lustre/mdt/mdt_open.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lustre/mdt/mdt_open.c
  32  *
  33  * Lustre Metadata Target (mdt) open/close file handling
  34  *
  35  * Author: Huang Hua <huanghua@clusterfs.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_MDS
  39
  40 #include <lustre_acl.h>
  41 #include <lustre_mds.h>
  42 #include <lustre_swab.h>
  43 #include "mdt_internal.h"
  44 #include <lustre_nodemap.h>
  45
  46 static const char mfd_open_handle_owner[] = "mdt";
  47
  48 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep,
  49                            struct mdt_lock_handle *lhc);
  50
  51 /* Create a new mdt_file_data struct, initialize it,
  52  * and insert it to global hash table */
  53 struct mdt_file_data *mdt_mfd_new(const struct mdt_export_data *med)
  54 {
  55         struct mdt_file_data *mfd;
  56
  57         ENTRY;
  58         OBD_ALLOC_PTR(mfd);
  59         if (mfd != NULL) {
  60                 refcount_set(&mfd->mfd_open_handle.h_ref, 1);
  61                 INIT_HLIST_NODE(&mfd->mfd_open_handle.h_link);
  62                 mfd->mfd_owner = med;
  63                 INIT_LIST_HEAD(&mfd->mfd_list);
  64                 class_handle_hash(&mfd->mfd_open_handle, mfd_open_handle_owner);
  65         }
  66
  67         RETURN(mfd);
  68 }
  69
  70 /*
  71  * Find the mfd pointed to by handle in global hash table.
  72  * In case of replay the handle is obsoleted
  73  * but mfd can be found in mfd list by that handle.
  74  * Callers need to be holding med_open_lock.
  75  */
  76 struct mdt_file_data *mdt_open_handle2mfd(struct mdt_export_data *med,
  77                                         const struct lustre_handle *open_handle,
  78                                         bool is_replay_or_resent)
  79 {
  80         struct mdt_file_data   *mfd;
  81
  82         ENTRY;
  83         LASSERT(open_handle != NULL);
  84         mfd = class_handle2object(open_handle->cookie, mfd_open_handle_owner);
  85         if (mfd)
  86                 refcount_dec(&mfd->mfd_open_handle.h_ref);
  87
  88         /* during dw/setattr replay the mfd can be found by old handle */
  89         if ((!mfd || mfd->mfd_owner != med) && is_replay_or_resent) {
  90                 list_for_each_entry(mfd, &med->med_open_head, mfd_list) {
  91                         if (mfd->mfd_open_handle_old.cookie ==
  92                             open_handle->cookie)
  93                                 RETURN(mfd);
  94                 }
  95                 mfd = NULL;
  96         }
  97
  98         RETURN(mfd);
  99 }
 100
 101 /* free mfd */
 102 void mdt_mfd_free(struct mdt_file_data *mfd)
 103 {
 104         LASSERT(refcount_read(&mfd->mfd_open_handle.h_ref) == 1);
 105         LASSERT(list_empty(&mfd->mfd_list));
 106         OBD_FREE_PRE(mfd, sizeof(*mfd), "kfree_rcu");
 107         kfree_rcu(mfd, mfd_open_handle.h_rcu);
 108 }
 109
 110 static int mdt_create_data(struct mdt_thread_info *info,
 111                            struct mdt_object *p, struct mdt_object *o)
 112 {
 113         struct md_op_spec *spec = &info->mti_spec;
 114         struct md_attr *ma = &info->mti_attr;
 115         int rc = 0;
 116
 117         ENTRY;
 118         if (!md_should_create(spec->sp_cr_flags))
 119                 RETURN(0);
 120
 121         ma->ma_need = MA_INODE | MA_LOV;
 122         ma->ma_valid = 0;
 123         mutex_lock(&o->mot_lov_mutex);
 124         if (!o->mot_lov_created) {
 125                 rc = mdo_create_data(info->mti_env,
 126                                      p ? mdt_object_child(p) : NULL,
 127                                      mdt_object_child(o), spec, ma);
 128                 if (rc == 0)
 129                         rc = mdt_attr_get_complex(info, o, ma);
 130
 131                 if (rc == 0 && ma->ma_valid & MA_LOV)
 132                         o->mot_lov_created = 1;
 133         }
 134
 135         mutex_unlock(&o->mot_lov_mutex);
 136         RETURN(rc);
 137 }
 138
 139 int mdt_write_read(struct mdt_object *o)
 140 {
 141         int rc = 0;
 142
 143         ENTRY;
 144         spin_lock(&o->mot_write_lock);
 145         rc = o->mot_write_count;
 146         spin_unlock(&o->mot_write_lock);
 147         RETURN(rc);
 148 }
 149
 150 int mdt_write_get(struct mdt_object *o)
 151 {
 152         int rc = 0;
 153
 154         ENTRY;
 155         spin_lock(&o->mot_write_lock);
 156         if (o->mot_write_count < 0)
 157                 rc = -ETXTBSY;
 158         else
 159                 o->mot_write_count++;
 160         spin_unlock(&o->mot_write_lock);
 161
 162         RETURN(rc);
 163 }
 164
 165 void mdt_write_put(struct mdt_object *o)
 166 {
 167         ENTRY;
 168         spin_lock(&o->mot_write_lock);
 169         o->mot_write_count--;
 170         spin_unlock(&o->mot_write_lock);
 171         EXIT;
 172 }
 173
 174 static int mdt_write_deny(struct mdt_object *o)
 175 {
 176         int rc = 0;
 177
 178         ENTRY;
 179         spin_lock(&o->mot_write_lock);
 180         if (o->mot_write_count > 0)
 181                 rc = -ETXTBSY;
 182         else
 183                 o->mot_write_count--;
 184         spin_unlock(&o->mot_write_lock);
 185         RETURN(rc);
 186 }
 187
 188 static void mdt_write_allow(struct mdt_object *o)
 189 {
 190         ENTRY;
 191         spin_lock(&o->mot_write_lock);
 192         o->mot_write_count++;
 193         spin_unlock(&o->mot_write_lock);
 194         EXIT;
 195 }
 196
 197 /* there can be no real transaction so prepare the fake one */
 198 static void mdt_empty_transno(struct mdt_thread_info *info, int rc)
 199 {
 200         struct mdt_device *mdt = info->mti_mdt;
 201         struct ptlrpc_request *req = mdt_info_req(info);
 202         struct tg_export_data *ted;
 203         struct lsd_client_data *lcd;
 204
 205         ENTRY;
 206         if (mdt_rdonly(req->rq_export))
 207                 RETURN_EXIT;
 208
 209         /* transaction has occurred already */
 210         if (lustre_msg_get_transno(req->rq_repmsg) != 0)
 211                 RETURN_EXIT;
 212
 213         if (tgt_is_multimodrpcs_client(req->rq_export)) {
 214                 struct thandle *th;
 215
 216                 /* generate an empty transaction to get a transno
 217                  * and reply data */
 218                 th = dt_trans_create(info->mti_env, mdt->mdt_bottom);
 219                 if (!IS_ERR(th)) {
 220                         rc = dt_trans_start(info->mti_env, mdt->mdt_bottom, th);
 221                         dt_trans_stop(info->mti_env, mdt->mdt_bottom, th);
 222                 }
 223                 RETURN_EXIT;
 224         }
 225
 226         spin_lock(&mdt->mdt_lut.lut_translock);
 227         if (rc != 0) {
 228                 if (info->mti_transno != 0) {
 229                         struct obd_export *exp = req->rq_export;
 230
 231                         CERROR("%s: replay trans %llu NID %s: rc = %d\n",
 232                                mdt_obd_name(mdt), info->mti_transno,
 233                                obd_export_nid2str(exp), rc);
 234                         spin_unlock(&mdt->mdt_lut.lut_translock);
 235                         RETURN_EXIT;
 236                 }
 237         } else if (info->mti_transno == 0) {
 238                 info->mti_transno = ++mdt->mdt_lut.lut_last_transno;
 239         } else {
 240                 /* should be replay */
 241                 if (info->mti_transno > mdt->mdt_lut.lut_last_transno)
 242                         mdt->mdt_lut.lut_last_transno = info->mti_transno;
 243         }
 244         spin_unlock(&mdt->mdt_lut.lut_translock);
 245
 246         CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n",
 247                info->mti_transno,
 248                req->rq_export->exp_obd->obd_last_committed);
 249
 250         req->rq_transno = info->mti_transno;
 251         lustre_msg_set_transno(req->rq_repmsg, info->mti_transno);
 252
 253         /* update lcd in memory only for resent cases */
 254         ted = &req->rq_export->exp_target_data;
 255         LASSERT(ted);
 256         mutex_lock(&ted->ted_lcd_lock);
 257         lcd = ted->ted_lcd;
 258         if (info->mti_transno < lcd->lcd_last_transno &&
 259             info->mti_transno != 0) {
 260                 /* This should happen during replay. Do not update
 261                  * last rcvd info if replay req transno < last transno,
 262                  * otherwise the following resend(after replay) can not
 263                  * be checked correctly by xid */
 264                 mutex_unlock(&ted->ted_lcd_lock);
 265                 CDEBUG(D_HA, "%s: transno = %llu < last_transno = %llu\n",
 266                        mdt_obd_name(mdt), info->mti_transno,
 267                        lcd->lcd_last_transno);
 268                 RETURN_EXIT;
 269         }
 270
 271         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
 272                 if (info->mti_transno != 0)
 273                         lcd->lcd_last_close_transno = info->mti_transno;
 274                 lcd->lcd_last_close_xid = req->rq_xid;
 275                 lcd->lcd_last_close_result = rc;
 276         } else {
 277                 /* VBR: save versions in last_rcvd for reconstruct. */
 278                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 279
 280                 if (pre_versions) {
 281                         lcd->lcd_pre_versions[0] = pre_versions[0];
 282                         lcd->lcd_pre_versions[1] = pre_versions[1];
 283                         lcd->lcd_pre_versions[2] = pre_versions[2];
 284                         lcd->lcd_pre_versions[3] = pre_versions[3];
 285                 }
 286                 if (info->mti_transno != 0)
 287                         lcd->lcd_last_transno = info->mti_transno;
 288
 289                 lcd->lcd_last_xid = req->rq_xid;
 290                 lcd->lcd_last_result = rc;
 291                 lcd->lcd_last_data = info->mti_opdata;
 292         }
 293         mutex_unlock(&ted->ted_lcd_lock);
 294
 295         EXIT;
 296 }
 297
 298 void mdt_mfd_set_mode(struct mdt_file_data *mfd, u64 open_flags)
 299 {
 300         LASSERT(mfd != NULL);
 301
 302         CDEBUG(D_DENTRY, DFID " Change mfd open_flags %#llo -> %#llo.\n",
 303                PFID(mdt_object_fid(mfd->mfd_object)), mfd->mfd_open_flags,
 304                open_flags);
 305
 306         mfd->mfd_open_flags = open_flags;
 307 }
 308
 309 /**
 310  * prep ma_lmm/ma_lmv for md_attr from reply
 311  */
 312 void mdt_prep_ma_buf_from_rep(struct mdt_thread_info *info,
 313                               struct mdt_object *obj, struct md_attr *ma,
 314                               __u64 open_flags)
 315 {
 316         struct req_capsule *pill;
 317
 318         if (ma->ma_lmv || ma->ma_lmm) {
 319                 CDEBUG(D_INFO, DFID " %s already set.\n",
 320                        PFID(mdt_object_fid(obj)),
 321                        ma->ma_lmv ? (ma->ma_lmm ? "ma_lmv and ma_lmm"
 322                                                 : "ma_lmv")
 323                                   : "ma_lmm");
 324                 return;
 325         }
 326
 327         pill = info->mti_pill;
 328         if (S_ISDIR(obj->mot_header.loh_attr)) {
 329                 ma->ma_lmv = req_capsule_server_get(pill, &RMF_MDT_MD);
 330                 ma->ma_lmv_size = req_capsule_get_size(pill, &RMF_MDT_MD,
 331                                                        RCL_SERVER);
 332                 if (ma->ma_lmv_size > 0)
 333                         ma->ma_need |= MA_LMV;
 334
 335                 if (open_flags & MDS_OPEN_DEFAULT_LMV) {
 336                         ma->ma_default_lmv = req_capsule_server_get(pill,
 337                                                         &RMF_DEFAULT_MDT_MD);
 338                         ma->ma_default_lmv_size = req_capsule_get_size(pill,
 339                                                         &RMF_DEFAULT_MDT_MD,
 340                                                         RCL_SERVER);
 341                         if (ma->ma_default_lmv_size > 0)
 342                                 ma->ma_need |= MA_LMV_DEF;
 343                 }
 344         } else {
 345                 ma->ma_lmm = req_capsule_server_get(pill, &RMF_MDT_MD);
 346                 ma->ma_lmm_size = req_capsule_get_size(pill, &RMF_MDT_MD,
 347                                                        RCL_SERVER);
 348                 if (ma->ma_lmm_size > 0)
 349                         ma->ma_need |= MA_LOV;
 350         }
 351 }
 352
 353 static int mdt_mfd_open(struct mdt_thread_info *info, struct mdt_object *p,
 354                         struct mdt_object *o, u64 open_flags, int created,
 355                         struct ldlm_reply *rep)
 356 {
 357         struct ptlrpc_request *req = mdt_info_req(info);
 358         struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 359         struct mdt_file_data *mfd;
 360         struct md_attr *ma  = &info->mti_attr;
 361         struct lu_attr *la  = &ma->ma_attr;
 362         struct mdt_body *repbody;
 363         bool isdir, isreg;
 364         int rc = 0;
 365
 366         ENTRY;
 367         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 368
 369         isreg = S_ISREG(la->la_mode);
 370         isdir = S_ISDIR(la->la_mode);
 371         if (isreg && !(ma->ma_valid & MA_LOV) &&
 372             !(open_flags & MDS_OPEN_RELEASE)) {
 373                 /*
 374                  * No EA, check whether it is will set regEA and dirEA since in
 375                  * above attr get, these size might be zero, so reset it, to
 376                  * retrieve the MD after create obj.
