 259             info->mti_transno != 0) {
 260                 /* This should happen during replay. Do not update
 261                  * last rcvd info if replay req transno < last transno,
 262                  * otherwise the following resend(after replay) can not
 263                  * be checked correctly by xid */
 264                 mutex_unlock(&ted->ted_lcd_lock);
 265                 CDEBUG(D_HA, "%s: transno = %llu < last_transno = %llu\n",
 266                        mdt_obd_name(mdt), info->mti_transno,
 267                        lcd->lcd_last_transno);
 268                 RETURN_EXIT;
 269         }
 270
 271         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
 272                 if (info->mti_transno != 0)
 273                         lcd->lcd_last_close_transno = info->mti_transno;
 274                 lcd->lcd_last_close_xid = req->rq_xid;
 275                 lcd->lcd_last_close_result = rc;
 276         } else {
 277                 /* VBR: save versions in last_rcvd for reconstruct. */
 278                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 279
 280                 if (pre_versions) {
 281                         lcd->lcd_pre_versions[0] = pre_versions[0];
 282                         lcd->lcd_pre_versions[1] = pre_versions[1];
 283                         lcd->lcd_pre_versions[2] = pre_versions[2];
 284                         lcd->lcd_pre_versions[3] = pre_versions[3];
 285                 }
 286                 if (info->mti_transno != 0)
 287                         lcd->lcd_last_transno = info->mti_transno;
 288
 289                 lcd->lcd_last_xid = req->rq_xid;
 290                 lcd->lcd_last_result = rc;
 291                 lcd->lcd_last_data = info->mti_opdata;
 292         }
 293         mutex_unlock(&ted->ted_lcd_lock);
 294
 295         EXIT;
 296 }
 297
 298 void mdt_mfd_set_mode(struct mdt_file_data *mfd, u64 open_flags)
 299 {
 300         LASSERT(mfd != NULL);
 301
 302         CDEBUG(D_DENTRY, DFID " Change mfd open_flags %#llo -> %#llo.\n",
 303                PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_open_flags,
 304                open_flags);
 305
 306         mfd->mfd_open_flags = open_flags;
 307 }
 308
 309 /**
 310  * prep ma_lmm/ma_lmv for md_attr from reply
 311  */
 312 static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
 313                                      struct mdt_object *obj,
 314                                      struct md_attr *ma, __u64 open_flags)
 315 {
 316         struct req_capsule *pill;
 317
 318         if (ma->ma_lmv || ma->ma_lmm) {
 319                 CDEBUG(D_INFO, DFID " %s already set.\n",
 320                        PFID(mdt_object_fid(obj)),
 321                        ma->ma_lmv ? (ma->ma_lmm ? "ma_lmv and ma_lmm"
 322                                                 : "ma_lmv")
 323                                   : "ma_lmm");
 324                 return;
 325         }
 326
 327         pill = info->mti_pill;
 328         if (S_ISDIR(obj->mot_header.loh_attr)) {
 329                 ma->ma_lmv = req_capsule_server_get(pill, &RMF_MDT_MD);
 330                 ma->ma_lmv_size = req_capsule_get_size(pill, &RMF_MDT_MD,
 331                                                        RCL_SERVER);
 332                 if (ma->ma_lmv_size > 0)
 333                         ma->ma_need |= MA_LMV;
 334
 335                 if (open_flags & MDS_OPEN_DEFAULT_LMV) {
 336                         ma->ma_default_lmv = req_capsule_server_get(pill,
 337                                                         &RMF_DEFAULT_MDT_MD);
 338                         ma->ma_default_lmv_size = req_capsule_get_size(pill,
 339                                                         &RMF_DEFAULT_MDT_MD,
 340                                                         RCL_SERVER);
 341                         if (ma->ma_default_lmv_size > 0)
 342                                 ma->ma_need |= MA_LMV_DEF;
 343                 }
 344         } else {
 345                 ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
 346                 ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
 347                                                        RCL_SERVER);
 348                 if (ma->ma_lmm_size > 0)
 349                         ma->ma_need |= MA_LOV;
 350         }
 351 }
 352
 353 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 354                         struct mdt_object *o, u64 open_flags, int created,
 355                         struct ldlm_reply *rep)
 356 {
 357         struct ptlrpc_request *req = mdt_info_req(info);
 358         struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 359         struct mdt_file_data *mfd;
 360         struct md_attr *ma  = &info->mti_attr;
 361         struct lu_attr *la  = &ma->ma_attr;
 362         struct mdt_body *repbody;
 363         bool isdir, isreg;
 364         int rc = 0;
 365
 366         ENTRY;
 367         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 368
 369         isreg = S_ISREG(la->la_mode);
 370         isdir = S_ISDIR(la->la_mode);
 371         if (isreg && !(ma->ma_valid & MA_LOV) &&
 372             !(open_flags & MDS_OPEN_RELEASE)) {
 373                 /*
 374                  * No EA, check whether it is will set regEA and dirEA since in
 375                  * above attr get, these size might be zero, so reset it, to
 376                  * retrieve the MD after create obj.
 377                  */
 378                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 379                                                        &RMF_MDT_MD,
 380                                                        RCL_SERVER);
 381                 /* in replay case, p == NULL */
 382                 rc = mdt_create_data(info, p, o);
 383                 if (rc)
 384                         RETURN(rc);
 385
 386                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
 387                         mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
 388         }
 389
 390         CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n",
 391                ma->ma_valid, ma->ma_lmm_size);
 392
 393         if (ma->ma_valid & MA_LOV) {
 394                 LASSERT(ma->ma_lmm_size != 0);
 395                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 396                 if (isdir)
 397                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 398                 else
 399                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 400         }
 401
 402         if (ma->ma_valid & MA_LMV) {
 403                 LASSERT(ma->ma_lmv_size != 0);
 404                 repbody->mbo_eadatasize = ma->ma_lmv_size;
 405                 LASSERT(isdir);
 406                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
 407         }
 408
 409         if (ma->ma_valid & MA_LMV_DEF) {
 410                 LASSERT(ma->ma_default_lmv_size != 0);
 411                 LASSERT(isdir);
 412                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_DEFAULT_MEA;
 413         }
 414
 415         if (open_flags & MDS_FMODE_WRITE)
 416                 rc = mdt_write_get(o);
 417         else if (open_flags & MDS_FMODE_EXEC)
 418                 rc = mdt_write_deny(o);
 419
 420         if (rc)
 421                 RETURN(rc);
 422
 423         rc = mo_open(info->mti_env, mdt_object_child(o),
 424                      created ? open_flags | MDS_OPEN_CREATED : open_flags,
 425                      &info->mti_spec);
 426         if (rc != 0) {
 427                 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
 428                  * client does not know which suppgid should be sent to the MDS,
 429                  * or some other(s) changed the target file's GID after this RPC
 430                  * sent to the MDS with the suppgid as the original GID, then we
 431                  * give the client another chance to send the right suppgid. */
 432                 if (rc == -EACCES &&
 433                     allow_client_chgrp(info, lu_ucred(info->mti_env)))
 434                         mdt_set_disposition(info, rep, DISP_OPEN_DENY);
 435
 436                 GOTO(err_out, rc);
 437         }
 438
 439         mfd = mdt_mfd_new(med);
 440         if (mfd == NULL)
 441                 GOTO(err_out, rc = -ENOMEM);
 442
 443         /*
 444          * Keep a reference on this object for this open, and is
 445          * released by mdt_mfd_close().
 446          */
 447         mdt_object_get(info->mti_env, o);
 448         mfd->mfd_object = o;
 449         mfd->mfd_xid = req->rq_xid;
 450
 451         /*
 452          * @open_flags is always not zero. At least it should be FMODE_READ,
 453          * FMODE_WRITE or MDS_FMODE_EXEC.
 454          */
 455         LASSERT(open_flags != 0);
 456
 457         /* Open handling. */
 458         mdt_mfd_set_mode(mfd, open_flags);
 459
 460         atomic_inc(&o->mot_open_count);
 461         if (open_flags & MDS_OPEN_LEASE)
 462                 atomic_inc(&o->mot_lease_count);
 463
 464         /* replay handle */
 465         if (req_is_replay(req)) {
 466                 struct mdt_file_data *old_mfd;
 467                 /* Check wheather old cookie already exist in
 468                  * the list, becasue when do recovery, client
 469                  * might be disconnected from server, and
 470                  * restart replay, so there maybe some orphan
 471                  * mfd here, we should remove them */
 472                 LASSERT(info->mti_rr.rr_open_handle != NULL);
 473                 spin_lock(&med->med_open_lock);
 474                 old_mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 475                                               true);
 476                 if (old_mfd != NULL) {
 477                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 478                                "cookie = %#llx\n", mfd,
 479                                PFID(mdt_object_fid(mfd->mfd_object)),
 480                                info->mti_rr.rr_open_handle->cookie);
 481                         class_handle_unhash(&old_mfd->mfd_open_handle);
 482                         list_del_init(&old_mfd->mfd_list);
 483                         spin_unlock(&med->med_open_lock);
 484                         /* no attr update for that close */
 485                         la->la_valid = 0;
 486                         ma->ma_valid |= MA_FLAGS;
 487                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 488                         mdt_mfd_close(info, old_mfd);
 489                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 490                         ma->ma_valid &= ~MA_FLAGS;
 491                 } else {
 492                         spin_unlock(&med->med_open_lock);
 493                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 494                                "cookie = %#llx\n",
 495                                PFID(mdt_object_fid(mfd->mfd_object)),
 496                                info->mti_rr.rr_open_handle->cookie);
 497                 }
 498
 499                 CDEBUG(D_HA, "Store old cookie %#llx in new mfd\n",
 500                        info->mti_rr.rr_open_handle->cookie);
 501
 502                 mfd->mfd_open_handle_old = *info->mti_rr.rr_open_handle;
 503         }
 504
 505         repbody->mbo_open_handle.cookie = mfd->mfd_open_handle.h_cookie;
 506
 507         if (req->rq_export->exp_disconnected) {
 508                 spin_lock(&med->med_open_lock);
 509                 class_handle_unhash(&mfd->mfd_open_handle);
 510                 list_del_init(&mfd->mfd_list);
 511                 spin_unlock(&med->med_open_lock);
 512                 mdt_mfd_close(info, mfd);
 513         } else {
 514                 spin_lock(&med->med_open_lock);
 515                 list_add_tail(&mfd->mfd_list, &med->med_open_head);
 516                 spin_unlock(&med->med_open_lock);
 517         }
 518
 519         mdt_empty_transno(info, rc);
 520
 521         RETURN(rc);
 522
 523 err_out:
 524         if (open_flags & MDS_FMODE_WRITE)
 525                 mdt_write_put(o);
 526         else if (open_flags & MDS_FMODE_EXEC)
 527                 mdt_write_allow(o);
 528
 529         return rc;
 530 }
 531
 532 static int mdt_finish_open(struct mdt_thread_info *info,
 533                            struct mdt_object *p, struct mdt_object *o,
 534                            u64 open_flags,
 535                            struct ldlm_reply *rep)
 536 {
 537         struct ptlrpc_request *req = mdt_info_req(info);
 538         struct obd_export *exp = req->rq_export;
 539         struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 540         struct md_attr *ma  = &info->mti_attr;
 541         struct lu_attr *la  = &ma->ma_attr;
 542         struct mdt_file_data *mfd;
 543         struct mdt_body *repbody;
 544         int created;
 545         int rc = 0;
 546         int isreg, isdir, islnk;
 547         struct list_head *t;
 548
 549         ENTRY;
 550         LASSERT(ma->ma_valid & MA_INODE);
 551         created = mdt_get_disposition(rep, DISP_OPEN_CREATE);
 552
 553         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 554
 555         isreg = S_ISREG(la->la_mode);
 556         isdir = S_ISDIR(la->la_mode);
 557         islnk = S_ISLNK(la->la_mode);
 558         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 559
 560         /* compatibility check for 2.10 clients when it tries to open mirrored
 561          * files. 2.10 clients don't verify overlapping components so they
 562          * would read and write mirrored files just as if they were normal
 563          * PFL files, which will cause the problem that sycned mirrors actually
 564          * contain different data.
