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(ns, &lh->mlh_reg_lh, lh->mlh_reg_mode, policy,
1247                           res, dlmflags, &info->mti_exp->exp_handle.h_cookie);
1248         return rc;
1249 }
1250
1251 /**
1252  * Get BFL lock for rename or migrate process.
1253  **/
1254 static int mdt_rename_lock(struct mdt_thread_info *info,
1255                            struct lustre_handle *lh)
1256 {
1257         int     rc;
1258         ENTRY;
1259
1260         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1261                 struct lu_fid *fid = &info->mti_tmp_fid1;
1262                 struct mdt_object *obj;
1263
1264                 /* XXX, right now, it has to use object API to
1265                  * enqueue lock cross MDT, so it will enqueue
1266                  * rename lock(with LUSTRE_BFL_FID) by root object */
1267                 lu_root_fid(fid);
1268                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1269                 if (IS_ERR(obj))
1270                         RETURN(PTR_ERR(obj));
1271
1272                 rc = mdt_remote_object_lock(info, obj,
1273                                             &LUSTRE_BFL_FID, lh,
1274                                             LCK_EX,
1275                                             MDS_INODELOCK_UPDATE, false);
1276                 mdt_object_put(info->mti_env, obj);
1277         } else {
1278                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1279                 union ldlm_policy_data *policy = &info->mti_policy;
1280                 struct ldlm_res_id *res_id = &info->mti_res_id;
1281                 __u64 flags = 0;
1282
1283                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1284                 memset(policy, 0, sizeof *policy);
1285                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1286                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1287                 rc = ldlm_cli_enqueue_local(ns, res_id, LDLM_IBITS, policy,
1288                                            LCK_EX, &flags, ldlm_blocking_ast,
1289                                            ldlm_completion_ast, NULL, NULL, 0,
1290                                            LVB_T_NONE,
1291                                            &info->mti_exp->exp_handle.h_cookie,
1292                                            lh);
1293                 RETURN(rc);
1294         }
1295         RETURN(rc);
1296 }
1297
1298 static void mdt_rename_unlock(struct lustre_handle *lh)
1299 {
1300         ENTRY;
1301         LASSERT(lustre_handle_is_used(lh));
1302         /* Cancel the single rename lock right away */
1303         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1304         EXIT;
1305 }
1306
1307 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1308                                                 const struct lu_fid *fid,
1309                                                 int idx)
1310 {
1311         struct mdt_object *dir;
1312         int rc;
1313
1314         ENTRY;
1315
1316         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1317         if (IS_ERR(dir))
1318                 RETURN(dir);
1319
1320         /* check early, the real version will be saved after locking */
1321         rc = mdt_version_get_check(info, dir, idx);
1322         if (rc)
1323                 GOTO(out_put, rc);
1324
1325         RETURN(dir);
1326 out_put:
1327         mdt_object_put(info->mti_env, dir);
1328         return ERR_PTR(rc);
1329 }
1330
1331 /*
1332  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1333  * herein we don't really check it, just do revoke.
1334  */
1335 static int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1336                                          struct mdt_object *pobj,
1337                                          struct mdt_object *obj)
1338 {
1339         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1340         int rc;
1341
1342         mdt_lock_handle_init(lh);
1343         mdt_lock_reg_init(lh, LCK_EX);
1344
1345         if (mdt_object_remote(pobj)) {
1346                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1347                                             &lh->mlh_rreg_lh, LCK_EX,
1348                                             MDS_INODELOCK_LOOKUP, false);
1349         } else {
1350                 struct ldlm_res_id *res = &info->mti_res_id;
1351                 union ldlm_policy_data *policy = &info->mti_policy;
1352                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1353                                  LDLM_FL_COS_INCOMPAT;
1354
1355                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1356                 memset(policy, 0, sizeof(*policy));
1357                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1358                 rc = mdt_fid_lock(info->mti_mdt->mdt_namespace, &lh->mlh_reg_lh,
1359                                   LCK_EX, policy, res, dlmflags, NULL);
1360         }
1361
1362         if (rc != ELDLM_OK)
1363                 return rc;
1364
1365         /*
1366          * TODO, currently we don't save this lock because there is no place to
1367          * hold this lock handle, but to avoid race we need to save this lock.
1368          */
1369         mdt_object_unlock(info, NULL, lh, 1);
1370
1371         return 0;
1372 }
1373
1374 /*
1375  * operation may takes locks of linkea, or directory stripes, group them in
1376  * different list.
1377  */
1378 struct mdt_sub_lock {
1379         struct mdt_object      *msl_obj;
1380         struct mdt_lock_handle  msl_lh;
1381         struct list_head        msl_linkage;
1382 };
1383
1384 static void mdt_unlock_list(struct mdt_thread_info *info,
1385                             struct list_head *list, int decref)
1386 {
1387         struct mdt_sub_lock *msl;
1388         struct mdt_sub_lock *tmp;
1389
1390         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1391                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1392                 list_del(&msl->msl_linkage);
1393                 OBD_FREE_PTR(msl);
1394         }
1395 }
1396
1397 /*
1398  * lock parents of links, and also check whether total locks don't exceed
1399  * RS_MAX_LOCKS.