 377                  */
 378                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
 379                                                        &RMF_MDT_MD,
 380                                                        RCL_SERVER);
 381                 /* in replay case, p == NULL */
 382                 rc = mdt_create_data(info, p, o);
 383                 if (rc)
 384                         RETURN(rc);
 385
 386                 if (exp_connect_flags(req->rq_export) & OBD_CONNECT_DISP_STRIPE)
 387                         mdt_set_disposition(info, rep, DISP_OPEN_STRIPE);
 388         }
 389
 390         CDEBUG(D_INODE, "after open, ma_valid bit = %#llx lmm_size = %d\n",
 391                ma->ma_valid, ma->ma_lmm_size);
 392
 393         if (ma->ma_valid & MA_LOV) {
 394                 LASSERT(ma->ma_lmm_size != 0);
 395                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 396                 if (isdir)
 397                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 398                 else
 399                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 400         }
 401
 402         if (ma->ma_valid & MA_LMV) {
 403                 LASSERT(ma->ma_lmv_size != 0);
 404                 repbody->mbo_eadatasize = ma->ma_lmv_size;
 405                 LASSERT(isdir);
 406                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_MEA;
 407         }
 408
 409         if (ma->ma_valid & MA_LMV_DEF) {
 410                 LASSERT(ma->ma_default_lmv_size != 0);
 411                 LASSERT(isdir);
 412                 repbody->mbo_valid |= OBD_MD_FLDIREA | OBD_MD_DEFAULT_MEA;
 413         }
 414
 415         if (open_flags & MDS_FMODE_WRITE)
 416                 rc = mdt_write_get(o);
 417         else if (open_flags & MDS_FMODE_EXEC)
 418                 rc = mdt_write_deny(o);
 419
 420         if (rc)
 421                 RETURN(rc);
 422
 423         rc = mo_open(info->mti_env, mdt_object_child(o),
 424                      created ? open_flags | MDS_OPEN_CREATED : open_flags,
 425                      &info->mti_spec);
 426         if (rc != 0) {
 427                 /* If we allow the client to chgrp (CFS_SETGRP_PERM), but the
 428                  * client does not know which suppgid should be sent to the MDS,
 429                  * or some other(s) changed the target file's GID after this RPC
 430                  * sent to the MDS with the suppgid as the original GID, then we
 431                  * give the client another chance to send the right suppgid. */
 432                 if (rc == -EACCES &&
 433                     allow_client_chgrp(info, lu_ucred(info->mti_env)))
 434                         mdt_set_disposition(info, rep, DISP_OPEN_DENY);
 435
 436                 GOTO(err_out, rc);
 437         }
 438
 439         mfd = mdt_mfd_new(med);
 440         if (mfd == NULL)
 441                 GOTO(err_out, rc = -ENOMEM);
 442
 443         /*
 444          * Keep a reference on this object for this open, and is
 445          * released by mdt_mfd_close().
 446          */
 447         mdt_object_get(info->mti_env, o);
 448         mfd->mfd_object = o;
 449         mfd->mfd_xid = req->rq_xid;
 450
 451         /*
 452          * @open_flags is always not zero. At least it should be FMODE_READ,
 453          * FMODE_WRITE or MDS_FMODE_EXEC.
 454          */
 455         LASSERT(open_flags != 0);
 456
 457         /* Open handling. */
 458         mdt_mfd_set_mode(mfd, open_flags);
 459
 460         atomic_inc(&o->mot_open_count);
 461         if (open_flags & MDS_OPEN_LEASE)
 462                 atomic_inc(&o->mot_lease_count);
 463
 464         /* replay handle */
 465         if (req_is_replay(req)) {
 466                 struct mdt_file_data *old_mfd;
 467                 /* Check wheather old cookie already exist in
 468                  * the list, becasue when do recovery, client
 469                  * might be disconnected from server, and
 470                  * restart replay, so there maybe some orphan
 471                  * mfd here, we should remove them */
 472                 LASSERT(info->mti_rr.rr_open_handle != NULL);
 473                 spin_lock(&med->med_open_lock);
 474                 old_mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 475                                               true);
 476                 if (old_mfd != NULL) {
 477                         CDEBUG(D_HA, "delete orphan mfd = %p, fid = "DFID", "
 478                                "cookie = %#llx\n", mfd,
 479                                PFID(mdt_object_fid(mfd->mfd_object)),
 480                                info->mti_rr.rr_open_handle->cookie);
 481                         class_handle_unhash(&old_mfd->mfd_open_handle);
 482                         list_del_init(&old_mfd->mfd_list);
 483                         spin_unlock(&med->med_open_lock);
 484                         /* no attr update for that close */
 485                         la->la_valid = 0;
 486                         ma->ma_valid |= MA_FLAGS;
 487                         ma->ma_attr_flags |= MDS_RECOV_OPEN;
 488                         mdt_mfd_close(info, old_mfd);
 489                         ma->ma_attr_flags &= ~MDS_RECOV_OPEN;
 490                         ma->ma_valid &= ~MA_FLAGS;
 491                 } else {
 492                         spin_unlock(&med->med_open_lock);
 493                         CDEBUG(D_HA, "orphan mfd not found, fid = "DFID", "
 494                                "cookie = %#llx\n",
 495                                PFID(mdt_object_fid(mfd->mfd_object)),
 496                                info->mti_rr.rr_open_handle->cookie);
 497                 }
 498
 499                 CDEBUG(D_HA, "Store old cookie %#llx in new mfd\n",
 500                        info->mti_rr.rr_open_handle->cookie);
 501
 502                 mfd->mfd_open_handle_old = *info->mti_rr.rr_open_handle;
 503         }
 504
 505         repbody->mbo_open_handle.cookie = mfd->mfd_open_handle.h_cookie;
 506
 507         if (req->rq_export->exp_disconnected) {
 508                 spin_lock(&med->med_open_lock);
 509                 class_handle_unhash(&mfd->mfd_open_handle);
 510                 list_del_init(&mfd->mfd_list);
 511                 spin_unlock(&med->med_open_lock);
 512                 mdt_mfd_close(info, mfd);
 513         } else {
 514                 spin_lock(&med->med_open_lock);
 515                 list_add_tail(&mfd->mfd_list, &med->med_open_head);
 516                 spin_unlock(&med->med_open_lock);
 517         }
 518
 519         mdt_empty_transno(info, rc);
 520
 521         RETURN(rc);
 522
 523 err_out:
 524         if (open_flags & MDS_FMODE_WRITE)
 525                 mdt_write_put(o);
 526         else if (open_flags & MDS_FMODE_EXEC)
 527                 mdt_write_allow(o);
 528
 529         return rc;
 530 }
 531
 532 static int mdt_finish_open(struct mdt_thread_info *info,
 533                            struct mdt_object *p, struct mdt_object *o,
 534                            u64 open_flags,
 535                            struct ldlm_reply *rep)
 536 {
 537         struct ptlrpc_request *req = mdt_info_req(info);
 538         struct obd_export *exp = req->rq_export;
 539         struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 540         struct md_attr *ma  = &info->mti_attr;
 541         struct lu_attr *la  = &ma->ma_attr;
 542         struct mdt_file_data *mfd;
 543         struct mdt_body *repbody;
 544         int created;
 545         int rc = 0;
 546         int isreg, isdir, islnk;
 547         struct list_head *t;
 548
 549         ENTRY;
 550         LASSERT(ma->ma_valid & MA_INODE);
 551         created = mdt_get_disposition(rep, DISP_OPEN_CREATE);
 552
 553         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 554
 555         isreg = S_ISREG(la->la_mode);
 556         isdir = S_ISDIR(la->la_mode);
 557         islnk = S_ISLNK(la->la_mode);
 558         mdt_pack_attr2body(info, repbody, la, mdt_object_fid(o));
 559
 560         /* compatibility check for 2.10 clients when it tries to open mirrored
 561          * files. 2.10 clients don't verify overlapping components so they
 562          * would read and write mirrored files just as if they were normal
 563          * PFL files, which will cause the problem that sycned mirrors actually
 564          * contain different data.
 565          * Older clients are not a concern here because they don't even
 566          * understand PFL layout. */
 567         if (isreg && !exp_connect_flr(exp) && ma->ma_valid & MA_LOV &&
 568             mdt_lmm_is_flr(ma->ma_lmm)) {
 569                 /* LU-10286: for simplicity clients who don't understand
 570                  * mirrored layout(with connect flag OBD_CONNECT2_FLR) won't
 571                  * be able to open mirrored files */
 572                 RETURN(-EOPNOTSUPP);
 573         }
 574
 575         /* Overstriped files can crash older clients */
 576         if (isreg && !exp_connect_overstriping(exp) &&
 577             mdt_lmm_is_overstriping(ma->ma_lmm))
 578                 RETURN(-EOPNOTSUPP);
 579
 580         /* LU-2275, simulate broken behaviour (esp. prevalent in
 581          * pre-2.4 servers where a very strange reply is sent on error
 582          * that looks like it was actually almost successful and a
 583          * failure at the same time.) */
 584         if (CFS_FAIL_CHECK(OBD_FAIL_MDS_NEGATIVE_POSITIVE)) {
 585                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN |
 586                                                DISP_LOOKUP_NEG |
 587                                                DISP_LOOKUP_POS);
 588
 589                 if (open_flags & MDS_OPEN_LOCK)
 590                         mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
 591
 592                 RETURN(-ENOENT);
 593         }
 594
 595 #ifdef CONFIG_LUSTRE_FS_POSIX_ACL
 596         if (exp_connect_flags(exp) & OBD_CONNECT_ACL) {
 597                 struct lu_nodemap *nodemap = nodemap_get_from_exp(exp);
 598                 if (IS_ERR(nodemap))
 599                         RETURN(PTR_ERR(nodemap));
 600
 601                 rc = mdt_pack_acl2body(info, repbody, o, nodemap);
 602                 nodemap_putref(nodemap);
 603                 if (rc)
 604                         RETURN(rc);
 605         }
 606 #endif
 607
 608         /*
 609          * If we are following a symlink, don't open; and do not return open
 610          * handle for special nodes as client required.
 611          */
 612         if (islnk || (!isreg && !isdir &&
 613             (exp_connect_flags(req->rq_export) & OBD_CONNECT_NODEVOH))) {
 614                 lustre_msg_set_transno(req->rq_repmsg, 0);
 615                 RETURN(0);
 616         }
 617
 618         /*
 619          * We need to return the existing object's fid back, so it is done here,
 620          * after preparing the reply.
 621          */
 622         if (!created && (open_flags & MDS_OPEN_EXCL) &&
 623             (open_flags & MDS_OPEN_CREAT))
 624                 RETURN(-EEXIST);
 625
 626         /* This can't be done earlier, we need to return reply body */
 627         if (isdir) {
 628                 if (open_flags & (MDS_OPEN_CREAT | MDS_FMODE_WRITE)) {
 629                         /* We are trying to create or write an existing dir. */
 630                         RETURN(-EISDIR);
 631                 }
 632         } else if (open_flags & MDS_OPEN_DIRECTORY)
 633                 RETURN(-ENOTDIR);
 634
 635         if (CFS_FAIL_CHECK_RESET(OBD_FAIL_MDS_OPEN_CREATE,
 636                                  OBD_FAIL_MDS_LDLM_REPLY_NET | CFS_FAIL_ONCE))
 637                 RETURN(-EAGAIN);
 638
 639         mfd = NULL;
 640         if (info->mti_rr.rr_flags & MRF_OPEN_RESEND) {
 641                 spin_lock(&med->med_open_lock);
 642                 list_for_each(t, &med->med_open_head) {
 643                         mfd = list_entry(t, struct mdt_file_data, mfd_list);
 644                         if (mfd->mfd_xid == req->rq_xid) {
 645                                 repbody->mbo_open_handle.cookie =
 646                                                 mfd->mfd_open_handle.h_cookie;
 647                                 break;
 648                         }
 649                         mfd = NULL;
 650                 }
 651                 spin_unlock(&med->med_open_lock);
 652
 653                 if (mfd != NULL) {
 654                         /* set repbody->ea_size for resent case */
 655                         if (ma->ma_valid & MA_LOV) {
 656                                 LASSERT(ma->ma_lmm_size != 0);
 657                                 repbody->mbo_eadatasize = ma->ma_lmm_size;
 658                                 if (isdir)
 659                                         repbody->mbo_valid |= OBD_MD_FLDIREA;
 660                                 else
 661                                         repbody->mbo_valid |= OBD_MD_FLEASIZE;
 662                         }
 663                         mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 664                         RETURN(0);
 665                 }
 666                 /* if we have a real resend (not a resend afrer failover), it
 667                  * means close is already happend, so lets return error
 668                  */
 669                 RETURN(-ESTALE);
 670         }
 671
 672         rc = mdt_mfd_open(info, p, o, open_flags, created, rep);
 673         if (!rc)
 674                 mdt_set_disposition(info, rep, DISP_OPEN_OPEN);
 675
 676         RETURN(rc);
 677 }
 678
 679 void mdt_reconstruct_open(struct mdt_thread_info *info,
 680                           struct mdt_lock_handle *lhc)
 681 {
 682         struct req_capsule *pill = info->mti_pill;
 683         struct ptlrpc_request *req = mdt_info_req(info);
 684         struct mdt_reint_record *rr = &info->mti_rr;
 685         struct md_attr *ma = &info->mti_attr;
 686         struct ldlm_reply *ldlm_rep;
 687         u64 opdata;
 688         int rc;
 689
 690         ENTRY;
 691         LASSERT(pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
 692         ldlm_rep = req_capsule_server_get(pill, &RMF_DLM_REP);
 693
 694         ma->ma_need = MA_INODE | MA_HSM;
 695         ma->ma_valid = 0;
 696         opdata = mdt_req_from_lrd(req, info->mti_reply_data);
 697         mdt_set_disposition(info, ldlm_rep, opdata);
 698
 699         CDEBUG(D_INODE, "This is reconstruct open: disp=%#llx, result=%d\n",
 700                ldlm_rep->lock_policy_res1, req->rq_status);
 701         if (req->rq_status)
 702                 /* We did not create successfully, return error to client. */
 703                 GOTO(out, rc = req->rq_status);
 704
 705         /* tg_reply_data is just memory only  structure, so any non zero fid
 706          * means a real resend not a resend after recovery which need to be
 707          * handled as regular open
 708          */
 709         if (likely(!fid_is_zero(&info->mti_reply_data->trd_object))) {
 710                 rr->rr_fid2 = &info->mti_reply_data->trd_object;
 711                 rr->rr_flags |= MRF_OPEN_RESEND;
 712                 rc = mdt_open_by_fid(info, ldlm_rep, lhc);
 713                 if (rc)
 714                         lustre_msg_set_transno(req->rq_repmsg, 0);
 715         } else {
 716                 /* We did not try to create, so we are a pure open */
 717                 rc = mdt_reint_open(info, lhc);
 718         }
 719         EXIT;
 720 out:
 721         req->rq_status = rc;
 722         lustre_msg_set_status(req->rq_repmsg, req->rq_status);
 723         LASSERT(ergo(rc < 0, lustre_msg_get_transno(req->rq_repmsg) == 0));
 724 }
 725
 726 static int mdt_open_by_fid(struct mdt_thread_info *info, struct ldlm_reply *rep,
 727                            struct mdt_lock_handle *lhc)
 728 {
 729         u64 open_flags = info->mti_spec.sp_cr_flags;
 730         struct mdt_reint_record *rr = &info->mti_rr;
 731         struct md_attr *ma = &info->mti_attr;
 732         struct mdt_object *o;
 733         int rc;
 734
 735         ENTRY;
 736         o = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
 737         if (IS_ERR(o))
 738                 RETURN(rc = PTR_ERR(o));
 739
 740         rc = mdt_check_enc(info, o);
 741         if (rc)
 742                 GOTO(out, rc);
 743
 744         if (unlikely(mdt_object_remote(o))) {
 745                 /* the child object was created on remote server */
 746                 struct mdt_body *repbody;
 747
 748                 mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 749                                                 DISP_LOOKUP_EXECD |
 750                                                 DISP_LOOKUP_POS));
 751                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 752                 repbody->mbo_fid1 = *rr->rr_fid2;
 753                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 754                 rc = 0;
 755         } else {
 756                 if (mdt_object_exists(o)) {
 757                         tgt_open_obj_set(info->mti_env, mdt_obj2dt(o));
 758                         mdt_set_disposition(info, rep, (DISP_IT_EXECD |
 759                                                         DISP_LOOKUP_EXECD |
 760                                                         DISP_LOOKUP_POS));
 761                         if ((open_flags & MDS_OPEN_EXCL) &&
 762                             (open_flags & MDS_OPEN_CREAT))
 763                                 mdt_set_disposition(info, rep,
 764                                                     DISP_OPEN_CREATE);
 765
 766                         mdt_prep_ma_buf_from_rep(info, o, ma, open_flags);
 767                         rc = mdt_attr_get_complex(info, o, ma);
 768                         if (rc)
 769                                 GOTO(out, rc);
 770                         rc = mdt_finish_open(info, NULL, o, open_flags, rep);
 771                         if (rc)
 772                                 GOTO(out, rc);
 773                         mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
 774                 } else {
 775                         rc = -ENOENT;
 776                 }
 777         }
 778
 779 out:
 780         mdt_object_put(info->mti_env, o);
 781         RETURN(rc);
 782 }
 783
 784 /* lock object for open */
 785 static int mdt_object_open_lock(struct mdt_thread_info *info,
 786                                 struct mdt_object *obj,
 787                                 struct mdt_lock_handle *lhc,
 788                                 __u64 *ibits)
 789 {
 790         struct md_attr *ma = &info->mti_attr;
 791         __u64 open_flags = info->mti_spec.sp_cr_flags;
 792         __u64 trybits = 0;
 793         enum ldlm_mode lm = LCK_PR;
 794         bool acq_lease = !!(open_flags & MDS_OPEN_LEASE);
 795         bool try_layout = false;
 796         bool create_layout = false;
 797         int rc = 0;
 798         __u32 dom_stripe = 0;
 799         unsigned int dom_only = 0;
 800         unsigned int dom_lock = 0;
 801
 802         ENTRY;
 803         *ibits = 0;
 804         if (req_is_replay(mdt_info_req(info)))
 805                 RETURN(0);
 806
 807         if (S_ISREG(lu_object_attr(&obj->mot_obj))) {
 808                 if (ma->ma_need & MA_LOV && !(ma->ma_valid & MA_LOV) &&
 809                     md_should_create(open_flags))
 810                         create_layout = true;
 811                 if (exp_connect_layout(info->mti_exp) && !create_layout &&
 812                     ma->ma_need & MA_LOV)
 813                         try_layout = true;
 814
 815                 /* DoM files can take IO lock at OPEN when it makes sense,
 816                  * check if file has DoM stripe and ask for lock if client
 817                  * no lock on that resource yet.