 565          * Older clients are not a concern here because they don't even
 566          * understand PFL layout. */
 567         if (isreg && !exp_connect_flr(exp) && ma->ma_valid & MA_LOV &&
 568             mdt_lmm_is_flr(ma->ma_lmm)) {
 569                 /* LU-10286: for simplicity clients who don't understand
 570                  * mirrored layout(with connect flag OBD_CONNECT2_FLR) won't
 571                  * be able to open mirrored files */
 572                 RETURN(-EOPNOTSUPP);
 573         }
 574
 575         /* Overstriped files can crash older clients */
 576         if (isreg && !exp_connect_overstriping(exp) &&
 577             mdt_lmm_is_overstriping(ma->ma_lmm))
 578                 RETURN(-EOPNOTSUPP);
 579
 580         /* LU-2275, simulate broken behaviour (esp. prevalent in
 581          * pre-2.4 servers where a very strange reply is sent on error
 582          * that looks like it was actually almost successful and a
 583          * failure at the same time.) */
 584         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 585                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 586                                                DISP_LOOKUP_NEG |
 587                                                DISP_LOOKUP_POS);
 588
 589                 if (open_flags & MDS_OPEN_LOCK)
 590                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 591
 592                 RETURN(-ENOENT);
 593         }
 594
 595 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
 596         if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 597                 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
 598                 if (IS_ERR(nodemap))
 599                         RETURN(PTR_ERR(nodemap));
 600
 601                 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
 602                 nodemap_putref(nodemap);
 603                 if (rc)
 604                         RETURN(rc);
 605         }
 606 #endif
 607
 608         /*
 609          * If we are following a symlink, don't open; and do not return open
 610          * handle for special nodes as client required.
 611          */
 612         if (islnk || (!isreg && !isdir &&
 613             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 614                 lustre_msg_set_transno(req->rq_repmsg, 0);
 615                 RETURN(0);
 616         }
 617
 618         /*
 619          * We need to return the existing object's fid back, so it is done here,
 620          * after preparing the reply.
 621          */
 622         if (!created && (open_flags & MDS_OPEN_EXCL) &&
 623             (open_flags & MDS_OPEN_CREAT))
 624                 RETURN(-EEXIST);
 625
 626         /* This can't be done earlier, we need to return reply body */
 627         if (isdir) {
 628                 if (open_flags & (MDS_OPEN_CREAT | MDS_FMODE_WRITE)) {
 629                         /* We are trying to create or write an existing dir. */
 630                         RETURN(-EISDIR);
 631                 }
 632         } else if (open_flags & MDS_OPEN_DIRECTORY)
 633                 RETURN(-ENOTDIR);
 634
 635         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 636                                  OBD_FAIL_MDS_LDLM_REPLY_NET | OBD_FAIL_ONCE))
 637                 RETURN(-EAGAIN);
 638
 639         mfd = NULL;
 640         if (info->mti_rr.rr_flags & MRF_OPEN_RESEND) {
 641                 spin_lock(&med->med_open_lock);
 642                 list_for_each(t, &med->med_open_head) {
 643                         mfd = list_entry(t, struct mdt_file_data, mfd_list);
 644                         if (mfd->mfd_xid == req->rq_xid) {
 645                                 repbody->mbo_open_handle.cookie =
 646                                                 mfd->mfd_open_handle.h_cookie;
 647                                 break;
 648                         }
 649                         mfd = NULL;
 650                 }
 651                 spin_unlock(&med->med_open_lock);
 652
 653                 if (mfd != NULL) {
 654                         /* set repbody->ea_size for resent case */
 655                         if (ma->ma_valid & MA_LOV) {
 656                                 LASSERT(ma->ma_lmm_size != 0);
 657                                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 658                                 if (isdir)
 659                                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 660                                 else
 661                                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 662                         }
 663                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 664                         RETURN(0);
 665                 }
 666                 /* if we have a real resend (not a resend afrer failover), it
 667                  * means close is already happend, so lets return error
 668                  */
 669                 RETURN(-ESTALE);
 670         }
 671
 672         rc = mdt_mfd_open(info, p, o, open_flags, created, rep);
 673         if (!rc)
 674                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 675
 676         RETURN(rc);
 677 }
 678
 679 void mdt_reconstruct_open(struct mdt_thread_info *info,
 680                           struct mdt_lock_handle *lhc)
 681 {
 682         struct req_capsule *pill = info->mti_pill;
 683         struct ptlrpc_request *req = mdt_info_req(info);
 684         struct mdt_reint_record *rr = &info->mti_rr;
 685         struct md_attr *ma = &info->mti_attr;
 686         struct ldlm_reply *ldlm_rep;
 687         u64 opdata;
 688         int rc;
 689
 690         ENTRY;
 691         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 692         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 693
 694         ma->ma_need = MA_INODE | MA_HSM;
 695         ma->ma_valid = 0;
 696         opdata = mdt_req_from_lrd(req, info->mti_reply_data);
 697         mdt_set_disposition(info, ldlm_rep, opdata);
 698
 699         CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n",
 700                ldlm_rep->lock_policy_res1, req->rq_status);
 701         if (req->rq_status)
 702                 /* We did not create successfully, return error to client. */
 703                 GOTO(out, rc = req->rq_status);
 704
 705         /* tg_reply_data is just memory only  structure, so any non zero fid
 706          * means a real resend not a resend after recovery which need to be
 707          * handled as regular open
 708          */
 709         if (likely(!fid_is_zero(&info->mti_reply_data->trd_object))) {
 710                 rr->rr_fid2 = &info->mti_reply_data->trd_object;
 711                 rr->rr_flags |= MRF_OPEN_RESEND;
 712                 rc = mdt_open_by_fid(info, ldlm_rep, lhc);
 713                 if (rc)
 714                         lustre_msg_set_transno(req->rq_repmsg, 0);
 715         } else {
 716                 /* We did not try to create, so we are a pure open */
 717                 rc = mdt_reint_open(info, lhc);
 718         }
 719         EXIT;
 720 out:
 721         req->rq_status = rc;
 722         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 723         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
 724 }
 725
 726 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep,
 727                            struct mdt_lock_handle *lhc)
 728 {
 729         u64 open_flags = info->mti_spec.sp_cr_flags;
 730         struct mdt_reint_record *rr = &info->mti_rr;
 731         struct md_attr *ma = &info->mti_attr;
 732         struct mdt_object *o;
 733         int rc;
 734
 735         ENTRY;
 736         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
 737         if (IS_ERR(o))
 738                 RETURN(rc = PTR_ERR(o));
 739
 740         rc = mdt_check_enc(info, o);
 741         if (rc)
 742                 GOTO(out, rc);
 743
 744         if (unlikely(mdt_object_remote(o))) {
 745                 /* the child object was created on remote server */
 746                 struct mdt_body *repbody;
 747
 748                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 749                                                 DISP_LOOKUP_EXECD |
 750                                                 DISP_LOOKUP_POS));
 751                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 752                 repbody->mbo_fid1 = *rr->rr_fid2;
 753                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 754                 rc = 0;
 755         } else {
 756                 if (mdt_object_exists(o)) {
 757                         tgt_open_obj_set(info->mti_env, mdt_obj2dt(o));
 758                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 759                                                         DISP_LOOKUP_EXECD |
 760                                                         DISP_LOOKUP_POS));
 761                         if ((open_flags & MDS_OPEN_EXCL) &&
 762                             (open_flags & MDS_OPEN_CREAT))
 763                                 mdt_set_disposition(info, rep,
 764                                                     DISP_OPEN_CREATE);
 765
 766                         mdt_prep_ma_buf_from_rep(info, o, ma, open_flags);
 767                         rc = mdt_attr_get_complex(info, o, ma);
 768                         if (rc)
 769                                 GOTO(out, rc);
 770                         rc = mdt_finish_open(info, NULL, o, open_flags, rep);
 771                         if (rc)
 772                                 GOTO(out, rc);
 773                         mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
 774                 } else {
 775                         rc = -ENOENT;
 776                 }
 777         }
 778
 779 out:
 780         mdt_object_put(info->mti_env, o);
 781         RETURN(rc);
 782 }
 783
 784 /* lock object for open */
 785 static int mdt_object_open_lock(struct mdt_thread_info *info,
 786                                 struct mdt_object *obj,
 787                                 struct mdt_lock_handle *lhc,
 788                                 __u64 *ibits)
 789 {
 790         struct md_attr *ma = &info->mti_attr;
 791         __u64 open_flags = info->mti_spec.sp_cr_flags;
 792         __u64 trybits = 0;
 793         enum ldlm_mode lm = LCK_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 (!OBD_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                                          false);
 922
 923         CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
 924                ", open_flags = %#llo, try_layout = %d : rc = %d\n",
 925                mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
 926                *ibits, trybits, open_flags, try_layout, rc);
 927
 928         /* will change layout, revoke layout locks by enqueuing EX lock. */
 929         if (rc == 0 && create_layout) {
 930                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
 931
 932                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
 933                         ", open_flags = %#llo\n",
 934                         PFID(mdt_object_fid(obj)), open_flags);
 935
 936                 /* We cannot enqueue another lock for the same resource we
 937                  * already have a lock for, due to mechanics of waiting list
 938                  * iterating in ldlm, see LU-3601.
