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  *
  31  * lustre/mdt/mdt_reint.c
  32  *
  33  * Lustre Metadata Target (mdt) reintegration routines
  34  *
  35  * Author: Peter Braam <braam@clusterfs.com>
  36  * Author: Andreas Dilger <adilger@clusterfs.com>
  37  * Author: Phil Schwan <phil@clusterfs.com>
  38  * Author: Huang Hua <huanghua@clusterfs.com>
  39  * Author: Yury Umanets <umka@clusterfs.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_MDS
  43
  44 #include <lprocfs_status.h>
  45 #include "mdt_internal.h"
  46 #include <lustre_lmv.h>
  47
  48 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
  49                                      struct md_attr *ma)
  50 {
  51         ma->ma_need = MA_INODE;
  52         ma->ma_valid = 0;
  53 }
  54
  55 /**
  56  * Get version of object by fid.
  57  *
  58  * Return real version or ENOENT_VERSION if object doesn't exist
  59  */
  60 static void mdt_obj_version_get(struct mdt_thread_info *info,
  61                                 struct mdt_object *o, __u64 *version)
  62 {
  63         LASSERT(o);
  64
  65         if (mdt_object_exists(o) && !mdt_object_remote(o) &&
  66             !fid_is_obf(mdt_object_fid(o)))
  67                 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
  68         else
  69                 *version = ENOENT_VERSION;
  70         CDEBUG(D_INODE, "FID "DFID" version is %#llx\n",
  71                PFID(mdt_object_fid(o)), *version);
  72 }
  73
  74 /**
  75  * Check version is correct.
  76  *
  77  * Should be called only during replay.
  78  */
  79 static int mdt_version_check(struct ptlrpc_request *req,
  80                              __u64 version, int idx)
  81 {
  82         __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
  83
  84         ENTRY;
  85         if (!exp_connect_vbr(req->rq_export))
  86                 RETURN(0);
  87
  88         LASSERT(req_is_replay(req));
  89         /** VBR: version is checked always because costs nothing */
  90         LASSERT(idx < PTLRPC_NUM_VERSIONS);
  91         /** Sanity check for malformed buffers */
  92         if (pre_ver == NULL) {
  93                 CERROR("No versions in request buffer\n");
  94                 spin_lock(&req->rq_export->exp_lock);
  95                 req->rq_export->exp_vbr_failed = 1;
  96                 spin_unlock(&req->rq_export->exp_lock);
  97                 RETURN(-EOVERFLOW);
  98         } else if (pre_ver[idx] != version) {
  99                 CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n",
 100                        pre_ver[idx], version);
 101                 spin_lock(&req->rq_export->exp_lock);
 102                 req->rq_export->exp_vbr_failed = 1;
 103                 spin_unlock(&req->rq_export->exp_lock);
 104                 RETURN(-EOVERFLOW);
 105         }
 106         RETURN(0);
 107 }
 108
 109 /**
 110  * Save pre-versions in reply.
 111  */
 112 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
 113                              int idx)
 114 {
 115         __u64 *reply_ver;
 116
 117         if (!exp_connect_vbr(req->rq_export))
 118                 return;
 119
 120         LASSERT(!req_is_replay(req));
 121         LASSERT(req->rq_repmsg != NULL);
 122         reply_ver = lustre_msg_get_versions(req->rq_repmsg);
 123         if (reply_ver)
 124                 reply_ver[idx] = version;
 125 }
 126
 127 /**
 128  * Save enoent version, it is needed when it is obvious that object doesn't
 129  * exist, e.g. child during create.
 130  */
 131 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
 132 {
 133         /* save version of file name for replay, it must be ENOENT here */
 134         if (!req_is_replay(mdt_info_req(info))) {
 135                 info->mti_ver[idx] = ENOENT_VERSION;
 136                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 137         }
 138 }
 139
 140 /**
 141  * Get version from disk and save in reply buffer.
 142  *
 143  * Versions are saved in reply only during normal operations not replays.
 144  */
 145 void mdt_version_get_save(struct mdt_thread_info *info,
 146                           struct mdt_object *mto, int idx)
 147 {
 148         /* don't save versions during replay */
 149         if (!req_is_replay(mdt_info_req(info))) {
 150                 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 151                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 152         }
 153 }
 154
 155 /**
 156  * Get version from disk and check it, no save in reply.
 157  */
 158 int mdt_version_get_check(struct mdt_thread_info *info,
 159                           struct mdt_object *mto, int idx)
 160 {
 161         /* only check versions during replay */
 162         if (!req_is_replay(mdt_info_req(info)))
 163                 return 0;
 164
 165         mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 166         return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
 167 }
 168
 169 /**
 170  * Get version from disk and check if recovery or just save.
 171  */
 172 int mdt_version_get_check_save(struct mdt_thread_info *info,
 173                                struct mdt_object *mto, int idx)
 174 {
 175         int rc = 0;
 176
 177         mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 178         if (req_is_replay(mdt_info_req(info)))
 179                 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
 180                                        idx);
 181         else
 182                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 183         return rc;
 184 }
 185
 186 /**
 187  * Lookup with version checking.
 188  *
 189  * This checks version of 'name'. Many reint functions uses 'name' for child not
 190  * FID, therefore we need to get object by name and check its version.
 191  */
 192 int mdt_lookup_version_check(struct mdt_thread_info *info,
 193                              struct mdt_object *p,
 194                              const struct lu_name *lname,
 195                              struct lu_fid *fid, int idx)
 196 {
 197         int rc, vbrc;
 198
 199         rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
 200                         &info->mti_spec);
 201         /* Check version only during replay */
 202         if (!req_is_replay(mdt_info_req(info)))
 203                 return rc;
 204
 205         info->mti_ver[idx] = ENOENT_VERSION;
 206         if (rc == 0) {
 207                 struct mdt_object *child;
 208
 209                 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
 210                 if (likely(!IS_ERR(child))) {
 211                         mdt_obj_version_get(info, child, &info->mti_ver[idx]);
 212                         mdt_object_put(info->mti_env, child);
 213                 }
 214         }
 215         vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
 216         return vbrc ? vbrc : rc;
 217
 218 }
 219
 220 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
 221                              struct mdt_object *obj,
 222                              struct ldlm_enqueue_info *einfo,
 223                              int decref)
 224 {
 225         union ldlm_policy_data *policy = &mti->mti_policy;
 226         struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
 227         struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
 228         int i;
 229
 230         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 231         LASSERT(slave_locks);
 232
 233         memset(policy, 0, sizeof(*policy));
 234         policy->l_inodebits.bits = einfo->ei_inodebits;
 235         mdt_lock_handle_init(lh);
 236         mdt_lock_reg_init(lh, einfo->ei_mode);
 237         for (i = 0; i < slave_locks->ha_count; i++) {
 238                 if (test_bit(i, (void *)slave_locks->ha_map))
 239                         lh->mlh_rreg_lh = slave_locks->ha_handles[i];
 240                 else
 241                         lh->mlh_reg_lh = slave_locks->ha_handles[i];
 242                 mdt_object_unlock(mti, NULL, lh, decref);
 243                 slave_locks->ha_handles[i].cookie = 0ull;
 244         }
 245
 246         return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
 247                                 policy);
 248 }
 249
 250 static inline int mdt_object_striped(struct mdt_thread_info *mti,
 251                                      struct mdt_object *obj)
 252 {
 253         struct lu_device *bottom_dev;
 254         struct lu_object *bottom_obj;
 255         int rc;
 256
 257         if (!S_ISDIR(obj->mot_header.loh_attr))
 258                 return 0;
 259
 260         /* getxattr from bottom obj to avoid reading in shard FIDs */
 261         bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom);
 262         bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev,
 263                                           mdt_object_fid(obj), NULL);
 264         if (IS_ERR(bottom_obj))
 265                 return PTR_ERR(bottom_obj);
 266
 267         rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL,
 268                           XATTR_NAME_LMV);
 269         lu_object_put(mti->mti_env, bottom_obj);
 270
 271         return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc;
 272 }
 273
 274 /**
 275  * Lock slave stripes if necessary, the lock handles of slave stripes
 276  * will be stored in einfo->ei_cbdata.
 277  **/
 278 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
 279                            enum ldlm_mode mode, __u64 ibits,
 280                            struct ldlm_enqueue_info *einfo)
 281 {
 282         union ldlm_policy_data *policy = &mti->mti_policy;
 283
 284         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 285
 286         einfo->ei_type = LDLM_IBITS;
 287         einfo->ei_mode = mode;
 288         einfo->ei_cb_bl = mdt_remote_blocking_ast;
 289         einfo->ei_cb_local_bl = mdt_blocking_ast;
 290         einfo->ei_cb_cp = ldlm_completion_ast;
 291         einfo->ei_enq_slave = 1;
 292         einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
 293         einfo->ei_inodebits = ibits;
 294         memset(policy, 0, sizeof(*policy));
 295         policy->l_inodebits.bits = ibits;
 296
 297         return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
 298                               policy);
 299 }
 300
 301 int mdt_reint_striped_lock(struct mdt_thread_info *info,
 302                            struct mdt_object *o,
 303                            struct mdt_lock_handle *lh,
 304                            __u64 ibits,
 305                            struct ldlm_enqueue_info *einfo,
 306                            bool cos_incompat)
 307 {
 308         int rc;
 309
 310         LASSERT(!mdt_object_remote(o));
 311
 312         memset(einfo, 0, sizeof(*einfo));
 313
 314         rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
 315         if (rc)
 316                 return rc;
 317
 318         rc = mdt_object_striped(info, o);
 319         if (rc != 1) {
 320                 if (rc < 0)
 321                         mdt_object_unlock(info, o, lh, rc);
 322                 return rc;
 323         }
 324
 325         rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
 326         if (rc) {
 327                 mdt_object_unlock(info, o, lh, rc);
 328                 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
 329                         rc = 0;
 330         }
 331
 332         return rc;
 333 }
 334
 335 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
 336                               struct mdt_object *o,
 337                               struct mdt_lock_handle *lh,
 338                               struct ldlm_enqueue_info *einfo, int decref)
 339 {
 340         if (einfo->ei_cbdata)
 341                 mdt_unlock_slaves(info, o, einfo, decref);
 342         mdt_object_unlock(info, o, lh, decref);
 343 }
 344
 345 static int mdt_restripe(struct mdt_thread_info *info,
 346                         struct mdt_object *parent,
 347                         const struct lu_name *lname,
 348                         const struct lu_fid *tfid,
 349                         struct md_op_spec *spec,
 350                         struct md_attr *ma)
 351 {
 352         struct mdt_device *mdt = info->mti_mdt;
 353         struct lu_fid *fid = &info->mti_tmp_fid2;
 354         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 355         struct lmv_mds_md_v1 *lmv;
 356         struct mdt_object *child;
 357         struct mdt_lock_handle *lhp;
 358         struct mdt_lock_handle *lhc;
 359         struct mdt_body *repbody;
 360         int rc;
 361
 362         ENTRY;
 363         if (!mdt->mdt_enable_dir_restripe)
 364                 RETURN(-EPERM);
 365
 366         rc = mdt_version_get_check_save(info, parent, 0);
 367         if (rc)
 368                 RETURN(rc);
 369
 370         lhp = &info->mti_lh[MDT_LH_PARENT];
 371         mdt_lock_pdo_init(lhp, LCK_PW, lname);
 372         rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
 373                                    true);
 374         if (rc)
 375                 RETURN(rc);
 376
 377         rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV);
 378         if (rc)
 379                 GOTO(unlock_parent, rc);
 380
 381         if (ma->ma_valid & MA_LMV) {
 382                 /* don't allow restripe if parent dir layout is changing */
 383                 lmv = &ma->ma_lmv->lmv_md_v1;
 384                 if (!lmv_is_sane(lmv))
 385                         GOTO(unlock_parent, rc = -EBADF);
 386
 387                 if (lmv_is_layout_changing(lmv))
 388                         GOTO(unlock_parent, rc = -EBUSY);
 389         }
 390
 391         fid_zero(fid);
 392         rc = mdt_lookup_version_check(info, parent, lname, fid, 1);
 393         if (rc)
 394                 GOTO(unlock_parent, rc);
 395
 396         child = mdt_object_find(info->mti_env, mdt, fid);
 397         if (IS_ERR(child))
 398                 GOTO(unlock_parent, rc = PTR_ERR(child));
 399
 400         if (!mdt_object_exists(child))
 401                 GOTO(out_child, rc = -ENOENT);
 402
 403         if (mdt_object_remote(child)) {
 404                 struct mdt_body *repbody;
 405
 406                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 407                 if (!repbody)
 408                         GOTO(out_child, rc = -EPROTO);
 409
 410                 repbody->mbo_fid1 = *fid;
 411                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 412                 GOTO(out_child, rc = -EREMOTE);
 413         }
 414
 415         /* lock object */
 416         lhc = &info->mti_lh[MDT_LH_CHILD];
 417         mdt_lock_reg_init(lhc, LCK_EX);
 418
 419         /* enqueue object remote LOOKUP lock */
 420         if (mdt_object_remote(parent)) {
 421                 rc = mdt_remote_object_lock(info, parent, fid,
 422                                             &lhc->mlh_rreg_lh,
 423                                             lhc->mlh_rreg_mode,
 424                                             MDS_INODELOCK_LOOKUP, false);
 425                 if (rc != ELDLM_OK)
 426                         GOTO(out_child, rc);
 427         }
 428
 429         rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo,
 430                                     true);
 431         if (rc)
 432                 GOTO(unlock_child, rc);
 433
 434         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 435         rc = mdt_version_get_check_save(info, child, 1);
 436         if (rc)
 437                 GOTO(unlock_child, rc);
 438
 439         spin_lock(&mdt->mdt_restriper.mdr_lock);
 440         if (child->mot_restriping) {
 441                 /* race? */
 442                 spin_unlock(&mdt->mdt_restriper.mdr_lock);
 443                 GOTO(unlock_child, rc = -EBUSY);
 444         }
 445         child->mot_restriping = 1;
 446         spin_unlock(&mdt->mdt_restriper.mdr_lock);
 447
 448         *fid = *tfid;
 449         rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
 450         if (rc)
 451                 GOTO(restriping_clear, rc);
 452
 453         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 454         if (!repbody)
 455                 GOTO(restriping_clear, rc = -EPROTO);
 456
 457         mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid);
 458         EXIT;
 459
 460 restriping_clear:
 461         child->mot_restriping = 0;
 462 unlock_child:
 463         mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 464 out_child:
 465         mdt_object_put(info->mti_env, child);
 466 unlock_parent:
 467         mdt_object_unlock(info, parent, lhp, rc);
 468
 469         return rc;
 470 }
 471
 472 /*
 473  * VBR: we save three versions in reply:
 474  * 0 - parent. Check that parent version is the same during replay.