1400  *
1401  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1402  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1403  * \retval      -ev negative errno upon error
1404  */
1405 static int mdt_lock_links(struct mdt_thread_info *info,
1406                           struct mdt_object *pobj,
1407                           const struct md_attr *ma,
1408                           struct mdt_object *obj,
1409                           struct list_head *link_locks)
1410 {
1411         struct mdt_device *mdt = info->mti_mdt;
1412         struct lu_buf *buf = &info->mti_big_buf;
1413         struct lu_name *lname = &info->mti_name;
1414         struct linkea_data ldata = { NULL };
1415         bool blocked = false;
1416         int retries = 5;
1417         int local_lnkp_cnt = 0;
1418         int rc;
1419
1420         ENTRY;
1421
1422         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1423                 RETURN(0);
1424
1425         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1426         if (buf->lb_buf == NULL)
1427                 RETURN(-ENOMEM);
1428
1429         ldata.ld_buf = buf;
1430         rc = mdt_links_read(info, obj, &ldata);
1431         if (rc) {
1432                 if (rc == -ENOENT || rc == -ENODATA)
1433                         rc = 0;
1434                 RETURN(rc);
1435         }
1436
1437 repeat:
1438         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1439              linkea_next_entry(&ldata)) {
1440                 struct mdt_object *lnkp;
1441                 struct mdt_sub_lock *msl;
1442                 struct lu_fid fid;
1443                 __u64 ibits;
1444
1445                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1446                                     &fid);
1447
1448                 /* check if it's also linked to parent */
1449                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1450                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1451                                PFID(&fid), PNAME(lname));
1452                         /* in case link is remote object, revoke LOOKUP lock */
1453                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1454                         continue;
1455                 }
1456
1457                 lnkp = NULL;
1458
1459                 /* check if it's linked to a stripe of parent */
1460                 if (ma->ma_valid & MA_LMV) {
1461                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1462                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1463                         int j = 0;
1464
1465                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1466                                 fid_le_to_cpu(stripe_fid,
1467                                               &lmv->lmv_stripe_fids[j]);
1468                                 if (lu_fid_eq(stripe_fid, &fid)) {
1469                                         CDEBUG(D_INFO, "skip stripe "DFID
1470                                                ", reovke "DNAME"\n",
1471                                                PFID(&fid), PNAME(lname));
1472                                         lnkp = mdt_object_find(info->mti_env,
1473                                                                mdt, &fid);
1474                                         if (IS_ERR(lnkp))
1475                                                 GOTO(out, rc = PTR_ERR(lnkp));
1476                                         break;
1477                                 }
1478                         }
1479
1480                         if (lnkp) {
1481                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1482                                                                    obj);
1483                                 mdt_object_put(info->mti_env, lnkp);
1484                                 continue;
1485                         }
1486                 }
1487
1488                 /* Check if it's already locked */
1489                 list_for_each_entry(msl, link_locks, msl_linkage) {
1490                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1491                                 CDEBUG(D_INFO,
1492                                        DFID" was locked, revoke "DNAME"\n",
1493                                        PFID(&fid), PNAME(lname));
1494                                 lnkp = msl->msl_obj;
1495                                 break;
1496                         }
1497                 }
1498
1499                 if (lnkp) {
1500                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1501                         continue;
1502                 }
1503
1504                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1505                        PFID(&fid), PNAME(lname));
1506
1507                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1508                 if (IS_ERR(lnkp)) {
1509                         CWARN("%s: cannot find obj "DFID": %ld\n",
1510                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1511                         continue;
1512                 }
1513
1514                 if (!mdt_object_exists(lnkp)) {
1515                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1516                               PFID(&fid), PNAME(lname));
1517                         mdt_object_put(info->mti_env, lnkp);
1518                         continue;
1519                 }
1520
1521                 if (!mdt_object_remote(lnkp))
1522                         local_lnkp_cnt++;
1523
1524                 OBD_ALLOC_PTR(msl);
1525                 if (msl == NULL)
1526                         GOTO(out, rc = -ENOMEM);
1527
1528                 /*
1529                  * we can't follow parent-child lock order like other MD
1530                  * operations, use lock_try here to avoid deadlock, if the lock
1531                  * cannot be taken, drop all locks taken, revoke the blocked
1532                  * one, and continue processing the remaining entries, and in
1533                  * the end of the loop restart from beginning.
1534                  */
1535                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1536                 ibits = 0;
1537                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1538                                          MDS_INODELOCK_UPDATE, true);
1539                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1540                         blocked = true;
1541
1542                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME" retry %d\n",
1543                                PFID(&fid), PNAME(lname), retries);
1544
1545                         mdt_unlock_list(info, link_locks, 1);
1546
1547                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1548                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1549                                              MDS_INODELOCK_UPDATE);
1550                         if (rc) {
1551                                 mdt_object_put(info->mti_env, lnkp);
1552                                 OBD_FREE_PTR(msl);
1553                                 GOTO(out, rc);
1554                         }
1555
1556                         if (mdt_object_remote(lnkp)) {
1557                                 struct ldlm_lock *lock;
1558
1559                                 /*
1560                                  * for remote object, set lock cb_atomic,
1561                                  * so lock can be released in blocking_ast()
1562                                  * immediately, then the next lock_try will
1563                                  * have better chance of success.
1564                                  */
1565                                 lock = ldlm_handle2lock(
1566                                                 &msl->msl_lh.mlh_rreg_lh);
1567                                 LASSERT(lock != NULL);
1568                                 lock_res_and_lock(lock);
1569                                 ldlm_set_atomic_cb(lock);
1570                                 unlock_res_and_lock(lock);
1571                                 LDLM_LOCK_PUT(lock);
1572                         }
1573
1574                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1575                         OBD_FREE_PTR(msl);
1576                         continue;
1577                 }
1578
1579                 INIT_LIST_HEAD(&msl->msl_linkage);
1580                 msl->msl_obj = lnkp;
1581                 list_add_tail(&msl->msl_linkage, link_locks);
1582
1583                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1584         }
1585
1586         if (blocked) {
1587                 rc = -EBUSY;
1588                 if (--retries > 0) {
1589                         mdt_unlock_list(info, link_locks, rc);
1590                         blocked = false;
1591                         local_lnkp_cnt = 0;
1592                         goto repeat;
1593                 }
1594         }
1595
1596         EXIT;
1597 out:
1598         if (rc)
1599                 mdt_unlock_list(info, link_locks, rc);
1600         else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1601                 /*
1602                  * parent may have 3 local objects: master object and 2 stripes
1603                  * (if it's being migrated too); source may have 2 local
1604                  * objects: master and 1 stripe; target has 1 local object.