 939                  * As such we'll drop the open lock we just got above here,
 940                  * it's ok not to have this open lock as it's main purpose is to
 941                  * flush unused cached client open handles. */
 942                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 943                         mdt_object_unlock(info, obj, lhc, 1);
 944
 945                 LASSERT(!try_layout);
 946                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
 947                                      LCK_EX, false);
 948
 949                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
 950         }
 951
 952         /* Check if there is any other open handles after acquiring
 953          * open lock. At this point, caching open handles have been revoked
 954          * by open lock.
 955          * XXX: Now only exclusive open is supported. Need to check the
 956          * type of open for generic lease support. */
 957         if (rc == 0 && acq_lease) {
 958                 struct ptlrpc_request *req = mdt_info_req(info);
 959                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 960                 struct mdt_file_data *mfd;
 961                 bool is_replay_or_resent;
 962                 int open_count = 0;
 963
 964                 /* For lease: application can open a file and then apply lease,
 965                  * @handle contains original open handle in that case.
 966                  * In recovery, open REQ will be replayed and the lease REQ may
 967                  * be resent that means the open handle is already stale, so we
 968                  * need to fix it up here by finding new handle. */
 969                 is_replay_or_resent = req_is_replay(req) ||
 970                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
 971
 972                 /* if the request is _not_ a replay request, rr_open_handle
 973                  * may be used to hold an open file handle which is issuing the
 974                  * lease request, so that this openhandle doesn't count. */
 975                 mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 976                                           is_replay_or_resent);
 977                 if (mfd != NULL)
 978                         ++open_count;
 979
 980                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
 981                         PFID(mdt_object_fid(obj)),
 982                         atomic_read(&obj->mot_open_count), open_count);
 983
 984                 if (atomic_read(&obj->mot_open_count) > open_count) {
 985                         /* fail if anyone *else* has opened file for write */
 986                         if (mdt_write_read(obj) > 1)
 987                                 GOTO(out, rc = -EBUSY);
 988                 }
 989         }
 990         GOTO(out, rc);
 991
 992 out:
 993         RETURN(rc);
 994 }
 995
 996 static void mdt_object_open_unlock(struct mdt_thread_info *info,
 997                                    struct mdt_object *obj,
 998                                    struct mdt_lock_handle *lhc,
 999                                    __u64 ibits, int rc)
1000 {
1001         __u64 open_flags = info->mti_spec.sp_cr_flags;
1002         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1003
1004         ENTRY;
1005         if (req_is_replay(mdt_info_req(info)))
1006                 RETURN_EXIT;
1007
1008         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1009          * return to client side. */
1010         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1011                 mdt_object_unlock(info, obj, ll, 1);
1012
1013         ll = &info->mti_lh[MDT_LH_LAYOUT];
1014         /* Release local layout lock, layout was created */
1015         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1016                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1017                 mdt_object_unlock(info, obj, ll, 1);
1018         }
1019
1020         if (open_flags & MDS_OPEN_LEASE)
1021                 up_write(&obj->mot_open_sem);
1022         else
1023                 up_read(&obj->mot_open_sem);
1024
1025         /* Cross-ref case, the lock should be returned to the client */
1026         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1027                 RETURN_EXIT;
1028
1029         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1030             !(ibits & MDS_INODELOCK_DOM)) {
1031                 /* for the open request, the lock will only return to client
1032                  * if open or layout lock is granted. */
1033                 rc = 1;
1034         }
1035
1036         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1037                 struct ldlm_reply       *ldlm_rep;
1038
1039                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1040                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1041                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1042                         mdt_object_unlock(info, obj, lhc, 1);
1043         }
1044         RETURN_EXIT;
1045 }
1046
1047 /**
1048  * Check release is permitted for the current HSM flags.
1049  */
1050 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1051 {
1052         if (!(ma->ma_valid & MA_HSM))
1053                 return false;
1054
1055         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1056                 return false;
1057
1058         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1059                 return false;
1060
1061         return true;
1062 }
1063
1064 static int mdt_refetch_lovea(struct mdt_thread_info *info,
1065                              struct mdt_object *o, struct md_attr *ma,
1066                              u64 ibits)
1067 {
1068         struct mdt_body *repbody;
1069         int rc;
1070
1071         if ((ibits & MDS_INODELOCK_LAYOUT) == 0)
1072                 return 0;
1073         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
1074                 return 0;
1075
1076         if ((ma->ma_valid & MA_LOV) == 0)
1077                 return 0;
1078
1079         ma->ma_valid &= ~MA_LOV;
1080         info->mti_big_lmm_used = 0;
1081         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1082         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1083                                                RCL_SERVER);
1084         rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1085         if (rc)
1086                 return rc;
1087
1088         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1089         repbody->mbo_eadatasize = ma->ma_lmm_size;
1090         return 0;
1091 }
1092
1093 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1094                                 struct ldlm_reply *rep,
1095                                 struct mdt_lock_handle *lhc)
1096 {
1097         const struct lu_env *env = info->mti_env;
1098         struct mdt_device *mdt = info->mti_mdt;
1099         u64 open_flags = info->mti_spec.sp_cr_flags;
1100         struct mdt_reint_record *rr = &info->mti_rr;
1101         struct md_attr *ma = &info->mti_attr;
1102         struct mdt_object *parent = NULL;
1103         struct mdt_object *o;
1104         bool object_locked = false;
1105         u64 ibits = 0;
1106         int rc;
1107
1108         ENTRY;
1109         if (md_should_create(open_flags)) {
1110                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1111                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1112                         if (IS_ERR(parent)) {
1113                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1114                                        " for anonymous created %ld, try to"
1115                                        " use server-side parent.\n",
1116                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1117                                 parent = NULL;
1118                         }
1119                 }
1120                 if (parent == NULL)
1121                         ma->ma_need |= MA_PFID;
1122         }
1123
1124         o = mdt_object_find(env, mdt, rr->rr_fid2);
1125         if (IS_ERR(o))
1126                 GOTO(out_parent_put, rc = PTR_ERR(o));
1127
1128         if (mdt_object_remote(o)) {
1129                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1130                        mdt_obd_name(info->mti_mdt),
1131                        PFID(rr->rr_fid2));
1132                 GOTO(out, rc = -EREMOTE);
1133         } else if (!mdt_object_exists(o)) {
1134                 mdt_set_disposition(info, rep,
1135                                     DISP_IT_EXECD |
1136                                     DISP_LOOKUP_EXECD |
1137                                     DISP_LOOKUP_NEG);
1138                 GOTO(out, rc = -ENOENT);
1139         }
1140
1141         /* do not check enc for directory: always allow open */
1142         if (!S_ISDIR(lu_object_attr(&o->mot_obj))) {
1143                 rc = mdt_check_enc(info, o);
1144                 if (rc)
1145                         GOTO(out, rc);
1146         }
1147
1148         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1149
1150         mdt_prep_ma_buf_from_rep(info, o, ma, open_flags);
1151         if (open_flags & MDS_OPEN_RELEASE)
1152                 ma->ma_need |= MA_HSM;
1153         rc = mdt_attr_get_complex(info, o, ma);
1154         if (rc)
1155                 GOTO(out, rc);
1156
1157         /* We should not change file's existing LOV EA */
1158         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1159             open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1160                 GOTO(out, rc = -EEXIST);
1161
1162         /* If a release request, check file open flags are fine and ask for an
1163          * exclusive open access. */
1164         if (open_flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1165                 GOTO(out, rc = -EPERM);
1166
1167         rc = mdt_check_resent_lock(info, o, lhc);
1168         if (rc < 0) {
1169                 GOTO(out, rc);
1170         } else if (rc > 0) {
1171                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1172                 object_locked = true;
1173                 if (rc)
1174                         GOTO(out_unlock, rc);
1175         }
1176
1177         if (ma->ma_valid & MA_PFID) {
1178                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1179                 if (IS_ERR(parent)) {
1180                         CDEBUG(D_INODE, "Fail to find parent "DFID
1181                                " for anonymous created %ld, try to"
1182                                " use system default.\n",
1183                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1184                         parent = NULL;
1185                 }
1186         }
1187
1188         tgt_open_obj_set(info->mti_env, mdt_obj2dt(o));
1189         rc = mdt_finish_open(info, parent, o, open_flags, rep);
1190         if (rc)
1191                 GOTO(out_unlock, rc);
1192
1193         mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1194         if (open_flags & MDS_OPEN_LOCK)
1195                 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1196         if (open_flags & MDS_OPEN_LEASE)
1197                 mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1198
1199         /*
1200          * if layout lock is granted, then we should re-fetch LOVEA
1201          * which was originally taken w/o the lock
1202          */
1203         rc = mdt_refetch_lovea(info, o, ma, ibits);
1204
1205         GOTO(out_unlock, rc);
1206
1207 out_unlock:
1208         if (object_locked)
1209                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1210 out:
1211         mdt_object_put(env, o);
1212         if (rc == 0)
1213                 mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
1214 out_parent_put:
1215         if (parent != NULL)
1216                 mdt_object_put(env, parent);
1217         return rc;
1218 }
1219
1220 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1221 static int mdt_cross_open(struct mdt_thread_info *info,
1222                           const struct lu_fid *parent_fid,
1223                           const struct lu_fid *fid,
1224                           struct ldlm_reply *rep, u64 open_flags)
1225 {
1226         struct md_attr *ma = &info->mti_attr;
1227         struct mdt_object *o;
1228         int rc;
1229
1230         ENTRY;
1231         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1232         if (IS_ERR(o))
1233                 RETURN(rc = PTR_ERR(o));
1234
1235         rc = mdt_check_enc(info, o);
1236         if (rc)
1237                 GOTO(out, rc);
1238
1239         if (mdt_object_remote(o)) {
1240                 /* Something is wrong here, the object is on another MDS! */
1241                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1242                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1243                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1244                                 &o->mot_obj,
1245                                 "Object isn't on this server! FLD error?");
1246                 rc = -EFAULT;
1247         } else {
1248                 if (mdt_object_exists(o)) {
1249                         int mask;
1250
1251                         /* Do permission check for cross-open after converting
1252                          * MDS_OPEN_* flags to MAY_* permission mask.