 475  * 1 - name. Version of 'name' if file exists with the same name or
 476  * ENOENT_VERSION, it is needed because file may appear due to missed replays.
 477  * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
 478  * check.
 479  */
 480 static int mdt_create(struct mdt_thread_info *info)
 481 {
 482         struct mdt_device *mdt = info->mti_mdt;
 483         struct mdt_object *parent;
 484         struct mdt_object *child;
 485         struct mdt_lock_handle *lh;
 486         struct mdt_body *repbody;
 487         struct md_attr *ma = &info->mti_attr;
 488         struct mdt_reint_record *rr = &info->mti_rr;
 489         struct md_op_spec *spec = &info->mti_spec;
 490         bool restripe = false;
 491         int rc;
 492
 493         ENTRY;
 494         DEBUG_REQ(D_INODE, mdt_info_req(info),
 495                   "Create ("DNAME"->"DFID") in "DFID,
 496                   PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
 497
 498         if (!fid_is_md_operative(rr->rr_fid1))
 499                 RETURN(-EPERM);
 500
 501         if (S_ISDIR(ma->ma_attr.la_mode) &&
 502             spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
 503                 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
 504                 struct lu_ucred *uc = mdt_ucred(info);
 505                 struct obd_export *exp = mdt_info_req(info)->rq_export;
 506
 507                 /* Only new clients can create remote dir( >= 2.4) and
 508                  * striped dir(>= 2.6), old client will return -ENOTSUPP
 509                  */
 510                 if (!mdt_is_dne_client(exp))
 511                         RETURN(-ENOTSUPP);
 512
 513                 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
 514                         if (!mdt_is_striped_client(exp))
 515                                 RETURN(-ENOTSUPP);
 516
 517                         if (!mdt->mdt_enable_striped_dir)
 518                                 RETURN(-EPERM);
 519                 } else if (!mdt->mdt_enable_remote_dir) {
 520                         RETURN(-EPERM);
 521                 }
 522
 523                 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
 524                     (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) ==
 525                     LMV_HASH_TYPE_CRUSH)
 526                         RETURN(-EPROTO);
 527
 528                 if (!md_capable(uc, CAP_SYS_ADMIN) &&
 529                     uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 530                     mdt->mdt_enable_remote_dir_gid != -1)
 531                         RETURN(-EPERM);
 532
 533                 /* restripe if later found dir exists, MDS_OPEN_CREAT means
 534                  * this is create only, don't try restripe.
 535                  */
 536                 if (mdt->mdt_enable_dir_restripe &&
 537                     le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT &&
 538                     !(spec->sp_cr_flags & MDS_OPEN_CREAT))
 539                         restripe = true;
 540         }
 541
 542         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 543
 544         parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 545         if (IS_ERR(parent))
 546                 RETURN(PTR_ERR(parent));
 547
 548         if (!mdt_object_exists(parent))
 549                 GOTO(put_parent, rc = -ENOENT);
 550
 551         /*
 552          * LU-10235: check if name exists locklessly first to avoid massive
 553          * lock recalls on existing directories.
 554          */
 555         rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
 556                                       &info->mti_tmp_fid1, 1);
 557         if (rc == 0) {
 558                 if (!restripe)
 559                         GOTO(put_parent, rc = -EEXIST);
 560
 561                 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
 562                                   ma);
 563         }
 564
 565         /* -ENOENT is expected here */
 566         if (rc != -ENOENT)
 567                 GOTO(put_parent, rc);
 568
 569         /* save version of file name for replay, it must be ENOENT here */
 570         mdt_enoent_version_save(info, 1);
 571
 572         OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
 573
 574         lh = &info->mti_lh[MDT_LH_PARENT];
 575         mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 576         rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
 577         if (rc)
 578                 GOTO(put_parent, rc);
 579
 580         if (!mdt_object_remote(parent)) {
 581                 rc = mdt_version_get_check_save(info, parent, 0);
 582                 if (rc)
 583                         GOTO(unlock_parent, rc);
 584         }
 585
 586         child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
 587         if (unlikely(IS_ERR(child)))
 588                 GOTO(unlock_parent, rc = PTR_ERR(child));
 589
 590         ma->ma_need = MA_INODE;
 591         ma->ma_valid = 0;
 592
 593         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 594                         OBD_FAIL_MDS_REINT_CREATE_WRITE);
 595
 596         /* Version of child will be updated on disk. */
 597         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 598         rc = mdt_version_get_check_save(info, child, 2);
 599         if (rc)
 600                 GOTO(put_child, rc);
 601
 602         /* Let lower layer know current lock mode. */
 603         info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
 604
 605         /*
 606          * Do not perform lookup sanity check. We know that name does
 607          * not exist.
 608          */
 609         info->mti_spec.sp_cr_lookup = 0;
 610         info->mti_spec.sp_feat = &dt_directory_features;
 611
 612         rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
 613                         mdt_object_child(child), &info->mti_spec, ma);
 614         if (rc == 0)
 615                 rc = mdt_attr_get_complex(info, child, ma);
 616
 617         if (rc < 0)
 618                 GOTO(put_child, rc);
 619
 620         /*
 621          * On DNE, we need to eliminate dependey between 'mkdir a' and
 622          * 'mkdir a/b' if b is a striped directory, to achieve this, two
 623          * things are done below:
 624          * 1. save child and slaves lock.
 625          * 2. if the child is a striped directory, relock parent so to
 626          *    compare against with COS locks to ensure parent was
 627          *    committed to disk.
 628          */
 629         if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
 630                 struct mdt_lock_handle *lhc;
 631                 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 632                 bool cos_incompat;
 633
 634                 rc = mdt_object_striped(info, child);
 635                 if (rc < 0)
 636                         GOTO(put_child, rc);
 637
 638                 cos_incompat = rc;
 639                 if (cos_incompat) {
 640                         if (!mdt_object_remote(parent)) {
 641                                 mdt_object_unlock(info, parent, lh, 1);
 642                                 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 643                                 rc = mdt_reint_object_lock(info, parent, lh,
 644                                                            MDS_INODELOCK_UPDATE,
 645                                                            true);
 646                                 if (rc)
 647                                         GOTO(put_child, rc);
 648                         }
 649                 }
 650
 651                 lhc = &info->mti_lh[MDT_LH_CHILD];
 652                 mdt_lock_handle_init(lhc);
 653                 mdt_lock_reg_init(lhc, LCK_PW);
 654                 rc = mdt_reint_striped_lock(info, child, lhc,
 655                                             MDS_INODELOCK_UPDATE, einfo,
 656                                             cos_incompat);
 657                 if (rc)
 658                         GOTO(put_child, rc);
 659
 660                 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 661         }
 662
 663         /* Return fid & attr to client. */
 664         if (ma->ma_valid & MA_INODE)
 665                 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
 666                                    mdt_object_fid(child));
 667         EXIT;
 668 put_child:
 669         mdt_object_put(info->mti_env, child);
 670 unlock_parent:
 671         mdt_object_unlock(info, parent, lh, rc);
 672 put_parent:
 673         mdt_object_put(info->mti_env, parent);
 674         return rc;
 675 }
 676
 677 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
 678                         struct md_attr *ma)
 679 {
 680         struct mdt_lock_handle  *lh;
 681         int do_vbr = ma->ma_attr.la_valid &
 682                         (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
 683         __u64 lockpart = MDS_INODELOCK_UPDATE;
 684         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 685         bool cos_incompat;
 686         int rc;
 687
 688         ENTRY;
 689         rc = mdt_object_striped(info, mo);
 690         if (rc < 0)
 691                 RETURN(rc);
 692
 693         cos_incompat = rc;
 694
 695         lh = &info->mti_lh[MDT_LH_PARENT];
 696         mdt_lock_reg_init(lh, LCK_PW);
 697
 698         /* Even though the new MDT will grant PERM lock to the old
 699          * client, but the old client will almost ignore that during
 700          * So it needs to revoke both LOOKUP and PERM lock here, so
 701          * both new and old client can cancel the dcache
 702          */
 703         if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
 704                 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
 705
 706         rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
 707                                     cos_incompat);
 708         if (rc != 0)
 709                 RETURN(rc);
 710
 711         /* all attrs are packed into mti_attr in unpack_setattr */
 712         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 713                        OBD_FAIL_MDS_REINT_SETATTR_WRITE);
 714
 715         /* VBR: update version if attr changed are important for recovery */
 716         if (do_vbr) {
 717                 /* update on-disk version of changed object */
 718                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
 719                 rc = mdt_version_get_check_save(info, mo, 0);
 720                 if (rc)
 721                         GOTO(out_unlock, rc);
 722         }
 723
 724         /* Ensure constant striping during chown(). See LU-2789. */
 725         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 726                 mutex_lock(&mo->mot_lov_mutex);
 727
 728         /* all attrs are packed into mti_attr in unpack_setattr */
 729         rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
 730
 731         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 732                 mutex_unlock(&mo->mot_lov_mutex);
 733
 734         if (rc != 0)
 735                 GOTO(out_unlock, rc);
 736         mdt_dom_obj_lvb_update(info->mti_env, mo, false);
 737         EXIT;
 738 out_unlock:
 739         mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
 740         return rc;
 741 }
 742
 743 /**
 744  * Check HSM flags and add HS_DIRTY flag if relevant.
 745  *
 746  * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
 747  * and is not RELEASED.
 748  */
 749 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
 750                         struct md_attr *ma)
 751 {
 752         struct lu_ucred *uc = mdt_ucred(info);
 753         cfs_cap_t cap_saved;
 754         int rc;
 755
 756         ENTRY;
 757         /* If the file was modified, add the dirty flag */
 758         ma->ma_need = MA_HSM;
 759         rc = mdt_attr_get_complex(info, mo, ma);
 760         if (rc) {
 761                 CERROR("file attribute read error for "DFID": %d.\n",
 762                         PFID(mdt_object_fid(mo)), rc);
 763                 RETURN(rc);
 764         }
 765
 766         /* If an up2date copy exists in the backend, add dirty flag */
 767         if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
 768             && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
 769                 ma->ma_hsm.mh_flags |= HS_DIRTY;
 770
 771                 /* Bump cap so that closes from non-owner writers can
 772                  * set the HSM state to dirty.