 818                  */
 819                 if (ma->ma_valid & MA_LOV && ma->ma_lmm != NULL)
 820                         dom_stripe = mdt_lmm_dom_entry_check(ma->ma_lmm,
 821                                                              &dom_only);
 822                 /* If only DOM stripe is being used then we can expect IO
 823                  * to it after OPEN and will return corresponding DOM ibit
 824                  * using default strategy from mdt_opts.mo_dom_lock.
 825                  * Otherwise trylock mode is used always and DOM ibit will
 826                  * be returned optionally.
 827                  */
 828                 if (dom_stripe &&
 829                     !mdt_dom_client_has_lock(info, mdt_object_fid(obj)))
 830                         dom_lock = !dom_only ? TRYLOCK_DOM_ON_OPEN :
 831                                    info->mti_mdt->mdt_opts.mo_dom_lock;
 832         }
 833
 834         if (acq_lease) {
 835                 /* lease open, acquire write mode of open sem */
 836                 down_write(&obj->mot_open_sem);
 837
 838                 /* Lease exists and ask for new lease */
 839                 if (atomic_read(&obj->mot_lease_count) > 0) {
 840                         /* only exclusive open is supported, so lease
 841                          * are conflicted to each other */
 842                         GOTO(out, rc = -EBUSY);
 843                 }
 844
 845                 /* Lease must be with open lock */
 846                 if (!(open_flags & MDS_OPEN_LOCK)) {
 847                         CERROR("%s: Request lease for file:"DFID ", but open lock "
 848                                "is missed, open_flags = %#llo : rc = %d\n",
 849                                mdt_obd_name(info->mti_mdt),
 850                                PFID(mdt_object_fid(obj)), open_flags, -EPROTO);
 851                         GOTO(out, rc = -EPROTO);
 852                 }
 853
 854                 /* should conflict with new opens for write/execute */
 855                 lm = LCK_PW;
 856                 *ibits = MDS_INODELOCK_OPEN;
 857
 858                 /* never grant LCK_EX layout lock to client */
 859                 try_layout = false;
 860         } else { /* normal open */
 861                 /* normal open holds read mode of open sem */
 862                 down_read(&obj->mot_open_sem);
 863
 864                 if (open_flags & MDS_OPEN_LOCK) {
 865                         if (open_flags & MDS_FMODE_WRITE)
 866                                 lm = LCK_CW;
 867                         else if (open_flags & MDS_FMODE_EXEC)
 868                                 lm = LCK_PR;
 869                         else
 870                                 lm = LCK_CR;
 871
 872                         *ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_OPEN;
 873                 } else if (atomic_read(&obj->mot_lease_count) > 0) {
 874                         if (open_flags & MDS_FMODE_WRITE)
 875                                 lm = LCK_CW;
 876                         else
 877                                 lm = LCK_CR;
 878
 879                         /* revoke lease */
 880                         *ibits = MDS_INODELOCK_OPEN;
 881                         try_layout = false;
 882
 883                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 884                 } else if (dom_lock) {
 885                         lm = (open_flags & MDS_FMODE_WRITE) ? LCK_PW : LCK_PR;
 886                         trybits |= MDS_INODELOCK_DOM | MDS_INODELOCK_LAYOUT;
 887                 }
 888
 889                 CDEBUG(D_INODE, "normal open:"DFID" lease count: %d, lm: %d\n",
 890                         PFID(mdt_object_fid(obj)),
 891                         atomic_read(&obj->mot_lease_count), lm);
 892         }
 893
 894         /* Return lookup lock to validate inode at the client side.
 895          * This is pretty important otherwise MDT will return layout
 896          * lock for each open.
 897          * Since trylocks are opportunistic now and we also have
 898          * ability to selectively cancel lock bits from clients on conflict,
 899          * we would also return other useful bits: PERM and UPDATE - both
 900          * necessary to ensure subsequent stats from the client would not
 901          * need to talk to us again. Historically speaking both permissions
 902          * and other inode data rarely changes after creation anyway, so
 903          * there should not be any downsides from doing it for normal
 904          * operations now.
 905          */
 906         if (!CFS_FAIL_CHECK(OBD_FAIL_MDS_NO_LL_OPEN) && try_layout) {
 907                 if (!(*ibits & MDS_INODELOCK_LOOKUP))
 908                         trybits |= MDS_INODELOCK_LOOKUP;
 909                 trybits |= MDS_INODELOCK_LAYOUT;
 910                 /* PERM|UPDATE ibits in CR mode don't make sense, since
 911                  * all modifications lock these ibits in PW mode.
 912                  * Don't lock UPDATE|PERM ibits upon PCC attach, otherwise
 913                  * PCC will be detached when these ibits are revoked.
 914                  */
 915                 if (lm != LCK_CR && !(open_flags & MDS_OPEN_PCC))
 916                         trybits |= MDS_INODELOCK_UPDATE | MDS_INODELOCK_PERM;
 917         }
 918
 919         if (*ibits | trybits)
 920                 rc = mdt_object_lock_try(info, obj, lhc, ibits, trybits, lm);
 921
 922         CDEBUG(D_INODE, "%s: Requested bits lock:"DFID ", ibits = %#llx/%#llx"
 923                ", open_flags = %#llo, try_layout = %d : rc = %d\n",
 924                mdt_obd_name(info->mti_mdt), PFID(mdt_object_fid(obj)),
 925                *ibits, trybits, open_flags, try_layout, rc);
 926
 927         /* will change layout, revoke layout locks by enqueuing EX lock. */
 928         if (rc == 0 && create_layout) {
 929                 struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LAYOUT];
 930
 931                 CDEBUG(D_INODE, "Will create layout, get EX layout lock:"DFID
 932                         ", open_flags = %#llo\n",
 933                         PFID(mdt_object_fid(obj)), open_flags);
 934
 935                 /* We cannot enqueue another lock for the same resource we
 936                  * already have a lock for, due to mechanics of waiting list
 937                  * iterating in ldlm, see LU-3601.
 938                  * As such we'll drop the open lock we just got above here,
 939                  * it's ok not to have this open lock as it's main purpose is to
 940                  * flush unused cached client open handles. */
 941                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
 942                         mdt_object_unlock(info, obj, lhc, 1);
 943
 944                 LASSERT(!try_layout);
 945                 rc = mdt_object_lock(info, obj, ll, MDS_INODELOCK_LAYOUT,
 946                                      LCK_EX);
 947
 948                 CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_LL_BLOCK, 2);
 949         }
 950
 951         /* Check if there is any other open handles after acquiring
 952          * open lock. At this point, caching open handles have been revoked
 953          * by open lock.
 954          * XXX: Now only exclusive open is supported. Need to check the
 955          * type of open for generic lease support. */
 956         if (rc == 0 && acq_lease) {
 957                 struct ptlrpc_request *req = mdt_info_req(info);
 958                 struct mdt_export_data *med = &req->rq_export->exp_mdt_data;
 959                 struct mdt_file_data *mfd;
 960                 bool is_replay_or_resent;
 961                 int open_count = 0;
 962
 963                 /* For lease: application can open a file and then apply lease,
 964                  * @handle contains original open handle in that case.
 965                  * In recovery, open REQ will be replayed and the lease REQ may
 966                  * be resent that means the open handle is already stale, so we
 967                  * need to fix it up here by finding new handle. */
 968                 is_replay_or_resent = req_is_replay(req) ||
 969                         lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT;
 970
 971                 /* if the request is _not_ a replay request, rr_open_handle
 972                  * may be used to hold an open file handle which is issuing the
 973                  * lease request, so that this openhandle doesn't count. */
 974                 mfd = mdt_open_handle2mfd(med, info->mti_rr.rr_open_handle,
 975                                           is_replay_or_resent);
 976                 if (mfd != NULL)
 977                         ++open_count;
 978
 979                 CDEBUG(D_INODE, "acq_lease "DFID": openers: %d, want: %d\n",
 980                         PFID(mdt_object_fid(obj)),
 981                         atomic_read(&obj->mot_open_count), open_count);
 982
 983                 if (atomic_read(&obj->mot_open_count) > open_count) {
 984                         /* fail if anyone *else* has opened file for write */
 985                         if (mdt_write_read(obj) > 1)
 986                                 GOTO(out, rc = -EBUSY);
 987                 }
 988         }
 989         GOTO(out, rc);
 990
 991 out:
 992         RETURN(rc);
 993 }
 994
 995 static void mdt_object_open_unlock(struct mdt_thread_info *info,
 996                                    struct mdt_object *obj,
 997                                    struct mdt_lock_handle *lhc,
 998                                    __u64 ibits, int rc)
 999 {
1000         __u64 open_flags = info->mti_spec.sp_cr_flags;
1001         struct mdt_lock_handle *ll = &info->mti_lh[MDT_LH_LOCAL];
1002
1003         ENTRY;
1004         if (req_is_replay(mdt_info_req(info)))
1005                 RETURN_EXIT;
1006
1007         /* Release local lock - the lock put in MDT_LH_LOCAL will never
1008          * return to client side. */
1009         if (lustre_handle_is_used(&ll->mlh_reg_lh))
1010                 mdt_object_unlock(info, obj, ll, 1);
1011
1012         ll = &info->mti_lh[MDT_LH_LAYOUT];
1013         /* Release local layout lock, layout was created */
1014         if (lustre_handle_is_used(&ll->mlh_reg_lh)) {
1015                 LASSERT(!(ibits & MDS_INODELOCK_LAYOUT));
1016                 mdt_object_unlock(info, obj, ll, 1);
1017         }
1018
1019         if (open_flags & MDS_OPEN_LEASE)
1020                 up_write(&obj->mot_open_sem);
1021         else
1022                 up_read(&obj->mot_open_sem);
1023
1024         /* Cross-ref case, the lock should be returned to the client */
1025         if (ibits == 0 || rc == -MDT_EREMOTE_OPEN)
1026                 RETURN_EXIT;
1027
1028         if (!(open_flags & MDS_OPEN_LOCK) && !(ibits & MDS_INODELOCK_LAYOUT) &&
1029             !(ibits & MDS_INODELOCK_DOM)) {
1030                 /* for the open request, the lock will only return to client
1031                  * if open or layout lock is granted. */
1032                 rc = 1;
1033         }
1034
1035         if (rc != 0 || !lustre_handle_is_used(&lhc->mlh_reg_lh)) {
1036                 struct ldlm_reply       *ldlm_rep;
1037
1038                 ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1039                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1040                 if (lustre_handle_is_used(&lhc->mlh_reg_lh))
1041                         mdt_object_unlock(info, obj, lhc, 1);
1042         }
1043         RETURN_EXIT;
1044 }
1045
1046 /**
1047  * Check release is permitted for the current HSM flags.