1605                  */
1606                 rc = 1;
1607         return rc;
1608 }
1609
1610 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1611                                   struct mdt_object *obj,
1612                                   const struct md_attr *ma,
1613                                   struct list_head *slave_locks)
1614 {
1615         struct mdt_device *mdt = info->mti_mdt;
1616         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1617         struct lu_fid *fid = &info->mti_tmp_fid1;
1618         struct mdt_object *slave;
1619         struct mdt_sub_lock *msl;
1620         int i;
1621         int rc;
1622
1623         ENTRY;
1624
1625         LASSERT(mdt_object_remote(obj));
1626         LASSERT(ma->ma_valid & MA_LMV);
1627         LASSERT(lmv);
1628
1629         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1630                 RETURN(-EINVAL);
1631
1632         if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1633                 RETURN(0);
1634
1635         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1636                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1637
1638                 slave = mdt_object_find(info->mti_env, mdt, fid);
1639                 if (IS_ERR(slave))
1640                         GOTO(out, rc = PTR_ERR(slave));
1641
1642                 OBD_ALLOC_PTR(msl);
1643                 if (!msl) {
1644                         mdt_object_put(info->mti_env, slave);
1645                         GOTO(out, rc = -ENOMEM);
1646                 }
1647
1648                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1649                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1650                                            MDS_INODELOCK_UPDATE, true);
1651                 if (rc) {
1652                         OBD_FREE_PTR(msl);
1653                         mdt_object_put(info->mti_env, slave);
1654                         GOTO(out, rc);
1655                 }
1656
1657                 INIT_LIST_HEAD(&msl->msl_linkage);
1658                 msl->msl_obj = slave;
1659                 list_add_tail(&msl->msl_linkage, slave_locks);
1660
1661         }
1662         EXIT;
1663
1664 out:
1665         if (rc)
1666                 mdt_unlock_list(info, slave_locks, rc);
1667         return rc;
1668 }
1669
1670 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1671                                              struct mdt_object *obj,
1672                                              struct mdt_lock_handle *lh,
1673                                              struct ldlm_enqueue_info *einfo,
1674                                              struct list_head *slave_locks,
1675                                              int decref)
1676 {
1677         if (mdt_object_remote(obj)) {
1678                 mdt_unlock_list(info, slave_locks, decref);
1679                 mdt_object_unlock(info, obj, lh, decref);
1680         } else {
1681                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1682         }
1683 }
1684
1685 /* lock parent and its stripes */
1686 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1687                                    struct mdt_object *obj,
1688                                    const struct md_attr *ma,
1689                                    struct mdt_lock_handle *lh,
1690                                    struct ldlm_enqueue_info *einfo,
1691                                    struct list_head *slave_locks)
1692 {
1693         int rc;
1694
1695         if (mdt_object_remote(obj)) {
1696                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1697                                             &lh->mlh_rreg_lh, LCK_PW,
1698                                             MDS_INODELOCK_UPDATE, false);
1699                 if (rc != ELDLM_OK)
1700                         return rc;
1701
1702                 /*
1703                  * if obj is remote and striped, lock its stripes explicitly
1704                  * because it's not striped in LOD layer on this MDT.
1705                  */
1706                 if (ma->ma_valid & MA_LMV) {
1707                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1708                         if (rc)
1709                                 mdt_object_unlock(info, obj, lh, rc);
1710                 }
1711         } else {
1712                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1713                                             einfo, true);
1714         }
1715
1716         return rc;
1717 }
1718
1719 /*
1720  * in migration, object may be remote, and we need take full lock of it and its
1721  * stripes if it's directory, besides, object may be a remote object on its
1722  * parent, revoke its LOOKUP lock on where its parent is located.
1723  */
1724 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1725                                    struct mdt_object *pobj,
1726                                    struct mdt_object *obj,
1727                                    struct mdt_lock_handle *lh,
1728                                    struct ldlm_enqueue_info *einfo,
1729                                    struct list_head *slave_locks)
1730 {
1731         int rc;
1732
1733         if (mdt_object_remote(obj)) {
1734                 /* don't bother to check if pobj and obj are on the same MDT. */
1735                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1736                 if (rc)
1737                         return rc;
1738
1739                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1740                                             &lh->mlh_rreg_lh, LCK_EX,
1741                                             MDS_INODELOCK_FULL, false);
1742                 if (rc != ELDLM_OK)
1743                         return rc;
1744
1745                 /*
1746                  * if obj is remote and striped, lock its stripes explicitly
1747                  * because it's not striped in LOD layer on this MDT.
1748                  */
1749                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1750                         struct md_attr *ma = &info->mti_attr;
1751
1752                         ma->ma_lmv = info->mti_big_lmm;
1753                         ma->ma_lmv_size = info->mti_big_lmmsize;
1754                         ma->ma_valid = 0;
1755                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1756                         if (rc) {
1757                                 mdt_object_unlock(info, obj, lh, rc);
1758                                 return rc;
1759                         }
1760
1761                         if (ma->ma_valid & MA_LMV) {
1762                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1763                                                             slave_locks);
1764                                 if (rc)
1765                                         mdt_object_unlock(info, obj, lh, rc);
1766                         }
1767                 }
1768         } else {
1769                 if (mdt_object_remote(pobj)) {
1770                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1771                         if (rc)
1772                                 return rc;
1773                 }
1774
1775                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1776                                             einfo, true);
1777         }
1778
1779         return rc;
1780 }
1781
1782 /*
1783  * lookup source by name, if parent is striped directory, we need to find the
1784  * corresponding stripe where source is located, and then lookup there.
1785  *
1786  * besides, if parent is migrating too, and file is already in target stripe,
1787  * this should be a redo of 'lfs migrate' on client side.
1788  */
1789 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1790                               struct mdt_object *pobj,
1791                               const struct md_attr *ma,
1792                               const struct lu_name *lname,
1793                               struct mdt_object **spobj,
1794                               struct mdt_object **sobj)
1795 {
1796         const struct lu_env *env = info->mti_env;
1797         struct lu_fid *fid = &info->mti_tmp_fid1;
1798         struct mdt_object *stripe;
1799         int rc;
1800
1801         if (ma->ma_valid & MA_LMV) {
1802                 /* if parent is striped, lookup on corresponding stripe */
1803                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1804                 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1805                 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1806                 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1807                                     LMV_HASH_FLAG_MIGRATION;
1808
1809                 if (is_migrating) {
1810                         hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1811                         stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1812                 }
1813
1814                 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1815                                               lname->ln_name,
1816                                               lname->ln_namelen);
1817                 if (rc < 0)
1818                         return rc;
1819
1820                 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1821                         rc += le32_to_cpu(lmv->lmv_migrate_offset);
1822
1823                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1824
1825                 stripe = mdt_object_find(env, info->mti_mdt, fid);
1826                 if (IS_ERR(stripe))
1827                         return PTR_ERR(stripe);
1828
1829                 fid_zero(fid);
1830                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1831                                 &info->mti_spec);
1832                 if (rc == -ENOENT && is_migrating) {
1833                         /*
1834                          * if parent is migrating, and lookup child failed on
1835                          * source stripe, lookup again on target stripe, if it
1836                          * exists, it means previous migration was interrupted,
1837                          * and current file was migrated already.