1253                          */
1254                         mask = mds_accmode(open_flags);
1255
1256                         rc = mo_permission(info->mti_env, NULL,
1257                                            mdt_object_child(o), NULL, mask);
1258                         if (rc)
1259                                 goto out;
1260
1261                         mdt_prep_ma_buf_from_rep(info, o, ma, open_flags);
1262                         rc = mdt_attr_get_complex(info, o, ma);
1263                         if (rc != 0)
1264                                 GOTO(out, rc);
1265
1266                         rc = mdt_pack_secctx_in_reply(info, o);
1267                         if (unlikely(rc))
1268                                 GOTO(out, rc);
1269
1270                         rc = mdt_pack_encctx_in_reply(info, o);
1271                         if (unlikely(rc))
1272                                 GOTO(out, rc);
1273
1274                         rc = mdt_finish_open(info, NULL, o, open_flags, rep);
1275                 } else {
1276                         /*
1277                          * Something is wrong here. lookup was positive but
1278                          * there is no object!
1279                          */
1280                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1281                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1282                         rc = -EFAULT;
1283                 }
1284         }
1285 out:
1286         mdt_object_put(info->mti_env, o);
1287
1288         RETURN(rc);
1289 }
1290
1291 /*
1292  * find root object and take its xattr lock if it's on remote MDT, later create
1293  * may use fs default striping (which is stored in root xattr).
1294  */
1295 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1296                                struct mdt_device *mdt)
1297 {
1298         struct mdt_object *md_root = mdt->mdt_md_root;
1299         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1300         __u64 ibits = MDS_INODELOCK_XATTR;
1301         int rc;
1302
1303         if (md_root == NULL) {
1304                 lu_root_fid(&info->mti_tmp_fid1);
1305                 md_root = mdt_object_find(info->mti_env, mdt,
1306                                           &info->mti_tmp_fid1);
1307                 if (IS_ERR(md_root))
1308                         return PTR_ERR(md_root);
1309
1310                 spin_lock(&mdt->mdt_lock);
1311                 if (mdt->mdt_md_root != NULL) {
1312                         spin_unlock(&mdt->mdt_lock);
1313
1314                         LASSERTF(mdt->mdt_md_root == md_root,
1315                                  "Different root object ("
1316                                  DFID") instances, %p, %p\n",
1317                                  PFID(&info->mti_tmp_fid1),
1318                                  mdt->mdt_md_root, md_root);
1319                         LASSERT(atomic_read(
1320                                 &md_root->mot_obj.lo_header->loh_ref) > 1);
1321
1322                         mdt_object_put(info->mti_env, md_root);
1323                 } else {
1324                         mdt->mdt_md_root = md_root;
1325                         spin_unlock(&mdt->mdt_lock);
1326                 }
1327         }
1328
1329         if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1330                 return 0;
1331
1332         mdt_lock_reg_init(lh, LCK_PR);
1333         rc = mdt_object_lock_internal(info, md_root, mdt_object_fid(md_root),
1334                                       lh, &ibits, 0, true, false);
1335         if (rc < 0)
1336                 return rc;
1337
1338         md_root->mot_cache_attr = 1;
1339
1340         /* don't cancel this lock, so that we know the cached xattr is valid. */
1341         ldlm_lock_decref(&lh->mlh_rreg_lh, LCK_PR);
1342         lh->mlh_rreg_lh.cookie = 0ull;
1343
1344         return 0;
1345 }
1346
1347 static inline enum ldlm_mode mdt_open_lock_mode(struct mdt_thread_info *info,
1348                                                 struct mdt_object *p,
1349                                                 struct lu_name *name,
1350                                                 u64 open_flags)
1351 {
1352         int result;
1353         struct lu_fid fid;
1354
1355         /* We don't need to take the DLM lock for a volatile */
1356         if (open_flags & MDS_OPEN_VOLATILE)
1357                 return LCK_NL;
1358
1359         if (!(open_flags & MDS_OPEN_CREAT))
1360                 return LCK_PR;
1361
1362         result = mdo_lookup(info->mti_env, mdt_object_child(p), name, &fid,
1363                             &info->mti_spec);
1364
1365         /* If the file exists we only need a read lock on the parent */
1366         return (result == 0) ? LCK_PR : LCK_PW;
1367 }
1368
1369 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1370 {
1371         struct mdt_device *mdt = info->mti_mdt;
1372         struct ptlrpc_request *req = mdt_info_req(info);
1373         struct mdt_object *parent;
1374         struct mdt_object *child;
1375         struct mdt_lock_handle *lh = NULL;
1376         struct ldlm_reply *ldlm_rep;
1377         struct mdt_body *repbody;
1378         struct lu_fid *child_fid = &info->mti_tmp_fid1;
1379         struct lu_ucred *uc = mdt_ucred(info);
1380         struct md_attr *ma = &info->mti_attr;
1381         u64 open_flags = info->mti_spec.sp_cr_flags;
1382         u64 ibits = 0;
1383         struct mdt_reint_record *rr = &info->mti_rr;
1384         int result, rc;
1385         int created = 0;
1386         int object_locked = 0;
1387         enum ldlm_mode lock_mode;
1388         u32 msg_flags;
1389         ktime_t kstart = ktime_get();
1390
1391         ENTRY;
1392         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1393                                (obd_timeout + 1) / 4);
1394
1395         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1396
1397         ma->ma_need = MA_INODE;
1398         ma->ma_valid = 0;
1399
1400         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1401         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1402
1403         if (unlikely(open_flags & MDS_OPEN_JOIN_FILE)) {
1404                 CERROR("file join is not supported anymore.\n");
1405                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1406         }
1407         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1408
1409         if ((open_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1410             info->mti_spec.u.sp_ea.eadata == NULL)
1411                 GOTO(out, result = err_serious(-EINVAL));
1412
1413         if (open_flags & MDS_FMODE_WRITE &&
1414             exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1415                 GOTO(out, result = -EROFS);
1416
1417         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1418                "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1419                PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2),
1420                open_flags, ma->ma_attr.la_mode, msg_flags);
1421
1422         /* Prevent by-fid operation if parent fid is .lustre/fid.
1423          * Also, we want rbac roles to have precedence over any other
1424          * permission or capability checks
1425          */
1426         if (lu_fid_eq(rr->rr_fid1, &LU_OBF_FID) && !uc->uc_rbac_byfid_ops)
1427                 GOTO(out, result = -EPERM);
1428
1429         if (info->mti_cross_ref) {
1430                 /* This is cross-ref open */
1431                 mdt_set_disposition(info, ldlm_rep,
1432                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1433                              DISP_LOOKUP_POS));
1434                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1435                                         ldlm_rep, open_flags);
1436                 GOTO(out, result);
1437         } else if (req_is_replay(req)) {
1438                 result = mdt_open_by_fid(info, ldlm_rep, lhc);
1439
1440                 if (result != -ENOENT)
1441                         GOTO(out, result);
1442
1443                 /* We didn't find the correct object, so we need to re-create it
1444                  * via a regular replay. */
1445                 if (!(open_flags & MDS_OPEN_CREAT)) {
1446                         DEBUG_REQ(D_ERROR, req,
1447                                   "OPEN & CREAT not in open replay/by_fid");
1448                         GOTO(out, result = -EFAULT);
1449                 }
1450                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1451         } else if (open_flags & MDS_OPEN_BY_FID) {
1452                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1453                 if (result < 0)
1454                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1455                                PFID(rr->rr_fid2), result);
1456                 GOTO(out, result);
1457         }
1458
1459         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1460                 GOTO(out, result = err_serious(-ENOMEM));
1461
1462         mdt_set_disposition(info, ldlm_rep,
1463                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1464
1465         if (!lu_name_is_valid(&rr->rr_name))
1466                 GOTO(out, result = -EPROTO);
1467
1468         result = mdt_lock_root_xattr(info, mdt);
1469         if (result < 0)
1470                 GOTO(out, result);
1471
1472         parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1473         if (IS_ERR(parent))
1474                 GOTO(out, result = PTR_ERR(parent));
1475
1476         /* get and check version of parent */
1477         result = mdt_version_get_check(info, parent, 0);
1478         if (result) {
1479                 mdt_object_put(info->mti_env, parent);
1480                 GOTO(out, result);
1481         }
1482
1483         if (!uc->uc_rbac_fscrypt_admin &&
1484             parent->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD &&
1485             open_flags & MDS_OPEN_CREAT)
1486                 GOTO(out_parent, result = -EPERM);
1487
1488         result = mdt_check_enc(info, parent);
1489         if (result)
1490                 GOTO(out_parent, result);
1491
1492         fid_zero(child_fid);
1493         result = -ENOENT;
1494         lock_mode = mdt_open_lock_mode(info, parent, &rr->rr_name, open_flags);
1495
1496         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
1497 again_pw:
1498         if (lock_mode != LCK_NL) {
1499                 lh = &info->mti_lh[MDT_LH_PARENT];
1500                 result = mdt_parent_lock(info, parent, lh, &rr->rr_name,
1501                                          lock_mode, false);
1502                 if (result != 0)
1503                         GOTO(out_parent, result);
1504
1505                 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1506                                     &rr->rr_name, child_fid, &info->mti_spec);
1507         }
1508
1509         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1510                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1511                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1512                  PFID(child_fid));
1513
1514         if (result != 0 && result != -ENOENT)
1515                 GOTO(out_parent_unlock, result);
1516
1517         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1518
1519         if (result == -ENOENT) {
1520                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1521                 if (!(open_flags & MDS_OPEN_CREAT))
1522                         GOTO(out_parent_unlock, result);
1523                 if (mdt_rdonly(req->rq_export))
1524                         GOTO(out_parent_unlock, result = -EROFS);
1525
1526                 LASSERT(equi(lh == NULL, lock_mode == LCK_NL));
1527
1528                 if (lock_mode == LCK_PR) {
1529                         /* unlink vs create race: get write lock and restart */
1530                         mdt_object_unlock(info, parent, lh, 1);
1531                         mdt_clear_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1532                         lock_mode = LCK_PW;
1533                         goto again_pw;
1534                 }
1535
1536                 *child_fid = *info->mti_rr.rr_fid2;
1537                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1538                          PFID(child_fid));
1539                 /* In the function below, .hs_keycmp resolves to
1540                  * lu_obj_hop_keycmp() */
1541                 /* coverity[overrun-buffer-val] */
1542                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1543         } else {
1544                 /*
1545                  * Check for O_EXCL is moved to the mdt_finish_open(),
1546                  * we need to return FID back in that case.
