lustre/mdt/mdt_reint.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/mdt/mdt_reint.c
  33  *
  34  * Lustre Metadata Target (mdt) reintegration routines
  35  *
  36  * Author: Peter Braam <braam@clusterfs.com>
  37  * Author: Andreas Dilger <adilger@clusterfs.com>
  38  * Author: Phil Schwan <phil@clusterfs.com>
  39  * Author: Huang Hua <huanghua@clusterfs.com>
  40  * Author: Yury Umanets <umka@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_MDS
  44
  45 #include <lprocfs_status.h>
  46 #include "mdt_internal.h"
  47 #include <lustre_lmv.h>
  48
  49 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
  50                                      struct md_attr *ma)
  51 {
  52         ma->ma_need = MA_INODE;
  53         ma->ma_valid = 0;
  54 }
  55
  56 /**
  57  * Get version of object by fid.
  58  *
  59  * Return real version or ENOENT_VERSION if object doesn't exist
  60  */
  61 static void mdt_obj_version_get(struct mdt_thread_info *info,
  62                                 struct mdt_object *o, __u64 *version)
  63 {
  64         LASSERT(o);
  65
  66         if (mdt_object_exists(o) && !mdt_object_remote(o) &&
  67             !fid_is_obf(mdt_object_fid(o)))
  68                 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
  69         else
  70                 *version = ENOENT_VERSION;
  71         CDEBUG(D_INODE, "FID "DFID" version is %#llx\n",
  72                PFID(mdt_object_fid(o)), *version);
  73 }
  74
  75 /**
  76  * Check version is correct.
  77  *
  78  * Should be called only during replay.
  79  */
  80 static int mdt_version_check(struct ptlrpc_request *req,
  81                              __u64 version, int idx)
  82 {
  83         __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
  84
  85         ENTRY;
  86         if (!exp_connect_vbr(req->rq_export))
  87                 RETURN(0);
  88
  89         LASSERT(req_is_replay(req));
  90         /** VBR: version is checked always because costs nothing */
  91         LASSERT(idx < PTLRPC_NUM_VERSIONS);
  92         /** Sanity check for malformed buffers */
  93         if (pre_ver == NULL) {
  94                 CERROR("No versions in request buffer\n");
  95                 spin_lock(&req->rq_export->exp_lock);
  96                 req->rq_export->exp_vbr_failed = 1;
  97                 spin_unlock(&req->rq_export->exp_lock);
  98                 RETURN(-EOVERFLOW);
  99         } else if (pre_ver[idx] != version) {
 100                 CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n",
 101                        pre_ver[idx], version);
 102                 spin_lock(&req->rq_export->exp_lock);
 103                 req->rq_export->exp_vbr_failed = 1;
 104                 spin_unlock(&req->rq_export->exp_lock);
 105                 RETURN(-EOVERFLOW);
 106         }
 107         RETURN(0);
 108 }
 109
 110 /**
 111  * Save pre-versions in reply.
 112  */
 113 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
 114                              int idx)
 115 {
 116         __u64 *reply_ver;
 117
 118         if (!exp_connect_vbr(req->rq_export))
 119                 return;
 120
 121         LASSERT(!req_is_replay(req));
 122         LASSERT(req->rq_repmsg != NULL);
 123         reply_ver = lustre_msg_get_versions(req->rq_repmsg);
 124         if (reply_ver)
 125                 reply_ver[idx] = version;
 126 }
 127
 128 /**
 129  * Save enoent version, it is needed when it is obvious that object doesn't
 130  * exist, e.g. child during create.
 131  */
 132 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
 133 {
 134         /* save version of file name for replay, it must be ENOENT here */
 135         if (!req_is_replay(mdt_info_req(info))) {
 136                 info->mti_ver[idx] = ENOENT_VERSION;
 137                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 138         }
 139 }
 140
 141 /**
 142  * Get version from disk and save in reply buffer.
 143  *
 144  * Versions are saved in reply only during normal operations not replays.
 145  */
 146 void mdt_version_get_save(struct mdt_thread_info *info,
 147                           struct mdt_object *mto, int idx)
 148 {
 149         /* don't save versions during replay */
 150         if (!req_is_replay(mdt_info_req(info))) {
 151                 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 152                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 153         }
 154 }
 155
 156 /**
 157  * Get version from disk and check it, no save in reply.
 158  */
 159 int mdt_version_get_check(struct mdt_thread_info *info,
 160                           struct mdt_object *mto, int idx)
 161 {
 162         /* only check versions during replay */
 163         if (!req_is_replay(mdt_info_req(info)))
 164                 return 0;
 165
 166         mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 167         return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
 168 }
 169
 170 /**
 171  * Get version from disk and check if recovery or just save.
 172  */
 173 int mdt_version_get_check_save(struct mdt_thread_info *info,
 174                                struct mdt_object *mto, int idx)
 175 {
 176         int rc = 0;
 177
 178         mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 179         if (req_is_replay(mdt_info_req(info)))
 180                 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
 181                                        idx);
 182         else
 183                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 184         return rc;
 185 }
 186
 187 /**
 188  * Lookup with version checking.
 189  *
 190  * This checks version of 'name'. Many reint functions uses 'name' for child not
 191  * FID, therefore we need to get object by name and check its version.
 192  */
 193 int mdt_lookup_version_check(struct mdt_thread_info *info,
 194                              struct mdt_object *p,
 195                              const struct lu_name *lname,
 196                              struct lu_fid *fid, int idx)
 197 {
 198         int rc, vbrc;
 199
 200         rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
 201                         &info->mti_spec);
 202         /* Check version only during replay */
 203         if (!req_is_replay(mdt_info_req(info)))
 204                 return rc;
 205
 206         info->mti_ver[idx] = ENOENT_VERSION;
 207         if (rc == 0) {
 208                 struct mdt_object *child;
 209
 210                 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
 211                 if (likely(!IS_ERR(child))) {
 212                         mdt_obj_version_get(info, child, &info->mti_ver[idx]);
 213                         mdt_object_put(info->mti_env, child);
 214                 }
 215         }
 216         vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
 217         return vbrc ? vbrc : rc;
 218
 219 }
 220
 221 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
 222                              struct mdt_object *obj,
 223                              struct ldlm_enqueue_info *einfo,
 224                              int decref)
 225 {
 226         union ldlm_policy_data *policy = &mti->mti_policy;
 227         struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
 228         struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
 229         int i;
 230
 231         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 232         LASSERT(slave_locks);
 233
 234         memset(policy, 0, sizeof(*policy));
 235         policy->l_inodebits.bits = einfo->ei_inodebits;
 236         mdt_lock_handle_init(lh);
 237         mdt_lock_reg_init(lh, einfo->ei_mode);
 238         for (i = 0; i < slave_locks->ha_count; i++) {
 239                 if (test_bit(i, (void *)slave_locks->ha_map))
 240                         lh->mlh_rreg_lh = slave_locks->ha_handles[i];
 241                 else
 242                         lh->mlh_reg_lh = slave_locks->ha_handles[i];
 243                 mdt_object_unlock(mti, NULL, lh, decref);
 244                 slave_locks->ha_handles[i].cookie = 0ull;
 245         }
 246
 247         return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
 248                                 policy);
 249 }
 250
 251 static inline int mdt_object_striped(struct mdt_thread_info *mti,
 252                                      struct mdt_object *obj)
 253 {
 254         struct lu_device *bottom_dev;
 255         struct lu_object *bottom_obj;
 256         int rc;
 257
 258         if (!S_ISDIR(obj->mot_header.loh_attr))
 259                 return 0;
 260
 261         /* getxattr from bottom obj to avoid reading in shard FIDs */
 262         bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom);
 263         bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev,
 264                                           mdt_object_fid(obj), NULL);
 265         if (IS_ERR(bottom_obj))
 266                 return PTR_ERR(bottom_obj);
 267
 268         rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL,
 269                           XATTR_NAME_LMV);
 270         lu_object_put(mti->mti_env, bottom_obj);
 271
 272         return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc;
 273 }
 274
 275 /**
 276  * Lock slave stripes if necessary, the lock handles of slave stripes
 277  * will be stored in einfo->ei_cbdata.
