lustre/mdt/mdt_reint.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  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/mdt/mdt_reint.c
  33  *
  34  * Lustre Metadata Target (mdt) reintegration routines
  35  *
  36  * Author: Peter Braam <braam@clusterfs.com>
  37  * Author: Andreas Dilger <adilger@clusterfs.com>
  38  * Author: Phil Schwan <phil@clusterfs.com>
  39  * Author: Huang Hua <huanghua@clusterfs.com>
  40  * Author: Yury Umanets <umka@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_MDS
  44
  45 #include <lprocfs_status.h>
  46 #include "mdt_internal.h"
  47 #include <lustre_lmv.h>
  48
  49 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
  50                                      struct md_attr *ma)
  51 {
  52         ma->ma_need = MA_INODE;
  53         ma->ma_valid = 0;
  54 }
  55
  56 /**
  57  * Get version of object by fid.
  58  *
  59  * Return real version or ENOENT_VERSION if object doesn't exist
  60  */
  61 static void mdt_obj_version_get(struct mdt_thread_info *info,
  62                                 struct mdt_object *o, __u64 *version)
  63 {
  64         LASSERT(o);
  65
  66         if (mdt_object_exists(o) && !mdt_object_remote(o) &&
  67             !fid_is_obf(mdt_object_fid(o)))
  68                 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
  69         else
  70                 *version = ENOENT_VERSION;
  71         CDEBUG(D_INODE, "FID "DFID" version is %#llx\n",
  72                PFID(mdt_object_fid(o)), *version);
  73 }
  74
  75 /**
  76  * Check version is correct.
  77  *
  78  * Should be called only during replay.
  79  */
  80 static int mdt_version_check(struct ptlrpc_request *req,
  81                              __u64 version, int idx)
  82 {
  83         __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
  84
  85         ENTRY;
  86         if (!exp_connect_vbr(req->rq_export))
  87                 RETURN(0);
  88
  89         LASSERT(req_is_replay(req));
  90         /** VBR: version is checked always because costs nothing */
  91         LASSERT(idx < PTLRPC_NUM_VERSIONS);
  92         /** Sanity check for malformed buffers */
  93         if (pre_ver == NULL) {
  94                 CERROR("No versions in request buffer\n");
  95                 spin_lock(&req->rq_export->exp_lock);
  96                 req->rq_export->exp_vbr_failed = 1;
  97                 spin_unlock(&req->rq_export->exp_lock);
  98                 RETURN(-EOVERFLOW);
  99         } else if (pre_ver[idx] != version) {
 100                 CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n",
 101                        pre_ver[idx], version);
 102                 spin_lock(&req->rq_export->exp_lock);
 103                 req->rq_export->exp_vbr_failed = 1;
 104                 spin_unlock(&req->rq_export->exp_lock);
 105                 RETURN(-EOVERFLOW);
 106         }
 107         RETURN(0);
 108 }
 109
 110 /**
 111  * Save pre-versions in reply.
 112  */
 113 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
 114                              int idx)
 115 {
 116         __u64 *reply_ver;
 117
 118         if (!exp_connect_vbr(req->rq_export))
 119                 return;
 120
 121         LASSERT(!req_is_replay(req));
 122         LASSERT(req->rq_repmsg != NULL);
 123         reply_ver = lustre_msg_get_versions(req->rq_repmsg);
 124         if (reply_ver)
 125                 reply_ver[idx] = version;
 126 }
 127
 128 /**
 129  * Save enoent version, it is needed when it is obvious that object doesn't
 130  * exist, e.g. child during create.
 131  */
 132 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
 133 {
 134         /* save version of file name for replay, it must be ENOENT here */
 135         if (!req_is_replay(mdt_info_req(info))) {
 136                 info->mti_ver[idx] = ENOENT_VERSION;
 137                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 138         }
 139 }
 140
 141 /**
 142  * Get version from disk and save in reply buffer.
 143  *
 144  * Versions are saved in reply only during normal operations not replays.
 145  */
 146 void mdt_version_get_save(struct mdt_thread_info *info,
 147                           struct mdt_object *mto, int idx)
 148 {
 149         /* don't save versions during replay */
 150         if (!req_is_replay(mdt_info_req(info))) {
 151                 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 152                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 153         }
 154 }
 155
 156 /**
 157  * Get version from disk and check it, no save in reply.
 158  */
 159 int mdt_version_get_check(struct mdt_thread_info *info,
 160                           struct mdt_object *mto, int idx)
 161 {
 162         /* only check versions during replay */
 163         if (!req_is_replay(mdt_info_req(info)))
 164                 return 0;
 165
 166         mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 167         return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
 168 }
 169
 170 /**
 171  * Get version from disk and check if recovery or just save.
 172  */
 173 int mdt_version_get_check_save(struct mdt_thread_info *info,
 174                                struct mdt_object *mto, int idx)
 175 {
 176         int rc = 0;
 177
 178         mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 179         if (req_is_replay(mdt_info_req(info)))
 180                 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
 181                                        idx);
 182         else
 183                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 184         return rc;
 185 }
 186
 187 /**
 188  * Lookup with version checking.
 189  *
 190  * This checks version of 'name'. Many reint functions uses 'name' for child not
 191  * FID, therefore we need to get object by name and check its version.
 192  */
 193 int mdt_lookup_version_check(struct mdt_thread_info *info,
 194                              struct mdt_object *p,
 195                              const struct lu_name *lname,
 196                              struct lu_fid *fid, int idx)
 197 {
 198         int rc, vbrc;
 199
 200         rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
 201                         &info->mti_spec);
 202         /* Check version only during replay */
 203         if (!req_is_replay(mdt_info_req(info)))
 204                 return rc;
 205
 206         info->mti_ver[idx] = ENOENT_VERSION;
 207         if (rc == 0) {
 208                 struct mdt_object *child;
 209
 210                 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
 211                 if (likely(!IS_ERR(child))) {
 212                         mdt_obj_version_get(info, child, &info->mti_ver[idx]);
 213                         mdt_object_put(info->mti_env, child);
 214                 }
 215         }
 216         vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
 217         return vbrc ? vbrc : rc;
 218
 219 }
 220
 221 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
 222                              struct mdt_object *obj,
 223                              struct ldlm_enqueue_info *einfo,
 224                              int decref)
 225 {
 226         union ldlm_policy_data *policy = &mti->mti_policy;
 227         struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
 228         struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
 229         int i;
 230
 231         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 232         LASSERT(slave_locks);
 233
 234         memset(policy, 0, sizeof(*policy));
 235         policy->l_inodebits.bits = einfo->ei_inodebits;
 236         mdt_lock_handle_init(lh);
 237         mdt_lock_reg_init(lh, einfo->ei_mode);
 238         for (i = 0; i < slave_locks->ha_count; i++) {
 239                 if (test_bit(i, (void *)slave_locks->ha_map))
 240                         lh->mlh_rreg_lh = slave_locks->ha_handles[i];
 241                 else
 242                         lh->mlh_reg_lh = slave_locks->ha_handles[i];
 243                 mdt_object_unlock(mti, NULL, lh, decref);
 244                 slave_locks->ha_handles[i].cookie = 0ull;
 245         }
 246
 247         return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
 248                                 policy);
 249 }
 250
 251 static inline int mdt_object_striped(struct mdt_thread_info *mti,
 252                                      struct mdt_object *obj)
 253 {
 254         struct lu_device *bottom_dev;
 255         struct lu_object *bottom_obj;
 256         int rc;
 257
 258         if (!S_ISDIR(obj->mot_header.loh_attr))
 259                 return 0;
 260
 261         /* getxattr from bottom obj to avoid reading in shard FIDs */
 262         bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom);
 263         bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev,
 264                                           mdt_object_fid(obj), NULL);
 265         if (IS_ERR(bottom_obj))
 266                 return PTR_ERR(bottom_obj);
 267
 268         rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL,
 269                           XATTR_NAME_LMV);
 270         lu_object_put(mti->mti_env, bottom_obj);
 271
 272         return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc;
 273 }
 274
 275 /**
 276  * Lock slave stripes if necessary, the lock handles of slave stripes
 277  * will be stored in einfo->ei_cbdata.
 278  **/
 279 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
 280                            enum ldlm_mode mode, __u64 ibits,
 281                            struct ldlm_enqueue_info *einfo)
 282 {
 283         union ldlm_policy_data *policy = &mti->mti_policy;
 284
 285         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 286
 287         einfo->ei_type = LDLM_IBITS;
 288         einfo->ei_mode = mode;
 289         einfo->ei_cb_bl = mdt_remote_blocking_ast;
 290         einfo->ei_cb_local_bl = mdt_blocking_ast;
 291         einfo->ei_cb_cp = ldlm_completion_ast;
 292         einfo->ei_enq_slave = 1;
 293         einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
 294         einfo->ei_inodebits = ibits;
 295         memset(policy, 0, sizeof(*policy));
 296         policy->l_inodebits.bits = ibits;
 297
 298         return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
 299                               policy);
 300 }
 301
 302 int mdt_reint_striped_lock(struct mdt_thread_info *info,
 303                            struct mdt_object *o,
 304                            struct mdt_lock_handle *lh,
 305                            __u64 ibits,
 306                            struct ldlm_enqueue_info *einfo,
 307                            bool cos_incompat)
 308 {
 309         int rc;
 310
 311         LASSERT(!mdt_object_remote(o));
 312
 313         memset(einfo, 0, sizeof(*einfo));
 314
 315         rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
 316         if (rc)
 317                 return rc;
 318
 319         rc = mdt_object_striped(info, o);
 320         if (rc != 1) {
 321                 if (rc < 0)
 322                         mdt_object_unlock(info, o, lh, rc);
 323                 return rc;
 324         }
 325
 326         rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
 327         if (rc) {
 328                 mdt_object_unlock(info, o, lh, rc);
 329                 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
 330                         rc = 0;
 331         }
 332
 333         return rc;
 334 }
 335
 336 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
 337                               struct mdt_object *o,
 338                               struct mdt_lock_handle *lh,
 339                               struct ldlm_enqueue_info *einfo, int decref)
 340 {
 341         if (einfo->ei_cbdata)
 342                 mdt_unlock_slaves(info, o, einfo, decref);
 343         mdt_object_unlock(info, o, lh, decref);
 344 }
 345
 346 static int mdt_restripe(struct mdt_thread_info *info,
 347                         struct mdt_object *parent,
 348                         const struct lu_name *lname,
 349                         const struct lu_fid *tfid,
 350                         struct md_op_spec *spec,
 351                         struct md_attr *ma)
 352 {
 353         struct mdt_device *mdt = info->mti_mdt;
 354         struct lu_fid *fid = &info->mti_tmp_fid2;
 355         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 356         struct lmv_mds_md_v1 *lmv;
 357         struct mdt_object *child;
 358         struct mdt_lock_handle *lhp;
 359         struct mdt_lock_handle *lhc;
 360         struct mdt_body *repbody;
 361         int rc;
 362
 363         ENTRY;
 364         if (!mdt->mdt_enable_dir_restripe)
 365                 RETURN(-EPERM);
 366
 367         rc = mdt_version_get_check_save(info, parent, 0);
 368         if (rc)
 369                 RETURN(rc);
 370
 371         lhp = &info->mti_lh[MDT_LH_PARENT];
 372         mdt_lock_pdo_init(lhp, LCK_PW, lname);
 373         rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
 374                                    true);
 375         if (rc)
 376                 RETURN(rc);
 377
 378         rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV);
 379         if (rc)
 380                 GOTO(unlock_parent, rc);
 381
 382         if (ma->ma_valid & MA_LMV) {
 383                 /* don't allow restripe if parent dir layout is changing */
 384                 lmv = &ma->ma_lmv->lmv_md_v1;
 385                 if (!lmv_is_sane(lmv))
 386                         GOTO(unlock_parent, rc = -EBADF);
 387
 388                 if (lmv_is_layout_changing(lmv))
 389                         GOTO(unlock_parent, rc = -EBUSY);
 390         }
 391
 392         fid_zero(fid);
 393         rc = mdt_lookup_version_check(info, parent, lname, fid, 1);
 394         if (rc)
 395                 GOTO(unlock_parent, rc);
 396
 397         child = mdt_object_find(info->mti_env, mdt, fid);
 398         if (IS_ERR(child))
 399                 GOTO(unlock_parent, rc = PTR_ERR(child));
 400
 401         if (!mdt_object_exists(child))
 402                 GOTO(out_child, rc = -ENOENT);
 403
 404         if (mdt_object_remote(child)) {
 405                 struct mdt_body *repbody;
 406
 407                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 408                 if (!repbody)
 409                         GOTO(out_child, rc = -EPROTO);
 410
 411                 repbody->mbo_fid1 = *fid;
 412                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 413                 GOTO(out_child, rc = -EREMOTE);
 414         }
 415
 416         /* lock object */
 417         lhc = &info->mti_lh[MDT_LH_CHILD];
 418         mdt_lock_reg_init(lhc, LCK_EX);
 419
 420         /* enqueue object remote LOOKUP lock */
 421         if (mdt_object_remote(parent)) {
 422                 rc = mdt_remote_object_lock(info, parent, fid,
 423                                             &lhc->mlh_rreg_lh,
 424                                             lhc->mlh_rreg_mode,
 425                                             MDS_INODELOCK_LOOKUP, false);
 426                 if (rc != ELDLM_OK)
 427                         GOTO(out_child, rc);
 428         }
 429
 430         rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo,
 431                                     true);
 432         if (rc)
 433                 GOTO(unlock_child, rc);
 434
 435         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 436         rc = mdt_version_get_check_save(info, child, 1);
 437         if (rc)
 438                 GOTO(unlock_child, rc);
 439
 440         spin_lock(&mdt->mdt_restriper.mdr_lock);
 441         if (child->mot_restriping) {
 442                 /* race? */
 443                 spin_unlock(&mdt->mdt_restriper.mdr_lock);
 444                 GOTO(unlock_child, rc = -EBUSY);
 445         }
 446         child->mot_restriping = 1;
 447         spin_unlock(&mdt->mdt_restriper.mdr_lock);
 448
 449         *fid = *tfid;
 450         rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
 451         if (rc)
 452                 GOTO(restriping_clear, rc);
 453
 454         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 455         if (!repbody)
 456                 GOTO(restriping_clear, rc = -EPROTO);
 457
 458         mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid);
 459         EXIT;
 460
 461 restriping_clear:
 462         child->mot_restriping = 0;
 463 unlock_child:
 464         mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 465 out_child:
 466         mdt_object_put(info->mti_env, child);
 467 unlock_parent:
 468         mdt_object_unlock(info, parent, lhp, rc);
 469
 470         return rc;
 471 }
 472
 473 /*
 474  * VBR: we save three versions in reply:
 475  * 0 - parent. Check that parent version is the same during replay.