 773                  */
 774                 cap_saved = uc->uc_cap;
 775                 uc->uc_cap |= MD_CAP_TO_MASK(CAP_FOWNER);
 776                 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
 777                 uc->uc_cap = cap_saved;
 778                 if (rc)
 779                         CERROR("file attribute change error for "DFID": %d\n",
 780                                 PFID(mdt_object_fid(mo)), rc);
 781         }
 782
 783         RETURN(rc);
 784 }
 785
 786 static int mdt_reint_setattr(struct mdt_thread_info *info,
 787                              struct mdt_lock_handle *lhc)
 788 {
 789         struct mdt_device *mdt = info->mti_mdt;
 790         struct md_attr *ma = &info->mti_attr;
 791         struct mdt_reint_record *rr = &info->mti_rr;
 792         struct ptlrpc_request *req = mdt_info_req(info);
 793         struct mdt_object *mo;
 794         struct mdt_body *repbody;
 795         ktime_t kstart = ktime_get();
 796         int rc, rc2;
 797
 798         ENTRY;
 799         DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
 800                   (unsigned int)ma->ma_attr.la_valid);
 801
 802         if (info->mti_dlm_req)
 803                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 804
 805         OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
 806
 807         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 808         mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
 809         if (IS_ERR(mo))
 810                 GOTO(out, rc = PTR_ERR(mo));
 811
 812         if (!mdt_object_exists(mo))
 813                 GOTO(out_put, rc = -ENOENT);
 814
 815         if (mdt_object_remote(mo))
 816                 GOTO(out_put, rc = -EREMOTE);
 817
 818         ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
 819         /* revoke lease lock if size is going to be changed */
 820         if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
 821                      !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
 822                      atomic_read(&mo->mot_lease_count) > 0)) {
 823                 down_read(&mo->mot_open_sem);
 824
 825                 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
 826                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 827                         mdt_lock_reg_init(lhc, LCK_CW);
 828
 829                         rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
 830                         if (rc != 0) {
 831                                 up_read(&mo->mot_open_sem);
 832                                 GOTO(out_put, rc);
 833                         }
 834
 835                         /* revoke lease lock */
 836                         mdt_object_unlock(info, mo, lhc, 1);
 837                 }
 838                 up_read(&mo->mot_open_sem);
 839         }
 840
 841         if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
 842                 /* Check write access for the O_TRUNC case */
 843                 if (mdt_write_read(mo) < 0)
 844                         GOTO(out_put, rc = -ETXTBSY);
 845
 846                 /* LU-10286: compatibility check for FLR.
 847                  * Please check the comment in mdt_finish_open() for details
 848                  */
 849                 if (!exp_connect_flr(info->mti_exp) ||
 850                     !exp_connect_overstriping(info->mti_exp)) {
 851                         rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
 852                         if (rc < 0 && rc != -ENODATA)
 853                                 GOTO(out_put, rc);
 854
 855                         if (!exp_connect_flr(info->mti_exp)) {
 856                                 if (rc > 0 &&
 857                                     mdt_lmm_is_flr(info->mti_big_lmm))
 858                                         GOTO(out_put, rc = -EOPNOTSUPP);
 859                         }
 860
 861                         if (!exp_connect_overstriping(info->mti_exp)) {
 862                                 if (rc > 0 &&
 863                                     mdt_lmm_is_overstriping(info->mti_big_lmm))
 864                                         GOTO(out_put, rc = -EOPNOTSUPP);
 865                         }
 866                 }
 867
 868                 /* For truncate, the file size sent from client
 869                  * is believable, but the blocks are incorrect,
 870                  * which makes the block size in LSOM attribute
 871                  * inconsisent with the real block size.
 872                  */
 873                 rc = mdt_lsom_update(info, mo, true);
 874                 if (rc)
 875                         GOTO(out_put, rc);
 876         }
 877
 878         if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
 879                 if (ma->ma_valid & MA_LOV)
 880                         GOTO(out_put, rc = -EPROTO);
 881
 882                 /* MDT supports FMD for regular files due to Data-on-MDT */
 883                 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
 884                     ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
 885                         tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
 886                                        req->rq_xid);
 887
 888                         if (ma->ma_attr.la_valid & LA_MTIME) {
 889                                 rc = mdt_attr_get_pfid(info, mo, &ma->ma_pfid);
 890                                 if (!rc)
 891                                         ma->ma_valid |= MA_PFID;
 892                         }
 893                 }
 894
 895                 rc = mdt_attr_set(info, mo, ma);
 896                 if (rc)
 897                         GOTO(out_put, rc);
 898         } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
 899                    (ma->ma_valid & MA_INODE)) {
 900                 struct lu_buf *buf = &info->mti_buf;
 901                 struct lu_ucred *uc = mdt_ucred(info);
 902                 struct mdt_lock_handle *lh;
 903                 const char *name;
 904                 __u64 lockpart = MDS_INODELOCK_XATTR;
 905
 906                 /* reject if either remote or striped dir is disabled */
 907                 if (ma->ma_valid & MA_LMV) {
 908                         if (!mdt->mdt_enable_remote_dir ||
 909                             !mdt->mdt_enable_striped_dir)
 910                                 GOTO(out_put, rc = -EPERM);
 911
 912                         if (!md_capable(uc, CAP_SYS_ADMIN) &&
 913                             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 914                             mdt->mdt_enable_remote_dir_gid != -1)
 915                                 GOTO(out_put, rc = -EPERM);
 916                 }
 917
 918                 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
 919                         GOTO(out_put, rc = -ENOTDIR);
 920
 921                 if (ma->ma_attr.la_valid != 0)
 922                         GOTO(out_put, rc = -EPROTO);
 923
 924                 if (ma->ma_valid & MA_LOV) {
 925                         buf->lb_buf = ma->ma_lmm;
 926                         buf->lb_len = ma->ma_lmm_size;
 927                         name = XATTR_NAME_LOV;
 928                 } else {
 929                         struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
 930
 931                         buf->lb_buf = lmu;
 932                         buf->lb_len = ma->ma_lmv_size;
 933                         name = XATTR_NAME_DEFAULT_LMV;
 934                         /* force client to update dir default layout */
 935                         lockpart |= MDS_INODELOCK_LOOKUP;
 936                 }
 937
 938                 lh = &info->mti_lh[MDT_LH_PARENT];
 939                 mdt_lock_reg_init(lh, LCK_PW);
 940
 941                 rc = mdt_object_lock(info, mo, lh, lockpart);
 942                 if (rc != 0)
 943                         GOTO(out_put, rc);
 944
 945                 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
 946                                   name, 0);
 947
 948                 mdt_object_unlock(info, mo, lh, rc);
 949                 if (rc)
 950                         GOTO(out_put, rc);
 951         } else {
 952                 GOTO(out_put, rc = -EPROTO);
 953         }
 954
 955         /* If file data is modified, add the dirty flag */
 956         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
 957                 rc = mdt_add_dirty_flag(info, mo, ma);
 958
 959         ma->ma_need = MA_INODE;
 960         ma->ma_valid = 0;
 961         rc = mdt_attr_get_complex(info, mo, ma);
 962         if (rc != 0)
 963                 GOTO(out_put, rc);
 964
 965         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
 966
 967         EXIT;
 968 out_put:
 969         mdt_object_put(info->mti_env, mo);
 970 out:
 971         if (rc == 0)
 972                 mdt_counter_incr(req, LPROC_MDT_SETATTR,
 973                                  ktime_us_delta(ktime_get(), kstart));
 974
 975         mdt_client_compatibility(info);
 976         rc2 = mdt_fix_reply(info);
 977         if (rc == 0)
 978                 rc = rc2;
 979         return rc;
 980 }
 981
 982 static int mdt_reint_create(struct mdt_thread_info *info,
 983                             struct mdt_lock_handle *lhc)
 984 {
 985         struct ptlrpc_request   *req = mdt_info_req(info);
 986         ktime_t                 kstart = ktime_get();
 987         int                     rc;
 988
 989         ENTRY;
 990         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 991                 RETURN(err_serious(-ESTALE));
 992
 993         if (info->mti_dlm_req)
 994                 ldlm_request_cancel(mdt_info_req(info),
 995                                     info->mti_dlm_req, 0, LATF_SKIP);
 996
 997         if (!lu_name_is_valid(&info->mti_rr.rr_name))
 998                 RETURN(-EPROTO);
 999
1000         switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
1001         case S_IFDIR:
1002         case S_IFREG:
1003         case S_IFLNK:
1004         case S_IFCHR:
1005         case S_IFBLK:
1006         case S_IFIFO:
1007         case S_IFSOCK:
1008                 break;
1009         default:
1010                 CERROR("%s: Unsupported mode %o\n",
1011                        mdt_obd_name(info->mti_mdt),
1012                        info->mti_attr.ma_attr.la_mode);
1013                 RETURN(err_serious(-EOPNOTSUPP));
1014         }
1015
1016         rc = mdt_create(info);
1017         if (rc == 0) {
1018                 if ((info->mti_attr.ma_attr.la_mode & S_IFMT) == S_IFDIR)
1019                         mdt_counter_incr(req, LPROC_MDT_MKDIR,
1020                                          ktime_us_delta(ktime_get(), kstart));
1021                 else
1022                         /* Special file should stay on the same node as parent*/
1023                         mdt_counter_incr(req, LPROC_MDT_MKNOD,
1024                                          ktime_us_delta(ktime_get(), kstart));
1025         }
1026
1027         RETURN(rc);
1028 }
1029
1030 /*
1031  * VBR: save parent version in reply and child version getting by its name.
1032  * Version of child is getting and checking during its lookup. If
1033  */
1034 static int mdt_reint_unlink(struct mdt_thread_info *info,
1035                             struct mdt_lock_handle *lhc)
1036 {
1037         struct mdt_reint_record *rr = &info->mti_rr;
1038         struct ptlrpc_request *req = mdt_info_req(info);
1039         struct md_attr *ma = &info->mti_attr;
1040         struct lu_fid *child_fid = &info->mti_tmp_fid1;
1041         struct mdt_object *mp;
1042         struct mdt_object *mc;
1043         struct mdt_lock_handle *parent_lh;
1044         struct mdt_lock_handle *child_lh;
1045         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
1046         __u64 lock_ibits;
1047         bool cos_incompat = false;
1048         int no_name = 0;
1049         ktime_t kstart = ktime_get();
1050         int rc;
1051
1052         ENTRY;
1053         DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1054                   PNAME(&rr->rr_name));
1055
1056         if (info->mti_dlm_req)
1057                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1058
1059         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1060                 RETURN(err_serious(-ENOENT));
1061
1062         if (!fid_is_md_operative(rr->rr_fid1))
1063                 RETURN(-EPERM);
1064
1065         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1066         if (IS_ERR(mp))
1067                 RETURN(PTR_ERR(mp));
1068
1069         if (mdt_object_remote(mp)) {
1070                 cos_incompat = true;
1071         } else {
1072                 rc = mdt_version_get_check_save(info, mp, 0);
1073                 if (rc)
1074                         GOTO(put_parent, rc);
1075         }
1076
1077         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
1078         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1079 relock:
1080         parent_lh = &info->mti_lh[MDT_LH_PARENT];
1081         mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
1082         rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
1083                                    cos_incompat);
1084         if (rc != 0)
1085                 GOTO(put_parent, rc);
1086
1087         /* lookup child object along with version checking */
1088         fid_zero(child_fid);
1089         rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
1090         if (rc != 0) {
1091                 /* Name might not be able to find during resend of
1092                  * remote unlink, considering following case.
1093                  * dir_A is a remote directory, the name entry of
1094                  * dir_A is on MDT0, the directory is on MDT1,
1095                  *
1096                  * 1. client sends unlink req to MDT1.
1097                  * 2. MDT1 sends name delete update to MDT0.
1098                  * 3. name entry is being deleted in MDT0 synchronously.
1099                  * 4. MDT1 is restarted.
1100                  * 5. client resends unlink req to MDT1. So it can not
1101                  *    find the name entry on MDT0 anymore.
1102                  * In this case, MDT1 only needs to destory the local
1103                  * directory.
