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         ENTRY;
  85
  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 static 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                 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
 210                 if (likely(!IS_ERR(child))) {
 211                         mdt_obj_version_get(info, child, &info->mti_ver[idx]);
 212                         mdt_object_put(info->mti_env, child);
 213                 }
 214         }
 215         vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
 216         return vbrc ? vbrc : rc;
 217
 218 }
 219
 220 static inline int mdt_remote_permission_check(struct mdt_thread_info *info)
 221 {
 222         struct lu_ucred *uc  = mdt_ucred(info);
 223         struct mdt_device *mdt = info->mti_mdt;
 224
 225         if (!md_capable(uc, CFS_CAP_SYS_ADMIN)) {
 226                 if (uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 227                     mdt->mdt_enable_remote_dir_gid != -1)
 228                         return -EPERM;
 229         }
 230
 231         return 0;
 232 }
 233
 234 /**
 235  * mdt_remote_permission: Check whether the remote operation is permitted,
 236  *
 237  * Only sysadmin can create remote directory / striped directory,
 238  * migrate directory and set default stripedEA on directory, unless
 239  *
 240  * lctl set_param mdt.*.enable_remote_dir_gid=allow_gid.
 241  *
 242  * param[in] info: mdt_thread_info.
 243  *
 244  * retval       = 0 remote operation is allowed.
 245  *              < 0 remote operation is denied.
 246  */
 247 int mdt_remote_permission(struct mdt_thread_info *info)
 248 {
 249         struct md_op_spec *spec = &info->mti_spec;
 250         struct lu_attr *attr = &info->mti_attr.ma_attr;
 251         struct obd_export *exp = mdt_info_req(info)->rq_export;
 252         int rc;
 253
 254         if (info->mti_rr.rr_opcode == REINT_MIGRATE) {
 255                 rc = mdt_remote_permission_check(info);
 256                 if (rc != 0)
 257                         return rc;
 258         }
 259
 260         if (info->mti_rr.rr_opcode == REINT_CREATE &&
 261             (S_ISDIR(attr->la_mode) && spec->u.sp_ea.eadata != NULL &&
 262              spec->u.sp_ea.eadatalen != 0)) {
 263                 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
 264
 265                 /* Only new clients can create remote dir( >= 2.4) and
 266                  * striped dir(>= 2.6), old client will return -ENOTSUPP */
 267                 if (!mdt_is_dne_client(exp))
 268                         return -ENOTSUPP;
 269
 270                 if (le32_to_cpu(lum->lum_stripe_count) > 1 &&
 271                     !mdt_is_striped_client(exp))
 272                         return -ENOTSUPP;
 273
 274                 rc = mdt_remote_permission_check(info);
 275                 if (rc != 0)
 276                         return rc;
 277         }
 278
 279         if (info->mti_rr.rr_opcode == REINT_SETATTR) {
 280                 struct md_attr *ma = &info->mti_attr;
 281
 282                 if ((ma->ma_valid & MA_LMV)) {
 283                         rc = mdt_remote_permission_check(info);
 284                         if (rc != 0)
 285                                 return rc;
 286                 }
 287         }
 288
 289         return 0;
 290 }
 291
 292 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
 293                              struct mdt_object *obj,
 294                              struct ldlm_enqueue_info *einfo,
 295                              int decref)
 296 {
 297         union ldlm_policy_data *policy = &mti->mti_policy;
 298         struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
 299         struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
 300         int i;
 301
 302         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 303         LASSERT(slave_locks);
 304
 305         memset(policy, 0, sizeof(*policy));
 306         policy->l_inodebits.bits = einfo->ei_inodebits;
 307         mdt_lock_handle_init(lh);
 308         mdt_lock_reg_init(lh, einfo->ei_mode);
 309         for (i = 0; i < slave_locks->ha_count; i++) {
 310                 if (test_bit(i, (void *)slave_locks->ha_map))
 311                         lh->mlh_rreg_lh = slave_locks->ha_handles[i];
 312                 else
 313                         lh->mlh_reg_lh = slave_locks->ha_handles[i];
 314                 mdt_object_unlock(mti, NULL, lh, decref);
 315                 slave_locks->ha_handles[i].cookie = 0ull;
 316         }
 317
 318         return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
 319                                 policy);
 320 }
 321
 322 static inline int mdt_object_striped(struct mdt_thread_info *mti,
 323                                      struct mdt_object *obj)
 324 {
 325         int rc;
 326
 327         if (!S_ISDIR(obj->mot_header.loh_attr))
 328                 return 0;
 329
 330         rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), &LU_BUF_NULL,
 331                           XATTR_NAME_LMV);
 332         if (rc <= 0)
 333                 return rc == -ENODATA ? 0 : rc;
 334
 335         return 1;
 336 }
 337
 338 /**
 339  * Lock slave stripes if necessary, the lock handles of slave stripes
 340  * will be stored in einfo->ei_cbdata.
 341  **/
 342 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
 343                            enum ldlm_mode mode, __u64 ibits,
 344                            struct ldlm_enqueue_info *einfo)
 345 {
 346         union ldlm_policy_data *policy = &mti->mti_policy;
 347
 348         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 349
 350         einfo->ei_type = LDLM_IBITS;
 351         einfo->ei_mode = mode;
 352         einfo->ei_cb_bl = mdt_remote_blocking_ast;
 353         einfo->ei_cb_local_bl = mdt_blocking_ast;
 354         einfo->ei_cb_cp = ldlm_completion_ast;
 355         einfo->ei_enq_slave = 1;
 356         einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
 357         einfo->ei_inodebits = ibits;
 358         memset(policy, 0, sizeof(*policy));
 359         policy->l_inodebits.bits = ibits;
 360
 361         return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
 362                               policy);
 363 }
 364
 365 int mdt_reint_striped_lock(struct mdt_thread_info *info,
 366                            struct mdt_object *o,
 367                            struct mdt_lock_handle *lh,
 368                            __u64 ibits,
 369                            struct ldlm_enqueue_info *einfo,
 370                            bool cos_incompat)
 371 {
 372         int rc;
 373
 374         LASSERT(!mdt_object_remote(o));
 375
 376         memset(einfo, 0, sizeof(*einfo));
 377
 378         rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
 379         if (rc)
 380                 return rc;
 381
 382         rc = mdt_object_striped(info, o);
 383         if (rc != 1) {
 384                 if (rc < 0)
 385                         mdt_object_unlock(info, o, lh, rc);
 386                 return rc;
 387         }
 388
 389         rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
 390         if (rc) {
 391                 mdt_object_unlock(info, o, lh, rc);
 392                 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
 393                         rc = 0;
 394         }
 395
 396         return rc;
 397 }
 398
 399 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
 400                               struct mdt_object *o,
 401                               struct mdt_lock_handle *lh,
 402                               struct ldlm_enqueue_info *einfo, int decref)
 403 {
 404         if (einfo->ei_cbdata)
 405                 mdt_unlock_slaves(info, o, einfo, decref);
 406         mdt_object_unlock(info, o, lh, decref);
 407 }
 408
 409 /*
 410  * VBR: we save three versions in reply:
 411  * 0 - parent. Check that parent version is the same during replay.
 412  * 1 - name. Version of 'name' if file exists with the same name or
 413  * ENOENT_VERSION, it is needed because file may appear due to missed replays.
 414  * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
 415  * check.
 416  */
 417 static int mdt_create(struct mdt_thread_info *info)
 418 {
 419         struct mdt_device       *mdt = info->mti_mdt;
 420         struct mdt_object       *parent;
 421         struct mdt_object       *child;
 422         struct mdt_lock_handle  *lh;
 423         struct mdt_body         *repbody;
 424         struct md_attr          *ma = &info->mti_attr;
 425         struct mdt_reint_record *rr = &info->mti_rr;
 426         int rc;
 427         ENTRY;
 428
 429         DEBUG_REQ(D_INODE, mdt_info_req(info), "Create  ("DNAME"->"DFID") "
 430                   "in "DFID,
 431                   PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
 432
 433         if (!fid_is_md_operative(rr->rr_fid1))
 434                 RETURN(-EPERM);
 435
 436         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 437
 438         parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 439         if (IS_ERR(parent))
 440                 RETURN(PTR_ERR(parent));
 441
 442         if (!mdt_object_exists(parent))
 443                 GOTO(put_parent, rc = -ENOENT);
 444
 445         lh = &info->mti_lh[MDT_LH_PARENT];
 446         mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 447         rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
 448         if (rc)
 449                 GOTO(put_parent, rc);
 450
 451         if (!mdt_object_remote(parent)) {
 452                 rc = mdt_version_get_check_save(info, parent, 0);
 453                 if (rc)
 454                         GOTO(unlock_parent, rc);
 455         }
 456
 457         /*
 458          * Check child name version during replay.
 459          * During create replay a file may exist with same name.
 460          */
 461         rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
 462                                       &info->mti_tmp_fid1, 1);
 463         if (rc == 0)
 464                 GOTO(unlock_parent, rc = -EEXIST);
 465
 466         /* -ENOENT is expected here */
 467         if (rc != -ENOENT)
 468                 GOTO(unlock_parent, rc);
 469
 470         /* save version of file name for replay, it must be ENOENT here */
 471         mdt_enoent_version_save(info, 1);
 472
 473         child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
 474         if (unlikely(IS_ERR(child)))
 475                 GOTO(unlock_parent, rc = PTR_ERR(child));
 476
 477         rc = mdt_remote_permission(info);
 478         if (rc != 0)
 479                 GOTO(put_child, rc);
 480
 481         ma->ma_need = MA_INODE;
 482         ma->ma_valid = 0;
 483
 484         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 485                         OBD_FAIL_MDS_REINT_CREATE_WRITE);
 486
 487         /* Version of child will be updated on disk. */
 488         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 489         rc = mdt_version_get_check_save(info, child, 2);
 490         if (rc)
 491                 GOTO(put_child, rc);
 492
 493         /* Let lower layer know current lock mode. */
 494         info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
 495
 496         /*
 497          * Do not perform lookup sanity check. We know that name does
 498          * not exist.
 499          */
 500         info->mti_spec.sp_cr_lookup = 0;
 501         info->mti_spec.sp_feat = &dt_directory_features;
 502
 503         rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
 504                         mdt_object_child(child), &info->mti_spec, ma);
 505         if (rc == 0)
 506                 rc = mdt_attr_get_complex(info, child, ma);
 507
 508         if (rc < 0)
 509                 GOTO(put_child, rc);
 510
 511         /*
 512          * On DNE, we need to eliminate dependey between 'mkdir a' and
 513          * 'mkdir a/b' if b is a striped directory, to achieve this, two
 514          * things are done below:
 515          * 1. save child and slaves lock.