 278  **/
 279 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
 280                            enum ldlm_mode mode, __u64 ibits,
 281                            struct ldlm_enqueue_info *einfo)
 282 {
 283         union ldlm_policy_data *policy = &mti->mti_policy;
 284
 285         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 286
 287         einfo->ei_type = LDLM_IBITS;
 288         einfo->ei_mode = mode;
 289         einfo->ei_cb_bl = mdt_remote_blocking_ast;
 290         einfo->ei_cb_local_bl = mdt_blocking_ast;
 291         einfo->ei_cb_cp = ldlm_completion_ast;
 292         einfo->ei_enq_slave = 1;
 293         einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
 294         einfo->ei_inodebits = ibits;
 295         memset(policy, 0, sizeof(*policy));
 296         policy->l_inodebits.bits = ibits;
 297
 298         return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
 299                               policy);
 300 }
 301
 302 int mdt_reint_striped_lock(struct mdt_thread_info *info,
 303                            struct mdt_object *o,
 304                            struct mdt_lock_handle *lh,
 305                            __u64 ibits,
 306                            struct ldlm_enqueue_info *einfo,
 307                            bool cos_incompat)
 308 {
 309         int rc;
 310
 311         LASSERT(!mdt_object_remote(o));
 312
 313         memset(einfo, 0, sizeof(*einfo));
 314
 315         rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
 316         if (rc)
 317                 return rc;
 318
 319         rc = mdt_object_striped(info, o);
 320         if (rc != 1) {
 321                 if (rc < 0)
 322                         mdt_object_unlock(info, o, lh, rc);
 323                 return rc;
 324         }
 325
 326         rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
 327         if (rc) {
 328                 mdt_object_unlock(info, o, lh, rc);
 329                 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
 330                         rc = 0;
 331         }
 332
 333         return rc;
 334 }
 335
 336 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
 337                               struct mdt_object *o,
 338                               struct mdt_lock_handle *lh,
 339                               struct ldlm_enqueue_info *einfo, int decref)
 340 {
 341         if (einfo->ei_cbdata)
 342                 mdt_unlock_slaves(info, o, einfo, decref);
 343         mdt_object_unlock(info, o, lh, decref);
 344 }
 345
 346 static int mdt_restripe(struct mdt_thread_info *info,
 347                         struct mdt_object *parent,
 348                         const struct lu_name *lname,
 349                         const struct lu_fid *tfid,
 350                         struct md_op_spec *spec,
 351                         struct md_attr *ma)
 352 {
 353         struct mdt_device *mdt = info->mti_mdt;
 354         struct lu_fid *fid = &info->mti_tmp_fid2;
 355         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 356         struct lmv_mds_md_v1 *lmv;
 357         struct mdt_object *child;
 358         struct mdt_lock_handle *lhp;
 359         struct mdt_lock_handle *lhc;
 360         struct mdt_body *repbody;
 361         int rc;
 362
 363         ENTRY;
 364         if (!mdt->mdt_enable_dir_restripe)
 365                 RETURN(-EPERM);
 366
 367         rc = mdt_version_get_check_save(info, parent, 0);
 368         if (rc)
 369                 RETURN(rc);
 370
 371         lhp = &info->mti_lh[MDT_LH_PARENT];
 372         mdt_lock_pdo_init(lhp, LCK_PW, lname);
 373         rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
 374                                    true);
 375         if (rc)
 376                 RETURN(rc);
 377
 378         rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV);
 379         if (rc)
 380                 GOTO(unlock_parent, rc);
 381
 382         if (ma->ma_valid & MA_LMV) {
 383                 /* don't allow restripe if parent dir layout is changing */
 384                 lmv = &ma->ma_lmv->lmv_md_v1;
 385                 if (!lmv_is_sane(lmv))
 386                         GOTO(unlock_parent, rc = -EBADF);
 387
 388                 if (lmv_is_layout_changing(lmv))
 389                         GOTO(unlock_parent, rc = -EBUSY);
 390         }
 391
 392         fid_zero(fid);
 393         rc = mdt_lookup_version_check(info, parent, lname, fid, 1);
 394         if (rc)
 395                 GOTO(unlock_parent, rc);
 396
 397         child = mdt_object_find(info->mti_env, mdt, fid);
 398         if (IS_ERR(child))
 399                 GOTO(unlock_parent, rc = PTR_ERR(child));
 400
 401         if (!mdt_object_exists(child))
 402                 GOTO(out_child, rc = -ENOENT);
 403
 404         if (mdt_object_remote(child)) {
 405                 struct mdt_body *repbody;
 406
 407                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 408                 if (!repbody)
 409                         GOTO(out_child, rc = -EPROTO);
 410
 411                 repbody->mbo_fid1 = *fid;
 412                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 413                 GOTO(out_child, rc = -EREMOTE);
 414         }
 415
 416         /* lock object */
 417         lhc = &info->mti_lh[MDT_LH_CHILD];
 418         mdt_lock_reg_init(lhc, LCK_EX);
 419
 420         /* enqueue object remote LOOKUP lock */
 421         if (mdt_object_remote(parent)) {
 422                 rc = mdt_remote_object_lock(info, parent, fid,
 423                                             &lhc->mlh_rreg_lh,
 424                                             lhc->mlh_rreg_mode,
 425                                             MDS_INODELOCK_LOOKUP, false);
 426                 if (rc != ELDLM_OK)
 427                         GOTO(out_child, rc);
 428         }
 429
 430         rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo,
 431                                     true);
 432         if (rc)
 433                 GOTO(unlock_child, rc);
 434
 435         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 436         rc = mdt_version_get_check_save(info, child, 1);
 437         if (rc)
 438                 GOTO(unlock_child, rc);
 439
 440         spin_lock(&mdt->mdt_restriper.mdr_lock);
 441         if (child->mot_restriping) {
 442                 /* race? */
 443                 spin_unlock(&mdt->mdt_restriper.mdr_lock);
 444                 GOTO(unlock_child, rc = -EBUSY);
 445         }
 446         child->mot_restriping = 1;
 447         spin_unlock(&mdt->mdt_restriper.mdr_lock);
 448
 449         *fid = *tfid;
 450         rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
 451         if (rc)
 452                 GOTO(restriping_clear, rc);
 453
 454         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 455         if (!repbody)
 456                 GOTO(restriping_clear, rc = -EPROTO);
 457
 458         mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid);
 459         EXIT;
 460
 461 restriping_clear:
 462         child->mot_restriping = 0;
 463 unlock_child:
 464         mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 465 out_child:
 466         mdt_object_put(info->mti_env, child);
 467 unlock_parent:
 468         mdt_object_unlock(info, parent, lhp, rc);
 469
 470         return rc;
 471 }
 472
 473 /*
 474  * VBR: we save three versions in reply:
 475  * 0 - parent. Check that parent version is the same during replay.
 476  * 1 - name. Version of 'name' if file exists with the same name or
 477  * ENOENT_VERSION, it is needed because file may appear due to missed replays.
 478  * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
 479  * check.
 480  */
 481 static int mdt_create(struct mdt_thread_info *info)
 482 {
 483         struct mdt_device *mdt = info->mti_mdt;
 484         struct mdt_object *parent;
 485         struct mdt_object *child;
 486         struct mdt_lock_handle *lh;
 487         struct mdt_body *repbody;
 488         struct md_attr *ma = &info->mti_attr;
 489         struct mdt_reint_record *rr = &info->mti_rr;
 490         struct md_op_spec *spec = &info->mti_spec;
 491         bool restripe = false;
 492         int rc;
 493
 494         ENTRY;
 495         DEBUG_REQ(D_INODE, mdt_info_req(info),
 496                   "Create ("DNAME"->"DFID") in "DFID,
 497                   PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
 498
 499         if (!fid_is_md_operative(rr->rr_fid1))
 500                 RETURN(-EPERM);
 501
 502         if (S_ISDIR(ma->ma_attr.la_mode) &&
 503             spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
 504                 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
 505                 struct lu_ucred *uc = mdt_ucred(info);
 506                 struct obd_export *exp = mdt_info_req(info)->rq_export;
 507
 508                 /* Only new clients can create remote dir( >= 2.4) and
 509                  * striped dir(>= 2.6), old client will return -ENOTSUPP
 510                  */
 511                 if (!mdt_is_dne_client(exp))
 512                         RETURN(-ENOTSUPP);
 513
 514                 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
 515                         if (!mdt_is_striped_client(exp))
 516                                 RETURN(-ENOTSUPP);
 517
 518                         if (!mdt->mdt_enable_striped_dir)
 519                                 RETURN(-EPERM);
 520                 } else if (!mdt->mdt_enable_remote_dir) {
 521                         RETURN(-EPERM);
 522                 }
 523
 524                 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
 525                     (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) ==
 526                     LMV_HASH_TYPE_CRUSH)
 527                         RETURN(-EPROTO);
 528
 529                 if (!md_capable(uc, CAP_SYS_ADMIN) &&
 530                     uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 531                     mdt->mdt_enable_remote_dir_gid != -1)
 532                         RETURN(-EPERM);
 533
 534                 /* restripe if later found dir exists */
 535                 if (le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT)
 536                         restripe = true;
 537         }
 538
 539         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 540
 541         parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 542         if (IS_ERR(parent))
 543                 RETURN(PTR_ERR(parent));
 544
 545         if (!mdt_object_exists(parent))
 546                 GOTO(put_parent, rc = -ENOENT);
 547
 548         /*
 549          * LU-10235: check if name exists locklessly first to avoid massive
 550          * lock recalls on existing directories.
 551          */
 552         rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
 553                                       &info->mti_tmp_fid1, 1);
 554         if (rc == 0) {
 555                 if (!restripe)
 556                         GOTO(put_parent, rc = -EEXIST);
 557
 558                 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
 559                                   ma);
 560         }
 561
 562         /* -ENOENT is expected here */
 563         if (rc != -ENOENT)
 564                 GOTO(put_parent, rc);
 565
 566         /* save version of file name for replay, it must be ENOENT here */
 567         mdt_enoent_version_save(info, 1);
 568
 569         OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
 570
 571         lh = &info->mti_lh[MDT_LH_PARENT];
 572         mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 573         rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
 574         if (rc)
 575                 GOTO(put_parent, rc);
 576
 577         if (!mdt_object_remote(parent)) {
 578                 rc = mdt_version_get_check_save(info, parent, 0);
 579                 if (rc)
 580                         GOTO(unlock_parent, rc);
 581         }
 582
 583         child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
 584         if (unlikely(IS_ERR(child)))
 585                 GOTO(unlock_parent, rc = PTR_ERR(child));
 586
 587         ma->ma_need = MA_INODE;
 588         ma->ma_valid = 0;
 589
 590         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 591                         OBD_FAIL_MDS_REINT_CREATE_WRITE);
 592
 593         /* Version of child will be updated on disk. */
 594         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 595         rc = mdt_version_get_check_save(info, child, 2);
 596         if (rc)
 597                 GOTO(put_child, rc);
 598
 599         /* Let lower layer know current lock mode. */
 600         info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
 601
 602         /*
 603          * Do not perform lookup sanity check. We know that name does
 604          * not exist.
 605          */
 606         info->mti_spec.sp_cr_lookup = 0;
 607         info->mti_spec.sp_feat = &dt_directory_features;
 608
 609         rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
 610                         mdt_object_child(child), &info->mti_spec, ma);
 611         if (rc == 0)
 612                 rc = mdt_attr_get_complex(info, child, ma);
 613
 614         if (rc < 0)
 615                 GOTO(put_child, rc);
 616
 617         /*
 618          * On DNE, we need to eliminate dependey between 'mkdir a' and
 619          * 'mkdir a/b' if b is a striped directory, to achieve this, two
 620          * things are done below:
 621          * 1. save child and slaves lock.
 622          * 2. if the child is a striped directory, relock parent so to
 623          *    compare against with COS locks to ensure parent was
 624          *    committed to disk.
