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 /* Create a new mdt_file_data struct, initialize it,
  49  * and insert it to global hash table */
  50 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
  51 {
  52         struct mdt_file_data *mfd;
  53         ENTRY;
  54
  55         OBD_ALLOC_PTR(mfd);
  56         if (mfd != NULL) {
  57                 refcount_set(&mfd->mfd_open_handle.h_ref, 1);
  58                 INIT_HLIST_NODE(&mfd->mfd_open_handle.h_link);
  59                 mfd->mfd_owner = med;
  60                 INIT_LIST_HEAD(&mfd->mfd_list);
  61                 class_handle_hash(&mfd->mfd_open_handle, mfd_open_handle_owner);
  62         }
  63
  64         RETURN(mfd);
  65 }
  66
  67 /*
  68  * Find the mfd pointed to by handle in global hash table.
  69  * In case of replay the handle is obsoleted
  70  * but mfd can be found in mfd list by that handle.
  71  * Callers need to be holding med_open_lock.
  72  */
  73 struct mdt_file_data *mdt_open_handle2mfd(struct mdt_export_data *med,
  74                                         const struct lustre_handle *open_handle,
  75                                         bool is_replay_or_resent)
  76 {
  77         struct mdt_file_data   *mfd;
  78         ENTRY;
  79
  80         LASSERT(open_handle != NULL);
  81         mfd = class_handle2object(open_handle->cookie, mfd_open_handle_owner);
  82         if (mfd)
  83                 refcount_dec(&mfd->mfd_open_handle.h_ref);
  84
  85         /* during dw/setattr replay the mfd can be found by old handle */
  86         if ((!mfd || mfd->mfd_owner != med) && is_replay_or_resent) {
  87                 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  88                         if (mfd->mfd_open_handle_old.cookie ==
  89                             open_handle->cookie)
  90                                 RETURN(mfd);
  91                 }
  92                 mfd = NULL;
  93         }
  94
  95         RETURN(mfd);
  96 }
  97
  98 /* free mfd */
  99 void mdt_mfd_free(struct mdt_file_data *mfd)
 100 {
 101         LASSERT(refcount_read(&mfd->mfd_open_handle.h_ref) == 1);
 102         LASSERT(list_empty(&mfd->mfd_list));
 103         OBD_FREE_PRE(mfd, sizeof(*mfd), "rcu");
 104         kfree_rcu(mfd, mfd_open_handle.h_rcu);
 105 }
 106
 107 static int mdt_create_data(struct mdt_thread_info *info,
 108                            struct mdt_object *p, struct mdt_object *o)
 109 {
 110         struct md_op_spec     *spec = &info->mti_spec;
 111         struct md_attr        *ma   = &info->mti_attr;
 112         int                    rc   = 0;
 113         ENTRY;
 114
 115         if (!md_should_create(spec->sp_cr_flags))
 116                 RETURN(0);
 117
 118         ma->ma_need = MA_INODE | MA_LOV;
 119         ma->ma_valid = 0;
 120         mutex_lock(&o->mot_lov_mutex);
 121         if (!o->mot_lov_created) {
 122                 rc = mdo_create_data(info->mti_env,
 123                                      p ? mdt_object_child(p) : NULL,
 124                                      mdt_object_child(o), spec, ma);
 125                 if (rc == 0)
 126                         rc = mdt_attr_get_complex(info, o, ma);
 127
 128                 if (rc == 0 && ma->ma_valid & MA_LOV)
 129                         o->mot_lov_created = 1;
 130         }
 131
 132         mutex_unlock(&o->mot_lov_mutex);
 133         RETURN(rc);
 134 }
 135
 136 int mdt_write_read(struct mdt_object *o)
 137 {
 138         int rc = 0;
 139         ENTRY;
 140         spin_lock(&o->mot_write_lock);
 141         rc = o->mot_write_count;
 142         spin_unlock(&o->mot_write_lock);
 143         RETURN(rc);
 144 }
 145
 146 int mdt_write_get(struct mdt_object *o)
 147 {
 148         int rc = 0;
 149         ENTRY;
 150         spin_lock(&o->mot_write_lock);
 151         if (o->mot_write_count < 0)
 152                 rc = -ETXTBSY;
 153         else
 154                 o->mot_write_count++;
 155         spin_unlock(&o->mot_write_lock);
 156
 157         RETURN(rc);
 158 }
 159
 160 void mdt_write_put(struct mdt_object *o)
 161 {
 162         ENTRY;
 163         spin_lock(&o->mot_write_lock);
 164         o->mot_write_count--;
 165         spin_unlock(&o->mot_write_lock);
 166         EXIT;
 167 }
 168
 169 static int mdt_write_deny(struct mdt_object *o)
 170 {
 171         int rc = 0;
 172         ENTRY;
 173         spin_lock(&o->mot_write_lock);
 174         if (o->mot_write_count > 0)
 175                 rc = -ETXTBSY;
 176         else
 177                 o->mot_write_count--;
 178         spin_unlock(&o->mot_write_lock);
 179         RETURN(rc);
 180 }
 181
 182 static void mdt_write_allow(struct mdt_object *o)
 183 {
 184         ENTRY;
 185         spin_lock(&o->mot_write_lock);
 186         o->mot_write_count++;
 187         spin_unlock(&o->mot_write_lock);
 188         EXIT;
 189 }
 190
 191 /* there can be no real transaction so prepare the fake one */
 192 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 193 {
 194         struct mdt_device      *mdt = info->mti_mdt;
 195         struct ptlrpc_request  *req = mdt_info_req(info);
 196         struct tg_export_data  *ted;
 197         struct lsd_client_data *lcd;
 198         ENTRY;
 199
 200         if (mdt_rdonly(req->rq_export))
 201                 RETURN_EXIT;
 202
 203         /* transaction has occurred already */
 204         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 205                 RETURN_EXIT;
 206
 207         if (tgt_is_multimodrpcs_client(req->rq_export)) {
 208                 struct thandle         *th;
 209
 210                 /* generate an empty transaction to get a transno
 211                  * and reply data */
 212                 th = dt_trans_create(info->mti_env, mdt->mdt_bottom);
 213                 if (!IS_ERR(th)) {
 214                         rc = dt_trans_start(info->mti_env, mdt->mdt_bottom, th);
 215                         dt_trans_stop(info->mti_env, mdt->mdt_bottom, th);
 216                 }
 217                 RETURN_EXIT;
 218         }
 219
 220         spin_lock(&mdt->mdt_lut.lut_translock);
 221         if (rc != 0) {
 222                 if (info->mti_transno != 0) {
 223                         struct obd_export *exp = req->rq_export;
 224
 225                         CERROR("%s: replay trans %llu NID %s: rc = %d\n",
 226                                mdt_obd_name(mdt), info->mti_transno,
 227                                obd_export_nid2str(exp), rc);
 228                         spin_unlock(&mdt->mdt_lut.lut_translock);
 229                         RETURN_EXIT;
 230                 }
 231         } else if (info->mti_transno == 0) {
 232                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
 233         } else {
 234                 /* should be replay */
 235                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 236                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 237         }
 238         spin_unlock(&mdt->mdt_lut.lut_translock);
 239
 240         CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
 241                info->mti_transno,
 242                req->rq_export->exp_obd->obd_last_committed);
 243
 244         req->rq_transno = info->mti_transno;
 245         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 246
 247         /* update lcd in memory only for resent cases */
 248         ted = &req->rq_export->exp_target_data;
 249         LASSERT(ted);
 250         mutex_lock(&ted->ted_lcd_lock);
 251         lcd = ted->ted_lcd;
 252         if (info->mti_transno < lcd->lcd_last_transno &&
 253             info->mti_transno != 0) {
 254                 /* This should happen during replay. Do not update
 255                  * last rcvd info if replay req transno < last transno,
 256                  * otherwise the following resend(after replay) can not
 257                  * be checked correctly by xid */
 258                 mutex_unlock(&ted->ted_lcd_lock);
 259                 CDEBUG(D_HA, "%s: transno = %llu < last_transno = %llu\n",
 260                        mdt_obd_name(mdt), info->mti_transno,
 261                        lcd->lcd_last_transno);
 262                 RETURN_EXIT;
 263         }
 264
 265         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
 266                 if (info->mti_transno != 0)
 267                         lcd->lcd_last_close_transno = info->mti_transno;
 268                 lcd->lcd_last_close_xid = req->rq_xid;
 269                 lcd->lcd_last_close_result = rc;
 270         } else {
 271                 /* VBR: save versions in last_rcvd for reconstruct. */
 272                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 273                 if (pre_versions) {
 274                         lcd->lcd_pre_versions[0] = pre_versions[0];
 275                         lcd->lcd_pre_versions[1] = pre_versions[1];
 276                         lcd->lcd_pre_versions[2] = pre_versions[2];
 277                         lcd->lcd_pre_versions[3] = pre_versions[3];
 278                 }
 279                 if (info->mti_transno != 0)
 280                         lcd->lcd_last_transno = info->mti_transno;
 281
 282                 lcd->lcd_last_xid = req->rq_xid;
 283                 lcd->lcd_last_result = rc;
 284                 lcd->lcd_last_data = info->mti_opdata;
 285         }
 286         mutex_unlock(&ted->ted_lcd_lock);
 287
 288         EXIT;
 289 }
 290
 291 void mdt_mfd_set_mode(struct mdt_file_data *mfd, u64 open_flags)
 292 {
 293         LASSERT(mfd != NULL);
 294
 295         CDEBUG(D_DENTRY, DFID " Change mfd open_flags %#llo -> %#llo.\n",
 296                PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_open_flags,
 297                open_flags);
 298
 299         mfd->mfd_open_flags = open_flags;
 300 }
 301
 302 /**
 303  * prep ma_lmm/ma_lmv for md_attr from reply
 304  */
 305 static void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
 306                                      struct mdt_object *obj,
 307                                      struct md_attr *ma)
 308 {
 309         if (ma->ma_lmv || ma->ma_lmm) {
 310                 CDEBUG(D_INFO, DFID " %s already set.\n",
 311                        PFID(mdt_object_fid(obj)),
 312                        ma->ma_lmv ? (ma->ma_lmm ? "ma_lmv and ma_lmm"
 313                                                 : "ma_lmv")
 314                                   : "ma_lmm");
 315                 return;
 316         }
 317
 318         if (S_ISDIR(obj->mot_header.loh_attr)) {
 319                 ma->ma_lmv = req_capsule_server_get(info->mti_pill,
 320                                                     &RMF_MDT_MD);
 321                 ma->ma_lmv_size = req_capsule_get_size(info->mti_pill,
 322                                                        &RMF_MDT_MD,
 323                                                        RCL_SERVER);
 324                 if (ma->ma_lmv_size > 0)
 325                         ma->ma_need |= MA_LMV;
 326         } else {
 327                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
 328                                                     &RMF_MDT_MD);
 329                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 330                                                        &RMF_MDT_MD,
 331                                                        RCL_SERVER);
 332                 if (ma->ma_lmm_size > 0)
 333                         ma->ma_need |= MA_LOV;
 334         }
 335 }
 336
 337 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 338                         struct mdt_object *o, u64 open_flags, int created,
 339                         struct ldlm_reply *rep)
 340 {
 341         struct ptlrpc_request *req = mdt_info_req(info);
 342         struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 343         struct mdt_file_data *mfd;
 344         struct md_attr *ma  = &info->mti_attr;
 345         struct lu_attr *la  = &ma->ma_attr;
 346         struct mdt_body *repbody;
 347         bool isdir, isreg;
 348         int rc = 0;
 349
 350         ENTRY;
 351         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 352
 353         isreg = S_ISREG(la->la_mode);
 354         isdir = S_ISDIR(la->la_mode);
 355         if (isreg && !(ma->ma_valid & MA_LOV) &&
 356             !(open_flags & MDS_OPEN_RELEASE)) {
 357                 /*
 358                  * No EA, check whether it is will set regEA and dirEA since in
 359                  * above attr get, these size might be zero, so reset it, to
 360                  * retrieve the MD after create obj.