 476  * 1 - name. Version of 'name' if file exists with the same name or
 477  * ENOENT_VERSION, it is needed because file may appear due to missed replays.
 478  * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
 479  * check.
 480  */
 481 static int mdt_create(struct mdt_thread_info *info)
 482 {
 483         struct mdt_device *mdt = info->mti_mdt;
 484         struct mdt_object *parent;
 485         struct mdt_object *child;
 486         struct mdt_lock_handle *lh;
 487         struct mdt_body *repbody;
 488         struct md_attr *ma = &info->mti_attr;
 489         struct mdt_reint_record *rr = &info->mti_rr;
 490         struct md_op_spec *spec = &info->mti_spec;
 491         bool restripe = false;
 492         int rc;
 493
 494         ENTRY;
 495         DEBUG_REQ(D_INODE, mdt_info_req(info),
 496                   "Create ("DNAME"->"DFID") in "DFID,
 497                   PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
 498
 499         if (!fid_is_md_operative(rr->rr_fid1))
 500                 RETURN(-EPERM);
 501
 502         if (S_ISDIR(ma->ma_attr.la_mode) &&
 503             spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
 504                 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
 505                 struct lu_ucred *uc = mdt_ucred(info);
 506                 struct obd_export *exp = mdt_info_req(info)->rq_export;
 507
 508                 /* Only new clients can create remote dir( >= 2.4) and
 509                  * striped dir(>= 2.6), old client will return -ENOTSUPP
 510                  */
 511                 if (!mdt_is_dne_client(exp))
 512                         RETURN(-ENOTSUPP);
 513
 514                 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
 515                         if (!mdt_is_striped_client(exp))
 516                                 RETURN(-ENOTSUPP);
 517
 518                         if (!mdt->mdt_enable_striped_dir)
 519                                 RETURN(-EPERM);
 520                 } else if (!mdt->mdt_enable_remote_dir) {
 521                         RETURN(-EPERM);
 522                 }
 523
 524                 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
 525                     (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) ==
 526                     LMV_HASH_TYPE_CRUSH)
 527                         RETURN(-EPROTO);
 528
 529                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 530                     uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 531                     mdt->mdt_enable_remote_dir_gid != -1)
 532                         RETURN(-EPERM);
 533
 534                 /* restripe if later found dir exists */
 535                 if (le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT)
 536                         restripe = true;
 537         }
 538
 539         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 540
 541         parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 542         if (IS_ERR(parent))
 543                 RETURN(PTR_ERR(parent));
 544
 545         if (!mdt_object_exists(parent))
 546                 GOTO(put_parent, rc = -ENOENT);
 547
 548         /*
 549          * LU-10235: check if name exists locklessly first to avoid massive
 550          * lock recalls on existing directories.
 551          */
 552         rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
 553                                       &info->mti_tmp_fid1, 1);
 554         if (rc == 0) {
 555                 if (!restripe)
 556                         GOTO(put_parent, rc = -EEXIST);
 557
 558                 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
 559                                   ma);
 560         }
 561
 562         /* -ENOENT is expected here */
 563         if (rc != -ENOENT)
 564                 GOTO(put_parent, rc);
 565
 566         /* save version of file name for replay, it must be ENOENT here */
 567         mdt_enoent_version_save(info, 1);
 568
 569         OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
 570
 571         lh = &info->mti_lh[MDT_LH_PARENT];
 572         mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 573         rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
 574         if (rc)
 575                 GOTO(put_parent, rc);
 576
 577         if (!mdt_object_remote(parent)) {
 578                 rc = mdt_version_get_check_save(info, parent, 0);
 579                 if (rc)
 580                         GOTO(unlock_parent, rc);
 581         }
 582
 583         child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
 584         if (unlikely(IS_ERR(child)))
 585                 GOTO(unlock_parent, rc = PTR_ERR(child));
 586
 587         ma->ma_need = MA_INODE;
 588         ma->ma_valid = 0;
 589
 590         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 591                         OBD_FAIL_MDS_REINT_CREATE_WRITE);
 592
 593         /* Version of child will be updated on disk. */
 594         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 595         rc = mdt_version_get_check_save(info, child, 2);
 596         if (rc)
 597                 GOTO(put_child, rc);
 598
 599         /* Let lower layer know current lock mode. */
 600         info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
 601
 602         /*
 603          * Do not perform lookup sanity check. We know that name does
 604          * not exist.
 605          */
 606         info->mti_spec.sp_cr_lookup = 0;
 607         info->mti_spec.sp_feat = &dt_directory_features;
 608
 609         rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
 610                         mdt_object_child(child), &info->mti_spec, ma);
 611         if (rc == 0)
 612                 rc = mdt_attr_get_complex(info, child, ma);
 613
 614         if (rc < 0)
 615                 GOTO(put_child, rc);
 616
 617         /*
 618          * On DNE, we need to eliminate dependey between 'mkdir a' and
 619          * 'mkdir a/b' if b is a striped directory, to achieve this, two
 620          * things are done below:
 621          * 1. save child and slaves lock.
 622          * 2. if the child is a striped directory, relock parent so to
 623          *    compare against with COS locks to ensure parent was
 624          *    committed to disk.
 625          */
 626         if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
 627                 struct mdt_lock_handle *lhc;
 628                 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 629                 bool cos_incompat;
 630
 631                 rc = mdt_object_striped(info, child);
 632                 if (rc < 0)
 633                         GOTO(put_child, rc);
 634
 635                 cos_incompat = rc;
 636                 if (cos_incompat) {
 637                         if (!mdt_object_remote(parent)) {
 638                                 mdt_object_unlock(info, parent, lh, 1);
 639                                 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 640                                 rc = mdt_reint_object_lock(info, parent, lh,
 641                                                            MDS_INODELOCK_UPDATE,
 642                                                            true);
 643                                 if (rc)
 644                                         GOTO(put_child, rc);
 645                         }
 646                 }
 647
 648                 lhc = &info->mti_lh[MDT_LH_CHILD];
 649                 mdt_lock_handle_init(lhc);
 650                 mdt_lock_reg_init(lhc, LCK_PW);
 651                 rc = mdt_reint_striped_lock(info, child, lhc,
 652                                             MDS_INODELOCK_UPDATE, einfo,
 653                                             cos_incompat);
 654                 if (rc)
 655                         GOTO(put_child, rc);
 656
 657                 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 658         }
 659
 660         /* Return fid & attr to client. */
 661         if (ma->ma_valid & MA_INODE)
 662                 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
 663                                    mdt_object_fid(child));
 664         EXIT;
 665 put_child:
 666         mdt_object_put(info->mti_env, child);
 667 unlock_parent:
 668         mdt_object_unlock(info, parent, lh, rc);
 669 put_parent:
 670         mdt_object_put(info->mti_env, parent);
 671         return rc;
 672 }
 673
 674 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
 675                         struct md_attr *ma)
 676 {
 677         struct mdt_lock_handle  *lh;
 678         int do_vbr = ma->ma_attr.la_valid &
 679                         (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
 680         __u64 lockpart = MDS_INODELOCK_UPDATE;
 681         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 682         bool cos_incompat;
 683         int rc;
 684
 685         ENTRY;
 686         rc = mdt_object_striped(info, mo);
 687         if (rc < 0)
 688                 RETURN(rc);
 689
 690         cos_incompat = rc;
 691
 692         lh = &info->mti_lh[MDT_LH_PARENT];
 693         mdt_lock_reg_init(lh, LCK_PW);
 694
 695         /* Even though the new MDT will grant PERM lock to the old
 696          * client, but the old client will almost ignore that during
 697          * So it needs to revoke both LOOKUP and PERM lock here, so
 698          * both new and old client can cancel the dcache
 699          */
 700         if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
 701                 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
 702
 703         rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
 704                                     cos_incompat);
 705         if (rc != 0)
 706                 RETURN(rc);
 707
 708         /* all attrs are packed into mti_attr in unpack_setattr */
 709         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 710                        OBD_FAIL_MDS_REINT_SETATTR_WRITE);
 711
 712         /* VBR: update version if attr changed are important for recovery */
 713         if (do_vbr) {
 714                 /* update on-disk version of changed object */
 715                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
 716                 rc = mdt_version_get_check_save(info, mo, 0);
 717                 if (rc)
 718                         GOTO(out_unlock, rc);
 719         }
 720
 721         /* Ensure constant striping during chown(). See LU-2789. */
 722         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 723                 mutex_lock(&mo->mot_lov_mutex);
 724
 725         /* all attrs are packed into mti_attr in unpack_setattr */
 726         rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
 727
 728         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 729                 mutex_unlock(&mo->mot_lov_mutex);
 730
 731         if (rc != 0)
 732                 GOTO(out_unlock, rc);
 733         mdt_dom_obj_lvb_update(info->mti_env, mo, false);
 734         EXIT;
 735 out_unlock:
 736         mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
 737         return rc;
 738 }
 739
 740 /**
 741  * Check HSM flags and add HS_DIRTY flag if relevant.
 742  *
 743  * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
 744  * and is not RELEASED.
 745  */
 746 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
 747                         struct md_attr *ma)
 748 {
 749         struct lu_ucred *uc = mdt_ucred(info);
 750         cfs_cap_t cap_saved;
 751         int rc;
 752
 753         ENTRY;
 754         /* If the file was modified, add the dirty flag */
 755         ma->ma_need = MA_HSM;
 756         rc = mdt_attr_get_complex(info, mo, ma);
 757         if (rc) {
 758                 CERROR("file attribute read error for "DFID": %d.\n",
 759                         PFID(mdt_object_fid(mo)), rc);
 760                 RETURN(rc);
 761         }
 762
 763         /* If an up2date copy exists in the backend, add dirty flag */
 764         if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
 765             && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
 766                 ma->ma_hsm.mh_flags |= HS_DIRTY;
 767
 768                 /* Bump cap so that closes from non-owner writers can
 769                  * set the HSM state to dirty.