 625          */
 626         if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
 627                 struct mdt_lock_handle *lhc;
 628                 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 629                 bool cos_incompat;
 630
 631                 rc = mdt_object_striped(info, child);
 632                 if (rc < 0)
 633                         GOTO(put_child, rc);
 634
 635                 cos_incompat = rc;
 636                 if (cos_incompat) {
 637                         if (!mdt_object_remote(parent)) {
 638                                 mdt_object_unlock(info, parent, lh, 1);
 639                                 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 640                                 rc = mdt_reint_object_lock(info, parent, lh,
 641                                                            MDS_INODELOCK_UPDATE,
 642                                                            true);
 643                                 if (rc)
 644                                         GOTO(put_child, rc);
 645                         }
 646                 }
 647
 648                 lhc = &info->mti_lh[MDT_LH_CHILD];
 649                 mdt_lock_handle_init(lhc);
 650                 mdt_lock_reg_init(lhc, LCK_PW);
 651                 rc = mdt_reint_striped_lock(info, child, lhc,
 652                                             MDS_INODELOCK_UPDATE, einfo,
 653                                             cos_incompat);
 654                 if (rc)
 655                         GOTO(put_child, rc);
 656
 657                 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 658         }
 659
 660         /* Return fid & attr to client. */
 661         if (ma->ma_valid & MA_INODE)
 662                 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
 663                                    mdt_object_fid(child));
 664         EXIT;
 665 put_child:
 666         mdt_object_put(info->mti_env, child);
 667 unlock_parent:
 668         mdt_object_unlock(info, parent, lh, rc);
 669 put_parent:
 670         mdt_object_put(info->mti_env, parent);
 671         return rc;
 672 }
 673
 674 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
 675                         struct md_attr *ma)
 676 {
 677         struct mdt_lock_handle  *lh;
 678         int do_vbr = ma->ma_attr.la_valid &
 679                         (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
 680         __u64 lockpart = MDS_INODELOCK_UPDATE;
 681         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 682         bool cos_incompat;
 683         int rc;
 684
 685         ENTRY;
 686         rc = mdt_object_striped(info, mo);
 687         if (rc < 0)
 688                 RETURN(rc);
 689
 690         cos_incompat = rc;
 691
 692         lh = &info->mti_lh[MDT_LH_PARENT];
 693         mdt_lock_reg_init(lh, LCK_PW);
 694
 695         /* Even though the new MDT will grant PERM lock to the old
 696          * client, but the old client will almost ignore that during
 697          * So it needs to revoke both LOOKUP and PERM lock here, so
 698          * both new and old client can cancel the dcache
 699          */
 700         if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
 701                 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
 702
 703         rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
 704                                     cos_incompat);
 705         if (rc != 0)
 706                 RETURN(rc);
 707
 708         /* all attrs are packed into mti_attr in unpack_setattr */
 709         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 710                        OBD_FAIL_MDS_REINT_SETATTR_WRITE);
 711
 712         /* VBR: update version if attr changed are important for recovery */
 713         if (do_vbr) {
 714                 /* update on-disk version of changed object */
 715                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
 716                 rc = mdt_version_get_check_save(info, mo, 0);
 717                 if (rc)
 718                         GOTO(out_unlock, rc);
 719         }
 720
 721         /* Ensure constant striping during chown(). See LU-2789. */
 722         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 723                 mutex_lock(&mo->mot_lov_mutex);
 724
 725         /* all attrs are packed into mti_attr in unpack_setattr */
 726         rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
 727
 728         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 729                 mutex_unlock(&mo->mot_lov_mutex);
 730
 731         if (rc != 0)
 732                 GOTO(out_unlock, rc);
 733         mdt_dom_obj_lvb_update(info->mti_env, mo, false);
 734         EXIT;
 735 out_unlock:
 736         mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
 737         return rc;
 738 }
 739
 740 /**
 741  * Check HSM flags and add HS_DIRTY flag if relevant.
 742  *
 743  * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
 744  * and is not RELEASED.
 745  */
 746 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
 747                         struct md_attr *ma)
 748 {
 749         struct lu_ucred *uc = mdt_ucred(info);
 750         cfs_cap_t cap_saved;
 751         int rc;
 752
 753         ENTRY;
 754         /* If the file was modified, add the dirty flag */
 755         ma->ma_need = MA_HSM;
 756         rc = mdt_attr_get_complex(info, mo, ma);
 757         if (rc) {
 758                 CERROR("file attribute read error for "DFID": %d.\n",
 759                         PFID(mdt_object_fid(mo)), rc);
 760                 RETURN(rc);
 761         }
 762
 763         /* If an up2date copy exists in the backend, add dirty flag */
 764         if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
 765             && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
 766                 ma->ma_hsm.mh_flags |= HS_DIRTY;
 767
 768                 /* Bump cap so that closes from non-owner writers can
 769                  * set the HSM state to dirty.
 770                  */
 771                 cap_saved = uc->uc_cap;
 772                 uc->uc_cap |= MD_CAP_TO_MASK(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                 req->rq_no_reply = 1;
1297                 RETURN(err_serious(-ENOENT));
1298         }
1299
1300         if (info->mti_dlm_req)
1301                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1302
1303         /* Invalid case so return error immediately instead of
1304          * processing it
1305          */
1306         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1307                 RETURN(-EPERM);
1308
1309         if (!fid_is_md_operative(rr->rr_fid1) ||
1310             !fid_is_md_operative(rr->rr_fid2))
1311                 RETURN(-EPERM);
1312
1313         /* step 1: find target parent dir */
1314         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1315         if (IS_ERR(mp))
1316                 RETURN(PTR_ERR(mp));
1317
1318         rc = mdt_version_get_check_save(info, mp, 0);
1319         if (rc)
1320                 GOTO(put_parent, rc);
1321
1322         /* step 2: find source */
1323         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1324         if (IS_ERR(ms))
1325                 GOTO(put_parent, rc = PTR_ERR(ms));
1326
1327         if (!mdt_object_exists(ms)) {
1328                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1329                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1330                 GOTO(put_source, rc = -ENOENT);
1331         }
1332
1333         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1334
1335         lhp = &info->mti_lh[MDT_LH_PARENT];
1336         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1337         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1338                                    cos_incompat);
1339         if (rc != 0)
1340                 GOTO(put_source, rc);
1341
1342         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1343
1344         lhs = &info->mti_lh[MDT_LH_CHILD];
1345         mdt_lock_reg_init(lhs, LCK_EX);
1346         rc = mdt_reint_object_lock(info, ms, lhs,
1347                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1348                                    cos_incompat);
1349         if (rc != 0)
1350                 GOTO(unlock_parent, rc);
1351
1352         /* step 3: link it */
1353         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1354                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1355
1356         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1357         rc = mdt_version_get_check_save(info, ms, 1);
1358         if (rc)
1359                 GOTO(unlock_source, rc);
1360
1361         /** check target version by name during replay */
1362         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1363                                       &info->mti_tmp_fid1, 2);
1364         if (rc != 0 && rc != -ENOENT)
1365                 GOTO(unlock_source, rc);
1366         /* save version of file name for replay, it must be ENOENT here */
1367         if (!req_is_replay(mdt_info_req(info))) {
1368                 if (rc != -ENOENT) {
1369                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1370                                PNAME(&rr->rr_name));
1371                         GOTO(unlock_source, rc = -EEXIST);
1372                 }
1373                 info->mti_ver[2] = ENOENT_VERSION;
1374                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1375         }
1376
1377         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1378                       mdt_object_child(ms), &rr->rr_name, ma);
1379
1380         if (rc == 0)
1381                 mdt_counter_incr(req, LPROC_MDT_LINK,
1382                                  ktime_us_delta(ktime_get(), kstart));
1383
1384         EXIT;
1385 unlock_source:
1386         mdt_object_unlock(info, ms, lhs, rc);
1387 unlock_parent:
1388         mdt_object_unlock(info, mp, lhp, rc);
1389 put_source:
1390         mdt_object_put(info->mti_env, ms);
1391 put_parent:
1392         mdt_object_put(info->mti_env, mp);
1393         return rc;
1394 }
1395 /**
1396  * lock the part of the directory according to the hash of the name
1397  * (lh->mlh_pdo_hash) in parallel directory lock.
1398  */
1399 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1400                               struct mdt_lock_handle *lh,
1401                               struct mdt_object *obj, __u64 ibits,
1402                               bool cos_incompat)
1403 {
1404         struct ldlm_res_id *res = &info->mti_res_id;
1405         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1406         union ldlm_policy_data *policy = &info->mti_policy;
1407         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1408         int rc;
1409
1410         /*
1411          * Finish res_id initializing by name hash marking part of
1412          * directory which is taking modification.
1413          */
1414         LASSERT(lh->mlh_pdo_hash != 0);
1415         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1416         memset(policy, 0, sizeof(*policy));
1417         policy->l_inodebits.bits = ibits;
1418         if (cos_incompat &&
1419             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1420                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1421         /*
1422          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1423          * going to be sent to client. If it is - mdt_intent_policy() path will
1424          * fix it up and turn FL_LOCAL flag off.
1425          */
1426         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1427                           policy, res, dlmflags,
1428                           &info->mti_exp->exp_handle.h_cookie);
1429         return rc;
1430 }
1431
1432 /**
1433  * Get BFL lock for rename or migrate process.
1434  **/
1435 static int mdt_rename_lock(struct mdt_thread_info *info,
1436                            struct lustre_handle *lh)
1437 {
1438         int     rc;
1439
1440         ENTRY;
1441         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1442                 struct lu_fid *fid = &info->mti_tmp_fid1;
1443                 struct mdt_object *obj;
1444
1445                 /* XXX, right now, it has to use object API to
1446                  * enqueue lock cross MDT, so it will enqueue
1447                  * rename lock(with LUSTRE_BFL_FID) by root object
1448                  */
1449                 lu_root_fid(fid);
1450                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1451                 if (IS_ERR(obj))
1452                         RETURN(PTR_ERR(obj));
1453
1454                 rc = mdt_remote_object_lock(info, obj,
1455                                             &LUSTRE_BFL_FID, lh,
1456                                             LCK_EX,
1457                                             MDS_INODELOCK_UPDATE, false);
1458                 mdt_object_put(info->mti_env, obj);
1459         } else {
1460                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1461                 union ldlm_policy_data *policy = &info->mti_policy;
1462                 struct ldlm_res_id *res_id = &info->mti_res_id;
1463                 __u64 flags = 0;
1464
1465                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1466                 memset(policy, 0, sizeof(*policy));
1467                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1468                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1469                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1470                                             LDLM_IBITS, policy, LCK_EX, &flags,
1471                                             ldlm_blocking_ast,
1472                                             ldlm_completion_ast, NULL, NULL, 0,
1473                                             LVB_T_NONE,
1474                                             &info->mti_exp->exp_handle.h_cookie,
1475                                             lh);
1476                 RETURN(rc);
1477         }
1478         RETURN(rc);
1479 }
1480
1481 static void mdt_rename_unlock(struct lustre_handle *lh)
1482 {
1483         ENTRY;
1484         LASSERT(lustre_handle_is_used(lh));
1485         /* Cancel the single rename lock right away */
1486         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1487         EXIT;
1488 }
1489
1490 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1491                                                 const struct lu_fid *fid,
1492                                                 int idx)
1493 {
1494         struct mdt_object *dir;
1495         int rc;
1496
1497         ENTRY;
1498         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1499         if (IS_ERR(dir))
1500                 RETURN(dir);
1501
1502         /* check early, the real version will be saved after locking */
1503         rc = mdt_version_get_check(info, dir, idx);
1504         if (rc)
1505                 GOTO(out_put, rc);
1506
1507         if (!mdt_object_exists(dir))
1508                 GOTO(out_put, rc = -ENOENT);
1509
1510         if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1511                 GOTO(out_put, rc = -ENOTDIR);
1512
1513         RETURN(dir);
1514 out_put:
1515         mdt_object_put(info->mti_env, dir);
1516         return ERR_PTR(rc);
1517 }
1518
1519 /*
1520  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1521  * herein we don't really check it, just do revoke.
