lustre/mdt/mdt_reint.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/mdt/mdt_reint.c
  33  *
  34  * Lustre Metadata Target (mdt) reintegration routines
  35  *
  36  * Author: Peter Braam <braam@clusterfs.com>
  37  * Author: Andreas Dilger <adilger@clusterfs.com>
  38  * Author: Phil Schwan <phil@clusterfs.com>
  39  * Author: Huang Hua <huanghua@clusterfs.com>
  40  * Author: Yury Umanets <umka@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_MDS
  44
  45 #include <lprocfs_status.h>
  46 #include "mdt_internal.h"
  47 #include <lustre_lmv.h>
  48
  49 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
  50                                      struct md_attr *ma)
  51 {
  52         ma->ma_need = MA_INODE;
  53         ma->ma_valid = 0;
  54 }
  55
  56 /**
  57  * Get version of object by fid.
  58  *
  59  * Return real version or ENOENT_VERSION if object doesn't exist
  60  */
  61 static void mdt_obj_version_get(struct mdt_thread_info *info,
  62                                 struct mdt_object *o, __u64 *version)
  63 {
  64         LASSERT(o);
  65
  66         if (mdt_object_exists(o) && !mdt_object_remote(o) &&
  67             !fid_is_obf(mdt_object_fid(o)))
  68                 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
  69         else
  70                 *version = ENOENT_VERSION;
  71         CDEBUG(D_INODE, "FID "DFID" version is %#llx\n",
  72                PFID(mdt_object_fid(o)), *version);
  73 }
  74
  75 /**
  76  * Check version is correct.
  77  *
  78  * Should be called only during replay.
  79  */
  80 static int mdt_version_check(struct ptlrpc_request *req,
  81                              __u64 version, int idx)
  82 {
  83         __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
  84
  85         ENTRY;
  86         if (!exp_connect_vbr(req->rq_export))
  87                 RETURN(0);
  88
  89         LASSERT(req_is_replay(req));
  90         /** VBR: version is checked always because costs nothing */
  91         LASSERT(idx < PTLRPC_NUM_VERSIONS);
  92         /** Sanity check for malformed buffers */
  93         if (pre_ver == NULL) {
  94                 CERROR("No versions in request buffer\n");
  95                 spin_lock(&req->rq_export->exp_lock);
  96                 req->rq_export->exp_vbr_failed = 1;
  97                 spin_unlock(&req->rq_export->exp_lock);
  98                 RETURN(-EOVERFLOW);
  99         } else if (pre_ver[idx] != version) {
 100                 CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n",
 101                        pre_ver[idx], version);
 102                 spin_lock(&req->rq_export->exp_lock);
 103                 req->rq_export->exp_vbr_failed = 1;
 104                 spin_unlock(&req->rq_export->exp_lock);
 105                 RETURN(-EOVERFLOW);
 106         }
 107         RETURN(0);
 108 }
 109
 110 /**
 111  * Save pre-versions in reply.
 112  */
 113 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
 114                              int idx)
 115 {
 116         __u64 *reply_ver;
 117
 118         if (!exp_connect_vbr(req->rq_export))
 119                 return;
 120
 121         LASSERT(!req_is_replay(req));
 122         LASSERT(req->rq_repmsg != NULL);
 123         reply_ver = lustre_msg_get_versions(req->rq_repmsg);
 124         if (reply_ver)
 125                 reply_ver[idx] = version;
 126 }
 127
 128 /**
 129  * Save enoent version, it is needed when it is obvious that object doesn't
 130  * exist, e.g. child during create.
 131  */
 132 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
 133 {
 134         /* save version of file name for replay, it must be ENOENT here */
 135         if (!req_is_replay(mdt_info_req(info))) {
 136                 info->mti_ver[idx] = ENOENT_VERSION;
 137                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 138         }
 139 }
 140
 141 /**
 142  * Get version from disk and save in reply buffer.
 143  *
 144  * Versions are saved in reply only during normal operations not replays.
 145  */
 146 void mdt_version_get_save(struct mdt_thread_info *info,
 147                           struct mdt_object *mto, int idx)
 148 {
 149         /* don't save versions during replay */
 150         if (!req_is_replay(mdt_info_req(info))) {
 151                 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 152                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 153         }
 154 }
 155
 156 /**
 157  * Get version from disk and check it, no save in reply.
 158  */
 159 int mdt_version_get_check(struct mdt_thread_info *info,
 160                           struct mdt_object *mto, int idx)
 161 {
 162         /* only check versions during replay */
 163         if (!req_is_replay(mdt_info_req(info)))
 164                 return 0;
 165
 166         mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 167         return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
 168 }
 169
 170 /**
 171  * Get version from disk and check if recovery or just save.
 172  */
 173 int mdt_version_get_check_save(struct mdt_thread_info *info,
 174                                struct mdt_object *mto, int idx)
 175 {
 176         int rc = 0;
 177
 178         mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 179         if (req_is_replay(mdt_info_req(info)))
 180                 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
 181                                        idx);
 182         else
 183                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 184         return rc;
 185 }
 186
 187 /**
 188  * Lookup with version checking.
 189  *
 190  * This checks version of 'name'. Many reint functions uses 'name' for child not
 191  * FID, therefore we need to get object by name and check its version.
 192  */
 193 int mdt_lookup_version_check(struct mdt_thread_info *info,
 194                              struct mdt_object *p,
 195                              const struct lu_name *lname,
 196                              struct lu_fid *fid, int idx)
 197 {
 198         int rc, vbrc;
 199
 200         rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
 201                         &info->mti_spec);
 202         /* Check version only during replay */
 203         if (!req_is_replay(mdt_info_req(info)))
 204                 return rc;
 205
 206         info->mti_ver[idx] = ENOENT_VERSION;
 207         if (rc == 0) {
 208                 struct mdt_object *child;
 209
 210                 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
 211                 if (likely(!IS_ERR(child))) {
 212                         mdt_obj_version_get(info, child, &info->mti_ver[idx]);
 213                         mdt_object_put(info->mti_env, child);
 214                 }
 215         }
 216         vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
 217         return vbrc ? vbrc : rc;
 218
 219 }
 220
 221 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
 222                              struct mdt_object *obj,
 223                              struct ldlm_enqueue_info *einfo,
 224                              int decref)
 225 {
 226         union ldlm_policy_data *policy = &mti->mti_policy;
 227         struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
 228         struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
 229         int i;
 230
 231         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 232         LASSERT(slave_locks);
 233
 234         memset(policy, 0, sizeof(*policy));
 235         policy->l_inodebits.bits = einfo->ei_inodebits;
 236         mdt_lock_handle_init(lh);
 237         mdt_lock_reg_init(lh, einfo->ei_mode);
 238         for (i = 0; i < slave_locks->ha_count; i++) {
 239                 if (test_bit(i, (void *)slave_locks->ha_map))
 240                         lh->mlh_rreg_lh = slave_locks->ha_handles[i];
 241                 else
 242                         lh->mlh_reg_lh = slave_locks->ha_handles[i];
 243                 mdt_object_unlock(mti, NULL, lh, decref);
 244                 slave_locks->ha_handles[i].cookie = 0ull;
 245         }
 246
 247         return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
 248                                 policy);
 249 }
 250
 251 static inline int mdt_object_striped(struct mdt_thread_info *mti,
 252                                      struct mdt_object *obj)
 253 {
 254         struct lu_device *bottom_dev;
 255         struct lu_object *bottom_obj;
 256         int rc;
 257
 258         if (!S_ISDIR(obj->mot_header.loh_attr))
 259                 return 0;
 260
 261         /* getxattr from bottom obj to avoid reading in shard FIDs */
 262         bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom);
 263         bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev,
 264                                           mdt_object_fid(obj), NULL);
 265         if (IS_ERR(bottom_obj))
 266                 return PTR_ERR(bottom_obj);
 267
 268         rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL,
 269                           XATTR_NAME_LMV);
 270         lu_object_put(mti->mti_env, bottom_obj);
 271
 272         return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc;
 273 }
 274
 275 /**
 276  * Lock slave stripes if necessary, the lock handles of slave stripes
 277  * will be stored in einfo->ei_cbdata.