 516          * 2. if the child is a striped directory, relock parent so to
 517          *    compare against with COS locks to ensure parent was
 518          *    committed to disk.
 519          */
 520         if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
 521                 struct mdt_lock_handle *lhc;
 522                 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 523                 bool cos_incompat;
 524
 525                 rc = mdt_object_striped(info, child);
 526                 if (rc < 0)
 527                         GOTO(put_child, rc);
 528
 529                 cos_incompat = rc;
 530                 if (cos_incompat) {
 531                         if (!mdt_object_remote(parent)) {
 532                                 mdt_object_unlock(info, parent, lh, 1);
 533                                 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 534                                 rc = mdt_reint_object_lock(info, parent, lh,
 535                                                            MDS_INODELOCK_UPDATE,
 536                                                            true);
 537                                 if (rc)
 538                                         GOTO(put_child, rc);
 539                         }
 540                 }
 541
 542                 lhc = &info->mti_lh[MDT_LH_CHILD];
 543                 mdt_lock_handle_init(lhc);
 544                 mdt_lock_reg_init(lhc, LCK_PW);
 545                 rc = mdt_reint_striped_lock(info, child, lhc,
 546                                             MDS_INODELOCK_UPDATE, einfo,
 547                                             cos_incompat);
 548                 if (rc)
 549                         GOTO(put_child, rc);
 550
 551                 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 552         }
 553
 554         /* Return fid & attr to client. */
 555         if (ma->ma_valid & MA_INODE)
 556                 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
 557                                    mdt_object_fid(child));
 558 put_child:
 559         mdt_object_put(info->mti_env, child);
 560 unlock_parent:
 561         mdt_object_unlock(info, parent, lh, rc);
 562 put_parent:
 563         mdt_object_put(info->mti_env, parent);
 564         RETURN(rc);
 565 }
 566
 567 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
 568                         struct md_attr *ma)
 569 {
 570         struct mdt_lock_handle  *lh;
 571         int do_vbr = ma->ma_attr.la_valid &
 572                         (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
 573         __u64 lockpart = MDS_INODELOCK_UPDATE;
 574         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 575         bool cos_incompat;
 576         int rc;
 577         ENTRY;
 578
 579         rc = mdt_object_striped(info, mo);
 580         if (rc < 0)
 581                 RETURN(rc);
 582
 583         cos_incompat = rc;
 584
 585         lh = &info->mti_lh[MDT_LH_PARENT];
 586         mdt_lock_reg_init(lh, LCK_PW);
 587
 588         /* Even though the new MDT will grant PERM lock to the old
 589          * client, but the old client will almost ignore that during
 590          * So it needs to revoke both LOOKUP and PERM lock here, so
 591          * both new and old client can cancel the dcache */
 592         if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
 593                 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
 594
 595         rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
 596                                     cos_incompat);
 597         if (rc != 0)
 598                 RETURN(rc);
 599
 600         /* all attrs are packed into mti_attr in unpack_setattr */
 601         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 602                        OBD_FAIL_MDS_REINT_SETATTR_WRITE);
 603
 604         /* VBR: update version if attr changed are important for recovery */
 605         if (do_vbr) {
 606                 /* update on-disk version of changed object */
 607                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
 608                 rc = mdt_version_get_check_save(info, mo, 0);
 609                 if (rc)
 610                         GOTO(out_unlock, rc);
 611         }
 612
 613         /* Ensure constant striping during chown(). See LU-2789. */
 614         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 615                 mutex_lock(&mo->mot_lov_mutex);
 616
 617         /* all attrs are packed into mti_attr in unpack_setattr */
 618         rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
 619
 620         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 621                 mutex_unlock(&mo->mot_lov_mutex);
 622
 623         if (rc != 0)
 624                 GOTO(out_unlock, rc);
 625         mdt_dom_obj_lvb_update(info->mti_env, mo, false);
 626         EXIT;
 627 out_unlock:
 628         mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
 629         return rc;
 630 }
 631
 632 /**
 633  * Check HSM flags and add HS_DIRTY flag if relevant.
 634  *
 635  * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
 636  * and is not RELEASED.
 637  */
 638 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
 639                         struct md_attr *ma)
 640 {
 641         int rc;
 642         ENTRY;
 643
 644         /* If the file was modified, add the dirty flag */
 645         ma->ma_need = MA_HSM;
 646         rc = mdt_attr_get_complex(info, mo, ma);
 647         if (rc) {
 648                 CERROR("file attribute read error for "DFID": %d.\n",
 649                         PFID(mdt_object_fid(mo)), rc);
 650                 RETURN(rc);
 651         }
 652
 653         /* If an up2date copy exists in the backend, add dirty flag */
 654         if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
 655             && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
 656                 ma->ma_hsm.mh_flags |= HS_DIRTY;
 657
 658                 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
 659                 if (rc)
 660                         CERROR("file attribute change error for "DFID": %d\n",
 661                                 PFID(mdt_object_fid(mo)), rc);
 662         }
 663
 664         RETURN(rc);
 665 }
 666
 667 static int mdt_reint_setattr(struct mdt_thread_info *info,
 668                              struct mdt_lock_handle *lhc)
 669 {
 670         struct md_attr          *ma = &info->mti_attr;
 671         struct mdt_reint_record *rr = &info->mti_rr;
 672         struct ptlrpc_request   *req = mdt_info_req(info);
 673         struct mdt_object       *mo;
 674         struct mdt_body         *repbody;
 675         int                      rc, rc2;
 676         ENTRY;
 677
 678         DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
 679                   (unsigned int)ma->ma_attr.la_valid);
 680
 681         if (info->mti_dlm_req)
 682                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 683
 684         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 685         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 686         if (IS_ERR(mo))
 687                 GOTO(out, rc = PTR_ERR(mo));
 688
 689         if (!mdt_object_exists(mo))
 690                 GOTO(out_put, rc = -ENOENT);
 691
 692         if (mdt_object_remote(mo))
 693                 GOTO(out_put, rc = -EREMOTE);
 694
 695         /* revoke lease lock if size is going to be changed */
 696         if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
 697                      !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
 698                      atomic_read(&mo->mot_lease_count) > 0)) {
 699                 down_read(&mo->mot_open_sem);
 700
 701                 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
 702                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 703                         mdt_lock_reg_init(lhc, LCK_CW);
 704
 705                         rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
 706                         if (rc != 0) {
 707                                 up_read(&mo->mot_open_sem);
 708                                 GOTO(out_put, rc);
 709                         }
 710
 711                         /* revoke lease lock */
 712                         mdt_object_unlock(info, mo, lhc, 1);
 713                 }
 714                 up_read(&mo->mot_open_sem);
 715         }
 716
 717         if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
 718                 /* Check write access for the O_TRUNC case */
 719                 if (mdt_write_read(mo) < 0)
 720                         GOTO(out_put, rc = -ETXTBSY);
 721
 722                 /* LU-10286: compatibility check for FLR.
 723                  * Please check the comment in mdt_finish_open() for details */
 724                 if (!exp_connect_flr(info->mti_exp)) {
 725                         rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
 726                         if (rc < 0 && rc != -ENODATA)
 727                                 GOTO(out_put, rc);
 728
 729                         if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm))
 730                                 GOTO(out_put, rc = -EOPNOTSUPP);
 731                 }
 732
 733                 /* For truncate, the file size sent from client
 734                  * is believable, but the blocks are incorrect,
 735                  * which makes the block size in LSOM attribute
 736                  * inconsisent with the real block size.
 737                  */
 738                 rc = mdt_lsom_update(info, mo, true);
 739                 if (rc)
 740                         GOTO(out_put, rc);
 741         }
 742
 743         if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
 744                 if (ma->ma_valid & MA_LOV)
 745                         GOTO(out_put, rc = -EPROTO);
 746
 747                 rc = mdt_attr_set(info, mo, ma);
 748                 if (rc)
 749                         GOTO(out_put, rc);
 750         } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
 751                    (ma->ma_valid & MA_INODE)) {
 752                 struct lu_buf *buf  = &info->mti_buf;
 753                 struct mdt_lock_handle  *lh;
 754
 755                 rc = mdt_remote_permission(info);
 756                 if (rc < 0)
 757                         GOTO(out_put, rc);
 758
 759                 if (ma->ma_attr.la_valid != 0)
 760                         GOTO(out_put, rc = -EPROTO);
 761
 762                 lh = &info->mti_lh[MDT_LH_PARENT];
 763                 mdt_lock_reg_init(lh, LCK_PW);
 764
 765                 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
 766                 if (rc != 0)
 767                         GOTO(out_put, rc);
 768
 769                 if (ma->ma_valid & MA_LOV) {
 770                         buf->lb_buf = ma->ma_lmm;
 771                         buf->lb_len = ma->ma_lmm_size;
 772                 } else {
 773                         buf->lb_buf = ma->ma_lmv;
 774                         buf->lb_len = ma->ma_lmv_size;
 775                 }
 776                 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
 777                                   (ma->ma_valid & MA_LOV) ?
 778                                         XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
 779                                   0);
 780
 781                 mdt_object_unlock(info, mo, lh, rc);
 782                 if (rc)
 783                         GOTO(out_put, rc);
 784         } else {
 785                 GOTO(out_put, rc = -EPROTO);
 786         }
 787
 788         /* If file data is modified, add the dirty flag */
 789         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
 790                 rc = mdt_add_dirty_flag(info, mo, ma);
 791
 792         ma->ma_need = MA_INODE;
 793         ma->ma_valid = 0;
 794         rc = mdt_attr_get_complex(info, mo, ma);
 795         if (rc != 0)
 796                 GOTO(out_put, rc);
 797
 798         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
 799
 800         EXIT;
 801 out_put:
 802         mdt_object_put(info->mti_env, mo);
 803 out:
 804         if (rc == 0)
 805                 mdt_counter_incr(req, LPROC_MDT_SETATTR);
 806
 807         mdt_client_compatibility(info);
 808         rc2 = mdt_fix_reply(info);
 809         if (rc == 0)
 810                 rc = rc2;
 811         return rc;
 812 }
 813
 814 static int mdt_reint_create(struct mdt_thread_info *info,
 815                             struct mdt_lock_handle *lhc)
 816 {
 817         struct ptlrpc_request   *req = mdt_info_req(info);
 818         int                     rc;
 819         ENTRY;
 820
 821         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 822                 RETURN(err_serious(-ESTALE));
 823
 824         if (info->mti_dlm_req)
 825                 ldlm_request_cancel(mdt_info_req(info),
 826                                     info->mti_dlm_req, 0, LATF_SKIP);
 827
 828         if (!lu_name_is_valid(&info->mti_rr.rr_name))
 829                 RETURN(-EPROTO);
 830
 831         switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
 832         case S_IFDIR:
 833                 mdt_counter_incr(req, LPROC_MDT_MKDIR);
 834                 break;
 835         case S_IFREG:
 836         case S_IFLNK:
 837         case S_IFCHR:
 838         case S_IFBLK:
 839         case S_IFIFO:
 840         case S_IFSOCK:
 841                 /* Special file should stay on the same node as parent. */
 842                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
 843                 break;
 844         default:
 845                 CERROR("%s: Unsupported mode %o\n",
 846                        mdt_obd_name(info->mti_mdt),
 847                        info->mti_attr.ma_attr.la_mode);
 848                 RETURN(err_serious(-EOPNOTSUPP));
 849         }
 850
 851         rc = mdt_create(info);
 852         RETURN(rc);
 853 }
 854
 855 /*
 856  * VBR: save parent version in reply and child version getting by its name.