 770                  */
 771                 cap_saved = uc->uc_cap;
 772                 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
 773                 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
 774                 uc->uc_cap = cap_saved;
 775                 if (rc)
 776                         CERROR("file attribute change error for "DFID": %d\n",
 777                                 PFID(mdt_object_fid(mo)), rc);
 778         }
 779
 780         RETURN(rc);
 781 }
 782
 783 static int mdt_reint_setattr(struct mdt_thread_info *info,
 784                              struct mdt_lock_handle *lhc)
 785 {
 786         struct mdt_device *mdt = info->mti_mdt;
 787         struct md_attr *ma = &info->mti_attr;
 788         struct mdt_reint_record *rr = &info->mti_rr;
 789         struct ptlrpc_request *req = mdt_info_req(info);
 790         struct mdt_object *mo;
 791         struct mdt_body *repbody;
 792         int rc, rc2;
 793
 794         ENTRY;
 795         DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
 796                   (unsigned int)ma->ma_attr.la_valid);
 797
 798         if (info->mti_dlm_req)
 799                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 800
 801         OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
 802
 803         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 804         mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
 805         if (IS_ERR(mo))
 806                 GOTO(out, rc = PTR_ERR(mo));
 807
 808         if (!mdt_object_exists(mo))
 809                 GOTO(out_put, rc = -ENOENT);
 810
 811         if (mdt_object_remote(mo))
 812                 GOTO(out_put, rc = -EREMOTE);
 813
 814         ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
 815         /* revoke lease lock if size is going to be changed */
 816         if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
 817                      !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
 818                      atomic_read(&mo->mot_lease_count) > 0)) {
 819                 down_read(&mo->mot_open_sem);
 820
 821                 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
 822                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 823                         mdt_lock_reg_init(lhc, LCK_CW);
 824
 825                         rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
 826                         if (rc != 0) {
 827                                 up_read(&mo->mot_open_sem);
 828                                 GOTO(out_put, rc);
 829                         }
 830
 831                         /* revoke lease lock */
 832                         mdt_object_unlock(info, mo, lhc, 1);
 833                 }
 834                 up_read(&mo->mot_open_sem);
 835         }
 836
 837         if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
 838                 /* Check write access for the O_TRUNC case */
 839                 if (mdt_write_read(mo) < 0)
 840                         GOTO(out_put, rc = -ETXTBSY);
 841
 842                 /* LU-10286: compatibility check for FLR.
 843                  * Please check the comment in mdt_finish_open() for details
 844                  */
 845                 if (!exp_connect_flr(info->mti_exp) ||
 846                     !exp_connect_overstriping(info->mti_exp)) {
 847                         rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
 848                         if (rc < 0 && rc != -ENODATA)
 849                                 GOTO(out_put, rc);
 850
 851                         if (!exp_connect_flr(info->mti_exp)) {
 852                                 if (rc > 0 &&
 853                                     mdt_lmm_is_flr(info->mti_big_lmm))
 854                                         GOTO(out_put, rc = -EOPNOTSUPP);
 855                         }
 856
 857                         if (!exp_connect_overstriping(info->mti_exp)) {
 858                                 if (rc > 0 &&
 859                                     mdt_lmm_is_overstriping(info->mti_big_lmm))
 860                                         GOTO(out_put, rc = -EOPNOTSUPP);
 861                         }
 862                 }
 863
 864                 /* For truncate, the file size sent from client
 865                  * is believable, but the blocks are incorrect,
 866                  * which makes the block size in LSOM attribute
 867                  * inconsisent with the real block size.
 868                  */
 869                 rc = mdt_lsom_update(info, mo, true);
 870                 if (rc)
 871                         GOTO(out_put, rc);
 872         }
 873
 874         if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
 875                 if (ma->ma_valid & MA_LOV)
 876                         GOTO(out_put, rc = -EPROTO);
 877
 878                 /* MDT supports FMD for regular files due to Data-on-MDT */
 879                 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
 880                     ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
 881                         tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
 882                                        req->rq_xid);
 883
 884                 rc = mdt_attr_set(info, mo, ma);
 885                 if (rc)
 886                         GOTO(out_put, rc);
 887         } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
 888                    (ma->ma_valid & MA_INODE)) {
 889                 struct lu_buf *buf = &info->mti_buf;
 890                 struct lu_ucred *uc = mdt_ucred(info);
 891                 struct mdt_lock_handle *lh;
 892                 const char *name;
 893                 __u64 lockpart = MDS_INODELOCK_XATTR;
 894
 895                 /* reject if either remote or striped dir is disabled */
 896                 if (ma->ma_valid & MA_LMV) {
 897                         if (!mdt->mdt_enable_remote_dir ||
 898                             !mdt->mdt_enable_striped_dir)
 899                                 GOTO(out_put, rc = -EPERM);
 900
 901                         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 902                             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 903                             mdt->mdt_enable_remote_dir_gid != -1)
 904                                 GOTO(out_put, rc = -EPERM);
 905                 }
 906
 907                 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
 908                         GOTO(out_put, rc = -ENOTDIR);
 909
 910                 if (ma->ma_attr.la_valid != 0)
 911                         GOTO(out_put, rc = -EPROTO);
 912
 913                 if (ma->ma_valid & MA_LOV) {
 914                         buf->lb_buf = ma->ma_lmm;
 915                         buf->lb_len = ma->ma_lmm_size;
 916                         name = XATTR_NAME_LOV;
 917                 } else {
 918                         struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
 919
 920                         buf->lb_buf = lmu;
 921                         buf->lb_len = ma->ma_lmv_size;
 922                         name = XATTR_NAME_DEFAULT_LMV;
 923                         /* force client to update dir default layout */
 924                         lockpart |= MDS_INODELOCK_LOOKUP;
 925                 }
 926
 927                 lh = &info->mti_lh[MDT_LH_PARENT];
 928                 mdt_lock_reg_init(lh, LCK_PW);
 929
 930                 rc = mdt_object_lock(info, mo, lh, lockpart);
 931                 if (rc != 0)
 932                         GOTO(out_put, rc);
 933
 934                 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
 935                                   name, 0);
 936
 937                 mdt_object_unlock(info, mo, lh, rc);
 938                 if (rc)
 939                         GOTO(out_put, rc);
 940         } else {
 941                 GOTO(out_put, rc = -EPROTO);
 942         }
 943
 944         /* If file data is modified, add the dirty flag */
 945         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
 946                 rc = mdt_add_dirty_flag(info, mo, ma);
 947
 948         ma->ma_need = MA_INODE;
 949         ma->ma_valid = 0;
 950         rc = mdt_attr_get_complex(info, mo, ma);
 951         if (rc != 0)
 952                 GOTO(out_put, rc);
 953
 954         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
 955
 956         EXIT;
 957 out_put:
 958         mdt_object_put(info->mti_env, mo);
 959 out:
 960         if (rc == 0)
 961                 mdt_counter_incr(req, LPROC_MDT_SETATTR);
 962
 963         mdt_client_compatibility(info);
 964         rc2 = mdt_fix_reply(info);
 965         if (rc == 0)
 966                 rc = rc2;
 967         return rc;
 968 }
 969
 970 static int mdt_reint_create(struct mdt_thread_info *info,
 971                             struct mdt_lock_handle *lhc)
 972 {
 973         struct ptlrpc_request   *req = mdt_info_req(info);
 974         int                     rc;
 975
 976         ENTRY;
 977         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 978                 RETURN(err_serious(-ESTALE));
 979
 980         if (info->mti_dlm_req)
 981                 ldlm_request_cancel(mdt_info_req(info),
 982                                     info->mti_dlm_req, 0, LATF_SKIP);
 983
 984         if (!lu_name_is_valid(&info->mti_rr.rr_name))
 985                 RETURN(-EPROTO);
 986
 987         switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
 988         case S_IFDIR:
 989                 mdt_counter_incr(req, LPROC_MDT_MKDIR);
 990                 break;
 991         case S_IFREG:
 992         case S_IFLNK:
 993         case S_IFCHR:
 994         case S_IFBLK:
 995         case S_IFIFO:
 996         case S_IFSOCK:
 997                 /* Special file should stay on the same node as parent. */
 998                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
 999                 break;
1000         default:
1001                 CERROR("%s: Unsupported mode %o\n",
1002                        mdt_obd_name(info->mti_mdt),
1003                        info->mti_attr.ma_attr.la_mode);
1004                 RETURN(err_serious(-EOPNOTSUPP));
1005         }
1006
1007         rc = mdt_create(info);
1008         RETURN(rc);
1009 }
1010
1011 /*
1012  * VBR: save parent version in reply and child version getting by its name.
1013  * Version of child is getting and checking during its lookup. If
1014  */
1015 static int mdt_reint_unlink(struct mdt_thread_info *info,
1016                             struct mdt_lock_handle *lhc)
1017 {
1018         struct mdt_reint_record *rr = &info->mti_rr;
1019         struct ptlrpc_request *req = mdt_info_req(info);
1020         struct md_attr *ma = &info->mti_attr;
1021         struct lu_fid *child_fid = &info->mti_tmp_fid1;
1022         struct mdt_object *mp;
1023         struct mdt_object *mc;
1024         struct mdt_lock_handle *parent_lh;
1025         struct mdt_lock_handle *child_lh;
1026         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
1027         __u64 lock_ibits;
1028         bool cos_incompat = false;
1029         int no_name = 0;
1030         int rc;
1031
1032         ENTRY;
1033         DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1034                   PNAME(&rr->rr_name));
1035
1036         if (info->mti_dlm_req)
1037                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1038
1039         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1040                 RETURN(err_serious(-ENOENT));
1041
1042         if (!fid_is_md_operative(rr->rr_fid1))
1043                 RETURN(-EPERM);
1044
1045         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1046         if (IS_ERR(mp))
1047                 RETURN(PTR_ERR(mp));
1048
1049         if (mdt_object_remote(mp)) {
1050                 cos_incompat = true;
1051         } else {
1052                 rc = mdt_version_get_check_save(info, mp, 0);
1053                 if (rc)
1054                         GOTO(put_parent, rc);
1055         }
1056
1057 relock:
1058         parent_lh = &info->mti_lh[MDT_LH_PARENT];
1059         mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
1060         rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
1061                                    cos_incompat);
1062         if (rc != 0)
1063                 GOTO(put_parent, rc);
1064
1065         /* lookup child object along with version checking */
1066         fid_zero(child_fid);
1067         rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
1068         if (rc != 0) {
1069                 /* Name might not be able to find during resend of
1070                  * remote unlink, considering following case.
1071                  * dir_A is a remote directory, the name entry of
1072                  * dir_A is on MDT0, the directory is on MDT1,
1073                  *
1074                  * 1. client sends unlink req to MDT1.
1075                  * 2. MDT1 sends name delete update to MDT0.
1076                  * 3. name entry is being deleted in MDT0 synchronously.
1077                  * 4. MDT1 is restarted.
1078                  * 5. client resends unlink req to MDT1. So it can not
1079                  *    find the name entry on MDT0 anymore.
1080                  * In this case, MDT1 only needs to destory the local
1081                  * directory.