 278  **/
 279 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
 280                            enum ldlm_mode mode, __u64 ibits,
 281                            struct ldlm_enqueue_info *einfo)
 282 {
 283         union ldlm_policy_data *policy = &mti->mti_policy;
 284
 285         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 286
 287         einfo->ei_type = LDLM_IBITS;
 288         einfo->ei_mode = mode;
 289         einfo->ei_cb_bl = mdt_remote_blocking_ast;
 290         einfo->ei_cb_local_bl = mdt_blocking_ast;
 291         einfo->ei_cb_cp = ldlm_completion_ast;
 292         einfo->ei_enq_slave = 1;
 293         einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
 294         einfo->ei_inodebits = ibits;
 295         memset(policy, 0, sizeof(*policy));
 296         policy->l_inodebits.bits = ibits;
 297
 298         return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
 299                               policy);
 300 }
 301
 302 int mdt_reint_striped_lock(struct mdt_thread_info *info,
 303                            struct mdt_object *o,
 304                            struct mdt_lock_handle *lh,
 305                            __u64 ibits,
 306                            struct ldlm_enqueue_info *einfo,
 307                            bool cos_incompat)
 308 {
 309         int rc;
 310
 311         LASSERT(!mdt_object_remote(o));
 312
 313         memset(einfo, 0, sizeof(*einfo));
 314
 315         rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
 316         if (rc)
 317                 return rc;
 318
 319         rc = mdt_object_striped(info, o);
 320         if (rc != 1) {
 321                 if (rc < 0)
 322                         mdt_object_unlock(info, o, lh, rc);
 323                 return rc;
 324         }
 325
 326         rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
 327         if (rc) {
 328                 mdt_object_unlock(info, o, lh, rc);
 329                 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
 330                         rc = 0;
 331         }
 332
 333         return rc;
 334 }
 335
 336 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
 337                               struct mdt_object *o,
 338                               struct mdt_lock_handle *lh,
 339                               struct ldlm_enqueue_info *einfo, int decref)
 340 {
 341         if (einfo->ei_cbdata)
 342                 mdt_unlock_slaves(info, o, einfo, decref);
 343         mdt_object_unlock(info, o, lh, decref);
 344 }
 345
 346 static int mdt_restripe(struct mdt_thread_info *info,
 347                         struct mdt_object *parent,
 348                         const struct lu_name *lname,
 349                         const struct lu_fid *tfid,
 350                         struct md_op_spec *spec,
 351                         struct md_attr *ma)
 352 {
 353         struct mdt_device *mdt = info->mti_mdt;
 354         struct lu_fid *fid = &info->mti_tmp_fid2;
 355         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 356         struct lmv_mds_md_v1 *lmv;
 357         struct mdt_object *child;
 358         struct mdt_lock_handle *lhp;
 359         struct mdt_lock_handle *lhc;
 360         struct mdt_body *repbody;
 361         int rc;
 362
 363         ENTRY;
 364         if (!mdt->mdt_enable_dir_restripe)
 365                 RETURN(-EPERM);
 366
 367         rc = mdt_version_get_check_save(info, parent, 0);
 368         if (rc)
 369                 RETURN(rc);
 370
 371         lhp = &info->mti_lh[MDT_LH_PARENT];
 372         mdt_lock_pdo_init(lhp, LCK_PW, lname);
 373         rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
 374                                    true);
 375         if (rc)
 376                 RETURN(rc);
 377
 378         rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV);
 379         if (rc)
 380                 GOTO(unlock_parent, rc);
 381
 382         if (ma->ma_valid & MA_LMV) {
 383                 /* don't allow restripe if parent dir layout is changing */
 384                 lmv = &ma->ma_lmv->lmv_md_v1;
 385                 if (!lmv_is_sane(lmv))
 386                         GOTO(unlock_parent, rc = -EBADF);
 387
 388                 if (lmv_is_layout_changing(lmv))
 389                         GOTO(unlock_parent, rc = -EBUSY);
 390         }
 391
 392         fid_zero(fid);
 393         rc = mdt_lookup_version_check(info, parent, lname, fid, 1);
 394         if (rc)
 395                 GOTO(unlock_parent, rc);
 396
 397         child = mdt_object_find(info->mti_env, mdt, fid);
 398         if (IS_ERR(child))
 399                 GOTO(unlock_parent, rc = PTR_ERR(child));
 400
 401         if (!mdt_object_exists(child))
 402                 GOTO(out_child, rc = -ENOENT);
 403
 404         if (mdt_object_remote(child)) {
 405                 struct mdt_body *repbody;
 406
 407                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 408                 if (!repbody)
 409                         GOTO(out_child, rc = -EPROTO);
 410
 411                 repbody->mbo_fid1 = *fid;
 412                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 413                 GOTO(out_child, rc = -EREMOTE);
 414         }
 415
 416         /* lock object */
 417         lhc = &info->mti_lh[MDT_LH_CHILD];
 418         mdt_lock_reg_init(lhc, LCK_EX);
 419
 420         /* enqueue object remote LOOKUP lock */
 421         if (mdt_object_remote(parent)) {
 422                 rc = mdt_remote_object_lock(info, parent, fid,
 423                                             &lhc->mlh_rreg_lh,
 424                                             lhc->mlh_rreg_mode,
 425                                             MDS_INODELOCK_LOOKUP, false);
 426                 if (rc != ELDLM_OK)
 427                         GOTO(out_child, rc);
 428         }
 429
 430         rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo,
 431                                     true);
 432         if (rc)
 433                 GOTO(unlock_child, rc);
 434
 435         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 436         rc = mdt_version_get_check_save(info, child, 1);
 437         if (rc)
 438                 GOTO(unlock_child, rc);
 439
 440         spin_lock(&mdt->mdt_restriper.mdr_lock);
 441         if (child->mot_restriping) {
 442                 /* race? */
 443                 spin_unlock(&mdt->mdt_restriper.mdr_lock);
 444                 GOTO(unlock_child, rc = -EBUSY);
 445         }
 446         child->mot_restriping = 1;
 447         spin_unlock(&mdt->mdt_restriper.mdr_lock);
 448
 449         *fid = *tfid;
 450         rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
 451         if (rc)
 452                 GOTO(restriping_clear, rc);
 453
 454         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 455         if (!repbody)
 456                 GOTO(restriping_clear, rc = -EPROTO);
 457
 458         mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid);
 459         EXIT;
 460
 461 restriping_clear:
 462         child->mot_restriping = 0;
 463 unlock_child:
 464         mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 465 out_child:
 466         mdt_object_put(info->mti_env, child);
 467 unlock_parent:
 468         mdt_object_unlock(info, parent, lhp, rc);
 469
 470         return rc;
 471 }
 472
 473 /*
 474  * VBR: we save three versions in reply:
 475  * 0 - parent. Check that parent version is the same during replay.
 476  * 1 - name. Version of 'name' if file exists with the same name or
 477  * ENOENT_VERSION, it is needed because file may appear due to missed replays.
 478  * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
 479  * check.
 480  */
 481 static int mdt_create(struct mdt_thread_info *info)
 482 {
 483         struct mdt_device *mdt = info->mti_mdt;
 484         struct mdt_object *parent;
 485         struct mdt_object *child;
 486         struct mdt_lock_handle *lh;
 487         struct mdt_body *repbody;
 488         struct md_attr *ma = &info->mti_attr;
 489         struct mdt_reint_record *rr = &info->mti_rr;
 490         struct md_op_spec *spec = &info->mti_spec;
 491         bool restripe = false;
 492         int rc;
 493
 494         ENTRY;
 495         DEBUG_REQ(D_INODE, mdt_info_req(info),
 496                   "Create ("DNAME"->"DFID") in "DFID,
 497                   PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
 498
 499         if (!fid_is_md_operative(rr->rr_fid1))
 500                 RETURN(-EPERM);
 501
 502         if (S_ISDIR(ma->ma_attr.la_mode) &&
 503             spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
 504                 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
 505                 struct lu_ucred *uc = mdt_ucred(info);
 506                 struct obd_export *exp = mdt_info_req(info)->rq_export;
 507
 508                 /* Only new clients can create remote dir( >= 2.4) and
 509                  * striped dir(>= 2.6), old client will return -ENOTSUPP
 510                  */
 511                 if (!mdt_is_dne_client(exp))
 512                         RETURN(-ENOTSUPP);
 513
 514                 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
 515                         if (!mdt_is_striped_client(exp))
 516                                 RETURN(-ENOTSUPP);
 517
 518                         if (!mdt->mdt_enable_striped_dir)
 519                                 RETURN(-EPERM);
 520                 } else if (!mdt->mdt_enable_remote_dir) {
 521                         RETURN(-EPERM);
 522                 }
 523
 524                 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
 525                     (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) ==
 526                     LMV_HASH_TYPE_CRUSH)
 527                         RETURN(-EPROTO);
 528
 529                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 530                     uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 531                     mdt->mdt_enable_remote_dir_gid != -1)
 532                         RETURN(-EPERM);
 533
 534                 /* restripe if later found dir exists */
 535                 if (le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT)
 536                         restripe = true;
 537         }
 538
 539         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 540
 541         parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 542         if (IS_ERR(parent))
 543                 RETURN(PTR_ERR(parent));
 544
 545         if (!mdt_object_exists(parent))
 546                 GOTO(put_parent, rc = -ENOENT);
 547
 548         /*
 549          * LU-10235: check if name exists locklessly first to avoid massive
 550          * lock recalls on existing directories.
 551          */
 552         rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
 553                                       &info->mti_tmp_fid1, 1);
 554         if (rc == 0) {
 555                 if (!restripe)
 556                         GOTO(put_parent, rc = -EEXIST);
 557
 558                 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
 559                                   ma);
 560         }
 561
 562         /* -ENOENT is expected here */
 563         if (rc != -ENOENT)
 564                 GOTO(put_parent, rc);
 565
 566         /* save version of file name for replay, it must be ENOENT here */
 567         mdt_enoent_version_save(info, 1);
 568
 569         OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
 570
 571         lh = &info->mti_lh[MDT_LH_PARENT];
 572         mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 573         rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
 574         if (rc)
 575                 GOTO(put_parent, rc);
 576
 577         if (!mdt_object_remote(parent)) {
 578                 rc = mdt_version_get_check_save(info, parent, 0);
 579                 if (rc)
 580                         GOTO(unlock_parent, rc);
 581         }
 582
 583         child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
 584         if (unlikely(IS_ERR(child)))
 585                 GOTO(unlock_parent, rc = PTR_ERR(child));
 586
 587         ma->ma_need = MA_INODE;
 588         ma->ma_valid = 0;
 589
 590         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 591                         OBD_FAIL_MDS_REINT_CREATE_WRITE);
 592
 593         /* Version of child will be updated on disk. */
 594         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 595         rc = mdt_version_get_check_save(info, child, 2);
 596         if (rc)
 597                 GOTO(put_child, rc);
 598
 599         /* Let lower layer know current lock mode. */
 600         info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
 601
 602         /*
 603          * Do not perform lookup sanity check. We know that name does
 604          * not exist.
 605          */
 606         info->mti_spec.sp_cr_lookup = 0;
 607         info->mti_spec.sp_feat = &dt_directory_features;
 608
 609         rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
 610                         mdt_object_child(child), &info->mti_spec, ma);
 611         if (rc == 0)
 612                 rc = mdt_attr_get_complex(info, child, ma);
 613
 614         if (rc < 0)
 615                 GOTO(put_child, rc);
 616
 617         /*
 618          * On DNE, we need to eliminate dependey between 'mkdir a' and
 619          * 'mkdir a/b' if b is a striped directory, to achieve this, two
 620          * things are done below:
 621          * 1. save child and slaves lock.
 622          * 2. if the child is a striped directory, relock parent so to
 623          *    compare against with COS locks to ensure parent was
 624          *    committed to disk.
 625          */
 626         if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
 627                 struct mdt_lock_handle *lhc;
 628                 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 629                 bool cos_incompat;
 630
 631                 rc = mdt_object_striped(info, child);
 632                 if (rc < 0)
 633                         GOTO(put_child, rc);
 634
 635                 cos_incompat = rc;
 636                 if (cos_incompat) {
 637                         if (!mdt_object_remote(parent)) {
 638                                 mdt_object_unlock(info, parent, lh, 1);
 639                                 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 640                                 rc = mdt_reint_object_lock(info, parent, lh,
 641                                                            MDS_INODELOCK_UPDATE,
 642                                                            true);
 643                                 if (rc)
 644                                         GOTO(put_child, rc);
 645                         }
 646                 }
 647
 648                 lhc = &info->mti_lh[MDT_LH_CHILD];
 649                 mdt_lock_handle_init(lhc);
 650                 mdt_lock_reg_init(lhc, LCK_PW);
 651                 rc = mdt_reint_striped_lock(info, child, lhc,
 652                                             MDS_INODELOCK_UPDATE, einfo,
 653                                             cos_incompat);
 654                 if (rc)
 655                         GOTO(put_child, rc);
 656
 657                 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 658         }
 659
 660         /* Return fid & attr to client. */
 661         if (ma->ma_valid & MA_INODE)
 662                 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
 663                                    mdt_object_fid(child));
 664         EXIT;
 665 put_child:
 666         mdt_object_put(info->mti_env, child);
 667 unlock_parent:
 668         mdt_object_unlock(info, parent, lh, rc);
 669 put_parent:
 670         mdt_object_put(info->mti_env, parent);
 671         return rc;
 672 }
 673
 674 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
 675                         struct md_attr *ma)
 676 {
 677         struct mdt_lock_handle  *lh;
 678         int do_vbr = ma->ma_attr.la_valid &
 679                         (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
 680         __u64 lockpart = MDS_INODELOCK_UPDATE;
 681         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 682         bool cos_incompat;
 683         int rc;
 684
 685         ENTRY;
 686         rc = mdt_object_striped(info, mo);
 687         if (rc < 0)
 688                 RETURN(rc);
 689
 690         cos_incompat = rc;
 691
 692         lh = &info->mti_lh[MDT_LH_PARENT];
 693         mdt_lock_reg_init(lh, LCK_PW);
 694
 695         /* Even though the new MDT will grant PERM lock to the old
 696          * client, but the old client will almost ignore that during
 697          * So it needs to revoke both LOOKUP and PERM lock here, so
 698          * both new and old client can cancel the dcache
 699          */
 700         if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
 701                 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
 702
 703         rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
 704                                     cos_incompat);
 705         if (rc != 0)
 706                 RETURN(rc);
 707
 708         /* all attrs are packed into mti_attr in unpack_setattr */
 709         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 710                        OBD_FAIL_MDS_REINT_SETATTR_WRITE);
 711
 712         /* VBR: update version if attr changed are important for recovery */
 713         if (do_vbr) {
 714                 /* update on-disk version of changed object */
 715                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
 716                 rc = mdt_version_get_check_save(info, mo, 0);
 717                 if (rc)
 718                         GOTO(out_unlock, rc);
 719         }
 720
 721         /* Ensure constant striping during chown(). See LU-2789. */
 722         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 723                 mutex_lock(&mo->mot_lov_mutex);
 724
 725         /* all attrs are packed into mti_attr in unpack_setattr */
 726         rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
 727
 728         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 729                 mutex_unlock(&mo->mot_lov_mutex);
 730
 731         if (rc != 0)
 732                 GOTO(out_unlock, rc);
 733         mdt_dom_obj_lvb_update(info->mti_env, mo, false);
 734         EXIT;
 735 out_unlock:
 736         mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
 737         return rc;
 738 }
 739
 740 /**
 741  * Check HSM flags and add HS_DIRTY flag if relevant.