1522  */
1523 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1524                                   struct mdt_object *pobj,
1525                                   struct mdt_object *obj)
1526 {
1527         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1528         int rc;
1529
1530         mdt_lock_handle_init(lh);
1531         mdt_lock_reg_init(lh, LCK_EX);
1532
1533         if (mdt_object_remote(pobj)) {
1534                 /* don't bother to check if pobj and obj are on the same MDT. */
1535                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1536                                             &lh->mlh_rreg_lh, LCK_EX,
1537                                             MDS_INODELOCK_LOOKUP, false);
1538         } else if (mdt_object_remote(obj)) {
1539                 struct ldlm_res_id *res = &info->mti_res_id;
1540                 union ldlm_policy_data *policy = &info->mti_policy;
1541                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1542                                  LDLM_FL_COS_INCOMPAT;
1543
1544                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1545                 memset(policy, 0, sizeof(*policy));
1546                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1547                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1548                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1549                                   dlmflags, NULL);
1550         } else {
1551                 /* do nothing if both are local */
1552                 return 0;
1553         }
1554
1555         if (rc != ELDLM_OK)
1556                 return rc;
1557
1558         /*
1559          * TODO, currently we don't save this lock because there is no place to
1560          * hold this lock handle, but to avoid race we need to save this lock.
1561          */
1562         mdt_object_unlock(info, NULL, lh, 1);
1563
1564         return 0;
1565 }
1566
1567 /*
1568  * operation may takes locks of linkea, or directory stripes, group them in
1569  * different list.
1570  */
1571 struct mdt_sub_lock {
1572         struct mdt_object *msl_obj;
1573         struct mdt_lock_handle msl_lh;
1574         struct list_head msl_linkage;
1575 };
1576
1577 static void mdt_unlock_list(struct mdt_thread_info *info,
1578                             struct list_head *list, int decref)
1579 {
1580         struct mdt_sub_lock *msl;
1581         struct mdt_sub_lock *tmp;
1582
1583         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1584                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1585                 list_del(&msl->msl_linkage);
1586                 OBD_FREE_PTR(msl);
1587         }
1588 }
1589
1590 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1591                                              struct mdt_object *obj,
1592                                              struct mdt_lock_handle *lh,
1593                                              struct ldlm_enqueue_info *einfo,
1594                                              struct list_head *slave_locks,
1595                                              int decref)
1596 {
1597         if (mdt_object_remote(obj)) {
1598                 mdt_unlock_list(info, slave_locks, decref);
1599                 mdt_object_unlock(info, obj, lh, decref);
1600         } else {
1601                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1602         }
1603 }
1604
1605 /*
1606  * lock parents of links, and also check whether total locks don't exceed
1607  * RS_MAX_LOCKS.
1608  *
1609  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1610  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1611  * \retval      -ev negative errno upon error
1612  */
1613 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1614                                  struct mdt_object *pobj,
1615                                  const struct md_attr *ma,
1616                                  struct mdt_object *obj,
1617                                  struct mdt_lock_handle *lhp,
1618                                  struct ldlm_enqueue_info *peinfo,
1619                                  struct list_head *parent_slave_locks,
1620                                  struct list_head *link_locks)
1621 {
1622         struct mdt_device *mdt = info->mti_mdt;
1623         struct lu_buf *buf = &info->mti_big_buf;
1624         struct lu_name *lname = &info->mti_name;
1625         struct linkea_data ldata = { NULL };
1626         bool blocked = false;
1627         int local_lnkp_cnt = 0;
1628         int rc;
1629
1630         ENTRY;
1631         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1632                 RETURN(0);
1633
1634         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1635         if (buf->lb_buf == NULL)
1636                 RETURN(-ENOMEM);
1637
1638         ldata.ld_buf = buf;
1639         rc = mdt_links_read(info, obj, &ldata);
1640         if (rc) {
1641                 if (rc == -ENOENT || rc == -ENODATA)
1642                         rc = 0;
1643                 RETURN(rc);
1644         }
1645
1646         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1647              linkea_next_entry(&ldata)) {
1648                 struct mdt_object *lnkp;
1649                 struct mdt_sub_lock *msl;
1650                 struct lu_fid fid;
1651                 __u64 ibits;
1652
1653                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1654                                     &fid);
1655
1656                 /* check if it's also linked to parent */
1657                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1658                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1659                                PFID(&fid), PNAME(lname));
1660                         /* in case link is remote object, revoke LOOKUP lock */
1661                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1662                         continue;
1663                 }
1664
1665                 lnkp = NULL;
1666
1667                 /* check if it's linked to a stripe of parent */
1668                 if (ma->ma_valid & MA_LMV) {
1669                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1670                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1671                         int j = 0;
1672
1673                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1674                                 fid_le_to_cpu(stripe_fid,
1675                                               &lmv->lmv_stripe_fids[j]);
1676                                 if (lu_fid_eq(stripe_fid, &fid)) {
1677                                         CDEBUG(D_INFO, "skip stripe "DFID
1678                                                ", reovke "DNAME"\n",
1679                                                PFID(&fid), PNAME(lname));
1680                                         lnkp = mdt_object_find(info->mti_env,
1681                                                                mdt, &fid);
1682                                         if (IS_ERR(lnkp))
1683                                                 GOTO(out, rc = PTR_ERR(lnkp));
1684                                         break;
1685                                 }
1686                         }
1687
1688                         if (lnkp) {
1689                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1690                                                                    obj);
1691                                 mdt_object_put(info->mti_env, lnkp);
1692                                 continue;
1693                         }
1694                 }
1695
1696                 /* Check if it's already locked */
1697                 list_for_each_entry(msl, link_locks, msl_linkage) {
1698                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1699                                 CDEBUG(D_INFO,
1700                                        DFID" was locked, revoke "DNAME"\n",
1701                                        PFID(&fid), PNAME(lname));
1702                                 lnkp = msl->msl_obj;
1703                                 break;
1704                         }
1705                 }
1706
1707                 if (lnkp) {
1708                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1709                         continue;
1710                 }
1711
1712                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1713                        PFID(&fid), PNAME(lname));
1714
1715                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1716                 if (IS_ERR(lnkp)) {
1717                         CWARN("%s: cannot find obj "DFID": %ld\n",
1718                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1719                         continue;
1720                 }
1721
1722                 if (!mdt_object_exists(lnkp)) {
1723                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1724                               PFID(&fid), PNAME(lname));
1725                         mdt_object_put(info->mti_env, lnkp);
1726                         continue;
1727                 }
1728
1729                 if (!mdt_object_remote(lnkp))
1730                         local_lnkp_cnt++;
1731
1732                 OBD_ALLOC_PTR(msl);
1733                 if (msl == NULL)
1734                         GOTO(out, rc = -ENOMEM);
1735
1736                 /*
1737                  * we can't follow parent-child lock order like other MD
1738                  * operations, use lock_try here to avoid deadlock, if the lock
1739                  * cannot be taken, drop all locks taken, revoke the blocked
1740                  * one, and continue processing the remaining entries, and in
1741                  * the end of the loop restart from beginning.
1742                  */
1743                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1744                 ibits = 0;
1745                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1746                                          MDS_INODELOCK_UPDATE, true);
1747                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1748
1749                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1750                                PFID(&fid), PNAME(lname));
1751
1752                         mdt_unlock_list(info, link_locks, 1);
1753                         /* also unlock parent locks to avoid deadlock */
1754                         if (!blocked)
1755                                 mdt_migrate_object_unlock(info, pobj, lhp,
1756                                                           peinfo,
1757                                                           parent_slave_locks,
1758                                                           1);
1759
1760                         blocked = true;
1761
1762                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1763                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1764                                              MDS_INODELOCK_UPDATE);
1765                         if (rc) {
1766                                 mdt_object_put(info->mti_env, lnkp);
1767                                 OBD_FREE_PTR(msl);
1768                                 GOTO(out, rc);
1769                         }
1770
1771                         if (mdt_object_remote(lnkp)) {
1772                                 struct ldlm_lock *lock;
1773
1774                                 /*
1775                                  * for remote object, set lock cb_atomic,
1776                                  * so lock can be released in blocking_ast()
1777                                  * immediately, then the next lock_try will
1778                                  * have better chance of success.
1779                                  */
1780                                 lock = ldlm_handle2lock(
1781                                                 &msl->msl_lh.mlh_rreg_lh);
1782                                 LASSERT(lock != NULL);
1783                                 lock_res_and_lock(lock);
1784                                 ldlm_set_atomic_cb(lock);
1785                                 unlock_res_and_lock(lock);
1786                                 LDLM_LOCK_PUT(lock);
1787                         }
1788
1789                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1790                         OBD_FREE_PTR(msl);
1791                         continue;
1792                 }
1793
1794                 INIT_LIST_HEAD(&msl->msl_linkage);
1795                 msl->msl_obj = lnkp;
1796                 list_add_tail(&msl->msl_linkage, link_locks);
1797
1798                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1799         }
1800
1801         if (blocked)
1802                 GOTO(out, rc = -EBUSY);
1803
1804         EXIT;
1805 out:
1806         if (rc) {
1807                 mdt_unlock_list(info, link_locks, rc);
1808         } else if (local_lnkp_cnt > RS_MAX_LOCKS - 5) {
1809                 CDEBUG(D_INFO, "Too many links (%d), sync operations\n",
1810                        local_lnkp_cnt);
1811                 /*
1812                  * parent may have 3 local objects: master object and 2 stripes
1813                  * (if it's being migrated too); source may have 1 local objects
1814                  * as regular file; target has 1 local object.
1815                  * Note, source may have 2 local locks if it is directory but it
1816                  * can't have hardlinks, so it is not considered here.