1082                  */
1083                 if (mdt_object_remote(mp) && rc == -ENOENT &&
1084                     !fid_is_zero(rr->rr_fid2) &&
1085                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1086                         no_name = 1;
1087                         *child_fid = *rr->rr_fid2;
1088                 } else {
1089                         GOTO(unlock_parent, rc);
1090                 }
1091         }
1092
1093         if (!fid_is_md_operative(child_fid))
1094                 GOTO(unlock_parent, rc = -EPERM);
1095
1096         /* We will lock the child regardless it is local or remote. No harm. */
1097         mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1098         if (IS_ERR(mc))
1099                 GOTO(unlock_parent, rc = PTR_ERR(mc));
1100
1101         if (!cos_incompat) {
1102                 rc = mdt_object_striped(info, mc);
1103                 if (rc < 0)
1104                         GOTO(put_child, rc);
1105
1106                 cos_incompat = rc;
1107                 if (cos_incompat) {
1108                         mdt_object_put(info->mti_env, mc);
1109                         mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1110                         goto relock;
1111                 }
1112         }
1113
1114         child_lh = &info->mti_lh[MDT_LH_CHILD];
1115         mdt_lock_reg_init(child_lh, LCK_EX);
1116         if (info->mti_spec.sp_rm_entry) {
1117                 struct lu_ucred *uc  = mdt_ucred(info);
1118
1119                 if (!mdt_is_dne_client(req->rq_export))
1120                         /* Return -ENOTSUPP for old client */
1121                         GOTO(put_child, rc = -ENOTSUPP);
1122
1123                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
1124                         GOTO(put_child, rc = -EPERM);
1125
1126                 ma->ma_need = MA_INODE;
1127                 ma->ma_valid = 0;
1128                 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1129                                 NULL, &rr->rr_name, ma, no_name);
1130                 GOTO(put_child, rc);
1131         }
1132
1133         if (mdt_object_remote(mc)) {
1134                 struct mdt_body  *repbody;
1135
1136                 if (!fid_is_zero(rr->rr_fid2)) {
1137                         CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1138                                mdt_obd_name(info->mti_mdt),
1139                                PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1140                         GOTO(put_child, rc = -ENOENT);
1141                 }
1142                 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1143                        mdt_obd_name(info->mti_mdt),
1144                        PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1145
1146                 if (!mdt_is_dne_client(req->rq_export))
1147                         /* Return -ENOTSUPP for old client */
1148                         GOTO(put_child, rc = -ENOTSUPP);
1149
1150                 /* Revoke the LOOKUP lock of the remote object granted by
1151                  * this MDT. Since the unlink will happen on another MDT,
1152                  * it will release the LOOKUP lock right away. Then What
1153                  * would happen if another client try to grab the LOOKUP
1154                  * lock at the same time with unlink XXX
1155                  */
1156                 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1157                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1158                 LASSERT(repbody != NULL);
1159                 repbody->mbo_fid1 = *mdt_object_fid(mc);
1160                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1161                 GOTO(unlock_child, rc = -EREMOTE);
1162         }
1163         /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1164          * this now because a running HSM restore on the child (unlink
1165          * victim) will hold the layout lock. See LU-4002.
1166          */
1167         lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1168         if (mdt_object_remote(mp)) {
1169                 /* Enqueue lookup lock from parent MDT */
1170                 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1171                                             &child_lh->mlh_rreg_lh,
1172                                             child_lh->mlh_rreg_mode,
1173                                             MDS_INODELOCK_LOOKUP, false);
1174                 if (rc != ELDLM_OK)
1175                         GOTO(put_child, rc);
1176
1177                 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1178         }
1179
1180         rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1181                                     cos_incompat);
1182         if (rc != 0)
1183                 GOTO(put_child, rc);
1184
1185         /*
1186          * Now we can only make sure we need MA_INODE, in mdd layer, will check
1187          * whether need MA_LOV and MA_COOKIE.
1188          */
1189         ma->ma_need = MA_INODE;
1190         ma->ma_valid = 0;
1191
1192         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1193                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1194         /* save version when object is locked */
1195         mdt_version_get_save(info, mc, 1);
1196
1197         mutex_lock(&mc->mot_lov_mutex);
1198
1199         rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1200                         mdt_object_child(mc), &rr->rr_name, ma, no_name);
1201
1202         mutex_unlock(&mc->mot_lov_mutex);
1203         if (rc != 0)
1204                 GOTO(unlock_child, rc);
1205
1206         if (!lu_object_is_dying(&mc->mot_header)) {
1207                 rc = mdt_attr_get_complex(info, mc, ma);
1208                 if (rc)
1209                         GOTO(out_stat, rc);
1210         } else if (mdt_dom_check_for_discard(info, mc)) {
1211                 mdt_dom_discard_data(info, mc);
1212         }
1213         mdt_handle_last_unlink(info, mc, ma);
1214
1215 out_stat:
1216         if (ma->ma_valid & MA_INODE) {
1217                 switch (ma->ma_attr.la_mode & S_IFMT) {
1218                 case S_IFDIR:
1219                         mdt_counter_incr(req, LPROC_MDT_RMDIR);
1220                         break;
1221                 case S_IFREG:
1222                 case S_IFLNK:
1223                 case S_IFCHR:
1224                 case S_IFBLK:
1225                 case S_IFIFO:
1226                 case S_IFSOCK:
1227                         mdt_counter_incr(req, LPROC_MDT_UNLINK);
1228                         break;
1229                 default:
1230                         LASSERTF(0, "bad file type %o unlinking\n",
1231                                 ma->ma_attr.la_mode);
1232                 }
1233         }
1234
1235         EXIT;
1236
1237 unlock_child:
1238         mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1239 put_child:
1240         mdt_object_put(info->mti_env, mc);
1241 unlock_parent:
1242         mdt_object_unlock(info, mp, parent_lh, rc);
1243 put_parent:
1244         mdt_object_put(info->mti_env, mp);
1245         CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1246         return rc;
1247 }
1248
1249 /*
1250  * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1251  * name.
1252  */
1253 static int mdt_reint_link(struct mdt_thread_info *info,
1254                           struct mdt_lock_handle *lhc)
1255 {
1256         struct mdt_reint_record *rr = &info->mti_rr;
1257         struct ptlrpc_request   *req = mdt_info_req(info);
1258         struct md_attr          *ma = &info->mti_attr;
1259         struct mdt_object       *ms;
1260         struct mdt_object       *mp;
1261         struct mdt_lock_handle  *lhs;
1262         struct mdt_lock_handle  *lhp;
1263         bool cos_incompat;
1264         int rc;
1265
1266         ENTRY;
1267         DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1268                   PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1269
1270         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1271                 RETURN(err_serious(-ENOENT));
1272
1273         if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1274                 req->rq_no_reply = 1;
1275                 RETURN(err_serious(-ENOENT));
1276         }
1277
1278         if (info->mti_dlm_req)
1279                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1280
1281         /* Invalid case so return error immediately instead of
1282          * processing it
1283          */
1284         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1285                 RETURN(-EPERM);
1286
1287         if (!fid_is_md_operative(rr->rr_fid1) ||
1288             !fid_is_md_operative(rr->rr_fid2))
1289                 RETURN(-EPERM);
1290
1291         /* step 1: find target parent dir */
1292         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1293         if (IS_ERR(mp))
1294                 RETURN(PTR_ERR(mp));
1295
1296         rc = mdt_version_get_check_save(info, mp, 0);
1297         if (rc)
1298                 GOTO(put_parent, rc);
1299
1300         /* step 2: find source */
1301         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1302         if (IS_ERR(ms))
1303                 GOTO(put_parent, rc = PTR_ERR(ms));
1304
1305         if (!mdt_object_exists(ms)) {
1306                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1307                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1308                 GOTO(put_source, rc = -ENOENT);
1309         }
1310
1311         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1312
1313         lhp = &info->mti_lh[MDT_LH_PARENT];
1314         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1315         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1316                                    cos_incompat);
1317         if (rc != 0)
1318                 GOTO(put_source, rc);
1319
1320         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1321
1322         lhs = &info->mti_lh[MDT_LH_CHILD];
1323         mdt_lock_reg_init(lhs, LCK_EX);
1324         rc = mdt_reint_object_lock(info, ms, lhs,
1325                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1326                                    cos_incompat);
1327         if (rc != 0)
1328                 GOTO(unlock_parent, rc);
1329
1330         /* step 3: link it */
1331         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1332                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1333
1334         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1335         rc = mdt_version_get_check_save(info, ms, 1);
1336         if (rc)
1337                 GOTO(unlock_source, rc);
1338
1339         /** check target version by name during replay */
1340         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1341                                       &info->mti_tmp_fid1, 2);
1342         if (rc != 0 && rc != -ENOENT)
1343                 GOTO(unlock_source, rc);
1344         /* save version of file name for replay, it must be ENOENT here */
1345         if (!req_is_replay(mdt_info_req(info))) {
1346                 if (rc != -ENOENT) {
1347                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1348                                PNAME(&rr->rr_name));
1349                         GOTO(unlock_source, rc = -EEXIST);
1350                 }
1351                 info->mti_ver[2] = ENOENT_VERSION;
1352                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1353         }
1354
1355         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1356                       mdt_object_child(ms), &rr->rr_name, ma);
1357
1358         if (rc == 0)
1359                 mdt_counter_incr(req, LPROC_MDT_LINK);
1360
1361         EXIT;
1362 unlock_source:
1363         mdt_object_unlock(info, ms, lhs, rc);
1364 unlock_parent:
1365         mdt_object_unlock(info, mp, lhp, rc);
1366 put_source:
1367         mdt_object_put(info->mti_env, ms);
1368 put_parent:
1369         mdt_object_put(info->mti_env, mp);
1370         return rc;
1371 }
1372 /**
1373  * lock the part of the directory according to the hash of the name
1374  * (lh->mlh_pdo_hash) in parallel directory lock.
1375  */
1376 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1377                               struct mdt_lock_handle *lh,
1378                               struct mdt_object *obj, __u64 ibits,
1379                               bool cos_incompat)
1380 {
1381         struct ldlm_res_id *res = &info->mti_res_id;
1382         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1383         union ldlm_policy_data *policy = &info->mti_policy;
1384         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1385         int rc;
1386
1387         /*
1388          * Finish res_id initializing by name hash marking part of
1389          * directory which is taking modification.
1390          */
1391         LASSERT(lh->mlh_pdo_hash != 0);
1392         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1393         memset(policy, 0, sizeof(*policy));
1394         policy->l_inodebits.bits = ibits;
1395         if (cos_incompat &&
1396             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1397                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1398         /*
1399          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1400          * going to be sent to client. If it is - mdt_intent_policy() path will
1401          * fix it up and turn FL_LOCAL flag off.
1402          */
1403         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1404                           policy, res, dlmflags,
1405                           &info->mti_exp->exp_handle.h_cookie);
1406         return rc;
1407 }
1408
1409 /**
1410  * Get BFL lock for rename or migrate process.
1411  **/
1412 static int mdt_rename_lock(struct mdt_thread_info *info,
1413                            struct lustre_handle *lh)
1414 {
1415         int     rc;
1416
1417         ENTRY;
1418         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1419                 struct lu_fid *fid = &info->mti_tmp_fid1;
1420                 struct mdt_object *obj;
1421
1422                 /* XXX, right now, it has to use object API to
1423                  * enqueue lock cross MDT, so it will enqueue
1424                  * rename lock(with LUSTRE_BFL_FID) by root object
1425                  */
1426                 lu_root_fid(fid);
1427                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1428                 if (IS_ERR(obj))
1429                         RETURN(PTR_ERR(obj));
1430
1431                 rc = mdt_remote_object_lock(info, obj,
1432                                             &LUSTRE_BFL_FID, lh,
1433                                             LCK_EX,
1434                                             MDS_INODELOCK_UPDATE, false);
1435                 mdt_object_put(info->mti_env, obj);
1436         } else {
1437                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1438                 union ldlm_policy_data *policy = &info->mti_policy;
1439                 struct ldlm_res_id *res_id = &info->mti_res_id;
1440                 __u64 flags = 0;
1441
1442                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1443                 memset(policy, 0, sizeof(*policy));
1444                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1445                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1446                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1447                                             LDLM_IBITS, policy, LCK_EX, &flags,
1448                                             ldlm_blocking_ast,
1449                                             ldlm_completion_ast, NULL, NULL, 0,
1450                                             LVB_T_NONE,
1451                                             &info->mti_exp->exp_handle.h_cookie,
1452                                             lh);
1453                 RETURN(rc);
1454         }
1455         RETURN(rc);
1456 }
1457
1458 static void mdt_rename_unlock(struct lustre_handle *lh)
1459 {
1460         ENTRY;
1461         LASSERT(lustre_handle_is_used(lh));
1462         /* Cancel the single rename lock right away */
1463         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1464         EXIT;
1465 }
1466
1467 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1468                                                 const struct lu_fid *fid,
1469                                                 int idx)
1470 {
1471         struct mdt_object *dir;
1472         int rc;
1473
1474         ENTRY;
1475         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1476         if (IS_ERR(dir))
1477                 RETURN(dir);
1478
1479         /* check early, the real version will be saved after locking */
1480         rc = mdt_version_get_check(info, dir, idx);
1481         if (rc)
1482                 GOTO(out_put, rc);
1483
1484         if (!mdt_object_exists(dir))
1485                 GOTO(out_put, rc = -ENOENT);
1486
1487         if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1488                 GOTO(out_put, rc = -ENOTDIR);
1489
1490         RETURN(dir);
1491 out_put:
1492         mdt_object_put(info->mti_env, dir);
1493         return ERR_PTR(rc);
1494 }
1495
1496 /*
1497  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1498  * herein we don't really check it, just do revoke.