1104                  */
1105                 if (mdt_object_remote(mp) && rc == -ENOENT &&
1106                     !fid_is_zero(rr->rr_fid2) &&
1107                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1108                         no_name = 1;
1109                         *child_fid = *rr->rr_fid2;
1110                 } else {
1111                         GOTO(unlock_parent, rc);
1112                 }
1113         }
1114
1115         if (!fid_is_md_operative(child_fid))
1116                 GOTO(unlock_parent, rc = -EPERM);
1117
1118         /* We will lock the child regardless it is local or remote. No harm. */
1119         mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1120         if (IS_ERR(mc))
1121                 GOTO(unlock_parent, rc = PTR_ERR(mc));
1122
1123         if (!cos_incompat) {
1124                 rc = mdt_object_striped(info, mc);
1125                 if (rc < 0)
1126                         GOTO(put_child, rc);
1127
1128                 cos_incompat = rc;
1129                 if (cos_incompat) {
1130                         mdt_object_put(info->mti_env, mc);
1131                         mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1132                         goto relock;
1133                 }
1134         }
1135
1136         child_lh = &info->mti_lh[MDT_LH_CHILD];
1137         mdt_lock_reg_init(child_lh, LCK_EX);
1138         if (info->mti_spec.sp_rm_entry) {
1139                 struct lu_ucred *uc  = mdt_ucred(info);
1140
1141                 if (!mdt_is_dne_client(req->rq_export))
1142                         /* Return -ENOTSUPP for old client */
1143                         GOTO(put_child, rc = -ENOTSUPP);
1144
1145                 if (!md_capable(uc, CAP_SYS_ADMIN))
1146                         GOTO(put_child, rc = -EPERM);
1147
1148                 ma->ma_need = MA_INODE;
1149                 ma->ma_valid = 0;
1150                 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1151                                 NULL, &rr->rr_name, ma, no_name);
1152                 GOTO(put_child, rc);
1153         }
1154
1155         if (mdt_object_remote(mc)) {
1156                 struct mdt_body  *repbody;
1157
1158                 if (!fid_is_zero(rr->rr_fid2)) {
1159                         CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1160                                mdt_obd_name(info->mti_mdt),
1161                                PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1162                         GOTO(put_child, rc = -ENOENT);
1163                 }
1164                 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1165                        mdt_obd_name(info->mti_mdt),
1166                        PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1167
1168                 if (!mdt_is_dne_client(req->rq_export))
1169                         /* Return -ENOTSUPP for old client */
1170                         GOTO(put_child, rc = -ENOTSUPP);
1171
1172                 /* Revoke the LOOKUP lock of the remote object granted by
1173                  * this MDT. Since the unlink will happen on another MDT,
1174                  * it will release the LOOKUP lock right away. Then What
1175                  * would happen if another client try to grab the LOOKUP
1176                  * lock at the same time with unlink XXX
1177                  */
1178                 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1179                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1180                 LASSERT(repbody != NULL);
1181                 repbody->mbo_fid1 = *mdt_object_fid(mc);
1182                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1183                 GOTO(unlock_child, rc = -EREMOTE);
1184         }
1185         /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1186          * this now because a running HSM restore on the child (unlink
1187          * victim) will hold the layout lock. See LU-4002.
1188          */
1189         lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1190         if (mdt_object_remote(mp)) {
1191                 /* Enqueue lookup lock from parent MDT */
1192                 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1193                                             &child_lh->mlh_rreg_lh,
1194                                             child_lh->mlh_rreg_mode,
1195                                             MDS_INODELOCK_LOOKUP, false);
1196                 if (rc != ELDLM_OK)
1197                         GOTO(put_child, rc);
1198
1199                 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1200         }
1201
1202         rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1203                                     cos_incompat);
1204         if (rc != 0)
1205                 GOTO(put_child, rc);
1206
1207         /*
1208          * Now we can only make sure we need MA_INODE, in mdd layer, will check
1209          * whether need MA_LOV and MA_COOKIE.
1210          */
1211         ma->ma_need = MA_INODE;
1212         ma->ma_valid = 0;
1213
1214         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1215                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1216         /* save version when object is locked */
1217         mdt_version_get_save(info, mc, 1);
1218
1219         mutex_lock(&mc->mot_lov_mutex);
1220
1221         rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1222                         mdt_object_child(mc), &rr->rr_name, ma, no_name);
1223
1224         mutex_unlock(&mc->mot_lov_mutex);
1225         if (rc != 0)
1226                 GOTO(unlock_child, rc);
1227
1228         if (!lu_object_is_dying(&mc->mot_header)) {
1229                 rc = mdt_attr_get_complex(info, mc, ma);
1230                 if (rc)
1231                         GOTO(out_stat, rc);
1232         } else if (mdt_dom_check_for_discard(info, mc)) {
1233                 mdt_dom_discard_data(info, mc);
1234         }
1235         mdt_handle_last_unlink(info, mc, ma);
1236
1237 out_stat:
1238         if (ma->ma_valid & MA_INODE) {
1239                 switch (ma->ma_attr.la_mode & S_IFMT) {
1240                 case S_IFDIR:
1241                         mdt_counter_incr(req, LPROC_MDT_RMDIR,
1242                                          ktime_us_delta(ktime_get(), kstart));
1243                         break;
1244                 case S_IFREG:
1245                 case S_IFLNK:
1246                 case S_IFCHR:
1247                 case S_IFBLK:
1248                 case S_IFIFO:
1249                 case S_IFSOCK:
1250                         mdt_counter_incr(req, LPROC_MDT_UNLINK,
1251                                          ktime_us_delta(ktime_get(), kstart));
1252                         break;
1253                 default:
1254                         LASSERTF(0, "bad file type %o unlinking\n",
1255                                 ma->ma_attr.la_mode);
1256                 }
1257         }
1258
1259         EXIT;
1260
1261 unlock_child:
1262         mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1263 put_child:
1264         mdt_object_put(info->mti_env, mc);
1265 unlock_parent:
1266         mdt_object_unlock(info, mp, parent_lh, rc);
1267 put_parent:
1268         mdt_object_put(info->mti_env, mp);
1269         CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1270         return rc;
1271 }
1272
1273 /*
1274  * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1275  * name.
1276  */
1277 static int mdt_reint_link(struct mdt_thread_info *info,
1278                           struct mdt_lock_handle *lhc)
1279 {
1280         struct mdt_reint_record *rr = &info->mti_rr;
1281         struct ptlrpc_request   *req = mdt_info_req(info);
1282         struct md_attr          *ma = &info->mti_attr;
1283         struct mdt_object       *ms;
1284         struct mdt_object       *mp;
1285         struct mdt_lock_handle  *lhs;
1286         struct mdt_lock_handle  *lhp;
1287         ktime_t kstart = ktime_get();
1288         bool cos_incompat;
1289         int rc;
1290
1291         ENTRY;
1292         DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1293                   PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1294
1295         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1296                 RETURN(err_serious(-ENOENT));
1297
1298         if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE) ||
1299             OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND)) {
1300                 req->rq_no_reply = 1;
1301                 RETURN(err_serious(-ENOENT));
1302         }
1303
1304         if (info->mti_dlm_req)
1305                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1306
1307         /* Invalid case so return error immediately instead of
1308          * processing it
1309          */
1310         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1311                 RETURN(-EPERM);
1312
1313         if (!fid_is_md_operative(rr->rr_fid1) ||
1314             !fid_is_md_operative(rr->rr_fid2))
1315                 RETURN(-EPERM);
1316
1317         /* step 1: find target parent dir */
1318         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1319         if (IS_ERR(mp))
1320                 RETURN(PTR_ERR(mp));
1321
1322         rc = mdt_version_get_check_save(info, mp, 0);
1323         if (rc)
1324                 GOTO(put_parent, rc);
1325
1326         /* step 2: find source */
1327         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1328         if (IS_ERR(ms))
1329                 GOTO(put_parent, rc = PTR_ERR(ms));
1330
1331         if (!mdt_object_exists(ms)) {
1332                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1333                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1334                 GOTO(put_source, rc = -ENOENT);
1335         }
1336
1337         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1338
1339         lhp = &info->mti_lh[MDT_LH_PARENT];
1340         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1341         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1342                                    cos_incompat);
1343         if (rc != 0)
1344                 GOTO(put_source, rc);
1345
1346         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1347
1348         lhs = &info->mti_lh[MDT_LH_CHILD];
1349         mdt_lock_reg_init(lhs, LCK_EX);
1350         rc = mdt_reint_object_lock(info, ms, lhs,
1351                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1352                                    cos_incompat);
1353         if (rc != 0)
1354                 GOTO(unlock_parent, rc);
1355
1356         /* step 3: link it */
1357         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1358                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1359
1360         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1361         rc = mdt_version_get_check_save(info, ms, 1);
1362         if (rc)
1363                 GOTO(unlock_source, rc);
1364
1365         /** check target version by name during replay */
1366         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1367                                       &info->mti_tmp_fid1, 2);
1368         if (rc != 0 && rc != -ENOENT)
1369                 GOTO(unlock_source, rc);
1370         /* save version of file name for replay, it must be ENOENT here */
1371         if (!req_is_replay(mdt_info_req(info))) {
1372                 if (rc != -ENOENT) {
1373                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1374                                PNAME(&rr->rr_name));
1375                         GOTO(unlock_source, rc = -EEXIST);
1376                 }
1377                 info->mti_ver[2] = ENOENT_VERSION;
1378                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1379         }
1380
1381         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1382                       mdt_object_child(ms), &rr->rr_name, ma);
1383
1384         if (rc == 0)
1385                 mdt_counter_incr(req, LPROC_MDT_LINK,
1386                                  ktime_us_delta(ktime_get(), kstart));
1387
1388         EXIT;
1389 unlock_source:
1390         mdt_object_unlock(info, ms, lhs, rc);
1391 unlock_parent:
1392         mdt_object_unlock(info, mp, lhp, rc);
1393 put_source:
1394         mdt_object_put(info->mti_env, ms);
1395 put_parent:
1396         mdt_object_put(info->mti_env, mp);
1397         return rc;
1398 }
1399 /**
1400  * lock the part of the directory according to the hash of the name
1401  * (lh->mlh_pdo_hash) in parallel directory lock.
1402  */
1403 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1404                               struct mdt_lock_handle *lh,
1405                               struct mdt_object *obj, __u64 ibits,
1406                               bool cos_incompat)
1407 {
1408         struct ldlm_res_id *res = &info->mti_res_id;
1409         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1410         union ldlm_policy_data *policy = &info->mti_policy;
1411         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1412         int rc;
1413
1414         /*
1415          * Finish res_id initializing by name hash marking part of
1416          * directory which is taking modification.
1417          */
1418         LASSERT(lh->mlh_pdo_hash != 0);
1419         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1420         memset(policy, 0, sizeof(*policy));
1421         policy->l_inodebits.bits = ibits;
1422         if (cos_incompat &&
1423             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1424                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1425         /*
1426          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1427          * going to be sent to client. If it is - mdt_intent_policy() path will
1428          * fix it up and turn FL_LOCAL flag off.
1429          */
1430         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1431                           policy, res, dlmflags,
1432                           &info->mti_exp->exp_handle.h_cookie);
1433         return rc;
1434 }
1435
1436 /**
1437  * Get BFL lock for rename or migrate process.
1438  **/
1439 static int mdt_rename_lock(struct mdt_thread_info *info,
1440                            struct lustre_handle *lh)
1441 {
1442         int     rc;
1443
1444         ENTRY;
1445         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1446                 struct lu_fid *fid = &info->mti_tmp_fid1;
1447                 struct mdt_object *obj;
1448
1449                 /* XXX, right now, it has to use object API to
1450                  * enqueue lock cross MDT, so it will enqueue
1451                  * rename lock(with LUSTRE_BFL_FID) by root object
1452                  */
1453                 lu_root_fid(fid);
1454                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1455                 if (IS_ERR(obj))
1456                         RETURN(PTR_ERR(obj));
1457
1458                 rc = mdt_remote_object_lock(info, obj,
1459                                             &LUSTRE_BFL_FID, lh,
1460                                             LCK_EX,
1461                                             MDS_INODELOCK_UPDATE, false);
1462                 mdt_object_put(info->mti_env, obj);
1463         } else {
1464                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1465                 union ldlm_policy_data *policy = &info->mti_policy;
1466                 struct ldlm_res_id *res_id = &info->mti_res_id;
1467                 __u64 flags = 0;
1468
1469                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1470                 memset(policy, 0, sizeof(*policy));
1471                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1472                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1473                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1474                                             LDLM_IBITS, policy, LCK_EX, &flags,
1475                                             ldlm_blocking_ast,
1476                                             ldlm_completion_ast, NULL, NULL, 0,
1477                                             LVB_T_NONE,
1478                                             &info->mti_exp->exp_handle.h_cookie,
1479                                             lh);
1480                 RETURN(rc);
1481         }
1482         RETURN(rc);
1483 }
1484
1485 static void mdt_rename_unlock(struct lustre_handle *lh)
1486 {
1487         ENTRY;
1488         LASSERT(lustre_handle_is_used(lh));
1489         /* Cancel the single rename lock right away */
1490         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1491         EXIT;
1492 }
1493
1494 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1495                                                 const struct lu_fid *fid,
1496                                                 int idx)
1497 {
1498         struct mdt_object *dir;
1499         int rc;
1500
1501         ENTRY;
1502         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1503         if (IS_ERR(dir))
1504                 RETURN(dir);
1505
1506         /* check early, the real version will be saved after locking */
1507         rc = mdt_version_get_check(info, dir, idx);
1508         if (rc)
1509                 GOTO(out_put, rc);
1510
1511         if (!mdt_object_exists(dir))
1512                 GOTO(out_put, rc = -ENOENT);
1513
1514         if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1515                 GOTO(out_put, rc = -ENOTDIR);
1516
1517         RETURN(dir);
1518 out_put:
1519         mdt_object_put(info->mti_env, dir);
1520         return ERR_PTR(rc);
1521 }
1522
1523 /*
1524  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1525  * herein we don't really check it, just do revoke.