 361                  */
 362                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 363                                                        &RMF_MDT_MD,
 364                                                        RCL_SERVER);
 365                 /* in replay case, p == NULL */
 366                 rc = mdt_create_data(info, p, o);
 367                 if (rc)
 368                         RETURN(rc);
 369
 370                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
 371                         mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
 372         }
 373
 374         CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n",
 375                ma->ma_valid, ma->ma_lmm_size);
 376
 377         if (ma->ma_valid & MA_LOV) {
 378                 LASSERT(ma->ma_lmm_size != 0);
 379                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 380                 if (isdir)
 381                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 382                 else
 383                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 384         }
 385
 386         if (ma->ma_valid & MA_LMV) {
 387                 LASSERT(ma->ma_lmv_size != 0);
 388                 repbody->mbo_eadatasize = ma->ma_lmv_size;
 389                 LASSERT(isdir);
 390                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
 391         }
 392
 393         if (open_flags & MDS_FMODE_WRITE)
 394                 rc = mdt_write_get(o);
 395         else if (open_flags & MDS_FMODE_EXEC)
 396                 rc = mdt_write_deny(o);
 397
 398         if (rc)
 399                 RETURN(rc);
 400
 401         rc = mo_open(info->mti_env, mdt_object_child(o),
 402                      created ? open_flags | MDS_OPEN_CREATED : open_flags,
 403                      &info->mti_spec);
 404         if (rc != 0) {
 405                 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
 406                  * client does not know which suppgid should be sent to the MDS,
 407                  * or some other(s) changed the target file's GID after this RPC
 408                  * sent to the MDS with the suppgid as the original GID, then we
 409                  * give the client another chance to send the right suppgid. */
 410                 if (rc == -EACCES &&
 411                     allow_client_chgrp(info, lu_ucred(info->mti_env)))
 412                         mdt_set_disposition(info, rep, DISP_OPEN_DENY);
 413
 414                 GOTO(err_out, rc);
 415         }
 416
 417         mfd = mdt_mfd_new(med);
 418         if (mfd == NULL)
 419                 GOTO(err_out, rc = -ENOMEM);
 420
 421         /*
 422          * Keep a reference on this object for this open, and is
 423          * released by mdt_mfd_close().
 424          */
 425         mdt_object_get(info->mti_env, o);
 426         mfd->mfd_object = o;
 427         mfd->mfd_xid = req->rq_xid;
 428
 429         /*
 430          * @open_flags is always not zero. At least it should be FMODE_READ,
 431          * FMODE_WRITE or MDS_FMODE_EXEC.
 432          */
 433         LASSERT(open_flags != 0);
 434
 435         /* Open handling. */
 436         mdt_mfd_set_mode(mfd, open_flags);
 437
 438         atomic_inc(&o->mot_open_count);
 439         if (open_flags & MDS_OPEN_LEASE)
 440                 atomic_inc(&o->mot_lease_count);
 441
 442         /* replay handle */
 443         if (req_is_replay(req)) {
 444                 struct mdt_file_data *old_mfd;
 445                 /* Check wheather old cookie already exist in
 446                  * the list, becasue when do recovery, client
 447                  * might be disconnected from server, and
 448                  * restart replay, so there maybe some orphan
 449                  * mfd here, we should remove them */
 450                 LASSERT(info->mti_rr.rr_open_handle != NULL);
 451                 spin_lock(&med->med_open_lock);
 452                 old_mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 453                                               true);
 454                 if (old_mfd != NULL) {
 455                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 456                                "cookie = %#llx\n", mfd,
 457                                PFID(mdt_object_fid(mfd->mfd_object)),
 458                                info->mti_rr.rr_open_handle->cookie);
 459                         class_handle_unhash(&old_mfd->mfd_open_handle);
 460                         list_del_init(&old_mfd->mfd_list);
 461                         spin_unlock(&med->med_open_lock);
 462                         /* no attr update for that close */
 463                         la->la_valid = 0;
 464                         ma->ma_valid |= MA_FLAGS;
 465                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 466                         mdt_mfd_close(info, old_mfd);
 467                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 468                         ma->ma_valid &= ~MA_FLAGS;
 469                 } else {
 470                         spin_unlock(&med->med_open_lock);
 471                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 472                                "cookie = %#llx\n",
 473                                PFID(mdt_object_fid(mfd->mfd_object)),
 474                                info->mti_rr.rr_open_handle->cookie);
 475                 }
 476
 477                 CDEBUG(D_HA, "Store old cookie %#llx in new mfd\n",
 478                        info->mti_rr.rr_open_handle->cookie);
 479
 480                 mfd->mfd_open_handle_old = *info->mti_rr.rr_open_handle;
 481         }
 482
 483         repbody->mbo_open_handle.cookie = mfd->mfd_open_handle.h_cookie;
 484
 485         if (req->rq_export->exp_disconnected) {
 486                 spin_lock(&med->med_open_lock);
 487                 class_handle_unhash(&mfd->mfd_open_handle);
 488                 list_del_init(&mfd->mfd_list);
 489                 spin_unlock(&med->med_open_lock);
 490                 mdt_mfd_close(info, mfd);
 491         } else {
 492                 spin_lock(&med->med_open_lock);
 493                 list_add_tail(&mfd->mfd_list, &med->med_open_head);
 494                 spin_unlock(&med->med_open_lock);
 495         }
 496
 497         mdt_empty_transno(info, rc);
 498
 499         RETURN(rc);
 500
 501 err_out:
 502         if (open_flags & MDS_FMODE_WRITE)
 503                 mdt_write_put(o);
 504         else if (open_flags & MDS_FMODE_EXEC)
 505                 mdt_write_allow(o);
 506
 507         return rc;
 508 }
 509
 510 static int mdt_finish_open(struct mdt_thread_info *info,
 511                            struct mdt_object *p, struct mdt_object *o,
 512                            u64 open_flags, int created,
 513                            struct ldlm_reply *rep)
 514 {
 515         struct ptlrpc_request   *req = mdt_info_req(info);
 516         struct obd_export       *exp = req->rq_export;
 517         struct mdt_export_data  *med = &req->rq_export->exp_mdt_data;
 518         struct md_attr          *ma  = &info->mti_attr;
 519         struct lu_attr          *la  = &ma->ma_attr;
 520         struct mdt_file_data    *mfd;
 521         struct mdt_body         *repbody;
 522         int                      rc = 0;
 523         int                      isreg, isdir, islnk;
 524         struct list_head        *t;
 525         ENTRY;
 526
 527         LASSERT(ma->ma_valid & MA_INODE);
 528
 529         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 530
 531         isreg = S_ISREG(la->la_mode);
 532         isdir = S_ISDIR(la->la_mode);
 533         islnk = S_ISLNK(la->la_mode);
 534         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 535
 536         /* compatibility check for 2.10 clients when it tries to open mirrored
 537          * files. 2.10 clients don't verify overlapping components so they
 538          * would read and write mirrored files just as if they were normal
 539          * PFL files, which will cause the problem that sycned mirrors actually
 540          * contain different data.
 541          * Older clients are not a concern here because they don't even
 542          * understand PFL layout. */
 543         if (isreg && !exp_connect_flr(exp) && ma->ma_valid & MA_LOV &&
 544             mdt_lmm_is_flr(ma->ma_lmm)) {
 545                 /* LU-10286: for simplicity clients who don't understand
 546                  * mirrored layout(with connect flag OBD_CONNECT2_FLR) won't
 547                  * be able to open mirrored files */
 548                 RETURN(-EOPNOTSUPP);
 549         }
 550
 551         /* Overstriped files can crash older clients */
 552         if (isreg && !exp_connect_overstriping(exp) &&
 553             mdt_lmm_is_overstriping(ma->ma_lmm))
 554                 RETURN(-EOPNOTSUPP);
 555
 556         /* LU-2275, simulate broken behaviour (esp. prevalent in
 557          * pre-2.4 servers where a very strange reply is sent on error
 558          * that looks like it was actually almost successful and a
 559          * failure at the same time.) */
 560         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 561                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 562                                                DISP_LOOKUP_NEG |
 563                                                DISP_LOOKUP_POS);
 564
 565                 if (open_flags & MDS_OPEN_LOCK)
 566                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 567
 568                 RETURN(-ENOENT);
 569         }
 570
 571 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
 572         if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 573                 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
 574                 if (IS_ERR(nodemap))
 575                         RETURN(PTR_ERR(nodemap));
 576
 577                 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
 578                 nodemap_putref(nodemap);
 579                 if (rc)
 580                         RETURN(rc);
 581         }
 582 #endif
 583
 584         /*
 585          * If we are following a symlink, don't open; and do not return open
 586          * handle for special nodes as client required.
 587          */
 588         if (islnk || (!isreg && !isdir &&
 589             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 590                 lustre_msg_set_transno(req->rq_repmsg, 0);
 591                 RETURN(0);
 592         }
 593
 594         /*
 595          * We need to return the existing object's fid back, so it is done here,
 596          * after preparing the reply.
 597          */
 598         if (!created && (open_flags & MDS_OPEN_EXCL) &&
 599             (open_flags & MDS_OPEN_CREAT))
 600                 RETURN(-EEXIST);
 601
 602         /* This can't be done earlier, we need to return reply body */
 603         if (isdir) {
 604                 if (open_flags & (MDS_OPEN_CREAT | MDS_FMODE_WRITE)) {
 605                         /* We are trying to create or write an existing dir. */
 606                         RETURN(-EISDIR);
 607                 }
 608         } else if (open_flags & MDS_OPEN_DIRECTORY)
 609                 RETURN(-ENOTDIR);
 610
 611         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 612                                  OBD_FAIL_MDS_LDLM_REPLY_NET | OBD_FAIL_ONCE))
 613                 RETURN(-EAGAIN);
 614
 615         mfd = NULL;
 616         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 617                 spin_lock(&med->med_open_lock);
 618                 list_for_each(t, &med->med_open_head) {
 619                         mfd = list_entry(t, struct mdt_file_data, mfd_list);
 620                         if (mfd->mfd_xid == req->rq_xid)
 621                                 break;
 622                         mfd = NULL;
 623                 }
 624                 spin_unlock(&med->med_open_lock);
 625
 626                 if (mfd != NULL) {
 627                         repbody->mbo_open_handle.cookie =
 628                                 mfd->mfd_open_handle.h_cookie;
 629                         /* set repbody->ea_size for resent case */
 630                         if (ma->ma_valid & MA_LOV) {
 631                                 LASSERT(ma->ma_lmm_size != 0);
 632                                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 633                                 if (isdir)
 634                                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 635                                 else
 636                                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 637                         }
 638                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 639                         RETURN(0);
 640                 }
 641         }
 642
 643         rc = mdt_mfd_open(info, p, o, open_flags, created, rep);
 644         if (!rc)
 645                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 646
 647         RETURN(rc);
 648 }
 649
 650 void mdt_reconstruct_open(struct mdt_thread_info *info,
 651                           struct mdt_lock_handle *lhc)
 652 {
 653         const struct lu_env *env = info->mti_env;
 654         struct mdt_device *mdt = info->mti_mdt;
 655         struct req_capsule *pill = info->mti_pill;
 656         struct ptlrpc_request *req = mdt_info_req(info);
 657         struct md_attr *ma = &info->mti_attr;
 658         struct mdt_reint_record *rr = &info->mti_rr;
 659         u64 open_flags = info->mti_spec.sp_cr_flags;
 660         struct ldlm_reply *ldlm_rep;
 661         struct mdt_object *parent;
 662         struct mdt_object *child;
 663         struct mdt_body *repbody;
 664         u64 opdata;
 665         int rc;
 666         ENTRY;
 667
 668         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 669         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 670         repbody = req_capsule_server_get(pill, &RMF_MDT_BODY);
 671
 672         ma->ma_need = MA_INODE | MA_HSM;
 673         ma->ma_valid = 0;
 674
 675         opdata = mdt_req_from_lrd(req, info->mti_reply_data);
 676         mdt_set_disposition(info, ldlm_rep, opdata);
 677
 678         CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n",
 679                ldlm_rep->lock_policy_res1, req->rq_status);
 680
 681         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE) &&
 682             req->rq_status != 0)
 683                 /* We did not create successfully, return error to client. */
 684                 GOTO(out, rc = req->rq_status);
 685
 686         if (mdt_get_disposition(ldlm_rep, DISP_OPEN_CREATE)) {
 687                 struct obd_export *exp = req->rq_export;
 688                 /*
 689                  * We failed after creation, but we do not know in which step
 690                  * we failed. So try to check the child object.