1499  */
1500 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1501                                   struct mdt_object *pobj,
1502                                   struct mdt_object *obj)
1503 {
1504         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1505         int rc;
1506
1507         mdt_lock_handle_init(lh);
1508         mdt_lock_reg_init(lh, LCK_EX);
1509
1510         if (mdt_object_remote(pobj)) {
1511                 /* don't bother to check if pobj and obj are on the same MDT. */
1512                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1513                                             &lh->mlh_rreg_lh, LCK_EX,
1514                                             MDS_INODELOCK_LOOKUP, false);
1515         } else if (mdt_object_remote(obj)) {
1516                 struct ldlm_res_id *res = &info->mti_res_id;
1517                 union ldlm_policy_data *policy = &info->mti_policy;
1518                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1519                                  LDLM_FL_COS_INCOMPAT;
1520
1521                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1522                 memset(policy, 0, sizeof(*policy));
1523                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1524                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1525                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1526                                   dlmflags, NULL);
1527         } else {
1528                 /* do nothing if both are local */
1529                 return 0;
1530         }
1531
1532         if (rc != ELDLM_OK)
1533                 return rc;
1534
1535         /*
1536          * TODO, currently we don't save this lock because there is no place to
1537          * hold this lock handle, but to avoid race we need to save this lock.
1538          */
1539         mdt_object_unlock(info, NULL, lh, 1);
1540
1541         return 0;
1542 }
1543
1544 /*
1545  * operation may takes locks of linkea, or directory stripes, group them in
1546  * different list.
1547  */
1548 struct mdt_sub_lock {
1549         struct mdt_object *msl_obj;
1550         struct mdt_lock_handle msl_lh;
1551         struct list_head msl_linkage;
1552 };
1553
1554 static void mdt_unlock_list(struct mdt_thread_info *info,
1555                             struct list_head *list, int decref)
1556 {
1557         struct mdt_sub_lock *msl;
1558         struct mdt_sub_lock *tmp;
1559
1560         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1561                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1562                 list_del(&msl->msl_linkage);
1563                 OBD_FREE_PTR(msl);
1564         }
1565 }
1566
1567 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1568                                              struct mdt_object *obj,
1569                                              struct mdt_lock_handle *lh,
1570                                              struct ldlm_enqueue_info *einfo,
1571                                              struct list_head *slave_locks,
1572                                              int decref)
1573 {
1574         if (mdt_object_remote(obj)) {
1575                 mdt_unlock_list(info, slave_locks, decref);
1576                 mdt_object_unlock(info, obj, lh, decref);
1577         } else {
1578                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1579         }
1580 }
1581
1582 /*
1583  * lock parents of links, and also check whether total locks don't exceed
1584  * RS_MAX_LOCKS.
1585  *
1586  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1587  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1588  * \retval      -ev negative errno upon error
1589  */
1590 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1591                                  struct mdt_object *pobj,
1592                                  const struct md_attr *ma,
1593                                  struct mdt_object *obj,
1594                                  struct mdt_lock_handle *lhp,
1595                                  struct ldlm_enqueue_info *peinfo,
1596                                  struct list_head *parent_slave_locks,
1597                                  struct list_head *link_locks)
1598 {
1599         struct mdt_device *mdt = info->mti_mdt;
1600         struct lu_buf *buf = &info->mti_big_buf;
1601         struct lu_name *lname = &info->mti_name;
1602         struct linkea_data ldata = { NULL };
1603         bool blocked = false;
1604         int local_lnkp_cnt = 0;
1605         int rc;
1606
1607         ENTRY;
1608         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1609                 RETURN(0);
1610
1611         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1612         if (buf->lb_buf == NULL)
1613                 RETURN(-ENOMEM);
1614
1615         ldata.ld_buf = buf;
1616         rc = mdt_links_read(info, obj, &ldata);
1617         if (rc) {
1618                 if (rc == -ENOENT || rc == -ENODATA)
1619                         rc = 0;
1620                 RETURN(rc);
1621         }
1622
1623         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1624              linkea_next_entry(&ldata)) {
1625                 struct mdt_object *lnkp;
1626                 struct mdt_sub_lock *msl;
1627                 struct lu_fid fid;
1628                 __u64 ibits;
1629
1630                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1631                                     &fid);
1632
1633                 /* check if it's also linked to parent */
1634                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1635                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1636                                PFID(&fid), PNAME(lname));
1637                         /* in case link is remote object, revoke LOOKUP lock */
1638                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1639                         continue;
1640                 }
1641
1642                 lnkp = NULL;
1643
1644                 /* check if it's linked to a stripe of parent */
1645                 if (ma->ma_valid & MA_LMV) {
1646                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1647                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1648                         int j = 0;
1649
1650                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1651                                 fid_le_to_cpu(stripe_fid,
1652                                               &lmv->lmv_stripe_fids[j]);
1653                                 if (lu_fid_eq(stripe_fid, &fid)) {
1654                                         CDEBUG(D_INFO, "skip stripe "DFID
1655                                                ", reovke "DNAME"\n",
1656                                                PFID(&fid), PNAME(lname));
1657                                         lnkp = mdt_object_find(info->mti_env,
1658                                                                mdt, &fid);
1659                                         if (IS_ERR(lnkp))
1660                                                 GOTO(out, rc = PTR_ERR(lnkp));
1661                                         break;
1662                                 }
1663                         }
1664
1665                         if (lnkp) {
1666                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1667                                                                    obj);
1668                                 mdt_object_put(info->mti_env, lnkp);
1669                                 continue;
1670                         }
1671                 }
1672
1673                 /* Check if it's already locked */
1674                 list_for_each_entry(msl, link_locks, msl_linkage) {
1675                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1676                                 CDEBUG(D_INFO,
1677                                        DFID" was locked, revoke "DNAME"\n",
1678                                        PFID(&fid), PNAME(lname));
1679                                 lnkp = msl->msl_obj;
1680                                 break;
1681                         }
1682                 }
1683
1684                 if (lnkp) {
1685                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1686                         continue;
1687                 }
1688
1689                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1690                        PFID(&fid), PNAME(lname));
1691
1692                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1693                 if (IS_ERR(lnkp)) {
1694                         CWARN("%s: cannot find obj "DFID": %ld\n",
1695                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1696                         continue;
1697                 }
1698
1699                 if (!mdt_object_exists(lnkp)) {
1700                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1701                               PFID(&fid), PNAME(lname));
1702                         mdt_object_put(info->mti_env, lnkp);
1703                         continue;
1704                 }
1705
1706                 if (!mdt_object_remote(lnkp))
1707                         local_lnkp_cnt++;
1708
1709                 OBD_ALLOC_PTR(msl);
1710                 if (msl == NULL)
1711                         GOTO(out, rc = -ENOMEM);
1712
1713                 /*
1714                  * we can't follow parent-child lock order like other MD
1715                  * operations, use lock_try here to avoid deadlock, if the lock
1716                  * cannot be taken, drop all locks taken, revoke the blocked
1717                  * one, and continue processing the remaining entries, and in
1718                  * the end of the loop restart from beginning.
1719                  */
1720                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1721                 ibits = 0;
1722                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1723                                          MDS_INODELOCK_UPDATE, true);
1724                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1725
1726                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1727                                PFID(&fid), PNAME(lname));
1728
1729                         mdt_unlock_list(info, link_locks, 1);
1730                         /* also unlock parent locks to avoid deadlock */
1731                         if (!blocked)
1732                                 mdt_migrate_object_unlock(info, pobj, lhp,
1733                                                           peinfo,
1734                                                           parent_slave_locks,
1735                                                           1);
1736
1737                         blocked = true;
1738
1739                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1740                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1741                                              MDS_INODELOCK_UPDATE);
1742                         if (rc) {
1743                                 mdt_object_put(info->mti_env, lnkp);
1744                                 OBD_FREE_PTR(msl);
1745                                 GOTO(out, rc);
1746                         }
1747
1748                         if (mdt_object_remote(lnkp)) {
1749                                 struct ldlm_lock *lock;
1750
1751                                 /*
1752                                  * for remote object, set lock cb_atomic,
1753                                  * so lock can be released in blocking_ast()
1754                                  * immediately, then the next lock_try will
1755                                  * have better chance of success.
1756                                  */
1757                                 lock = ldlm_handle2lock(
1758                                                 &msl->msl_lh.mlh_rreg_lh);
1759                                 LASSERT(lock != NULL);
1760                                 lock_res_and_lock(lock);
1761                                 ldlm_set_atomic_cb(lock);
1762                                 unlock_res_and_lock(lock);
1763                                 LDLM_LOCK_PUT(lock);
1764                         }
1765
1766                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1767                         OBD_FREE_PTR(msl);
1768                         continue;
1769                 }
1770
1771                 INIT_LIST_HEAD(&msl->msl_linkage);
1772                 msl->msl_obj = lnkp;
1773                 list_add_tail(&msl->msl_linkage, link_locks);
1774
1775                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1776         }
1777
1778         if (blocked)
1779                 GOTO(out, rc = -EBUSY);
1780
1781         EXIT;
1782 out:
1783         if (rc)
1784                 mdt_unlock_list(info, link_locks, rc);
1785         else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1786                 /*
1787                  * parent may have 3 local objects: master object and 2 stripes
1788                  * (if it's being migrated too); source may have 2 local
1789                  * objects: master and 1 stripe; target has 1 local object.
1790                  */
1791                 rc = 1;
1792         return rc;
1793 }
1794
1795 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1796                                   struct mdt_object *obj,
1797                                   const struct md_attr *ma,
1798                                   struct list_head *slave_locks)
1799 {
1800         struct mdt_device *mdt = info->mti_mdt;
1801         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1802         struct lu_fid *fid = &info->mti_tmp_fid1;
1803         struct mdt_object *slave;
1804         struct mdt_sub_lock *msl;
1805         int i;
1806         int rc;
1807
1808         ENTRY;
1809         LASSERT(mdt_object_remote(obj));
1810         LASSERT(ma->ma_valid & MA_LMV);
1811         LASSERT(lmv);
1812
1813         if (!lmv_is_sane(lmv))
1814                 RETURN(-EINVAL);
1815
1816         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1817                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1818
1819                 if (!fid_is_sane(fid))
1820                         continue;
1821
1822                 slave = mdt_object_find(info->mti_env, mdt, fid);
1823                 if (IS_ERR(slave))
1824                         GOTO(out, rc = PTR_ERR(slave));
1825
1826                 OBD_ALLOC_PTR(msl);
1827                 if (!msl) {
1828                         mdt_object_put(info->mti_env, slave);
1829                         GOTO(out, rc = -ENOMEM);
1830                 }
1831
1832                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1833                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1834                                            MDS_INODELOCK_UPDATE, true);
1835                 if (rc) {
1836                         OBD_FREE_PTR(msl);
1837                         mdt_object_put(info->mti_env, slave);
1838                         GOTO(out, rc);
1839                 }
1840
1841                 INIT_LIST_HEAD(&msl->msl_linkage);
1842                 msl->msl_obj = slave;
1843                 list_add_tail(&msl->msl_linkage, slave_locks);
1844         }
1845         EXIT;
1846
1847 out:
1848         if (rc)
1849                 mdt_unlock_list(info, slave_locks, rc);
1850         return rc;
1851 }
1852
1853 /* lock parent and its stripes */
1854 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1855                                    struct mdt_object *obj,
1856                                    const struct md_attr *ma,
1857                                    struct mdt_lock_handle *lh,
1858                                    struct ldlm_enqueue_info *einfo,
1859                                    struct list_head *slave_locks)
1860 {
1861         int rc;
1862
1863         if (mdt_object_remote(obj)) {
1864                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1865                                             &lh->mlh_rreg_lh, LCK_PW,
1866                                             MDS_INODELOCK_UPDATE, false);
1867                 if (rc != ELDLM_OK)
1868                         return rc;
1869
1870                 /*
1871                  * if obj is remote and striped, lock its stripes explicitly
1872                  * because it's not striped in LOD layer on this MDT.