1547                  */
1548                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1549                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1550         }
1551         if (IS_ERR(child))
1552                 GOTO(out_parent_unlock, result = PTR_ERR(child));
1553
1554         /** check version of child  */
1555         rc = mdt_version_get_check(info, child, 1);
1556         if (rc)
1557                 GOTO(out_child, result = rc);
1558
1559         tgt_open_obj_set(info->mti_env, mdt_obj2dt(child));
1560
1561         if (result == -ENOENT) {
1562                 /* Create under OBF and .lustre is not permitted */
1563                 if (!fid_is_md_operative(rr->rr_fid1) &&
1564                     (open_flags & MDS_OPEN_VOLATILE) == 0)
1565                         GOTO(out_child, result = -EPERM);
1566
1567                 /* save versions in reply */
1568                 mdt_version_get_save(info, parent, 0);
1569                 mdt_version_get_save(info, child, 1);
1570
1571                 /* version of child will be changed */
1572                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1573
1574                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1575                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1576
1577                 /* Don't do lookup sanity check. We know name doesn't exist. */
1578                 info->mti_spec.sp_cr_lookup = 0;
1579                 info->mti_spec.sp_feat = &dt_directory_features;
1580
1581                 result = mdo_create(info->mti_env, mdt_object_child(parent),
1582                                     &rr->rr_name, mdt_object_child(child),
1583                                     &info->mti_spec, &info->mti_attr);
1584                 if (result == -ERESTART) {
1585                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1586                         GOTO(out_child, result);
1587                 } else {
1588                         mdt_prep_ma_buf_from_rep(info, child, ma, open_flags);
1589                         /* XXX: we should call this once, see few lines below */
1590                         if (result == 0)
1591                                 result = mdt_attr_get_complex(info, child, ma);
1592
1593                         if (result != 0)
1594                                 GOTO(out_child, result);
1595                 }
1596                 created = 1;
1597                 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1598                                  ktime_us_delta(ktime_get(), kstart));
1599         } else {
1600                 /*
1601                  * The object is on remote node, return its FID for remote open.
1602                  */
1603                 if (mdt_object_remote(child)) {
1604                         /*
1605                          * Check if this lock already was sent to client and
1606                          * this is resent case. For resent case do not take lock
1607                          * again, use what is already granted.
1608                          */
1609                         LASSERT(lhc != NULL);
1610
1611                         rc = mdt_check_resent_lock(info, child, lhc);
1612                         if (rc < 0)
1613                                 GOTO(out_child, result = rc);
1614                         else if (rc > 0)
1615                                 rc = mdt_object_lookup_lock(info, NULL, child,
1616                                                             lhc, LCK_PR, false);
1617                         repbody->mbo_fid1 = *mdt_object_fid(child);
1618                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1619                         if (rc != 0)
1620                                 result = rc;
1621                         else
1622                                 result = -MDT_EREMOTE_OPEN;
1623                         GOTO(out_child, result);
1624                 } else if (mdt_object_exists(child)) {
1625                         /* Check early for MDS_OPEN_DIRECTORY/O_DIRECTORY to
1626                          * avoid opening regular files from lfs getstripe
1627                          * since doing so breaks the leases used by lfs
1628                          * mirror. See LU-13693. */
1629                         if (open_flags & MDS_OPEN_DIRECTORY &&
1630                             S_ISREG(lu_object_attr(&child->mot_obj)))
1631                                 GOTO(out_child, result = -ENOTDIR);
1632
1633                         /* We have to get attr & LOV EA & HSM for this
1634                          * object. */
1635                         mdt_prep_ma_buf_from_rep(info, child, ma, open_flags);
1636                         ma->ma_need |= MA_HSM;
1637                         result = mdt_attr_get_complex(info, child, ma);
1638                         if (result != 0)
1639                                 GOTO(out_child, result);
1640                 } else {
1641                         /* Object does not exist. Likely FS corruption. */
1642                         CERROR("%s: name '"DNAME"' present, but FID "
1643                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1644                                PNAME(&rr->rr_name), PFID(child_fid));
1645                         GOTO(out_child, result = -EIO);
1646                 }
1647         }
1648
1649         repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1650         repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1651
1652         rc = mdt_pack_secctx_in_reply(info, child);
1653         if (unlikely(rc))
1654                 GOTO(out_child, result = rc);
1655
1656         rc = mdt_pack_encctx_in_reply(info, child);
1657         if (unlikely(rc))
1658                 GOTO(out_child, result = rc);
1659
1660         rc = mdt_check_resent_lock(info, child, lhc);
1661         if (rc < 0) {
1662                 GOTO(out_child, result = rc);
1663         } else if (rc == 0) {
1664                 /* the open lock might already be gotten in
1665                  * ldlm_handle_enqueue() */
1666                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1667                 if (open_flags & MDS_OPEN_LOCK)
1668                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1669         } else {
1670                 /* get openlock if this isn't replay and client requested it */
1671                 if (!req_is_replay(req)) {
1672                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_DELAY_OPEN, cfs_fail_val);
1673                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1674                         object_locked = 1;
1675                         if (rc != 0)
1676                                 GOTO(out_child_unlock, result = rc);
1677                         else if (open_flags & MDS_OPEN_LOCK)
1678                                 mdt_set_disposition(info, ldlm_rep,
1679                                                     DISP_OPEN_LOCK);
1680                 }
1681         }
1682         /* Try to open it now. */
1683         rc = mdt_finish_open(info, parent, child, open_flags, ldlm_rep);
1684         if (rc) {
1685                 result = rc;
1686                 /* openlock will be released if mdt_finish_open() failed */
1687                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1688
1689                 if (created && (open_flags & MDS_OPEN_VOLATILE)) {
1690                         CERROR("%s: cannot open volatile file "DFID", orphan "
1691                                "file will be left in PENDING directory until "
1692                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1693                                PFID(mdt_object_fid(child)), rc);
1694                         GOTO(out_child_unlock, result);
1695                 }
1696
1697                 if (created) {
1698                         ma->ma_need = 0;
1699                         ma->ma_valid = 0;
1700                         rc = mdo_unlink(info->mti_env,
1701                                         mdt_object_child(parent),
1702                                         mdt_object_child(child),
1703                                         &rr->rr_name,
1704                                         &info->mti_attr, 0);
1705                         if (rc != 0)
1706                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1707                                        mdt_obd_name(info->mti_mdt),
1708                                        PFID(mdt_object_fid(child)), rc);
1709                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1710                 }
1711                 GOTO(out_child_unlock, result);
1712         }
1713
1714         /*
1715          * if layout lock is granted, then we should re-fetch LOVEA
1716          * which was originally taken w/o the lock.
1717          */
1718         result = mdt_refetch_lovea(info, child, ma, ibits);
1719
1720         mdt_counter_incr(req, LPROC_MDT_OPEN,
1721                          ktime_us_delta(ktime_get(), kstart));
1722
1723         GOTO(out_child_unlock, result);
1724
1725 out_child_unlock:
1726         if (object_locked)
1727                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1728 out_child:
1729         mdt_object_put(info->mti_env, child);
1730         if (result == 0)
1731                 mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh);
1732 out_parent_unlock:
1733         if (lh != NULL)
1734                 mdt_object_unlock(info, parent, lh, result || !created);
1735
1736 out_parent:
1737         mdt_object_put(info->mti_env, parent);
1738 out:
1739         if (result)
1740                 lustre_msg_set_transno(req->rq_repmsg, 0);
1741         return result;
1742 }
1743
1744 /**
1745  * Create an orphan object use local root.