1817                  */
1818                 rc = 1;
1819         }
1820         return rc;
1821 }
1822
1823 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1824                                   struct mdt_object *obj,
1825                                   const struct md_attr *ma,
1826                                   struct list_head *slave_locks)
1827 {
1828         struct mdt_device *mdt = info->mti_mdt;
1829         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1830         struct lu_fid *fid = &info->mti_tmp_fid1;
1831         struct mdt_object *slave;
1832         struct mdt_sub_lock *msl;
1833         int i;
1834         int rc;
1835
1836         ENTRY;
1837         LASSERT(mdt_object_remote(obj));
1838         LASSERT(ma->ma_valid & MA_LMV);
1839         LASSERT(lmv);
1840
1841         if (!lmv_is_sane(lmv))
1842                 RETURN(-EINVAL);
1843
1844         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1845                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1846
1847                 if (!fid_is_sane(fid))
1848                         continue;
1849
1850                 slave = mdt_object_find(info->mti_env, mdt, fid);
1851                 if (IS_ERR(slave))
1852                         GOTO(out, rc = PTR_ERR(slave));
1853
1854                 OBD_ALLOC_PTR(msl);
1855                 if (!msl) {
1856                         mdt_object_put(info->mti_env, slave);
1857                         GOTO(out, rc = -ENOMEM);
1858                 }
1859
1860                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1861                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1862                                            MDS_INODELOCK_UPDATE, true);
1863                 if (rc) {
1864                         OBD_FREE_PTR(msl);
1865                         mdt_object_put(info->mti_env, slave);
1866                         GOTO(out, rc);
1867                 }
1868
1869                 INIT_LIST_HEAD(&msl->msl_linkage);
1870                 msl->msl_obj = slave;
1871                 list_add_tail(&msl->msl_linkage, slave_locks);
1872         }
1873         EXIT;
1874
1875 out:
1876         if (rc)
1877                 mdt_unlock_list(info, slave_locks, rc);
1878         return rc;
1879 }
1880
1881 /* lock parent and its stripes */
1882 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1883                                    struct mdt_object *obj,
1884                                    const struct md_attr *ma,
1885                                    struct mdt_lock_handle *lh,
1886                                    struct ldlm_enqueue_info *einfo,
1887                                    struct list_head *slave_locks)
1888 {
1889         int rc;
1890
1891         if (mdt_object_remote(obj)) {
1892                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1893                                             &lh->mlh_rreg_lh, LCK_PW,
1894                                             MDS_INODELOCK_UPDATE, false);
1895                 if (rc != ELDLM_OK)
1896                         return rc;
1897
1898                 /*
1899                  * if obj is remote and striped, lock its stripes explicitly
1900                  * because it's not striped in LOD layer on this MDT.
1901                  */
1902                 if (ma->ma_valid & MA_LMV) {
1903                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1904                         if (rc)
1905                                 mdt_object_unlock(info, obj, lh, rc);
1906                 }
1907         } else {
1908                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1909                                             einfo, true);
1910         }
1911
1912         return rc;
1913 }
1914
1915 /*
1916  * in migration, object may be remote, and we need take full lock of it and its
1917  * stripes if it's directory, besides, object may be a remote object on its
1918  * parent, revoke its LOOKUP lock on where its parent is located.
1919  */
1920 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1921                                    struct mdt_object *pobj,
1922                                    struct mdt_object *obj,
1923                                    struct mdt_lock_handle *lh,
1924                                    struct ldlm_enqueue_info *einfo,
1925                                    struct list_head *slave_locks)
1926 {
1927         int rc;
1928
1929         if (mdt_object_remote(obj)) {
1930                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1931                 if (rc)
1932                         return rc;
1933
1934                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1935                                             &lh->mlh_rreg_lh, LCK_EX,
1936                                             MDS_INODELOCK_FULL, false);
1937                 if (rc != ELDLM_OK)
1938                         return rc;
1939
1940                 /*
1941                  * if obj is remote and striped, lock its stripes explicitly
1942                  * because it's not striped in LOD layer on this MDT.
1943                  */
1944                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1945                         struct md_attr *ma = &info->mti_attr;
1946
1947                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1948                         if (rc) {
1949                                 mdt_object_unlock(info, obj, lh, rc);
1950                                 return rc;
1951                         }
1952
1953                         if (ma->ma_valid & MA_LMV) {
1954                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1955                                                             slave_locks);
1956                                 if (rc)
1957                                         mdt_object_unlock(info, obj, lh, rc);
1958                         }
1959                 }
1960         } else {
1961                 if (mdt_object_remote(pobj)) {
1962                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1963                         if (rc)
1964                                 return rc;
1965                 }
1966
1967                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1968                                             einfo, true);
1969         }
1970
1971         return rc;
1972 }
1973
1974 /*
1975  * lookup source by name, if parent is striped directory, we need to find the
1976  * corresponding stripe where source is located, and then lookup there.
1977  *
1978  * besides, if parent is migrating too, and file is already in target stripe,
1979  * this should be a redo of 'lfs migrate' on client side.
1980  */
1981 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1982                               struct mdt_object *pobj,
1983                               const struct md_attr *ma,
1984                               const struct lu_name *lname,
1985                               struct mdt_object **spobj,
1986                               struct mdt_object **sobj)
1987 {
1988         const struct lu_env *env = info->mti_env;
1989         struct lu_fid *fid = &info->mti_tmp_fid1;
1990         struct mdt_object *stripe;
1991         int rc;
1992
1993         if (ma->ma_valid & MA_LMV) {
1994                 /* if parent is striped, lookup on corresponding stripe */
1995                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1996
1997                 if (!lmv_is_sane(lmv))
1998                         return -EBADF;
1999
2000                 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
2001                                                   lname->ln_namelen);
2002                 if (rc < 0)
2003                         return rc;
2004
2005                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2006
2007                 stripe = mdt_object_find(env, info->mti_mdt, fid);
2008                 if (IS_ERR(stripe))
2009                         return PTR_ERR(stripe);
2010
2011                 fid_zero(fid);
2012                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
2013                                 &info->mti_spec);
2014                 if (rc == -ENOENT && lmv_is_layout_changing(lmv)) {
2015                         /*
2016                          * if parent layout is changeing, and lookup child
2017                          * failed on source stripe, lookup again on target
2018                          * stripe, if it exists, it means previous migration
2019                          * was interrupted, and current file was migrated
2020                          * already.
2021                          */
2022                         mdt_object_put(env, stripe);
2023
2024                         rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
2025                                                       lname->ln_namelen);
2026                         if (rc < 0)
2027                                 return rc;
2028
2029                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2030
2031                         stripe = mdt_object_find(env, info->mti_mdt, fid);
2032                         if (IS_ERR(stripe))
2033                                 return PTR_ERR(stripe);
2034
2035                         fid_zero(fid);
2036                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
2037                                         fid, &info->mti_spec);
2038                         mdt_object_put(env, stripe);
2039                         return rc ?: -EALREADY;
2040                 } else if (rc) {
2041                         mdt_object_put(env, stripe);
2042                         return rc;
2043                 }
2044         } else {
2045                 fid_zero(fid);
2046                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2047                                 &info->mti_spec);
2048                 if (rc)
2049                         return rc;
2050
2051                 stripe = pobj;
2052                 mdt_object_get(env, stripe);
2053         }
2054
2055         *spobj = stripe;
2056
2057         *sobj = mdt_object_find(env, info->mti_mdt, fid);
2058         if (IS_ERR(*sobj)) {
2059                 mdt_object_put(env, stripe);
2060                 rc = PTR_ERR(*sobj);
2061                 *spobj = NULL;
2062                 *sobj = NULL;
2063         }
2064
2065         return rc;
2066 }
2067
2068 /* end lease and close file for regular file */
2069 static int mdd_migrate_close(struct mdt_thread_info *info,
2070                              struct mdt_object *obj)
2071 {
2072         struct close_data *data;
2073         struct mdt_body *repbody;
2074         struct ldlm_lock *lease;
2075         int rc;
2076         int rc2;
2077
2078         rc = -EPROTO;
2079         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2080                                       RCL_CLIENT) ||
2081             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2082                                       RCL_CLIENT))
2083                 goto close;
2084
2085         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2086         if (!data)
2087                 goto close;
2088
2089         rc = -ESTALE;
2090         lease = ldlm_handle2lock(&data->cd_handle);
2091         if (!lease)
2092                 goto close;
2093
2094         /* check if the lease was already canceled */
2095         lock_res_and_lock(lease);
2096         rc = ldlm_is_cancel(lease);
2097         unlock_res_and_lock(lease);
2098
2099         if (rc) {
2100                 rc = -EAGAIN;
2101                 LDLM_DEBUG(lease, DFID" lease broken",
2102                            PFID(mdt_object_fid(obj)));
2103         }
2104
2105         /*
2106          * cancel server side lease, client side counterpart should have been
2107          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2108          */
2109         ldlm_lock_cancel(lease);
2110         ldlm_reprocess_all(lease->l_resource, lease);
2111         LDLM_LOCK_PUT(lease);
2112
2113 close:
2114         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2115         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2116         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2117
2118         return rc ?: rc2;
2119 }
2120
2121 /*
2122  * migrate file in below steps:
2123  *  1. lock parent and its stripes
2124  *  2. lookup source by name
2125  *  3. lock parents of source links if source is not directory
2126  *  4. reject if source is in HSM
2127  *  5. take source open_sem and close file if source is regular file
2128  *  6. lock source and its stripes if it's directory
2129  *  7. lock target so subsequent change to it can trigger COS
2130  *  8. migrate file
2131  *  9. unlock above locks
2132  * 10. sync device if source has links
2133  */
2134 int mdt_reint_migrate(struct mdt_thread_info *info,
2135                       struct mdt_lock_handle *unused)
2136 {
2137         const struct lu_env *env = info->mti_env;
2138         struct mdt_device *mdt = info->mti_mdt;
2139         struct ptlrpc_request *req = mdt_info_req(info);
2140         struct mdt_reint_record *rr = &info->mti_rr;
2141         struct lu_ucred *uc = mdt_ucred(info);
2142         struct md_attr *ma = &info->mti_attr;
2143         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
2144         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
2145         struct mdt_object *pobj;
2146         struct mdt_object *spobj = NULL;
2147         struct mdt_object *sobj = NULL;
2148         struct mdt_object *tobj;
2149         struct lustre_handle rename_lh = { 0 };
2150         struct mdt_lock_handle *lhp;
2151         struct mdt_lock_handle *lhs;
2152         struct mdt_lock_handle *lht;
2153         LIST_HEAD(parent_slave_locks);
2154         LIST_HEAD(child_slave_locks);
2155         LIST_HEAD(link_locks);
2156         int lock_retries = 5;
2157         bool open_sem_locked = false;
2158         bool do_sync = false;
2159         int rc;
2160
2161         ENTRY;
2162         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2163                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2164
2165         if (info->mti_dlm_req)
2166                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2167
2168         if (!fid_is_md_operative(rr->rr_fid1) ||
2169             !fid_is_md_operative(rr->rr_fid2))
2170                 RETURN(-EPERM);
2171
2172         /* don't allow migrate . or .. */
2173         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2174                 RETURN(-EBUSY);
2175
2176         if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2177                 RETURN(-EPERM);
2178
2179         if (uc && !md_capable(uc, CAP_SYS_ADMIN) &&
2180             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2181             mdt->mdt_enable_remote_dir_gid != -1)
2182                 RETURN(-EPERM);
2183
2184         /*
2185          * Note: do not enqueue rename lock for replay request, because
2186          * if other MDT holds rename lock, but being blocked to wait for
2187          * this MDT to finish its recovery, and the failover MDT can not
2188          * get rename lock, which will cause deadlock.
