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