 742  *
 743  * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
 744  * and is not RELEASED.
 745  */
 746 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
 747                         struct md_attr *ma)
 748 {
 749         struct lu_ucred *uc = mdt_ucred(info);
 750         cfs_cap_t cap_saved;
 751         int rc;
 752
 753         ENTRY;
 754         /* If the file was modified, add the dirty flag */
 755         ma->ma_need = MA_HSM;
 756         rc = mdt_attr_get_complex(info, mo, ma);
 757         if (rc) {
 758                 CERROR("file attribute read error for "DFID": %d.\n",
 759                         PFID(mdt_object_fid(mo)), rc);
 760                 RETURN(rc);
 761         }
 762
 763         /* If an up2date copy exists in the backend, add dirty flag */
 764         if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
 765             && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
 766                 ma->ma_hsm.mh_flags |= HS_DIRTY;
 767
 768                 /* Bump cap so that closes from non-owner writers can
 769                  * set the HSM state to dirty.
 770                  */
 771                 cap_saved = uc->uc_cap;
 772                 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
 773                 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
 774                 uc->uc_cap = cap_saved;
 775                 if (rc)
 776                         CERROR("file attribute change error for "DFID": %d\n",
 777                                 PFID(mdt_object_fid(mo)), rc);
 778         }
 779
 780         RETURN(rc);
 781 }
 782
 783 static int mdt_reint_setattr(struct mdt_thread_info *info,
 784                              struct mdt_lock_handle *lhc)
 785 {
 786         struct mdt_device *mdt = info->mti_mdt;
 787         struct md_attr *ma = &info->mti_attr;
 788         struct mdt_reint_record *rr = &info->mti_rr;
 789         struct ptlrpc_request *req = mdt_info_req(info);
 790         struct mdt_object *mo;
 791         struct mdt_body *repbody;
 792         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, CFS_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 relock:
1075         parent_lh = &info->mti_lh[MDT_LH_PARENT];
1076         mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
1077         rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
1078                                    cos_incompat);
1079         if (rc != 0)
1080                 GOTO(put_parent, rc);
1081
1082         /* lookup child object along with version checking */
1083         fid_zero(child_fid);
1084         rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
1085         if (rc != 0) {
1086                 /* Name might not be able to find during resend of
1087                  * remote unlink, considering following case.
1088                  * dir_A is a remote directory, the name entry of
1089                  * dir_A is on MDT0, the directory is on MDT1,
1090                  *
1091                  * 1. client sends unlink req to MDT1.
1092                  * 2. MDT1 sends name delete update to MDT0.
1093                  * 3. name entry is being deleted in MDT0 synchronously.
1094                  * 4. MDT1 is restarted.
1095                  * 5. client resends unlink req to MDT1. So it can not
1096                  *    find the name entry on MDT0 anymore.
1097                  * In this case, MDT1 only needs to destory the local
1098                  * directory.
1099                  */
1100                 if (mdt_object_remote(mp) && rc == -ENOENT &&
1101                     !fid_is_zero(rr->rr_fid2) &&
1102                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1103                         no_name = 1;
1104                         *child_fid = *rr->rr_fid2;
1105                 } else {
1106                         GOTO(unlock_parent, rc);
1107                 }
1108         }
1109
1110         if (!fid_is_md_operative(child_fid))
1111                 GOTO(unlock_parent, rc = -EPERM);
1112
1113         /* We will lock the child regardless it is local or remote. No harm. */
1114         mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1115         if (IS_ERR(mc))
1116                 GOTO(unlock_parent, rc = PTR_ERR(mc));
1117
1118         if (!cos_incompat) {
1119                 rc = mdt_object_striped(info, mc);
1120                 if (rc < 0)
1121                         GOTO(put_child, rc);
1122
1123                 cos_incompat = rc;
1124                 if (cos_incompat) {
1125                         mdt_object_put(info->mti_env, mc);
1126                         mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1127                         goto relock;
1128                 }
1129         }
1130
1131         child_lh = &info->mti_lh[MDT_LH_CHILD];
1132         mdt_lock_reg_init(child_lh, LCK_EX);
1133         if (info->mti_spec.sp_rm_entry) {
1134                 struct lu_ucred *uc  = mdt_ucred(info);
1135
1136                 if (!mdt_is_dne_client(req->rq_export))
1137                         /* Return -ENOTSUPP for old client */
1138                         GOTO(put_child, rc = -ENOTSUPP);
1139
1140                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
1141                         GOTO(put_child, rc = -EPERM);
1142
1143                 ma->ma_need = MA_INODE;
1144                 ma->ma_valid = 0;
1145                 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1146                                 NULL, &rr->rr_name, ma, no_name);
1147                 GOTO(put_child, rc);
1148         }
1149
1150         if (mdt_object_remote(mc)) {
1151                 struct mdt_body  *repbody;
1152
1153                 if (!fid_is_zero(rr->rr_fid2)) {
1154                         CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1155                                mdt_obd_name(info->mti_mdt),
1156                                PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1157                         GOTO(put_child, rc = -ENOENT);
1158                 }
1159                 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1160                        mdt_obd_name(info->mti_mdt),
1161                        PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1162
1163                 if (!mdt_is_dne_client(req->rq_export))
1164                         /* Return -ENOTSUPP for old client */
1165                         GOTO(put_child, rc = -ENOTSUPP);
1166
1167                 /* Revoke the LOOKUP lock of the remote object granted by
1168                  * this MDT. Since the unlink will happen on another MDT,
1169                  * it will release the LOOKUP lock right away. Then What
1170                  * would happen if another client try to grab the LOOKUP
1171                  * lock at the same time with unlink XXX
1172                  */
1173                 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1174                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1175                 LASSERT(repbody != NULL);
1176                 repbody->mbo_fid1 = *mdt_object_fid(mc);
1177                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1178                 GOTO(unlock_child, rc = -EREMOTE);
1179         }
1180         /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1181          * this now because a running HSM restore on the child (unlink
1182          * victim) will hold the layout lock. See LU-4002.
1183          */
1184         lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1185         if (mdt_object_remote(mp)) {
1186                 /* Enqueue lookup lock from parent MDT */
1187                 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1188                                             &child_lh->mlh_rreg_lh,
1189                                             child_lh->mlh_rreg_mode,
1190                                             MDS_INODELOCK_LOOKUP, false);
1191                 if (rc != ELDLM_OK)
1192                         GOTO(put_child, rc);
1193
1194                 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1195         }
1196
1197         rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1198                                     cos_incompat);
1199         if (rc != 0)
1200                 GOTO(put_child, rc);
1201
1202         /*
1203          * Now we can only make sure we need MA_INODE, in mdd layer, will check
1204          * whether need MA_LOV and MA_COOKIE.
1205          */
1206         ma->ma_need = MA_INODE;
1207         ma->ma_valid = 0;
1208
1209         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1210                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1211         /* save version when object is locked */
1212         mdt_version_get_save(info, mc, 1);
1213
1214         mutex_lock(&mc->mot_lov_mutex);
1215
1216         rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1217                         mdt_object_child(mc), &rr->rr_name, ma, no_name);
1218
1219         mutex_unlock(&mc->mot_lov_mutex);
1220         if (rc != 0)
1221                 GOTO(unlock_child, rc);
1222
1223         if (!lu_object_is_dying(&mc->mot_header)) {
1224                 rc = mdt_attr_get_complex(info, mc, ma);
1225                 if (rc)
1226                         GOTO(out_stat, rc);
1227         } else if (mdt_dom_check_for_discard(info, mc)) {
1228                 mdt_dom_discard_data(info, mc);
1229         }
1230         mdt_handle_last_unlink(info, mc, ma);
1231
1232 out_stat:
1233         if (ma->ma_valid & MA_INODE) {
1234                 switch (ma->ma_attr.la_mode & S_IFMT) {
1235                 case S_IFDIR:
1236                         mdt_counter_incr(req, LPROC_MDT_RMDIR,
1237                                          ktime_us_delta(ktime_get(), kstart));
1238                         break;
1239                 case S_IFREG:
1240                 case S_IFLNK:
1241                 case S_IFCHR:
1242                 case S_IFBLK:
1243                 case S_IFIFO:
1244                 case S_IFSOCK:
1245                         mdt_counter_incr(req, LPROC_MDT_UNLINK,
1246                                          ktime_us_delta(ktime_get(), kstart));
1247                         break;
1248                 default:
1249                         LASSERTF(0, "bad file type %o unlinking\n",
1250                                 ma->ma_attr.la_mode);
1251                 }
1252         }
1253
1254         EXIT;
1255
1256 unlock_child:
1257         mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1258 put_child:
1259         mdt_object_put(info->mti_env, mc);
1260 unlock_parent:
1261         mdt_object_unlock(info, mp, parent_lh, rc);
1262 put_parent:
1263         mdt_object_put(info->mti_env, mp);
1264         CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1265         return rc;
1266 }
1267
1268 /*
1269  * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1270  * name.