1838                          */
1839                         mdt_object_put(env, stripe);
1840
1841                         hash_type = le32_to_cpu(lmv->lmv_hash_type);
1842                         stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1843
1844                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1845                                                       lname->ln_name,
1846                                                       lname->ln_namelen);
1847                         if (rc < 0)
1848                                 return rc;
1849
1850                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1851
1852                         stripe = mdt_object_find(env, info->mti_mdt, fid);
1853                         if (IS_ERR(stripe))
1854                                 return PTR_ERR(stripe);
1855
1856                         fid_zero(fid);
1857                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1858                                         fid, &info->mti_spec);
1859                         mdt_object_put(env, stripe);
1860                         return rc ?: -EALREADY;
1861                 } else if (rc) {
1862                         mdt_object_put(env, stripe);
1863                         return rc;
1864                 }
1865         } else {
1866                 fid_zero(fid);
1867                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1868                                 &info->mti_spec);
1869                 if (rc)
1870                         return rc;
1871
1872                 stripe = pobj;
1873                 mdt_object_get(env, stripe);
1874         }
1875
1876         *spobj = stripe;
1877
1878         *sobj = mdt_object_find(env, info->mti_mdt, fid);
1879         if (IS_ERR(*sobj)) {
1880                 mdt_object_put(env, stripe);
1881                 rc = PTR_ERR(*sobj);
1882                 *spobj = NULL;
1883                 *sobj = NULL;
1884         }
1885
1886         return rc;
1887 }
1888
1889 /* end lease and close file for regular file */
1890 static int mdd_migrate_close(struct mdt_thread_info *info,
1891                              struct mdt_object *obj)
1892 {
1893         struct close_data *data;
1894         struct mdt_body *repbody;
1895         struct ldlm_lock *lease;
1896         int rc;
1897         int rc2;
1898
1899         rc = -EPROTO;
1900         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1901                                       RCL_CLIENT) ||
1902             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1903                                       RCL_CLIENT))
1904                 goto close;
1905
1906         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1907         if (!data)
1908                 goto close;
1909
1910         rc = -ESTALE;
1911         lease = ldlm_handle2lock(&data->cd_handle);
1912         if (!lease)
1913                 goto close;
1914
1915         /* check if the lease was already canceled */
1916         lock_res_and_lock(lease);
1917         rc = ldlm_is_cancel(lease);
1918         unlock_res_and_lock(lease);
1919
1920         if (rc) {
1921                 rc = -EAGAIN;
1922                 LDLM_DEBUG(lease, DFID" lease broken",
1923                            PFID(mdt_object_fid(obj)));
1924         }
1925
1926         /*
1927          * cancel server side lease, client side counterpart should have been
1928          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1929          */
1930         ldlm_lock_cancel(lease);
1931         ldlm_reprocess_all(lease->l_resource);
1932         LDLM_LOCK_PUT(lease);
1933
1934 close:
1935         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1936         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1937         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1938
1939         return rc ?: rc2;
1940 }
1941
1942 /*
1943  * migrate file in below steps:
1944  *  1. lock parent and its stripes
1945  *  2. lookup source by name
1946  *  3. lock parents of source links if source is not directory
1947  *  4. reject if source is in HSM
1948  *  5. take source open_sem and close file if source is regular file
1949  *  6. lock source and its stripes if it's directory
1950  *  7. lock target so subsequent change to it can trigger COS
1951  *  8. migrate file
1952  *  9. unlock above locks
1953  * 10. sync device if source has links
1954  */
1955 static int mdt_reint_migrate_internal(struct mdt_thread_info *info)
1956 {
1957         const struct lu_env *env = info->mti_env;
1958         struct mdt_device *mdt = info->mti_mdt;
1959         struct mdt_reint_record *rr = &info->mti_rr;
1960         struct md_attr *ma = &info->mti_attr;
1961         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1962         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1963         struct mdt_object *pobj;
1964         struct mdt_object *spobj = NULL;
1965         struct mdt_object *sobj = NULL;
1966         struct mdt_object *tobj;
1967         struct mdt_lock_handle *lhp;
1968         struct mdt_lock_handle *lhs;
1969         struct mdt_lock_handle *lht;
1970         LIST_HEAD(parent_slave_locks);
1971         LIST_HEAD(child_slave_locks);
1972         LIST_HEAD(link_locks);
1973         bool open_sem_locked = false;
1974         bool do_sync = false;
1975         int rc;
1976         ENTRY;
1977
1978         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1979                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1980
1981         /* don't allow migrate . or .. */
1982         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
1983                 RETURN(-EBUSY);
1984
1985         rc = mdt_remote_permission(info);
1986         if (rc)
1987                 RETURN(rc);
1988
1989         /* pobj is master object of parent */
1990         pobj = mdt_object_find_check(info, rr->rr_fid1, 0);
1991         if (IS_ERR(pobj))
1992                 RETURN(PTR_ERR(pobj));
1993
1994         if (unlikely(!info->mti_big_lmm)) {
1995                 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
1996                 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
1997                 if (!info->mti_big_lmm)
1998                         GOTO(put_parent, rc = -ENOMEM);
1999         }
2000
2001         ma->ma_lmv = info->mti_big_lmm;
2002         ma->ma_lmv_size = info->mti_big_lmmsize;
2003         ma->ma_valid = 0;
2004         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2005         if (rc)
2006                 GOTO(put_parent, rc);
2007
2008         /* lock parent object */
2009         lhp = &info->mti_lh[MDT_LH_PARENT];
2010         mdt_lock_reg_init(lhp, LCK_PW);
2011         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2012                                      &parent_slave_locks);
2013         if (rc)
2014                 GOTO(put_parent, rc);
2015
2016         /*
2017          * spobj is the corresponding stripe against name if pobj is striped
2018          * directory, which is the real parent, and no need to lock, because
2019          * we've taken full lock of pobj.
2020          */
2021         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2022         if (rc)
2023                 GOTO(unlock_parent, rc);
2024
2025         /* lock parents of source links, and revoke LOOKUP lock of links */
2026         rc = mdt_lock_links(info, pobj, ma, sobj, &link_locks);
2027         if (rc < 0)
2028                 GOTO(put_source, rc);
2029
2030         /*
2031          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2032          * with one request, if total lock count exceeds this limit, we will
2033          * drop all locks after migration, and synchronous device in the end.