 857  * Version of child is getting and checking during its lookup. If
 858  */
 859 static int mdt_reint_unlink(struct mdt_thread_info *info,
 860                             struct mdt_lock_handle *lhc)
 861 {
 862         struct mdt_reint_record *rr = &info->mti_rr;
 863         struct ptlrpc_request *req = mdt_info_req(info);
 864         struct md_attr *ma = &info->mti_attr;
 865         struct lu_fid *child_fid = &info->mti_tmp_fid1;
 866         struct mdt_object *mp;
 867         struct mdt_object *mc;
 868         struct mdt_lock_handle *parent_lh;
 869         struct mdt_lock_handle *child_lh;
 870         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 871         __u64 lock_ibits;
 872         bool cos_incompat = false, discard = false;
 873         int no_name = 0;
 874         int rc;
 875
 876         ENTRY;
 877
 878         DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
 879                   PNAME(&rr->rr_name));
 880
 881         if (info->mti_dlm_req)
 882                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 883
 884         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
 885                 RETURN(err_serious(-ENOENT));
 886
 887         if (!fid_is_md_operative(rr->rr_fid1))
 888                 RETURN(-EPERM);
 889
 890         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 891         if (IS_ERR(mp))
 892                 RETURN(PTR_ERR(mp));
 893
 894         if (mdt_object_remote(mp)) {
 895                 cos_incompat = true;
 896         } else {
 897                 rc = mdt_version_get_check_save(info, mp, 0);
 898                 if (rc)
 899                         GOTO(put_parent, rc);
 900         }
 901
 902 relock:
 903         parent_lh = &info->mti_lh[MDT_LH_PARENT];
 904         mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
 905         rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
 906                                    cos_incompat);
 907         if (rc != 0)
 908                 GOTO(put_parent, rc);
 909
 910         /* lookup child object along with version checking */
 911         fid_zero(child_fid);
 912         rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
 913         if (rc != 0) {
 914                 /* Name might not be able to find during resend of
 915                  * remote unlink, considering following case.
 916                  * dir_A is a remote directory, the name entry of
 917                  * dir_A is on MDT0, the directory is on MDT1,
 918                  *
 919                  * 1. client sends unlink req to MDT1.
 920                  * 2. MDT1 sends name delete update to MDT0.
 921                  * 3. name entry is being deleted in MDT0 synchronously.
 922                  * 4. MDT1 is restarted.
 923                  * 5. client resends unlink req to MDT1. So it can not
 924                  *    find the name entry on MDT0 anymore.
 925                  * In this case, MDT1 only needs to destory the local
 926                  * directory.
 927                  * */
 928                 if (mdt_object_remote(mp) && rc == -ENOENT &&
 929                     !fid_is_zero(rr->rr_fid2) &&
 930                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 931                         no_name = 1;
 932                         *child_fid = *rr->rr_fid2;
 933                  } else {
 934                         GOTO(unlock_parent, rc);
 935                  }
 936         }
 937
 938         if (!fid_is_md_operative(child_fid))
 939                 GOTO(unlock_parent, rc = -EPERM);
 940
 941         /* We will lock the child regardless it is local or remote. No harm. */
 942         mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
 943         if (IS_ERR(mc))
 944                 GOTO(unlock_parent, rc = PTR_ERR(mc));
 945
 946         if (!cos_incompat) {
 947                 rc = mdt_object_striped(info, mc);
 948                 if (rc < 0)
 949                         GOTO(unlock_parent, rc = PTR_ERR(mc));
 950
 951                 cos_incompat = rc;
 952                 if (cos_incompat) {
 953                         mdt_object_put(info->mti_env, mc);
 954                         mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
 955                         goto relock;
 956                 }
 957         }
 958
 959         child_lh = &info->mti_lh[MDT_LH_CHILD];
 960         mdt_lock_reg_init(child_lh, LCK_EX);
 961         if (info->mti_spec.sp_rm_entry) {
 962                 struct lu_ucred *uc  = mdt_ucred(info);
 963
 964                 if (!mdt_is_dne_client(req->rq_export))
 965                         /* Return -ENOTSUPP for old client */
 966                         GOTO(put_child, rc = -ENOTSUPP);
 967
 968                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
 969                         GOTO(put_child, rc = -EPERM);
 970
 971                 ma->ma_need = MA_INODE;
 972                 ma->ma_valid = 0;
 973                 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
 974                                 NULL, &rr->rr_name, ma, no_name);
 975                 GOTO(put_child, rc);
 976         }
 977
 978         if (mdt_object_remote(mc)) {
 979                 struct mdt_body  *repbody;
 980
 981                 if (!fid_is_zero(rr->rr_fid2)) {
 982                         CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
 983                                mdt_obd_name(info->mti_mdt),
 984                                PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 985                         GOTO(put_child, rc = -ENOENT);
 986                 }
 987                 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
 988                        mdt_obd_name(info->mti_mdt),
 989                        PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 990
 991                 if (!mdt_is_dne_client(req->rq_export))
 992                         /* Return -ENOTSUPP for old client */
 993                         GOTO(put_child, rc = -ENOTSUPP);
 994
 995                 /* Revoke the LOOKUP lock of the remote object granted by
 996                  * this MDT. Since the unlink will happen on another MDT,
 997                  * it will release the LOOKUP lock right away. Then What
 998                  * would happen if another client try to grab the LOOKUP
 999                  * lock at the same time with unlink XXX */
1000                 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1001                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1002                 LASSERT(repbody != NULL);
1003                 repbody->mbo_fid1 = *mdt_object_fid(mc);
1004                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1005                 GOTO(unlock_child, rc = -EREMOTE);
1006         }
1007         /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1008          * this now because a running HSM restore on the child (unlink
1009          * victim) will hold the layout lock. See LU-4002. */
1010         lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1011         if (mdt_object_remote(mp)) {
1012                 /* Enqueue lookup lock from parent MDT */
1013                 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1014                                             &child_lh->mlh_rreg_lh,
1015                                             child_lh->mlh_rreg_mode,
1016                                             MDS_INODELOCK_LOOKUP, false);
1017                 if (rc != ELDLM_OK)
1018                         GOTO(put_child, rc);
1019
1020                 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1021         }
1022
1023         rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1024                                     cos_incompat);
1025         if (rc != 0)
1026                 GOTO(put_child, rc);
1027
1028         /*
1029          * Now we can only make sure we need MA_INODE, in mdd layer, will check
1030          * whether need MA_LOV and MA_COOKIE.
1031          */
1032         ma->ma_need = MA_INODE;
1033         ma->ma_valid = 0;
1034
1035         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1036                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1037         /* save version when object is locked */
1038         mdt_version_get_save(info, mc, 1);
1039
1040         mutex_lock(&mc->mot_lov_mutex);
1041
1042         rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1043                         mdt_object_child(mc), &rr->rr_name, ma, no_name);
1044
1045         mutex_unlock(&mc->mot_lov_mutex);
1046         if (rc != 0)
1047                 GOTO(unlock_child, rc);
1048
1049         if (!lu_object_is_dying(&mc->mot_header)) {
1050                 rc = mdt_attr_get_complex(info, mc, ma);
1051                 if (rc)
1052                         GOTO(out_stat, rc);
1053         } else {
1054                 discard = mdt_dom_check_for_discard(info, mc);
1055         }
1056         mdt_handle_last_unlink(info, mc, ma);
1057
1058 out_stat:
1059         if (ma->ma_valid & MA_INODE) {
1060                 switch (ma->ma_attr.la_mode & S_IFMT) {
1061                 case S_IFDIR:
1062                         mdt_counter_incr(req, LPROC_MDT_RMDIR);
1063                         break;
1064                 case S_IFREG:
1065                 case S_IFLNK:
1066                 case S_IFCHR:
1067                 case S_IFBLK:
1068                 case S_IFIFO:
1069                 case S_IFSOCK:
1070                         mdt_counter_incr(req, LPROC_MDT_UNLINK);
1071                         break;
1072                 default:
1073                         LASSERTF(0, "bad file type %o unlinking\n",
1074                                 ma->ma_attr.la_mode);
1075                 }
1076         }
1077
1078         EXIT;
1079
1080 unlock_child:
1081         mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1082 put_child:
1083         mdt_object_put(info->mti_env, mc);
1084 unlock_parent:
1085         mdt_object_unlock(info, mp, parent_lh, rc);
1086 put_parent:
1087         mdt_object_put(info->mti_env, mp);
1088
1089         /* discard is just a PW DOM lock to drop the data on a client
1090          * no need to keep objects being get and locked, do that after all.
1091          */
1092         if (discard)
1093                 mdt_dom_discard_data(info, child_fid);
1094
1095         return rc;
1096 }
1097
1098 /*
1099  * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1100  * name.