1271  */
1272 static int mdt_reint_link(struct mdt_thread_info *info,
1273                           struct mdt_lock_handle *lhc)
1274 {
1275         struct mdt_reint_record *rr = &info->mti_rr;
1276         struct ptlrpc_request   *req = mdt_info_req(info);
1277         struct md_attr          *ma = &info->mti_attr;
1278         struct mdt_object       *ms;
1279         struct mdt_object       *mp;
1280         struct mdt_lock_handle  *lhs;
1281         struct mdt_lock_handle  *lhp;
1282         ktime_t kstart = ktime_get();
1283         bool cos_incompat;
1284         int rc;
1285
1286         ENTRY;
1287         DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1288                   PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1289
1290         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1291                 RETURN(err_serious(-ENOENT));
1292
1293         if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1294                 req->rq_no_reply = 1;
1295                 RETURN(err_serious(-ENOENT));
1296         }
1297
1298         if (info->mti_dlm_req)
1299                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1300
1301         /* Invalid case so return error immediately instead of
1302          * processing it
1303          */
1304         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1305                 RETURN(-EPERM);
1306
1307         if (!fid_is_md_operative(rr->rr_fid1) ||
1308             !fid_is_md_operative(rr->rr_fid2))
1309                 RETURN(-EPERM);
1310
1311         /* step 1: find target parent dir */
1312         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1313         if (IS_ERR(mp))
1314                 RETURN(PTR_ERR(mp));
1315
1316         rc = mdt_version_get_check_save(info, mp, 0);
1317         if (rc)
1318                 GOTO(put_parent, rc);
1319
1320         /* step 2: find source */
1321         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1322         if (IS_ERR(ms))
1323                 GOTO(put_parent, rc = PTR_ERR(ms));
1324
1325         if (!mdt_object_exists(ms)) {
1326                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1327                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1328                 GOTO(put_source, rc = -ENOENT);
1329         }
1330
1331         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1332
1333         lhp = &info->mti_lh[MDT_LH_PARENT];
1334         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1335         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1336                                    cos_incompat);
1337         if (rc != 0)
1338                 GOTO(put_source, rc);
1339
1340         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1341
1342         lhs = &info->mti_lh[MDT_LH_CHILD];
1343         mdt_lock_reg_init(lhs, LCK_EX);
1344         rc = mdt_reint_object_lock(info, ms, lhs,
1345                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1346                                    cos_incompat);
1347         if (rc != 0)
1348                 GOTO(unlock_parent, rc);
1349
1350         /* step 3: link it */
1351         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1352                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1353
1354         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1355         rc = mdt_version_get_check_save(info, ms, 1);
1356         if (rc)
1357                 GOTO(unlock_source, rc);
1358
1359         /** check target version by name during replay */
1360         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1361                                       &info->mti_tmp_fid1, 2);
1362         if (rc != 0 && rc != -ENOENT)
1363                 GOTO(unlock_source, rc);
1364         /* save version of file name for replay, it must be ENOENT here */
1365         if (!req_is_replay(mdt_info_req(info))) {
1366                 if (rc != -ENOENT) {
1367                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1368                                PNAME(&rr->rr_name));
1369                         GOTO(unlock_source, rc = -EEXIST);
1370                 }
1371                 info->mti_ver[2] = ENOENT_VERSION;
1372                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1373         }
1374
1375         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1376                       mdt_object_child(ms), &rr->rr_name, ma);
1377
1378         if (rc == 0)
1379                 mdt_counter_incr(req, LPROC_MDT_LINK,
1380                                  ktime_us_delta(ktime_get(), kstart));
1381
1382         EXIT;
1383 unlock_source:
1384         mdt_object_unlock(info, ms, lhs, rc);
1385 unlock_parent:
1386         mdt_object_unlock(info, mp, lhp, rc);
1387 put_source:
1388         mdt_object_put(info->mti_env, ms);
1389 put_parent:
1390         mdt_object_put(info->mti_env, mp);
1391         return rc;
1392 }
1393 /**
1394  * lock the part of the directory according to the hash of the name
1395  * (lh->mlh_pdo_hash) in parallel directory lock.
1396  */
1397 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1398                               struct mdt_lock_handle *lh,
1399                               struct mdt_object *obj, __u64 ibits,
1400                               bool cos_incompat)
1401 {
1402         struct ldlm_res_id *res = &info->mti_res_id;
1403         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1404         union ldlm_policy_data *policy = &info->mti_policy;
1405         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1406         int rc;
1407
1408         /*
1409          * Finish res_id initializing by name hash marking part of
1410          * directory which is taking modification.
1411          */
1412         LASSERT(lh->mlh_pdo_hash != 0);
1413         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1414         memset(policy, 0, sizeof(*policy));
1415         policy->l_inodebits.bits = ibits;
1416         if (cos_incompat &&
1417             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1418                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1419         /*
1420          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1421          * going to be sent to client. If it is - mdt_intent_policy() path will
1422          * fix it up and turn FL_LOCAL flag off.
1423          */
1424         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1425                           policy, res, dlmflags,
1426                           &info->mti_exp->exp_handle.h_cookie);
1427         return rc;
1428 }
1429
1430 /**
1431  * Get BFL lock for rename or migrate process.
1432  **/
1433 static int mdt_rename_lock(struct mdt_thread_info *info,
1434                            struct lustre_handle *lh)
1435 {
1436         int     rc;
1437
1438         ENTRY;
1439         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1440                 struct lu_fid *fid = &info->mti_tmp_fid1;
1441                 struct mdt_object *obj;
1442
1443                 /* XXX, right now, it has to use object API to
1444                  * enqueue lock cross MDT, so it will enqueue
1445                  * rename lock(with LUSTRE_BFL_FID) by root object
1446                  */
1447                 lu_root_fid(fid);
1448                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1449                 if (IS_ERR(obj))
1450                         RETURN(PTR_ERR(obj));
1451
1452                 rc = mdt_remote_object_lock(info, obj,
1453                                             &LUSTRE_BFL_FID, lh,
1454                                             LCK_EX,
1455                                             MDS_INODELOCK_UPDATE, false);
1456                 mdt_object_put(info->mti_env, obj);
1457         } else {
1458                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1459                 union ldlm_policy_data *policy = &info->mti_policy;
1460                 struct ldlm_res_id *res_id = &info->mti_res_id;
1461                 __u64 flags = 0;
1462
1463                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1464                 memset(policy, 0, sizeof(*policy));
1465                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1466                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1467                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1468                                             LDLM_IBITS, policy, LCK_EX, &flags,
1469                                             ldlm_blocking_ast,
1470                                             ldlm_completion_ast, NULL, NULL, 0,
1471                                             LVB_T_NONE,
1472                                             &info->mti_exp->exp_handle.h_cookie,
1473                                             lh);
1474                 RETURN(rc);
1475         }
1476         RETURN(rc);
1477 }
1478
1479 static void mdt_rename_unlock(struct lustre_handle *lh)
1480 {
1481         ENTRY;
1482         LASSERT(lustre_handle_is_used(lh));
1483         /* Cancel the single rename lock right away */
1484         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1485         EXIT;
1486 }
1487
1488 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1489                                                 const struct lu_fid *fid,
1490                                                 int idx)
1491 {
1492         struct mdt_object *dir;
1493         int rc;
1494
1495         ENTRY;
1496         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1497         if (IS_ERR(dir))
1498                 RETURN(dir);
1499
1500         /* check early, the real version will be saved after locking */
1501         rc = mdt_version_get_check(info, dir, idx);
1502         if (rc)
1503                 GOTO(out_put, rc);
1504
1505         if (!mdt_object_exists(dir))
1506                 GOTO(out_put, rc = -ENOENT);
1507
1508         if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1509                 GOTO(out_put, rc = -ENOTDIR);
1510
1511         RETURN(dir);
1512 out_put:
1513         mdt_object_put(info->mti_env, dir);
1514         return ERR_PTR(rc);
1515 }
1516
1517 /*
1518  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1519  * herein we don't really check it, just do revoke.
1520  */
1521 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1522                                   struct mdt_object *pobj,
1523                                   struct mdt_object *obj)
1524 {
1525         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1526         int rc;
1527
1528         mdt_lock_handle_init(lh);
1529         mdt_lock_reg_init(lh, LCK_EX);
1530
1531         if (mdt_object_remote(pobj)) {
1532                 /* don't bother to check if pobj and obj are on the same MDT. */
1533                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1534                                             &lh->mlh_rreg_lh, LCK_EX,
1535                                             MDS_INODELOCK_LOOKUP, false);
1536         } else if (mdt_object_remote(obj)) {
1537                 struct ldlm_res_id *res = &info->mti_res_id;
1538                 union ldlm_policy_data *policy = &info->mti_policy;
1539                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1540                                  LDLM_FL_COS_INCOMPAT;
1541
1542                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1543                 memset(policy, 0, sizeof(*policy));
1544                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1545                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1546                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1547                                   dlmflags, NULL);
1548         } else {
1549                 /* do nothing if both are local */
1550                 return 0;
1551         }
1552
1553         if (rc != ELDLM_OK)
1554                 return rc;
1555
1556         /*
1557          * TODO, currently we don't save this lock because there is no place to
1558          * hold this lock handle, but to avoid race we need to save this lock.
1559          */
1560         mdt_object_unlock(info, NULL, lh, 1);
1561
1562         return 0;
1563 }
1564
1565 /*
1566  * operation may takes locks of linkea, or directory stripes, group them in
1567  * different list.
1568  */
1569 struct mdt_sub_lock {
1570         struct mdt_object *msl_obj;
1571         struct mdt_lock_handle msl_lh;
1572         struct list_head msl_linkage;
1573 };
1574
1575 static void mdt_unlock_list(struct mdt_thread_info *info,
1576                             struct list_head *list, int decref)
1577 {
1578         struct mdt_sub_lock *msl;
1579         struct mdt_sub_lock *tmp;
1580
1581         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1582                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1583                 list_del(&msl->msl_linkage);
1584                 OBD_FREE_PTR(msl);
1585         }
1586 }
1587
1588 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1589                                              struct mdt_object *obj,
1590                                              struct mdt_lock_handle *lh,
1591                                              struct ldlm_enqueue_info *einfo,
1592                                              struct list_head *slave_locks,
1593                                              int decref)
1594 {
1595         if (mdt_object_remote(obj)) {
1596                 mdt_unlock_list(info, slave_locks, decref);
1597                 mdt_object_unlock(info, obj, lh, decref);
1598         } else {
1599                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1600         }
1601 }
1602
1603 /*
1604  * lock parents of links, and also check whether total locks don't exceed
1605  * RS_MAX_LOCKS.
1606  *
1607  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1608  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1609  * \retval      -ev negative errno upon error
1610  */
1611 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1612                                  struct mdt_object *pobj,
1613                                  const struct md_attr *ma,
1614                                  struct mdt_object *obj,
1615                                  struct mdt_lock_handle *lhp,
1616                                  struct ldlm_enqueue_info *peinfo,
1617                                  struct list_head *parent_slave_locks,
1618                                  struct list_head *link_locks)
1619 {
1620         struct mdt_device *mdt = info->mti_mdt;
1621         struct lu_buf *buf = &info->mti_big_buf;
1622         struct lu_name *lname = &info->mti_name;
1623         struct linkea_data ldata = { NULL };
1624         bool blocked = false;
1625         int local_lnkp_cnt = 0;
1626         int rc;
1627
1628         ENTRY;
1629         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1630                 RETURN(0);
1631
1632         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1633         if (buf->lb_buf == NULL)
1634                 RETURN(-ENOMEM);
1635
1636         ldata.ld_buf = buf;
1637         rc = mdt_links_read(info, obj, &ldata);
1638         if (rc) {
1639                 if (rc == -ENOENT || rc == -ENODATA)
1640                         rc = 0;
1641                 RETURN(rc);
1642         }
1643
1644         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1645              linkea_next_entry(&ldata)) {
1646                 struct mdt_object *lnkp;
1647                 struct mdt_sub_lock *msl;
1648                 struct lu_fid fid;
1649                 __u64 ibits;
1650
1651                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1652                                     &fid);
1653
1654                 /* check if it's also linked to parent */
1655                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1656                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1657                                PFID(&fid), PNAME(lname));
1658                         /* in case link is remote object, revoke LOOKUP lock */
1659                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1660                         continue;
1661                 }
1662
1663                 lnkp = NULL;
1664
1665                 /* check if it's linked to a stripe of parent */
1666                 if (ma->ma_valid & MA_LMV) {
1667                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1668                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1669                         int j = 0;
1670
1671                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1672                                 fid_le_to_cpu(stripe_fid,
1673                                               &lmv->lmv_stripe_fids[j]);
1674                                 if (lu_fid_eq(stripe_fid, &fid)) {
1675                                         CDEBUG(D_INFO, "skip stripe "DFID
1676                                                ", reovke "DNAME"\n",
1677                                                PFID(&fid), PNAME(lname));
1678                                         lnkp = mdt_object_find(info->mti_env,
1679                                                                mdt, &fid);
1680                                         if (IS_ERR(lnkp))
1681                                                 GOTO(out, rc = PTR_ERR(lnkp));
1682                                         break;
1683                                 }
1684                         }
1685
1686                         if (lnkp) {
1687                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1688                                                                    obj);
1689                                 mdt_object_put(info->mti_env, lnkp);
1690                                 continue;
1691                         }
1692                 }
1693
1694                 /* Check if it's already locked */
1695                 list_for_each_entry(msl, link_locks, msl_linkage) {
1696                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1697                                 CDEBUG(D_INFO,
1698                                        DFID" was locked, revoke "DNAME"\n",
1699                                        PFID(&fid), PNAME(lname));
1700                                 lnkp = msl->msl_obj;
1701                                 break;
1702                         }
1703                 }
1704
1705                 if (lnkp) {
1706                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1707                         continue;
1708                 }
1709
1710                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1711                        PFID(&fid), PNAME(lname));
1712
1713                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1714                 if (IS_ERR(lnkp)) {
1715                         CWARN("%s: cannot find obj "DFID": %ld\n",
1716                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1717                         continue;
1718                 }
1719
1720                 if (!mdt_object_exists(lnkp)) {
1721                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1722                               PFID(&fid), PNAME(lname));
1723                         mdt_object_put(info->mti_env, lnkp);
1724                         continue;
1725                 }
1726
1727                 if (!mdt_object_remote(lnkp))
1728                         local_lnkp_cnt++;
1729
1730                 OBD_ALLOC_PTR(msl);
1731                 if (msl == NULL)
1732                         GOTO(out, rc = -ENOMEM);
1733
1734                 /*
1735                  * we can't follow parent-child lock order like other MD
1736                  * operations, use lock_try here to avoid deadlock, if the lock
1737                  * cannot be taken, drop all locks taken, revoke the blocked
1738                  * one, and continue processing the remaining entries, and in
1739                  * the end of the loop restart from beginning.