2034          */
2035         do_sync = rc;
2036
2037         /* if migration HSM is allowed */
2038         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2039                 ma->ma_need = MA_HSM;
2040                 ma->ma_valid = 0;
2041                 rc = mdt_attr_get_complex(info, sobj, ma);
2042                 if (rc)
2043                         GOTO(unlock_links, rc);
2044
2045                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2046                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2047         }
2048
2049         /* end lease and close file for regular file */
2050         if (info->mti_spec.sp_migrate_close) {
2051                 /* try to hold open_sem so that nobody else can open the file */
2052                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2053                         /* close anyway */
2054                         mdd_migrate_close(info, sobj);
2055                         GOTO(unlock_links, rc = -EBUSY);
2056                 } else {
2057                         open_sem_locked = true;
2058                         rc = mdd_migrate_close(info, sobj);
2059                         if (rc)
2060                                 GOTO(unlock_open_sem, rc);
2061                 }
2062         }
2063
2064         /* lock source */
2065         lhs = &info->mti_lh[MDT_LH_OLD];
2066         mdt_lock_reg_init(lhs, LCK_EX);
2067         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2068                                      &child_slave_locks);
2069         if (rc)
2070                 GOTO(unlock_open_sem, rc);
2071
2072         /* lock target */
2073         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2074         if (IS_ERR(tobj))
2075                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2076
2077         lht = &info->mti_lh[MDT_LH_NEW];
2078         mdt_lock_reg_init(lht, LCK_EX);
2079         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2080         if (rc)
2081                 GOTO(put_target, rc);
2082
2083         /* Don't do lookup sanity check. We know name doesn't exist. */
2084         info->mti_spec.sp_cr_lookup = 0;
2085         info->mti_spec.sp_feat = &dt_directory_features;
2086
2087         rc = mdo_migrate(env, mdt_object_child(pobj),
2088                          mdt_object_child(sobj), &rr->rr_name,
2089                          mdt_object_child(tobj), &info->mti_spec, ma);
2090         EXIT;
2091
2092         mdt_object_unlock(info, tobj, lht, rc);
2093 put_target:
2094         mdt_object_put(env, tobj);
2095 unlock_source:
2096         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2097                                   &child_slave_locks, rc);
2098 unlock_open_sem:
2099         if (open_sem_locked)
2100                 up_write(&sobj->mot_open_sem);
2101 unlock_links:
2102         mdt_unlock_list(info, &link_locks, rc);
2103 put_source:
2104         mdt_object_put(env, sobj);
2105         mdt_object_put(env, spobj);
2106 unlock_parent:
2107         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2108                                   &parent_slave_locks, rc);
2109 put_parent:
2110         mdt_object_put(env, pobj);
2111
2112         if (!rc && do_sync)
2113                 mdt_device_sync(env, mdt);
2114
2115         return rc;
2116 }
2117
2118 static int mdt_object_lock_save(struct mdt_thread_info *info,
2119                                 struct mdt_object *dir,
2120                                 struct mdt_lock_handle *lh,
2121                                 int idx, bool cos_incompat)
2122 {
2123         int rc;
2124
2125         /* we lock the target dir if it is local */
2126         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2127                                    cos_incompat);
2128         if (rc != 0)
2129                 return rc;
2130
2131         /* get and save correct version after locking */
2132         mdt_version_get_save(info, dir, idx);
2133         return 0;
2134 }
2135
2136 /*
2137  * determine lock order of sobj and tobj
2138  *
2139  * there are two situations we need to lock tobj before sobj:
2140  * 1. sobj is child of tobj
2141  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2142  *    larger than that of tobj
2143  *
2144  * \retval      1 lock tobj before sobj
2145  * \retval      0 lock sobj before tobj
2146  * \retval      -ev negative errno upon error
2147  */
2148 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2149                                            struct mdt_object *sobj,
2150                                            struct mdt_object *tobj)
2151 {
2152         struct md_attr *ma = &info->mti_attr;
2153         struct lu_fid *spfid = &info->mti_tmp_fid1;
2154         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2155         struct lmv_mds_md_v1 *lmv;
2156         __u32 sindex;
2157         __u32 tindex;
2158         int rc;
2159
2160         /* sobj and tobj are the same */
2161         if (sobj == tobj)
2162                 return 0;
2163
2164         if (fid_is_root(mdt_object_fid(sobj)))
2165                 return 0;
2166
2167         if (fid_is_root(mdt_object_fid(tobj)))
2168                 return 1;
2169
2170         /* check whether sobj is child of tobj */
2171         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2172                            mdt_object_fid(tobj));
2173         if (rc < 0)
2174                 return rc;
2175
2176         if (rc == 1)
2177                 return 1;
2178
2179         /* check whether sobj and tobj are children of the same parent */
2180         rc = mdt_attr_get_pfid(info, sobj, spfid);
2181         if (rc)
2182                 return rc;
2183
2184         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2185         if (rc)
2186                 return rc;
2187
2188         if (!lu_fid_eq(spfid, tpfid))
2189                 return 0;
2190
2191         /* check whether sobj and tobj are sibling stripes */
2192         ma->ma_need = MA_LMV;
2193         ma->ma_valid = 0;
2194         ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2195         ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2196         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2197         if (rc)
2198                 return rc;
2199
2200         if (!(ma->ma_valid & MA_LMV))
2201                 return 0;
2202
2203         lmv = &ma->ma_lmv->lmv_md_v1;
2204         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2205                 return 0;
2206         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2207
2208         ma->ma_valid = 0;
2209         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2210         if (rc)
2211                 return rc;
2212
2213         if (!(ma->ma_valid & MA_LMV))
2214                 return -ENODATA;
2215
2216         lmv = &ma->ma_lmv->lmv_md_v1;
2217         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2218                 return -EINVAL;
2219         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2220
2221         /* check stripe index of sobj and tobj */
2222         if (sindex == tindex)
2223                 return -EINVAL;
2224
2225         return sindex < tindex ? 0 : 1;
2226 }
2227
2228 /*
2229  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2230  * 2 - srcdir child; 3 - tgtdir child.