 691                  */
 692                 parent = mdt_object_find(env, mdt, rr->rr_fid1);
 693                 if (IS_ERR(parent)) {
 694                         rc = PTR_ERR(parent);
 695                         LCONSOLE_WARN("Parent "DFID" lookup error %d."
 696                                       " Evicting client %s with export %s.\n",
 697                                       PFID(rr->rr_fid1), rc,
 698                                       obd_uuid2str(&exp->exp_client_uuid),
 699                                       obd_export_nid2str(exp));
 700                         mdt_export_evict(exp);
 701                         RETURN_EXIT;
 702                 }
 703
 704                 child = mdt_object_find(env, mdt, rr->rr_fid2);
 705                 if (IS_ERR(child)) {
 706                         rc = PTR_ERR(child);
 707                         LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; "
 708                                       "evicting client %s with export %s\n",
 709                                       PFID(rr->rr_fid2), rc,
 710                                       obd_uuid2str(&exp->exp_client_uuid),
 711                                       obd_export_nid2str(exp));
 712                         mdt_object_put(env, parent);
 713                         mdt_export_evict(exp);
 714                         RETURN_EXIT;
 715                 }
 716
 717                 if (unlikely(mdt_object_remote(child))) {
 718                         mdt_object_put(env, parent);
 719                         mdt_object_put(env, child);
 720                         /* the child object was created on remote server */
 721                         if (!mdt_is_dne_client(exp))
 722                                 /* Return -EIO for old client */
 723                                 GOTO(out, rc = -EIO);
 724                         repbody->mbo_fid1 = *rr->rr_fid2;
 725                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 726                         GOTO(out, rc = 0);
 727                 }
 728                 if (mdt_object_exists(child)) {
 729                         mdt_prep_ma_buf_from_rep(info, child, ma);
 730                         rc = mdt_attr_get_complex(info, child, ma);
 731                         if (!rc)
 732                                 rc = mdt_finish_open(info, parent, child,
 733                                                      open_flags, 1, ldlm_rep);
 734                         mdt_object_put(env, parent);
 735                         mdt_object_put(env, child);
 736                         if (!rc)
 737                                 mdt_pack_size2body(info, rr->rr_fid2,
 738                                                    &lhc->mlh_reg_lh);
 739                         GOTO(out, rc);
 740                 }
 741                 /* the child does not exist, we should do regular open */
 742                 mdt_object_put(env, parent);
 743                 mdt_object_put(env, child);
 744         }
 745         /* We did not try to create, so we are a pure open */
 746         rc = mdt_reint_open(info, lhc);
 747         EXIT;
 748 out:
 749         req->rq_status = rc;
 750         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 751         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
 752 }
 753
 754 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep)
 755 {
 756         u64 open_flags = info->mti_spec.sp_cr_flags;
 757         struct mdt_reint_record *rr = &info->mti_rr;
 758         struct md_attr *ma = &info->mti_attr;
 759         struct mdt_object *o;
 760         int rc;
 761
 762         ENTRY;
 763         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
 764         if (IS_ERR(o))
 765                 RETURN(rc = PTR_ERR(o));
 766
 767         if (unlikely(mdt_object_remote(o))) {
 768                 /* the child object was created on remote server */
 769                 struct mdt_body *repbody;
 770
 771                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 772                                                 DISP_LOOKUP_EXECD |
 773                                                 DISP_LOOKUP_POS));
 774                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 775                 repbody->mbo_fid1 = *rr->rr_fid2;
 776                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 777                 rc = 0;
 778         } else {
 779                 if (mdt_object_exists(o)) {
 780                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 781                                                         DISP_LOOKUP_EXECD |
 782                                                         DISP_LOOKUP_POS));
 783                         if ((open_flags & MDS_OPEN_EXCL) &&
 784                             (open_flags & MDS_OPEN_CREAT))
 785                                 mdt_set_disposition(info, rep,
 786                                                     DISP_OPEN_CREATE);
 787
 788                         mdt_prep_ma_buf_from_rep(info, o, ma);
 789                         rc = mdt_attr_get_complex(info, o, ma);
 790                         if (rc == 0)
 791                                 rc = mdt_finish_open(info, NULL, o, open_flags,
 792                                                      0, rep);
 793                 } else {
 794                         rc = -ENOENT;
 795                 }
 796         }
 797
 798         mdt_object_put(info->mti_env, o);
 799         RETURN(rc);
 800 }
 801
 802 /* lock object for open */
 803 static int mdt_object_open_lock(struct mdt_thread_info *info,
 804                                 struct mdt_object *obj,
 805                                 struct mdt_lock_handle *lhc,
 806                                 __u64 *ibits)
 807 {
 808         struct md_attr *ma = &info->mti_attr;
 809         __u64 open_flags = info->mti_spec.sp_cr_flags;
 810         __u64 trybits = 0;
 811         enum ldlm_mode lm = LCK_CR;
 812         bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
 813         bool try_layout = false;
 814         bool create_layout = false;
 815         int rc = 0;
 816         __u32 dom_stripe = 0;
 817         unsigned int dom_only = 0;
 818         unsigned int dom_lock = 0;
 819
 820         ENTRY;
 821
 822         *ibits = 0;
 823         mdt_lock_handle_init(lhc);
 824
 825         if (req_is_replay(mdt_info_req(info)))
 826                 RETURN(0);
 827
 828         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
 829                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
 830                     md_should_create(open_flags))
 831                         create_layout = true;
 832                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
 833                     ma->ma_need & MA_LOV)
 834                         try_layout = true;
 835
 836                 /* DoM files can take IO lock at OPEN when it makes sense,
 837                  * check if file has DoM stripe and ask for lock if client
 838                  * no lock on that resource yet.
 839                  */
 840                 if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
 841                         dom_stripe = mdt_lmm_dom_entry_check(ma->ma_lmm,
 842                                                              &dom_only);
 843                 /* If only DOM stripe is being used then we can expect IO
 844                  * to it after OPEN and will return corresponding DOM ibit
 845                  * using default strategy from mdt_opts.mo_dom_lock.
 846                  * Otherwise trylock mode is used always and DOM ibit will
 847                  * be returned optionally.
 848                  */
 849                 if (dom_stripe &&
 850                     !mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
 851                         dom_lock = !dom_only ? TRYLOCK_DOM_ON_OPEN :
 852                                    info->mti_mdt->mdt_opts.mo_dom_lock;
 853         }
 854
 855         if (acq_lease) {
 856                 /* lease open, acquire write mode of open sem */
 857                 down_write(&obj->mot_open_sem);
 858
 859                 /* Lease exists and ask for new lease */
 860                 if (atomic_read(&obj->mot_lease_count) > 0) {
 861                         /* only exclusive open is supported, so lease
 862                          * are conflicted to each other */
 863                         GOTO(out, rc = -EBUSY);
 864                 }
 865
 866                 /* Lease must be with open lock */
 867                 if (!(open_flags & MDS_OPEN_LOCK)) {
 868                         CERROR("%s: Request lease for file:"DFID ", but open lock "
 869                                "is missed, open_flags = %#llo : rc = %d\n",
 870                                mdt_obd_name(info->mti_mdt),
 871                                PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
 872                         GOTO(out, rc = -EPROTO);
 873                 }
 874
 875                 /* should conflict with new opens for write/execute */
 876                 lm = LCK_PW;
 877                 *ibits = MDS_INODELOCK_OPEN;
 878
 879                 /* never grant LCK_EX layout lock to client */
 880                 try_layout = false;
 881         } else { /* normal open */
 882                 /* normal open holds read mode of open sem */
 883                 down_read(&obj->mot_open_sem);
 884
 885                 if (open_flags & MDS_OPEN_LOCK) {
 886                         if (open_flags & MDS_FMODE_WRITE)
 887                                 lm = LCK_CW;
 888                         else if (open_flags & MDS_FMODE_EXEC)
 889                                 lm = LCK_PR;
 890                         else
 891                                 lm = LCK_CR;
 892
 893                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
 894                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
 895                         if (open_flags & MDS_FMODE_WRITE)
 896                                 lm = LCK_CW;
 897                         else
 898                                 lm = LCK_CR;
 899
 900                         /* revoke lease */
 901                         *ibits = MDS_INODELOCK_OPEN;
 902                         try_layout = false;
 903
 904                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 905                 } else if (dom_lock) {
 906                         lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR;
 907                         trybits |= MDS_INODELOCK_DOM | MDS_INODELOCK_LAYOUT;
 908                 }
 909
 910                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
 911                         PFID(mdt_object_fid(obj)),
 912                         atomic_read(&obj->mot_lease_count), lm);
 913         }
 914
 915         mdt_lock_reg_init(lhc, lm);
 916
 917         /* Return lookup lock to validate inode at the client side.
 918          * This is pretty important otherwise MDT will return layout
 919          * lock for each open.
 920          * However this is a double-edged sword because changing
 921          * permission will revoke a huge number of LOOKUP locks.
 922          */
 923         if (!OBD_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
 924                 if (!(*ibits & MDS_INODELOCK_LOOKUP))
 925                         trybits |= MDS_INODELOCK_LOOKUP;
 926                 trybits |= MDS_INODELOCK_LAYOUT;
 927         }
 928
 929         if (*ibits | trybits)
 930                 rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, false);
 931
 932         CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
 933                ", open_flags = %#llo, try_layout = %d : rc = %d\n",
 934                mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
 935                *ibits, trybits, open_flags, try_layout, rc);
 936
 937         /* will change layout, revoke layout locks by enqueuing EX lock. */
 938         if (rc == 0 && create_layout) {
 939                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
 940
 941                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
 942                         ", open_flags = %#llo\n",
 943                         PFID(mdt_object_fid(obj)), open_flags);
 944
 945                 /* We cannot enqueue another lock for the same resource we
 946                  * already have a lock for, due to mechanics of waiting list
 947                  * iterating in ldlm, see LU-3601.