1740                  */
1741                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1742                 ibits = 0;
1743                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1744                                          MDS_INODELOCK_UPDATE, true);
1745                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1746
1747                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1748                                PFID(&fid), PNAME(lname));
1749
1750                         mdt_unlock_list(info, link_locks, 1);
1751                         /* also unlock parent locks to avoid deadlock */
1752                         if (!blocked)
1753                                 mdt_migrate_object_unlock(info, pobj, lhp,
1754                                                           peinfo,
1755                                                           parent_slave_locks,
1756                                                           1);
1757
1758                         blocked = true;
1759
1760                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1761                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1762                                              MDS_INODELOCK_UPDATE);
1763                         if (rc) {
1764                                 mdt_object_put(info->mti_env, lnkp);
1765                                 OBD_FREE_PTR(msl);
1766                                 GOTO(out, rc);
1767                         }
1768
1769                         if (mdt_object_remote(lnkp)) {
1770                                 struct ldlm_lock *lock;
1771
1772                                 /*
1773                                  * for remote object, set lock cb_atomic,
1774                                  * so lock can be released in blocking_ast()
1775                                  * immediately, then the next lock_try will
1776                                  * have better chance of success.
1777                                  */
1778                                 lock = ldlm_handle2lock(
1779                                                 &msl->msl_lh.mlh_rreg_lh);
1780                                 LASSERT(lock != NULL);
1781                                 lock_res_and_lock(lock);
1782                                 ldlm_set_atomic_cb(lock);
1783                                 unlock_res_and_lock(lock);
1784                                 LDLM_LOCK_PUT(lock);
1785                         }
1786
1787                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1788                         OBD_FREE_PTR(msl);
1789                         continue;
1790                 }
1791
1792                 INIT_LIST_HEAD(&msl->msl_linkage);
1793                 msl->msl_obj = lnkp;
1794                 list_add_tail(&msl->msl_linkage, link_locks);
1795
1796                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1797         }
1798
1799         if (blocked)
1800                 GOTO(out, rc = -EBUSY);
1801
1802         EXIT;
1803 out:
1804         if (rc) {
1805                 mdt_unlock_list(info, link_locks, rc);
1806         } else if (local_lnkp_cnt > RS_MAX_LOCKS - 5) {
1807                 CDEBUG(D_INFO, "Too many links (%d), sync operations\n",
1808                        local_lnkp_cnt);
1809                 /*
1810                  * parent may have 3 local objects: master object and 2 stripes
1811                  * (if it's being migrated too); source may have 1 local objects
1812                  * as regular file; target has 1 local object.
1813                  * Note, source may have 2 local locks if it is directory but it
1814                  * can't have hardlinks, so it is not considered here.
1815                  */
1816                 rc = 1;
1817         }
1818         return rc;
1819 }
1820
1821 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1822                                   struct mdt_object *obj,
1823                                   const struct md_attr *ma,
1824                                   struct list_head *slave_locks)
1825 {
1826         struct mdt_device *mdt = info->mti_mdt;
1827         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1828         struct lu_fid *fid = &info->mti_tmp_fid1;
1829         struct mdt_object *slave;
1830         struct mdt_sub_lock *msl;
1831         int i;
1832         int rc;
1833
1834         ENTRY;
1835         LASSERT(mdt_object_remote(obj));
1836         LASSERT(ma->ma_valid & MA_LMV);
1837         LASSERT(lmv);
1838
1839         if (!lmv_is_sane(lmv))
1840                 RETURN(-EINVAL);
1841
1842         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1843                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1844
1845                 if (!fid_is_sane(fid))
1846                         continue;
1847
1848                 slave = mdt_object_find(info->mti_env, mdt, fid);
1849                 if (IS_ERR(slave))
1850                         GOTO(out, rc = PTR_ERR(slave));
1851
1852                 OBD_ALLOC_PTR(msl);
1853                 if (!msl) {
1854                         mdt_object_put(info->mti_env, slave);
1855                         GOTO(out, rc = -ENOMEM);
1856                 }
1857
1858                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1859                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1860                                            MDS_INODELOCK_UPDATE, true);
1861                 if (rc) {
1862                         OBD_FREE_PTR(msl);
1863                         mdt_object_put(info->mti_env, slave);
1864                         GOTO(out, rc);
1865                 }
1866
1867                 INIT_LIST_HEAD(&msl->msl_linkage);
1868                 msl->msl_obj = slave;
1869                 list_add_tail(&msl->msl_linkage, slave_locks);
1870         }
1871         EXIT;
1872
1873 out:
1874         if (rc)
1875                 mdt_unlock_list(info, slave_locks, rc);
1876         return rc;
1877 }
1878
1879 /* lock parent and its stripes */
1880 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1881                                    struct mdt_object *obj,
1882                                    const struct md_attr *ma,
1883                                    struct mdt_lock_handle *lh,
1884                                    struct ldlm_enqueue_info *einfo,
1885                                    struct list_head *slave_locks)
1886 {
1887         int rc;
1888
1889         if (mdt_object_remote(obj)) {
1890                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1891                                             &lh->mlh_rreg_lh, LCK_PW,
1892                                             MDS_INODELOCK_UPDATE, false);
1893                 if (rc != ELDLM_OK)
1894                         return rc;
1895
1896                 /*
1897                  * if obj is remote and striped, lock its stripes explicitly
1898                  * because it's not striped in LOD layer on this MDT.
1899                  */
1900                 if (ma->ma_valid & MA_LMV) {
1901                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1902                         if (rc)
1903                                 mdt_object_unlock(info, obj, lh, rc);
1904                 }
1905         } else {
1906                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1907                                             einfo, true);
1908         }
1909
1910         return rc;
1911 }
1912
1913 /*
1914  * in migration, object may be remote, and we need take full lock of it and its
1915  * stripes if it's directory, besides, object may be a remote object on its
1916  * parent, revoke its LOOKUP lock on where its parent is located.
1917  */
1918 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1919                                    struct mdt_object *pobj,
1920                                    struct mdt_object *obj,
1921                                    struct mdt_lock_handle *lh,
1922                                    struct ldlm_enqueue_info *einfo,
1923                                    struct list_head *slave_locks)
1924 {
1925         int rc;
1926
1927         if (mdt_object_remote(obj)) {
1928                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1929                 if (rc)
1930                         return rc;
1931
1932                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1933                                             &lh->mlh_rreg_lh, LCK_EX,
1934                                             MDS_INODELOCK_FULL, false);
1935                 if (rc != ELDLM_OK)
1936                         return rc;
1937
1938                 /*
1939                  * if obj is remote and striped, lock its stripes explicitly
1940                  * because it's not striped in LOD layer on this MDT.
1941                  */
1942                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1943                         struct md_attr *ma = &info->mti_attr;
1944
1945                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1946                         if (rc) {
1947                                 mdt_object_unlock(info, obj, lh, rc);
1948                                 return rc;
1949                         }
1950
1951                         if (ma->ma_valid & MA_LMV) {
1952                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1953                                                             slave_locks);
1954                                 if (rc)
1955                                         mdt_object_unlock(info, obj, lh, rc);
1956                         }
1957                 }
1958         } else {
1959                 if (mdt_object_remote(pobj)) {
1960                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1961                         if (rc)
1962                                 return rc;
1963                 }
1964
1965                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1966                                             einfo, true);
1967         }
1968
1969         return rc;
1970 }
1971
1972 /*
1973  * lookup source by name, if parent is striped directory, we need to find the
1974  * corresponding stripe where source is located, and then lookup there.
1975  *
1976  * besides, if parent is migrating too, and file is already in target stripe,
1977  * this should be a redo of 'lfs migrate' on client side.
1978  */
1979 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1980                               struct mdt_object *pobj,
1981                               const struct md_attr *ma,
1982                               const struct lu_name *lname,
1983                               struct mdt_object **spobj,
1984                               struct mdt_object **sobj)
1985 {
1986         const struct lu_env *env = info->mti_env;
1987         struct lu_fid *fid = &info->mti_tmp_fid1;
1988         struct mdt_object *stripe;
1989         int rc;
1990
1991         if (ma->ma_valid & MA_LMV) {
1992                 /* if parent is striped, lookup on corresponding stripe */
1993                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1994
1995                 if (!lmv_is_sane(lmv))
1996                         return -EBADF;
1997
1998                 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
1999                                                   lname->ln_namelen);
2000                 if (rc < 0)
2001                         return rc;
2002
2003                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2004
2005                 stripe = mdt_object_find(env, info->mti_mdt, fid);
2006                 if (IS_ERR(stripe))
2007                         return PTR_ERR(stripe);
2008
2009                 fid_zero(fid);
2010                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
2011                                 &info->mti_spec);
2012                 if (rc == -ENOENT && lmv_is_layout_changing(lmv)) {
2013                         /*
2014                          * if parent layout is changeing, and lookup child
2015                          * failed on source stripe, lookup again on target
2016                          * stripe, if it exists, it means previous migration
2017                          * was interrupted, and current file was migrated
2018                          * already.