1873                  */
1874                 if (ma->ma_valid & MA_LMV) {
1875                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1876                         if (rc)
1877                                 mdt_object_unlock(info, obj, lh, rc);
1878                 }
1879         } else {
1880                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1881                                             einfo, true);
1882         }
1883
1884         return rc;
1885 }
1886
1887 /*
1888  * in migration, object may be remote, and we need take full lock of it and its
1889  * stripes if it's directory, besides, object may be a remote object on its
1890  * parent, revoke its LOOKUP lock on where its parent is located.
1891  */
1892 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1893                                    struct mdt_object *pobj,
1894                                    struct mdt_object *obj,
1895                                    struct mdt_lock_handle *lh,
1896                                    struct ldlm_enqueue_info *einfo,
1897                                    struct list_head *slave_locks)
1898 {
1899         int rc;
1900
1901         if (mdt_object_remote(obj)) {
1902                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1903                 if (rc)
1904                         return rc;
1905
1906                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1907                                             &lh->mlh_rreg_lh, LCK_EX,
1908                                             MDS_INODELOCK_FULL, false);
1909                 if (rc != ELDLM_OK)
1910                         return rc;
1911
1912                 /*
1913                  * if obj is remote and striped, lock its stripes explicitly
1914                  * because it's not striped in LOD layer on this MDT.
1915                  */
1916                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1917                         struct md_attr *ma = &info->mti_attr;
1918
1919                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1920                         if (rc) {
1921                                 mdt_object_unlock(info, obj, lh, rc);
1922                                 return rc;
1923                         }
1924
1925                         if (ma->ma_valid & MA_LMV) {
1926                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1927                                                             slave_locks);
1928                                 if (rc)
1929                                         mdt_object_unlock(info, obj, lh, rc);
1930                         }
1931                 }
1932         } else {
1933                 if (mdt_object_remote(pobj)) {
1934                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1935                         if (rc)
1936                                 return rc;
1937                 }
1938
1939                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1940                                             einfo, true);
1941         }
1942
1943         return rc;
1944 }
1945
1946 /*
1947  * lookup source by name, if parent is striped directory, we need to find the
1948  * corresponding stripe where source is located, and then lookup there.
1949  *
1950  * besides, if parent is migrating too, and file is already in target stripe,
1951  * this should be a redo of 'lfs migrate' on client side.
1952  */
1953 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1954                               struct mdt_object *pobj,
1955                               const struct md_attr *ma,
1956                               const struct lu_name *lname,
1957                               struct mdt_object **spobj,
1958                               struct mdt_object **sobj)
1959 {
1960         const struct lu_env *env = info->mti_env;
1961         struct lu_fid *fid = &info->mti_tmp_fid1;
1962         struct mdt_object *stripe;
1963         int rc;
1964
1965         if (ma->ma_valid & MA_LMV) {
1966                 /* if parent is striped, lookup on corresponding stripe */
1967                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1968
1969                 if (!lmv_is_sane(lmv))
1970                         return -EBADF;
1971
1972                 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
1973                                                   lname->ln_namelen);
1974                 if (rc < 0)
1975                         return rc;
1976
1977                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1978
1979                 stripe = mdt_object_find(env, info->mti_mdt, fid);
1980                 if (IS_ERR(stripe))
1981                         return PTR_ERR(stripe);
1982
1983                 fid_zero(fid);
1984                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1985                                 &info->mti_spec);
1986                 if (rc == -ENOENT && lmv_is_layout_changing(lmv)) {
1987                         /*
1988                          * if parent layout is changeing, and lookup child
1989                          * failed on source stripe, lookup again on target
1990                          * stripe, if it exists, it means previous migration
1991                          * was interrupted, and current file was migrated
1992                          * already.
1993                          */
1994                         mdt_object_put(env, stripe);
1995
1996                         rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
1997                                                       lname->ln_namelen);
1998                         if (rc < 0)
1999                                 return rc;
2000
2001                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2002
2003                         stripe = mdt_object_find(env, info->mti_mdt, fid);
2004                         if (IS_ERR(stripe))
2005                                 return PTR_ERR(stripe);
2006
2007                         fid_zero(fid);
2008                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
2009                                         fid, &info->mti_spec);
2010                         mdt_object_put(env, stripe);
2011                         return rc ?: -EALREADY;
2012                 } else if (rc) {
2013                         mdt_object_put(env, stripe);
2014                         return rc;
2015                 }
2016         } else {
2017                 fid_zero(fid);
2018                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2019                                 &info->mti_spec);
2020                 if (rc)
2021                         return rc;
2022
2023                 stripe = pobj;
2024                 mdt_object_get(env, stripe);
2025         }
2026
2027         *spobj = stripe;
2028
2029         *sobj = mdt_object_find(env, info->mti_mdt, fid);
2030         if (IS_ERR(*sobj)) {
2031                 mdt_object_put(env, stripe);
2032                 rc = PTR_ERR(*sobj);
2033                 *spobj = NULL;
2034                 *sobj = NULL;
2035         }
2036
2037         return rc;
2038 }
2039
2040 /* end lease and close file for regular file */
2041 static int mdd_migrate_close(struct mdt_thread_info *info,
2042                              struct mdt_object *obj)
2043 {
2044         struct close_data *data;
2045         struct mdt_body *repbody;
2046         struct ldlm_lock *lease;
2047         int rc;
2048         int rc2;
2049
2050         rc = -EPROTO;
2051         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2052                                       RCL_CLIENT) ||
2053             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2054                                       RCL_CLIENT))
2055                 goto close;
2056
2057         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2058         if (!data)
2059                 goto close;
2060
2061         rc = -ESTALE;
2062         lease = ldlm_handle2lock(&data->cd_handle);
2063         if (!lease)
2064                 goto close;
2065
2066         /* check if the lease was already canceled */
2067         lock_res_and_lock(lease);
2068         rc = ldlm_is_cancel(lease);
2069         unlock_res_and_lock(lease);
2070
2071         if (rc) {
2072                 rc = -EAGAIN;
2073                 LDLM_DEBUG(lease, DFID" lease broken",
2074                            PFID(mdt_object_fid(obj)));
2075         }
2076
2077         /*
2078          * cancel server side lease, client side counterpart should have been
2079          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2080          */
2081         ldlm_lock_cancel(lease);
2082         ldlm_reprocess_all(lease->l_resource, lease);
2083         LDLM_LOCK_PUT(lease);
2084
2085 close:
2086         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2087         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2088         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2089
2090         return rc ?: rc2;
2091 }
2092
2093 /*
2094  * migrate file in below steps:
2095  *  1. lock parent and its stripes
2096  *  2. lookup source by name
2097  *  3. lock parents of source links if source is not directory
2098  *  4. reject if source is in HSM
2099  *  5. take source open_sem and close file if source is regular file
2100  *  6. lock source and its stripes if it's directory
2101  *  7. lock target so subsequent change to it can trigger COS
2102  *  8. migrate file
2103  *  9. unlock above locks
2104  * 10. sync device if source has links
2105  */
2106 int mdt_reint_migrate(struct mdt_thread_info *info,
2107                       struct mdt_lock_handle *unused)
2108 {
2109         const struct lu_env *env = info->mti_env;
2110         struct mdt_device *mdt = info->mti_mdt;
2111         struct ptlrpc_request *req = mdt_info_req(info);
2112         struct mdt_reint_record *rr = &info->mti_rr;
2113         struct lu_ucred *uc = mdt_ucred(info);
2114         struct md_attr *ma = &info->mti_attr;
2115         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
2116         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
2117         struct mdt_object *pobj;
2118         struct mdt_object *spobj = NULL;
2119         struct mdt_object *sobj = NULL;
2120         struct mdt_object *tobj;
2121         struct lustre_handle rename_lh = { 0 };
2122         struct mdt_lock_handle *lhp;
2123         struct mdt_lock_handle *lhs;
2124         struct mdt_lock_handle *lht;
2125         LIST_HEAD(parent_slave_locks);
2126         LIST_HEAD(child_slave_locks);
2127         LIST_HEAD(link_locks);
2128         int lock_retries = 5;
2129         bool open_sem_locked = false;
2130         bool do_sync = false;
2131         int rc;
2132
2133         ENTRY;
2134         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2135                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2136
2137         if (info->mti_dlm_req)
2138                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2139
2140         if (!fid_is_md_operative(rr->rr_fid1) ||
2141             !fid_is_md_operative(rr->rr_fid2))
2142                 RETURN(-EPERM);
2143
2144         /* don't allow migrate . or .. */
2145         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2146                 RETURN(-EBUSY);
2147
2148         if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2149                 RETURN(-EPERM);
2150
2151         if (uc && !md_capable(uc, CFS_CAP_SYS_ADMIN) &&
2152             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2153             mdt->mdt_enable_remote_dir_gid != -1)
2154                 RETURN(-EPERM);
2155
2156         /*
2157          * Note: do not enqueue rename lock for replay request, because
2158          * if other MDT holds rename lock, but being blocked to wait for
2159          * this MDT to finish its recovery, and the failover MDT can not
2160          * get rename lock, which will cause deadlock.
2161          *
2162          * req is NULL if this is called by directory auto-split.
2163          */
2164         if (req && !req_is_replay(req)) {
2165                 rc = mdt_rename_lock(info, &rename_lh);
2166                 if (rc != 0) {
2167                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2168                                mdt_obd_name(info->mti_mdt), rc);
2169                         RETURN(rc);
2170                 }
2171         }
2172
2173         /* pobj is master object of parent */
2174         pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2175         if (IS_ERR(pobj))
2176                 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2177
2178         if (req) {
2179                 rc = mdt_version_get_check(info, pobj, 0);
2180                 if (rc)
2181                         GOTO(put_parent, rc);
2182         }
2183
2184         if (!mdt_object_exists(pobj))
2185                 GOTO(put_parent, rc = -ENOENT);
2186
2187         if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2188                 GOTO(put_parent, rc = -ENOTDIR);
2189
2190         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2191         if (rc)
2192                 GOTO(put_parent, rc);
2193
2194 lock_parent:
2195         /* lock parent object */
2196         lhp = &info->mti_lh[MDT_LH_PARENT];
2197         mdt_lock_reg_init(lhp, LCK_PW);
2198         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2199                                      &parent_slave_locks);
2200         if (rc)
2201                 GOTO(put_parent, rc);
2202
2203         /*
2204          * spobj is the corresponding stripe against name if pobj is striped
2205          * directory, which is the real parent, and no need to lock, because
2206          * we've taken full lock of pobj.
2207          */
2208         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2209         if (rc)
2210                 GOTO(unlock_parent, rc);
2211
2212         /* lock parents of source links, and revoke LOOKUP lock of links */
2213         rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2214                                    &parent_slave_locks, &link_locks);
2215         if (rc == -EBUSY && lock_retries-- > 0) {
2216                 mdt_object_put(env, sobj);
2217                 mdt_object_put(env, spobj);
2218                 goto lock_parent;
2219         }
2220
2221         if (rc < 0)
2222                 GOTO(put_source, rc);
2223
2224         /*
2225          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2226          * with one request, if total lock count exceeds this limit, we will
2227          * drop all locks after migration, and synchronous device in the end.