1746  */
1747 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1748                                           struct mdt_device *mdt,
1749                                           const struct lu_fid *fid,
1750                                           struct md_attr *attr, fmode_t fmode)
1751 {
1752         const struct lu_env *env = info->mti_env;
1753         struct md_op_spec *spec = &info->mti_spec;
1754         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1755         struct mdt_object *obj = NULL;
1756         struct mdt_object *local_root;
1757         static const struct lu_name lname = {
1758                 .ln_name = "i_am_nobody",
1759                 .ln_namelen = sizeof("i_am_nobody") - 1,
1760         };
1761         struct lu_ucred *uc;
1762         kernel_cap_t uc_cap_save;
1763         int rc;
1764
1765         ENTRY;
1766         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1767         if (rc != 0)
1768                 RETURN(ERR_PTR(rc));
1769
1770         local_root = mdt_object_find(env, mdt, local_root_fid);
1771         if (IS_ERR(local_root))
1772                 RETURN(local_root);
1773
1774         obj = mdt_object_new(env, mdt, fid);
1775         if (IS_ERR(obj))
1776                 GOTO(out, rc = PTR_ERR(obj));
1777
1778         spec->sp_cr_lookup = 0;
1779         spec->sp_feat = &dt_directory_features;
1780         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1781         if (attr->ma_valid & MA_LOV) {
1782                 spec->u.sp_ea.eadata = attr->ma_lmm;
1783                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1784                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1785         } else {
1786                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1787         }
1788
1789         uc = lu_ucred(env);
1790         uc_cap_save = uc->uc_cap;
1791         cap_raise(uc->uc_cap, CAP_DAC_OVERRIDE);
1792         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1793                         mdt_object_child(obj), spec, attr);
1794         uc->uc_cap = uc_cap_save;
1795         if (rc < 0) {
1796                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1797                        mdt_obd_name(mdt), PFID(fid), rc);
1798                 GOTO(out, rc);
1799         }
1800
1801         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED, spec);
1802         if (rc < 0)
1803                 CERROR("%s: cannot open volatile file "DFID", orphan "
1804                        "file will be left in PENDING directory until "
1805                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1806                        PFID(fid), rc);
1807         GOTO(out, rc);
1808
1809 out:
1810         if (rc < 0) {
1811                 if (!IS_ERR(obj))
1812                         mdt_object_put(env, obj);
1813                 obj = ERR_PTR(rc);
1814         }
1815         mdt_object_put(env, local_root);
1816         return obj;
1817 }
1818
1819 /* XXX Look into layout in MDT layer. */
1820 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1821 {
1822         struct lov_comp_md_v1 *comp_v1;
1823         struct lov_mds_md *v1;
1824         __u32 off;
1825         int i;
1826
1827         if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1828                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1829
1830                 if (comp_v1->lcm_entry_count == 0)
1831                         return -EINVAL;
1832
1833                 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1834                         off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1835                         v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1836                         v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1837                 }
1838         } else {
1839                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1840         }
1841         return 0;
1842 }
1843
1844 static inline int mdt_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1845 {
1846         struct lov_comp_md_v1 *comp_v1;
1847
1848         if (le32_to_cpu(lmm->lmm_magic == LOV_MAGIC_COMP_V1)) {
1849                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1850                 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1851         } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1852                    le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1853                 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1854         } else {
1855                 return -EINVAL;
1856         }
1857         return 0;
1858 }
1859
1860 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1861                            struct md_attr *ma)
1862 {
1863         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1864         struct lu_ucred *uc = mdt_ucred(info);
1865         struct close_data *data;
1866         struct ldlm_lock *lease;
1867         struct mdt_object *orphan;
1868         struct md_attr *orp_ma;
1869         struct lu_buf *buf;
1870         kernel_cap_t cap;
1871         bool lease_broken;
1872         int rc;
1873         int rc2;
1874
1875         ENTRY;
1876         if (mdt_rdonly(info->mti_exp))
1877                 RETURN(-EROFS);
1878
1879         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1880         if (data == NULL)
1881                 RETURN(-EPROTO);
1882
1883         lease = ldlm_handle2lock(&data->cd_handle);
1884         if (lease == NULL)
1885                 RETURN(-ESTALE);
1886
1887         /* try to hold open_sem so that nobody else can open the file */
1888         if (!down_write_trylock(&o->mot_open_sem)) {
1889                 ldlm_lock_cancel(lease);
1890                 GOTO(out_reprocess, rc = -EBUSY);
1891         }
1892
1893         /* Check if the lease open lease has already canceled */
1894         lock_res_and_lock(lease);
1895         lease_broken = ldlm_is_cancel(lease);
1896         unlock_res_and_lock(lease);
1897
1898         LDLM_DEBUG(lease, DFID " lease broken? %d",
1899                    PFID(mdt_object_fid(o)), lease_broken);
1900
1901         /* Cancel server side lease. Client side counterpart should
1902          * have been cancelled. It's okay to cancel it now as we've
1903          * held mot_open_sem. */
1904         ldlm_lock_cancel(lease);
1905
1906         if (lease_broken) /* don't perform release task */
1907                 GOTO(out_unlock, rc = -ESTALE);
1908
1909         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1910                 GOTO(out_unlock, rc = -EINVAL);
1911
1912         /* ma_need was set before but it seems fine to change it in order to
1913          * avoid modifying the one from RPC */
1914         ma->ma_need = MA_HSM;
1915         rc = mdt_attr_get_complex(info, o, ma);
1916         if (rc != 0)
1917                 GOTO(out_unlock, rc);
1918
1919         if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
1920                 if (ma->ma_valid & MA_HSM) {
1921                         if (ma->ma_hsm.mh_flags & HS_RELEASED)
1922                                 GOTO(out_unlock, rc = -EALREADY);
1923
1924                         if (ma->ma_hsm.mh_arch_id != data->cd_archive_id)
1925                                 CDEBUG(D_CACHE,
1926                                        DFID" archive id diff: %llu:%u\n",
1927                                        PFID(mdt_object_fid(o)),
1928                                        ma->ma_hsm.mh_arch_id,
1929                                        data->cd_archive_id);
1930
1931                         if (!(ma->ma_hsm.mh_flags & HS_DIRTY) &&
1932                             ma->ma_hsm.mh_arch_ver == data->cd_data_version) {
1933                                 CDEBUG(D_CACHE,
1934                                        DFID" data version matches: packed=%llu "
1935                                        "and on-disk=%llu\n",
1936                                        PFID(mdt_object_fid(o)),
1937                                        data->cd_data_version,
1938                                        ma->ma_hsm.mh_arch_ver);
1939                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1940                         }
1941
1942                         if (ma->ma_hsm.mh_flags & HS_DIRTY)
1943                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1944                 } else {
1945                         /* Set up HSM attribte for PCC archived object */
1946                         BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1947                                      sizeof(info->mti_xattr_buf));
1948                         buf = &info->mti_buf;
1949                         buf->lb_buf = info->mti_xattr_buf;
1950                         buf->lb_len = sizeof(struct hsm_attrs);
1951                         memset(&ma->ma_hsm, 0, sizeof(ma->ma_hsm));
1952                         ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1953                         ma->ma_hsm.mh_arch_id = data->cd_archive_id;
1954                         ma->ma_hsm.mh_arch_ver = data->cd_data_version;
1955                         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1956
1957                         rc = mo_xattr_set(info->mti_env, mdt_object_child(o),
1958                                           buf, XATTR_NAME_HSM, 0);
1959                         if (rc)
1960                                 GOTO(out_unlock, rc);
1961                 }
1962         } else {
1963                 if (!mdt_hsm_release_allow(ma))
1964                         GOTO(out_unlock, rc = -EPERM);
1965
1966                 /* already released? */
1967                 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1968                         GOTO(out_unlock, rc = 0);
1969
1970                 /* Compare on-disk and packed data_version */
1971                 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1972                         CDEBUG(D_HSM, DFID" data_version mismatches: "
1973                                "packed=%llu and on-disk=%llu\n",
1974                                PFID(mdt_object_fid(o)),
1975                                data->cd_data_version,
1976                                ma->ma_hsm.mh_arch_ver);
1977                         GOTO(out_unlock, rc = -EPERM);
1978                 }
1979         }
1980
1981         ma->ma_valid = MA_INODE;
1982         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1983         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1984         if (rc < 0)
1985                 GOTO(out_unlock, rc);
1986
1987         mutex_lock(&o->mot_som_mutex);
1988         rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size,
1989                            ma->ma_attr.la_blocks);
1990         mutex_unlock(&o->mot_som_mutex);
1991         if (rc2 < 0)
1992                 CDEBUG(D_INODE,
1993                        "%s: File "DFID" SOM update failed: rc = %d\n",
1994                        mdt_obd_name(info->mti_mdt),
1995                        PFID(mdt_object_fid(o)), rc2);
1996
1997
1998         ma->ma_need = MA_INODE | MA_LOV;
1999         rc = mdt_attr_get_complex(info, o, ma);
2000         if (rc < 0)
2001                 GOTO(out_unlock, rc);
2002
2003         if (!(ma->ma_valid & MA_LOV)) {
2004                 /* Even empty file are released */
2005                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2006                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
2007                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2008                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2009                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2010         } else {
2011                 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
2012                  * ma_lmm as lov_user_md, then it will be confused by union of
2013                  * layout_gen and stripe_offset. */
2014                 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
2015                     LOV_MAGIC_MAGIC)
2016                         ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2017                 else
2018                         GOTO(out_unlock, rc = -EINVAL);
2019         }
2020
2021         /* Set file as released. */
2022         rc = mdt_hsm_set_released(ma->ma_lmm);
2023         if (rc)
2024                 GOTO(out_unlock, rc);
2025
2026         orp_ma = &info->mti_u.hsm.attr;
2027         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2028         /* We use root ownership to bypass potential quota
2029          * restrictions on the user and group of the file. */
2030         orp_ma->ma_attr.la_uid = 0;
2031         orp_ma->ma_attr.la_gid = 0;
2032         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2033         orp_ma->ma_lmm = ma->ma_lmm;
2034         orp_ma->ma_lmm_size = ma->ma_lmm_size;
2035         orp_ma->ma_valid = MA_INODE | MA_LOV;
2036         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2037                                  MDS_FMODE_WRITE);
2038         if (IS_ERR(orphan)) {
2039                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2040                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2041                        PTR_ERR(orphan));
2042                 GOTO(out_unlock, rc = PTR_ERR(orphan));
2043         }
2044
2045         /* Set up HSM attribute for orphan object */
2046         BUILD_BUG_ON(sizeof(struct hsm_attrs) > sizeof(info->mti_xattr_buf));
2047         buf = &info->mti_buf;
2048         buf->lb_buf = info->mti_xattr_buf;
2049         buf->lb_len = sizeof(struct hsm_attrs);
2050         ma->ma_hsm.mh_flags |= HS_RELEASED;
2051         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2052         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2053
2054         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2055                              MDS_INODELOCK_XATTR, LCK_EX, false);
2056         if (rc != 0)
2057                 GOTO(out_close, rc);
2058
2059         /* The orphan has root ownership so we need to raise
2060          * CAP_FOWNER to set the HSM attributes. */
2061         cap = uc->uc_cap;
2062         cap_raise(uc->uc_cap, CAP_FOWNER);
2063         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2064                           XATTR_NAME_HSM, 0);
2065         uc->uc_cap = cap;
2066         if (rc != 0)
2067                 GOTO(out_layout_lock, rc);
2068
2069         /* Swap layout with orphan objects. */
2070         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2071                              mdt_object_child(orphan),
2072                              SWAP_LAYOUTS_MDS_HSM);
2073
2074         if (!rc && ma->ma_attr_flags & MDS_PCC_ATTACH) {
2075                 ma->ma_need = MA_LOV;
2076                 rc = mdt_attr_get_complex(info, o, ma);
2077         }
2078
2079         EXIT;
2080
2081 out_layout_lock:
2082         /* Release exclusive LL */
2083         mdt_object_unlock(info, o, lh, 1);
2084 out_close:
2085         /* Close orphan object anyway */
2086         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2087                        MDS_FMODE_WRITE);
2088         if (rc2 < 0)
2089                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2090                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2091         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2092                         "object closed");
2093         mdt_object_put(info->mti_env, orphan);
2094
2095 out_unlock:
2096         up_write(&o->mot_open_sem);
2097
2098         /* already released */
2099         if (rc == 0) {
2100                 struct mdt_body *repbody;
2101
2102                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2103                 LASSERT(repbody != NULL);
2104                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2105                 if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
2106                         LASSERT(ma->ma_valid & MA_LOV);
2107                         rc = mdt_get_lmm_gen(ma->ma_lmm,
2108                                              &repbody->mbo_layout_gen);
2109                         if (!rc)
2110                                 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
2111                 }
2112         }
2113
2114 out_reprocess:
2115         ldlm_reprocess_all(lease->l_resource,
2116                            lease->l_policy_data.l_inodebits.bits);
2117         LDLM_LOCK_PUT(lease);
2118
2119         ma->ma_valid = 0;
2120         ma->ma_need = 0;
2121
2122         return rc;
2123 }
2124
2125 int mdt_close_handle_layouts(struct mdt_thread_info *info,
2126                              struct mdt_object *o, struct md_attr *ma)
2127 {
2128         struct mdt_lock_handle *lh1 = &info->mti_lh[MDT_LH_NEW];
2129         struct mdt_lock_handle *lh2 = &info->mti_lh[MDT_LH_OLD];
2130         struct close_data *data;
2131         struct ldlm_lock *lease;
2132         struct mdt_object *o1 = o, *o2 = NULL;
2133         bool lease_broken;
2134         bool swap_objects = false;
2135         int rc;
2136
2137         ENTRY;
2138         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2139                 RETURN(-EROFS);
2140
2141         if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
2142                 RETURN(-EINVAL);
2143
2144         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2145         if (data == NULL)
2146                 RETURN(-EPROTO);
2147
2148         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2149                 RETURN(-EINVAL);
2150
2151         rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
2152         if (rc == 0) {
2153                 /**
2154                  * only MDS_CLOSE_LAYOUT_SPLIT use the same fid to indicate
2155                  * mirror deletion, so we'd zero cd_fid, and keeps o2 be NULL.