1048  */
1049 static bool mdt_hsm_release_allow(const struct md_attr *ma)
1050 {
1051         if (!(ma->ma_valid & MA_HSM))
1052                 return false;
1053
1054         if (ma->ma_hsm.mh_flags & (HS_DIRTY|HS_NORELEASE|HS_LOST))
1055                 return false;
1056
1057         if (!(ma->ma_hsm.mh_flags & HS_ARCHIVED))
1058                 return false;
1059
1060         return true;
1061 }
1062
1063 static int mdt_refetch_lovea(struct mdt_thread_info *info,
1064                              struct mdt_object *o, struct md_attr *ma,
1065                              u64 ibits)
1066 {
1067         struct mdt_body *repbody;
1068         int rc;
1069
1070         if ((ibits & MDS_INODELOCK_LAYOUT) == 0)
1071                 return 0;
1072         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
1073                 return 0;
1074
1075         if ((ma->ma_valid & MA_LOV) == 0)
1076                 return 0;
1077
1078         ma->ma_valid &= ~MA_LOV;
1079         info->mti_big_lmm_used = 0;
1080         ma->ma_lmm = req_capsule_server_get(info->mti_pill, &RMF_MDT_MD);
1081         ma->ma_lmm_size = req_capsule_get_size(info->mti_pill, &RMF_MDT_MD,
1082                                                RCL_SERVER);
1083         rc = __mdt_stripe_get(info, o, ma, XATTR_NAME_LOV);
1084         if (rc)
1085                 return rc;
1086
1087         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1088         repbody->mbo_eadatasize = ma->ma_lmm_size;
1089         return 0;
1090 }
1091
1092 static int mdt_open_by_fid_lock(struct mdt_thread_info *info,
1093                                 struct ldlm_reply *rep,
1094                                 struct mdt_lock_handle *lhc)
1095 {
1096         const struct lu_env *env = info->mti_env;
1097         struct mdt_device *mdt = info->mti_mdt;
1098         u64 open_flags = info->mti_spec.sp_cr_flags;
1099         struct mdt_reint_record *rr = &info->mti_rr;
1100         struct md_attr *ma = &info->mti_attr;
1101         struct mdt_object *parent = NULL;
1102         struct mdt_object *o;
1103         bool object_locked = false;
1104         u64 ibits = 0;
1105         int rc;
1106
1107         ENTRY;
1108         if (md_should_create(open_flags)) {
1109                 if (!lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
1110                         parent = mdt_object_find(env, mdt, rr->rr_fid1);
1111                         if (IS_ERR(parent)) {
1112                                 CDEBUG(D_INODE, "Fail to find parent "DFID
1113                                        " for anonymous created %ld, try to"
1114                                        " use server-side parent.\n",
1115                                        PFID(rr->rr_fid1), PTR_ERR(parent));
1116                                 parent = NULL;
1117                         }
1118                 }
1119                 if (parent == NULL)
1120                         ma->ma_need |= MA_PFID;
1121         }
1122
1123         o = mdt_object_find(env, mdt, rr->rr_fid2);
1124         if (IS_ERR(o))
1125                 GOTO(out_parent_put, rc = PTR_ERR(o));
1126
1127         if (mdt_object_remote(o)) {
1128                 CDEBUG(D_INFO, "%s: "DFID" is on remote MDT.\n",
1129                        mdt_obd_name(info->mti_mdt),
1130                        PFID(rr->rr_fid2));
1131                 GOTO(out, rc = -EREMOTE);
1132         } else if (!mdt_object_exists(o)) {
1133                 mdt_set_disposition(info, rep,
1134                                     DISP_IT_EXECD |
1135                                     DISP_LOOKUP_EXECD |
1136                                     DISP_LOOKUP_NEG);
1137                 GOTO(out, rc = -ENOENT);
1138         }
1139
1140         /* do not check enc for directory: always allow open */
1141         if (!S_ISDIR(lu_object_attr(&o->mot_obj))) {
1142                 rc = mdt_check_enc(info, o);
1143                 if (rc)
1144                         GOTO(out, rc);
1145         }
1146
1147         mdt_set_disposition(info, rep, (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1148
1149         mdt_prep_ma_buf_from_rep(info, o, ma, open_flags);
1150         if (open_flags & MDS_OPEN_RELEASE)
1151                 ma->ma_need |= MA_HSM;
1152         rc = mdt_attr_get_complex(info, o, ma);
1153         if (rc)
1154                 GOTO(out, rc);
1155
1156         /* We should not change file's existing LOV EA */
1157         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
1158             open_flags & MDS_OPEN_HAS_EA && ma->ma_valid & MA_LOV)
1159                 GOTO(out, rc = -EEXIST);
1160
1161         /* If a release request, check file open flags are fine and ask for an
1162          * exclusive open access. */
1163         if (open_flags & MDS_OPEN_RELEASE && !mdt_hsm_release_allow(ma))
1164                 GOTO(out, rc = -EPERM);
1165
1166         rc = mdt_check_resent_lock(info, o, lhc);
1167         if (rc < 0) {
1168                 GOTO(out, rc);
1169         } else if (rc > 0) {
1170                 rc = mdt_object_open_lock(info, o, lhc, &ibits);
1171                 object_locked = true;
1172                 if (rc)
1173                         GOTO(out_unlock, rc);
1174         }
1175
1176         if (ma->ma_valid & MA_PFID) {
1177                 parent = mdt_object_find(env, mdt, &ma->ma_pfid);
1178                 if (IS_ERR(parent)) {
1179                         CDEBUG(D_INODE, "Fail to find parent "DFID
1180                                " for anonymous created %ld, try to"
1181                                " use system default.\n",
1182                                PFID(&ma->ma_pfid), PTR_ERR(parent));
1183                         parent = NULL;
1184                 }
1185         }
1186
1187         tgt_open_obj_set(info->mti_env, mdt_obj2dt(o));
1188         rc = mdt_finish_open(info, parent, o, open_flags, rep);
1189         if (rc)
1190                 GOTO(out_unlock, rc);
1191
1192         mdt_set_disposition(info, rep, DISP_LOOKUP_POS);
1193         if (open_flags & MDS_OPEN_LOCK)
1194                 mdt_set_disposition(info, rep, DISP_OPEN_LOCK);
1195         if (open_flags & MDS_OPEN_LEASE)
1196                 mdt_set_disposition(info, rep, DISP_OPEN_LEASE);
1197
1198         /*
1199          * if layout lock is granted, then we should re-fetch LOVEA
1200          * which was originally taken w/o the lock
1201          */
1202         rc = mdt_refetch_lovea(info, o, ma, ibits);
1203
1204         GOTO(out_unlock, rc);
1205
1206 out_unlock:
1207         if (object_locked)
1208                 mdt_object_open_unlock(info, o, lhc, ibits, rc);
1209 out:
1210         mdt_object_put(env, o);
1211         if (rc == 0)
1212                 mdt_pack_size2body(info, rr->rr_fid2, &lhc->mlh_reg_lh);
1213 out_parent_put:
1214         if (parent != NULL)
1215                 mdt_object_put(env, parent);
1216         return rc;
1217 }
1218
1219 /* Cross-ref request. Currently it can only be a pure open (w/o create) */
1220 static int mdt_cross_open(struct mdt_thread_info *info,
1221                           const struct lu_fid *parent_fid,
1222                           const struct lu_fid *fid,
1223                           struct ldlm_reply *rep, u64 open_flags)
1224 {
1225         struct md_attr *ma = &info->mti_attr;
1226         struct mdt_object *o;
1227         int rc;
1228
1229         ENTRY;
1230         o = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1231         if (IS_ERR(o))
1232                 RETURN(rc = PTR_ERR(o));
1233
1234         rc = mdt_check_enc(info, o);
1235         if (rc)
1236                 GOTO(out, rc);
1237
1238         if (mdt_object_remote(o)) {
1239                 /* Something is wrong here, the object is on another MDS! */
1240                 CERROR("%s: "DFID" isn't on this server!: rc = %d\n",
1241                        mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1242                 LU_OBJECT_DEBUG(D_WARNING, info->mti_env,
1243                                 &o->mot_obj,
1244                                 "Object isn't on this server! FLD error?");
1245                 rc = -EFAULT;
1246         } else {
1247                 if (mdt_object_exists(o)) {
1248                         int mask;
1249
1250                         /* Do permission check for cross-open after converting
1251                          * MDS_OPEN_* flags to MAY_* permission mask.
1252                          */
1253                         mask = mds_accmode(open_flags);
1254
1255                         rc = mo_permission(info->mti_env, NULL,
1256                                            mdt_object_child(o), NULL, mask);
1257                         if (rc)
1258                                 goto out;
1259
1260                         mdt_prep_ma_buf_from_rep(info, o, ma, open_flags);
1261                         rc = mdt_attr_get_complex(info, o, ma);
1262                         if (rc != 0)
1263                                 GOTO(out, rc);
1264
1265                         rc = mdt_pack_secctx_in_reply(info, o);
1266                         if (unlikely(rc))
1267                                 GOTO(out, rc);
1268
1269                         rc = mdt_pack_encctx_in_reply(info, o);
1270                         if (unlikely(rc))
1271                                 GOTO(out, rc);
1272
1273                         rc = mdt_finish_open(info, NULL, o, open_flags, rep);
1274                 } else {
1275                         /*
1276                          * Something is wrong here. lookup was positive but
1277                          * there is no object!
1278                          */
1279                         CERROR("%s: "DFID" doesn't exist!: rc = %d\n",
1280                               mdt_obd_name(info->mti_mdt), PFID(fid), -EFAULT);
1281                         rc = -EFAULT;
1282                 }
1283         }
1284 out:
1285         mdt_object_put(info->mti_env, o);
1286
1287         RETURN(rc);
1288 }
1289
1290 /*
1291  * find root object and take its xattr lock if it's on remote MDT, later create
1292  * may use fs default striping (which is stored in root xattr).
1293  */
1294 static int mdt_lock_root_xattr(struct mdt_thread_info *info,
1295                                struct mdt_device *mdt)
1296 {
1297         struct mdt_object *md_root = mdt->mdt_md_root;
1298         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1299         __u64 ibits = MDS_INODELOCK_XATTR;
1300         int rc;
1301
1302         if (md_root == NULL) {
1303                 lu_root_fid(&info->mti_tmp_fid1);
1304                 md_root = mdt_object_find(info->mti_env, mdt,
1305                                           &info->mti_tmp_fid1);
1306                 if (IS_ERR(md_root))
1307                         return PTR_ERR(md_root);
1308
1309                 spin_lock(&mdt->mdt_lock);
1310                 if (mdt->mdt_md_root != NULL) {
1311                         spin_unlock(&mdt->mdt_lock);
1312
1313                         LASSERTF(mdt->mdt_md_root == md_root,
1314                                  "Different root object ("
1315                                  DFID") instances, %px, %px\n",
1316                                  PFID(&info->mti_tmp_fid1),
1317                                  mdt->mdt_md_root, md_root);
1318                         LASSERT(atomic_read(
1319                                 &md_root->mot_obj.lo_header->loh_ref) > 1);
1320
1321                         mdt_object_put(info->mti_env, md_root);
1322                 } else {
1323                         mdt->mdt_md_root = md_root;
1324                         spin_unlock(&mdt->mdt_lock);
1325                 }
1326         }
1327
1328         if (md_root->mot_cache_attr || !mdt_object_remote(md_root))
1329                 return 0;
1330
1331         mdt_lock_reg_init(lh, LCK_PR);
1332         rc = mdt_object_lock_internal(info, md_root, mdt_object_fid(md_root),
1333                                       lh, &ibits, 0, true);
1334         if (rc < 0)
1335                 return rc;
1336
1337         md_root->mot_cache_attr = 1;
1338
1339         /* don't cancel this lock, so that we know the cached xattr is valid. */
1340         ldlm_lock_decref(&lh->mlh_rreg_lh, LCK_PR);
1341         lh->mlh_rreg_lh.cookie = 0ull;
1342
1343         return 0;
1344 }
1345
1346 static inline enum ldlm_mode mdt_open_lock_mode(struct mdt_thread_info *info,
1347                                                 struct mdt_object *p,
1348                                                 struct lu_name *name,
1349                                                 u64 open_flags)
1350 {
1351         int result;
1352         struct lu_fid fid;
1353
1354         /* We don't need to take the DLM lock for a volatile */
1355         if (open_flags & MDS_OPEN_VOLATILE)
1356                 return LCK_NL;
1357
1358         if (!(open_flags & MDS_OPEN_CREAT))
1359                 return LCK_PR;
1360
1361         result = mdo_lookup(info->mti_env, mdt_object_child(p), name, &fid,
1362                             &info->mti_spec);
1363
1364         /* If the file exists we only need a read lock on the parent */
1365         return (result == 0) ? LCK_PR : LCK_PW;
1366 }
1367
1368 int mdt_reint_open(struct mdt_thread_info *info, struct mdt_lock_handle *lhc)
1369 {
1370         struct mdt_device *mdt = info->mti_mdt;
1371         struct ptlrpc_request *req = mdt_info_req(info);
1372         struct mdt_object *parent;
1373         struct mdt_object *child;
1374         struct mdt_lock_handle *lh = NULL;
1375         struct ldlm_reply *ldlm_rep;
1376         struct mdt_body *repbody;
1377         struct lu_fid *child_fid = &info->mti_tmp_fid1;
1378         struct lu_ucred *uc = mdt_ucred(info);
1379         struct md_attr *ma = &info->mti_attr;
1380         u64 open_flags = info->mti_spec.sp_cr_flags;
1381         u64 ibits = 0;
1382         struct mdt_reint_record *rr = &info->mti_rr;
1383         int result, rc;
1384         int created = 0;
1385         int object_locked = 0;
1386         enum ldlm_mode lock_mode;
1387         u32 msg_flags;
1388         ktime_t kstart = ktime_get();
1389
1390         ENTRY;
1391         CFS_FAIL_TIMEOUT_ORSET(OBD_FAIL_MDS_PAUSE_OPEN, CFS_FAIL_ONCE,
1392                                (obd_timeout + 1) / 4);
1393
1394         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1395
1396         ma->ma_need = MA_INODE;
1397         ma->ma_valid = 0;
1398
1399         LASSERT(info->mti_pill->rc_fmt == &RQF_LDLM_INTENT_OPEN);
1400         ldlm_rep = req_capsule_server_get(info->mti_pill, &RMF_DLM_REP);
1401
1402         if (unlikely(open_flags & MDS_OPEN_JOIN_FILE)) {
1403                 CERROR("file join is not supported anymore.\n");
1404                 GOTO(out, result = err_serious(-EOPNOTSUPP));
1405         }
1406         msg_flags = lustre_msg_get_flags(req->rq_reqmsg);
1407
1408         if ((open_flags & (MDS_OPEN_HAS_EA | MDS_OPEN_HAS_OBJS)) &&
1409             info->mti_spec.u.sp_ea.eadata == NULL)
1410                 GOTO(out, result = err_serious(-EINVAL));
1411
1412         if (open_flags & MDS_FMODE_WRITE &&
1413             exp_connect_flags(req->rq_export) & OBD_CONNECT_RDONLY)
1414                 GOTO(out, result = -EROFS);
1415
1416         CDEBUG(D_INODE, "I am going to open "DFID"/("DNAME"->"DFID") "
1417                "cr_flag=%#llo mode=0%06o msg_flag=0x%x\n",
1418                PFID(rr->rr_fid1), PNAME(&rr->rr_name), PFID(rr->rr_fid2),
1419                open_flags, ma->ma_attr.la_mode, msg_flags);
1420
1421         /* Prevent by-fid operation if parent fid is .lustre/fid.