2189          *
2190          * req is NULL if this is called by directory auto-split.
2191          */
2192         if (req && !req_is_replay(req)) {
2193                 rc = mdt_rename_lock(info, &rename_lh);
2194                 if (rc != 0) {
2195                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2196                                mdt_obd_name(info->mti_mdt), rc);
2197                         RETURN(rc);
2198                 }
2199         }
2200
2201         /* pobj is master object of parent */
2202         pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2203         if (IS_ERR(pobj))
2204                 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2205
2206         if (req) {
2207                 rc = mdt_version_get_check(info, pobj, 0);
2208                 if (rc)
2209                         GOTO(put_parent, rc);
2210         }
2211
2212         if (!mdt_object_exists(pobj))
2213                 GOTO(put_parent, rc = -ENOENT);
2214
2215         if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2216                 GOTO(put_parent, rc = -ENOTDIR);
2217
2218         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2219         if (rc)
2220                 GOTO(put_parent, rc);
2221
2222 lock_parent:
2223         /* lock parent object */
2224         lhp = &info->mti_lh[MDT_LH_PARENT];
2225         mdt_lock_reg_init(lhp, LCK_PW);
2226         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2227                                      &parent_slave_locks);
2228         if (rc)
2229                 GOTO(put_parent, rc);
2230
2231         /*
2232          * spobj is the corresponding stripe against name if pobj is striped
2233          * directory, which is the real parent, and no need to lock, because
2234          * we've taken full lock of pobj.
2235          */
2236         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2237         if (rc)
2238                 GOTO(unlock_parent, rc);
2239
2240         /* lock parents of source links, and revoke LOOKUP lock of links */
2241         rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2242                                    &parent_slave_locks, &link_locks);
2243         if (rc == -EBUSY && lock_retries-- > 0) {
2244                 mdt_object_put(env, sobj);
2245                 mdt_object_put(env, spobj);
2246                 goto lock_parent;
2247         }
2248
2249         if (rc < 0)
2250                 GOTO(put_source, rc);
2251
2252         /*
2253          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2254          * with one request, if total lock count exceeds this limit, we will
2255          * drop all locks after migration, and synchronous device in the end.
2256          */
2257         do_sync = rc;
2258
2259         /* TODO: DoM migration is not supported, migrate dirent only */
2260         if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2261                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2262                 if (rc)
2263                         GOTO(unlock_links, rc);
2264
2265                 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2266                         info->mti_spec.sp_migrate_nsonly = 1;
2267         }
2268
2269         /* if migration HSM is allowed */
2270         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2271                 ma->ma_need = MA_HSM;
2272                 ma->ma_valid = 0;
2273                 rc = mdt_attr_get_complex(info, sobj, ma);
2274                 if (rc)
2275                         GOTO(unlock_links, rc);
2276
2277                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2278                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2279         }
2280
2281         /* end lease and close file for regular file */
2282         if (info->mti_spec.sp_migrate_close) {
2283                 /* try to hold open_sem so that nobody else can open the file */
2284                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2285                         /* close anyway */
2286                         mdd_migrate_close(info, sobj);
2287                         GOTO(unlock_links, rc = -EBUSY);
2288                 } else {
2289                         open_sem_locked = true;
2290                         rc = mdd_migrate_close(info, sobj);
2291                         if (rc)
2292                                 GOTO(unlock_open_sem, rc);
2293                 }
2294         }
2295
2296         /* lock source */
2297         lhs = &info->mti_lh[MDT_LH_OLD];
2298         mdt_lock_reg_init(lhs, LCK_EX);
2299         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2300                                      &child_slave_locks);
2301         if (rc)
2302                 GOTO(unlock_open_sem, rc);
2303
2304         /* lock target */
2305         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2306         if (IS_ERR(tobj))
2307                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2308
2309         lht = &info->mti_lh[MDT_LH_NEW];
2310         mdt_lock_reg_init(lht, LCK_EX);
2311         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2312         if (rc)
2313                 GOTO(put_target, rc);
2314
2315         /* Don't do lookup sanity check. We know name doesn't exist. */
2316         info->mti_spec.sp_cr_lookup = 0;
2317         info->mti_spec.sp_feat = &dt_directory_features;
2318
2319         rc = mdo_migrate(env, mdt_object_child(pobj),
2320                          mdt_object_child(sobj), &rr->rr_name,
2321                          mdt_object_child(tobj),
2322                          &info->mti_spec, ma);
2323         if (!rc)
2324                 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2325                                      LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2326         EXIT;
2327
2328         mdt_object_unlock(info, tobj, lht, rc);
2329 put_target:
2330         mdt_object_put(env, tobj);
2331 unlock_source:
2332         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2333                                   &child_slave_locks, rc);
2334 unlock_open_sem:
2335         if (open_sem_locked)
2336                 up_write(&sobj->mot_open_sem);
2337 unlock_links:
2338         /* if we've got too many locks to save into RPC,
2339          * then just commit before the locks are released
2340          */
2341         if (!rc && do_sync)
2342                 mdt_device_sync(env, mdt);
2343         mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2344 put_source:
2345         mdt_object_put(env, sobj);
2346         mdt_object_put(env, spobj);
2347 unlock_parent:
2348         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2349                                   &parent_slave_locks, rc);
2350 put_parent:
2351         mdt_object_put(env, pobj);
2352 unlock_rename:
2353         if (lustre_handle_is_used(&rename_lh))
2354                 mdt_rename_unlock(&rename_lh);
2355
2356         return rc;
2357 }
2358
2359 static int mdt_object_lock_save(struct mdt_thread_info *info,
2360                                 struct mdt_object *dir,
2361                                 struct mdt_lock_handle *lh,
2362                                 int idx, bool cos_incompat)
2363 {
2364         int rc;
2365
2366         /* we lock the target dir if it is local */
2367         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2368                                    cos_incompat);
2369         if (rc != 0)
2370                 return rc;
2371
2372         /* get and save correct version after locking */
2373         mdt_version_get_save(info, dir, idx);
2374         return 0;
2375 }
2376
2377 /*
2378  * determine lock order of sobj and tobj
2379  *
2380  * there are two situations we need to lock tobj before sobj:
2381  * 1. sobj is child of tobj
2382  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2383  *    larger than that of tobj
2384  *
2385  * \retval      1 lock tobj before sobj
2386  * \retval      0 lock sobj before tobj
2387  * \retval      -ev negative errno upon error
2388  */
2389 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2390                                            struct mdt_object *sobj,
2391                                            struct mdt_object *tobj)
2392 {
2393         struct md_attr *ma = &info->mti_attr;
2394         struct lu_fid *spfid = &info->mti_tmp_fid1;
2395         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2396         struct lmv_mds_md_v1 *lmv;
2397         __u32 sindex;
2398         __u32 tindex;
2399         int rc;
2400
2401         /* sobj and tobj are the same */
2402         if (sobj == tobj)
2403                 return 0;
2404
2405         if (fid_is_root(mdt_object_fid(sobj)))
2406                 return 0;
2407
2408         if (fid_is_root(mdt_object_fid(tobj)))
2409                 return 1;
2410
2411         /* check whether sobj is child of tobj */
2412         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2413                            mdt_object_fid(tobj));
2414         if (rc < 0)
2415                 return rc;
2416
2417         if (rc == 1)
2418                 return 1;
2419
2420         /* check whether sobj and tobj are children of the same parent */
2421         rc = mdt_attr_get_pfid(info, sobj, spfid);
2422         if (rc)
2423                 return rc;
2424
2425         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2426         if (rc)
2427                 return rc;
2428
2429         if (!lu_fid_eq(spfid, tpfid))
2430                 return 0;
2431
2432         /* check whether sobj and tobj are sibling stripes */
2433         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2434         if (rc)
2435                 return rc;
2436
2437         if (!(ma->ma_valid & MA_LMV))
2438                 return 0;
2439
2440         lmv = &ma->ma_lmv->lmv_md_v1;
2441         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2442                 return 0;
2443         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2444
2445         ma->ma_valid = 0;
2446         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2447         if (rc)
2448                 return rc;
2449
2450         if (!(ma->ma_valid & MA_LMV))
2451                 return -ENODATA;
2452
2453         lmv = &ma->ma_lmv->lmv_md_v1;
2454         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2455                 return -EINVAL;
2456         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2457
2458         /* check stripe index of sobj and tobj */
2459         if (sindex == tindex)
2460                 return -EINVAL;
2461
2462         return sindex < tindex ? 0 : 1;
2463 }
2464
2465 /*
2466  * lock rename source object.
2467  *
2468  * Both source and source parent may be remote, and source may be a remote
2469  * object on source parent, to avoid overriding lock handle, store remote
2470  * LOOKUP lock separately in @lhr.