2156                  */
2157                 if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT))
2158                         RETURN(-EINVAL);
2159
2160                 /* zero cd_fid to keeps o2 be NULL */
2161                 fid_zero(&data->cd_fid);
2162         } else if (rc < 0) {
2163                 /* Exchange o1 and o2, to enforce locking order */
2164                 swap_objects = true;
2165         }
2166
2167         lease = ldlm_handle2lock(&data->cd_handle);
2168         if (lease == NULL)
2169                 RETURN(-ESTALE);
2170
2171         if (!fid_is_zero(&data->cd_fid)) {
2172                 o2 = mdt_object_find(info->mti_env, info->mti_mdt,
2173                                      &data->cd_fid);
2174                 if (IS_ERR(o2))
2175                         GOTO(out_lease, rc = PTR_ERR(o2));
2176
2177                 if (!mdt_object_exists(o2))
2178                         GOTO(out_obj, rc = -ENOENT);
2179
2180                 if (!S_ISREG(lu_object_attr(&o2->mot_obj)))
2181                         GOTO(out_obj, rc = -EINVAL);
2182
2183                 if (swap_objects)
2184                         swap(o1, o2);
2185         }
2186
2187         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
2188                            MAY_WRITE);
2189         if (rc < 0)
2190                 GOTO(out_obj, rc);
2191
2192         if (o2) {
2193                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2),
2194                                    NULL, MAY_WRITE);
2195                 if (rc < 0)
2196                         GOTO(out_obj, rc);
2197         }
2198
2199         /* try to hold open_sem so that nobody else can open the file */
2200         if (!down_write_trylock(&o->mot_open_sem)) {
2201                 ldlm_lock_cancel(lease);
2202                 GOTO(out_obj, rc = -EBUSY);
2203         }
2204
2205         /* Check if the lease open lease has already canceled */
2206         lock_res_and_lock(lease);
2207         lease_broken = ldlm_is_cancel(lease);
2208         unlock_res_and_lock(lease);
2209
2210         LDLM_DEBUG(lease, DFID " lease broken? %d",
2211                    PFID(mdt_object_fid(o)), lease_broken);
2212
2213         /* Cancel server side lease. Client side counterpart should
2214          * have been cancelled. It's okay to cancel it now as we've
2215          * held mot_open_sem. */
2216         ldlm_lock_cancel(lease);
2217
2218         if (lease_broken)
2219                 GOTO(out_unlock_sem, rc = -ESTALE);
2220
2221         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
2222                              MDS_INODELOCK_XATTR, LCK_EX, false);
2223         if (rc < 0)
2224                 GOTO(out_unlock_sem, rc);
2225
2226         if (o2) {
2227                 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2228                                      MDS_INODELOCK_XATTR, LCK_EX, false);
2229                 if (rc < 0)
2230                         GOTO(out_unlock1, rc);
2231         }
2232
2233         /* Swap layout with orphan object */
2234         if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2235                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2236                                      mdt_object_child(o2), 0);
2237         } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2238                    ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2239                 struct lu_buf *buf = &info->mti_buf;
2240                 struct md_rejig_data mrd;
2241
2242                 if (o2) {
2243                         mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2244                 } else {
2245                         if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)) {
2246                                 /* paranoid check again */
2247                                 CERROR(DFID
2248                                   ":only mirror split support NULL o2 object\n",
2249                                         PFID(mdt_object_fid(o)));
2250                                 GOTO(out_unlock1, rc = -EINVAL);
2251                         }
2252
2253                         /* set NULL mrd_obj for deleting mirror objects */
2254                         mrd.mrd_obj = NULL;
2255                 }
2256
2257                 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2258                         mrd.mrd_mirror_id = data->cd_mirror_id;
2259                         /* set a small enough blocks in the SoM */
2260                         ma->ma_attr.la_blocks >>= 1;
2261                 }
2262
2263                 buf->lb_len = sizeof(mrd);
2264                 buf->lb_buf = &mrd;
2265                 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2266                                   XATTR_LUSTRE_LOV,
2267                                   ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2268                                   LU_XATTR_SPLIT : LU_XATTR_MERGE);
2269                 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS |
2270                                                        LA_LSIZE | LA_LBLOCKS)) {
2271                         int rc2;
2272                         enum lustre_som_flags lsf;
2273
2274                         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS))
2275                                 lsf = SOM_FL_STRICT;
2276                         else
2277                                 lsf = SOM_FL_LAZY;
2278
2279                         mutex_lock(&o->mot_som_mutex);
2280                         rc2 = mdt_set_som(info, o, lsf,
2281                                           ma->ma_attr.la_size,
2282                                           ma->ma_attr.la_blocks);
2283                         mutex_unlock(&o->mot_som_mutex);
2284                         if (rc2 < 0)
2285                                 CERROR(DFID": Setting i_blocks error: %d, "
2286                                        "i_blocks will be reported wrongly and "
2287                                        "can only be fixed in next resync\n",
2288                                        PFID(mdt_object_fid(o)), rc2);
2289                 }
2290         }
2291         if (rc < 0)
2292                 GOTO(out_unlock2, rc);
2293
2294         EXIT;
2295
2296 out_unlock2:
2297         /* Release exclusive LL */
2298         if (o2)
2299                 mdt_object_unlock(info, o2, lh2, 1);
2300
2301 out_unlock1:
2302         mdt_object_unlock(info, o1, lh1, 1);
2303
2304 out_unlock_sem:
2305         up_write(&o->mot_open_sem);
2306
2307         /* already swapped */
2308         if (rc == 0) {
2309                 struct mdt_body *repbody;
2310
2311                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2312                 LASSERT(repbody != NULL);
2313                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2314         }
2315
2316 out_obj:
2317         if (o1 != o)
2318                 /* the 2nd object has been used, and swapped to o1 */
2319                 mdt_object_put(info->mti_env, o1);
2320         else if (o2)
2321                 /* the 2nd object has been used, and not swapped */
2322                 mdt_object_put(info->mti_env, o2);
2323
2324         ldlm_reprocess_all(lease->l_resource,
2325                            lease->l_policy_data.l_inodebits.bits);
2326
2327 out_lease:
2328         LDLM_LOCK_PUT(lease);
2329
2330         if (ma != NULL) {
2331                 ma->ma_valid = 0;
2332                 ma->ma_need = 0;
2333         }
2334
2335         return rc;
2336 }
2337
2338 static int mdt_close_resync_done(struct mdt_thread_info *info,
2339                                  struct mdt_object *o, struct md_attr *ma)
2340 {
2341         struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LOCAL];
2342         struct close_data *data;
2343         struct ldlm_lock *lease;
2344         struct md_layout_change layout = { 0 };
2345         __u32 *resync_ids = NULL;
2346         size_t resync_count = 0;
2347         bool lease_broken;
2348         int rc;
2349
2350         ENTRY;
2351         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2352                 RETURN(-EROFS);
2353
2354         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2355                 RETURN(-EINVAL);
2356
2357         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2358         if (data == NULL)
2359                 RETURN(-EPROTO);
2360
2361         if (req_capsule_req_need_swab(info->mti_pill))
2362                 lustre_swab_close_data_resync_done(&data->cd_resync);
2363
2364         if (!fid_is_zero(&data->cd_fid))
2365                 RETURN(-EPROTO);
2366
2367         lease = ldlm_handle2lock(&data->cd_handle);
2368         if (lease == NULL)
2369                 RETURN(-ESTALE);
2370
2371         /* try to hold open_sem so that nobody else can open the file */
2372         if (!down_write_trylock(&o->mot_open_sem)) {
2373                 ldlm_lock_cancel(lease);
2374                 GOTO(out_reprocess, rc = -EBUSY);
2375         }
2376
2377         /* Check if the lease open lease has already canceled */
2378         lock_res_and_lock(lease);
2379         lease_broken = ldlm_is_cancel(lease);
2380         unlock_res_and_lock(lease);
2381
2382         LDLM_DEBUG(lease, DFID " lease broken? %d",
2383                    PFID(mdt_object_fid(o)), lease_broken);
2384
2385         /* Cancel server side lease. Client side counterpart should
2386          * have been cancelled. It's okay to cancel it now as we've
2387          * held mot_open_sem. */
2388         ldlm_lock_cancel(lease);
2389
2390         if (lease_broken) /* don't perform release task */
2391                 GOTO(out_unlock, rc = -ESTALE);
2392
2393         resync_count = data->cd_resync.resync_count;
2394
2395         if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2396                 void *data;
2397
2398                 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2399                                            RCL_CLIENT))
2400                         GOTO(out_unlock, rc = -EPROTO);
2401
2402                 OBD_ALLOC_PTR_ARRAY(resync_ids, resync_count);
2403                 if (!resync_ids)
2404                         GOTO(out_unlock, rc = -ENOMEM);
2405
2406                 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2407                 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2408
2409                 layout.mlc_resync_ids = resync_ids;
2410         } else {
2411                 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2412         }
2413
2414         layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2415         layout.mlc_resync_count = resync_count;
2416         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2417                 layout.mlc_som.lsa_valid = SOM_FL_STRICT;
2418                 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2419                 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2420         }
2421         rc = mdt_layout_change(info, o, lhc, &layout);
2422         if (rc)
2423                 GOTO(out_unlock, rc);
2424
2425         mdt_object_unlock(info, o, lhc, 0);
2426
2427         EXIT;
2428
2429 out_unlock:
2430         up_write(&o->mot_open_sem);
2431
2432         /* already released */
2433         if (rc == 0) {
2434                 struct mdt_body *repbody;
2435
2436                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2437                 LASSERT(repbody != NULL);
2438                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2439         }
2440
2441         if (resync_ids)
2442                 OBD_FREE_PTR_ARRAY(resync_ids, resync_count);