2231  * Update on disk version of srcdir child.
2232  */
2233 /**
2234  * For DNE phase I, only these renames are allowed
2235  *      mv src_p/src_c tgt_p/tgt_c
2236  * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
2237  * 2. src_p and tgt_p are same directory, and tgt_c does not
2238  *    exists. In this case, all of modification will happen
2239  *    in the MDT where ithesource parent is, only one remote
2240  *    update is needed, i.e. set c_time/m_time on the child.
2241  *    And tgt_c will be still in the same MDT as the original
2242  *    src_c.
2243  */
2244 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
2245                                      struct mdt_lock_handle *lhc)
2246 {
2247         struct mdt_reint_record *rr = &info->mti_rr;
2248         struct md_attr *ma = &info->mti_attr;
2249         struct ptlrpc_request *req = mdt_info_req(info);
2250         struct mdt_object *msrcdir = NULL;
2251         struct mdt_object *mtgtdir = NULL;
2252         struct mdt_object *mold;
2253         struct mdt_object *mnew = NULL;
2254         struct mdt_lock_handle *lh_srcdirp;
2255         struct mdt_lock_handle *lh_tgtdirp;
2256         struct mdt_lock_handle *lh_oldp = NULL;
2257         struct mdt_lock_handle *lh_newp = NULL;
2258         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2259         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2260         __u64 lock_ibits;
2261         bool reverse = false;
2262         bool cos_incompat, discard = false;
2263         int rc;
2264         ENTRY;
2265
2266         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2267                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2268                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2269
2270         /* find both parents. */
2271         msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
2272         if (IS_ERR(msrcdir))
2273                 RETURN(PTR_ERR(msrcdir));
2274
2275         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2276
2277         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2278                 mtgtdir = msrcdir;
2279                 mdt_object_get(info->mti_env, mtgtdir);
2280         } else {
2281                 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
2282                 if (IS_ERR(mtgtdir))
2283                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2284         }
2285
2286         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2287         if (rc < 0)
2288                 GOTO(out_put_tgtdir, rc);
2289
2290         reverse = rc;
2291
2292         /* source needs to be looked up after locking source parent, otherwise
2293          * this rename may race with unlink source, and cause rename hang, see
2294          * sanityn.sh 55b, so check parents first, if later we found source is
2295          * remote, relock parents. */
2296         cos_incompat = (mdt_object_remote(msrcdir) ||
2297                         mdt_object_remote(mtgtdir));
2298
2299         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2300
2301         /* lock parents in the proper order. */
2302         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2303         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2304
2305 relock:
2306         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2307         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2308
2309         if (reverse) {
2310                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2311                                           cos_incompat);
2312                 if (rc)
2313                         GOTO(out_put_tgtdir, rc);
2314
2315                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2316
2317                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2318                                           cos_incompat);
2319                 if (rc != 0) {
2320                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2321                         GOTO(out_put_tgtdir, rc);
2322                 }
2323         } else {
2324                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2325                                           cos_incompat);
2326                 if (rc)
2327                         GOTO(out_put_tgtdir, rc);
2328
2329                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2330
2331                 if (mtgtdir != msrcdir) {
2332                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2333                                                   cos_incompat);
2334                 } else if (lh_srcdirp->mlh_pdo_hash !=
2335                            lh_tgtdirp->mlh_pdo_hash) {
2336                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2337                                                 MDS_INODELOCK_UPDATE,
2338                                                 cos_incompat);
2339                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2340                 }
2341                 if (rc != 0) {
2342                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2343                         GOTO(out_put_tgtdir, rc);
2344                 }
2345         }
2346
2347         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2348         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2349
2350         /* find mold object. */
2351         fid_zero(old_fid);
2352         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2353         if (rc != 0)
2354                 GOTO(out_unlock_parents, rc);
2355
2356         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2357                 GOTO(out_unlock_parents, rc = -EINVAL);
2358
2359         if (!fid_is_md_operative(old_fid))
2360                 GOTO(out_unlock_parents, rc = -EPERM);
2361
2362         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2363         if (IS_ERR(mold))
2364                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2365
2366         if (!mdt_object_exists(mold)) {
2367                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2368                                 &mold->mot_obj,
2369                                 "object does not exist");
2370                 GOTO(out_put_old, rc = -ENOENT);
2371         }
2372
2373         /* Check if @mtgtdir is subdir of @mold, before locking child
2374          * to avoid reverse locking. */
2375         if (mtgtdir != msrcdir) {
2376                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2377                                    old_fid);
2378                 if (rc) {
2379                         if (rc == 1)
2380                                 rc = -EINVAL;
2381                         GOTO(out_put_old, rc);
2382                 }
2383         }
2384
2385         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2386         /* save version after locking */
2387         mdt_version_get_save(info, mold, 2);
2388
2389         if (!cos_incompat && mdt_object_remote(mold)) {
2390                 cos_incompat = true;
2391                 mdt_object_put(info->mti_env, mold);
2392                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2393                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2394                 goto relock;
2395         }
2396
2397         /* find mnew object:
2398          * mnew target object may not exist now
2399          * lookup with version checking */
2400         fid_zero(new_fid);
2401         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2402                                       3);
2403         if (rc == 0) {
2404                 /* the new_fid should have been filled at this moment */
2405                 if (lu_fid_eq(old_fid, new_fid))
2406                         GOTO(out_put_old, rc);
2407
2408                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2409                     lu_fid_eq(new_fid, rr->rr_fid2))
2410                         GOTO(out_put_old, rc = -EINVAL);