 948                  * As such we'll drop the open lock we just got above here,
 949                  * it's ok not to have this open lock as it's main purpose is to
 950                  * flush unused cached client open handles. */
 951                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 952                         mdt_object_unlock(info, obj, lhc, 1);
 953
 954                 LASSERT(!try_layout);
 955                 mdt_lock_handle_init(ll);
 956                 mdt_lock_reg_init(ll, LCK_EX);
 957                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT);
 958
 959                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
 960         }
 961
 962         /* Check if there is any other open handles after acquiring
 963          * open lock. At this point, caching open handles have been revoked
 964          * by open lock.
 965          * XXX: Now only exclusive open is supported. Need to check the
 966          * type of open for generic lease support. */
 967         if (rc == 0 && acq_lease) {
 968                 struct ptlrpc_request *req = mdt_info_req(info);
 969                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 970                 struct mdt_file_data *mfd;
 971                 bool is_replay_or_resent;
 972                 int open_count = 0;
 973
 974                 /* For lease: application can open a file and then apply lease,
 975                  * @handle contains original open handle in that case.
 976                  * In recovery, open REQ will be replayed and the lease REQ may
 977                  * be resent that means the open handle is already stale, so we
 978                  * need to fix it up here by finding new handle. */
 979                 is_replay_or_resent = req_is_replay(req) ||
 980                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
 981
 982                 /* if the request is _not_ a replay request, rr_open_handle
 983                  * may be used to hold an open file handle which is issuing the
 984                  * lease request, so that this openhandle doesn't count. */
 985                 mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 986                                           is_replay_or_resent);
 987                 if (mfd != NULL)
 988                         ++open_count;
 989
 990                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
 991                         PFID(mdt_object_fid(obj)),
 992                         atomic_read(&obj->mot_open_count), open_count);
 993
 994                 if (atomic_read(&obj->mot_open_count) > open_count) {
 995                         /* fail if anyone *else* has opened file for write */
 996                         if (mdt_write_read(obj) > 1)
 997                                 GOTO(out, rc = -EBUSY);
 998                 }
 999         }
1000         GOTO(out, rc);
1001
1002 out:
1003         RETURN(rc);
1004 }
1005
1006 static void mdt_object_open_unlock(struct mdt_thread_info *info,
1007                                    struct mdt_object *obj,
1008                                    struct mdt_lock_handle *lhc,
1009                                    __u64 ibits, int rc)
1010 {
1011         __u64 open_flags = info->mti_spec.sp_cr_flags;
1012         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1013         ENTRY;
1014
1015         if (req_is_replay(mdt_info_req(info)))
1016                 RETURN_EXIT;
1017
1018         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1019          * return to client side. */
1020         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1021                 mdt_object_unlock(info, obj, ll, 1);
1022
1023         ll = &info->mti_lh[MDT_LH_LAYOUT];
1024         /* Release local layout lock, layout was created */
1025         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1026                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1027                 mdt_object_unlock(info, obj, ll, 1);
1028         }
1029
1030         if (open_flags & MDS_OPEN_LEASE)
1031                 up_write(&obj->mot_open_sem);
1032         else
1033                 up_read(&obj->mot_open_sem);
1034
1035         /* Cross-ref case, the lock should be returned to the client */
1036         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1037                 RETURN_EXIT;
1038
1039         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1040             !(ibits & MDS_INODELOCK_DOM)) {
1041                 /* for the open request, the lock will only return to client
1042                  * if open or layout lock is granted. */
1043                 rc = 1;
1044         }
1045
1046         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1047                 struct ldlm_reply       *ldlm_rep;
1048
1049                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1050                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1051                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1052                         mdt_object_unlock(info, obj, lhc, 1);
1053         }
1054         RETURN_EXIT;
1055 }
1056
1057 /**
1058  * Check release is permitted for the current HSM flags.
1059  */
1060 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1061 {
1062         if (!(ma->ma_valid & MA_HSM))
1063                 return false;
1064
1065         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1066                 return false;
1067
1068         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1069                 return false;
1070
1071         return true;
1072 }
1073
1074 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1075                                 struct ldlm_reply *rep,
1076                                 struct mdt_lock_handle *lhc)
1077 {
1078         const struct lu_env *env = info->mti_env;
1079         struct mdt_device *mdt = info->mti_mdt;
1080         u64 open_flags = info->mti_spec.sp_cr_flags;
1081         struct mdt_reint_record *rr = &info->mti_rr;
1082         struct md_attr *ma = &info->mti_attr;
1083         struct mdt_object *parent = NULL;
1084         struct mdt_object *o;
1085         bool object_locked = false;
1086         u64 ibits = 0;
1087         int rc;
1088
1089         ENTRY;
1090         if (md_should_create(open_flags)) {
1091                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1092                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1093                         if (IS_ERR(parent)) {
1094                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1095                                        " for anonymous created %ld, try to"
1096                                        " use server-side parent.\n",
1097                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1098                                 parent = NULL;
1099                         }
1100                 }
1101                 if (parent == NULL)
1102                         ma->ma_need |= MA_PFID;
1103         }
1104
1105         o = mdt_object_find(env, mdt, rr->rr_fid2);
1106         if (IS_ERR(o))
1107                 GOTO(out_parent_put, rc = PTR_ERR(o));
1108
1109         if (mdt_object_remote(o)) {
1110                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1111                        mdt_obd_name(info->mti_mdt),
1112                        PFID(rr->rr_fid2));
1113                 GOTO(out, rc = -EREMOTE);
1114         } else if (!mdt_object_exists(o)) {
1115                 mdt_set_disposition(info, rep,
1116                                     DISP_IT_EXECD |
1117                                     DISP_LOOKUP_EXECD |
1118                                     DISP_LOOKUP_NEG);
1119                 GOTO(out, rc = -ENOENT);
1120         }
1121
1122         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1123
1124         mdt_prep_ma_buf_from_rep(info, o, ma);
1125         if (open_flags & MDS_OPEN_RELEASE)
1126                 ma->ma_need |= MA_HSM;
1127         rc = mdt_attr_get_complex(info, o, ma);
1128         if (rc)
1129                 GOTO(out, rc);
1130
1131         /* We should not change file's existing LOV EA */
1132         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1133             open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1134                 GOTO(out, rc = -EEXIST);
1135
1136         /* If a release request, check file open flags are fine and ask for an
1137          * exclusive open access. */
1138         if (open_flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1139                 GOTO(out, rc = -EPERM);
1140
1141         rc = mdt_check_resent_lock(info, o, lhc);
1142         if (rc < 0) {
1143                 GOTO(out, rc);
1144         } else if (rc > 0) {
1145                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1146                 object_locked = true;
1147                 if (rc)
1148                         GOTO(out_unlock, rc);
1149         }
1150
1151         if (ma->ma_valid & MA_PFID) {
1152                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1153                 if (IS_ERR(parent)) {
1154                         CDEBUG(D_INODE, "Fail to find parent "DFID
1155                                " for anonymous created %ld, try to"
1156                                " use system default.\n",
1157                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1158                         parent = NULL;
1159                 }
1160         }
1161
1162         rc = mdt_finish_open(info, parent, o, open_flags, 0, rep);
1163         if (!rc) {
1164                 mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1165                 if (open_flags & MDS_OPEN_LOCK)
1166                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1167                 if (open_flags & MDS_OPEN_LEASE)
1168                         mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1169         }
1170         GOTO(out_unlock, rc);
1171
1172 out_unlock:
1173         if (object_locked)
1174                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1175 out:
1176         mdt_object_put(env, o);
1177         if (rc == 0)
1178                 mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
1179 out_parent_put:
1180         if (parent != NULL)
1181                 mdt_object_put(env, parent);
1182         return rc;
1183 }
1184
1185 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1186 static int mdt_cross_open(struct mdt_thread_info *info,
1187                           const struct lu_fid *parent_fid,
1188                           const struct lu_fid *fid,
1189                           struct ldlm_reply *rep, u64 open_flags)
1190 {
1191         struct md_attr *ma = &info->mti_attr;
1192         struct mdt_object *o;
1193         int rc;
1194         ENTRY;
1195
1196         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1197         if (IS_ERR(o))
1198                 RETURN(rc = PTR_ERR(o));
1199
1200         if (mdt_object_remote(o)) {
1201                 /* Something is wrong here, the object is on another MDS! */
1202                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1203                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1204                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1205                                 &o->mot_obj,
1206                                 "Object isn't on this server! FLD error?");
1207                 rc = -EFAULT;
1208         } else {
1209                 if (mdt_object_exists(o)) {
1210                         int mask;
1211
1212                         /* Do permission check for cross-open after converting
1213                          * MDS_OPEN_* flags to MAY_* permission mask.
1214                          */
1215                         mask = mds_accmode(open_flags);
1216
1217                         rc = mo_permission(info->mti_env, NULL,
1218                                            mdt_object_child(o), NULL, mask);
1219                         if (rc)
1220                                 goto out;
1221
1222                         mdt_prep_ma_buf_from_rep(info, o, ma);
1223                         rc = mdt_attr_get_complex(info, o, ma);
1224                         if (rc != 0)
1225                                 GOTO(out, rc);
1226
1227                         rc = mdt_pack_secctx_in_reply(info, o);
1228                         if (unlikely(rc))
1229                                 GOTO(out, rc);
1230
1231                         rc = mdt_pack_encctx_in_reply(info, o);
1232                         if (unlikely(rc))
1233                                 GOTO(out, rc);
1234
1235                         rc = mdt_finish_open(info, NULL, o, open_flags, 0, rep);
1236                 } else {
1237                         /*
1238                          * Something is wrong here. lookup was positive but
1239                          * there is no object!
1240                          */
1241                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1242                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1243                         rc = -EFAULT;
1244                 }
1245         }
1246 out:
1247         mdt_object_put(info->mti_env, o);
1248
1249         RETURN(rc);
1250 }
1251
1252 /*
1253  * find root object and take its xattr lock if it's on remote MDT, later create
1254  * may use fs default striping (which is stored in root xattr).