1422          * Also, we want rbac roles to have precedence over any other
1423          * permission or capability checks
1424          */
1425         if (lu_fid_eq(rr->rr_fid1, &LU_OBF_FID) && !uc->uc_rbac_byfid_ops)
1426                 GOTO(out, result = -EPERM);
1427
1428         if (info->mti_cross_ref) {
1429                 /* This is cross-ref open */
1430                 mdt_set_disposition(info, ldlm_rep,
1431                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD |
1432                              DISP_LOOKUP_POS));
1433                 result = mdt_cross_open(info, rr->rr_fid2, rr->rr_fid1,
1434                                         ldlm_rep, open_flags);
1435                 GOTO(out, result);
1436         } else if (req_is_replay(req)) {
1437                 result = mdt_open_by_fid(info, ldlm_rep, lhc);
1438
1439                 if (result != -ENOENT)
1440                         GOTO(out, result);
1441
1442                 /* We didn't find the correct object, so we need to re-create it
1443                  * via a regular replay. */
1444                 if (!(open_flags & MDS_OPEN_CREAT)) {
1445                         DEBUG_REQ(D_ERROR, req,
1446                                   "OPEN & CREAT not in open replay/by_fid");
1447                         GOTO(out, result = -EFAULT);
1448                 }
1449                 CDEBUG(D_INFO, "No object(1), continue as regular open.\n");
1450         } else if (open_flags & MDS_OPEN_BY_FID) {
1451                 result = mdt_open_by_fid_lock(info, ldlm_rep, lhc);
1452                 if (result < 0)
1453                         CDEBUG(D_INFO, "no object for "DFID": %d\n",
1454                                PFID(rr->rr_fid2), result);
1455                 GOTO(out, result);
1456         }
1457
1458         if (CFS_FAIL_CHECK(OBD_FAIL_MDS_OPEN_PACK))
1459                 GOTO(out, result = err_serious(-ENOMEM));
1460
1461         mdt_set_disposition(info, ldlm_rep,
1462                             (DISP_IT_EXECD | DISP_LOOKUP_EXECD));
1463
1464         if (!lu_name_is_valid(&rr->rr_name))
1465                 GOTO(out, result = -EPROTO);
1466
1467         result = mdt_lock_root_xattr(info, mdt);
1468         if (result < 0)
1469                 GOTO(out, result);
1470
1471         parent = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
1472         if (IS_ERR(parent))
1473                 GOTO(out, result = PTR_ERR(parent));
1474
1475         /* get and check version of parent */
1476         result = mdt_version_get_check(info, parent, 0);
1477         if (result) {
1478                 mdt_object_put(info->mti_env, parent);
1479                 GOTO(out, result);
1480         }
1481
1482         if (!uc->uc_rbac_fscrypt_admin &&
1483             parent->mot_obj.lo_header->loh_attr & LOHA_FSCRYPT_MD &&
1484             open_flags & MDS_OPEN_CREAT)
1485                 GOTO(out_parent, result = -EPERM);
1486
1487         result = mdt_check_enc(info, parent);
1488         if (result)
1489                 GOTO(out_parent, result);
1490
1491         fid_zero(child_fid);
1492         result = -ENOENT;
1493         lock_mode = mdt_open_lock_mode(info, parent, &rr->rr_name, open_flags);
1494
1495         CFS_RACE(OBD_FAIL_MDS_REINT_OPEN);
1496 again_pw:
1497         if (lock_mode != LCK_NL) {
1498                 lh = &info->mti_lh[MDT_LH_PARENT];
1499                 result = mdt_parent_lock(info, parent, lh, &rr->rr_name,
1500                                          lock_mode);
1501                 if (result != 0)
1502                         GOTO(out_parent, result);
1503
1504                 result = mdo_lookup(info->mti_env, mdt_object_child(parent),
1505                                     &rr->rr_name, child_fid, &info->mti_spec);
1506         }
1507
1508         LASSERTF(ergo(result == 0, fid_is_sane(child_fid)),
1509                  "looking for "DFID"/"DNAME", found FID = "DFID"\n",
1510                  PFID(mdt_object_fid(parent)), PNAME(&rr->rr_name),
1511                  PFID(child_fid));
1512
1513         if (result != 0 && result != -ENOENT)
1514                 GOTO(out_parent_unlock, result);
1515
1516         CFS_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1517
1518         if (result == -ENOENT) {
1519                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1520                 if (!(open_flags & MDS_OPEN_CREAT))
1521                         GOTO(out_parent_unlock, result);
1522                 if (mdt_rdonly(req->rq_export))
1523                         GOTO(out_parent_unlock, result = -EROFS);
1524
1525                 LASSERT(equi(lh == NULL, lock_mode == LCK_NL));
1526
1527                 if (lock_mode == LCK_PR) {
1528                         /* unlink vs create race: get write lock and restart */
1529                         mdt_object_unlock(info, parent, lh, 1);
1530                         mdt_clear_disposition(info, ldlm_rep, DISP_LOOKUP_NEG);
1531                         lock_mode = LCK_PW;
1532                         goto again_pw;
1533                 }
1534
1535                 *child_fid = *info->mti_rr.rr_fid2;
1536                 LASSERTF(fid_is_sane(child_fid), "fid="DFID"\n",
1537                          PFID(child_fid));
1538                 /* In the function below, .hs_keycmp resolves to
1539                  * lu_obj_hop_keycmp() */
1540                 child = mdt_object_new(info->mti_env, mdt, child_fid);
1541         } else {
1542                 /*
1543                  * Check for O_EXCL is moved to the mdt_finish_open(),
1544                  * we need to return FID back in that case.
1545                  */
1546                 mdt_set_disposition(info, ldlm_rep, DISP_LOOKUP_POS);
1547                 child = mdt_object_find(info->mti_env, mdt, child_fid);
1548         }
1549         if (IS_ERR(child))
1550                 GOTO(out_parent_unlock, result = PTR_ERR(child));
1551
1552         /** check version of child  */
1553         rc = mdt_version_get_check(info, child, 1);
1554         if (rc)
1555                 GOTO(out_child, result = rc);
1556
1557         tgt_open_obj_set(info->mti_env, mdt_obj2dt(child));
1558
1559         if (result == -ENOENT) {
1560                 /* Create under OBF and .lustre is not permitted */
1561                 if (!fid_is_md_operative(rr->rr_fid1) &&
1562                     (open_flags & MDS_OPEN_VOLATILE) == 0)
1563                         GOTO(out_child, result = -EPERM);
1564
1565                 /* save versions in reply */
1566                 mdt_version_get_save(info, parent, 0);
1567                 mdt_version_get_save(info, child, 1);
1568
1569                 /* version of child will be changed */
1570                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
1571
1572                 /* Not found and with MDS_OPEN_CREAT: let's create it. */
1573                 mdt_set_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1574
1575                 /* Don't do lookup sanity check. We know name doesn't exist. */
1576                 info->mti_spec.sp_cr_lookup = 0;
1577                 info->mti_spec.sp_feat = &dt_directory_features;
1578
1579                 /* set jobid xattr name from sysfs parameter */
1580                 strncpy(info->mti_spec.sp_cr_job_xattr, mdt->mdt_job_xattr,
1581                         XATTR_JOB_MAX_LEN);
1582
1583                 result = mdo_create(info->mti_env, mdt_object_child(parent),
1584                                     &rr->rr_name, mdt_object_child(child),
1585                                     &info->mti_spec, &info->mti_attr);
1586                 if (result == -ERESTART) {
1587                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1588                         GOTO(out_child, result);
1589                 } else {
1590                         mdt_prep_ma_buf_from_rep(info, child, ma, open_flags);
1591                         /* XXX: we should call this once, see few lines below */
1592                         if (result == 0)
1593                                 result = mdt_attr_get_complex(info, child, ma);
1594
1595                         if (result != 0)
1596                                 GOTO(out_child, result);
1597                 }
1598                 created = 1;
1599                 mdt_counter_incr(req, LPROC_MDT_MKNOD,
1600                                  ktime_us_delta(ktime_get(), kstart));
1601         } else {
1602                 /*
1603                  * The object is on remote node, return its FID for remote open.
1604                  */
1605                 if (mdt_object_remote(child)) {
1606                         /*
1607                          * Check if this lock already was sent to client and
1608                          * this is resent case. For resent case do not take lock
1609                          * again, use what is already granted.
1610                          */
1611                         LASSERT(lhc != NULL);
1612
1613                         rc = mdt_check_resent_lock(info, child, lhc);
1614                         if (rc < 0)
1615                                 GOTO(out_child, result = rc);
1616                         else if (rc > 0)
1617                                 rc = mdt_object_lookup_lock(info, NULL, child,
1618                                                             lhc, LCK_PR);
1619                         repbody->mbo_fid1 = *mdt_object_fid(child);
1620                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1621                         if (rc != 0)
1622                                 result = rc;
1623                         else
1624                                 result = -MDT_EREMOTE_OPEN;
1625                         GOTO(out_child, result);
1626                 } else if (mdt_object_exists(child)) {
1627                         /* Check early for MDS_OPEN_DIRECTORY/O_DIRECTORY to
1628                          * avoid opening regular files from lfs getstripe
1629                          * since doing so breaks the leases used by lfs
1630                          * mirror. See LU-13693. */
1631                         if (open_flags & MDS_OPEN_DIRECTORY &&
1632                             S_ISREG(lu_object_attr(&child->mot_obj)))
1633                                 GOTO(out_child, result = -ENOTDIR);
1634
1635                         /* We have to get attr & LOV EA & HSM for this
1636                          * object. */
1637                         mdt_prep_ma_buf_from_rep(info, child, ma, open_flags);
1638                         ma->ma_need |= MA_HSM;
1639                         result = mdt_attr_get_complex(info, child, ma);
1640                         if (result != 0)
1641                                 GOTO(out_child, result);
1642                 } else {
1643                         /* Object does not exist. Likely FS corruption. */
1644                         CERROR("%s: name '"DNAME"' present, but FID "
1645                                DFID" is invalid\n", mdt_obd_name(info->mti_mdt),
1646                                PNAME(&rr->rr_name), PFID(child_fid));
1647                         GOTO(out_child, result = -EIO);
1648                 }
1649         }
1650
1651         repbody->mbo_max_mdsize = info->mti_mdt->mdt_max_mdsize;
1652         repbody->mbo_valid |= OBD_MD_FLMODEASIZE;
1653
1654         rc = mdt_pack_secctx_in_reply(info, child);
1655         if (unlikely(rc))
1656                 GOTO(out_child, result = rc);
1657
1658         rc = mdt_pack_encctx_in_reply(info, child);
1659         if (unlikely(rc))
1660                 GOTO(out_child, result = rc);
1661
1662         rc = mdt_check_resent_lock(info, child, lhc);
1663         if (rc < 0) {
1664                 GOTO(out_child, result = rc);
1665         } else if (rc == 0) {
1666                 /* the open lock might already be gotten in
1667                  * ldlm_handle_enqueue() */
1668                 LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
1669                 if (open_flags & MDS_OPEN_LOCK)
1670                         mdt_set_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1671         } else {
1672                 /* get openlock if this isn't replay and client requested it */
1673                 if (!req_is_replay(req)) {
1674                         CFS_FAIL_TIMEOUT(OBD_FAIL_MDS_DELAY_OPEN, cfs_fail_val);
1675                         rc = mdt_object_open_lock(info, child, lhc, &ibits);
1676                         object_locked = 1;
1677                         if (rc != 0)
1678                                 GOTO(out_child_unlock, result = rc);
1679                         else if (open_flags & MDS_OPEN_LOCK)
1680                                 mdt_set_disposition(info, ldlm_rep,
1681                                                     DISP_OPEN_LOCK);
1682                 }
1683         }
1684         /* Try to open it now. */
1685         rc = mdt_finish_open(info, parent, child, open_flags, ldlm_rep);
1686         if (rc) {
1687                 result = rc;
1688                 /* openlock will be released if mdt_finish_open() failed */
1689                 mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_LOCK);
1690
1691                 if (created && (open_flags & MDS_OPEN_VOLATILE)) {
1692                         CERROR("%s: cannot open volatile file "DFID", orphan "
1693                                "file will be left in PENDING directory until "
1694                                "next reboot, rc = %d\n", mdt_obd_name(mdt),
1695                                PFID(mdt_object_fid(child)), rc);
1696                         GOTO(out_child_unlock, result);
1697                 }
1698
1699                 if (created) {
1700                         ma->ma_need = 0;
1701                         ma->ma_valid = 0;
1702                         rc = mdo_unlink(info->mti_env,
1703                                         mdt_object_child(parent),
1704                                         mdt_object_child(child),
1705                                         &rr->rr_name,
1706                                         &info->mti_attr, 0);
1707                         if (rc != 0)
1708                                 CERROR("%s: "DFID" cleanup of open: rc = %d\n",
1709                                        mdt_obd_name(info->mti_mdt),
1710                                        PFID(mdt_object_fid(child)), rc);
1711                         mdt_clear_disposition(info, ldlm_rep, DISP_OPEN_CREATE);
1712                 }
1713                 GOTO(out_child_unlock, result);
1714         }
1715
1716         /*
1717          * if layout lock is granted, then we should re-fetch LOVEA
1718          * which was originally taken w/o the lock.