1101  */
1102 static int mdt_reint_link(struct mdt_thread_info *info,
1103                           struct mdt_lock_handle *lhc)
1104 {
1105         struct mdt_reint_record *rr = &info->mti_rr;
1106         struct ptlrpc_request   *req = mdt_info_req(info);
1107         struct md_attr          *ma = &info->mti_attr;
1108         struct mdt_object       *ms;
1109         struct mdt_object       *mp;
1110         struct mdt_lock_handle  *lhs;
1111         struct mdt_lock_handle  *lhp;
1112         bool cos_incompat;
1113         int rc;
1114         ENTRY;
1115
1116         DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1117                   PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1118
1119         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1120                 RETURN(err_serious(-ENOENT));
1121
1122         if (info->mti_dlm_req)
1123                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1124
1125         /* Invalid case so return error immediately instead of
1126          * processing it */
1127         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1128                 RETURN(-EPERM);
1129
1130         if (!fid_is_md_operative(rr->rr_fid1) ||
1131             !fid_is_md_operative(rr->rr_fid2))
1132                 RETURN(-EPERM);
1133
1134         /* step 1: find target parent dir */
1135         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1136         if (IS_ERR(mp))
1137                 RETURN(PTR_ERR(mp));
1138
1139         rc = mdt_version_get_check_save(info, mp, 0);
1140         if (rc)
1141                 GOTO(put_parent, rc);
1142
1143         /* step 2: find source */
1144         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1145         if (IS_ERR(ms))
1146                 GOTO(put_parent, rc = PTR_ERR(ms));
1147
1148         if (!mdt_object_exists(ms)) {
1149                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1150                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1151                 GOTO(put_source, rc = -ENOENT);
1152         }
1153
1154         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1155
1156         lhp = &info->mti_lh[MDT_LH_PARENT];
1157         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1158         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1159                                    cos_incompat);
1160         if (rc != 0)
1161                 GOTO(put_source, rc);
1162
1163         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1164
1165         lhs = &info->mti_lh[MDT_LH_CHILD];
1166         mdt_lock_reg_init(lhs, LCK_EX);
1167         rc = mdt_reint_object_lock(info, ms, lhs,
1168                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1169                                    cos_incompat);
1170         if (rc != 0)
1171                 GOTO(unlock_parent, rc);
1172
1173         /* step 3: link it */
1174         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1175                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1176
1177         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1178         rc = mdt_version_get_check_save(info, ms, 1);
1179         if (rc)
1180                 GOTO(unlock_source, rc);
1181
1182         /** check target version by name during replay */
1183         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1184                                       &info->mti_tmp_fid1, 2);
1185         if (rc != 0 && rc != -ENOENT)
1186                 GOTO(unlock_source, rc);
1187         /* save version of file name for replay, it must be ENOENT here */
1188         if (!req_is_replay(mdt_info_req(info))) {
1189                 if (rc != -ENOENT) {
1190                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1191                                PNAME(&rr->rr_name));
1192                         GOTO(unlock_source, rc = -EEXIST);
1193                 }
1194                 info->mti_ver[2] = ENOENT_VERSION;
1195                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1196         }
1197
1198         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1199                       mdt_object_child(ms), &rr->rr_name, ma);
1200
1201         if (rc == 0)
1202                 mdt_counter_incr(req, LPROC_MDT_LINK);
1203
1204         EXIT;
1205 unlock_source:
1206         mdt_object_unlock(info, ms, lhs, rc);
1207 unlock_parent:
1208         mdt_object_unlock(info, mp, lhp, rc);
1209 put_source:
1210         mdt_object_put(info->mti_env, ms);
1211 put_parent:
1212         mdt_object_put(info->mti_env, mp);
1213         return rc;
1214 }
1215 /**
1216  * lock the part of the directory according to the hash of the name
1217  * (lh->mlh_pdo_hash) in parallel directory lock.
1218  */
1219 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1220                               struct mdt_lock_handle *lh,
1221                               struct mdt_object *obj, __u64 ibits,
1222                               bool cos_incompat)
1223 {
1224         struct ldlm_res_id *res = &info->mti_res_id;
1225         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1226         union ldlm_policy_data *policy = &info->mti_policy;
1227         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1228         int rc;
1229
1230         /*
1231          * Finish res_id initializing by name hash marking part of
1232          * directory which is taking modification.
1233          */
1234         LASSERT(lh->mlh_pdo_hash != 0);
1235         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1236         memset(policy, 0, sizeof(*policy));
1237         policy->l_inodebits.bits = ibits;
1238         if (cos_incompat &&
1239             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1240                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1241         /*
1242          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1243          * going to be sent to client. If it is - mdt_intent_policy() path will
1244          * fix it up and turn FL_LOCAL flag off.
1245          */
1246         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1247                           policy, res, dlmflags,
1248                           &info->mti_exp->exp_handle.h_cookie);
1249         return rc;
1250 }
1251
1252 /**
1253  * Get BFL lock for rename or migrate process.
1254  **/
1255 static int mdt_rename_lock(struct mdt_thread_info *info,
1256                            struct lustre_handle *lh)
1257 {
1258         int     rc;
1259         ENTRY;
1260
1261         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1262                 struct lu_fid *fid = &info->mti_tmp_fid1;
1263                 struct mdt_object *obj;
1264
1265                 /* XXX, right now, it has to use object API to
1266                  * enqueue lock cross MDT, so it will enqueue
1267                  * rename lock(with LUSTRE_BFL_FID) by root object */
1268                 lu_root_fid(fid);
1269                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1270                 if (IS_ERR(obj))
1271                         RETURN(PTR_ERR(obj));
1272
1273                 rc = mdt_remote_object_lock(info, obj,
1274                                             &LUSTRE_BFL_FID, lh,
1275                                             LCK_EX,
1276                                             MDS_INODELOCK_UPDATE, false);
1277                 mdt_object_put(info->mti_env, obj);
1278         } else {
1279                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1280                 union ldlm_policy_data *policy = &info->mti_policy;
1281                 struct ldlm_res_id *res_id = &info->mti_res_id;
1282                 __u64 flags = 0;
1283
1284                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1285                 memset(policy, 0, sizeof *policy);
1286                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1287                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1288                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1289                                             LDLM_IBITS, policy, LCK_EX, &flags,
1290                                             ldlm_blocking_ast,
1291                                             ldlm_completion_ast, NULL, NULL, 0,
1292                                             LVB_T_NONE,
1293                                             &info->mti_exp->exp_handle.h_cookie,
1294                                             lh);
1295                 RETURN(rc);
1296         }
1297         RETURN(rc);
1298 }
1299
1300 static void mdt_rename_unlock(struct lustre_handle *lh)
1301 {
1302         ENTRY;
1303         LASSERT(lustre_handle_is_used(lh));
1304         /* Cancel the single rename lock right away */
1305         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1306         EXIT;
1307 }
1308
1309 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1310                                                 const struct lu_fid *fid,
1311                                                 int idx)
1312 {
1313         struct mdt_object *dir;
1314         int rc;
1315
1316         ENTRY;
1317
1318         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1319         if (IS_ERR(dir))
1320                 RETURN(dir);
1321
1322         /* check early, the real version will be saved after locking */
1323         rc = mdt_version_get_check(info, dir, idx);
1324         if (rc)
1325                 GOTO(out_put, rc);
1326
1327         RETURN(dir);
1328 out_put:
1329         mdt_object_put(info->mti_env, dir);
1330         return ERR_PTR(rc);
1331 }
1332
1333 /*
1334  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1335  * herein we don't really check it, just do revoke.
1336  */
1337 static int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1338                                          struct mdt_object *pobj,
1339                                          struct mdt_object *obj)
1340 {
1341         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1342         int rc;
1343
1344         mdt_lock_handle_init(lh);
1345         mdt_lock_reg_init(lh, LCK_EX);
1346
1347         if (mdt_object_remote(pobj)) {
1348                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1349                                             &lh->mlh_rreg_lh, LCK_EX,
1350                                             MDS_INODELOCK_LOOKUP, false);
1351         } else {
1352                 struct ldlm_res_id *res = &info->mti_res_id;
1353                 union ldlm_policy_data *policy = &info->mti_policy;
1354                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1355                                  LDLM_FL_COS_INCOMPAT;
1356
1357                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1358                 memset(policy, 0, sizeof(*policy));
1359                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1360                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1361                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1362                                   dlmflags, NULL);
1363         }
1364
1365         if (rc != ELDLM_OK)
1366                 return rc;
1367
1368         /*
1369          * TODO, currently we don't save this lock because there is no place to
1370          * hold this lock handle, but to avoid race we need to save this lock.
1371          */
1372         mdt_object_unlock(info, NULL, lh, 1);
1373
1374         return 0;
1375 }
1376
1377 /*
1378  * operation may takes locks of linkea, or directory stripes, group them in
1379  * different list.
1380  */
1381 struct mdt_sub_lock {
1382         struct mdt_object      *msl_obj;
1383         struct mdt_lock_handle  msl_lh;
1384         struct list_head        msl_linkage;
1385 };
1386
1387 static void mdt_unlock_list(struct mdt_thread_info *info,
1388                             struct list_head *list, int decref)
1389 {
1390         struct mdt_sub_lock *msl;
1391         struct mdt_sub_lock *tmp;
1392
1393         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1394                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1395                 list_del(&msl->msl_linkage);
1396                 OBD_FREE_PTR(msl);
1397         }
1398 }
1399
1400 /*
1401  * lock parents of links, and also check whether total locks don't exceed
1402  * RS_MAX_LOCKS.