1255  */
1256 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1257                                struct mdt_device *mdt)
1258 {
1259         struct mdt_object *md_root = mdt->mdt_md_root;
1260         struct lustre_handle lhroot;
1261         int rc;
1262
1263         if (md_root == NULL) {
1264                 lu_root_fid(&info->mti_tmp_fid1);
1265                 md_root = mdt_object_find(info->mti_env, mdt,
1266                                           &info->mti_tmp_fid1);
1267                 if (IS_ERR(md_root))
1268                         return PTR_ERR(md_root);
1269
1270                 spin_lock(&mdt->mdt_lock);
1271                 if (mdt->mdt_md_root != NULL) {
1272                         spin_unlock(&mdt->mdt_lock);
1273
1274                         LASSERTF(mdt->mdt_md_root == md_root,
1275                                  "Different root object ("
1276                                  DFID") instances, %p, %p\n",
1277                                  PFID(&info->mti_tmp_fid1),
1278                                  mdt->mdt_md_root, md_root);
1279                         LASSERT(atomic_read(
1280                                 &md_root->mot_obj.lo_header->loh_ref) > 1);
1281
1282                         mdt_object_put(info->mti_env, md_root);
1283                 } else {
1284                         mdt->mdt_md_root = md_root;
1285                         spin_unlock(&mdt->mdt_lock);
1286                 }
1287         }
1288
1289         if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1290                 return 0;
1291
1292         rc = mdt_remote_object_lock(info, md_root, mdt_object_fid(md_root),
1293                                     &lhroot, LCK_PR, MDS_INODELOCK_XATTR,
1294                                     true);
1295         if (rc < 0)
1296                 return rc;
1297
1298         md_root->mot_cache_attr = 1;
1299
1300         /* don't cancel this lock, so that we know the cached xattr is valid. */
1301         ldlm_lock_decref(&lhroot, LCK_PR);
1302
1303         return 0;
1304 }
1305
1306 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1307 {
1308         struct mdt_device *mdt = info->mti_mdt;
1309         struct ptlrpc_request *req = mdt_info_req(info);
1310         struct mdt_object *parent;
1311         struct mdt_object *child;
1312         struct mdt_lock_handle *lh = NULL;
1313         struct ldlm_reply *ldlm_rep;
1314         struct mdt_body *repbody;
1315         struct lu_fid *child_fid = &info->mti_tmp_fid1;
1316         struct md_attr *ma = &info->mti_attr;
1317         u64 open_flags = info->mti_spec.sp_cr_flags;
1318         u64 ibits = 0;
1319         struct mdt_reint_record *rr = &info->mti_rr;
1320         int result, rc;
1321         int created = 0;
1322         int object_locked = 0;
1323         enum ldlm_mode lock_mode = LCK_PR;
1324         u32 msg_flags;
1325         ktime_t kstart = ktime_get();
1326
1327         ENTRY;
1328         OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, OBD_FAIL_ONCE,
1329                                (obd_timeout + 1) / 4);
1330
1331         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1332
1333         ma->ma_need = MA_INODE;
1334         ma->ma_valid = 0;
1335
1336         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1337         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1338
1339         if (unlikely(open_flags & MDS_OPEN_JOIN_FILE)) {
1340                 CERROR("file join is not supported anymore.\n");
1341                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1342         }
1343         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1344
1345         if ((open_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1346             info->mti_spec.u.sp_ea.eadata == NULL)
1347                 GOTO(out, result = err_serious(-EINVAL));
1348
1349         if (open_flags & MDS_FMODE_WRITE &&
1350             exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1351                 GOTO(out, result = -EROFS);
1352
1353         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1354                "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1355                PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2),
1356                open_flags, ma->ma_attr.la_mode, msg_flags);
1357
1358         if (info->mti_cross_ref) {
1359                 /* This is cross-ref open */
1360                 mdt_set_disposition(info, ldlm_rep,
1361                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1362                              DISP_LOOKUP_POS));
1363                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1364                                         ldlm_rep, open_flags);
1365                 GOTO(out, result);
1366         } else if (req_is_replay(req)) {
1367                 result = mdt_open_by_fid(info, ldlm_rep);
1368
1369                 if (result != -ENOENT)
1370                         GOTO(out, result);
1371
1372                 /* We didn't find the correct object, so we need to re-create it
1373                  * via a regular replay. */
1374                 if (!(open_flags & MDS_OPEN_CREAT)) {
1375                         DEBUG_REQ(D_ERROR, req,
1376                                   "OPEN & CREAT not in open replay/by_fid");
1377                         GOTO(out, result = -EFAULT);
1378                 }
1379                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1380         } else if (open_flags & MDS_OPEN_BY_FID) {
1381                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1382                 if (result < 0)
1383                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1384                                PFID(rr->rr_fid2), result);
1385                 GOTO(out, result);
1386         }
1387
1388         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1389                 GOTO(out, result = err_serious(-ENOMEM));
1390
1391         mdt_set_disposition(info, ldlm_rep,
1392                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1393
1394         if (!lu_name_is_valid(&rr->rr_name))
1395                 GOTO(out, result = -EPROTO);
1396
1397         result = mdt_lock_root_xattr(info, mdt);
1398         if (result < 0)
1399                 GOTO(out, result);
1400
1401         parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1402         if (IS_ERR(parent))
1403                 GOTO(out, result = PTR_ERR(parent));
1404
1405         /* get and check version of parent */
1406         result = mdt_version_get_check(info, parent, 0);
1407         if (result) {
1408                 mdt_object_put(info->mti_env, parent);
1409                 GOTO(out, result);
1410         }
1411
1412         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
1413 again_pw:
1414         fid_zero(child_fid);
1415
1416         if (open_flags & MDS_OPEN_VOLATILE) {
1417                 lh = NULL;
1418                 result = -ENOENT;
1419         } else {
1420                 lh = &info->mti_lh[MDT_LH_PARENT];
1421                 mdt_lock_pdo_init(lh, lock_mode, &rr->rr_name);
1422                 result = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
1423                 if (result != 0) {
1424                         mdt_object_put(info->mti_env, parent);
1425                         GOTO(out, result);
1426                 }
1427
1428                 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1429                                     &rr->rr_name, child_fid, &info->mti_spec);
1430         }
1431
1432         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1433                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1434                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1435                  PFID(child_fid));
1436
1437         if (result != 0 && result != -ENOENT)
1438                 GOTO(out_parent, result);
1439
1440         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1441
1442         if (result == -ENOENT) {
1443                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1444                 if (!(open_flags & MDS_OPEN_CREAT))
1445                         GOTO(out_parent, result);
1446                 if (mdt_rdonly(req->rq_export))
1447                         GOTO(out_parent, result = -EROFS);
1448
1449                 LASSERT(equi(lh == NULL, open_flags & MDS_OPEN_VOLATILE));
1450
1451                 if (lh != NULL && lock_mode == LCK_PR) {
1452                         /* first pass: get write lock and restart */
1453                         mdt_object_unlock(info, parent, lh, 1);
1454                         mdt_clear_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1455                         mdt_lock_handle_init(lh);
1456                         lock_mode = LCK_PW;
1457                         goto again_pw;
1458                 }
1459
1460                 *child_fid = *info->mti_rr.rr_fid2;
1461                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1462                          PFID(child_fid));
1463                 /* In the function below, .hs_keycmp resolves to
1464                  * lu_obj_hop_keycmp() */
1465                 /* coverity[overrun-buffer-val] */
1466                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1467         } else {
1468                 /*
1469                  * Check for O_EXCL is moved to the mdt_finish_open(),
1470                  * we need to return FID back in that case.
1471                  */
1472                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1473                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1474         }
1475         if (IS_ERR(child))
1476                 GOTO(out_parent, result = PTR_ERR(child));
1477
1478         /** check version of child  */
1479         rc = mdt_version_get_check(info, child, 1);
1480         if (rc)
1481                 GOTO(out_child, result = rc);
1482
1483         if (result == -ENOENT) {
1484                 /* Create under OBF and .lustre is not permitted */
1485                 if (!fid_is_md_operative(rr->rr_fid1) &&
1486                     (open_flags & MDS_OPEN_VOLATILE) == 0)
1487                         GOTO(out_child, result = -EPERM);
1488
1489                 /* save versions in reply */
1490                 mdt_version_get_save(info, parent, 0);
1491                 mdt_version_get_save(info, child, 1);
1492
1493                 /* version of child will be changed */
1494                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1495
1496                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1497                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1498
1499                 /* Don't do lookup sanity check. We know name doesn't exist. */
1500                 info->mti_spec.sp_cr_lookup = 0;
1501                 info->mti_spec.sp_feat = &dt_directory_features;
1502
1503                 result = mdo_create(info->mti_env, mdt_object_child(parent),
1504                                     &rr->rr_name, mdt_object_child(child),
1505                                     &info->mti_spec, &info->mti_attr);
1506                 if (result == -ERESTART) {
1507                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1508                         GOTO(out_child, result);
1509                 } else {
1510                         mdt_prep_ma_buf_from_rep(info, child, ma);
1511                         /* XXX: we should call this once, see few lines below */
1512                         if (result == 0)
1513                                 result = mdt_attr_get_complex(info, child, ma);
1514
1515                         if (result != 0)
1516                                 GOTO(out_child, result);
1517                 }
1518                 created = 1;
1519                 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1520                                  ktime_us_delta(ktime_get(), kstart));
1521         } else {
1522                 /*
1523                  * The object is on remote node, return its FID for remote open.
1524                  */
1525                 if (mdt_object_remote(child)) {
1526                         /*
1527                          * Check if this lock already was sent to client and
1528                          * this is resent case. For resent case do not take lock
1529                          * again, use what is already granted.
1530                          */
1531                         LASSERT(lhc != NULL);
1532
1533                         rc = mdt_check_resent_lock(info, child, lhc);
1534                         if (rc < 0) {
1535                                 GOTO(out_child, result = rc);
1536                         } else if (rc > 0) {
1537                                 mdt_lock_handle_init(lhc);
1538                                 mdt_lock_reg_init(lhc, LCK_PR);
1539
1540                                 rc = mdt_object_lock(info, child, lhc,
1541                                                      MDS_INODELOCK_LOOKUP);
1542                         }
1543                         repbody->mbo_fid1 = *mdt_object_fid(child);
1544                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1545                         if (rc != 0)
1546                                 result = rc;
1547                         else
1548                                 result = -MDT_EREMOTE_OPEN;
1549                         GOTO(out_child, result);
1550                 } else if (mdt_object_exists(child)) {
1551                         /* Check early for MDS_OPEN_DIRECTORY/O_DIRECTORY to
1552                          * avoid opening regular files from lfs getstripe
1553                          * since doing so breaks the leases used by lfs
1554                          * mirror. See LU-13693. */
1555                         if (open_flags & MDS_OPEN_DIRECTORY &&
1556                             S_ISREG(lu_object_attr(&child->mot_obj)))
1557                                 GOTO(out_child, result = -ENOTDIR);
1558
1559                         /* We have to get attr & LOV EA & HSM for this
1560                          * object. */
1561                         mdt_prep_ma_buf_from_rep(info, child, ma);
1562                         ma->ma_need |= MA_HSM;
1563                         result = mdt_attr_get_complex(info, child, ma);
1564                         if (result != 0)
1565                                 GOTO(out_child, result);
1566                 } else {
1567                         /* Object does not exist. Likely FS corruption. */
1568                         CERROR("%s: name '"DNAME"' present, but FID "
1569                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1570                                PNAME(&rr->rr_name), PFID(child_fid));
1571                         GOTO(out_child, result = -EIO);
1572                 }
1573         }
1574
1575         repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1576         repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1577
1578         rc = mdt_pack_secctx_in_reply(info, child);
1579         if (unlikely(rc))
1580                 GOTO(out_child, result = rc);
1581
1582         rc = mdt_pack_encctx_in_reply(info, child);
1583         if (unlikely(rc))
1584                 GOTO(out_child, result = rc);
1585
1586         rc = mdt_check_resent_lock(info, child, lhc);
1587         if (rc < 0) {
1588                 GOTO(out_child, result = rc);
1589         } else if (rc == 0) {
1590                 /* the open lock might already be gotten in
1591                  * ldlm_handle_enqueue() */
1592                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1593                 if (open_flags & MDS_OPEN_LOCK)
1594                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1595         } else {
1596                 /* get openlock if this isn't replay and client requested it */
1597                 if (!req_is_replay(req)) {
1598                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1599                         object_locked = 1;
1600                         if (rc != 0)
1601                                 GOTO(out_child_unlock, result = rc);
1602                         else if (open_flags & MDS_OPEN_LOCK)
1603                                 mdt_set_disposition(info, ldlm_rep,
1604                                                     DISP_OPEN_LOCK);
1605                 }
1606         }
1607         /* Try to open it now. */
1608         rc = mdt_finish_open(info, parent, child, open_flags,
1609                              created, ldlm_rep);
1610         if (rc) {
1611                 result = rc;
1612                 /* openlock will be released if mdt_finish_open() failed */
1613                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1614
1615                 if (created && (open_flags & MDS_OPEN_VOLATILE)) {
1616                         CERROR("%s: cannot open volatile file "DFID", orphan "
1617                                "file will be left in PENDING directory until "
1618                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1619                                PFID(mdt_object_fid(child)), rc);
1620                         GOTO(out_child_unlock, result);
1621                 }
1622
1623                 if (created) {
1624                         ma->ma_need = 0;
1625                         ma->ma_valid = 0;
1626                         rc = mdo_unlink(info->mti_env,
1627                                         mdt_object_child(parent),
1628                                         mdt_object_child(child),
1629                                         &rr->rr_name,
1630                                         &info->mti_attr, 0);
1631                         if (rc != 0)
1632                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1633                                        mdt_obd_name(info->mti_mdt),
1634                                        PFID(mdt_object_fid(child)), rc);
1635                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1636                 }
1637         }
1638
1639         mdt_counter_incr(req, LPROC_MDT_OPEN,
1640                          ktime_us_delta(ktime_get(), kstart));
1641
1642         EXIT;
1643 out_child_unlock:
1644         if (object_locked)
1645                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1646 out_child:
1647         mdt_object_put(info->mti_env, child);
1648         if (result == 0)
1649                 mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh);
1650 out_parent:
1651         if (lh != NULL)
1652                 mdt_object_unlock(info, parent, lh, result || !created);
1653
1654         mdt_object_put(info->mti_env, parent);
1655 out:
1656         if (result)
1657                 lustre_msg_set_transno(req->rq_repmsg, 0);
1658         return result;
1659 }
1660
1661 /**
1662  * Create an orphan object use local root.