2411
2412                 if (!fid_is_md_operative(new_fid))
2413                         GOTO(out_put_old, rc = -EPERM);
2414
2415                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2416                 if (IS_ERR(mnew))
2417                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2418
2419                 if (!mdt_object_exists(mnew)) {
2420                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2421                                         &mnew->mot_obj,
2422                                         "object does not exist");
2423                         GOTO(out_put_new, rc = -ENOENT);
2424                 }
2425
2426                 if (mdt_object_remote(mnew)) {
2427                         struct mdt_body  *repbody;
2428
2429                         /* Always send rename req to the target child MDT */
2430                         repbody = req_capsule_server_get(info->mti_pill,
2431                                                          &RMF_MDT_BODY);
2432                         LASSERT(repbody != NULL);
2433                         repbody->mbo_fid1 = *new_fid;
2434                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2435                         GOTO(out_put_new, rc = -EXDEV);
2436                 }
2437                 /* Before locking the target dir, check we do not replace
2438                  * a dir with a non-dir, otherwise it may deadlock with
2439                  * link op which tries to create a link in this dir
2440                  * back to this non-dir. */
2441                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2442                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2443                         GOTO(out_put_new, rc = -EISDIR);
2444
2445                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2446                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2447                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2448                 if (mdt_object_remote(msrcdir)) {
2449                         /* Enqueue lookup lock from the parent MDT */
2450                         rc = mdt_remote_object_lock(info, msrcdir,
2451                                                     mdt_object_fid(mold),
2452                                                     &lh_oldp->mlh_rreg_lh,
2453                                                     lh_oldp->mlh_rreg_mode,
2454                                                     MDS_INODELOCK_LOOKUP,
2455                                                     false);
2456                         if (rc != ELDLM_OK)
2457                                 GOTO(out_put_new, rc);
2458
2459                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2460                 }
2461
2462                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2463                                            cos_incompat);
2464                 if (rc != 0)
2465                         GOTO(out_unlock_old, rc);
2466
2467                 /* Check if @msrcdir is subdir of @mnew, before locking child
2468                  * to avoid reverse locking. */
2469                 if (mtgtdir != msrcdir) {
2470                         rc = mdo_is_subdir(info->mti_env,
2471                                            mdt_object_child(msrcdir), new_fid);
2472                         if (rc) {
2473                                 if (rc == 1)
2474                                         rc = -EINVAL;
2475                                 GOTO(out_unlock_old, rc);
2476                         }
2477                 }
2478
2479                 /* We used to acquire MDS_INODELOCK_FULL here but we
2480                  * can't do this now because a running HSM restore on
2481                  * the rename onto victim will hold the layout
2482                  * lock. See LU-4002. */
2483
2484                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2485                 mdt_lock_reg_init(lh_newp, LCK_EX);
2486                 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2487                                            MDS_INODELOCK_LOOKUP |
2488                                            MDS_INODELOCK_UPDATE,
2489                                            cos_incompat);
2490                 if (rc != 0)
2491                         GOTO(out_unlock_old, rc);
2492
2493                 /* get and save version after locking */
2494                 mdt_version_get_save(info, mnew, 3);
2495         } else if (rc != -EREMOTE && rc != -ENOENT) {
2496                 GOTO(out_put_old, rc);
2497         } else {
2498                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2499                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2500                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2501                 if (mdt_object_remote(msrcdir)) {
2502                         /* Enqueue lookup lock from the parent MDT */
2503                         rc = mdt_remote_object_lock(info, msrcdir,
2504                                                     mdt_object_fid(mold),
2505                                                     &lh_oldp->mlh_rreg_lh,
2506                                                     lh_oldp->mlh_rreg_mode,
2507                                                     MDS_INODELOCK_LOOKUP,
2508                                                     false);
2509                         if (rc != ELDLM_OK)
2510                                 GOTO(out_put_old, rc);
2511
2512                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2513                 }
2514
2515                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2516                                            cos_incompat);
2517                 if (rc != 0)
2518                         GOTO(out_unlock_old, rc);
2519
2520                 mdt_enoent_version_save(info, 3);
2521         }
2522
2523         /* step 5: rename it */
2524         mdt_reint_init_ma(info, ma);
2525
2526         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2527                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2528
2529         if (mnew != NULL)
2530                 mutex_lock(&mnew->mot_lov_mutex);
2531
2532         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2533                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2534                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2535                         &rr->rr_tgt_name, ma);
2536
2537         if (mnew != NULL)
2538                 mutex_unlock(&mnew->mot_lov_mutex);
2539
2540         /* handle last link of tgt object */
2541         if (rc == 0) {
2542                 mdt_counter_incr(req, LPROC_MDT_RENAME);
2543                 if (mnew) {
2544                         mdt_handle_last_unlink(info, mnew, ma);
2545                         discard = mdt_dom_check_for_discard(info, mnew);
2546                 }
2547
2548                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2549                                          msrcdir, mtgtdir);
2550         }
2551
2552         EXIT;
2553         if (mnew != NULL)
2554                 mdt_object_unlock(info, mnew, lh_newp, rc);
2555 out_unlock_old:
2556         mdt_object_unlock(info, mold, lh_oldp, rc);
2557 out_put_new:
2558         if (mnew != NULL)
2559                 mdt_object_put(info->mti_env, mnew);
2560 out_put_old:
2561         mdt_object_put(info->mti_env, mold);
2562 out_unlock_parents:
2563         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2564         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2565 out_put_tgtdir:
2566         mdt_object_put(info->mti_env, mtgtdir);
2567 out_put_srcdir:
2568         mdt_object_put(info->mti_env, msrcdir);
2569
2570         /* If 'discard' is set then new_fid must exits.
2571          * DOM data discard need neither object nor lock,
2572          * so do this at the end.