1719          */
1720         result = mdt_refetch_lovea(info, child, ma, ibits);
1721
1722         GOTO(out_child_unlock, result);
1723
1724 out_child_unlock:
1725         if (object_locked)
1726                 mdt_object_open_unlock(info, child, lhc, ibits, result);
1727 out_child:
1728         mdt_object_put(info->mti_env, child);
1729         if (result == 0)
1730                 mdt_pack_size2body(info, child_fid, &lhc->mlh_reg_lh);
1731 out_parent_unlock:
1732         if (lh != NULL)
1733                 mdt_object_unlock(info, parent, lh, result || !created);
1734
1735 out_parent:
1736         mdt_object_put(info->mti_env, parent);
1737 out:
1738         if (result)
1739                 lustre_msg_set_transno(req->rq_repmsg, 0);
1740         else
1741                 mdt_counter_incr(req, LPROC_MDT_OPEN,
1742                                  ktime_us_delta(ktime_get(), kstart));
1743
1744         return result;
1745 }
1746
1747 /**
1748  * Create an orphan object use local root.
1749  */
1750 static struct mdt_object *mdt_orphan_open(struct mdt_thread_info *info,
1751                                           struct mdt_device *mdt,
1752                                           const struct lu_fid *fid,
1753                                           struct md_attr *attr, fmode_t fmode)
1754 {
1755         const struct lu_env *env = info->mti_env;
1756         struct md_op_spec *spec = &info->mti_spec;
1757         struct lu_fid *local_root_fid = &info->mti_tmp_fid1;
1758         struct mdt_object *obj = NULL;
1759         struct mdt_object *local_root;
1760         static const struct lu_name lname = {
1761                 .ln_name = "i_am_nobody",
1762                 .ln_namelen = sizeof("i_am_nobody") - 1,
1763         };
1764         struct lu_ucred *uc;
1765         kernel_cap_t uc_cap_save;
1766         int rc;
1767
1768         ENTRY;
1769         rc = dt_root_get(env, mdt->mdt_bottom, local_root_fid);
1770         if (rc != 0)
1771                 RETURN(ERR_PTR(rc));
1772
1773         local_root = mdt_object_find(env, mdt, local_root_fid);
1774         if (IS_ERR(local_root))
1775                 RETURN(local_root);
1776
1777         obj = mdt_object_new(env, mdt, fid);
1778         if (IS_ERR(obj))
1779                 GOTO(out, rc = PTR_ERR(obj));
1780
1781         spec->sp_cr_lookup = 0;
1782         spec->sp_feat = &dt_directory_features;
1783         spec->sp_cr_flags = MDS_OPEN_VOLATILE | fmode;
1784         if (attr->ma_valid & MA_LOV) {
1785                 spec->u.sp_ea.eadata = attr->ma_lmm;
1786                 spec->u.sp_ea.eadatalen = attr->ma_lmm_size;
1787                 spec->sp_cr_flags |= MDS_OPEN_HAS_EA;
1788         } else {
1789                 spec->sp_cr_flags |= MDS_OPEN_DELAY_CREATE;
1790         }
1791
1792         uc = lu_ucred(env);
1793         uc_cap_save = uc->uc_cap;
1794         cap_raise(uc->uc_cap, CAP_DAC_OVERRIDE);
1795         rc = mdo_create(env, mdt_object_child(local_root), &lname,
1796                         mdt_object_child(obj), spec, attr);
1797         uc->uc_cap = uc_cap_save;
1798         if (rc < 0) {
1799                 CERROR("%s: cannot create volatile file "DFID": rc = %d\n",
1800                        mdt_obd_name(mdt), PFID(fid), rc);
1801                 GOTO(out, rc);
1802         }
1803
1804         rc = mo_open(env, mdt_object_child(obj), MDS_OPEN_CREATED, spec);
1805         if (rc < 0)
1806                 CERROR("%s: cannot open volatile file "DFID", orphan "
1807                        "file will be left in PENDING directory until "
1808                        "next reboot, rc = %d\n", mdt_obd_name(mdt),
1809                        PFID(fid), rc);
1810         GOTO(out, rc);
1811
1812 out:
1813         if (rc < 0) {
1814                 if (!IS_ERR(obj))
1815                         mdt_object_put(env, obj);
1816                 obj = ERR_PTR(rc);
1817         }
1818         mdt_object_put(env, local_root);
1819         return obj;
1820 }
1821
1822 /* XXX Look into layout in MDT layer. */
1823 static inline int mdt_hsm_set_released(struct lov_mds_md *lmm)
1824 {
1825         struct lov_comp_md_v1 *comp_v1;
1826         struct lov_mds_md *v1;
1827         __u32 off;
1828         int i;
1829
1830         if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_COMP_V1_DEFINED)) {
1831                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1832
1833                 if (comp_v1->lcm_entry_count == 0)
1834                         return -EINVAL;
1835
1836                 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1837                         off = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1838                         v1 = (struct lov_mds_md *)((char *)comp_v1 + off);
1839                         if (v1->lmm_magic == cpu_to_le32(LOV_MAGIC_FOREIGN))
1840                                 continue;
1841
1842                         v1->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1843                 }
1844         } else {
1845                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_RELEASED);
1846         }
1847         return 0;
1848 }
1849
1850 static inline int mdt_get_lmm_gen(struct lov_mds_md *lmm, __u32 *gen)
1851 {
1852         struct lov_comp_md_v1 *comp_v1;
1853
1854         if (le32_to_cpu(lmm->lmm_magic == LOV_MAGIC_COMP_V1)) {
1855                 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1856                 *gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1857         } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V1 ||
1858                    le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_V3) {
1859                 *gen = le16_to_cpu(lmm->lmm_layout_gen);
1860         } else {
1861                 return -EINVAL;
1862         }
1863         return 0;
1864 }
1865
1866 static int mdt_hsm_release(struct mdt_thread_info *info, struct mdt_object *o,
1867                            struct md_attr *ma)
1868 {
1869         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LAYOUT];
1870         struct lu_ucred *uc = mdt_ucred(info);
1871         struct close_data *data;
1872         struct ldlm_lock *lease;
1873         struct mdt_object *orphan;
1874         struct md_attr *orp_ma;
1875         struct lu_buf *buf;
1876         kernel_cap_t cap;
1877         bool lease_broken;
1878         int rc;
1879         int rc2;
1880
1881         ENTRY;
1882         if (mdt_rdonly(info->mti_exp))
1883                 RETURN(-EROFS);
1884
1885         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1886         if (data == NULL)
1887                 RETURN(-EPROTO);
1888
1889         if (req_capsule_req_need_swab(info->mti_pill))
1890                 lustre_swab_close_data_special(data, ma->ma_attr_flags);
1891
1892         lease = ldlm_handle2lock(&data->cd_handle);
1893         if (lease == NULL)
1894                 RETURN(-ESTALE);
1895
1896         /* try to hold open_sem so that nobody else can open the file */
1897         if (!down_write_trylock(&o->mot_open_sem)) {
1898                 ldlm_lock_cancel(lease);
1899                 GOTO(out_reprocess, rc = -EBUSY);
1900         }
1901
1902         /* Check if the lease open lease has already canceled */
1903         lock_res_and_lock(lease);
1904         lease_broken = ldlm_is_cancel(lease);
1905         unlock_res_and_lock(lease);
1906
1907         LDLM_DEBUG(lease, DFID " lease broken? %d",
1908                    PFID(mdt_object_fid(o)), lease_broken);
1909
1910         /* Cancel server side lease. Client side counterpart should
1911          * have been cancelled. It's okay to cancel it now as we've
1912          * held mot_open_sem. */
1913         ldlm_lock_cancel(lease);
1914
1915         if (lease_broken) /* don't perform release task */
1916                 GOTO(out_unlock, rc = -ESTALE);
1917
1918         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
1919                 GOTO(out_unlock, rc = -EINVAL);
1920
1921         /* ma_need was set before but it seems fine to change it in order to
1922          * avoid modifying the one from RPC */
1923         ma->ma_need = MA_HSM;
1924         rc = mdt_attr_get_complex(info, o, ma);
1925         if (rc != 0)
1926                 GOTO(out_unlock, rc);
1927
1928         if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
1929                 if (ma->ma_valid & MA_HSM) {
1930                         if (ma->ma_hsm.mh_flags & HS_RELEASED)
1931                                 GOTO(out_unlock, rc = -EALREADY);
1932
1933                         if (ma->ma_hsm.mh_arch_id != data->cd_archive_id)
1934                                 CDEBUG(D_CACHE,
1935                                        DFID" archive id diff: %llu:%u\n",
1936                                        PFID(mdt_object_fid(o)),
1937                                        ma->ma_hsm.mh_arch_id,
1938                                        data->cd_archive_id);
1939
1940                         if (!(ma->ma_hsm.mh_flags & HS_DIRTY) &&
1941                             ma->ma_hsm.mh_arch_ver == data->cd_data_version) {
1942                                 CDEBUG(D_CACHE,
1943                                        DFID" data version matches: packed=%llu "
1944                                        "and on-disk=%llu\n",
1945                                        PFID(mdt_object_fid(o)),
1946                                        data->cd_data_version,
1947                                        ma->ma_hsm.mh_arch_ver);
1948                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1949                         }
1950
1951                         if (ma->ma_hsm.mh_flags & HS_DIRTY)
1952                                 ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1953                 } else {
1954                         /* Set up HSM attribte for PCC archived object */
1955                         BUILD_BUG_ON(sizeof(struct hsm_attrs) >
1956                                      sizeof(info->mti_xattr_buf));
1957                         buf = &info->mti_buf;
1958                         buf->lb_buf = info->mti_xattr_buf;
1959                         buf->lb_len = sizeof(struct hsm_attrs);
1960                         memset(&ma->ma_hsm, 0, sizeof(ma->ma_hsm));
1961                         ma->ma_hsm.mh_flags = HS_ARCHIVED | HS_EXISTS;
1962                         ma->ma_hsm.mh_arch_id = data->cd_archive_id;
1963                         ma->ma_hsm.mh_arch_ver = data->cd_data_version;
1964                         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
1965
1966                         rc = mo_xattr_set(info->mti_env, mdt_object_child(o),
1967                                           buf, XATTR_NAME_HSM, 0);
1968                         if (rc)
1969                                 GOTO(out_unlock, rc);
1970                 }
1971         } else {
1972                 if (!mdt_hsm_release_allow(ma))
1973                         GOTO(out_unlock, rc = -EPERM);
1974
1975                 /* already released? */
1976                 if (ma->ma_hsm.mh_flags & HS_RELEASED)
1977                         GOTO(out_unlock, rc = 0);
1978
1979                 /* Compare on-disk and packed data_version */
1980                 if (data->cd_data_version != ma->ma_hsm.mh_arch_ver) {
1981                         CDEBUG(D_HSM, DFID" data_version mismatches: "
1982                                "packed=%llu and on-disk=%llu\n",
1983                                PFID(mdt_object_fid(o)),
1984                                data->cd_data_version,
1985                                ma->ma_hsm.mh_arch_ver);
1986                         GOTO(out_unlock, rc = -EPERM);
1987                 }
1988         }
1989
1990         ma->ma_valid = MA_INODE;
1991         ma->ma_attr.la_valid &= LA_ATIME | LA_MTIME | LA_CTIME | LA_SIZE;
1992         rc = mo_attr_set(info->mti_env, mdt_object_child(o), ma);
1993         if (rc < 0)
1994                 GOTO(out_unlock, rc);
1995
1996         mutex_lock(&o->mot_som_mutex);
1997         rc2 = mdt_set_som(info, o, SOM_FL_STRICT, ma->ma_attr.la_size,
1998                            ma->ma_attr.la_blocks);
1999         mutex_unlock(&o->mot_som_mutex);
2000         if (rc2 < 0)
2001                 CDEBUG(D_INODE,
2002                        "%s: File "DFID" SOM update failed: rc = %d\n",
2003                        mdt_obd_name(info->mti_mdt),
2004                        PFID(mdt_object_fid(o)), rc2);
2005
2006
2007         ma->ma_need = MA_INODE | MA_LOV;
2008         rc = mdt_attr_get_complex(info, o, ma);
2009         if (rc < 0)
2010                 GOTO(out_unlock, rc);
2011
2012         if (!(ma->ma_valid & MA_LOV)) {
2013                 /* Even empty file are released */
2014                 memset(ma->ma_lmm, 0, sizeof(*ma->ma_lmm));
2015                 ma->ma_lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1_DEFINED);
2016                 ma->ma_lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
2017                 ma->ma_lmm->lmm_stripe_size = cpu_to_le32(LOV_MIN_STRIPE_SIZE);
2018                 ma->ma_lmm_size = sizeof(*ma->ma_lmm);
2019         } else {
2020                 /* Magic must be LOV_MAGIC_*_DEFINED or LOD will interpret
2021                  * ma_lmm as lov_user_md, then it will be confused by union of
2022                  * layout_gen and stripe_offset. */
2023                 if ((le32_to_cpu(ma->ma_lmm->lmm_magic) & LOV_MAGIC_MASK) ==
2024                     LOV_MAGIC_MAGIC)
2025                         ma->ma_lmm->lmm_magic |= cpu_to_le32(LOV_MAGIC_DEFINED);
2026                 else
2027                         GOTO(out_unlock, rc = -EINVAL);
2028         }
2029
2030         /* Set file as released. */
2031         rc = mdt_hsm_set_released(ma->ma_lmm);
2032         if (rc)
2033                 GOTO(out_unlock, rc);
2034
2035         orp_ma = &info->mti_u.hsm.attr;
2036         orp_ma->ma_attr.la_mode = S_IFREG | S_IWUSR;
2037         /* We use root ownership to bypass potential quota
2038          * restrictions on the user and group of the file. */
2039         orp_ma->ma_attr.la_uid = 0;
2040         orp_ma->ma_attr.la_gid = 0;
2041         orp_ma->ma_attr.la_valid = LA_MODE | LA_UID | LA_GID;
2042         orp_ma->ma_lmm = ma->ma_lmm;
2043         orp_ma->ma_lmm_size = ma->ma_lmm_size;
2044         orp_ma->ma_valid = MA_INODE | MA_LOV;
2045         orphan = mdt_orphan_open(info, info->mti_mdt, &data->cd_fid, orp_ma,
2046                                  MDS_FMODE_WRITE);
2047         if (IS_ERR(orphan)) {