1403  *
1404  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1405  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1406  * \retval      -ev negative errno upon error
1407  */
1408 static int mdt_lock_links(struct mdt_thread_info *info,
1409                           struct mdt_object *pobj,
1410                           const struct md_attr *ma,
1411                           struct mdt_object *obj,
1412                           struct list_head *link_locks)
1413 {
1414         struct mdt_device *mdt = info->mti_mdt;
1415         struct lu_buf *buf = &info->mti_big_buf;
1416         struct lu_name *lname = &info->mti_name;
1417         struct linkea_data ldata = { NULL };
1418         bool blocked = false;
1419         int retries = 5;
1420         int local_lnkp_cnt = 0;
1421         int rc;
1422
1423         ENTRY;
1424
1425         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1426                 RETURN(0);
1427
1428         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1429         if (buf->lb_buf == NULL)
1430                 RETURN(-ENOMEM);
1431
1432         ldata.ld_buf = buf;
1433         rc = mdt_links_read(info, obj, &ldata);
1434         if (rc) {
1435                 if (rc == -ENOENT || rc == -ENODATA)
1436                         rc = 0;
1437                 RETURN(rc);
1438         }
1439
1440 repeat:
1441         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1442              linkea_next_entry(&ldata)) {
1443                 struct mdt_object *lnkp;
1444                 struct mdt_sub_lock *msl;
1445                 struct lu_fid fid;
1446                 __u64 ibits;
1447
1448                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1449                                     &fid);
1450
1451                 /* check if it's also linked to parent */
1452                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1453                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1454                                PFID(&fid), PNAME(lname));
1455                         /* in case link is remote object, revoke LOOKUP lock */
1456                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1457                         continue;
1458                 }
1459
1460                 lnkp = NULL;
1461
1462                 /* check if it's linked to a stripe of parent */
1463                 if (ma->ma_valid & MA_LMV) {
1464                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1465                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1466                         int j = 0;
1467
1468                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1469                                 fid_le_to_cpu(stripe_fid,
1470                                               &lmv->lmv_stripe_fids[j]);
1471                                 if (lu_fid_eq(stripe_fid, &fid)) {
1472                                         CDEBUG(D_INFO, "skip stripe "DFID
1473                                                ", reovke "DNAME"\n",
1474                                                PFID(&fid), PNAME(lname));
1475                                         lnkp = mdt_object_find(info->mti_env,
1476                                                                mdt, &fid);
1477                                         if (IS_ERR(lnkp))
1478                                                 GOTO(out, rc = PTR_ERR(lnkp));
1479                                         break;
1480                                 }
1481                         }
1482
1483                         if (lnkp) {
1484                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1485                                                                    obj);
1486                                 mdt_object_put(info->mti_env, lnkp);
1487                                 continue;
1488                         }
1489                 }
1490
1491                 /* Check if it's already locked */
1492                 list_for_each_entry(msl, link_locks, msl_linkage) {
1493                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1494                                 CDEBUG(D_INFO,
1495                                        DFID" was locked, revoke "DNAME"\n",
1496                                        PFID(&fid), PNAME(lname));
1497                                 lnkp = msl->msl_obj;
1498                                 break;
1499                         }
1500                 }
1501
1502                 if (lnkp) {
1503                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1504                         continue;
1505                 }
1506
1507                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1508                        PFID(&fid), PNAME(lname));
1509
1510                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1511                 if (IS_ERR(lnkp)) {
1512                         CWARN("%s: cannot find obj "DFID": %ld\n",
1513                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1514                         continue;
1515                 }
1516
1517                 if (!mdt_object_exists(lnkp)) {
1518                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1519                               PFID(&fid), PNAME(lname));
1520                         mdt_object_put(info->mti_env, lnkp);
1521                         continue;
1522                 }
1523
1524                 if (!mdt_object_remote(lnkp))
1525                         local_lnkp_cnt++;
1526
1527                 OBD_ALLOC_PTR(msl);
1528                 if (msl == NULL)
1529                         GOTO(out, rc = -ENOMEM);
1530
1531                 /*
1532                  * we can't follow parent-child lock order like other MD
1533                  * operations, use lock_try here to avoid deadlock, if the lock
1534                  * cannot be taken, drop all locks taken, revoke the blocked
1535                  * one, and continue processing the remaining entries, and in
1536                  * the end of the loop restart from beginning.
1537                  */
1538                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1539                 ibits = 0;
1540                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1541                                          MDS_INODELOCK_UPDATE, true);
1542                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1543                         blocked = true;
1544
1545                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME" retry %d\n",
1546                                PFID(&fid), PNAME(lname), retries);
1547
1548                         mdt_unlock_list(info, link_locks, 1);
1549
1550                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1551                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1552                                              MDS_INODELOCK_UPDATE);
1553                         if (rc) {
1554                                 mdt_object_put(info->mti_env, lnkp);
1555                                 OBD_FREE_PTR(msl);
1556                                 GOTO(out, rc);
1557                         }
1558
1559                         if (mdt_object_remote(lnkp)) {
1560                                 struct ldlm_lock *lock;
1561
1562                                 /*
1563                                  * for remote object, set lock cb_atomic,
1564                                  * so lock can be released in blocking_ast()
1565                                  * immediately, then the next lock_try will
1566                                  * have better chance of success.
1567                                  */
1568                                 lock = ldlm_handle2lock(
1569                                                 &msl->msl_lh.mlh_rreg_lh);
1570                                 LASSERT(lock != NULL);
1571                                 lock_res_and_lock(lock);
1572                                 ldlm_set_atomic_cb(lock);
1573                                 unlock_res_and_lock(lock);
1574                                 LDLM_LOCK_PUT(lock);
1575                         }
1576
1577                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1578                         OBD_FREE_PTR(msl);
1579                         continue;
1580                 }
1581
1582                 INIT_LIST_HEAD(&msl->msl_linkage);
1583                 msl->msl_obj = lnkp;
1584                 list_add_tail(&msl->msl_linkage, link_locks);
1585
1586                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1587         }
1588
1589         if (blocked) {
1590                 rc = -EBUSY;
1591                 if (--retries > 0) {
1592                         mdt_unlock_list(info, link_locks, rc);
1593                         blocked = false;
1594                         local_lnkp_cnt = 0;
1595                         goto repeat;
1596                 }
1597         }
1598
1599         EXIT;
1600 out:
1601         if (rc)
1602                 mdt_unlock_list(info, link_locks, rc);
1603         else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1604                 /*
1605                  * parent may have 3 local objects: master object and 2 stripes
1606                  * (if it's being migrated too); source may have 2 local
1607                  * objects: master and 1 stripe; target has 1 local object.
1608                  */
1609                 rc = 1;
1610         return rc;
1611 }
1612
1613 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1614                                   struct mdt_object *obj,
1615                                   const struct md_attr *ma,
1616                                   struct list_head *slave_locks)
1617 {
1618         struct mdt_device *mdt = info->mti_mdt;
1619         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1620         struct lu_fid *fid = &info->mti_tmp_fid1;
1621         struct mdt_object *slave;
1622         struct mdt_sub_lock *msl;
1623         int i;
1624         int rc;
1625
1626         ENTRY;
1627
1628         LASSERT(mdt_object_remote(obj));
1629         LASSERT(ma->ma_valid & MA_LMV);
1630         LASSERT(lmv);
1631
1632         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1633                 RETURN(-EINVAL);
1634
1635         if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1636                 RETURN(0);
1637
1638         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1639                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1640
1641                 slave = mdt_object_find(info->mti_env, mdt, fid);
1642                 if (IS_ERR(slave))
1643                         GOTO(out, rc = PTR_ERR(slave));
1644
1645                 OBD_ALLOC_PTR(msl);
1646                 if (!msl) {
1647                         mdt_object_put(info->mti_env, slave);
1648                         GOTO(out, rc = -ENOMEM);
1649                 }
1650
1651                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1652                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1653                                            MDS_INODELOCK_UPDATE, true);
1654                 if (rc) {
1655                         OBD_FREE_PTR(msl);
1656                         mdt_object_put(info->mti_env, slave);
1657                         GOTO(out, rc);
1658                 }
1659
1660                 INIT_LIST_HEAD(&msl->msl_linkage);
1661                 msl->msl_obj = slave;
1662                 list_add_tail(&msl->msl_linkage, slave_locks);
1663
1664         }
1665         EXIT;
1666
1667 out:
1668         if (rc)
1669                 mdt_unlock_list(info, slave_locks, rc);
1670         return rc;
1671 }
1672
1673 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1674                                              struct mdt_object *obj,
1675                                              struct mdt_lock_handle *lh,
1676                                              struct ldlm_enqueue_info *einfo,
1677                                              struct list_head *slave_locks,
1678                                              int decref)
1679 {
1680         if (mdt_object_remote(obj)) {
1681                 mdt_unlock_list(info, slave_locks, decref);
1682                 mdt_object_unlock(info, obj, lh, decref);
1683         } else {
1684                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1685         }
1686 }
1687
1688 /* lock parent and its stripes */
1689 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1690                                    struct mdt_object *obj,
1691                                    const struct md_attr *ma,
1692                                    struct mdt_lock_handle *lh,
1693                                    struct ldlm_enqueue_info *einfo,
1694                                    struct list_head *slave_locks)
1695 {
1696         int rc;
1697
1698         if (mdt_object_remote(obj)) {
1699                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1700                                             &lh->mlh_rreg_lh, LCK_PW,
1701                                             MDS_INODELOCK_UPDATE, false);
1702                 if (rc != ELDLM_OK)
1703                         return rc;
1704
1705                 /*
1706                  * if obj is remote and striped, lock its stripes explicitly
1707                  * because it's not striped in LOD layer on this MDT.
1708                  */
1709                 if (ma->ma_valid & MA_LMV) {
1710                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1711                         if (rc)
1712                                 mdt_object_unlock(info, obj, lh, rc);
1713                 }
1714         } else {
1715                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1716                                             einfo, true);
1717         }
1718
1719         return rc;
1720 }
1721
1722 /*
1723  * in migration, object may be remote, and we need take full lock of it and its
1724  * stripes if it's directory, besides, object may be a remote object on its
1725  * parent, revoke its LOOKUP lock on where its parent is located.
1726  */
1727 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1728                                    struct mdt_object *pobj,
1729                                    struct mdt_object *obj,
1730                                    struct mdt_lock_handle *lh,
1731                                    struct ldlm_enqueue_info *einfo,
1732                                    struct list_head *slave_locks)
1733 {
1734         int rc;
1735
1736         if (mdt_object_remote(obj)) {
1737                 /* don't bother to check if pobj and obj are on the same MDT. */
1738                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1739                 if (rc)
1740                         return rc;
1741
1742                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1743                                             &lh->mlh_rreg_lh, LCK_EX,
1744                                             MDS_INODELOCK_FULL, false);
1745                 if (rc != ELDLM_OK)
1746                         return rc;
1747
1748                 /*
1749                  * if obj is remote and striped, lock its stripes explicitly
1750                  * because it's not striped in LOD layer on this MDT.
1751                  */
1752                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1753                         struct md_attr *ma = &info->mti_attr;
1754
1755                         ma->ma_lmv = info->mti_big_lmm;
1756                         ma->ma_lmv_size = info->mti_big_lmmsize;
1757                         ma->ma_valid = 0;
1758                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1759                         if (rc) {
1760                                 mdt_object_unlock(info, obj, lh, rc);
1761                                 return rc;
1762                         }
1763
1764                         if (ma->ma_valid & MA_LMV) {
1765                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1766                                                             slave_locks);
1767                                 if (rc)
1768                                         mdt_object_unlock(info, obj, lh, rc);
1769                         }
1770                 }
1771         } else {
1772                 if (mdt_object_remote(pobj)) {
1773                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1774                         if (rc)
1775                                 return rc;
1776                 }
1777
1778                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1779                                             einfo, true);
1780         }
1781
1782         return rc;
1783 }
1784
1785 /*
1786  * lookup source by name, if parent is striped directory, we need to find the
1787  * corresponding stripe where source is located, and then lookup there.
1788  *
1789  * besides, if parent is migrating too, and file is already in target stripe,
1790  * this should be a redo of 'lfs migrate' on client side.