1526  */
1527 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1528                                   struct mdt_object *pobj,
1529                                   struct mdt_object *obj)
1530 {
1531         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1532         int rc;
1533
1534         mdt_lock_handle_init(lh);
1535         mdt_lock_reg_init(lh, LCK_EX);
1536
1537         if (mdt_object_remote(pobj)) {
1538                 /* don't bother to check if pobj and obj are on the same MDT. */
1539                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1540                                             &lh->mlh_rreg_lh, LCK_EX,
1541                                             MDS_INODELOCK_LOOKUP, false);
1542         } else if (mdt_object_remote(obj)) {
1543                 struct ldlm_res_id *res = &info->mti_res_id;
1544                 union ldlm_policy_data *policy = &info->mti_policy;
1545                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1546                                  LDLM_FL_COS_INCOMPAT;
1547
1548                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1549                 memset(policy, 0, sizeof(*policy));
1550                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1551                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1552                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1553                                   dlmflags, NULL);
1554         } else {
1555                 /* do nothing if both are local */
1556                 return 0;
1557         }
1558
1559         if (rc != ELDLM_OK)
1560                 return rc;
1561
1562         /*
1563          * TODO, currently we don't save this lock because there is no place to
1564          * hold this lock handle, but to avoid race we need to save this lock.
1565          */
1566         mdt_object_unlock(info, NULL, lh, 1);
1567
1568         return 0;
1569 }
1570
1571 /*
1572  * operation may takes locks of linkea, or directory stripes, group them in
1573  * different list.
1574  */
1575 struct mdt_sub_lock {
1576         struct mdt_object *msl_obj;
1577         struct mdt_lock_handle msl_lh;
1578         struct list_head msl_linkage;
1579 };
1580
1581 static void mdt_unlock_list(struct mdt_thread_info *info,
1582                             struct list_head *list, int decref)
1583 {
1584         struct mdt_sub_lock *msl;
1585         struct mdt_sub_lock *tmp;
1586
1587         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1588                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1589                 list_del(&msl->msl_linkage);
1590                 OBD_FREE_PTR(msl);
1591         }
1592 }
1593
1594 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1595                                              struct mdt_object *obj,
1596                                              struct mdt_lock_handle *lh,
1597                                              struct ldlm_enqueue_info *einfo,
1598                                              struct list_head *slave_locks,
1599                                              int decref)
1600 {
1601         if (mdt_object_remote(obj)) {
1602                 mdt_unlock_list(info, slave_locks, decref);
1603                 mdt_object_unlock(info, obj, lh, decref);
1604         } else {
1605                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1606         }
1607 }
1608
1609 /*
1610  * lock parents of links, and also check whether total locks don't exceed
1611  * RS_MAX_LOCKS.
1612  *
1613  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1614  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1615  * \retval      -ev negative errno upon error
1616  */
1617 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1618                                  struct mdt_object *pobj,
1619                                  const struct md_attr *ma,
1620                                  struct mdt_object *obj,
1621                                  struct mdt_lock_handle *lhp,
1622                                  struct ldlm_enqueue_info *peinfo,
1623                                  struct list_head *parent_slave_locks,
1624                                  struct list_head *link_locks)
1625 {
1626         struct mdt_device *mdt = info->mti_mdt;
1627         struct lu_buf *buf = &info->mti_big_buf;
1628         struct lu_name *lname = &info->mti_name;
1629         struct linkea_data ldata = { NULL };
1630         bool blocked = false;
1631         int local_lnkp_cnt = 0;
1632         int rc;
1633
1634         ENTRY;
1635         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1636                 RETURN(0);
1637
1638         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1639         if (buf->lb_buf == NULL)
1640                 RETURN(-ENOMEM);
1641
1642         ldata.ld_buf = buf;
1643         rc = mdt_links_read(info, obj, &ldata);
1644         if (rc) {
1645                 if (rc == -ENOENT || rc == -ENODATA)
1646                         rc = 0;
1647                 RETURN(rc);
1648         }
1649
1650         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1651              linkea_next_entry(&ldata)) {
1652                 struct mdt_object *lnkp;
1653                 struct mdt_sub_lock *msl;
1654                 struct lu_fid fid;
1655                 __u64 ibits;
1656
1657                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1658                                     &fid);
1659
1660                 /* check if it's also linked to parent */
1661                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1662                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1663                                PFID(&fid), PNAME(lname));
1664                         /* in case link is remote object, revoke LOOKUP lock */
1665                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1666                         continue;
1667                 }
1668
1669                 lnkp = NULL;
1670
1671                 /* check if it's linked to a stripe of parent */
1672                 if (ma->ma_valid & MA_LMV) {
1673                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1674                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1675                         int j = 0;
1676
1677                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1678                                 fid_le_to_cpu(stripe_fid,
1679                                               &lmv->lmv_stripe_fids[j]);
1680                                 if (lu_fid_eq(stripe_fid, &fid)) {
1681                                         CDEBUG(D_INFO, "skip stripe "DFID
1682                                                ", reovke "DNAME"\n",
1683                                                PFID(&fid), PNAME(lname));
1684                                         lnkp = mdt_object_find(info->mti_env,
1685                                                                mdt, &fid);
1686                                         if (IS_ERR(lnkp))
1687                                                 GOTO(out, rc = PTR_ERR(lnkp));
1688                                         break;
1689                                 }
1690                         }
1691
1692                         if (lnkp) {
1693                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1694                                                                    obj);
1695                                 mdt_object_put(info->mti_env, lnkp);
1696                                 continue;
1697                         }
1698                 }
1699
1700                 /* Check if it's already locked */
1701                 list_for_each_entry(msl, link_locks, msl_linkage) {
1702                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1703                                 CDEBUG(D_INFO,
1704                                        DFID" was locked, revoke "DNAME"\n",
1705                                        PFID(&fid), PNAME(lname));
1706                                 lnkp = msl->msl_obj;
1707                                 break;
1708                         }
1709                 }
1710
1711                 if (lnkp) {
1712                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1713                         continue;
1714                 }
1715
1716                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1717                        PFID(&fid), PNAME(lname));
1718
1719                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1720                 if (IS_ERR(lnkp)) {
1721                         CWARN("%s: cannot find obj "DFID": %ld\n",
1722                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1723                         continue;
1724                 }
1725
1726                 if (!mdt_object_exists(lnkp)) {
1727                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1728                               PFID(&fid), PNAME(lname));
1729                         mdt_object_put(info->mti_env, lnkp);
1730                         continue;
1731                 }
1732
1733                 if (!mdt_object_remote(lnkp))
1734                         local_lnkp_cnt++;
1735
1736                 OBD_ALLOC_PTR(msl);
1737                 if (msl == NULL)
1738                         GOTO(out, rc = -ENOMEM);
1739
1740                 /*
1741                  * we can't follow parent-child lock order like other MD
1742                  * operations, use lock_try here to avoid deadlock, if the lock
1743                  * cannot be taken, drop all locks taken, revoke the blocked
1744                  * one, and continue processing the remaining entries, and in
1745                  * the end of the loop restart from beginning.
1746                  */
1747                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1748                 ibits = 0;
1749                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1750                                          MDS_INODELOCK_UPDATE, true);
1751                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1752
1753                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1754                                PFID(&fid), PNAME(lname));
1755
1756                         mdt_unlock_list(info, link_locks, 1);
1757                         /* also unlock parent locks to avoid deadlock */
1758                         if (!blocked)
1759                                 mdt_migrate_object_unlock(info, pobj, lhp,
1760                                                           peinfo,
1761                                                           parent_slave_locks,
1762                                                           1);
1763
1764                         blocked = true;
1765
1766                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1767                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1768                                              MDS_INODELOCK_UPDATE);
1769                         if (rc) {
1770                                 mdt_object_put(info->mti_env, lnkp);
1771                                 OBD_FREE_PTR(msl);
1772                                 GOTO(out, rc);
1773                         }
1774
1775                         if (mdt_object_remote(lnkp)) {
1776                                 struct ldlm_lock *lock;
1777
1778                                 /*
1779                                  * for remote object, set lock cb_atomic,
1780                                  * so lock can be released in blocking_ast()
1781                                  * immediately, then the next lock_try will
1782                                  * have better chance of success.
1783                                  */
1784                                 lock = ldlm_handle2lock(
1785                                                 &msl->msl_lh.mlh_rreg_lh);
1786                                 LASSERT(lock != NULL);
1787                                 lock_res_and_lock(lock);
1788                                 ldlm_set_atomic_cb(lock);
1789                                 unlock_res_and_lock(lock);
1790                                 LDLM_LOCK_PUT(lock);
1791                         }
1792
1793                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1794                         OBD_FREE_PTR(msl);
1795                         continue;
1796                 }
1797
1798                 INIT_LIST_HEAD(&msl->msl_linkage);
1799                 msl->msl_obj = lnkp;
1800                 list_add_tail(&msl->msl_linkage, link_locks);
1801
1802                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1803         }
1804
1805         if (blocked)
1806                 GOTO(out, rc = -EBUSY);
1807
1808         EXIT;
1809 out:
1810         if (rc) {
1811                 mdt_unlock_list(info, link_locks, rc);
1812         } else if (local_lnkp_cnt > RS_MAX_LOCKS - 5) {
1813                 CDEBUG(D_INFO, "Too many links (%d), sync operations\n",
1814                        local_lnkp_cnt);
1815                 /*
1816                  * parent may have 3 local objects: master object and 2 stripes
1817                  * (if it's being migrated too); source may have 1 local objects
1818                  * as regular file; target has 1 local object.
1819                  * Note, source may have 2 local locks if it is directory but it
1820                  * can't have hardlinks, so it is not considered here.
1821                  */
1822                 rc = 1;
1823         }
1824         return rc;
1825 }
1826
1827 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1828                                   struct mdt_object *obj,
1829                                   const struct md_attr *ma,
1830                                   struct list_head *slave_locks)
1831 {
1832         struct mdt_device *mdt = info->mti_mdt;
1833         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1834         struct lu_fid *fid = &info->mti_tmp_fid1;
1835         struct mdt_object *slave;
1836         struct mdt_sub_lock *msl;
1837         int i;
1838         int rc;
1839
1840         ENTRY;
1841         LASSERT(mdt_object_remote(obj));
1842         LASSERT(ma->ma_valid & MA_LMV);
1843         LASSERT(lmv);
1844
1845         if (!lmv_is_sane(lmv))
1846                 RETURN(-EINVAL);
1847
1848         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1849                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1850
1851                 if (!fid_is_sane(fid))
1852                         continue;
1853
1854                 slave = mdt_object_find(info->mti_env, mdt, fid);
1855                 if (IS_ERR(slave))
1856                         GOTO(out, rc = PTR_ERR(slave));
1857
1858                 OBD_ALLOC_PTR(msl);
1859                 if (!msl) {
1860                         mdt_object_put(info->mti_env, slave);
1861                         GOTO(out, rc = -ENOMEM);
1862                 }
1863
1864                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1865                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1866                                            MDS_INODELOCK_UPDATE, true);
1867                 if (rc) {
1868                         OBD_FREE_PTR(msl);
1869                         mdt_object_put(info->mti_env, slave);
1870                         GOTO(out, rc);
1871                 }
1872
1873                 INIT_LIST_HEAD(&msl->msl_linkage);
1874                 msl->msl_obj = slave;
1875                 list_add_tail(&msl->msl_linkage, slave_locks);
1876         }
1877         EXIT;
1878
1879 out:
1880         if (rc)
1881                 mdt_unlock_list(info, slave_locks, rc);
1882         return rc;
1883 }
1884
1885 /* lock parent and its stripes */
1886 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1887                                    struct mdt_object *obj,
1888                                    const struct md_attr *ma,
1889                                    struct mdt_lock_handle *lh,
1890                                    struct ldlm_enqueue_info *einfo,
1891                                    struct list_head *slave_locks)
1892 {
1893         int rc;
1894
1895         if (mdt_object_remote(obj)) {
1896                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1897                                             &lh->mlh_rreg_lh, LCK_PW,
1898                                             MDS_INODELOCK_UPDATE, false);
1899                 if (rc != ELDLM_OK)
1900                         return rc;
1901
1902                 /*
1903                  * if obj is remote and striped, lock its stripes explicitly
1904                  * because it's not striped in LOD layer on this MDT.
1905                  */
1906                 if (ma->ma_valid & MA_LMV) {
1907                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1908                         if (rc)
1909                                 mdt_object_unlock(info, obj, lh, rc);
1910                 }
1911         } else {
1912                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1913                                             einfo, true);
1914         }
1915
1916         return rc;
1917 }
1918
1919 /*
1920  * in migration, object may be remote, and we need take full lock of it and its
1921  * stripes if it's directory, besides, object may be a remote object on its
1922  * parent, revoke its LOOKUP lock on where its parent is located.