1663  */
1664 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1665                                           struct mdt_device *mdt,
1666                                           const struct lu_fid *fid,
1667                                           struct md_attr *attr, fmode_t fmode)
1668 {
1669         const struct lu_env *env = info->mti_env;
1670         struct md_op_spec *spec = &info->mti_spec;
1671         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1672         struct mdt_object *obj = NULL;
1673         struct mdt_object *local_root;
1674         static const struct lu_name lname = {
1675                 .ln_name = "i_am_nobody",
1676                 .ln_namelen = sizeof("i_am_nobody") - 1,
1677         };
1678         struct lu_ucred *uc;
1679         kernel_cap_t uc_cap_save;
1680         int rc;
1681         ENTRY;
1682
1683         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1684         if (rc != 0)
1685                 RETURN(ERR_PTR(rc));
1686
1687         local_root = mdt_object_find(env, mdt, local_root_fid);
1688         if (IS_ERR(local_root))
1689                 RETURN(local_root);
1690
1691         obj = mdt_object_new(env, mdt, fid);
1692         if (IS_ERR(obj))
1693                 GOTO(out, rc = PTR_ERR(obj));
1694
1695         spec->sp_cr_lookup = 0;
1696         spec->sp_feat = &dt_directory_features;
1697         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1698         if (attr->ma_valid & MA_LOV) {
1699                 spec->u.sp_ea.eadata = attr->ma_lmm;
1700                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1701                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1702         } else {
1703                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1704         }
1705
1706         uc = lu_ucred(env);
1707         uc_cap_save = uc->uc_cap;
1708         cap_raise(uc->uc_cap, CAP_DAC_OVERRIDE);
1709         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1710                         mdt_object_child(obj), spec, attr);
1711         uc->uc_cap = uc_cap_save;
1712         if (rc < 0) {
1713                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1714                        mdt_obd_name(mdt), PFID(fid), rc);
1715                 GOTO(out, rc);
1716         }
1717
1718         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED, spec);
1719         if (rc < 0)
1720                 CERROR("%s: cannot open volatile file "DFID", orphan "
1721                        "file will be left in PENDING directory until "
1722                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1723                        PFID(fid), rc);
1724         GOTO(out, rc);
1725
1726 out:
1727         if (rc < 0) {
1728                 if (!IS_ERR(obj))
1729                         mdt_object_put(env, obj);
1730                 obj = ERR_PTR(rc);
1731         }
1732         mdt_object_put(env, local_root);
1733         return obj;
1734 }
1735
1736 /* XXX Look into layout in MDT layer. */
1737 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1738 {
1739         struct lov_comp_md_v1   *comp_v1;
1740         struct lov_mds_md       *v1;
1741         __u32   off;
1742         int     i;
1743
1744         if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1745                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1746
1747                 if (comp_v1->lcm_entry_count == 0)
1748                         return -EINVAL;
1749
1750                 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1751                         off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1752                         v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1753                         v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1754                 }
1755         } else {
1756                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1757         }
1758         return 0;
1759 }
1760
1761 static inline int mdt_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1762 {
1763         struct lov_comp_md_v1 *comp_v1;
1764
1765         if (le32_to_cpu(lmm->lmm_magic == LOV_MAGIC_COMP_V1)) {
1766                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1767                 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1768         } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1769                    le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1770                 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1771         } else {
1772                 return -EINVAL;
1773         }
1774         return 0;
1775 }
1776
1777 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1778                            struct md_attr *ma)
1779 {
1780         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1781         struct lu_ucred        *uc = mdt_ucred(info);
1782         struct close_data      *data;
1783         struct ldlm_lock       *lease;
1784         struct mdt_object      *orphan;
1785         struct md_attr         *orp_ma;
1786         struct lu_buf          *buf;
1787         kernel_cap_t cap;
1788         bool                    lease_broken;
1789         int                     rc;
1790         int                     rc2;
1791         ENTRY;
1792
1793         if (mdt_rdonly(info->mti_exp))
1794                 RETURN(-EROFS);
1795
1796         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1797         if (data == NULL)
1798                 RETURN(-EPROTO);
1799
1800         lease = ldlm_handle2lock(&data->cd_handle);
1801         if (lease == NULL)
1802                 RETURN(-ESTALE);
1803
1804         /* try to hold open_sem so that nobody else can open the file */
1805         if (!down_write_trylock(&o->mot_open_sem)) {
1806                 ldlm_lock_cancel(lease);
1807                 GOTO(out_reprocess, rc = -EBUSY);
1808         }
1809
1810         /* Check if the lease open lease has already canceled */
1811         lock_res_and_lock(lease);
1812         lease_broken = ldlm_is_cancel(lease);
1813         unlock_res_and_lock(lease);
1814
1815         LDLM_DEBUG(lease, DFID " lease broken? %d",
1816                    PFID(mdt_object_fid(o)), lease_broken);
1817
1818         /* Cancel server side lease. Client side counterpart should
1819          * have been cancelled. It's okay to cancel it now as we've
1820          * held mot_open_sem. */
1821         ldlm_lock_cancel(lease);
1822
1823         if (lease_broken) /* don't perform release task */
1824                 GOTO(out_unlock, rc = -ESTALE);
1825
1826         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1827                 GOTO(out_unlock, rc = -EINVAL);
1828
1829         /* ma_need was set before but it seems fine to change it in order to
1830          * avoid modifying the one from RPC */
1831         ma->ma_need = MA_HSM;
1832         rc = mdt_attr_get_complex(info, o, ma);
1833         if (rc != 0)
1834                 GOTO(out_unlock, rc);
1835
1836         if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
1837                 if (ma->ma_valid & MA_HSM) {
1838                         if (ma->ma_hsm.mh_flags & HS_RELEASED)
1839                                 GOTO(out_unlock, rc = -EALREADY);
1840
1841                         if (ma->ma_hsm.mh_arch_id != data->cd_archive_id)
1842                                 CDEBUG(D_CACHE,
1843                                        DFID" archive id diff: %llu:%u\n",
1844                                        PFID(mdt_object_fid(o)),
1845                                        ma->ma_hsm.mh_arch_id,
1846                                        data->cd_archive_id);
1847
1848                         if (!(ma->ma_hsm.mh_flags & HS_DIRTY) &&
1849                             ma->ma_hsm.mh_arch_ver == data->cd_data_version) {
1850                                 CDEBUG(D_CACHE,
1851                                        DFID" data version matches: packed=%llu "
1852                                        "and on-disk=%llu\n",
1853                                        PFID(mdt_object_fid(o)),
1854                                        data->cd_data_version,
1855                                        ma->ma_hsm.mh_arch_ver);
1856                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1857                         }
1858
1859                         if (ma->ma_hsm.mh_flags & HS_DIRTY)
1860                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1861                 } else {
1862                         /* Set up HSM attribte for PCC archived object */
1863                         BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1864                                      sizeof(info->mti_xattr_buf));
1865                         buf = &info->mti_buf;
1866                         buf->lb_buf = info->mti_xattr_buf;
1867                         buf->lb_len = sizeof(struct hsm_attrs);
1868                         memset(&ma->ma_hsm, 0, sizeof(ma->ma_hsm));
1869                         ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1870                         ma->ma_hsm.mh_arch_id = data->cd_archive_id;
1871                         ma->ma_hsm.mh_arch_ver = data->cd_data_version;
1872                         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1873
1874                         rc = mo_xattr_set(info->mti_env, mdt_object_child(o),
1875                                           buf, XATTR_NAME_HSM, 0);
1876                         if (rc)
1877                                 GOTO(out_unlock, rc);
1878                 }
1879         } else {
1880                 if (!mdt_hsm_release_allow(ma))
1881                         GOTO(out_unlock, rc = -EPERM);
1882
1883                 /* already released? */
1884                 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1885                         GOTO(out_unlock, rc = 0);
1886
1887                 /* Compare on-disk and packed data_version */
1888                 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1889                         CDEBUG(D_HSM, DFID" data_version mismatches: "
1890                                "packed=%llu and on-disk=%llu\n",
1891                                PFID(mdt_object_fid(o)),
1892                                data->cd_data_version,
1893                                ma->ma_hsm.mh_arch_ver);
1894                         GOTO(out_unlock, rc = -EPERM);
1895                 }
1896         }
1897
1898         ma->ma_valid = MA_INODE;
1899         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1900         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1901         if (rc < 0)
1902                 GOTO(out_unlock, rc);
1903
1904         mutex_lock(&o->mot_som_mutex);
1905         rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size,
1906                            ma->ma_attr.la_blocks);
1907         mutex_unlock(&o->mot_som_mutex);
1908         if (rc2 < 0)
1909                 CDEBUG(D_INODE,
1910                        "%s: File "DFID" SOM update failed: rc = %d\n",
1911                        mdt_obd_name(info->mti_mdt),
1912                        PFID(mdt_object_fid(o)), rc2);
1913
1914
1915         ma->ma_need = MA_INODE | MA_LOV;
1916         rc = mdt_attr_get_complex(info, o, ma);
1917         if (rc < 0)
1918                 GOTO(out_unlock, rc);
1919
1920         if (!(ma->ma_valid & MA_LOV)) {
1921                 /* Even empty file are released */
1922                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
1923                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
1924                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1925                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
1926                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
1927         } else {
1928                 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
1929                  * ma_lmm as lov_user_md, then it will be confused by union of
1930                  * layout_gen and stripe_offset. */
1931                 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
1932                     LOV_MAGIC_MAGIC)
1933                         ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
1934                 else
1935                         GOTO(out_unlock, rc = -EINVAL);
1936         }
1937
1938         /* Set file as released. */
1939         rc = mdt_hsm_set_released(ma->ma_lmm);
1940         if (rc)
1941                 GOTO(out_unlock, rc);
1942
1943         orp_ma = &info->mti_u.hsm.attr;
1944         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
1945         /* We use root ownership to bypass potential quota
1946          * restrictions on the user and group of the file. */
1947         orp_ma->ma_attr.la_uid = 0;
1948         orp_ma->ma_attr.la_gid = 0;
1949         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
1950         orp_ma->ma_lmm = ma->ma_lmm;
1951         orp_ma->ma_lmm_size = ma->ma_lmm_size;
1952         orp_ma->ma_valid = MA_INODE | MA_LOV;
1953         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
1954                                  MDS_FMODE_WRITE);
1955         if (IS_ERR(orphan)) {