2019                          */
2020                         mdt_object_put(env, stripe);
2021
2022                         rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
2023                                                       lname->ln_namelen);
2024                         if (rc < 0)
2025                                 return rc;
2026
2027                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2028
2029                         stripe = mdt_object_find(env, info->mti_mdt, fid);
2030                         if (IS_ERR(stripe))
2031                                 return PTR_ERR(stripe);
2032
2033                         fid_zero(fid);
2034                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
2035                                         fid, &info->mti_spec);
2036                         mdt_object_put(env, stripe);
2037                         return rc ?: -EALREADY;
2038                 } else if (rc) {
2039                         mdt_object_put(env, stripe);
2040                         return rc;
2041                 }
2042         } else {
2043                 fid_zero(fid);
2044                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2045                                 &info->mti_spec);
2046                 if (rc)
2047                         return rc;
2048
2049                 stripe = pobj;
2050                 mdt_object_get(env, stripe);
2051         }
2052
2053         *spobj = stripe;
2054
2055         *sobj = mdt_object_find(env, info->mti_mdt, fid);
2056         if (IS_ERR(*sobj)) {
2057                 mdt_object_put(env, stripe);
2058                 rc = PTR_ERR(*sobj);
2059                 *spobj = NULL;
2060                 *sobj = NULL;
2061         }
2062
2063         return rc;
2064 }
2065
2066 /* end lease and close file for regular file */
2067 static int mdd_migrate_close(struct mdt_thread_info *info,
2068                              struct mdt_object *obj)
2069 {
2070         struct close_data *data;
2071         struct mdt_body *repbody;
2072         struct ldlm_lock *lease;
2073         int rc;
2074         int rc2;
2075
2076         rc = -EPROTO;
2077         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2078                                       RCL_CLIENT) ||
2079             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2080                                       RCL_CLIENT))
2081                 goto close;
2082
2083         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2084         if (!data)
2085                 goto close;
2086
2087         rc = -ESTALE;
2088         lease = ldlm_handle2lock(&data->cd_handle);
2089         if (!lease)
2090                 goto close;
2091
2092         /* check if the lease was already canceled */
2093         lock_res_and_lock(lease);
2094         rc = ldlm_is_cancel(lease);
2095         unlock_res_and_lock(lease);
2096
2097         if (rc) {
2098                 rc = -EAGAIN;
2099                 LDLM_DEBUG(lease, DFID" lease broken",
2100                            PFID(mdt_object_fid(obj)));
2101         }
2102
2103         /*
2104          * cancel server side lease, client side counterpart should have been
2105          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2106          */
2107         ldlm_lock_cancel(lease);
2108         ldlm_reprocess_all(lease->l_resource, lease);
2109         LDLM_LOCK_PUT(lease);
2110
2111 close:
2112         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2113         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2114         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2115
2116         return rc ?: rc2;
2117 }
2118
2119 /*
2120  * migrate file in below steps:
2121  *  1. lock parent and its stripes
2122  *  2. lookup source by name
2123  *  3. lock parents of source links if source is not directory
2124  *  4. reject if source is in HSM
2125  *  5. take source open_sem and close file if source is regular file
2126  *  6. lock source and its stripes if it's directory
2127  *  7. lock target so subsequent change to it can trigger COS
2128  *  8. migrate file
2129  *  9. unlock above locks
2130  * 10. sync device if source has links
2131  */
2132 int mdt_reint_migrate(struct mdt_thread_info *info,
2133                       struct mdt_lock_handle *unused)
2134 {
2135         const struct lu_env *env = info->mti_env;
2136         struct mdt_device *mdt = info->mti_mdt;
2137         struct ptlrpc_request *req = mdt_info_req(info);
2138         struct mdt_reint_record *rr = &info->mti_rr;
2139         struct lu_ucred *uc = mdt_ucred(info);
2140         struct md_attr *ma = &info->mti_attr;
2141         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
2142         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
2143         struct mdt_object *pobj;
2144         struct mdt_object *spobj = NULL;
2145         struct mdt_object *sobj = NULL;
2146         struct mdt_object *tobj;
2147         struct lustre_handle rename_lh = { 0 };
2148         struct mdt_lock_handle *lhp;
2149         struct mdt_lock_handle *lhs;
2150         struct mdt_lock_handle *lht;
2151         LIST_HEAD(parent_slave_locks);
2152         LIST_HEAD(child_slave_locks);
2153         LIST_HEAD(link_locks);
2154         int lock_retries = 5;
2155         bool open_sem_locked = false;
2156         bool do_sync = false;
2157         int rc;
2158
2159         ENTRY;
2160         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2161                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2162
2163         if (info->mti_dlm_req)
2164                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2165
2166         if (!fid_is_md_operative(rr->rr_fid1) ||
2167             !fid_is_md_operative(rr->rr_fid2))
2168                 RETURN(-EPERM);
2169
2170         /* don't allow migrate . or .. */
2171         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2172                 RETURN(-EBUSY);
2173
2174         if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2175                 RETURN(-EPERM);
2176
2177         if (uc && !md_capable(uc, CFS_CAP_SYS_ADMIN) &&
2178             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2179             mdt->mdt_enable_remote_dir_gid != -1)
2180                 RETURN(-EPERM);
2181
2182         /*
2183          * Note: do not enqueue rename lock for replay request, because
2184          * if other MDT holds rename lock, but being blocked to wait for
2185          * this MDT to finish its recovery, and the failover MDT can not
2186          * get rename lock, which will cause deadlock.
2187          *
2188          * req is NULL if this is called by directory auto-split.
2189          */
2190         if (req && !req_is_replay(req)) {
2191                 rc = mdt_rename_lock(info, &rename_lh);
2192                 if (rc != 0) {
2193                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2194                                mdt_obd_name(info->mti_mdt), rc);
2195                         RETURN(rc);
2196                 }
2197         }
2198
2199         /* pobj is master object of parent */
2200         pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2201         if (IS_ERR(pobj))
2202                 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2203
2204         if (req) {
2205                 rc = mdt_version_get_check(info, pobj, 0);
2206                 if (rc)
2207                         GOTO(put_parent, rc);
2208         }
2209
2210         if (!mdt_object_exists(pobj))
2211                 GOTO(put_parent, rc = -ENOENT);
2212
2213         if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2214                 GOTO(put_parent, rc = -ENOTDIR);
2215
2216         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2217         if (rc)
2218                 GOTO(put_parent, rc);
2219
2220 lock_parent:
2221         /* lock parent object */
2222         lhp = &info->mti_lh[MDT_LH_PARENT];
2223         mdt_lock_reg_init(lhp, LCK_PW);
2224         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2225                                      &parent_slave_locks);
2226         if (rc)
2227                 GOTO(put_parent, rc);
2228
2229         /*
2230          * spobj is the corresponding stripe against name if pobj is striped
2231          * directory, which is the real parent, and no need to lock, because
2232          * we've taken full lock of pobj.
2233          */
2234         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2235         if (rc)
2236                 GOTO(unlock_parent, rc);
2237
2238         /* lock parents of source links, and revoke LOOKUP lock of links */
2239         rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2240                                    &parent_slave_locks, &link_locks);
2241         if (rc == -EBUSY && lock_retries-- > 0) {
2242                 mdt_object_put(env, sobj);
2243                 mdt_object_put(env, spobj);
2244                 goto lock_parent;
2245         }
2246
2247         if (rc < 0)
2248                 GOTO(put_source, rc);
2249
2250         /*
2251          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2252          * with one request, if total lock count exceeds this limit, we will
2253          * drop all locks after migration, and synchronous device in the end.
2254          */
2255         do_sync = rc;
2256
2257         /* TODO: DoM migration is not supported, migrate dirent only */
2258         if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2259                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2260                 if (rc)
2261                         GOTO(unlock_links, rc);
2262
2263                 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2264                         info->mti_spec.sp_migrate_nsonly = 1;
2265         }
2266
2267         /* if migration HSM is allowed */
2268         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2269                 ma->ma_need = MA_HSM;
2270                 ma->ma_valid = 0;
2271                 rc = mdt_attr_get_complex(info, sobj, ma);
2272                 if (rc)
2273                         GOTO(unlock_links, rc);
2274
2275                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2276                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2277         }
2278
2279         /* end lease and close file for regular file */
2280         if (info->mti_spec.sp_migrate_close) {
2281                 /* try to hold open_sem so that nobody else can open the file */
2282                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2283                         /* close anyway */
2284                         mdd_migrate_close(info, sobj);
2285                         GOTO(unlock_links, rc = -EBUSY);
2286                 } else {
2287                         open_sem_locked = true;
2288                         rc = mdd_migrate_close(info, sobj);
2289                         if (rc)
2290                                 GOTO(unlock_open_sem, rc);
2291                 }
2292         }
2293
2294         /* lock source */
2295         lhs = &info->mti_lh[MDT_LH_OLD];
2296         mdt_lock_reg_init(lhs, LCK_EX);
2297         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2298                                      &child_slave_locks);
2299         if (rc)
2300                 GOTO(unlock_open_sem, rc);
2301
2302         /* lock target */
2303         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2304         if (IS_ERR(tobj))
2305                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2306
2307         lht = &info->mti_lh[MDT_LH_NEW];
2308         mdt_lock_reg_init(lht, LCK_EX);
2309         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2310         if (rc)
2311                 GOTO(put_target, rc);
2312
2313         /* Don't do lookup sanity check. We know name doesn't exist. */
2314         info->mti_spec.sp_cr_lookup = 0;
2315         info->mti_spec.sp_feat = &dt_directory_features;
2316
2317         rc = mdo_migrate(env, mdt_object_child(pobj),
2318                          mdt_object_child(sobj), &rr->rr_name,
2319                          mdt_object_child(tobj),
2320                          &info->mti_spec, ma);
2321         if (!rc)
2322                 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2323                                      LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2324         EXIT;
2325
2326         mdt_object_unlock(info, tobj, lht, rc);
2327 put_target:
2328         mdt_object_put(env, tobj);
2329 unlock_source:
2330         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2331                                   &child_slave_locks, rc);
2332 unlock_open_sem:
2333         if (open_sem_locked)
2334                 up_write(&sobj->mot_open_sem);
2335 unlock_links:
2336         /* if we've got too many locks to save into RPC,
2337          * then just commit before the locks are released
2338          */
2339         if (!rc && do_sync)
2340                 mdt_device_sync(env, mdt);
2341         mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2342 put_source:
2343         mdt_object_put(env, sobj);
2344         mdt_object_put(env, spobj);
2345 unlock_parent:
2346         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2347                                   &parent_slave_locks, rc);
2348 put_parent:
2349         mdt_object_put(env, pobj);
2350 unlock_rename:
2351         if (lustre_handle_is_used(&rename_lh))
2352                 mdt_rename_unlock(&rename_lh);
2353
2354         return rc;
2355 }
2356
2357 static int mdt_object_lock_save(struct mdt_thread_info *info,
2358                                 struct mdt_object *dir,
2359                                 struct mdt_lock_handle *lh,
2360                                 int idx, bool cos_incompat)
2361 {
2362         int rc;
2363
2364         /* we lock the target dir if it is local */
2365         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2366                                    cos_incompat);
2367         if (rc != 0)
2368                 return rc;
2369
2370         /* get and save correct version after locking */
2371         mdt_version_get_save(info, dir, idx);
2372         return 0;
2373 }
2374
2375 /*
2376  * determine lock order of sobj and tobj
2377  *
2378  * there are two situations we need to lock tobj before sobj:
2379  * 1. sobj is child of tobj
2380  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2381  *    larger than that of tobj
2382  *
2383  * \retval      1 lock tobj before sobj
2384  * \retval      0 lock sobj before tobj
2385  * \retval      -ev negative errno upon error
2386  */
2387 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2388                                            struct mdt_object *sobj,
2389                                            struct mdt_object *tobj)
2390 {
2391         struct md_attr *ma = &info->mti_attr;
2392         struct lu_fid *spfid = &info->mti_tmp_fid1;
2393         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2394         struct lmv_mds_md_v1 *lmv;
2395         __u32 sindex;
2396         __u32 tindex;
2397         int rc;
2398
2399         /* sobj and tobj are the same */
2400         if (sobj == tobj)
2401                 return 0;
2402
2403         if (fid_is_root(mdt_object_fid(sobj)))
2404                 return 0;
2405
2406         if (fid_is_root(mdt_object_fid(tobj)))
2407                 return 1;
2408
2409         /* check whether sobj is child of tobj */
2410         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2411                            mdt_object_fid(tobj));
2412         if (rc < 0)
2413                 return rc;
2414
2415         if (rc == 1)
2416                 return 1;
2417
2418         /* check whether sobj and tobj are children of the same parent */
2419         rc = mdt_attr_get_pfid(info, sobj, spfid);
2420         if (rc)
2421                 return rc;
2422
2423         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2424         if (rc)
2425                 return rc;
2426
2427         if (!lu_fid_eq(spfid, tpfid))
2428                 return 0;
2429
2430         /* check whether sobj and tobj are sibling stripes */
2431         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2432         if (rc)
2433                 return rc;
2434
2435         if (!(ma->ma_valid & MA_LMV))
2436                 return 0;
2437
2438         lmv = &ma->ma_lmv->lmv_md_v1;
2439         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2440                 return 0;
2441         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2442
2443         ma->ma_valid = 0;
2444         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2445         if (rc)
2446                 return rc;
2447
2448         if (!(ma->ma_valid & MA_LMV))
2449                 return -ENODATA;
2450
2451         lmv = &ma->ma_lmv->lmv_md_v1;
2452         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2453                 return -EINVAL;
2454         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2455
2456         /* check stripe index of sobj and tobj */
2457         if (sindex == tindex)
2458                 return -EINVAL;
2459
2460         return sindex < tindex ? 0 : 1;
2461 }
2462
2463 /*
2464  * lock rename source object.