1923  */
1924 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1925                                    struct mdt_object *pobj,
1926                                    struct mdt_object *obj,
1927                                    struct mdt_lock_handle *lh,
1928                                    struct ldlm_enqueue_info *einfo,
1929                                    struct list_head *slave_locks)
1930 {
1931         int rc;
1932
1933         if (mdt_object_remote(obj)) {
1934                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1935                 if (rc)
1936                         return rc;
1937
1938                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1939                                             &lh->mlh_rreg_lh, LCK_EX,
1940                                             MDS_INODELOCK_FULL, false);
1941                 if (rc != ELDLM_OK)
1942                         return rc;
1943
1944                 /*
1945                  * if obj is remote and striped, lock its stripes explicitly
1946                  * because it's not striped in LOD layer on this MDT.
1947                  */
1948                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1949                         struct md_attr *ma = &info->mti_attr;
1950
1951                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1952                         if (rc) {
1953                                 mdt_object_unlock(info, obj, lh, rc);
1954                                 return rc;
1955                         }
1956
1957                         if (ma->ma_valid & MA_LMV) {
1958                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1959                                                             slave_locks);
1960                                 if (rc)
1961                                         mdt_object_unlock(info, obj, lh, rc);
1962                         }
1963                 }
1964         } else {
1965                 if (mdt_object_remote(pobj)) {
1966                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1967                         if (rc)
1968                                 return rc;
1969                 }
1970
1971                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1972                                             einfo, true);
1973         }
1974
1975         return rc;
1976 }
1977
1978 /*
1979  * lookup source by name, if parent is striped directory, we need to find the
1980  * corresponding stripe where source is located, and then lookup there.
1981  *
1982  * besides, if parent is migrating too, and file is already in target stripe,
1983  * this should be a redo of 'lfs migrate' on client side.
1984  */
1985 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1986                               struct mdt_object *pobj,
1987                               const struct md_attr *ma,
1988                               const struct lu_name *lname,
1989                               struct mdt_object **spobj,
1990                               struct mdt_object **sobj)
1991 {
1992         const struct lu_env *env = info->mti_env;
1993         struct lu_fid *fid = &info->mti_tmp_fid1;
1994         struct mdt_object *stripe;
1995         int rc;
1996
1997         if (ma->ma_valid & MA_LMV) {
1998                 /* if parent is striped, lookup on corresponding stripe */
1999                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
2000
2001                 if (!lmv_is_sane(lmv))
2002                         return -EBADF;
2003
2004                 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
2005                                                   lname->ln_namelen);
2006                 if (rc < 0)
2007                         return rc;
2008
2009                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2010
2011                 stripe = mdt_object_find(env, info->mti_mdt, fid);
2012                 if (IS_ERR(stripe))
2013                         return PTR_ERR(stripe);
2014
2015                 fid_zero(fid);
2016                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
2017                                 &info->mti_spec);
2018                 if (rc == -ENOENT && lmv_is_layout_changing(lmv)) {
2019                         /*
2020                          * if parent layout is changeing, and lookup child
2021                          * failed on source stripe, lookup again on target
2022                          * stripe, if it exists, it means previous migration
2023                          * was interrupted, and current file was migrated
2024                          * already.
2025                          */
2026                         mdt_object_put(env, stripe);
2027
2028                         rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
2029                                                       lname->ln_namelen);
2030                         if (rc < 0)
2031                                 return rc;
2032
2033                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2034
2035                         stripe = mdt_object_find(env, info->mti_mdt, fid);
2036                         if (IS_ERR(stripe))
2037                                 return PTR_ERR(stripe);
2038
2039                         fid_zero(fid);
2040                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
2041                                         fid, &info->mti_spec);
2042                         mdt_object_put(env, stripe);
2043                         return rc ?: -EALREADY;
2044                 } else if (rc) {
2045                         mdt_object_put(env, stripe);
2046                         return rc;
2047                 }
2048         } else {
2049                 fid_zero(fid);
2050                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2051                                 &info->mti_spec);
2052                 if (rc)
2053                         return rc;
2054
2055                 stripe = pobj;
2056                 mdt_object_get(env, stripe);
2057         }
2058
2059         *spobj = stripe;
2060
2061         *sobj = mdt_object_find(env, info->mti_mdt, fid);
2062         if (IS_ERR(*sobj)) {
2063                 mdt_object_put(env, stripe);
2064                 rc = PTR_ERR(*sobj);
2065                 *spobj = NULL;
2066                 *sobj = NULL;
2067         }
2068
2069         return rc;
2070 }
2071
2072 /* end lease and close file for regular file */
2073 static int mdd_migrate_close(struct mdt_thread_info *info,
2074                              struct mdt_object *obj)
2075 {
2076         struct close_data *data;
2077         struct mdt_body *repbody;
2078         struct ldlm_lock *lease;
2079         int rc;
2080         int rc2;
2081
2082         rc = -EPROTO;
2083         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2084                                       RCL_CLIENT) ||
2085             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2086                                       RCL_CLIENT))
2087                 goto close;
2088
2089         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2090         if (!data)
2091                 goto close;
2092
2093         rc = -ESTALE;
2094         lease = ldlm_handle2lock(&data->cd_handle);
2095         if (!lease)
2096                 goto close;
2097
2098         /* check if the lease was already canceled */
2099         lock_res_and_lock(lease);
2100         rc = ldlm_is_cancel(lease);
2101         unlock_res_and_lock(lease);
2102
2103         if (rc) {
2104                 rc = -EAGAIN;
2105                 LDLM_DEBUG(lease, DFID" lease broken",
2106                            PFID(mdt_object_fid(obj)));
2107         }
2108
2109         /*
2110          * cancel server side lease, client side counterpart should have been
2111          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2112          */
2113         ldlm_lock_cancel(lease);
2114         ldlm_reprocess_all(lease->l_resource, lease);
2115         LDLM_LOCK_PUT(lease);
2116
2117 close:
2118         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2119         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2120         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2121
2122         return rc ?: rc2;
2123 }
2124
2125 /*
2126  * migrate file in below steps:
2127  *  1. lock parent and its stripes
2128  *  2. lookup source by name
2129  *  3. lock parents of source links if source is not directory
2130  *  4. reject if source is in HSM
2131  *  5. take source open_sem and close file if source is regular file
2132  *  6. lock source and its stripes if it's directory
2133  *  7. lock target so subsequent change to it can trigger COS
2134  *  8. migrate file
2135  *  9. unlock above locks
2136  * 10. sync device if source has links
2137  */
2138 int mdt_reint_migrate(struct mdt_thread_info *info,
2139                       struct mdt_lock_handle *unused)
2140 {
2141         const struct lu_env *env = info->mti_env;
2142         struct mdt_device *mdt = info->mti_mdt;
2143         struct ptlrpc_request *req = mdt_info_req(info);
2144         struct mdt_reint_record *rr = &info->mti_rr;
2145         struct lu_ucred *uc = mdt_ucred(info);
2146         struct md_attr *ma = &info->mti_attr;
2147         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
2148         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
2149         struct mdt_object *pobj;
2150         struct mdt_object *spobj = NULL;
2151         struct mdt_object *sobj = NULL;
2152         struct mdt_object *tobj;
2153         struct lustre_handle rename_lh = { 0 };
2154         struct mdt_lock_handle *lhp;
2155         struct mdt_lock_handle *lhs;
2156         struct mdt_lock_handle *lht;
2157         LIST_HEAD(parent_slave_locks);
2158         LIST_HEAD(child_slave_locks);
2159         LIST_HEAD(link_locks);
2160         int lock_retries = 5;
2161         bool open_sem_locked = false;
2162         bool do_sync = false;
2163         int rc;
2164
2165         ENTRY;
2166         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2167                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2168
2169         if (info->mti_dlm_req)
2170                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2171
2172         if (!fid_is_md_operative(rr->rr_fid1) ||
2173             !fid_is_md_operative(rr->rr_fid2))
2174                 RETURN(-EPERM);
2175
2176         /* don't allow migrate . or .. */
2177         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2178                 RETURN(-EBUSY);
2179
2180         if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2181                 RETURN(-EPERM);
2182
2183         if (uc && !md_capable(uc, CAP_SYS_ADMIN) &&
2184             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2185             mdt->mdt_enable_remote_dir_gid != -1)
2186                 RETURN(-EPERM);
2187
2188         /*
2189          * Note: do not enqueue rename lock for replay request, because
2190          * if other MDT holds rename lock, but being blocked to wait for
2191          * this MDT to finish its recovery, and the failover MDT can not
2192          * get rename lock, which will cause deadlock.
2193          *
2194          * req is NULL if this is called by directory auto-split.
2195          */
2196         if (req && !req_is_replay(req)) {
2197                 rc = mdt_rename_lock(info, &rename_lh);
2198                 if (rc != 0) {
2199                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2200                                mdt_obd_name(info->mti_mdt), rc);
2201                         RETURN(rc);
2202                 }
2203         }
2204
2205         /* pobj is master object of parent */
2206         pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2207         if (IS_ERR(pobj))
2208                 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2209
2210         if (req) {
2211                 rc = mdt_version_get_check(info, pobj, 0);
2212                 if (rc)
2213                         GOTO(put_parent, rc);
2214         }
2215
2216         if (!mdt_object_exists(pobj))
2217                 GOTO(put_parent, rc = -ENOENT);
2218
2219         if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2220                 GOTO(put_parent, rc = -ENOTDIR);
2221
2222         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2223         if (rc)
2224                 GOTO(put_parent, rc);
2225
2226 lock_parent:
2227         /* lock parent object */
2228         lhp = &info->mti_lh[MDT_LH_PARENT];
2229         mdt_lock_reg_init(lhp, LCK_PW);
2230         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2231                                      &parent_slave_locks);
2232         if (rc)
2233                 GOTO(put_parent, rc);
2234
2235         /*
2236          * spobj is the corresponding stripe against name if pobj is striped
2237          * directory, which is the real parent, and no need to lock, because
2238          * we've taken full lock of pobj.
2239          */
2240         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2241         if (rc)
2242                 GOTO(unlock_parent, rc);
2243
2244         /* lock parents of source links, and revoke LOOKUP lock of links */
2245         rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2246                                    &parent_slave_locks, &link_locks);
2247         if (rc == -EBUSY && lock_retries-- > 0) {
2248                 mdt_object_put(env, sobj);
2249                 mdt_object_put(env, spobj);
2250                 goto lock_parent;
2251         }
2252
2253         if (rc < 0)
2254                 GOTO(put_source, rc);
2255
2256         /*
2257          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2258          * with one request, if total lock count exceeds this limit, we will
2259          * drop all locks after migration, and synchronous device in the end.