2228          */
2229         do_sync = rc;
2230
2231         /* TODO: DoM migration is not supported, migrate dirent only */
2232         if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2233                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2234                 if (rc)
2235                         GOTO(unlock_links, rc);
2236
2237                 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2238                         info->mti_spec.sp_migrate_nsonly = 1;
2239         }
2240
2241         /* if migration HSM is allowed */
2242         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2243                 ma->ma_need = MA_HSM;
2244                 ma->ma_valid = 0;
2245                 rc = mdt_attr_get_complex(info, sobj, ma);
2246                 if (rc)
2247                         GOTO(unlock_links, rc);
2248
2249                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2250                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2251         }
2252
2253         /* end lease and close file for regular file */
2254         if (info->mti_spec.sp_migrate_close) {
2255                 /* try to hold open_sem so that nobody else can open the file */
2256                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2257                         /* close anyway */
2258                         mdd_migrate_close(info, sobj);
2259                         GOTO(unlock_links, rc = -EBUSY);
2260                 } else {
2261                         open_sem_locked = true;
2262                         rc = mdd_migrate_close(info, sobj);
2263                         if (rc)
2264                                 GOTO(unlock_open_sem, rc);
2265                 }
2266         }
2267
2268         /* lock source */
2269         lhs = &info->mti_lh[MDT_LH_OLD];
2270         mdt_lock_reg_init(lhs, LCK_EX);
2271         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2272                                      &child_slave_locks);
2273         if (rc)
2274                 GOTO(unlock_open_sem, rc);
2275
2276         /* lock target */
2277         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2278         if (IS_ERR(tobj))
2279                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2280
2281         lht = &info->mti_lh[MDT_LH_NEW];
2282         mdt_lock_reg_init(lht, LCK_EX);
2283         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2284         if (rc)
2285                 GOTO(put_target, rc);
2286
2287         /* Don't do lookup sanity check. We know name doesn't exist. */
2288         info->mti_spec.sp_cr_lookup = 0;
2289         info->mti_spec.sp_feat = &dt_directory_features;
2290
2291         rc = mdo_migrate(env, mdt_object_child(pobj),
2292                          mdt_object_child(sobj), &rr->rr_name,
2293                          mdt_object_child(tobj),
2294                          &info->mti_spec, ma);
2295         if (!rc)
2296                 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2297                                      LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2298         EXIT;
2299
2300         mdt_object_unlock(info, tobj, lht, rc);
2301 put_target:
2302         mdt_object_put(env, tobj);
2303 unlock_source:
2304         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2305                                   &child_slave_locks, rc);
2306 unlock_open_sem:
2307         if (open_sem_locked)
2308                 up_write(&sobj->mot_open_sem);
2309 unlock_links:
2310         /* if we've got too many locks to save into RPC,
2311          * then just commit before the locks are released
2312          */
2313         if (!rc && do_sync)
2314                 mdt_device_sync(env, mdt);
2315         mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2316 put_source:
2317         mdt_object_put(env, sobj);
2318         mdt_object_put(env, spobj);
2319 unlock_parent:
2320         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2321                                   &parent_slave_locks, rc);
2322 put_parent:
2323         mdt_object_put(env, pobj);
2324 unlock_rename:
2325         if (lustre_handle_is_used(&rename_lh))
2326                 mdt_rename_unlock(&rename_lh);
2327
2328         return rc;
2329 }
2330
2331 static int mdt_object_lock_save(struct mdt_thread_info *info,
2332                                 struct mdt_object *dir,
2333                                 struct mdt_lock_handle *lh,
2334                                 int idx, bool cos_incompat)
2335 {
2336         int rc;
2337
2338         /* we lock the target dir if it is local */
2339         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2340                                    cos_incompat);
2341         if (rc != 0)
2342                 return rc;
2343
2344         /* get and save correct version after locking */
2345         mdt_version_get_save(info, dir, idx);
2346         return 0;
2347 }
2348
2349 /*
2350  * determine lock order of sobj and tobj
2351  *
2352  * there are two situations we need to lock tobj before sobj:
2353  * 1. sobj is child of tobj
2354  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2355  *    larger than that of tobj
2356  *
2357  * \retval      1 lock tobj before sobj
2358  * \retval      0 lock sobj before tobj
2359  * \retval      -ev negative errno upon error
2360  */
2361 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2362                                            struct mdt_object *sobj,
2363                                            struct mdt_object *tobj)
2364 {
2365         struct md_attr *ma = &info->mti_attr;
2366         struct lu_fid *spfid = &info->mti_tmp_fid1;
2367         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2368         struct lmv_mds_md_v1 *lmv;
2369         __u32 sindex;
2370         __u32 tindex;
2371         int rc;
2372
2373         /* sobj and tobj are the same */
2374         if (sobj == tobj)
2375                 return 0;
2376
2377         if (fid_is_root(mdt_object_fid(sobj)))
2378                 return 0;
2379
2380         if (fid_is_root(mdt_object_fid(tobj)))
2381                 return 1;
2382
2383         /* check whether sobj is child of tobj */
2384         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2385                            mdt_object_fid(tobj));
2386         if (rc < 0)
2387                 return rc;
2388
2389         if (rc == 1)
2390                 return 1;
2391
2392         /* check whether sobj and tobj are children of the same parent */
2393         rc = mdt_attr_get_pfid(info, sobj, spfid);
2394         if (rc)
2395                 return rc;
2396
2397         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2398         if (rc)
2399                 return rc;
2400
2401         if (!lu_fid_eq(spfid, tpfid))
2402                 return 0;
2403
2404         /* check whether sobj and tobj are sibling stripes */
2405         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2406         if (rc)
2407                 return rc;
2408
2409         if (!(ma->ma_valid & MA_LMV))
2410                 return 0;
2411
2412         lmv = &ma->ma_lmv->lmv_md_v1;
2413         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2414                 return 0;
2415         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2416
2417         ma->ma_valid = 0;
2418         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2419         if (rc)
2420                 return rc;
2421
2422         if (!(ma->ma_valid & MA_LMV))
2423                 return -ENODATA;
2424
2425         lmv = &ma->ma_lmv->lmv_md_v1;
2426         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2427                 return -EINVAL;
2428         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2429
2430         /* check stripe index of sobj and tobj */
2431         if (sindex == tindex)
2432                 return -EINVAL;
2433
2434         return sindex < tindex ? 0 : 1;
2435 }
2436
2437 /*
2438  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2439  * 2 - srcdir child; 3 - tgtdir child.
2440  * Update on disk version of srcdir child.
2441  */
2442 static int mdt_reint_rename(struct mdt_thread_info *info,
2443                             struct mdt_lock_handle *unused)
2444 {
2445         struct mdt_device *mdt = info->mti_mdt;
2446         struct mdt_reint_record *rr = &info->mti_rr;
2447         struct md_attr *ma = &info->mti_attr;
2448         struct ptlrpc_request *req = mdt_info_req(info);
2449         struct mdt_object *msrcdir = NULL;
2450         struct mdt_object *mtgtdir = NULL;
2451         struct mdt_object *mold;
2452         struct mdt_object *mnew = NULL;
2453         struct lustre_handle rename_lh = { 0 };
2454         struct mdt_lock_handle *lh_srcdirp;
2455         struct mdt_lock_handle *lh_tgtdirp;
2456         struct mdt_lock_handle *lh_oldp = NULL;
2457         struct mdt_lock_handle *lh_newp = NULL;
2458         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2459         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2460         __u64 lock_ibits;
2461         bool reverse = false, discard = false;
2462         bool cos_incompat;
2463         int rc;
2464
2465         ENTRY;
2466         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2467                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2468                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2469
2470         if (info->mti_dlm_req)
2471                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2472
2473         if (!fid_is_md_operative(rr->rr_fid1) ||
2474             !fid_is_md_operative(rr->rr_fid2))
2475                 RETURN(-EPERM);
2476
2477         /* find both parents. */
2478         msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2479         if (IS_ERR(msrcdir))
2480                 RETURN(PTR_ERR(msrcdir));
2481
2482         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2483
2484         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2485                 mtgtdir = msrcdir;
2486                 mdt_object_get(info->mti_env, mtgtdir);
2487         } else {
2488                 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2489                 if (IS_ERR(mtgtdir))
2490                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2491         }
2492
2493         /*
2494          * Note: do not enqueue rename lock for replay request, because
2495          * if other MDT holds rename lock, but being blocked to wait for
2496          * this MDT to finish its recovery, and the failover MDT can not
2497          * get rename lock, which will cause deadlock.
2498          */
2499         if (!req_is_replay(req)) {
2500                 /*
2501                  * Normally rename RPC is handled on the MDT with the target
2502                  * directory (if target exists, it's on the MDT with the
2503                  * target), if the source directory is remote, it's a hint that
2504                  * source is remote too (this may not be true, but it won't
2505                  * cause any issue), return -EXDEV early to avoid taking
2506                  * rename_lock.
2507                  */
2508                 if (!mdt->mdt_enable_remote_rename &&
2509                     mdt_object_remote(msrcdir))
2510                         GOTO(out_put_tgtdir, rc = -EXDEV);
2511
2512                 rc = mdt_rename_lock(info, &rename_lh);
2513                 if (rc != 0) {
2514                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2515                                mdt_obd_name(mdt), rc);
2516                         GOTO(out_put_tgtdir, rc);
2517                 }
2518         }
2519
2520         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2521         if (rc < 0)
2522                 GOTO(out_unlock_rename, rc);
2523
2524         reverse = rc;
2525
2526         /* source needs to be looked up after locking source parent, otherwise
2527          * this rename may race with unlink source, and cause rename hang, see
2528          * sanityn.sh 55b, so check parents first, if later we found source is
2529          * remote, relock parents.
2530          */
2531         cos_incompat = (mdt_object_remote(msrcdir) ||
2532                         mdt_object_remote(mtgtdir));
2533
2534         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2535
2536         /* lock parents in the proper order. */
2537         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2538         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2539
2540 relock:
2541         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2542         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2543
2544         if (reverse) {
2545                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2546                                           cos_incompat);
2547                 if (rc)
2548                         GOTO(out_unlock_rename, rc);
2549
2550                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2551
2552                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2553                                           cos_incompat);
2554                 if (rc != 0) {
2555                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2556                         GOTO(out_unlock_rename, rc);
2557                 }
2558         } else {
2559                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2560                                           cos_incompat);
2561                 if (rc)
2562                         GOTO(out_unlock_rename, rc);
2563
2564                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2565
2566                 if (mtgtdir != msrcdir) {
2567                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2568                                                   cos_incompat);
2569                 } else if (!mdt_object_remote(mtgtdir) &&
2570                            lh_srcdirp->mlh_pdo_hash !=
2571                            lh_tgtdirp->mlh_pdo_hash) {
2572                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2573                                                 MDS_INODELOCK_UPDATE,
2574                                                 cos_incompat);
2575                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2576                 }
2577                 if (rc != 0) {
2578                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2579                         GOTO(out_unlock_rename, rc);
2580                 }
2581         }
2582
2583         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2584         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2585
2586         /* find mold object. */
2587         fid_zero(old_fid);
2588         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2589         if (rc != 0)
2590                 GOTO(out_unlock_parents, rc);
2591
2592         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2593                 GOTO(out_unlock_parents, rc = -EINVAL);
2594
2595         if (!fid_is_md_operative(old_fid))
2596                 GOTO(out_unlock_parents, rc = -EPERM);
2597
2598         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2599         if (IS_ERR(mold))
2600                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2601
2602         if (!mdt_object_exists(mold)) {
2603                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2604                                 &mold->mot_obj,
2605                                 "object does not exist");
2606                 GOTO(out_put_old, rc = -ENOENT);
2607         }
2608
2609         if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2610                 GOTO(out_put_old, rc = -EXDEV);
2611
2612         /* Check if @mtgtdir is subdir of @mold, before locking child
2613          * to avoid reverse locking.