1956                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
1957                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
1958                        PTR_ERR(orphan));
1959                 GOTO(out_unlock, rc = PTR_ERR(orphan));
1960         }
1961
1962         /* Set up HSM attribute for orphan object */
1963         BUILD_BUG_ON(sizeof(struct hsm_attrs) > sizeof(info->mti_xattr_buf));
1964         buf = &info->mti_buf;
1965         buf->lb_buf = info->mti_xattr_buf;
1966         buf->lb_len = sizeof(struct hsm_attrs);
1967         ma->ma_hsm.mh_flags |= HS_RELEASED;
1968         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1969         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
1970
1971         mdt_lock_reg_init(lh, LCK_EX);
1972         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
1973                              MDS_INODELOCK_XATTR);
1974         if (rc != 0)
1975                 GOTO(out_close, rc);
1976
1977         /* The orphan has root ownership so we need to raise
1978          * CAP_FOWNER to set the HSM attributes. */
1979         cap = uc->uc_cap;
1980         cap_raise(uc->uc_cap, CAP_FOWNER);
1981         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
1982                           XATTR_NAME_HSM, 0);
1983         uc->uc_cap = cap;
1984         if (rc != 0)
1985                 GOTO(out_layout_lock, rc);
1986
1987         /* Swap layout with orphan objects. */
1988         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
1989                              mdt_object_child(orphan),
1990                              SWAP_LAYOUTS_MDS_HSM);
1991
1992         if (!rc && ma->ma_attr_flags & MDS_PCC_ATTACH) {
1993                 ma->ma_need = MA_LOV;
1994                 rc = mdt_attr_get_complex(info, o, ma);
1995         }
1996
1997         EXIT;
1998
1999 out_layout_lock:
2000         /* Release exclusive LL */
2001         mdt_object_unlock(info, o, lh, 1);
2002 out_close:
2003         /* Close orphan object anyway */
2004         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2005                        MDS_FMODE_WRITE);
2006         if (rc2 < 0)
2007                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2008                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2009         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2010                         "object closed");
2011         mdt_object_put(info->mti_env, orphan);
2012
2013 out_unlock:
2014         up_write(&o->mot_open_sem);
2015
2016         /* already released */
2017         if (rc == 0) {
2018                 struct mdt_body *repbody;
2019
2020                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2021                 LASSERT(repbody != NULL);
2022                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2023                 if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
2024                         LASSERT(ma->ma_valid & MA_LOV);
2025                         rc = mdt_get_lmm_gen(ma->ma_lmm,
2026                                              &repbody->mbo_layout_gen);
2027                         if (!rc)
2028                                 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
2029                 }
2030         }
2031
2032 out_reprocess:
2033         ldlm_reprocess_all(lease->l_resource,
2034                            lease->l_policy_data.l_inodebits.bits);
2035         LDLM_LOCK_PUT(lease);
2036
2037         ma->ma_valid = 0;
2038         ma->ma_need = 0;
2039
2040         return rc;
2041 }
2042
2043 int mdt_close_handle_layouts(struct mdt_thread_info *info,
2044                              struct mdt_object *o, struct md_attr *ma)
2045 {
2046         struct mdt_lock_handle  *lh1 = &info->mti_lh[MDT_LH_NEW];
2047         struct mdt_lock_handle  *lh2 = &info->mti_lh[MDT_LH_OLD];
2048         struct close_data       *data;
2049         struct ldlm_lock        *lease;
2050         struct mdt_object       *o1 = o, *o2 = NULL;
2051         bool                     lease_broken;
2052         bool                     swap_objects = false;
2053         int                      rc;
2054         ENTRY;
2055
2056         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2057                 RETURN(-EROFS);
2058
2059         if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
2060                 RETURN(-EINVAL);
2061
2062         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2063         if (data == NULL)
2064                 RETURN(-EPROTO);
2065
2066         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2067                 RETURN(-EINVAL);
2068
2069         rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
2070         if (rc == 0) {
2071                 /**
2072                  * only MDS_CLOSE_LAYOUT_SPLIT use the same fid to indicate
2073                  * mirror deletion, so we'd zero cd_fid, and keeps o2 be NULL.
2074                  */
2075                 if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT))
2076                         RETURN(-EINVAL);
2077
2078                 /* zero cd_fid to keeps o2 be NULL */
2079                 fid_zero(&data->cd_fid);
2080         } else if (rc < 0) {
2081                 /* Exchange o1 and o2, to enforce locking order */
2082                 swap_objects = true;
2083         }
2084
2085         lease = ldlm_handle2lock(&data->cd_handle);
2086         if (lease == NULL)
2087                 RETURN(-ESTALE);
2088
2089         if (!fid_is_zero(&data->cd_fid)) {
2090                 o2 = mdt_object_find(info->mti_env, info->mti_mdt,
2091                                      &data->cd_fid);
2092                 if (IS_ERR(o2))
2093                         GOTO(out_lease, rc = PTR_ERR(o2));
2094
2095                 if (!mdt_object_exists(o2))
2096                         GOTO(out_obj, rc = -ENOENT);
2097
2098                 if (!S_ISREG(lu_object_attr(&o2->mot_obj)))
2099                         GOTO(out_obj, rc = -EINVAL);
2100
2101                 if (swap_objects)
2102                         swap(o1, o2);
2103         }
2104
2105         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
2106                            MAY_WRITE);
2107         if (rc < 0)
2108                 GOTO(out_obj, rc);
2109
2110         if (o2) {
2111                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2),
2112                                    NULL, MAY_WRITE);
2113                 if (rc < 0)
2114                         GOTO(out_obj, rc);
2115         }
2116
2117         /* try to hold open_sem so that nobody else can open the file */
2118         if (!down_write_trylock(&o->mot_open_sem)) {
2119                 ldlm_lock_cancel(lease);
2120                 GOTO(out_obj, rc = -EBUSY);
2121         }
2122
2123         /* Check if the lease open lease has already canceled */
2124         lock_res_and_lock(lease);
2125         lease_broken = ldlm_is_cancel(lease);
2126         unlock_res_and_lock(lease);
2127
2128         LDLM_DEBUG(lease, DFID " lease broken? %d",
2129                    PFID(mdt_object_fid(o)), lease_broken);
2130
2131         /* Cancel server side lease. Client side counterpart should
2132          * have been cancelled. It's okay to cancel it now as we've
2133          * held mot_open_sem. */
2134         ldlm_lock_cancel(lease);
2135
2136         if (lease_broken)
2137                 GOTO(out_unlock_sem, rc = -ESTALE);
2138
2139         mdt_lock_reg_init(lh1, LCK_EX);
2140         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
2141                              MDS_INODELOCK_XATTR);
2142         if (rc < 0)
2143                 GOTO(out_unlock_sem, rc);
2144
2145         if (o2) {
2146                 mdt_lock_reg_init(lh2, LCK_EX);
2147                 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2148                                      MDS_INODELOCK_XATTR);
2149                 if (rc < 0)
2150                         GOTO(out_unlock1, rc);
2151         }
2152
2153         /* Swap layout with orphan object */
2154         if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2155                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2156                                      mdt_object_child(o2), 0);
2157         } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2158                    ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2159                 struct lu_buf *buf = &info->mti_buf;
2160                 struct md_rejig_data mrd;
2161
2162                 if (o2) {
2163                         mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2164                 } else {
2165                         if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)) {
2166                                 /* paranoid check again */
2167                                 CERROR(DFID
2168                                   ":only mirror split support NULL o2 object\n",
2169                                         PFID(mdt_object_fid(o)));
2170                                 GOTO(out_unlock1, rc = -EINVAL);
2171                         }
2172
2173                         /* set NULL mrd_obj for deleting mirror objects */
2174                         mrd.mrd_obj = NULL;
2175                 }
2176
2177                 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2178                         mrd.mrd_mirror_id = data->cd_mirror_id;
2179                         /* set a small enough blocks in the SoM */
2180                         ma->ma_attr.la_blocks >>= 1;
2181                 }
2182
2183                 buf->lb_len = sizeof(mrd);
2184                 buf->lb_buf = &mrd;
2185                 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2186                                   XATTR_LUSTRE_LOV,
2187                                   ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2188                                   LU_XATTR_SPLIT : LU_XATTR_MERGE);
2189                 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS |
2190                                                        LA_LSIZE | LA_LBLOCKS)) {
2191                         int rc2;
2192                         enum lustre_som_flags lsf;
2193
2194                         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS))
2195                                 lsf = SOM_FL_STRICT;
2196                         else
2197                                 lsf = SOM_FL_LAZY;
2198
2199                         mutex_lock(&o->mot_som_mutex);
2200                         rc2 = mdt_set_som(info, o, lsf,
2201                                           ma->ma_attr.la_size,
2202                                           ma->ma_attr.la_blocks);
2203                         mutex_unlock(&o->mot_som_mutex);
2204                         if (rc2 < 0)
2205                                 CERROR(DFID": Setting i_blocks error: %d, "
2206                                        "i_blocks will be reported wrongly and "
2207                                        "can only be fixed in next resync\n",
2208                                        PFID(mdt_object_fid(o)), rc2);
2209                 }
2210         }
2211         if (rc < 0)
2212                 GOTO(out_unlock2, rc);
2213
2214         EXIT;
2215
2216 out_unlock2:
2217         /* Release exclusive LL */
2218         if (o2)
2219                 mdt_object_unlock(info, o2, lh2, 1);
2220
2221 out_unlock1:
2222         mdt_object_unlock(info, o1, lh1, 1);
2223
2224 out_unlock_sem:
2225         up_write(&o->mot_open_sem);
2226
2227         /* already swapped */
2228         if (rc == 0) {
2229                 struct mdt_body *repbody;
2230
2231                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2232                 LASSERT(repbody != NULL);
2233                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2234         }
2235
2236 out_obj:
2237         if (o1 != o)
2238                 /* the 2nd object has been used, and swapped to o1 */
2239                 mdt_object_put(info->mti_env, o1);
2240         else if (o2)
2241                 /* the 2nd object has been used, and not swapped */
2242                 mdt_object_put(info->mti_env, o2);
2243
2244         ldlm_reprocess_all(lease->l_resource,
2245                            lease->l_policy_data.l_inodebits.bits);
2246
2247 out_lease:
2248         LDLM_LOCK_PUT(lease);
2249
2250         if (ma != NULL) {
2251                 ma->ma_valid = 0;
2252                 ma->ma_need = 0;
2253         }
2254
2255         return rc;
2256 }
2257
2258 static int mdt_close_resync_done(struct mdt_thread_info *info,
2259                                  struct mdt_object *o, struct md_attr *ma)
2260 {
2261         struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LOCAL];
2262         struct close_data *data;
2263         struct ldlm_lock *lease;
2264         struct md_layout_change layout = { 0 };
2265         __u32 *resync_ids = NULL;
2266         size_t resync_count = 0;
2267         bool lease_broken;
2268         int rc;
2269
2270         ENTRY;
2271
2272         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2273                 RETURN(-EROFS);
2274
2275         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2276                 RETURN(-EINVAL);