2048                 CERROR("%s: cannot open orphan file "DFID": rc = %ld\n",
2049                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid),
2050                        PTR_ERR(orphan));
2051                 GOTO(out_unlock, rc = PTR_ERR(orphan));
2052         }
2053
2054         /* Set up HSM attribute for orphan object */
2055         BUILD_BUG_ON(sizeof(struct hsm_attrs) > sizeof(info->mti_xattr_buf));
2056         buf = &info->mti_buf;
2057         buf->lb_buf = info->mti_xattr_buf;
2058         buf->lb_len = sizeof(struct hsm_attrs);
2059         ma->ma_hsm.mh_flags |= HS_RELEASED;
2060         lustre_hsm2buf(buf->lb_buf, &ma->ma_hsm);
2061         ma->ma_hsm.mh_flags &= ~HS_RELEASED;
2062
2063         rc = mdt_object_lock(info, o, lh, MDS_INODELOCK_LAYOUT |
2064                              MDS_INODELOCK_XATTR, LCK_EX);
2065         if (rc != 0)
2066                 GOTO(out_close, rc);
2067
2068         /* The orphan has root ownership so we need to raise
2069          * CAP_FOWNER to set the HSM attributes. */
2070         cap = uc->uc_cap;
2071         cap_raise(uc->uc_cap, CAP_FOWNER);
2072         rc = mo_xattr_set(info->mti_env, mdt_object_child(orphan), buf,
2073                           XATTR_NAME_HSM, 0);
2074         uc->uc_cap = cap;
2075         if (rc != 0)
2076                 GOTO(out_layout_lock, rc);
2077
2078         /* Swap layout with orphan objects. */
2079         rc = mo_swap_layouts(info->mti_env, mdt_object_child(o),
2080                              mdt_object_child(orphan), 0, 0,
2081                              SWAP_LAYOUTS_MDS_RELEASE);
2082
2083         if (!rc && ma->ma_attr_flags & MDS_PCC_ATTACH) {
2084                 ma->ma_need = MA_LOV;
2085                 rc = mdt_attr_get_complex(info, o, ma);
2086         }
2087
2088         EXIT;
2089
2090 out_layout_lock:
2091         /* Release exclusive LL */
2092         mdt_object_unlock(info, o, lh, 1);
2093 out_close:
2094         /* Close orphan object anyway */
2095         rc2 = mo_close(info->mti_env, mdt_object_child(orphan), orp_ma,
2096                        MDS_FMODE_WRITE);
2097         if (rc2 < 0)
2098                 CERROR("%s: error closing volatile file "DFID": rc = %d\n",
2099                        mdt_obd_name(info->mti_mdt), PFID(&data->cd_fid), rc2);
2100         LU_OBJECT_DEBUG(D_HSM, info->mti_env, &orphan->mot_obj,
2101                         "object closed");
2102         mdt_object_put(info->mti_env, orphan);
2103
2104 out_unlock:
2105         up_write(&o->mot_open_sem);
2106
2107         /* already released */
2108         if (rc == 0) {
2109                 struct mdt_body *repbody;
2110
2111                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2112                 LASSERT(repbody != NULL);
2113                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2114                 if (ma->ma_attr_flags & MDS_PCC_ATTACH) {
2115                         LASSERT(ma->ma_valid & MA_LOV);
2116                         rc = mdt_get_lmm_gen(ma->ma_lmm,
2117                                              &repbody->mbo_layout_gen);
2118                         if (!rc)
2119                                 repbody->mbo_valid |= OBD_MD_LAYOUT_VERSION;
2120                 }
2121         }
2122
2123 out_reprocess:
2124         ldlm_reprocess_all(lease->l_resource,
2125                            lease->l_policy_data.l_inodebits.bits);
2126         LDLM_LOCK_PUT(lease);
2127
2128         ma->ma_valid = 0;
2129         ma->ma_need = 0;
2130
2131         return rc;
2132 }
2133
2134 static int mdt_close_handle_layouts(struct mdt_thread_info *info,
2135                                     struct mdt_object *o, struct md_attr *ma)
2136 {
2137         struct mdt_lock_handle *lh1 = &info->mti_lh[MDT_LH_NEW];
2138         struct mdt_lock_handle *lh2 = &info->mti_lh[MDT_LH_OLD];
2139         struct close_data *data;
2140         struct ldlm_lock *lease;
2141         struct mdt_object *o1 = o, *o2 = NULL;
2142         bool lease_broken;
2143         bool swap_objects = false;
2144         int rc;
2145
2146         ENTRY;
2147         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2148                 RETURN(-EROFS);
2149
2150         if (!S_ISREG(lu_object_attr(&o1->mot_obj)))
2151                 RETURN(-EINVAL);
2152
2153         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2154         if (data == NULL)
2155                 RETURN(-EPROTO);
2156
2157         if (req_capsule_req_need_swab(info->mti_pill))
2158                 lustre_swab_close_data_special(data, ma->ma_attr_flags);
2159
2160         if (fid_is_zero(&data->cd_fid) || !fid_is_sane(&data->cd_fid))
2161                 RETURN(-EINVAL);
2162
2163         rc = lu_fid_cmp(&data->cd_fid, mdt_object_fid(o));
2164         if (rc == 0) {
2165                 /**
2166                  * only MDS_CLOSE_LAYOUT_SPLIT use the same fid to indicate
2167                  * mirror deletion, so we'd zero cd_fid, and keeps o2 be NULL.
2168                  */
2169                 if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT))
2170                         RETURN(-EINVAL);
2171
2172                 /* zero cd_fid to keeps o2 be NULL */
2173                 fid_zero(&data->cd_fid);
2174         } else if (rc < 0) {
2175                 /* Exchange o1 and o2, to enforce locking order */
2176                 swap_objects = true;
2177         }
2178
2179         lease = ldlm_handle2lock(&data->cd_handle);
2180         if (lease == NULL)
2181                 RETURN(-ESTALE);
2182
2183         if (!fid_is_zero(&data->cd_fid)) {
2184                 o2 = mdt_object_find(info->mti_env, info->mti_mdt,
2185                                      &data->cd_fid);
2186                 if (IS_ERR(o2))
2187                         GOTO(out_lease, rc = PTR_ERR(o2));
2188
2189                 if (!mdt_object_exists(o2))
2190                         GOTO(out_obj, rc = -ENOENT);
2191
2192                 if (!S_ISREG(lu_object_attr(&o2->mot_obj)))
2193                         GOTO(out_obj, rc = -EINVAL);
2194
2195                 if (swap_objects)
2196                         swap(o1, o2);
2197         }
2198
2199         rc = mo_permission(info->mti_env, NULL, mdt_object_child(o1), NULL,
2200                            MAY_WRITE);
2201         if (rc < 0)
2202                 GOTO(out_obj, rc);
2203
2204         if (o2) {
2205                 rc = mo_permission(info->mti_env, NULL, mdt_object_child(o2),
2206                                    NULL, MAY_WRITE);
2207                 if (rc < 0)
2208                         GOTO(out_obj, rc);
2209         }
2210
2211         /* try to hold open_sem so that nobody else can open the file */
2212         if (!down_write_trylock(&o->mot_open_sem)) {
2213                 ldlm_lock_cancel(lease);
2214                 GOTO(out_obj, rc = -EBUSY);
2215         }
2216
2217         /* Check if the lease open lease has already canceled */
2218         lock_res_and_lock(lease);
2219         lease_broken = ldlm_is_cancel(lease);
2220         unlock_res_and_lock(lease);
2221
2222         LDLM_DEBUG(lease, DFID " lease broken? %d",
2223                    PFID(mdt_object_fid(o)), lease_broken);
2224
2225         /* Cancel server side lease. Client side counterpart should
2226          * have been cancelled. It's okay to cancel it now as we've
2227          * held mot_open_sem. */
2228         ldlm_lock_cancel(lease);
2229
2230         if (lease_broken)
2231                 GOTO(out_unlock_sem, rc = -ESTALE);
2232
2233         rc = mdt_object_lock(info, o1, lh1, MDS_INODELOCK_LAYOUT |
2234                              MDS_INODELOCK_XATTR, LCK_EX);
2235         if (rc < 0)
2236                 GOTO(out_unlock_sem, rc);
2237
2238         if (o2) {
2239                 rc = mdt_object_lock(info, o2, lh2, MDS_INODELOCK_LAYOUT |
2240                                      MDS_INODELOCK_XATTR, LCK_EX);
2241                 if (rc < 0)
2242                         GOTO(out_unlock1, rc);
2243         }
2244
2245         /* Swap layout with orphan object */
2246         if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP) {
2247                 __u64 dv1 = data->cd_data_version;
2248                 __u64 dv2 = 0;
2249
2250                 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SWAP_HSM)
2251                         /* Compat: new clients send new dataversion in
2252                          * cd_data_version2 and old one in cd_data_version.
2253                          * Old clients sent cd_data_version = 0 and no
2254                          * cd_data_version2.
2255                          */
2256                         dv2 = data->cd_data_version2;
2257
2258                 if (swap_objects)
2259                         swap(dv1, dv2);
2260
2261                 rc = mo_swap_layouts(info->mti_env, mdt_object_child(o1),
2262                                      mdt_object_child(o2), dv1, dv2, 0);
2263         } else if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_MERGE ||
2264                    ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2265                 struct lu_buf *buf = &info->mti_buf;
2266                 struct md_rejig_data mrd;
2267
2268                 if (o2) {
2269                         mrd.mrd_obj = mdt_object_child(o == o1 ? o2 : o1);
2270                 } else {
2271                         if (!(ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT)) {
2272                                 /* paranoid check again */
2273                                 CERROR(DFID
2274                                   ":only mirror split support NULL o2 object\n",
2275                                         PFID(mdt_object_fid(o)));
2276                                 GOTO(out_unlock1, rc = -EINVAL);
2277                         }
2278
2279                         /* set NULL mrd_obj for deleting mirror objects */
2280                         mrd.mrd_obj = NULL;
2281                 }
2282
2283                 if (ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT) {
2284                         mrd.mrd_mirror_id = data->cd_mirror_id;
2285                         /* set a small enough blocks in the SoM */
2286                         ma->ma_attr.la_blocks >>= 1;
2287                 }
2288
2289                 buf->lb_len = sizeof(mrd);
2290                 buf->lb_buf = &mrd;
2291                 rc = mo_xattr_set(info->mti_env, mdt_object_child(o), buf,
2292                                   XATTR_LUSTRE_LOV,
2293                                   ma->ma_attr_flags & MDS_CLOSE_LAYOUT_SPLIT ?
2294                                   LU_XATTR_SPLIT : LU_XATTR_MERGE);
2295                 if (rc == 0 && ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS |
2296                                                        LA_LSIZE | LA_LBLOCKS)) {
2297                         int rc2;
2298                         enum lustre_som_flags lsf;
2299
2300                         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS))
2301                                 lsf = SOM_FL_STRICT;
2302                         else
2303                                 lsf = SOM_FL_LAZY;
2304
2305                         mutex_lock(&o->mot_som_mutex);
2306                         rc2 = mdt_set_som(info, o, lsf,
2307                                           ma->ma_attr.la_size,
2308                                           ma->ma_attr.la_blocks);
2309                         mutex_unlock(&o->mot_som_mutex);
2310                         if (rc2 < 0)
2311                                 CERROR(DFID": Setting i_blocks error: %d, "
2312                                        "i_blocks will be reported wrongly and "
2313                                        "can only be fixed in next resync\n",
2314                                        PFID(mdt_object_fid(o)), rc2);
2315                 }
2316         }
2317         if (rc < 0)
2318                 GOTO(out_unlock2, rc);
2319
2320         EXIT;
2321
2322 out_unlock2:
2323         /* Release exclusive LL */
2324         if (o2)
2325                 mdt_object_unlock(info, o2, lh2, 1);
2326
2327 out_unlock1:
2328         mdt_object_unlock(info, o1, lh1, 1);
2329
2330 out_unlock_sem:
2331         up_write(&o->mot_open_sem);
2332
2333         /* already swapped */
2334         if (rc == 0) {
2335                 struct mdt_body *repbody;
2336
2337                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2338                 LASSERT(repbody != NULL);
2339                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2340         }
2341
2342 out_obj:
2343         if (o1 != o)
2344                 /* the 2nd object has been used, and swapped to o1 */
2345                 mdt_object_put(info->mti_env, o1);
2346         else if (o2)
2347                 /* the 2nd object has been used, and not swapped */
2348                 mdt_object_put(info->mti_env, o2);
2349
2350         ldlm_reprocess_all(lease->l_resource,
2351                            lease->l_policy_data.l_inodebits.bits);
2352
2353 out_lease:
2354         LDLM_LOCK_PUT(lease);
2355
2356         if (ma != NULL) {
2357                 ma->ma_valid = 0;
2358                 ma->ma_need = 0;
2359         }
2360
2361         return rc;
2362 }
2363
2364 static int mdt_close_resync_done(struct mdt_thread_info *info,
2365                                  struct mdt_object *o, struct md_attr *ma)
2366 {
2367         struct mdt_lock_handle *lhc = &info->mti_lh[MDT_LH_LOCAL];
2368         struct close_data *data;
2369         struct ldlm_lock *lease;
2370         struct md_layout_change layout = { 0 };
2371         __u32 *resync_ids = NULL;
2372         size_t resync_count = 0;
2373         bool lease_broken;
2374         int rc;
2375
2376         ENTRY;
2377         if (exp_connect_flags(info->mti_exp) & OBD_CONNECT_RDONLY)
2378                 RETURN(-EROFS);
2379
2380         if (!S_ISREG(lu_object_attr(&o->mot_obj)))
2381                 RETURN(-EINVAL);
2382
2383         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2384         if (data == NULL)
2385                 RETURN(-EPROTO);
2386
2387         if (req_capsule_req_need_swab(info->mti_pill))
2388                 lustre_swab_close_data_special(data, ma->ma_attr_flags);
2389
2390         if (!fid_is_zero(&data->cd_fid))
2391                 RETURN(-EPROTO);
2392
2393         lease = ldlm_handle2lock(&data->cd_handle);
2394         if (lease == NULL)
2395                 RETURN(-ESTALE);
2396
2397         /* try to hold open_sem so that nobody else can open the file */