2465  *
2466  * Both source and source parent may be remote, and source may be a remote
2467  * object on source parent, to avoid overriding lock handle, store remote
2468  * LOOKUP lock separately in @lhr.
2469  *
2470  * \retval      0 on success
2471  * \retval      -ev negative errno upon error
2472  */
2473 static int mdt_rename_source_lock(struct mdt_thread_info *info,
2474                                   struct mdt_object *parent,
2475                                   struct mdt_object *child,
2476                                   struct mdt_lock_handle *lhc,
2477                                   struct mdt_lock_handle *lhr,
2478                                   __u64 ibits,
2479                                   bool cos_incompat)
2480 {
2481         int rc;
2482
2483         rc = mdt_is_remote_object(info, parent, child);
2484         if (rc < 0)
2485                 return rc;
2486
2487         if (rc) {
2488                 /* enqueue remote LOOKUP lock from the parent MDT */
2489                 __u64 rmt_ibits = MDS_INODELOCK_LOOKUP;
2490
2491                 if (mdt_object_remote(parent)) {
2492                         rc = mdt_remote_object_lock(info, parent,
2493                                                     mdt_object_fid(child),
2494                                                     &lhr->mlh_rreg_lh,
2495                                                     lhr->mlh_rreg_mode,
2496                                                     rmt_ibits, false);
2497                         if (rc != ELDLM_OK)
2498                                 return rc;
2499                 } else {
2500                         LASSERT(mdt_object_remote(child));
2501                         rc = mdt_object_local_lock(info, child, lhr,
2502                                                    &rmt_ibits, 0, true);
2503                         if (rc < 0)
2504                                 return rc;
2505                 }
2506
2507                 ibits &= ~MDS_INODELOCK_LOOKUP;
2508         }
2509
2510         if (mdt_object_remote(child)) {
2511                 rc = mdt_remote_object_lock(info, child, mdt_object_fid(child),
2512                                             &lhc->mlh_rreg_lh,
2513                                             lhc->mlh_rreg_mode,
2514                                             ibits, false);
2515                 if (rc == ELDLM_OK)
2516                         rc = 0;
2517         } else {
2518                 rc = mdt_reint_object_lock(info, child, lhc, ibits,
2519                                            cos_incompat);
2520         }
2521
2522         if (!rc)
2523                 mdt_object_unlock(info, child, lhr, rc);
2524
2525         return rc;
2526 }
2527
2528 /*
2529  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2530  * 2 - srcdir child; 3 - tgtdir child.
2531  * Update on disk version of srcdir child.
2532  */
2533 static int mdt_reint_rename(struct mdt_thread_info *info,
2534                             struct mdt_lock_handle *unused)
2535 {
2536         struct mdt_device *mdt = info->mti_mdt;
2537         struct mdt_reint_record *rr = &info->mti_rr;
2538         struct md_attr *ma = &info->mti_attr;
2539         struct ptlrpc_request *req = mdt_info_req(info);
2540         struct mdt_object *msrcdir = NULL;
2541         struct mdt_object *mtgtdir = NULL;
2542         struct mdt_object *mold;
2543         struct mdt_object *mnew = NULL;
2544         struct lustre_handle rename_lh = { 0 };
2545         struct mdt_lock_handle *lh_srcdirp;
2546         struct mdt_lock_handle *lh_tgtdirp;
2547         struct mdt_lock_handle *lh_oldp = NULL;
2548         struct mdt_lock_handle *lh_rmt = NULL;
2549         struct mdt_lock_handle *lh_newp = NULL;
2550         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2551         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2552         __u64 lock_ibits;
2553         bool reverse = false, discard = false;
2554         bool cos_incompat;
2555         ktime_t kstart = ktime_get();
2556         int rc;
2557
2558         ENTRY;
2559         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2560                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2561                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2562
2563         if (info->mti_dlm_req)
2564                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2565
2566         if (!fid_is_md_operative(rr->rr_fid1) ||
2567             !fid_is_md_operative(rr->rr_fid2))
2568                 RETURN(-EPERM);
2569
2570         /* find both parents. */
2571         msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2572         if (IS_ERR(msrcdir))
2573                 RETURN(PTR_ERR(msrcdir));
2574
2575         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2576
2577         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2578                 mtgtdir = msrcdir;
2579                 mdt_object_get(info->mti_env, mtgtdir);
2580         } else {
2581                 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2582                 if (IS_ERR(mtgtdir))
2583                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2584         }
2585
2586         /*
2587          * Note: do not enqueue rename lock for replay request, because
2588          * if other MDT holds rename lock, but being blocked to wait for
2589          * this MDT to finish its recovery, and the failover MDT can not
2590          * get rename lock, which will cause deadlock.
2591          */
2592         if (!req_is_replay(req)) {
2593                 /*
2594                  * Normally rename RPC is handled on the MDT with the target
2595                  * directory (if target exists, it's on the MDT with the
2596                  * target), if the source directory is remote, it's a hint that
2597                  * source is remote too (this may not be true, but it won't
2598                  * cause any issue), return -EXDEV early to avoid taking
2599                  * rename_lock.
2600                  */
2601                 if (!mdt->mdt_enable_remote_rename &&
2602                     mdt_object_remote(msrcdir))
2603                         GOTO(out_put_tgtdir, rc = -EXDEV);
2604
2605                 rc = mdt_rename_lock(info, &rename_lh);
2606                 if (rc != 0) {
2607                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2608                                mdt_obd_name(mdt), rc);
2609                         GOTO(out_put_tgtdir, rc);
2610                 }
2611         }
2612
2613         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2614         if (rc < 0)
2615                 GOTO(out_unlock_rename, rc);
2616
2617         reverse = rc;
2618
2619         /* source needs to be looked up after locking source parent, otherwise
2620          * this rename may race with unlink source, and cause rename hang, see
2621          * sanityn.sh 55b, so check parents first, if later we found source is
2622          * remote, relock parents.
2623          */
2624         cos_incompat = (mdt_object_remote(msrcdir) ||
2625                         mdt_object_remote(mtgtdir));
2626
2627         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2628
2629         /* lock parents in the proper order. */
2630         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2631         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2632
2633 relock:
2634         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2635         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2636
2637         if (reverse) {
2638                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2639                                           cos_incompat);
2640                 if (rc)
2641                         GOTO(out_unlock_rename, rc);
2642
2643                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2644
2645                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2646                                           cos_incompat);
2647                 if (rc != 0) {
2648                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2649                         GOTO(out_unlock_rename, rc);
2650                 }
2651         } else {
2652                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2653                                           cos_incompat);
2654                 if (rc)
2655                         GOTO(out_unlock_rename, rc);
2656
2657                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2658
2659                 if (mtgtdir != msrcdir) {
2660                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2661                                                   cos_incompat);
2662                 } else if (!mdt_object_remote(mtgtdir) &&
2663                            lh_srcdirp->mlh_pdo_hash !=
2664                            lh_tgtdirp->mlh_pdo_hash) {
2665                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2666                                                 MDS_INODELOCK_UPDATE,
2667                                                 cos_incompat);
2668                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2669                 }
2670                 if (rc != 0) {
2671                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2672                         GOTO(out_unlock_rename, rc);
2673                 }
2674         }
2675
2676         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2677         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2678
2679         /* find mold object. */
2680         fid_zero(old_fid);
2681         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2682         if (rc != 0)
2683                 GOTO(out_unlock_parents, rc);
2684
2685         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2686                 GOTO(out_unlock_parents, rc = -EINVAL);
2687
2688         if (!fid_is_md_operative(old_fid))
2689                 GOTO(out_unlock_parents, rc = -EPERM);
2690
2691         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2692         if (IS_ERR(mold))
2693                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2694
2695         if (!mdt_object_exists(mold)) {
2696                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2697                                 &mold->mot_obj,
2698                                 "object does not exist");
2699                 GOTO(out_put_old, rc = -ENOENT);
2700         }
2701
2702         if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2703                 GOTO(out_put_old, rc = -EXDEV);
2704
2705         /* Check if @mtgtdir is subdir of @mold, before locking child
2706          * to avoid reverse locking.
2707          */
2708         if (mtgtdir != msrcdir) {
2709                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2710                                    old_fid);
2711                 if (rc) {
2712                         if (rc == 1)
2713                                 rc = -EINVAL;
2714                         GOTO(out_put_old, rc);
2715                 }
2716         }
2717
2718         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2719         /* save version after locking */
2720         mdt_version_get_save(info, mold, 2);
2721
2722         if (!cos_incompat && mdt_object_remote(mold)) {
2723                 cos_incompat = true;
2724                 mdt_object_put(info->mti_env, mold);
2725                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2726                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2727                 goto relock;
2728         }
2729
2730         /* find mnew object:
2731          * mnew target object may not exist now
2732          * lookup with version checking
2733          */
2734         fid_zero(new_fid);
2735         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2736                                       3);
2737         if (rc == 0) {
2738                 /* the new_fid should have been filled at this moment */
2739                 if (lu_fid_eq(old_fid, new_fid))
2740                         GOTO(out_put_old, rc);
2741
2742                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2743                     lu_fid_eq(new_fid, rr->rr_fid2))
2744                         GOTO(out_put_old, rc = -EINVAL);
2745
2746                 if (!fid_is_md_operative(new_fid))
2747                         GOTO(out_put_old, rc = -EPERM);
2748
2749                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2750                 if (IS_ERR(mnew))
2751                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2752
2753                 if (!mdt_object_exists(mnew)) {
2754                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2755                                         &mnew->mot_obj,
2756                                         "object does not exist");
2757                         GOTO(out_put_new, rc = -ENOENT);
2758                 }
2759
2760                 if (mdt_object_remote(mnew)) {
2761                         struct mdt_body  *repbody;
2762
2763                         /* Always send rename req to the target child MDT */
2764                         repbody = req_capsule_server_get(info->mti_pill,
2765                                                          &RMF_MDT_BODY);
2766                         LASSERT(repbody != NULL);
2767                         repbody->mbo_fid1 = *new_fid;
2768                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2769                         GOTO(out_put_new, rc = -EXDEV);
2770                 }
2771                 /* Before locking the target dir, check we do not replace
2772                  * a dir with a non-dir, otherwise it may deadlock with
2773                  * link op which tries to create a link in this dir
2774                  * back to this non-dir.