1791  */
1792 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1793                               struct mdt_object *pobj,
1794                               const struct md_attr *ma,
1795                               const struct lu_name *lname,
1796                               struct mdt_object **spobj,
1797                               struct mdt_object **sobj)
1798 {
1799         const struct lu_env *env = info->mti_env;
1800         struct lu_fid *fid = &info->mti_tmp_fid1;
1801         struct mdt_object *stripe;
1802         int rc;
1803
1804         if (ma->ma_valid & MA_LMV) {
1805                 /* if parent is striped, lookup on corresponding stripe */
1806                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1807                 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1808                 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1809                 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1810                                     LMV_HASH_FLAG_MIGRATION;
1811
1812                 if (is_migrating) {
1813                         hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1814                         stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1815                 }
1816
1817                 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1818                                               lname->ln_name,
1819                                               lname->ln_namelen);
1820                 if (rc < 0)
1821                         return rc;
1822
1823                 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1824                         rc += le32_to_cpu(lmv->lmv_migrate_offset);
1825
1826                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1827
1828                 stripe = mdt_object_find(env, info->mti_mdt, fid);
1829                 if (IS_ERR(stripe))
1830                         return PTR_ERR(stripe);
1831
1832                 fid_zero(fid);
1833                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1834                                 &info->mti_spec);
1835                 if (rc == -ENOENT && is_migrating) {
1836                         /*
1837                          * if parent is migrating, and lookup child failed on
1838                          * source stripe, lookup again on target stripe, if it
1839                          * exists, it means previous migration was interrupted,
1840                          * and current file was migrated already.
1841                          */
1842                         mdt_object_put(env, stripe);
1843
1844                         hash_type = le32_to_cpu(lmv->lmv_hash_type);
1845                         stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1846
1847                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1848                                                       lname->ln_name,
1849                                                       lname->ln_namelen);
1850                         if (rc < 0)
1851                                 return rc;
1852
1853                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1854
1855                         stripe = mdt_object_find(env, info->mti_mdt, fid);
1856                         if (IS_ERR(stripe))
1857                                 return PTR_ERR(stripe);
1858
1859                         fid_zero(fid);
1860                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1861                                         fid, &info->mti_spec);
1862                         mdt_object_put(env, stripe);
1863                         return rc ?: -EALREADY;
1864                 } else if (rc) {
1865                         mdt_object_put(env, stripe);
1866                         return rc;
1867                 }
1868         } else {
1869                 fid_zero(fid);
1870                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1871                                 &info->mti_spec);
1872                 if (rc)
1873                         return rc;
1874
1875                 stripe = pobj;
1876                 mdt_object_get(env, stripe);
1877         }
1878
1879         *spobj = stripe;
1880
1881         *sobj = mdt_object_find(env, info->mti_mdt, fid);
1882         if (IS_ERR(*sobj)) {
1883                 mdt_object_put(env, stripe);
1884                 rc = PTR_ERR(*sobj);
1885                 *spobj = NULL;
1886                 *sobj = NULL;
1887         }
1888
1889         return rc;
1890 }
1891
1892 /* end lease and close file for regular file */
1893 static int mdd_migrate_close(struct mdt_thread_info *info,
1894                              struct mdt_object *obj)
1895 {
1896         struct close_data *data;
1897         struct mdt_body *repbody;
1898         struct ldlm_lock *lease;
1899         int rc;
1900         int rc2;
1901
1902         rc = -EPROTO;
1903         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1904                                       RCL_CLIENT) ||
1905             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1906                                       RCL_CLIENT))
1907                 goto close;
1908
1909         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1910         if (!data)
1911                 goto close;
1912
1913         rc = -ESTALE;
1914         lease = ldlm_handle2lock(&data->cd_handle);
1915         if (!lease)
1916                 goto close;
1917
1918         /* check if the lease was already canceled */
1919         lock_res_and_lock(lease);
1920         rc = ldlm_is_cancel(lease);
1921         unlock_res_and_lock(lease);
1922
1923         if (rc) {
1924                 rc = -EAGAIN;
1925                 LDLM_DEBUG(lease, DFID" lease broken",
1926                            PFID(mdt_object_fid(obj)));
1927         }
1928
1929         /*
1930          * cancel server side lease, client side counterpart should have been
1931          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1932          */
1933         ldlm_lock_cancel(lease);
1934         ldlm_reprocess_all(lease->l_resource);
1935         LDLM_LOCK_PUT(lease);
1936
1937 close:
1938         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1939         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1940         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1941
1942         return rc ?: rc2;
1943 }
1944
1945 /*
1946  * migrate file in below steps:
1947  *  1. lock parent and its stripes
1948  *  2. lookup source by name
1949  *  3. lock parents of source links if source is not directory
1950  *  4. reject if source is in HSM
1951  *  5. take source open_sem and close file if source is regular file
1952  *  6. lock source and its stripes if it's directory
1953  *  7. lock target so subsequent change to it can trigger COS
1954  *  8. migrate file
1955  *  9. unlock above locks
1956  * 10. sync device if source has links
1957  */
1958 static int mdt_reint_migrate_internal(struct mdt_thread_info *info)
1959 {
1960         const struct lu_env *env = info->mti_env;
1961         struct mdt_device *mdt = info->mti_mdt;
1962         struct mdt_reint_record *rr = &info->mti_rr;
1963         struct md_attr *ma = &info->mti_attr;
1964         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1965         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1966         struct mdt_object *pobj;
1967         struct mdt_object *spobj = NULL;
1968         struct mdt_object *sobj = NULL;
1969         struct mdt_object *tobj;
1970         struct mdt_lock_handle *lhp;
1971         struct mdt_lock_handle *lhs;
1972         struct mdt_lock_handle *lht;
1973         LIST_HEAD(parent_slave_locks);
1974         LIST_HEAD(child_slave_locks);
1975         LIST_HEAD(link_locks);
1976         bool open_sem_locked = false;
1977         bool do_sync = false;
1978         int rc;
1979         ENTRY;
1980
1981         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1982                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1983
1984         /* don't allow migrate . or .. */
1985         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
1986                 RETURN(-EBUSY);
1987
1988         rc = mdt_remote_permission(info);
1989         if (rc)
1990                 RETURN(rc);
1991
1992         /* pobj is master object of parent */
1993         pobj = mdt_object_find_check(info, rr->rr_fid1, 0);
1994         if (IS_ERR(pobj))
1995                 RETURN(PTR_ERR(pobj));
1996
1997         if (unlikely(!info->mti_big_lmm)) {
1998                 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
1999                 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
2000                 if (!info->mti_big_lmm)
2001                         GOTO(put_parent, rc = -ENOMEM);
2002         }
2003
2004         ma->ma_lmv = info->mti_big_lmm;
2005         ma->ma_lmv_size = info->mti_big_lmmsize;
2006         ma->ma_valid = 0;
2007         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2008         if (rc)
2009                 GOTO(put_parent, rc);
2010
2011         /* lock parent object */
2012         lhp = &info->mti_lh[MDT_LH_PARENT];
2013         mdt_lock_reg_init(lhp, LCK_PW);
2014         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2015                                      &parent_slave_locks);
2016         if (rc)
2017                 GOTO(put_parent, rc);
2018
2019         /*
2020          * spobj is the corresponding stripe against name if pobj is striped
2021          * directory, which is the real parent, and no need to lock, because
2022          * we've taken full lock of pobj.
2023          */
2024         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2025         if (rc)
2026                 GOTO(unlock_parent, rc);
2027
2028         /* lock parents of source links, and revoke LOOKUP lock of links */
2029         rc = mdt_lock_links(info, pobj, ma, sobj, &link_locks);
2030         if (rc < 0)
2031                 GOTO(put_source, rc);
2032
2033         /*
2034          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2035          * with one request, if total lock count exceeds this limit, we will
2036          * drop all locks after migration, and synchronous device in the end.
2037          */
2038         do_sync = rc;
2039
2040         /* if migration HSM is allowed */
2041         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2042                 ma->ma_need = MA_HSM;
2043                 ma->ma_valid = 0;
2044                 rc = mdt_attr_get_complex(info, sobj, ma);
2045                 if (rc)
2046                         GOTO(unlock_links, rc);
2047
2048                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2049                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2050         }
2051
2052         /* end lease and close file for regular file */
2053         if (info->mti_spec.sp_migrate_close) {
2054                 /* try to hold open_sem so that nobody else can open the file */
2055                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2056                         /* close anyway */
2057                         mdd_migrate_close(info, sobj);
2058                         GOTO(unlock_links, rc = -EBUSY);
2059                 } else {
2060                         open_sem_locked = true;
2061                         rc = mdd_migrate_close(info, sobj);
2062                         if (rc)
2063                                 GOTO(unlock_open_sem, rc);
2064                 }
2065         }
2066
2067         /* lock source */
2068         lhs = &info->mti_lh[MDT_LH_OLD];
2069         mdt_lock_reg_init(lhs, LCK_EX);
2070         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2071                                      &child_slave_locks);
2072         if (rc)
2073                 GOTO(unlock_open_sem, rc);
2074
2075         /* lock target */
2076         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2077         if (IS_ERR(tobj))
2078                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2079
2080         lht = &info->mti_lh[MDT_LH_NEW];
2081         mdt_lock_reg_init(lht, LCK_EX);
2082         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2083         if (rc)
2084                 GOTO(put_target, rc);
2085
2086         /* Don't do lookup sanity check. We know name doesn't exist. */
2087         info->mti_spec.sp_cr_lookup = 0;
2088         info->mti_spec.sp_feat = &dt_directory_features;
2089
2090         rc = mdo_migrate(env, mdt_object_child(pobj),
2091                          mdt_object_child(sobj), &rr->rr_name,
2092                          mdt_object_child(tobj), &info->mti_spec, ma);
2093         EXIT;
2094
2095         mdt_object_unlock(info, tobj, lht, rc);
2096 put_target:
2097         mdt_object_put(env, tobj);
2098 unlock_source:
2099         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2100                                   &child_slave_locks, rc);
2101 unlock_open_sem:
2102         if (open_sem_locked)
2103                 up_write(&sobj->mot_open_sem);
2104 unlock_links:
2105         mdt_unlock_list(info, &link_locks, rc);
2106 put_source:
2107         mdt_object_put(env, sobj);
2108         mdt_object_put(env, spobj);
2109 unlock_parent:
2110         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2111                                   &parent_slave_locks, rc);
2112 put_parent:
2113         mdt_object_put(env, pobj);
2114
2115         if (!rc && do_sync)
2116                 mdt_device_sync(env, mdt);
2117
2118         return rc;
2119 }
2120
2121 static int mdt_object_lock_save(struct mdt_thread_info *info,
2122                                 struct mdt_object *dir,
2123                                 struct mdt_lock_handle *lh,
2124                                 int idx, bool cos_incompat)
2125 {
2126         int rc;
2127
2128         /* we lock the target dir if it is local */
2129         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2130                                    cos_incompat);
2131         if (rc != 0)
2132                 return rc;
2133
2134         /* get and save correct version after locking */
2135         mdt_version_get_save(info, dir, idx);
2136         return 0;
2137 }
2138
2139 /*
2140  * determine lock order of sobj and tobj
2141  *
2142  * there are two situations we need to lock tobj before sobj:
2143  * 1. sobj is child of tobj
2144  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2145  *    larger than that of tobj
2146  *
2147  * \retval      1 lock tobj before sobj
2148  * \retval      0 lock sobj before tobj
2149  * \retval      -ev negative errno upon error
2150  */
2151 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2152                                            struct mdt_object *sobj,
2153                                            struct mdt_object *tobj)
2154 {
2155         struct md_attr *ma = &info->mti_attr;
2156         struct lu_fid *spfid = &info->mti_tmp_fid1;
2157         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2158         struct lmv_mds_md_v1 *lmv;
2159         __u32 sindex;
2160         __u32 tindex;
2161         int rc;
2162
2163         /* sobj and tobj are the same */
2164         if (sobj == tobj)
2165                 return 0;
2166
2167         if (fid_is_root(mdt_object_fid(sobj)))
2168                 return 0;
2169
2170         if (fid_is_root(mdt_object_fid(tobj)))
2171                 return 1;
2172
2173         /* check whether sobj is child of tobj */
2174         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2175                            mdt_object_fid(tobj));
2176         if (rc < 0)
2177                 return rc;
2178
2179         if (rc == 1)
2180                 return 1;
2181
2182         /* check whether sobj and tobj are children of the same parent */
2183         rc = mdt_attr_get_pfid(info, sobj, spfid);
2184         if (rc)
2185                 return rc;
2186
2187         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2188         if (rc)
2189                 return rc;
2190
2191         if (!lu_fid_eq(spfid, tpfid))
2192                 return 0;
2193
2194         /* check whether sobj and tobj are sibling stripes */
2195         ma->ma_need = MA_LMV;
2196         ma->ma_valid = 0;
2197         ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2198         ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2199         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2200         if (rc)
2201                 return rc;
2202
2203         if (!(ma->ma_valid & MA_LMV))
2204                 return 0;
2205
2206         lmv = &ma->ma_lmv->lmv_md_v1;
2207         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2208                 return 0;
2209         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2210
2211         ma->ma_valid = 0;
2212         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2213         if (rc)
2214                 return rc;
2215
2216         if (!(ma->ma_valid & MA_LMV))
2217                 return -ENODATA;
2218
2219         lmv = &ma->ma_lmv->lmv_md_v1;
2220         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2221                 return -EINVAL;
2222         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2223
2224         /* check stripe index of sobj and tobj */
2225         if (sindex == tindex)
2226                 return -EINVAL;
2227
2228         return sindex < tindex ? 0 : 1;
2229 }
2230
2231 /*
2232  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2233  * 2 - srcdir child; 3 - tgtdir child.