2277
2278         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2279         if (data == NULL)
2280                 RETURN(-EPROTO);
2281
2282         if (req_capsule_req_need_swab(info->mti_pill))
2283                 lustre_swab_close_data_resync_done(&data->cd_resync);
2284
2285         if (!fid_is_zero(&data->cd_fid))
2286                 RETURN(-EPROTO);
2287
2288         lease = ldlm_handle2lock(&data->cd_handle);
2289         if (lease == NULL)
2290                 RETURN(-ESTALE);
2291
2292         /* try to hold open_sem so that nobody else can open the file */
2293         if (!down_write_trylock(&o->mot_open_sem)) {
2294                 ldlm_lock_cancel(lease);
2295                 GOTO(out_reprocess, rc = -EBUSY);
2296         }
2297
2298         /* Check if the lease open lease has already canceled */
2299         lock_res_and_lock(lease);
2300         lease_broken = ldlm_is_cancel(lease);
2301         unlock_res_and_lock(lease);
2302
2303         LDLM_DEBUG(lease, DFID " lease broken? %d\n",
2304                    PFID(mdt_object_fid(o)), lease_broken);
2305
2306         /* Cancel server side lease. Client side counterpart should
2307          * have been cancelled. It's okay to cancel it now as we've
2308          * held mot_open_sem. */
2309         ldlm_lock_cancel(lease);
2310
2311         if (lease_broken) /* don't perform release task */
2312                 GOTO(out_unlock, rc = -ESTALE);
2313
2314         resync_count = data->cd_resync.resync_count;
2315
2316         if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2317                 void *data;
2318
2319                 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2320                                            RCL_CLIENT))
2321                         GOTO(out_unlock, rc = -EPROTO);
2322
2323                 OBD_ALLOC_PTR_ARRAY(resync_ids, resync_count);
2324                 if (!resync_ids)
2325                         GOTO(out_unlock, rc = -ENOMEM);
2326
2327                 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2328                 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2329
2330                 layout.mlc_resync_ids = resync_ids;
2331         } else {
2332                 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2333         }
2334
2335         layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2336         layout.mlc_resync_count = resync_count;
2337         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2338                 layout.mlc_som.lsa_valid = SOM_FL_STRICT;
2339                 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2340                 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2341         }
2342         rc = mdt_layout_change(info, o, lhc, &layout);
2343         if (rc)
2344                 GOTO(out_unlock, rc);
2345
2346         mdt_object_unlock(info, o, lhc, 0);
2347
2348         EXIT;
2349
2350 out_unlock:
2351         up_write(&o->mot_open_sem);
2352
2353         /* already released */
2354         if (rc == 0) {
2355                 struct mdt_body *repbody;
2356
2357                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2358                 LASSERT(repbody != NULL);
2359                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2360         }
2361
2362         if (resync_ids)
2363                 OBD_FREE_PTR_ARRAY(resync_ids, resync_count);
2364
2365 out_reprocess:
2366         ldlm_reprocess_all(lease->l_resource,
2367                            lease->l_policy_data.l_inodebits.bits);
2368         LDLM_LOCK_PUT(lease);
2369
2370         ma->ma_valid = 0;
2371         ma->ma_need = 0;
2372
2373         return rc;
2374 }
2375
2376 #define MFD_CLOSED(open_flags) ((open_flags) == MDS_FMODE_CLOSED)
2377
2378 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2379 {
2380         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_open_flags));
2381 }
2382
2383 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2384 {
2385         struct mdt_object *o = mfd->mfd_object;
2386         struct md_object *next = mdt_object_child(o);
2387         struct md_attr *ma = &info->mti_attr;
2388         struct lu_fid *ofid = &info->mti_tmp_fid1;
2389         int rc = 0;
2390         u64 open_flags;
2391         u64 intent;
2392
2393         ENTRY;
2394
2395         open_flags = mfd->mfd_open_flags;
2396         intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2397         *ofid = *mdt_object_fid(o);
2398
2399         /* the below message is checked in replay-single.sh test_46 */
2400         CDEBUG(D_INODE, "%s: %sclosing file handle "DFID" with intent: %llx\n",
2401                mdt_obd_name(info->mti_mdt),
2402                ma->ma_valid & MA_FORCE_LOG ? "force " : "", PFID(ofid), intent);
2403
2404         switch (intent) {
2405         case MDS_HSM_RELEASE: {
2406                 rc = mdt_hsm_release(info, o, ma);
2407                 if (rc < 0) {
2408                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2409                                mdt_obd_name(info->mti_mdt),
2410                                PFID(ofid), rc);
2411                         /* continue to close even error occurred. */
2412                 }
2413                 break;
2414         }
2415         case MDS_CLOSE_LAYOUT_MERGE:
2416         case MDS_CLOSE_LAYOUT_SPLIT:
2417         case MDS_CLOSE_LAYOUT_SWAP: {
2418                 rc = mdt_close_handle_layouts(info, o, ma);
2419                 if (rc < 0) {
2420                         CDEBUG(D_INODE,
2421                                "%s: cannot swap layout of "DFID": rc = %d\n",
2422                                mdt_obd_name(info->mti_mdt),
2423                                PFID(ofid), rc);
2424                         /* continue to close even if error occurred. */
2425                 }
2426                 break;
2427         }
2428         case MDS_CLOSE_RESYNC_DONE:
2429                 rc = mdt_close_resync_done(info, o, ma);
2430                 break;
2431         default:
2432                 /* nothing */
2433                 break;
2434         }
2435
2436         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
2437             ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) {
2438                 int rc2;
2439
2440                 rc2 = mdt_lsom_update(info, o, false);
2441                 if (rc2 < 0)
2442                         CDEBUG(D_INODE,
2443                                "%s: File " DFID " LSOM failed: rc = %d\n",
2444                                mdt_obd_name(info->mti_mdt),
2445                                PFID(ofid), rc2);
2446                         /* continue to close even if error occured. */
2447         }
2448
2449         if (open_flags & MDS_FMODE_WRITE)
2450                 mdt_write_put(o);
2451         else if (open_flags & MDS_FMODE_EXEC)
2452                 mdt_write_allow(o);
2453
2454         /* Update atime|mtime|ctime on close. */
2455         if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ ||
2456              open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
2457             (ma->ma_attr.la_valid & LA_ATIME ||
2458              ma->ma_attr.la_valid & LA_MTIME ||
2459              ma->ma_attr.la_valid & LA_CTIME)) {
2460                 ma->ma_valid = MA_INODE;
2461                 ma->ma_attr_flags |= MDS_CLOSE_UPDATE_TIMES;
2462                 ma->ma_attr.la_valid &= (LA_ATIME | LA_MTIME | LA_CTIME);
2463
2464                 if (ma->ma_attr.la_valid & LA_MTIME) {
2465                         rc = mdt_attr_get_pfid(info, o, &ma->ma_pfid);
2466                         if (!rc)
2467                                 ma->ma_valid |= MA_PFID;
2468                 }
2469
2470                 rc = mo_attr_set(info->mti_env, next, ma);
2471         }
2472
2473         /* If file data is modified, add the dirty flag. */
2474         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
2475                 rc = mdt_add_dirty_flag(info, o, ma);
2476
2477         ma->ma_need |= MA_INODE;
2478         ma->ma_valid &= ~MA_INODE;
2479
2480         LASSERT(atomic_read(&o->mot_open_count) > 0);
2481         atomic_dec(&o->mot_open_count);
2482         mdt_handle_last_unlink(info, o, ma);
2483
2484         if (!MFD_CLOSED(open_flags)) {
2485                 rc = mo_close(info->mti_env, next, ma, open_flags);
2486                 if (mdt_dom_check_for_discard(info, o))
2487                         mdt_dom_discard_data(info, o);
2488         }
2489
2490         /* adjust open and lease count */
2491         if (open_flags & MDS_OPEN_LEASE) {
2492                 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2493                 atomic_dec(&o->mot_lease_count);
2494         }
2495
2496         mdt_mfd_free(mfd);
2497         mdt_object_put(info->mti_env, o);
2498
2499         RETURN(rc);
2500 }
2501
2502 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2503                        struct mdt_body *repbody)
2504 {
2505         struct mdt_export_data *med;
2506         struct mdt_file_data   *mfd;
2507         int                     ret = 0;
2508         int                     rc = 0;
2509         ENTRY;
2510
2511         med = &req->rq_export->exp_mdt_data;
2512         spin_lock(&med->med_open_lock);
2513         mfd = mdt_open_handle2mfd(med, &info->mti_open_handle,
2514                                   req_is_replay(req));
2515         if (mdt_mfd_closed(mfd)) {
2516                 spin_unlock(&med->med_open_lock);
2517                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2518                        ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2519                        info->mti_open_handle.cookie);
2520                 /** not serious error since bug 3633 */
2521                 rc = -ESTALE;
2522         } else {
2523                 class_handle_unhash(&mfd->mfd_open_handle);
2524                 list_del_init(&mfd->mfd_list);
2525                 spin_unlock(&med->med_open_lock);
2526                 ret = mdt_mfd_close(info, mfd);
2527         }
2528
2529         RETURN(rc ? rc : ret);
2530 }
2531
2532 int mdt_close(struct tgt_session_info *tsi)
2533 {
2534         struct mdt_thread_info  *info = tsi2mdt_info(tsi);
2535         struct ptlrpc_request   *req = tgt_ses_req(tsi);
2536         struct md_attr         *ma = &info->mti_attr;
2537         struct mdt_body        *repbody = NULL;
2538         ktime_t                 kstart = ktime_get();
2539         int rc, ret = 0;
2540         ENTRY;
2541
2542         /* Close may come with the Size-on-MDS update. Unpack it. */
2543         rc = mdt_close_unpack(info);
2544         if (rc)
2545                 GOTO(out, rc = err_serious(rc));
2546
2547         /* These fields are no longer used and are left for compatibility.
2548          * size is always zero */
2549         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2550                              0);
2551         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2552                              0);
2553         rc = req_capsule_server_pack(info->mti_pill);
2554         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2555                 mdt_client_compatibility(info);
2556                 if (rc == 0)
2557                         mdt_fix_reply(info);
2558                 mdt_exit_ucred(info);
2559                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2560         }
2561
2562         /* Continue to close handle even if we can not pack reply */
2563         if (rc == 0) {
2564                 repbody = req_capsule_server_get(info->mti_pill,
2565                                                  &RMF_MDT_BODY);
2566                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2567                                                     &RMF_MDT_MD);
2568                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2569                                                        &RMF_MDT_MD,
2570                                                        RCL_SERVER);
2571                 ma->ma_need = MA_INODE | MA_LOV;
2572                 repbody->mbo_eadatasize = 0;
2573                 repbody->mbo_aclsize = 0;
2574         } else {
2575                 rc = err_serious(rc);
2576         }
2577
2578         rc = mdt_close_internal(info, req, repbody);
2579         if (rc != -ESTALE)
2580                 mdt_empty_transno(info, rc);
2581
2582         if (repbody != NULL) {
2583                 mdt_client_compatibility(info);
2584                 rc = mdt_fix_reply(info);
2585         }
2586
2587         mdt_exit_ucred(info);
2588         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2589                 GOTO(out, rc = err_serious(-ENOMEM));
2590
2591         if (OBD_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2592                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2593                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2594 out:
2595         mdt_thread_info_fini(info);
2596         if (rc == 0)
2597                 mdt_counter_incr(req, LPROC_MDT_CLOSE,
2598                                  ktime_us_delta(ktime_get(), kstart));
2599         RETURN(rc ? rc : ret);
2600 }