2775                  */
2776                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2777                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2778                         GOTO(out_put_new, rc = -EISDIR);
2779
2780                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2781                 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2782                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2783                 mdt_lock_reg_init(lh_rmt, LCK_EX);
2784                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2785                 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2786                                             lh_rmt, lock_ibits, cos_incompat);
2787                 if (rc < 0)
2788                         GOTO(out_put_new, rc);
2789
2790                 /* Check if @msrcdir is subdir of @mnew, before locking child
2791                  * to avoid reverse locking.
2792                  */
2793                 if (mtgtdir != msrcdir) {
2794                         rc = mdo_is_subdir(info->mti_env,
2795                                            mdt_object_child(msrcdir), new_fid);
2796                         if (rc) {
2797                                 if (rc == 1)
2798                                         rc = -EINVAL;
2799                                 GOTO(out_unlock_old, rc);
2800                         }
2801                 }
2802
2803                 /* We used to acquire MDS_INODELOCK_FULL here but we
2804                  * can't do this now because a running HSM restore on
2805                  * the rename onto victim will hold the layout
2806                  * lock. See LU-4002.
2807                  */
2808
2809                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2810                 mdt_lock_reg_init(lh_newp, LCK_EX);
2811                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
2812                 if (mdt_object_remote(mtgtdir)) {
2813                         rc = mdt_remote_object_lock(info, mtgtdir,
2814                                                     mdt_object_fid(mnew),
2815                                                     &lh_newp->mlh_rreg_lh,
2816                                                     lh_newp->mlh_rreg_mode,
2817                                                     MDS_INODELOCK_LOOKUP,
2818                                                     false);
2819                         if (rc != ELDLM_OK)
2820                                 GOTO(out_unlock_old, rc);
2821
2822                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2823                 }
2824                 rc = mdt_reint_object_lock(info, mnew, lh_newp, lock_ibits,
2825                                            cos_incompat);
2826                 if (rc != 0)
2827                         GOTO(out_unlock_new, rc);
2828
2829                 /* get and save version after locking */
2830                 mdt_version_get_save(info, mnew, 3);
2831         } else if (rc != -ENOENT) {
2832                 GOTO(out_put_old, rc);
2833         } else {
2834                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2835                 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2836                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2837                 mdt_lock_reg_init(lh_rmt, LCK_EX);
2838                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2839                 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2840                                             lh_rmt, lock_ibits, cos_incompat);
2841                 if (rc != 0)
2842                         GOTO(out_put_old, rc);
2843
2844                 mdt_enoent_version_save(info, 3);
2845         }
2846
2847         /* step 5: rename it */
2848         mdt_reint_init_ma(info, ma);
2849
2850         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2851                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2852
2853         if (mnew != NULL)
2854                 mutex_lock(&mnew->mot_lov_mutex);
2855
2856         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2857                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2858                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2859                         &rr->rr_tgt_name, ma);
2860
2861         if (mnew != NULL)
2862                 mutex_unlock(&mnew->mot_lov_mutex);
2863
2864         /* handle last link of tgt object */
2865         if (rc == 0) {
2866                 mdt_counter_incr(req, LPROC_MDT_RENAME,
2867                                  ktime_us_delta(ktime_get(), kstart));
2868                 if (mnew) {
2869                         mdt_handle_last_unlink(info, mnew, ma);
2870                         discard = mdt_dom_check_for_discard(info, mnew);
2871                 }
2872                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2873                                          msrcdir, mtgtdir,
2874                                          ktime_us_delta(ktime_get(), kstart));
2875         }
2876
2877         EXIT;
2878 out_unlock_new:
2879         if (mnew != NULL)
2880                 mdt_object_unlock(info, mnew, lh_newp, rc);
2881 out_unlock_old:
2882         mdt_object_unlock(info, NULL, lh_rmt, rc);
2883         mdt_object_unlock(info, mold, lh_oldp, rc);
2884 out_put_new:
2885         if (mnew && !discard)
2886                 mdt_object_put(info->mti_env, mnew);
2887 out_put_old:
2888         mdt_object_put(info->mti_env, mold);
2889 out_unlock_parents:
2890         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2891         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2892 out_unlock_rename:
2893         if (lustre_handle_is_used(&rename_lh))
2894                 mdt_rename_unlock(&rename_lh);
2895 out_put_tgtdir:
2896         mdt_object_put(info->mti_env, mtgtdir);
2897 out_put_srcdir:
2898         mdt_object_put(info->mti_env, msrcdir);
2899
2900         /* The DoM discard can be done right in the place above where it is
2901          * assigned, meanwhile it is done here after rename unlock due to
2902          * compatibility with old clients, for them the discard blocks
2903          * the main thread until completion. Check LU-11359 for details.
2904          */
2905         if (discard) {
2906                 mdt_dom_discard_data(info, mnew);
2907                 mdt_object_put(info->mti_env, mnew);
2908         }
2909         return rc;
2910 }
2911
2912 static int mdt_reint_resync(struct mdt_thread_info *info,
2913                             struct mdt_lock_handle *lhc)
2914 {
2915         struct mdt_reint_record *rr = &info->mti_rr;
2916         struct ptlrpc_request *req = mdt_info_req(info);
2917         struct md_attr *ma = &info->mti_attr;
2918         struct mdt_object *mo;
2919         struct ldlm_lock *lease;
2920         struct mdt_body *repbody;
2921         struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2922         bool lease_broken;
2923         int rc, rc2;
2924
2925         ENTRY;
2926         DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2927
2928         if (info->mti_dlm_req)
2929                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2930
2931         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2932         if (IS_ERR(mo))
2933                 GOTO(out, rc = PTR_ERR(mo));
2934
2935         if (!mdt_object_exists(mo))
2936                 GOTO(out_obj, rc = -ENOENT);
2937
2938         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2939                 GOTO(out_obj, rc = -EINVAL);
2940
2941         if (mdt_object_remote(mo))
2942                 GOTO(out_obj, rc = -EREMOTE);
2943
2944         lease = ldlm_handle2lock(rr->rr_lease_handle);
2945         if (lease == NULL)
2946                 GOTO(out_obj, rc = -ESTALE);
2947
2948         /* It's really necessary to grab open_sem and check if the lease lock
2949          * has been lost. There would exist a concurrent writer coming in and
2950          * generating some dirty data in memory cache, the writeback would fail
2951          * after the layout version is increased by MDS_REINT_RESYNC RPC.
2952          */
2953         if (!down_write_trylock(&mo->mot_open_sem))
2954                 GOTO(out_put_lease, rc = -EBUSY);
2955
2956         lock_res_and_lock(lease);
2957         lease_broken = ldlm_is_cancel(lease);
2958         unlock_res_and_lock(lease);
2959         if (lease_broken)
2960                 GOTO(out_unlock, rc = -EBUSY);
2961
2962         /* the file has yet opened by anyone else after we took the lease. */
2963         layout.mlc_opc = MD_LAYOUT_RESYNC;
2964         lhc = &info->mti_lh[MDT_LH_LOCAL];
2965         rc = mdt_layout_change(info, mo, lhc, &layout);
2966         if (rc)
2967                 GOTO(out_unlock, rc);
2968
2969         mdt_object_unlock(info, mo, lhc, 0);
2970
2971         ma->ma_need = MA_INODE;
2972         ma->ma_valid = 0;
2973         rc = mdt_attr_get_complex(info, mo, ma);
2974         if (rc != 0)
2975                 GOTO(out_unlock, rc);
2976
2977         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2978         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2979
2980         EXIT;
2981 out_unlock:
2982         up_write(&mo->mot_open_sem);
2983 out_put_lease:
2984         LDLM_LOCK_PUT(lease);
2985 out_obj:
2986         mdt_object_put(info->mti_env, mo);
2987 out:
2988         mdt_client_compatibility(info);
2989         rc2 = mdt_fix_reply(info);
2990         if (rc == 0)
2991                 rc = rc2;
2992         return rc;
2993 }
2994
2995 struct mdt_reinter {
2996         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2997         enum lprocfs_extra_opc mr_extra_opc;
2998 };
2999
3000 static const struct mdt_reinter mdt_reinters[] = {
3001         [REINT_SETATTR] = {
3002                 .mr_handler = &mdt_reint_setattr,
3003                 .mr_extra_opc = MDS_REINT_SETATTR,
3004         },
3005         [REINT_CREATE] = {
3006                 .mr_handler = &mdt_reint_create,
3007                 .mr_extra_opc = MDS_REINT_CREATE,
3008         },
3009         [REINT_LINK] = {
3010                 .mr_handler = &mdt_reint_link,
3011                 .mr_extra_opc = MDS_REINT_LINK,
3012         },
3013         [REINT_UNLINK] = {
3014                 .mr_handler = &mdt_reint_unlink,
3015                 .mr_extra_opc = MDS_REINT_UNLINK,
3016         },
3017         [REINT_RENAME] = {
3018                 .mr_handler = &mdt_reint_rename,
3019                 .mr_extra_opc = MDS_REINT_RENAME,
3020         },
3021         [REINT_OPEN] = {
3022                 .mr_handler = &mdt_reint_open,
3023                 .mr_extra_opc = MDS_REINT_OPEN,
3024         },
3025         [REINT_SETXATTR] = {
3026                 .mr_handler = &mdt_reint_setxattr,
3027                 .mr_extra_opc = MDS_REINT_SETXATTR,
3028         },
3029         [REINT_RMENTRY] = {
3030                 .mr_handler = &mdt_reint_unlink,
3031                 .mr_extra_opc = MDS_REINT_UNLINK,
3032         },
3033         [REINT_MIGRATE] = {
3034                 .mr_handler = &mdt_reint_migrate,
3035                 .mr_extra_opc = MDS_REINT_RENAME,
3036         },
3037         [REINT_RESYNC] = {
3038                 .mr_handler = &mdt_reint_resync,
3039                 .mr_extra_opc = MDS_REINT_RESYNC,
3040         },
3041 };
3042
3043 int mdt_reint_rec(struct mdt_thread_info *info,
3044                   struct mdt_lock_handle *lhc)
3045 {
3046         const struct mdt_reinter *mr;
3047         int rc;
3048
3049         ENTRY;
3050         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
3051                 RETURN(-EPROTO);
3052
3053         mr = &mdt_reinters[info->mti_rr.rr_opcode];
3054         if (mr->mr_handler == NULL)
3055                 RETURN(-EPROTO);
3056
3057         rc = (*mr->mr_handler)(info, lhc);
3058
3059         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
3060                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
3061
3062         RETURN(rc);
3063 }