2471  *
2472  * \retval      0 on success
2473  * \retval      -ev negative errno upon error
2474  */
2475 static int mdt_rename_source_lock(struct mdt_thread_info *info,
2476                                   struct mdt_object *parent,
2477                                   struct mdt_object *child,
2478                                   struct mdt_lock_handle *lhc,
2479                                   struct mdt_lock_handle *lhr,
2480                                   __u64 ibits,
2481                                   bool cos_incompat)
2482 {
2483         int rc;
2484
2485         rc = mdt_is_remote_object(info, parent, child);
2486         if (rc < 0)
2487                 return rc;
2488
2489         if (rc) {
2490                 /* enqueue remote LOOKUP lock from the parent MDT */
2491                 __u64 rmt_ibits = MDS_INODELOCK_LOOKUP;
2492
2493                 if (mdt_object_remote(parent)) {
2494                         rc = mdt_remote_object_lock(info, parent,
2495                                                     mdt_object_fid(child),
2496                                                     &lhr->mlh_rreg_lh,
2497                                                     lhr->mlh_rreg_mode,
2498                                                     rmt_ibits, false);
2499                         if (rc != ELDLM_OK)
2500                                 return rc;
2501                 } else {
2502                         LASSERT(mdt_object_remote(child));
2503                         rc = mdt_object_local_lock(info, child, lhr,
2504                                                    &rmt_ibits, 0, true);
2505                         if (rc < 0)
2506                                 return rc;
2507                 }
2508
2509                 ibits &= ~MDS_INODELOCK_LOOKUP;
2510         }
2511
2512         if (mdt_object_remote(child)) {
2513                 rc = mdt_remote_object_lock(info, child, mdt_object_fid(child),
2514                                             &lhc->mlh_rreg_lh,
2515                                             lhc->mlh_rreg_mode,
2516                                             ibits, false);
2517                 if (rc == ELDLM_OK)
2518                         rc = 0;
2519         } else {
2520                 rc = mdt_reint_object_lock(info, child, lhc, ibits,
2521                                            cos_incompat);
2522         }
2523
2524         if (!rc)
2525                 mdt_object_unlock(info, child, lhr, rc);
2526
2527         return rc;
2528 }
2529
2530 /*
2531  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2532  * 2 - srcdir child; 3 - tgtdir child.
2533  * Update on disk version of srcdir child.
2534  */
2535 static int mdt_reint_rename(struct mdt_thread_info *info,
2536                             struct mdt_lock_handle *unused)
2537 {
2538         struct mdt_device *mdt = info->mti_mdt;
2539         struct mdt_reint_record *rr = &info->mti_rr;
2540         struct md_attr *ma = &info->mti_attr;
2541         struct ptlrpc_request *req = mdt_info_req(info);
2542         struct mdt_object *msrcdir = NULL;
2543         struct mdt_object *mtgtdir = NULL;
2544         struct mdt_object *mold;
2545         struct mdt_object *mnew = NULL;
2546         struct lustre_handle rename_lh = { 0 };
2547         struct mdt_lock_handle *lh_srcdirp;
2548         struct mdt_lock_handle *lh_tgtdirp;
2549         struct mdt_lock_handle *lh_oldp = NULL;
2550         struct mdt_lock_handle *lh_rmt = NULL;
2551         struct mdt_lock_handle *lh_newp = NULL;
2552         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2553         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2554         __u64 lock_ibits;
2555         bool reverse = false, discard = false;
2556         bool cos_incompat;
2557         ktime_t kstart = ktime_get();
2558         int rc;
2559
2560         ENTRY;
2561         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2562                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2563                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2564
2565         if (info->mti_dlm_req)
2566                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2567
2568         if (!fid_is_md_operative(rr->rr_fid1) ||
2569             !fid_is_md_operative(rr->rr_fid2))
2570                 RETURN(-EPERM);
2571
2572         /* find both parents. */
2573         msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2574         if (IS_ERR(msrcdir))
2575                 RETURN(PTR_ERR(msrcdir));
2576
2577         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2578
2579         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2580                 mtgtdir = msrcdir;
2581                 mdt_object_get(info->mti_env, mtgtdir);
2582         } else {
2583                 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2584                 if (IS_ERR(mtgtdir))
2585                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2586         }
2587
2588         /*
2589          * Note: do not enqueue rename lock for replay request, because
2590          * if other MDT holds rename lock, but being blocked to wait for
2591          * this MDT to finish its recovery, and the failover MDT can not
2592          * get rename lock, which will cause deadlock.
2593          */
2594         if (!req_is_replay(req)) {
2595                 /*
2596                  * Normally rename RPC is handled on the MDT with the target
2597                  * directory (if target exists, it's on the MDT with the
2598                  * target), if the source directory is remote, it's a hint that
2599                  * source is remote too (this may not be true, but it won't
2600                  * cause any issue), return -EXDEV early to avoid taking
2601                  * rename_lock.
2602                  */
2603                 if (!mdt->mdt_enable_remote_rename &&
2604                     mdt_object_remote(msrcdir))
2605                         GOTO(out_put_tgtdir, rc = -EXDEV);
2606
2607                 /* This may be further relaxed in the future for different
2608                  * source and target parents. Start with only same-directory
2609                  * renames of non-directory files for simplicity and because
2610                  * this is by far the most common use case.
2611                  */
2612                 if (msrcdir != mtgtdir || S_ISDIR(ma->ma_attr.la_mode)) {
2613                         rc = mdt_rename_lock(info, &rename_lh);
2614                         if (rc != 0) {
2615                                 CERROR("%s: cannot lock for rename: rc = %d\n",
2616                                        mdt_obd_name(mdt), rc);
2617                                 GOTO(out_put_tgtdir, rc);
2618                         }
2619                 } else {
2620                         CDEBUG(D_INFO, "%s: samedir rename "DFID"/"DNAME"\n",
2621                                mdt_obd_name(mdt), PFID(rr->rr_fid1),
2622                                PNAME(&rr->rr_name));
2623                 }
2624         }
2625
2626         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2627         if (rc < 0)
2628                 GOTO(out_unlock_rename, rc);
2629
2630         reverse = rc;
2631
2632         /* source needs to be looked up after locking source parent, otherwise
2633          * this rename may race with unlink source, and cause rename hang, see
2634          * sanityn.sh 55b, so check parents first, if later we found source is
2635          * remote, relock parents.
2636          */
2637         cos_incompat = (mdt_object_remote(msrcdir) ||
2638                         mdt_object_remote(mtgtdir));
2639
2640         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2641
2642         /* lock parents in the proper order. */
2643         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2644         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2645
2646         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
2647         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
2648 relock:
2649         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2650         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2651
2652         if (reverse) {
2653                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2654                                           cos_incompat);
2655                 if (rc)
2656                         GOTO(out_unlock_rename, rc);
2657
2658                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2659
2660                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2661                                           cos_incompat);
2662                 if (rc != 0) {
2663                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2664                         GOTO(out_unlock_rename, rc);
2665                 }
2666         } else {
2667                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2668                                           cos_incompat);
2669                 if (rc)
2670                         GOTO(out_unlock_rename, rc);
2671
2672                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2673
2674                 if (mtgtdir != msrcdir) {
2675                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2676                                                   cos_incompat);
2677                 } else if (!mdt_object_remote(mtgtdir) &&
2678                            lh_srcdirp->mlh_pdo_hash !=
2679                            lh_tgtdirp->mlh_pdo_hash) {
2680                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2681                                                 MDS_INODELOCK_UPDATE,
2682                                                 cos_incompat);
2683                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2684                 }
2685                 if (rc != 0) {
2686                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2687                         GOTO(out_unlock_rename, rc);
2688                 }
2689         }
2690
2691         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2692         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2693
2694         /* find mold object. */
2695         fid_zero(old_fid);
2696         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2697         if (rc != 0)
2698                 GOTO(out_unlock_parents, rc);
2699
2700         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2701                 GOTO(out_unlock_parents, rc = -EINVAL);
2702
2703         if (!fid_is_md_operative(old_fid))
2704                 GOTO(out_unlock_parents, rc = -EPERM);
2705
2706         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2707         if (IS_ERR(mold))
2708                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2709
2710         if (!mdt_object_exists(mold)) {
2711                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2712                                 &mold->mot_obj,
2713                                 "object does not exist");
2714                 GOTO(out_put_old, rc = -ENOENT);
2715         }
2716
2717         if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2718                 GOTO(out_put_old, rc = -EXDEV);
2719
2720         /* Check if @mtgtdir is subdir of @mold, before locking child
2721          * to avoid reverse locking.
2722          */
2723         if (mtgtdir != msrcdir) {
2724                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2725                                    old_fid);
2726                 if (rc) {
2727                         if (rc == 1)
2728                                 rc = -EINVAL;
2729                         GOTO(out_put_old, rc);
2730                 }
2731         }
2732
2733         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2734         /* save version after locking */
2735         mdt_version_get_save(info, mold, 2);
2736
2737         if (!cos_incompat && mdt_object_remote(mold)) {
2738                 cos_incompat = true;
2739                 mdt_object_put(info->mti_env, mold);
2740                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2741                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2742                 goto relock;
2743         }
2744
2745         /* find mnew object:
2746          * mnew target object may not exist now
2747          * lookup with version checking
2748          */
2749         fid_zero(new_fid);
2750         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2751                                       3);
2752         if (rc == 0) {
2753                 /* the new_fid should have been filled at this moment */
2754                 if (lu_fid_eq(old_fid, new_fid))
2755                         GOTO(out_put_old, rc);
2756
2757                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2758                     lu_fid_eq(new_fid, rr->rr_fid2))
2759                         GOTO(out_put_old, rc = -EINVAL);
2760
2761                 if (!fid_is_md_operative(new_fid))
2762                         GOTO(out_put_old, rc = -EPERM);
2763
2764                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2765                 if (IS_ERR(mnew))
2766                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2767
2768                 if (!mdt_object_exists(mnew)) {
2769                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2770                                         &mnew->mot_obj,
2771                                         "object does not exist");
2772                         GOTO(out_put_new, rc = -ENOENT);
2773                 }
2774
2775                 if (mdt_object_remote(mnew)) {
2776                         struct mdt_body  *repbody;
2777
2778                         /* Always send rename req to the target child MDT */
2779                         repbody = req_capsule_server_get(info->mti_pill,
2780                                                          &RMF_MDT_BODY);
2781                         LASSERT(repbody != NULL);
2782                         repbody->mbo_fid1 = *new_fid;
2783                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2784                         GOTO(out_put_new, rc = -EXDEV);
2785                 }
2786                 /* Before locking the target dir, check we do not replace
2787                  * a dir with a non-dir, otherwise it may deadlock with
2788                  * link op which tries to create a link in this dir
2789                  * back to this non-dir.