2260          */
2261         do_sync = rc;
2262
2263         /* TODO: DoM migration is not supported, migrate dirent only */
2264         if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2265                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2266                 if (rc)
2267                         GOTO(unlock_links, rc);
2268
2269                 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2270                         info->mti_spec.sp_migrate_nsonly = 1;
2271         }
2272
2273         /* if migration HSM is allowed */
2274         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2275                 ma->ma_need = MA_HSM;
2276                 ma->ma_valid = 0;
2277                 rc = mdt_attr_get_complex(info, sobj, ma);
2278                 if (rc)
2279                         GOTO(unlock_links, rc);
2280
2281                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2282                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2283         }
2284
2285         /* end lease and close file for regular file */
2286         if (info->mti_spec.sp_migrate_close) {
2287                 /* try to hold open_sem so that nobody else can open the file */
2288                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2289                         /* close anyway */
2290                         mdd_migrate_close(info, sobj);
2291                         GOTO(unlock_links, rc = -EBUSY);
2292                 } else {
2293                         open_sem_locked = true;
2294                         rc = mdd_migrate_close(info, sobj);
2295                         if (rc)
2296                                 GOTO(unlock_open_sem, rc);
2297                 }
2298         }
2299
2300         /* lock source */
2301         lhs = &info->mti_lh[MDT_LH_OLD];
2302         mdt_lock_reg_init(lhs, LCK_EX);
2303         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2304                                      &child_slave_locks);
2305         if (rc)
2306                 GOTO(unlock_open_sem, rc);
2307
2308         /* lock target */
2309         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2310         if (IS_ERR(tobj))
2311                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2312
2313         lht = &info->mti_lh[MDT_LH_NEW];
2314         mdt_lock_reg_init(lht, LCK_EX);
2315         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2316         if (rc)
2317                 GOTO(put_target, rc);
2318
2319         /* Don't do lookup sanity check. We know name doesn't exist. */
2320         info->mti_spec.sp_cr_lookup = 0;
2321         info->mti_spec.sp_feat = &dt_directory_features;
2322
2323         rc = mdo_migrate(env, mdt_object_child(pobj),
2324                          mdt_object_child(sobj), &rr->rr_name,
2325                          mdt_object_child(tobj),
2326                          &info->mti_spec, ma);
2327         if (!rc)
2328                 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2329                                      LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2330         EXIT;
2331
2332         mdt_object_unlock(info, tobj, lht, rc);
2333 put_target:
2334         mdt_object_put(env, tobj);
2335 unlock_source:
2336         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2337                                   &child_slave_locks, rc);
2338 unlock_open_sem:
2339         if (open_sem_locked)
2340                 up_write(&sobj->mot_open_sem);
2341 unlock_links:
2342         /* if we've got too many locks to save into RPC,
2343          * then just commit before the locks are released
2344          */
2345         if (!rc && do_sync)
2346                 mdt_device_sync(env, mdt);
2347         mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2348 put_source:
2349         mdt_object_put(env, sobj);
2350         mdt_object_put(env, spobj);
2351 unlock_parent:
2352         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2353                                   &parent_slave_locks, rc);
2354 put_parent:
2355         mdt_object_put(env, pobj);
2356 unlock_rename:
2357         if (lustre_handle_is_used(&rename_lh))
2358                 mdt_rename_unlock(&rename_lh);
2359
2360         return rc;
2361 }
2362
2363 static int mdt_object_lock_save(struct mdt_thread_info *info,
2364                                 struct mdt_object *dir,
2365                                 struct mdt_lock_handle *lh,
2366                                 int idx, bool cos_incompat)
2367 {
2368         int rc;
2369
2370         /* we lock the target dir if it is local */
2371         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2372                                    cos_incompat);
2373         if (rc != 0)
2374                 return rc;
2375
2376         /* get and save correct version after locking */
2377         mdt_version_get_save(info, dir, idx);
2378         return 0;
2379 }
2380
2381 /*
2382  * determine lock order of sobj and tobj
2383  *
2384  * there are two situations we need to lock tobj before sobj:
2385  * 1. sobj is child of tobj
2386  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2387  *    larger than that of tobj
2388  *
2389  * \retval      1 lock tobj before sobj
2390  * \retval      0 lock sobj before tobj
2391  * \retval      -ev negative errno upon error
2392  */
2393 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2394                                            struct mdt_object *sobj,
2395                                            struct mdt_object *tobj)
2396 {
2397         struct md_attr *ma = &info->mti_attr;
2398         struct lu_fid *spfid = &info->mti_tmp_fid1;
2399         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2400         struct lmv_mds_md_v1 *lmv;
2401         __u32 sindex;
2402         __u32 tindex;
2403         int rc;
2404
2405         /* sobj and tobj are the same */
2406         if (sobj == tobj)
2407                 return 0;
2408
2409         if (fid_is_root(mdt_object_fid(sobj)))
2410                 return 0;
2411
2412         if (fid_is_root(mdt_object_fid(tobj)))
2413                 return 1;
2414
2415         /* check whether sobj is child of tobj */
2416         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2417                            mdt_object_fid(tobj));
2418         if (rc < 0)
2419                 return rc;
2420
2421         if (rc == 1)
2422                 return 1;
2423
2424         /* check whether sobj and tobj are children of the same parent */
2425         rc = mdt_attr_get_pfid(info, sobj, spfid);
2426         if (rc)
2427                 return rc;
2428
2429         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2430         if (rc)
2431                 return rc;
2432
2433         if (!lu_fid_eq(spfid, tpfid))
2434                 return 0;
2435
2436         /* check whether sobj and tobj are sibling stripes */
2437         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2438         if (rc)
2439                 return rc;
2440
2441         if (!(ma->ma_valid & MA_LMV))
2442                 return 0;
2443
2444         lmv = &ma->ma_lmv->lmv_md_v1;
2445         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2446                 return 0;
2447         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2448
2449         ma->ma_valid = 0;
2450         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2451         if (rc)
2452                 return rc;
2453
2454         if (!(ma->ma_valid & MA_LMV))
2455                 return -ENODATA;
2456
2457         lmv = &ma->ma_lmv->lmv_md_v1;
2458         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2459                 return -EINVAL;
2460         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2461
2462         /* check stripe index of sobj and tobj */
2463         if (sindex == tindex)
2464                 return -EINVAL;
2465
2466         return sindex < tindex ? 0 : 1;
2467 }
2468
2469 /*
2470  * lock rename source object.
2471  *
2472  * Both source and source parent may be remote, and source may be a remote
2473  * object on source parent, to avoid overriding lock handle, store remote
2474  * LOOKUP lock separately in @lhr.
2475  *
2476  * \retval      0 on success
2477  * \retval      -ev negative errno upon error
2478  */
2479 static int mdt_rename_source_lock(struct mdt_thread_info *info,
2480                                   struct mdt_object *parent,
2481                                   struct mdt_object *child,
2482                                   struct mdt_lock_handle *lhc,
2483                                   struct mdt_lock_handle *lhr,
2484                                   __u64 ibits,
2485                                   bool cos_incompat)
2486 {
2487         int rc;
2488
2489         rc = mdt_is_remote_object(info, parent, child);
2490         if (rc < 0)
2491                 return rc;
2492
2493         if (rc) {
2494                 /* enqueue remote LOOKUP lock from the parent MDT */
2495                 __u64 rmt_ibits = MDS_INODELOCK_LOOKUP;
2496
2497                 if (mdt_object_remote(parent)) {
2498                         rc = mdt_remote_object_lock(info, parent,
2499                                                     mdt_object_fid(child),
2500                                                     &lhr->mlh_rreg_lh,
2501                                                     lhr->mlh_rreg_mode,
2502                                                     rmt_ibits, false);
2503                         if (rc != ELDLM_OK)
2504                                 return rc;
2505                 } else {
2506                         LASSERT(mdt_object_remote(child));
2507                         rc = mdt_object_local_lock(info, child, lhr,
2508                                                    &rmt_ibits, 0, true);
2509                         if (rc < 0)
2510                                 return rc;
2511                 }
2512
2513                 ibits &= ~MDS_INODELOCK_LOOKUP;
2514         }
2515
2516         if (mdt_object_remote(child)) {
2517                 rc = mdt_remote_object_lock(info, child, mdt_object_fid(child),
2518                                             &lhc->mlh_rreg_lh,
2519                                             lhc->mlh_rreg_mode,
2520                                             ibits, false);
2521                 if (rc == ELDLM_OK)
2522                         rc = 0;
2523         } else {
2524                 rc = mdt_reint_object_lock(info, child, lhc, ibits,
2525                                            cos_incompat);
2526         }
2527
2528         if (!rc)
2529                 mdt_object_unlock(info, child, lhr, rc);
2530
2531         return rc;
2532 }
2533
2534 /*
2535  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2536  * 2 - srcdir child; 3 - tgtdir child.
2537  * Update on disk version of srcdir child.
2538  */
2539 static int mdt_reint_rename(struct mdt_thread_info *info,
2540                             struct mdt_lock_handle *unused)
2541 {
2542         struct mdt_device *mdt = info->mti_mdt;
2543         struct mdt_reint_record *rr = &info->mti_rr;
2544         struct md_attr *ma = &info->mti_attr;
2545         struct ptlrpc_request *req = mdt_info_req(info);
2546         struct mdt_object *msrcdir = NULL;
2547         struct mdt_object *mtgtdir = NULL;
2548         struct mdt_object *mold;
2549         struct mdt_object *mnew = NULL;
2550         struct lustre_handle rename_lh = { 0 };
2551         struct mdt_lock_handle *lh_srcdirp;
2552         struct mdt_lock_handle *lh_tgtdirp;
2553         struct mdt_lock_handle *lh_oldp = NULL;
2554         struct mdt_lock_handle *lh_rmt = NULL;
2555         struct mdt_lock_handle *lh_newp = NULL;
2556         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2557         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2558         __u64 lock_ibits;
2559         bool reverse = false, discard = false;
2560         bool cos_incompat;
2561         ktime_t kstart = ktime_get();
2562         int rc;
2563
2564         ENTRY;
2565         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2566                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2567                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2568
2569         if (info->mti_dlm_req)
2570                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2571
2572         if (!fid_is_md_operative(rr->rr_fid1) ||
2573             !fid_is_md_operative(rr->rr_fid2))
2574                 RETURN(-EPERM);
2575
2576         /* find both parents. */
2577         msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2578         if (IS_ERR(msrcdir))
2579                 RETURN(PTR_ERR(msrcdir));
2580
2581         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2582
2583         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2584                 mtgtdir = msrcdir;
2585                 mdt_object_get(info->mti_env, mtgtdir);
2586         } else {
2587                 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2588                 if (IS_ERR(mtgtdir))
2589                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2590         }
2591
2592         /*
2593          * Note: do not enqueue rename lock for replay request, because
2594          * if other MDT holds rename lock, but being blocked to wait for
2595          * this MDT to finish its recovery, and the failover MDT can not
2596          * get rename lock, which will cause deadlock.
2597          */
2598         if (!req_is_replay(req)) {
2599                 /*
2600                  * Normally rename RPC is handled on the MDT with the target
2601                  * directory (if target exists, it's on the MDT with the
2602                  * target), if the source directory is remote, it's a hint that
2603                  * source is remote too (this may not be true, but it won't
2604                  * cause any issue), return -EXDEV early to avoid taking
2605                  * rename_lock.
2606                  */
2607                 if (!mdt->mdt_enable_remote_rename &&
2608                     mdt_object_remote(msrcdir))
2609                         GOTO(out_put_tgtdir, rc = -EXDEV);
2610
2611                 /* This might be further relaxed in the future for regular file
2612                  * renames in different source and target parents. Start with
2613                  * only same-directory renames for simplicity and because this
2614                  * is by far the most the common use case.
2615                  */
2616                 if (msrcdir != mtgtdir) {
2617                         rc = mdt_rename_lock(info, &rename_lh);
2618                         if (rc != 0) {
2619                                 CERROR("%s: cannot lock for rename: rc = %d\n",
2620                                        mdt_obd_name(mdt), rc);
2621                                 GOTO(out_put_tgtdir, rc);
2622                         }
2623                 } else {
2624                         CDEBUG(D_INFO, "%s: samedir rename "DFID"/"DNAME"\n",
2625                                mdt_obd_name(mdt), PFID(rr->rr_fid1),
2626                                PNAME(&rr->rr_name));
2627                 }
2628         }
2629
2630         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2631         if (rc < 0)
2632                 GOTO(out_unlock_rename, rc);
2633
2634         reverse = rc;
2635
2636         /* source needs to be looked up after locking source parent, otherwise
2637          * this rename may race with unlink source, and cause rename hang, see
2638          * sanityn.sh 55b, so check parents first, if later we found source is
2639          * remote, relock parents.
2640          */
2641         cos_incompat = (mdt_object_remote(msrcdir) ||
2642                         mdt_object_remote(mtgtdir));
2643
2644         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2645
2646         /* lock parents in the proper order. */
2647         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2648         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2649
2650         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
2651         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
2652 relock:
2653         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2654         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2655
2656         if (reverse) {
2657                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2658                                           cos_incompat);
2659                 if (rc)
2660                         GOTO(out_unlock_rename, rc);
2661
2662                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2663
2664                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2665                                           cos_incompat);
2666                 if (rc != 0) {
2667                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2668                         GOTO(out_unlock_rename, rc);
2669                 }
2670         } else {
2671                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2672                                           cos_incompat);
2673                 if (rc)
2674                         GOTO(out_unlock_rename, rc);
2675
2676                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2677
2678                 if (mtgtdir != msrcdir) {
2679                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2680                                                   cos_incompat);
2681                 } else if (!mdt_object_remote(mtgtdir) &&
2682                            lh_srcdirp->mlh_pdo_hash !=
2683                            lh_tgtdirp->mlh_pdo_hash) {
2684                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2685                                                 MDS_INODELOCK_UPDATE,
2686                                                 cos_incompat);
2687                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2688                 }
2689                 if (rc != 0) {
2690                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2691                         GOTO(out_unlock_rename, rc);
2692                 }
2693         }
2694
2695         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2696         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2697
2698         /* find mold object. */
2699         fid_zero(old_fid);
2700         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2701         if (rc != 0)
2702                 GOTO(out_unlock_parents, rc);
2703
2704         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2705                 GOTO(out_unlock_parents, rc = -EINVAL);
2706
2707         if (!fid_is_md_operative(old_fid))
2708                 GOTO(out_unlock_parents, rc = -EPERM);
2709
2710         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2711         if (IS_ERR(mold))
2712                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2713
2714         if (!mdt_object_exists(mold)) {
2715                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2716                                 &mold->mot_obj,
2717                                 "object does not exist");
2718                 GOTO(out_put_old, rc = -ENOENT);
2719         }
2720
2721         if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2722                 GOTO(out_put_old, rc = -EXDEV);
2723
2724         /* Check if @mtgtdir is subdir of @mold, before locking child
2725          * to avoid reverse locking.