2614          */
2615         if (mtgtdir != msrcdir) {
2616                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2617                                    old_fid);
2618                 if (rc) {
2619                         if (rc == 1)
2620                                 rc = -EINVAL;
2621                         GOTO(out_put_old, rc);
2622                 }
2623         }
2624
2625         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2626         /* save version after locking */
2627         mdt_version_get_save(info, mold, 2);
2628
2629         if (!cos_incompat && mdt_object_remote(mold)) {
2630                 cos_incompat = true;
2631                 mdt_object_put(info->mti_env, mold);
2632                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2633                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2634                 goto relock;
2635         }
2636
2637         /* find mnew object:
2638          * mnew target object may not exist now
2639          * lookup with version checking
2640          */
2641         fid_zero(new_fid);
2642         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2643                                       3);
2644         if (rc == 0) {
2645                 /* the new_fid should have been filled at this moment */
2646                 if (lu_fid_eq(old_fid, new_fid))
2647                         GOTO(out_put_old, rc);
2648
2649                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2650                     lu_fid_eq(new_fid, rr->rr_fid2))
2651                         GOTO(out_put_old, rc = -EINVAL);
2652
2653                 if (!fid_is_md_operative(new_fid))
2654                         GOTO(out_put_old, rc = -EPERM);
2655
2656                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2657                 if (IS_ERR(mnew))
2658                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2659
2660                 if (!mdt_object_exists(mnew)) {
2661                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2662                                         &mnew->mot_obj,
2663                                         "object does not exist");
2664                         GOTO(out_put_new, rc = -ENOENT);
2665                 }
2666
2667                 if (mdt_object_remote(mnew)) {
2668                         struct mdt_body  *repbody;
2669
2670                         /* Always send rename req to the target child MDT */
2671                         repbody = req_capsule_server_get(info->mti_pill,
2672                                                          &RMF_MDT_BODY);
2673                         LASSERT(repbody != NULL);
2674                         repbody->mbo_fid1 = *new_fid;
2675                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2676                         GOTO(out_put_new, rc = -EXDEV);
2677                 }
2678                 /* Before locking the target dir, check we do not replace
2679                  * a dir with a non-dir, otherwise it may deadlock with
2680                  * link op which tries to create a link in this dir
2681                  * back to this non-dir.
2682                  */
2683                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2684                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2685                         GOTO(out_put_new, rc = -EISDIR);
2686
2687                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2688                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2689                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2690                 if (mdt_object_remote(msrcdir)) {
2691                         /* Enqueue lookup lock from the parent MDT */
2692                         rc = mdt_remote_object_lock(info, msrcdir,
2693                                                     mdt_object_fid(mold),
2694                                                     &lh_oldp->mlh_rreg_lh,
2695                                                     lh_oldp->mlh_rreg_mode,
2696                                                     MDS_INODELOCK_LOOKUP,
2697                                                     false);
2698                         if (rc != ELDLM_OK)
2699                                 GOTO(out_put_new, rc);
2700
2701                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2702                 }
2703
2704                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2705                                            cos_incompat);
2706                 if (rc != 0)
2707                         GOTO(out_unlock_old, rc);
2708
2709                 /* Check if @msrcdir is subdir of @mnew, before locking child
2710                  * to avoid reverse locking.
2711                  */
2712                 if (mtgtdir != msrcdir) {
2713                         rc = mdo_is_subdir(info->mti_env,
2714                                            mdt_object_child(msrcdir), new_fid);
2715                         if (rc) {
2716                                 if (rc == 1)
2717                                         rc = -EINVAL;
2718                                 GOTO(out_unlock_old, rc);
2719                         }
2720                 }
2721
2722                 /* We used to acquire MDS_INODELOCK_FULL here but we
2723                  * can't do this now because a running HSM restore on
2724                  * the rename onto victim will hold the layout
2725                  * lock. See LU-4002.
2726                  */
2727
2728                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2729                 mdt_lock_reg_init(lh_newp, LCK_EX);
2730                 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2731                                            MDS_INODELOCK_LOOKUP |
2732                                            MDS_INODELOCK_UPDATE,
2733                                            cos_incompat);
2734                 if (rc != 0)
2735                         GOTO(out_unlock_old, rc);
2736
2737                 /* get and save version after locking */
2738                 mdt_version_get_save(info, mnew, 3);
2739         } else if (rc != -EREMOTE && rc != -ENOENT) {
2740                 GOTO(out_put_old, rc);
2741         } else {
2742                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2743                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2744                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2745                 if (mdt_object_remote(msrcdir)) {
2746                         /* Enqueue lookup lock from the parent MDT */
2747                         rc = mdt_remote_object_lock(info, msrcdir,
2748                                                     mdt_object_fid(mold),
2749                                                     &lh_oldp->mlh_rreg_lh,
2750                                                     lh_oldp->mlh_rreg_mode,
2751                                                     MDS_INODELOCK_LOOKUP,
2752                                                     false);
2753                         if (rc != ELDLM_OK)
2754                                 GOTO(out_put_old, rc);
2755
2756                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2757                 }
2758
2759                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2760                                            cos_incompat);
2761                 if (rc != 0)
2762                         GOTO(out_unlock_old, rc);
2763
2764                 mdt_enoent_version_save(info, 3);
2765         }
2766
2767         /* step 5: rename it */
2768         mdt_reint_init_ma(info, ma);
2769
2770         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2771                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2772
2773         if (mnew != NULL)
2774                 mutex_lock(&mnew->mot_lov_mutex);
2775
2776         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2777                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2778                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2779                         &rr->rr_tgt_name, ma);
2780
2781         if (mnew != NULL)
2782                 mutex_unlock(&mnew->mot_lov_mutex);
2783
2784         /* handle last link of tgt object */
2785         if (rc == 0) {
2786                 mdt_counter_incr(req, LPROC_MDT_RENAME);
2787                 if (mnew) {
2788                         mdt_handle_last_unlink(info, mnew, ma);
2789                         discard = mdt_dom_check_for_discard(info, mnew);
2790                 }
2791                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2792                                          msrcdir, mtgtdir);
2793         }
2794
2795         EXIT;
2796         if (mnew != NULL)
2797                 mdt_object_unlock(info, mnew, lh_newp, rc);
2798 out_unlock_old:
2799         mdt_object_unlock(info, mold, lh_oldp, rc);
2800 out_put_new:
2801         if (mnew && !discard)
2802                 mdt_object_put(info->mti_env, mnew);
2803 out_put_old:
2804         mdt_object_put(info->mti_env, mold);
2805 out_unlock_parents:
2806         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2807         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2808 out_unlock_rename:
2809         if (lustre_handle_is_used(&rename_lh))
2810                 mdt_rename_unlock(&rename_lh);
2811 out_put_tgtdir:
2812         mdt_object_put(info->mti_env, mtgtdir);
2813 out_put_srcdir:
2814         mdt_object_put(info->mti_env, msrcdir);
2815
2816         /* The DoM discard can be done right in the place above where it is
2817          * assigned, meanwhile it is done here after rename unlock due to
2818          * compatibility with old clients, for them the discard blocks
2819          * the main thread until completion. Check LU-11359 for details.
2820          */
2821         if (discard) {
2822                 mdt_dom_discard_data(info, mnew);
2823                 mdt_object_put(info->mti_env, mnew);
2824         }
2825         return rc;
2826 }
2827
2828 static int mdt_reint_resync(struct mdt_thread_info *info,
2829                             struct mdt_lock_handle *lhc)
2830 {
2831         struct mdt_reint_record *rr = &info->mti_rr;
2832         struct ptlrpc_request *req = mdt_info_req(info);
2833         struct md_attr *ma = &info->mti_attr;
2834         struct mdt_object *mo;
2835         struct ldlm_lock *lease;
2836         struct mdt_body *repbody;
2837         struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2838         bool lease_broken;
2839         int rc, rc2;
2840
2841         ENTRY;
2842         DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2843
2844         if (info->mti_dlm_req)
2845                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2846
2847         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2848         if (IS_ERR(mo))
2849                 GOTO(out, rc = PTR_ERR(mo));
2850
2851         if (!mdt_object_exists(mo))
2852                 GOTO(out_obj, rc = -ENOENT);
2853
2854         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2855                 GOTO(out_obj, rc = -EINVAL);
2856
2857         if (mdt_object_remote(mo))
2858                 GOTO(out_obj, rc = -EREMOTE);
2859
2860         lease = ldlm_handle2lock(rr->rr_lease_handle);
2861         if (lease == NULL)
2862                 GOTO(out_obj, rc = -ESTALE);
2863
2864         /* It's really necessary to grab open_sem and check if the lease lock
2865          * has been lost. There would exist a concurrent writer coming in and
2866          * generating some dirty data in memory cache, the writeback would fail
2867          * after the layout version is increased by MDS_REINT_RESYNC RPC.
2868          */
2869         if (!down_write_trylock(&mo->mot_open_sem))
2870                 GOTO(out_put_lease, rc = -EBUSY);
2871
2872         lock_res_and_lock(lease);
2873         lease_broken = ldlm_is_cancel(lease);
2874         unlock_res_and_lock(lease);
2875         if (lease_broken)
2876                 GOTO(out_unlock, rc = -EBUSY);
2877
2878         /* the file has yet opened by anyone else after we took the lease. */
2879         layout.mlc_opc = MD_LAYOUT_RESYNC;
2880         lhc = &info->mti_lh[MDT_LH_LOCAL];
2881         rc = mdt_layout_change(info, mo, lhc, &layout);
2882         if (rc)
2883                 GOTO(out_unlock, rc);
2884
2885         mdt_object_unlock(info, mo, lhc, 0);
2886
2887         ma->ma_need = MA_INODE;
2888         ma->ma_valid = 0;
2889         rc = mdt_attr_get_complex(info, mo, ma);
2890         if (rc != 0)
2891                 GOTO(out_unlock, rc);
2892
2893         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2894         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2895
2896         EXIT;
2897 out_unlock:
2898         up_write(&mo->mot_open_sem);
2899 out_put_lease:
2900         LDLM_LOCK_PUT(lease);
2901 out_obj:
2902         mdt_object_put(info->mti_env, mo);
2903 out:
2904         mdt_client_compatibility(info);
2905         rc2 = mdt_fix_reply(info);
2906         if (rc == 0)
2907                 rc = rc2;
2908         return rc;
2909 }
2910
2911 struct mdt_reinter {
2912         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2913         enum lprocfs_extra_opc mr_extra_opc;
2914 };
2915
2916 static const struct mdt_reinter mdt_reinters[] = {
2917         [REINT_SETATTR] = {
2918                 .mr_handler = &mdt_reint_setattr,
2919                 .mr_extra_opc = MDS_REINT_SETATTR,
2920         },
2921         [REINT_CREATE] = {
2922                 .mr_handler = &mdt_reint_create,
2923                 .mr_extra_opc = MDS_REINT_CREATE,
2924         },
2925         [REINT_LINK] = {
2926                 .mr_handler = &mdt_reint_link,
2927                 .mr_extra_opc = MDS_REINT_LINK,
2928         },
2929         [REINT_UNLINK] = {
2930                 .mr_handler = &mdt_reint_unlink,
2931                 .mr_extra_opc = MDS_REINT_UNLINK,
2932         },
2933         [REINT_RENAME] = {
2934                 .mr_handler = &mdt_reint_rename,
2935                 .mr_extra_opc = MDS_REINT_RENAME,
2936         },
2937         [REINT_OPEN] = {
2938                 .mr_handler = &mdt_reint_open,
2939                 .mr_extra_opc = MDS_REINT_OPEN,
2940         },
2941         [REINT_SETXATTR] = {
2942                 .mr_handler = &mdt_reint_setxattr,
2943                 .mr_extra_opc = MDS_REINT_SETXATTR,
2944         },
2945         [REINT_RMENTRY] = {
2946                 .mr_handler = &mdt_reint_unlink,
2947                 .mr_extra_opc = MDS_REINT_UNLINK,
2948         },
2949         [REINT_MIGRATE] = {
2950                 .mr_handler = &mdt_reint_migrate,
2951                 .mr_extra_opc = MDS_REINT_RENAME,
2952         },
2953         [REINT_RESYNC] = {
2954                 .mr_handler = &mdt_reint_resync,
2955                 .mr_extra_opc = MDS_REINT_RESYNC,
2956         },
2957 };
2958
2959 int mdt_reint_rec(struct mdt_thread_info *info,
2960                   struct mdt_lock_handle *lhc)
2961 {
2962         const struct mdt_reinter *mr;
2963         int rc;
2964
2965         ENTRY;
2966         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2967                 RETURN(-EPROTO);
2968
2969         mr = &mdt_reinters[info->mti_rr.rr_opcode];
2970         if (mr->mr_handler == NULL)
2971                 RETURN(-EPROTO);
2972
2973         rc = (*mr->mr_handler)(info, lhc);
2974
2975         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2976                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
2977
2978         RETURN(rc);
2979 }