2443
2444 out_reprocess:
2445         ldlm_reprocess_all(lease->l_resource,
2446                            lease->l_policy_data.l_inodebits.bits);
2447         LDLM_LOCK_PUT(lease);
2448
2449         ma->ma_valid = 0;
2450         ma->ma_need = 0;
2451
2452         return rc;
2453 }
2454
2455 #define MFD_CLOSED(open_flags) ((open_flags) == MDS_FMODE_CLOSED)
2456
2457 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2458 {
2459         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_open_flags));
2460 }
2461
2462 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2463 {
2464         struct mdt_object *o = mfd->mfd_object;
2465         struct md_object *next = mdt_object_child(o);
2466         struct md_attr *ma = &info->mti_attr;
2467         struct lu_fid *ofid = &info->mti_tmp_fid1;
2468         int rc = 0;
2469         int rc2;
2470         u64 open_flags;
2471         u64 intent;
2472
2473         ENTRY;
2474         open_flags = mfd->mfd_open_flags;
2475         intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2476         *ofid = *mdt_object_fid(o);
2477
2478         /* the below message is checked in replay-single.sh test_46 */
2479         CDEBUG(D_INODE, "%s: %sclosing file handle "DFID" with intent: %llx\n",
2480                mdt_obd_name(info->mti_mdt),
2481                ma->ma_valid & MA_FORCE_LOG ? "force " : "", PFID(ofid), intent);
2482
2483         switch (intent) {
2484         case MDS_HSM_RELEASE: {
2485                 rc = mdt_hsm_release(info, o, ma);
2486                 if (rc < 0) {
2487                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2488                                mdt_obd_name(info->mti_mdt),
2489                                PFID(ofid), rc);
2490                         /* continue to close even error occurred. */
2491                 }
2492                 break;
2493         }
2494         case MDS_CLOSE_LAYOUT_MERGE:
2495         case MDS_CLOSE_LAYOUT_SPLIT:
2496         case MDS_CLOSE_LAYOUT_SWAP: {
2497                 rc = mdt_close_handle_layouts(info, o, ma);
2498                 if (rc < 0) {
2499                         CDEBUG(D_INODE,
2500                                "%s: cannot %s layout of "DFID": rc = %d\n",
2501                                mdt_obd_name(info->mti_mdt),
2502                                intent == MDS_CLOSE_LAYOUT_MERGE ? "merge" :
2503                                intent == MDS_CLOSE_LAYOUT_SPLIT ? "split" :
2504                                "swap",
2505                                PFID(ofid), rc);
2506                         /* continue to close even if error occurred. */
2507                 }
2508                 break;
2509         }
2510         case MDS_CLOSE_RESYNC_DONE:
2511                 rc = mdt_close_resync_done(info, o, ma);
2512                 if (rc < 0) {
2513                         CDEBUG(D_INODE,
2514                                "%s: cannot resync layout of "DFID": rc = %d\n",
2515                                mdt_obd_name(info->mti_mdt),
2516                                PFID(ofid), rc);
2517                         /* continue to close even if error occurred. */
2518                 }
2519                 break;
2520         default:
2521                 /* nothing */
2522                 break;
2523         }
2524
2525         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
2526             ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) {
2527                 rc2 = mdt_lsom_update(info, o, false);
2528                 if (rc2 < 0) {
2529                         CDEBUG(D_INODE,
2530                                "%s: File " DFID " LSOM failed: rc = %d\n",
2531                                mdt_obd_name(info->mti_mdt),
2532                                PFID(ofid), rc2);
2533                         if (rc == 0)
2534                                 rc = rc2;
2535                         /* continue to close even if error occurred. */
2536                 }
2537         }
2538
2539         if (open_flags & MDS_FMODE_WRITE)
2540                 mdt_write_put(o);
2541         else if (open_flags & MDS_FMODE_EXEC)
2542                 mdt_write_allow(o);
2543
2544         /* Update atime|mtime|ctime on close. */
2545         if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ ||
2546              open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
2547             (ma->ma_attr.la_valid & LA_ATIME ||
2548              ma->ma_attr.la_valid & LA_MTIME ||
2549              ma->ma_attr.la_valid & LA_CTIME)) {
2550                 ma->ma_valid = MA_INODE;
2551                 ma->ma_attr_flags |= MDS_CLOSE_UPDATE_TIMES;
2552                 ma->ma_attr.la_valid &= (LA_ATIME | LA_MTIME | LA_CTIME);
2553
2554                 if (ma->ma_attr.la_valid & LA_MTIME) {
2555                         if (mdt_attr_get_pfid(info, o, &ma->ma_pfid) == 0)
2556                                 ma->ma_valid |= MA_PFID;
2557                 }
2558
2559                 rc2 = mo_attr_set(info->mti_env, next, ma);
2560                 if (rc2 != 0) {
2561                         CDEBUG(D_INODE,
2562                            "%s: File "DFID" set attr (%#llx) failed: rc = %d\n",
2563                                mdt_obd_name(info->mti_mdt), PFID(ofid),
2564                                ma->ma_attr.la_valid, rc2);
2565                         if (rc == 0)
2566                                 rc = rc2;
2567                 }
2568         }
2569
2570         /* If file data is modified, add the dirty flag. */
2571         if (ma->ma_attr_flags & MDS_DATA_MODIFIED) {
2572                 rc2 = mdt_add_dirty_flag(info, o, ma);
2573                 if (rc2 != 0) {
2574                         CDEBUG(D_INODE,
2575                              "%s: File "DFID" add dirty flag failed: rc = %d\n",
2576                                mdt_obd_name(info->mti_mdt), PFID(ofid), rc2);
2577                         if (rc == 0)
2578                                 rc = rc2;
2579                 }
2580         }
2581
2582         ma->ma_need |= MA_INODE;
2583         ma->ma_valid &= ~MA_INODE;
2584
2585         LASSERT(atomic_read(&o->mot_open_count) > 0);
2586         atomic_dec(&o->mot_open_count);
2587         mdt_handle_last_unlink(info, o, ma);
2588
2589         if (!MFD_CLOSED(open_flags)) {
2590                 rc2 = mo_close(info->mti_env, next, ma, open_flags);
2591                 if (rc2 != 0) {
2592                         CDEBUG(D_INODE,
2593                                "%s: File "DFID" close failed: rc = %d\n",
2594                                mdt_obd_name(info->mti_mdt), PFID(ofid), rc2);
2595                         if (rc == 0)
2596                                 rc = rc2;
2597                 }
2598                 if (mdt_dom_check_for_discard(info, o))
2599                         mdt_dom_discard_data(info, o);
2600         }
2601
2602         /* adjust open and lease count */
2603         if (open_flags & MDS_OPEN_LEASE) {
2604                 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2605                 atomic_dec(&o->mot_lease_count);
2606         }
2607
2608         mdt_mfd_free(mfd);
2609         mdt_object_put(info->mti_env, o);
2610
2611         RETURN(rc);
2612 }
2613
2614 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2615                        struct mdt_body *repbody)
2616 {
2617         struct mdt_export_data *med;
2618         struct mdt_file_data   *mfd;
2619         int ret = 0;
2620         int rc = 0;
2621
2622         ENTRY;
2623         med = &req->rq_export->exp_mdt_data;
2624         spin_lock(&med->med_open_lock);
2625         mfd = mdt_open_handle2mfd(med, &info->mti_open_handle,
2626                                   req_is_replay(req));
2627         if (mdt_mfd_closed(mfd)) {
2628                 spin_unlock(&med->med_open_lock);
2629                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2630                        ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2631                        info->mti_open_handle.cookie);
2632                 /** not serious error since bug 3633 */
2633                 rc = -ESTALE;
2634         } else {
2635                 class_handle_unhash(&mfd->mfd_open_handle);
2636                 list_del_init(&mfd->mfd_list);
2637                 spin_unlock(&med->med_open_lock);
2638                 ret = mdt_mfd_close(info, mfd);
2639         }
2640
2641         RETURN(rc ? rc : ret);
2642 }
2643
2644 int mdt_close(struct tgt_session_info *tsi)
2645 {
2646         struct mdt_thread_info *info = tsi2mdt_info(tsi);
2647         struct ptlrpc_request *req = tgt_ses_req(tsi);
2648         struct md_attr *ma = &info->mti_attr;
2649         struct mdt_body *repbody = NULL;
2650         ktime_t kstart = ktime_get();
2651         int rc;
2652         int rc2;
2653
2654         ENTRY;
2655         /* Close may come with the Size-on-MDS update. Unpack it. */
2656         rc = mdt_close_unpack(info);
2657         if (rc)
2658                 GOTO(out, rc = err_serious(rc));
2659
2660         /* These fields are no longer used and are left for compatibility.
2661          * size is always zero */
2662         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2663                              0);
2664         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2665                              0);
2666         rc = req_capsule_server_pack(info->mti_pill);
2667         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2668                 mdt_client_compatibility(info);
2669                 if (rc == 0)
2670                         mdt_fix_reply(info);
2671                 mdt_exit_ucred(info);
2672                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2673         }
2674
2675         /* Continue to close handle even if we can not pack reply */
2676         if (rc == 0) {
2677                 repbody = req_capsule_server_get(info->mti_pill,
2678                                                  &RMF_MDT_BODY);
2679                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2680                                                     &RMF_MDT_MD);
2681                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2682                                                        &RMF_MDT_MD,
2683                                                        RCL_SERVER);
2684                 ma->ma_need = MA_INODE | MA_LOV;
2685                 repbody->mbo_eadatasize = 0;
2686                 repbody->mbo_aclsize = 0;
2687         } else {
2688                 rc = err_serious(rc);
2689         }
2690
2691         rc2 = mdt_close_internal(info, req, repbody);
2692         if (rc2 != -ESTALE)
2693                 mdt_empty_transno(info, rc2);
2694         if (rc2 != 0 && rc == 0)
2695                 rc = rc2;
2696
2697         if (repbody != NULL) {
2698                 mdt_client_compatibility(info);
2699                 rc2 = mdt_fix_reply(info);
2700                 if (rc2 != 0 && rc == 0)
2701                         rc = rc2;
2702         }
2703
2704         mdt_exit_ucred(info);
2705         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2706                 GOTO(out, rc = err_serious(-ENOMEM));
2707
2708         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2709                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2710                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2711 out:
2712         mdt_thread_info_fini(info);
2713         if (rc == 0)
2714                 mdt_counter_incr(req, LPROC_MDT_CLOSE,
2715                                  ktime_us_delta(ktime_get(), kstart));
2716         RETURN(rc);
2717 }