2726          */
2727         if (mtgtdir != msrcdir) {
2728                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2729                                    old_fid);
2730                 if (rc) {
2731                         if (rc == 1)
2732                                 rc = -EINVAL;
2733                         GOTO(out_put_old, rc);
2734                 }
2735         }
2736
2737         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2738         /* save version after locking */
2739         mdt_version_get_save(info, mold, 2);
2740
2741         if (!cos_incompat && mdt_object_remote(mold)) {
2742                 cos_incompat = true;
2743                 mdt_object_put(info->mti_env, mold);
2744                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2745                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2746                 goto relock;
2747         }
2748
2749         /* find mnew object:
2750          * mnew target object may not exist now
2751          * lookup with version checking
2752          */
2753         fid_zero(new_fid);
2754         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2755                                       3);
2756         if (rc == 0) {
2757                 /* the new_fid should have been filled at this moment */
2758                 if (lu_fid_eq(old_fid, new_fid))
2759                         GOTO(out_put_old, rc);
2760
2761                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2762                     lu_fid_eq(new_fid, rr->rr_fid2))
2763                         GOTO(out_put_old, rc = -EINVAL);
2764
2765                 if (!fid_is_md_operative(new_fid))
2766                         GOTO(out_put_old, rc = -EPERM);
2767
2768                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2769                 if (IS_ERR(mnew))
2770                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2771
2772                 if (!mdt_object_exists(mnew)) {
2773                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2774                                         &mnew->mot_obj,
2775                                         "object does not exist");
2776                         GOTO(out_put_new, rc = -ENOENT);
2777                 }
2778
2779                 if (mdt_object_remote(mnew)) {
2780                         struct mdt_body  *repbody;
2781
2782                         /* Always send rename req to the target child MDT */
2783                         repbody = req_capsule_server_get(info->mti_pill,
2784                                                          &RMF_MDT_BODY);
2785                         LASSERT(repbody != NULL);
2786                         repbody->mbo_fid1 = *new_fid;
2787                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2788                         GOTO(out_put_new, rc = -EXDEV);
2789                 }
2790                 /* Before locking the target dir, check we do not replace
2791                  * a dir with a non-dir, otherwise it may deadlock with
2792                  * link op which tries to create a link in this dir
2793                  * back to this non-dir.
2794                  */
2795                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2796                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2797                         GOTO(out_put_new, rc = -EISDIR);
2798
2799                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2800                 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2801                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2802                 mdt_lock_reg_init(lh_rmt, LCK_EX);
2803                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2804                 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2805                                             lh_rmt, lock_ibits, cos_incompat);
2806                 if (rc < 0)
2807                         GOTO(out_put_new, rc);
2808
2809                 /* Check if @msrcdir is subdir of @mnew, before locking child
2810                  * to avoid reverse locking.
2811                  */
2812                 if (mtgtdir != msrcdir) {
2813                         rc = mdo_is_subdir(info->mti_env,
2814                                            mdt_object_child(msrcdir), new_fid);
2815                         if (rc) {
2816                                 if (rc == 1)
2817                                         rc = -EINVAL;
2818                                 GOTO(out_unlock_old, rc);
2819                         }
2820                 }
2821
2822                 /* We used to acquire MDS_INODELOCK_FULL here but we
2823                  * can't do this now because a running HSM restore on
2824                  * the rename onto victim will hold the layout
2825                  * lock. See LU-4002.
2826                  */
2827
2828                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2829                 mdt_lock_reg_init(lh_newp, LCK_EX);
2830                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
2831                 if (mdt_object_remote(mtgtdir)) {
2832                         rc = mdt_remote_object_lock(info, mtgtdir,
2833                                                     mdt_object_fid(mnew),
2834                                                     &lh_newp->mlh_rreg_lh,
2835                                                     lh_newp->mlh_rreg_mode,
2836                                                     MDS_INODELOCK_LOOKUP,
2837                                                     false);
2838                         if (rc != ELDLM_OK)
2839                                 GOTO(out_unlock_old, rc);
2840
2841                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2842                 }
2843                 rc = mdt_reint_object_lock(info, mnew, lh_newp, lock_ibits,
2844                                            cos_incompat);
2845                 if (rc != 0)
2846                         GOTO(out_unlock_new, rc);
2847
2848                 /* get and save version after locking */
2849                 mdt_version_get_save(info, mnew, 3);
2850         } else if (rc != -ENOENT) {
2851                 GOTO(out_put_old, rc);
2852         } else {
2853                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2854                 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2855                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2856                 mdt_lock_reg_init(lh_rmt, LCK_EX);
2857                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2858                 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2859                                             lh_rmt, lock_ibits, cos_incompat);
2860                 if (rc != 0)
2861                         GOTO(out_put_old, rc);
2862
2863                 mdt_enoent_version_save(info, 3);
2864         }
2865
2866         /* step 5: rename it */
2867         mdt_reint_init_ma(info, ma);
2868
2869         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2870                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2871
2872         if (mnew != NULL)
2873                 mutex_lock(&mnew->mot_lov_mutex);
2874
2875         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2876                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2877                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2878                         &rr->rr_tgt_name, ma);
2879
2880         if (mnew != NULL)
2881                 mutex_unlock(&mnew->mot_lov_mutex);
2882
2883         /* handle last link of tgt object */
2884         if (rc == 0) {
2885                 mdt_counter_incr(req, LPROC_MDT_RENAME,
2886                                  ktime_us_delta(ktime_get(), kstart));
2887                 if (mnew) {
2888                         mdt_handle_last_unlink(info, mnew, ma);
2889                         discard = mdt_dom_check_for_discard(info, mnew);
2890                 }
2891                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2892                                          msrcdir, mtgtdir,
2893                                          ktime_us_delta(ktime_get(), kstart));
2894         }
2895
2896         EXIT;
2897 out_unlock_new:
2898         if (mnew != NULL)
2899                 mdt_object_unlock(info, mnew, lh_newp, rc);
2900 out_unlock_old:
2901         mdt_object_unlock(info, NULL, lh_rmt, rc);
2902         mdt_object_unlock(info, mold, lh_oldp, rc);
2903 out_put_new:
2904         if (mnew && !discard)
2905                 mdt_object_put(info->mti_env, mnew);
2906 out_put_old:
2907         mdt_object_put(info->mti_env, mold);
2908 out_unlock_parents:
2909         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2910         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2911 out_unlock_rename:
2912         if (lustre_handle_is_used(&rename_lh))
2913                 mdt_rename_unlock(&rename_lh);
2914 out_put_tgtdir:
2915         mdt_object_put(info->mti_env, mtgtdir);
2916 out_put_srcdir:
2917         mdt_object_put(info->mti_env, msrcdir);
2918
2919         /* The DoM discard can be done right in the place above where it is
2920          * assigned, meanwhile it is done here after rename unlock due to
2921          * compatibility with old clients, for them the discard blocks
2922          * the main thread until completion. Check LU-11359 for details.
2923          */
2924         if (discard) {
2925                 mdt_dom_discard_data(info, mnew);
2926                 mdt_object_put(info->mti_env, mnew);
2927         }
2928         return rc;
2929 }
2930
2931 static int mdt_reint_resync(struct mdt_thread_info *info,
2932                             struct mdt_lock_handle *lhc)
2933 {
2934         struct mdt_reint_record *rr = &info->mti_rr;
2935         struct ptlrpc_request *req = mdt_info_req(info);
2936         struct md_attr *ma = &info->mti_attr;
2937         struct mdt_object *mo;
2938         struct ldlm_lock *lease;
2939         struct mdt_body *repbody;
2940         struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2941         bool lease_broken;
2942         int rc, rc2;
2943
2944         ENTRY;
2945         DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2946
2947         if (info->mti_dlm_req)
2948                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2949
2950         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2951         if (IS_ERR(mo))
2952                 GOTO(out, rc = PTR_ERR(mo));
2953
2954         if (!mdt_object_exists(mo))
2955                 GOTO(out_obj, rc = -ENOENT);
2956
2957         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2958                 GOTO(out_obj, rc = -EINVAL);
2959
2960         if (mdt_object_remote(mo))
2961                 GOTO(out_obj, rc = -EREMOTE);
2962
2963         lease = ldlm_handle2lock(rr->rr_lease_handle);
2964         if (lease == NULL)
2965                 GOTO(out_obj, rc = -ESTALE);
2966
2967         /* It's really necessary to grab open_sem and check if the lease lock
2968          * has been lost. There would exist a concurrent writer coming in and
2969          * generating some dirty data in memory cache, the writeback would fail
2970          * after the layout version is increased by MDS_REINT_RESYNC RPC.
2971          */
2972         if (!down_write_trylock(&mo->mot_open_sem))
2973                 GOTO(out_put_lease, rc = -EBUSY);
2974
2975         lock_res_and_lock(lease);
2976         lease_broken = ldlm_is_cancel(lease);
2977         unlock_res_and_lock(lease);
2978         if (lease_broken)
2979                 GOTO(out_unlock, rc = -EBUSY);
2980
2981         /* the file has yet opened by anyone else after we took the lease. */
2982         layout.mlc_opc = MD_LAYOUT_RESYNC;
2983         lhc = &info->mti_lh[MDT_LH_LOCAL];
2984         rc = mdt_layout_change(info, mo, lhc, &layout);
2985         if (rc)
2986                 GOTO(out_unlock, rc);
2987
2988         mdt_object_unlock(info, mo, lhc, 0);
2989
2990         ma->ma_need = MA_INODE;
2991         ma->ma_valid = 0;
2992         rc = mdt_attr_get_complex(info, mo, ma);
2993         if (rc != 0)
2994                 GOTO(out_unlock, rc);
2995
2996         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2997         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2998
2999         EXIT;
3000 out_unlock:
3001         up_write(&mo->mot_open_sem);
3002 out_put_lease:
3003         LDLM_LOCK_PUT(lease);
3004 out_obj:
3005         mdt_object_put(info->mti_env, mo);
3006 out:
3007         mdt_client_compatibility(info);
3008         rc2 = mdt_fix_reply(info);
3009         if (rc == 0)
3010                 rc = rc2;
3011         return rc;
3012 }
3013
3014 struct mdt_reinter {
3015         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
3016         enum lprocfs_extra_opc mr_extra_opc;
3017 };
3018
3019 static const struct mdt_reinter mdt_reinters[] = {
3020         [REINT_SETATTR] = {
3021                 .mr_handler = &mdt_reint_setattr,
3022                 .mr_extra_opc = MDS_REINT_SETATTR,
3023         },
3024         [REINT_CREATE] = {
3025                 .mr_handler = &mdt_reint_create,
3026                 .mr_extra_opc = MDS_REINT_CREATE,
3027         },
3028         [REINT_LINK] = {
3029                 .mr_handler = &mdt_reint_link,
3030                 .mr_extra_opc = MDS_REINT_LINK,
3031         },
3032         [REINT_UNLINK] = {
3033                 .mr_handler = &mdt_reint_unlink,
3034                 .mr_extra_opc = MDS_REINT_UNLINK,
3035         },
3036         [REINT_RENAME] = {
3037                 .mr_handler = &mdt_reint_rename,
3038                 .mr_extra_opc = MDS_REINT_RENAME,
3039         },
3040         [REINT_OPEN] = {
3041                 .mr_handler = &mdt_reint_open,
3042                 .mr_extra_opc = MDS_REINT_OPEN,
3043         },
3044         [REINT_SETXATTR] = {
3045                 .mr_handler = &mdt_reint_setxattr,
3046                 .mr_extra_opc = MDS_REINT_SETXATTR,
3047         },
3048         [REINT_RMENTRY] = {
3049                 .mr_handler = &mdt_reint_unlink,
3050                 .mr_extra_opc = MDS_REINT_UNLINK,
3051         },
3052         [REINT_MIGRATE] = {
3053                 .mr_handler = &mdt_reint_migrate,
3054                 .mr_extra_opc = MDS_REINT_RENAME,
3055         },
3056         [REINT_RESYNC] = {
3057                 .mr_handler = &mdt_reint_resync,
3058                 .mr_extra_opc = MDS_REINT_RESYNC,
3059         },
3060 };
3061
3062 int mdt_reint_rec(struct mdt_thread_info *info,
3063                   struct mdt_lock_handle *lhc)
3064 {
3065         const struct mdt_reinter *mr;
3066         int rc;
3067
3068         ENTRY;
3069         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
3070                 RETURN(-EPROTO);
3071
3072         mr = &mdt_reinters[info->mti_rr.rr_opcode];
3073         if (mr->mr_handler == NULL)
3074                 RETURN(-EPROTO);
3075
3076         rc = (*mr->mr_handler)(info, lhc);
3077
3078         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
3079                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
3080
3081         RETURN(rc);
3082 }