2573          */
2574         if (discard)
2575                 mdt_dom_discard_data(info, new_fid);
2576
2577         return rc;
2578 }
2579
2580 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2581                                        struct mdt_lock_handle *lhc, bool rename)
2582 {
2583         struct mdt_reint_record *rr = &info->mti_rr;
2584         struct ptlrpc_request   *req = mdt_info_req(info);
2585         struct lustre_handle    rename_lh = { 0 };
2586         int                     rc;
2587         ENTRY;
2588
2589         if (info->mti_dlm_req)
2590                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2591
2592         if (!fid_is_md_operative(rr->rr_fid1) ||
2593             !fid_is_md_operative(rr->rr_fid2))
2594                 RETURN(-EPERM);
2595
2596         /* Note: do not enqueue rename lock for replay request, because
2597          * if other MDT holds rename lock, but being blocked to wait for
2598          * this MDT to finish its recovery, and the failover MDT can not
2599          * get rename lock, which will cause deadlock. */
2600         if (!req_is_replay(req)) {
2601                 rc = mdt_rename_lock(info, &rename_lh);
2602                 if (rc != 0) {
2603                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2604                                mdt_obd_name(info->mti_mdt), rc);
2605                         RETURN(rc);
2606                 }
2607         }
2608
2609         if (rename)
2610                 rc = mdt_reint_rename_internal(info, lhc);
2611         else
2612                 rc = mdt_reint_migrate_internal(info);
2613
2614         if (lustre_handle_is_used(&rename_lh))
2615                 mdt_rename_unlock(&rename_lh);
2616
2617         RETURN(rc);
2618 }
2619
2620 static int mdt_reint_rename(struct mdt_thread_info *info,
2621                             struct mdt_lock_handle *lhc)
2622 {
2623         return mdt_reint_rename_or_migrate(info, lhc, true);
2624 }
2625
2626 static int mdt_reint_migrate(struct mdt_thread_info *info,
2627                             struct mdt_lock_handle *lhc)
2628 {
2629         return mdt_reint_rename_or_migrate(info, lhc, false);
2630 }
2631
2632 static int mdt_reint_resync(struct mdt_thread_info *info,
2633                             struct mdt_lock_handle *lhc)
2634 {
2635         struct mdt_reint_record *rr = &info->mti_rr;
2636         struct ptlrpc_request   *req = mdt_info_req(info);
2637         struct md_attr          *ma = &info->mti_attr;
2638         struct mdt_object       *mo;
2639         struct ldlm_lock        *lease;
2640         struct mdt_body         *repbody;
2641         struct md_layout_change  layout = { 0 };
2642         bool                     lease_broken;
2643         int                      rc, rc2;
2644         ENTRY;
2645
2646         DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2647
2648         if (info->mti_dlm_req)
2649                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2650
2651         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2652         if (IS_ERR(mo))
2653                 GOTO(out, rc = PTR_ERR(mo));
2654
2655         if (!mdt_object_exists(mo))
2656                 GOTO(out_obj, rc = -ENOENT);
2657
2658         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2659                 GOTO(out_obj, rc = -EINVAL);
2660
2661         if (mdt_object_remote(mo))
2662                 GOTO(out_obj, rc = -EREMOTE);
2663
2664         lease = ldlm_handle2lock(rr->rr_handle);
2665         if (lease == NULL)
2666                 GOTO(out_obj, rc = -ESTALE);
2667
2668         /* It's really necessary to grab open_sem and check if the lease lock
2669          * has been lost. There would exist a concurrent writer coming in and
2670          * generating some dirty data in memory cache, the writeback would fail
2671          * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2672         if (!down_write_trylock(&mo->mot_open_sem))
2673                 GOTO(out_put_lease, rc = -EBUSY);
2674
2675         lock_res_and_lock(lease);
2676         lease_broken = ldlm_is_cancel(lease);
2677         unlock_res_and_lock(lease);
2678         if (lease_broken)
2679                 GOTO(out_unlock, rc = -EBUSY);
2680
2681         /* the file has yet opened by anyone else after we took the lease. */
2682         layout.mlc_opc = MD_LAYOUT_RESYNC;
2683         rc = mdt_layout_change(info, mo, &layout);
2684         if (rc)
2685                 GOTO(out_unlock, rc);
2686
2687         ma->ma_need = MA_INODE;
2688         ma->ma_valid = 0;
2689         rc = mdt_attr_get_complex(info, mo, ma);
2690         if (rc != 0)
2691                 GOTO(out_unlock, rc);
2692
2693         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2694         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2695
2696         EXIT;
2697 out_unlock:
2698         up_write(&mo->mot_open_sem);
2699 out_put_lease:
2700         LDLM_LOCK_PUT(lease);
2701 out_obj:
2702         mdt_object_put(info->mti_env, mo);
2703 out:
2704         mdt_client_compatibility(info);
2705         rc2 = mdt_fix_reply(info);
2706         if (rc == 0)
2707                 rc = rc2;
2708         return rc;
2709 }
2710
2711 struct mdt_reinter {
2712         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2713         enum lprocfs_extra_opc mr_extra_opc;
2714 };
2715
2716 static const struct mdt_reinter mdt_reinters[] = {
2717         [REINT_SETATTR] = {
2718                 .mr_handler = &mdt_reint_setattr,
2719                 .mr_extra_opc = MDS_REINT_SETATTR,
2720         },
2721         [REINT_CREATE] = {
2722                 .mr_handler = &mdt_reint_create,
2723                 .mr_extra_opc = MDS_REINT_CREATE,
2724         },
2725         [REINT_LINK] = {
2726                 .mr_handler = &mdt_reint_link,
2727                 .mr_extra_opc = MDS_REINT_LINK,
2728         },
2729         [REINT_UNLINK] = {
2730                 .mr_handler = &mdt_reint_unlink,
2731                 .mr_extra_opc = MDS_REINT_UNLINK,
2732         },
2733         [REINT_RENAME] = {
2734                 .mr_handler = &mdt_reint_rename,
2735                 .mr_extra_opc = MDS_REINT_RENAME,
2736         },
2737         [REINT_OPEN] = {
2738                 .mr_handler = &mdt_reint_open,
2739                 .mr_extra_opc = MDS_REINT_OPEN,
2740         },
2741         [REINT_SETXATTR] = {
2742                 .mr_handler = &mdt_reint_setxattr,
2743                 .mr_extra_opc = MDS_REINT_SETXATTR,
2744         },
2745         [REINT_RMENTRY] = {
2746                 .mr_handler = &mdt_reint_unlink,
2747                 .mr_extra_opc = MDS_REINT_UNLINK,
2748         },
2749         [REINT_MIGRATE] = {
2750                 .mr_handler = &mdt_reint_migrate,
2751                 .mr_extra_opc = MDS_REINT_RENAME,
2752         },
2753         [REINT_RESYNC] = {
2754                 .mr_handler = &mdt_reint_resync,
2755                 .mr_extra_opc = MDS_REINT_RESYNC,
2756         },
2757 };
2758
2759 int mdt_reint_rec(struct mdt_thread_info *info,
2760                   struct mdt_lock_handle *lhc)
2761 {
2762         const struct mdt_reinter *mr;
2763         int rc;
2764         ENTRY;
2765
2766         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2767                 RETURN(-EPROTO);
2768
2769         mr = &mdt_reinters[info->mti_rr.rr_opcode];
2770         if (mr->mr_handler == NULL)
2771                 RETURN(-EPROTO);
2772
2773         rc = (*mr->mr_handler)(info, lhc);
2774
2775         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2776                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
2777
2778         RETURN(rc);
2779 }