2234  * Update on disk version of srcdir child.
2235  */
2236 /**
2237  * For DNE phase I, only these renames are allowed
2238  *      mv src_p/src_c tgt_p/tgt_c
2239  * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
2240  * 2. src_p and tgt_p are same directory, and tgt_c does not
2241  *    exists. In this case, all of modification will happen
2242  *    in the MDT where ithesource parent is, only one remote
2243  *    update is needed, i.e. set c_time/m_time on the child.
2244  *    And tgt_c will be still in the same MDT as the original
2245  *    src_c.
2246  */
2247 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
2248                                      struct mdt_lock_handle *lhc)
2249 {
2250         struct mdt_reint_record *rr = &info->mti_rr;
2251         struct md_attr *ma = &info->mti_attr;
2252         struct ptlrpc_request *req = mdt_info_req(info);
2253         struct mdt_object *msrcdir = NULL;
2254         struct mdt_object *mtgtdir = NULL;
2255         struct mdt_object *mold;
2256         struct mdt_object *mnew = NULL;
2257         struct mdt_lock_handle *lh_srcdirp;
2258         struct mdt_lock_handle *lh_tgtdirp;
2259         struct mdt_lock_handle *lh_oldp = NULL;
2260         struct mdt_lock_handle *lh_newp = NULL;
2261         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2262         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2263         __u64 lock_ibits;
2264         bool reverse = false;
2265         bool cos_incompat, discard = false;
2266         int rc;
2267         ENTRY;
2268
2269         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2270                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2271                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2272
2273         /* find both parents. */
2274         msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
2275         if (IS_ERR(msrcdir))
2276                 RETURN(PTR_ERR(msrcdir));
2277
2278         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2279
2280         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2281                 mtgtdir = msrcdir;
2282                 mdt_object_get(info->mti_env, mtgtdir);
2283         } else {
2284                 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
2285                 if (IS_ERR(mtgtdir))
2286                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2287         }
2288
2289         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2290         if (rc < 0)
2291                 GOTO(out_put_tgtdir, rc);
2292
2293         reverse = rc;
2294
2295         /* source needs to be looked up after locking source parent, otherwise
2296          * this rename may race with unlink source, and cause rename hang, see
2297          * sanityn.sh 55b, so check parents first, if later we found source is
2298          * remote, relock parents. */
2299         cos_incompat = (mdt_object_remote(msrcdir) ||
2300                         mdt_object_remote(mtgtdir));
2301
2302         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2303
2304         /* lock parents in the proper order. */
2305         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2306         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2307
2308 relock:
2309         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2310         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2311
2312         if (reverse) {
2313                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2314                                           cos_incompat);
2315                 if (rc)
2316                         GOTO(out_put_tgtdir, rc);
2317
2318                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2319
2320                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2321                                           cos_incompat);
2322                 if (rc != 0) {
2323                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2324                         GOTO(out_put_tgtdir, rc);
2325                 }
2326         } else {
2327                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2328                                           cos_incompat);
2329                 if (rc)
2330                         GOTO(out_put_tgtdir, rc);
2331
2332                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2333
2334                 if (mtgtdir != msrcdir) {
2335                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2336                                                   cos_incompat);
2337                 } else if (lh_srcdirp->mlh_pdo_hash !=
2338                            lh_tgtdirp->mlh_pdo_hash) {
2339                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2340                                                 MDS_INODELOCK_UPDATE,
2341                                                 cos_incompat);
2342                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2343                 }
2344                 if (rc != 0) {
2345                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2346                         GOTO(out_put_tgtdir, rc);
2347                 }
2348         }
2349
2350         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2351         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2352
2353         /* find mold object. */
2354         fid_zero(old_fid);
2355         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2356         if (rc != 0)
2357                 GOTO(out_unlock_parents, rc);
2358
2359         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2360                 GOTO(out_unlock_parents, rc = -EINVAL);
2361
2362         if (!fid_is_md_operative(old_fid))
2363                 GOTO(out_unlock_parents, rc = -EPERM);
2364
2365         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2366         if (IS_ERR(mold))
2367                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2368
2369         if (!mdt_object_exists(mold)) {
2370                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2371                                 &mold->mot_obj,
2372                                 "object does not exist");
2373                 GOTO(out_put_old, rc = -ENOENT);
2374         }
2375
2376         /* Check if @mtgtdir is subdir of @mold, before locking child
2377          * to avoid reverse locking. */
2378         if (mtgtdir != msrcdir) {
2379                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2380                                    old_fid);
2381                 if (rc) {
2382                         if (rc == 1)
2383                                 rc = -EINVAL;
2384                         GOTO(out_put_old, rc);
2385                 }
2386         }
2387
2388         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2389         /* save version after locking */
2390         mdt_version_get_save(info, mold, 2);
2391
2392         if (!cos_incompat && mdt_object_remote(mold)) {
2393                 cos_incompat = true;
2394                 mdt_object_put(info->mti_env, mold);
2395                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2396                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2397                 goto relock;
2398         }
2399
2400         /* find mnew object:
2401          * mnew target object may not exist now
2402          * lookup with version checking */
2403         fid_zero(new_fid);
2404         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2405                                       3);
2406         if (rc == 0) {
2407                 /* the new_fid should have been filled at this moment */
2408                 if (lu_fid_eq(old_fid, new_fid))
2409                         GOTO(out_put_old, rc);
2410
2411                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2412                     lu_fid_eq(new_fid, rr->rr_fid2))
2413                         GOTO(out_put_old, rc = -EINVAL);
2414
2415                 if (!fid_is_md_operative(new_fid))
2416                         GOTO(out_put_old, rc = -EPERM);
2417
2418                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2419                 if (IS_ERR(mnew))
2420                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2421
2422                 if (!mdt_object_exists(mnew)) {
2423                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2424                                         &mnew->mot_obj,
2425                                         "object does not exist");
2426                         GOTO(out_put_new, rc = -ENOENT);
2427                 }
2428
2429                 if (mdt_object_remote(mnew)) {
2430                         struct mdt_body  *repbody;
2431
2432                         /* Always send rename req to the target child MDT */
2433                         repbody = req_capsule_server_get(info->mti_pill,
2434                                                          &RMF_MDT_BODY);
2435                         LASSERT(repbody != NULL);
2436                         repbody->mbo_fid1 = *new_fid;
2437                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2438                         GOTO(out_put_new, rc = -EXDEV);
2439                 }
2440                 /* Before locking the target dir, check we do not replace
2441                  * a dir with a non-dir, otherwise it may deadlock with
2442                  * link op which tries to create a link in this dir
2443                  * back to this non-dir. */
2444                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2445                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2446                         GOTO(out_put_new, rc = -EISDIR);
2447
2448                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2449                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2450                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2451                 if (mdt_object_remote(msrcdir)) {
2452                         /* Enqueue lookup lock from the parent MDT */
2453                         rc = mdt_remote_object_lock(info, msrcdir,
2454                                                     mdt_object_fid(mold),
2455                                                     &lh_oldp->mlh_rreg_lh,
2456                                                     lh_oldp->mlh_rreg_mode,
2457                                                     MDS_INODELOCK_LOOKUP,
2458                                                     false);
2459                         if (rc != ELDLM_OK)
2460                                 GOTO(out_put_new, rc);
2461
2462                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2463                 }
2464
2465                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2466                                            cos_incompat);
2467                 if (rc != 0)
2468                         GOTO(out_unlock_old, rc);
2469
2470                 /* Check if @msrcdir is subdir of @mnew, before locking child
2471                  * to avoid reverse locking. */
2472                 if (mtgtdir != msrcdir) {
2473                         rc = mdo_is_subdir(info->mti_env,
2474                                            mdt_object_child(msrcdir), new_fid);
2475                         if (rc) {
2476                                 if (rc == 1)
2477                                         rc = -EINVAL;
2478                                 GOTO(out_unlock_old, rc);
2479                         }
2480                 }
2481
2482                 /* We used to acquire MDS_INODELOCK_FULL here but we
2483                  * can't do this now because a running HSM restore on
2484                  * the rename onto victim will hold the layout
2485                  * lock. See LU-4002. */
2486
2487                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2488                 mdt_lock_reg_init(lh_newp, LCK_EX);
2489                 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2490                                            MDS_INODELOCK_LOOKUP |
2491                                            MDS_INODELOCK_UPDATE,
2492                                            cos_incompat);
2493                 if (rc != 0)
2494                         GOTO(out_unlock_old, rc);
2495
2496                 /* get and save version after locking */
2497                 mdt_version_get_save(info, mnew, 3);
2498         } else if (rc != -EREMOTE && rc != -ENOENT) {
2499                 GOTO(out_put_old, rc);
2500         } else {
2501                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2502                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2503                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2504                 if (mdt_object_remote(msrcdir)) {
2505                         /* Enqueue lookup lock from the parent MDT */
2506                         rc = mdt_remote_object_lock(info, msrcdir,
2507                                                     mdt_object_fid(mold),
2508                                                     &lh_oldp->mlh_rreg_lh,
2509                                                     lh_oldp->mlh_rreg_mode,
2510                                                     MDS_INODELOCK_LOOKUP,
2511                                                     false);
2512                         if (rc != ELDLM_OK)
2513                                 GOTO(out_put_old, rc);
2514
2515                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2516                 }
2517
2518                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2519                                            cos_incompat);
2520                 if (rc != 0)
2521                         GOTO(out_unlock_old, rc);
2522
2523                 mdt_enoent_version_save(info, 3);
2524         }
2525
2526         /* step 5: rename it */
2527         mdt_reint_init_ma(info, ma);
2528
2529         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2530                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2531
2532         if (mnew != NULL)
2533                 mutex_lock(&mnew->mot_lov_mutex);
2534
2535         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2536                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2537                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2538                         &rr->rr_tgt_name, ma);
2539
2540         if (mnew != NULL)
2541                 mutex_unlock(&mnew->mot_lov_mutex);
2542
2543         /* handle last link of tgt object */
2544         if (rc == 0) {
2545                 mdt_counter_incr(req, LPROC_MDT_RENAME);
2546                 if (mnew) {
2547                         mdt_handle_last_unlink(info, mnew, ma);
2548                         discard = mdt_dom_check_for_discard(info, mnew);
2549                 }
2550
2551                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2552                                          msrcdir, mtgtdir);
2553         }
2554
2555         EXIT;
2556         if (mnew != NULL)
2557                 mdt_object_unlock(info, mnew, lh_newp, rc);
2558 out_unlock_old:
2559         mdt_object_unlock(info, mold, lh_oldp, rc);
2560 out_put_new:
2561         if (mnew != NULL)
2562                 mdt_object_put(info->mti_env, mnew);
2563 out_put_old:
2564         mdt_object_put(info->mti_env, mold);
2565 out_unlock_parents:
2566         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2567         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2568 out_put_tgtdir:
2569         mdt_object_put(info->mti_env, mtgtdir);
2570 out_put_srcdir:
2571         mdt_object_put(info->mti_env, msrcdir);
2572
2573         /* If 'discard' is set then new_fid must exits.
