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