2790                  */
2791                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2792                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2793                         GOTO(out_put_new, rc = -EISDIR);
2794
2795                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2796                 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2797                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2798                 mdt_lock_reg_init(lh_rmt, LCK_EX);
2799                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2800                 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2801                                             lh_rmt, lock_ibits, cos_incompat);
2802                 if (rc < 0)
2803                         GOTO(out_put_new, rc);
2804
2805                 /* Check if @msrcdir is subdir of @mnew, before locking child
2806                  * to avoid reverse locking.
2807                  */
2808                 if (mtgtdir != msrcdir) {
2809                         rc = mdo_is_subdir(info->mti_env,
2810                                            mdt_object_child(msrcdir), new_fid);
2811                         if (rc) {
2812                                 if (rc == 1)
2813                                         rc = -EINVAL;
2814                                 GOTO(out_unlock_old, rc);
2815                         }
2816                 }
2817
2818                 /* We used to acquire MDS_INODELOCK_FULL here but we
2819                  * can't do this now because a running HSM restore on
2820                  * the rename onto victim will hold the layout
2821                  * lock. See LU-4002.
2822                  */
2823
2824                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2825                 mdt_lock_reg_init(lh_newp, LCK_EX);
2826                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
2827                 if (mdt_object_remote(mtgtdir)) {
2828                         rc = mdt_remote_object_lock(info, mtgtdir,
2829                                                     mdt_object_fid(mnew),
2830                                                     &lh_newp->mlh_rreg_lh,
2831                                                     lh_newp->mlh_rreg_mode,
2832                                                     MDS_INODELOCK_LOOKUP,
2833                                                     false);
2834                         if (rc != ELDLM_OK)
2835                                 GOTO(out_unlock_old, rc);
2836
2837                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2838                 }
2839                 rc = mdt_reint_object_lock(info, mnew, lh_newp, lock_ibits,
2840                                            cos_incompat);
2841                 if (rc != 0)
2842                         GOTO(out_unlock_new, rc);
2843
2844                 /* get and save version after locking */
2845                 mdt_version_get_save(info, mnew, 3);
2846         } else if (rc != -ENOENT) {
2847                 GOTO(out_put_old, rc);
2848         } else {
2849                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2850                 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2851                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2852                 mdt_lock_reg_init(lh_rmt, LCK_EX);
2853                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2854                 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2855                                             lh_rmt, lock_ibits, cos_incompat);
2856                 if (rc != 0)
2857                         GOTO(out_put_old, rc);
2858
2859                 mdt_enoent_version_save(info, 3);
2860         }
2861
2862         /* step 5: rename it */
2863         mdt_reint_init_ma(info, ma);
2864
2865         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2866                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2867
2868         if (mnew != NULL)
2869                 mutex_lock(&mnew->mot_lov_mutex);
2870
2871         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2872                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2873                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2874                         &rr->rr_tgt_name, ma);
2875
2876         if (mnew != NULL)
2877                 mutex_unlock(&mnew->mot_lov_mutex);
2878
2879         /* handle last link of tgt object */
2880         if (rc == 0) {
2881                 mdt_counter_incr(req, LPROC_MDT_RENAME,
2882                                  ktime_us_delta(ktime_get(), kstart));
2883                 if (mnew) {
2884                         mdt_handle_last_unlink(info, mnew, ma);
2885                         discard = mdt_dom_check_for_discard(info, mnew);
2886                 }
2887                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2888                                          msrcdir, mtgtdir,
2889                                          ktime_us_delta(ktime_get(), kstart));
2890         }
2891
2892         EXIT;
2893 out_unlock_new:
2894         if (mnew != NULL)
2895                 mdt_object_unlock(info, mnew, lh_newp, rc);
2896 out_unlock_old:
2897         mdt_object_unlock(info, NULL, lh_rmt, rc);
2898         mdt_object_unlock(info, mold, lh_oldp, rc);
2899 out_put_new:
2900         if (mnew && !discard)
2901                 mdt_object_put(info->mti_env, mnew);
2902 out_put_old:
2903         mdt_object_put(info->mti_env, mold);
2904 out_unlock_parents:
2905         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2906         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2907 out_unlock_rename:
2908         if (lustre_handle_is_used(&rename_lh))
2909                 mdt_rename_unlock(&rename_lh);
2910 out_put_tgtdir:
2911         mdt_object_put(info->mti_env, mtgtdir);
2912 out_put_srcdir:
2913         mdt_object_put(info->mti_env, msrcdir);
2914
2915         /* The DoM discard can be done right in the place above where it is
2916          * assigned, meanwhile it is done here after rename unlock due to
2917          * compatibility with old clients, for them the discard blocks
2918          * the main thread until completion. Check LU-11359 for details.
2919          */
2920         if (discard) {
2921                 mdt_dom_discard_data(info, mnew);
2922                 mdt_object_put(info->mti_env, mnew);
2923         }
2924         return rc;
2925 }
2926
2927 static int mdt_reint_resync(struct mdt_thread_info *info,
2928                             struct mdt_lock_handle *lhc)
2929 {
2930         struct mdt_reint_record *rr = &info->mti_rr;
2931         struct ptlrpc_request *req = mdt_info_req(info);
2932         struct md_attr *ma = &info->mti_attr;
2933         struct mdt_object *mo;
2934         struct ldlm_lock *lease;
2935         struct mdt_body *repbody;
2936         struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2937         bool lease_broken;
2938         int rc, rc2;
2939
2940         ENTRY;
2941         DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2942
2943         if (info->mti_dlm_req)
2944                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2945
2946         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2947         if (IS_ERR(mo))
2948                 GOTO(out, rc = PTR_ERR(mo));
2949
2950         if (!mdt_object_exists(mo))
2951                 GOTO(out_obj, rc = -ENOENT);
2952
2953         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2954                 GOTO(out_obj, rc = -EINVAL);
2955
2956         if (mdt_object_remote(mo))
2957                 GOTO(out_obj, rc = -EREMOTE);
2958
2959         lease = ldlm_handle2lock(rr->rr_lease_handle);
2960         if (lease == NULL)
2961                 GOTO(out_obj, rc = -ESTALE);
2962
2963         /* It's really necessary to grab open_sem and check if the lease lock
2964          * has been lost. There would exist a concurrent writer coming in and
2965          * generating some dirty data in memory cache, the writeback would fail
2966          * after the layout version is increased by MDS_REINT_RESYNC RPC.
2967          */
2968         if (!down_write_trylock(&mo->mot_open_sem))
2969                 GOTO(out_put_lease, rc = -EBUSY);
2970
2971         lock_res_and_lock(lease);
2972         lease_broken = ldlm_is_cancel(lease);
2973         unlock_res_and_lock(lease);
2974         if (lease_broken)
2975                 GOTO(out_unlock, rc = -EBUSY);
2976
2977         /* the file has yet opened by anyone else after we took the lease. */
2978         layout.mlc_opc = MD_LAYOUT_RESYNC;
2979         lhc = &info->mti_lh[MDT_LH_LOCAL];
2980         rc = mdt_layout_change(info, mo, lhc, &layout);
2981         if (rc)
2982                 GOTO(out_unlock, rc);
2983
2984         mdt_object_unlock(info, mo, lhc, 0);
2985
2986         ma->ma_need = MA_INODE;
2987         ma->ma_valid = 0;
2988         rc = mdt_attr_get_complex(info, mo, ma);
2989         if (rc != 0)
2990                 GOTO(out_unlock, rc);
2991
2992         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2993         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2994
2995         EXIT;
2996 out_unlock:
2997         up_write(&mo->mot_open_sem);
2998 out_put_lease:
2999         LDLM_LOCK_PUT(lease);
3000 out_obj:
3001         mdt_object_put(info->mti_env, mo);
3002 out:
3003         mdt_client_compatibility(info);
3004         rc2 = mdt_fix_reply(info);
3005         if (rc == 0)
3006                 rc = rc2;
3007         return rc;
3008 }
3009
3010 struct mdt_reinter {
3011         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
3012         enum lprocfs_extra_opc mr_extra_opc;
3013 };
3014
3015 static const struct mdt_reinter mdt_reinters[] = {
3016         [REINT_SETATTR] = {
3017                 .mr_handler = &mdt_reint_setattr,
3018                 .mr_extra_opc = MDS_REINT_SETATTR,
3019         },
3020         [REINT_CREATE] = {
3021                 .mr_handler = &mdt_reint_create,
3022                 .mr_extra_opc = MDS_REINT_CREATE,
3023         },
3024         [REINT_LINK] = {
3025                 .mr_handler = &mdt_reint_link,
3026                 .mr_extra_opc = MDS_REINT_LINK,
3027         },
3028         [REINT_UNLINK] = {
3029                 .mr_handler = &mdt_reint_unlink,
3030                 .mr_extra_opc = MDS_REINT_UNLINK,
3031         },
3032         [REINT_RENAME] = {
3033                 .mr_handler = &mdt_reint_rename,
3034                 .mr_extra_opc = MDS_REINT_RENAME,
3035         },
3036         [REINT_OPEN] = {
3037                 .mr_handler = &mdt_reint_open,
3038                 .mr_extra_opc = MDS_REINT_OPEN,
3039         },
3040         [REINT_SETXATTR] = {
3041                 .mr_handler = &mdt_reint_setxattr,
3042                 .mr_extra_opc = MDS_REINT_SETXATTR,
3043         },
3044         [REINT_RMENTRY] = {
3045                 .mr_handler = &mdt_reint_unlink,
3046                 .mr_extra_opc = MDS_REINT_UNLINK,
3047         },
3048         [REINT_MIGRATE] = {
3049                 .mr_handler = &mdt_reint_migrate,
3050                 .mr_extra_opc = MDS_REINT_RENAME,
3051         },
3052         [REINT_RESYNC] = {
3053                 .mr_handler = &mdt_reint_resync,
3054                 .mr_extra_opc = MDS_REINT_RESYNC,
3055         },
3056 };
3057
3058 int mdt_reint_rec(struct mdt_thread_info *info,
3059                   struct mdt_lock_handle *lhc)
3060 {
3061         const struct mdt_reinter *mr;
3062         int rc;
3063
3064         ENTRY;
3065         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
3066                 RETURN(-EPROTO);
3067
3068         mr = &mdt_reinters[info->mti_rr.rr_opcode];
3069         if (mr->mr_handler == NULL)
3070                 RETURN(-EPROTO);
3071
3072         rc = (*mr->mr_handler)(info, lhc);
3073
3074         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
3075                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
3076
3077         RETURN(rc);
3078 }