2398         if (!down_write_trylock(&o->mot_open_sem)) {
2399                 ldlm_lock_cancel(lease);
2400                 GOTO(out_reprocess, rc = -EBUSY);
2401         }
2402
2403         /* Check if the lease open lease has already canceled */
2404         lock_res_and_lock(lease);
2405         lease_broken = ldlm_is_cancel(lease);
2406         unlock_res_and_lock(lease);
2407
2408         LDLM_DEBUG(lease, DFID " lease broken? %d",
2409                    PFID(mdt_object_fid(o)), lease_broken);
2410
2411         /* Cancel server side lease. Client side counterpart should
2412          * have been cancelled. It's okay to cancel it now as we've
2413          * held mot_open_sem. */
2414         ldlm_lock_cancel(lease);
2415
2416         if (lease_broken) /* don't perform release task */
2417                 GOTO(out_unlock, rc = -ESTALE);
2418
2419         resync_count = data->cd_resync.resync_count;
2420
2421         if (resync_count > INLINE_RESYNC_ARRAY_SIZE) {
2422                 void *data;
2423
2424                 if (!req_capsule_has_field(info->mti_pill, &RMF_U32,
2425                                            RCL_CLIENT))
2426                         GOTO(out_unlock, rc = -EPROTO);
2427
2428                 OBD_ALLOC_PTR_ARRAY(resync_ids, resync_count);
2429                 if (!resync_ids)
2430                         GOTO(out_unlock, rc = -ENOMEM);
2431
2432                 data = req_capsule_client_get(info->mti_pill, &RMF_U32);
2433                 memcpy(resync_ids, data, resync_count * sizeof(__u32));
2434
2435                 layout.mlc_resync_ids = resync_ids;
2436         } else {
2437                 layout.mlc_resync_ids = data->cd_resync.resync_ids_inline;
2438         }
2439
2440         layout.mlc_opc = MD_LAYOUT_RESYNC_DONE;
2441         layout.mlc_resync_count = resync_count;
2442         if (ma->ma_attr.la_valid & (LA_SIZE | LA_BLOCKS)) {
2443                 layout.mlc_som.lsa_valid = SOM_FL_STRICT;
2444                 layout.mlc_som.lsa_size = ma->ma_attr.la_size;
2445                 layout.mlc_som.lsa_blocks = ma->ma_attr.la_blocks;
2446         }
2447         rc = mdt_layout_change(info, o, lhc, &layout);
2448         if (rc)
2449                 GOTO(out_unlock, rc);
2450
2451         mdt_object_unlock(info, o, lhc, 0);
2452
2453         EXIT;
2454
2455 out_unlock:
2456         up_write(&o->mot_open_sem);
2457
2458         /* already released */
2459         if (rc == 0) {
2460                 struct mdt_body *repbody;
2461
2462                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2463                 LASSERT(repbody != NULL);
2464                 repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2465         }
2466
2467         if (resync_ids)
2468                 OBD_FREE_PTR_ARRAY(resync_ids, resync_count);
2469
2470 out_reprocess:
2471         ldlm_reprocess_all(lease->l_resource,
2472                            lease->l_policy_data.l_inodebits.bits);
2473         LDLM_LOCK_PUT(lease);
2474
2475         ma->ma_valid = 0;
2476         ma->ma_need = 0;
2477
2478         return rc;
2479 }
2480
2481 #define MFD_CLOSED(open_flags) ((open_flags) == MDS_FMODE_CLOSED)
2482
2483 static int mdt_mfd_closed(struct mdt_file_data *mfd)
2484 {
2485         return ((mfd == NULL) || MFD_CLOSED(mfd->mfd_open_flags));
2486 }
2487
2488 int mdt_mfd_close(struct mdt_thread_info *info, struct mdt_file_data *mfd)
2489 {
2490         struct mdt_object *o = mfd->mfd_object;
2491         struct md_object *next = mdt_object_child(o);
2492         struct md_attr *ma = &info->mti_attr;
2493         struct lu_fid *ofid = &info->mti_tmp_fid1;
2494         int rc = 0;
2495         int rc2;
2496         u64 open_flags;
2497         u64 intent;
2498
2499         ENTRY;
2500         open_flags = mfd->mfd_open_flags;
2501         intent = ma->ma_attr_flags & MDS_CLOSE_INTENT;
2502         *ofid = *mdt_object_fid(o);
2503
2504         /* the below message is checked in replay-single.sh test_46 */
2505         CDEBUG(D_INODE, "%s: %sclosing file handle "DFID" with intent: %llx\n",
2506                mdt_obd_name(info->mti_mdt),
2507                ma->ma_valid & MA_FORCE_LOG ? "force " : "", PFID(ofid), intent);
2508
2509         switch (intent) {
2510         case MDS_HSM_RELEASE: {
2511                 rc = mdt_hsm_release(info, o, ma);
2512                 if (rc < 0) {
2513                         CDEBUG(D_HSM, "%s: File " DFID " release failed: %d\n",
2514                                mdt_obd_name(info->mti_mdt),
2515                                PFID(ofid), rc);
2516                         /* continue to close even error occurred. */
2517                 }
2518                 break;
2519         }
2520         case MDS_CLOSE_LAYOUT_MERGE:
2521         case MDS_CLOSE_LAYOUT_SPLIT:
2522         case MDS_CLOSE_LAYOUT_SWAP: {
2523                 rc = mdt_close_handle_layouts(info, o, ma);
2524                 if (rc < 0) {
2525                         CDEBUG(D_INODE,
2526                                "%s: cannot %s layout of "DFID": rc = %d\n",
2527                                mdt_obd_name(info->mti_mdt),
2528                                intent == MDS_CLOSE_LAYOUT_MERGE ? "merge" :
2529                                intent == MDS_CLOSE_LAYOUT_SPLIT ? "split" :
2530                                "swap",
2531                                PFID(ofid), rc);
2532                         /* continue to close even if error occurred. */
2533                 }
2534                 break;
2535         }
2536         case MDS_CLOSE_RESYNC_DONE:
2537                 rc = mdt_close_resync_done(info, o, ma);
2538                 if (rc < 0) {
2539                         CDEBUG(D_INODE,
2540                                "%s: cannot resync layout of "DFID": rc = %d\n",
2541                                mdt_obd_name(info->mti_mdt),
2542                                PFID(ofid), rc);
2543                         /* continue to close even if error occurred. */
2544                 }
2545                 break;
2546         default:
2547                 /* nothing */
2548                 break;
2549         }
2550
2551         if (S_ISREG(lu_object_attr(&o->mot_obj)) &&
2552             ma->ma_attr.la_valid & (LA_LSIZE | LA_LBLOCKS)) {
2553                 rc2 = mdt_lsom_update(info, o, false);
2554                 if (rc2 < 0) {
2555                         CDEBUG(D_INODE,
2556                                "%s: File " DFID " LSOM failed: rc = %d\n",
2557                                mdt_obd_name(info->mti_mdt),
2558                                PFID(ofid), rc2);
2559                         if (rc == 0)
2560                                 rc = rc2;
2561                         /* continue to close even if error occurred. */
2562                 }
2563         }
2564
2565         if (open_flags & MDS_FMODE_WRITE)
2566                 mdt_write_put(o);
2567         else if (open_flags & MDS_FMODE_EXEC)
2568                 mdt_write_allow(o);
2569
2570         /* Update atime|mtime|ctime on close. */
2571         if ((open_flags & MDS_FMODE_EXEC || open_flags & MDS_FMODE_READ ||
2572              open_flags & MDS_FMODE_WRITE) && (ma->ma_valid & MA_INODE) &&
2573             (ma->ma_attr.la_valid & LA_ATIME ||
2574              ma->ma_attr.la_valid & LA_MTIME ||
2575              ma->ma_attr.la_valid & LA_CTIME)) {
2576                 ma->ma_valid = MA_INODE;
2577                 ma->ma_attr_flags |= MDS_CLOSE_UPDATE_TIMES;
2578                 ma->ma_attr.la_valid &= (LA_ATIME | LA_MTIME | LA_CTIME);
2579
2580                 if (ma->ma_attr.la_valid & LA_MTIME) {
2581                         if (mdt_attr_get_pfid(info, o, &ma->ma_pfid) == 0)
2582                                 ma->ma_valid |= MA_PFID;
2583                 }
2584
2585                 rc2 = mo_attr_set(info->mti_env, next, ma);
2586                 if (rc2 != 0) {
2587                         CDEBUG(D_INODE,
2588                            "%s: File "DFID" set attr (%#llx) failed: rc = %d\n",
2589                                mdt_obd_name(info->mti_mdt), PFID(ofid),
2590                                ma->ma_attr.la_valid, rc2);
2591                         if (rc == 0)
2592                                 rc = rc2;
2593                 }
2594         }
2595
2596         /* If file data is modified, add the dirty flag. */
2597         if (ma->ma_attr_flags & MDS_DATA_MODIFIED) {
2598                 rc2 = mdt_add_dirty_flag(info, o, ma);
2599                 if (rc2 != 0) {
2600                         CDEBUG(D_INODE,
2601                              "%s: File "DFID" add dirty flag failed: rc = %d\n",
2602                                mdt_obd_name(info->mti_mdt), PFID(ofid), rc2);
2603                         if (rc == 0)
2604                                 rc = rc2;
2605                 }
2606         }
2607
2608         ma->ma_need |= MA_INODE;
2609         ma->ma_valid &= ~MA_INODE;
2610
2611         LASSERT(atomic_read(&o->mot_open_count) > 0);
2612         atomic_dec(&o->mot_open_count);
2613         mdt_handle_last_unlink(info, o, ma);
2614
2615         if (!MFD_CLOSED(open_flags)) {
2616                 rc2 = mo_close(info->mti_env, next, ma, open_flags);
2617                 if (rc2 != 0) {
2618                         CDEBUG(D_INODE,
2619                                "%s: File "DFID" close failed: rc = %d\n",
2620                                mdt_obd_name(info->mti_mdt), PFID(ofid), rc2);
2621                         if (rc == 0)
2622                                 rc = rc2;
2623                 }
2624                 if (mdt_dom_check_for_discard(info, o))
2625                         mdt_dom_discard_data(info, o);
2626         }
2627
2628         /* adjust open and lease count */
2629         if (open_flags & MDS_OPEN_LEASE) {
2630                 LASSERT(atomic_read(&o->mot_lease_count) > 0);
2631                 atomic_dec(&o->mot_lease_count);
2632         }
2633
2634         mdt_mfd_free(mfd);
2635         mdt_object_put(info->mti_env, o);
2636
2637         RETURN(rc);
2638 }
2639
2640 int mdt_close_internal(struct mdt_thread_info *info, struct ptlrpc_request *req,
2641                        struct mdt_body *repbody)
2642 {
2643         struct mdt_export_data *med;
2644         struct mdt_file_data   *mfd;
2645         int ret = 0;
2646         int rc = 0;
2647
2648         ENTRY;
2649         med = &req->rq_export->exp_mdt_data;
2650         spin_lock(&med->med_open_lock);
2651         mfd = mdt_open_handle2mfd(med, &info->mti_open_handle,
2652                                   req_is_replay(req));
2653         if (mdt_mfd_closed(mfd)) {
2654                 spin_unlock(&med->med_open_lock);
2655                 CDEBUG(D_INODE, "no handle for file close: fid = "DFID
2656                        ": cookie = %#llx\n", PFID(info->mti_rr.rr_fid1),
2657                        info->mti_open_handle.cookie);
2658                 /** not serious error since bug 3633 */
2659                 rc = -ESTALE;
2660         } else {
2661                 class_handle_unhash(&mfd->mfd_open_handle);
2662                 list_del_init(&mfd->mfd_list);
2663                 spin_unlock(&med->med_open_lock);
2664                 ret = mdt_mfd_close(info, mfd);
2665         }
2666
2667         RETURN(rc ? rc : ret);
2668 }
2669
2670 int mdt_close(struct tgt_session_info *tsi)
2671 {
2672         struct mdt_thread_info *info = tsi2mdt_info(tsi);
2673         struct ptlrpc_request *req = tgt_ses_req(tsi);
2674         struct md_attr *ma = &info->mti_attr;
2675         struct mdt_body *repbody = NULL;
2676         ktime_t kstart = ktime_get();
2677         int rc;
2678         int rc2;
2679
2680         ENTRY;
2681         /* Close may come with the Size-on-MDS update. Unpack it. */
2682         rc = mdt_close_unpack(info);
2683         if (rc)
2684                 GOTO(out, rc = err_serious(rc));
2685
2686         /* These fields are no longer used and are left for compatibility.
2687          * size is always zero */
2688         req_capsule_set_size(info->mti_pill, &RMF_MDT_MD, RCL_SERVER,
2689                              0);
2690         req_capsule_set_size(info->mti_pill, &RMF_LOGCOOKIES, RCL_SERVER,
2691                              0);
2692         rc = req_capsule_server_pack(info->mti_pill);
2693         if (mdt_check_resent(info, mdt_reconstruct_generic, NULL)) {
2694                 mdt_client_compatibility(info);
2695                 if (rc == 0)
2696                         mdt_fix_reply(info);
2697                 mdt_exit_ucred(info);
2698                 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2699         }
2700
2701         /* Continue to close handle even if we can not pack reply */
2702         if (rc == 0) {
2703                 repbody = req_capsule_server_get(info->mti_pill,
2704                                                  &RMF_MDT_BODY);
2705                 ma->ma_lmm = req_capsule_server_get(info->mti_pill,
2706                                                     &RMF_MDT_MD);
2707                 ma->ma_lmm_size = req_capsule_get_size(info->mti_pill,
2708                                                        &RMF_MDT_MD,
2709                                                        RCL_SERVER);
2710                 ma->ma_need = MA_INODE | MA_LOV;
2711                 repbody->mbo_eadatasize = 0;
2712                 repbody->mbo_aclsize = 0;
2713         } else {
2714                 rc = err_serious(rc);
2715         }
2716
2717         rc2 = mdt_close_internal(info, req, repbody);
2718         if (rc2 != -ESTALE)
2719                 mdt_empty_transno(info, rc2);
2720         if (rc2 != 0 && rc == 0)
2721                 rc = rc2;
2722
2723         if (repbody != NULL) {
2724                 mdt_client_compatibility(info);
2725                 rc2 = mdt_fix_reply(info);
2726                 if (rc2 != 0 && rc == 0)
2727                         rc = rc2;
2728         }
2729
2730         mdt_exit_ucred(info);
2731         if (CFS_FAIL_CHECK(OBD_FAIL_MDS_CLOSE_PACK))
2732                 GOTO(out, rc = err_serious(-ENOMEM));
2733
2734         if (CFS_FAIL_CHECK_RESET(OBD_FAIL_MDS_CLOSE_NET_REP,
2735                                  OBD_FAIL_MDS_CLOSE_NET_REP))
2736                 tsi->tsi_reply_fail_id = OBD_FAIL_MDS_CLOSE_NET_REP;
2737 out:
2738         mdt_thread_info_fini(info);
2739         if (rc == 0)
2740                 mdt_counter_incr(req, LPROC_MDT_CLOSE,
2741                                  ktime_us_delta(ktime_get(), kstart));
2742         RETURN(rc);
2743 }