2574          * DOM data discard need neither object nor lock,
2575          * so do this at the end.
2576          */
2577         if (discard)
2578                 mdt_dom_discard_data(info, new_fid);
2579
2580         return rc;
2581 }
2582
2583 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2584                                        struct mdt_lock_handle *lhc, bool rename)
2585 {
2586         struct mdt_reint_record *rr = &info->mti_rr;
2587         struct ptlrpc_request   *req = mdt_info_req(info);
2588         struct lustre_handle    rename_lh = { 0 };
2589         int                     rc;
2590         ENTRY;
2591
2592         if (info->mti_dlm_req)
2593                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2594
2595         if (!fid_is_md_operative(rr->rr_fid1) ||
2596             !fid_is_md_operative(rr->rr_fid2))
2597                 RETURN(-EPERM);
2598
2599         /* Note: do not enqueue rename lock for replay request, because
2600          * if other MDT holds rename lock, but being blocked to wait for
2601          * this MDT to finish its recovery, and the failover MDT can not
2602          * get rename lock, which will cause deadlock. */
2603         if (!req_is_replay(req)) {
2604                 rc = mdt_rename_lock(info, &rename_lh);
2605                 if (rc != 0) {
2606                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2607                                mdt_obd_name(info->mti_mdt), rc);
2608                         RETURN(rc);
2609                 }
2610         }
2611
2612         if (rename)
2613                 rc = mdt_reint_rename_internal(info, lhc);
2614         else
2615                 rc = mdt_reint_migrate_internal(info);
2616
2617         if (lustre_handle_is_used(&rename_lh))
2618                 mdt_rename_unlock(&rename_lh);
2619
2620         RETURN(rc);
2621 }
2622
2623 static int mdt_reint_rename(struct mdt_thread_info *info,
2624                             struct mdt_lock_handle *lhc)
2625 {
2626         return mdt_reint_rename_or_migrate(info, lhc, true);
2627 }
2628
2629 static int mdt_reint_migrate(struct mdt_thread_info *info,
2630                             struct mdt_lock_handle *lhc)
2631 {
2632         return mdt_reint_rename_or_migrate(info, lhc, false);
2633 }
2634
2635 static int mdt_reint_resync(struct mdt_thread_info *info,
2636                             struct mdt_lock_handle *lhc)
2637 {
2638         struct mdt_reint_record *rr = &info->mti_rr;
2639         struct ptlrpc_request   *req = mdt_info_req(info);
2640         struct md_attr          *ma = &info->mti_attr;
2641         struct mdt_object       *mo;
2642         struct ldlm_lock        *lease;
2643         struct mdt_body         *repbody;
2644         struct md_layout_change  layout = { 0 };
2645         bool                     lease_broken;
2646         int                      rc, rc2;
2647         ENTRY;
2648
2649         DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2650
2651         if (info->mti_dlm_req)
2652                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2653
2654         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2655         if (IS_ERR(mo))
2656                 GOTO(out, rc = PTR_ERR(mo));
2657
2658         if (!mdt_object_exists(mo))
2659                 GOTO(out_obj, rc = -ENOENT);
2660
2661         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2662                 GOTO(out_obj, rc = -EINVAL);
2663
2664         if (mdt_object_remote(mo))
2665                 GOTO(out_obj, rc = -EREMOTE);
2666
2667         lease = ldlm_handle2lock(rr->rr_lease_handle);
2668         if (lease == NULL)
2669                 GOTO(out_obj, rc = -ESTALE);
2670
2671         /* It's really necessary to grab open_sem and check if the lease lock
2672          * has been lost. There would exist a concurrent writer coming in and
2673          * generating some dirty data in memory cache, the writeback would fail
2674          * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2675         if (!down_write_trylock(&mo->mot_open_sem))
2676                 GOTO(out_put_lease, rc = -EBUSY);
2677
2678         lock_res_and_lock(lease);
2679         lease_broken = ldlm_is_cancel(lease);
2680         unlock_res_and_lock(lease);
2681         if (lease_broken)
2682                 GOTO(out_unlock, rc = -EBUSY);
2683
2684         /* the file has yet opened by anyone else after we took the lease. */
2685         layout.mlc_opc = MD_LAYOUT_RESYNC;
2686         rc = mdt_layout_change(info, mo, &layout);
2687         if (rc)
2688                 GOTO(out_unlock, rc);
2689
2690         ma->ma_need = MA_INODE;
2691         ma->ma_valid = 0;
2692         rc = mdt_attr_get_complex(info, mo, ma);
2693         if (rc != 0)
2694                 GOTO(out_unlock, rc);
2695
2696         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2697         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2698
2699         EXIT;
2700 out_unlock:
2701         up_write(&mo->mot_open_sem);
2702 out_put_lease:
2703         LDLM_LOCK_PUT(lease);
2704 out_obj:
2705         mdt_object_put(info->mti_env, mo);
2706 out:
2707         mdt_client_compatibility(info);
2708         rc2 = mdt_fix_reply(info);
2709         if (rc == 0)
2710                 rc = rc2;
2711         return rc;
2712 }
2713
2714 struct mdt_reinter {
2715         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2716         enum lprocfs_extra_opc mr_extra_opc;
2717 };
2718
2719 static const struct mdt_reinter mdt_reinters[] = {
2720         [REINT_SETATTR] = {
2721                 .mr_handler = &mdt_reint_setattr,
2722                 .mr_extra_opc = MDS_REINT_SETATTR,
2723         },
2724         [REINT_CREATE] = {
2725                 .mr_handler = &mdt_reint_create,
2726                 .mr_extra_opc = MDS_REINT_CREATE,
2727         },
2728         [REINT_LINK] = {
2729                 .mr_handler = &mdt_reint_link,
2730                 .mr_extra_opc = MDS_REINT_LINK,
2731         },
2732         [REINT_UNLINK] = {
2733                 .mr_handler = &mdt_reint_unlink,
2734                 .mr_extra_opc = MDS_REINT_UNLINK,
2735         },
2736         [REINT_RENAME] = {
2737                 .mr_handler = &mdt_reint_rename,
2738                 .mr_extra_opc = MDS_REINT_RENAME,
2739         },
2740         [REINT_OPEN] = {
2741                 .mr_handler = &mdt_reint_open,
2742                 .mr_extra_opc = MDS_REINT_OPEN,
2743         },
2744         [REINT_SETXATTR] = {
2745                 .mr_handler = &mdt_reint_setxattr,
2746                 .mr_extra_opc = MDS_REINT_SETXATTR,
2747         },
2748         [REINT_RMENTRY] = {
2749                 .mr_handler = &mdt_reint_unlink,
2750                 .mr_extra_opc = MDS_REINT_UNLINK,
2751         },
2752         [REINT_MIGRATE] = {
2753                 .mr_handler = &mdt_reint_migrate,
2754                 .mr_extra_opc = MDS_REINT_RENAME,
2755         },
2756         [REINT_RESYNC] = {
2757                 .mr_handler = &mdt_reint_resync,
2758                 .mr_extra_opc = MDS_REINT_RESYNC,
2759         },
2760 };
2761
2762 int mdt_reint_rec(struct mdt_thread_info *info,
2763                   struct mdt_lock_handle *lhc)
2764 {
2765         const struct mdt_reinter *mr;
2766         int rc;
2767         ENTRY;
2768
2769         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2770                 RETURN(-EPROTO);
2771
2772         mr = &mdt_reinters[info->mti_rr.rr_opcode];
2773         if (mr->mr_handler == NULL)
2774                 RETURN(-EPROTO);
2775
2776         rc = (*mr->mr_handler)(info, lhc);
2777
2778         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2779                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
2780
2781         RETURN(rc);
2782 }