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_dir_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_dir_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_dir_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_dir_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_dir_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_dir_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_dir_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         struct lu_ucred *uc = mdt_ucred(info);
 642         cfs_cap_t cap_saved;
 643         int rc;
 644         ENTRY;
 645
 646         /* If the file was modified, add the dirty flag */
 647         ma->ma_need = MA_HSM;
 648         rc = mdt_attr_get_complex(info, mo, ma);
 649         if (rc) {
 650                 CERROR("file attribute read error for "DFID": %d.\n",
 651                         PFID(mdt_object_fid(mo)), rc);
 652                 RETURN(rc);
 653         }
 654
 655         /* If an up2date copy exists in the backend, add dirty flag */
 656         if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
 657             && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
 658                 ma->ma_hsm.mh_flags |= HS_DIRTY;
 659
 660                 /* Bump cap so that closes from non-owner writers can
 661                  * set the HSM state to dirty. */
 662                 cap_saved = uc->uc_cap;
 663                 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
 664                 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
 665                 uc->uc_cap = cap_saved;
 666                 if (rc)
 667                         CERROR("file attribute change error for "DFID": %d\n",
 668                                 PFID(mdt_object_fid(mo)), rc);
 669         }
 670
 671         RETURN(rc);
 672 }
 673
 674 static int mdt_reint_setattr(struct mdt_thread_info *info,
 675                              struct mdt_lock_handle *lhc)
 676 {
 677         struct md_attr          *ma = &info->mti_attr;
 678         struct mdt_reint_record *rr = &info->mti_rr;
 679         struct ptlrpc_request   *req = mdt_info_req(info);
 680         struct mdt_object       *mo;
 681         struct mdt_body         *repbody;
 682         int                      rc, rc2;
 683         ENTRY;
 684
 685         DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
 686                   (unsigned int)ma->ma_attr.la_valid);
 687
 688         if (info->mti_dlm_req)
 689                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 690
 691         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 692         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 693         if (IS_ERR(mo))
 694                 GOTO(out, rc = PTR_ERR(mo));
 695
 696         if (!mdt_object_exists(mo))
 697                 GOTO(out_put, rc = -ENOENT);
 698
 699         if (mdt_object_remote(mo))
 700                 GOTO(out_put, rc = -EREMOTE);
 701
 702         /* revoke lease lock if size is going to be changed */
 703         if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
 704                      !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
 705                      atomic_read(&mo->mot_lease_count) > 0)) {
 706                 down_read(&mo->mot_open_sem);
 707
 708                 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
 709                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 710                         mdt_lock_reg_init(lhc, LCK_CW);
 711
 712                         rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
 713                         if (rc != 0) {
 714                                 up_read(&mo->mot_open_sem);
 715                                 GOTO(out_put, rc);
 716                         }
 717
 718                         /* revoke lease lock */
 719                         mdt_object_unlock(info, mo, lhc, 1);
 720                 }
 721                 up_read(&mo->mot_open_sem);
 722         }
 723
 724         if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
 725                 /* Check write access for the O_TRUNC case */
 726                 if (mdt_write_read(mo) < 0)
 727                         GOTO(out_put, rc = -ETXTBSY);
 728
 729                 /* LU-10286: compatibility check for FLR.
 730                  * Please check the comment in mdt_finish_open() for details */
 731                 if (!exp_connect_flr(info->mti_exp)) {
 732                         rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
 733                         if (rc < 0 && rc != -ENODATA)
 734                                 GOTO(out_put, rc);
 735
 736                         if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm))
 737                                 GOTO(out_put, rc = -EOPNOTSUPP);
 738                 }
 739
 740                 /* For truncate, the file size sent from client
 741                  * is believable, but the blocks are incorrect,
 742                  * which makes the block size in LSOM attribute
 743                  * inconsisent with the real block size.
 744                  */
 745                 rc = mdt_lsom_update(info, mo, true);
 746                 if (rc)
 747                         GOTO(out_put, rc);
 748         }
 749
 750         if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
 751                 if (ma->ma_valid & MA_LOV)
 752                         GOTO(out_put, rc = -EPROTO);
 753
 754                 rc = mdt_attr_set(info, mo, ma);
 755                 if (rc)
 756                         GOTO(out_put, rc);
 757         } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
 758                    (ma->ma_valid & MA_INODE)) {
 759                 struct lu_buf *buf  = &info->mti_buf;
 760                 struct mdt_lock_handle  *lh;
 761
 762                 rc = mdt_remote_dir_permission(info);
 763                 if (rc < 0)
 764                         GOTO(out_put, rc);
 765
 766                 if (ma->ma_attr.la_valid != 0)
 767                         GOTO(out_put, rc = -EPROTO);
 768
 769                 lh = &info->mti_lh[MDT_LH_PARENT];
 770                 mdt_lock_reg_init(lh, LCK_PW);
 771
 772                 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
 773                 if (rc != 0)
 774                         GOTO(out_put, rc);
 775
 776                 if (ma->ma_valid & MA_LOV) {
 777                         buf->lb_buf = ma->ma_lmm;
 778                         buf->lb_len = ma->ma_lmm_size;
 779                 } else {
 780                         buf->lb_buf = ma->ma_lmv;
 781                         buf->lb_len = ma->ma_lmv_size;
 782                 }
 783                 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
 784                                   (ma->ma_valid & MA_LOV) ?
 785                                         XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
 786                                   0);
 787
 788                 mdt_object_unlock(info, mo, lh, rc);
 789                 if (rc)
 790                         GOTO(out_put, rc);
 791         } else {
 792                 GOTO(out_put, rc = -EPROTO);
 793         }
 794
 795         /* If file data is modified, add the dirty flag */
 796         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
 797                 rc = mdt_add_dirty_flag(info, mo, ma);
 798
 799         ma->ma_need = MA_INODE;
 800         ma->ma_valid = 0;
 801         rc = mdt_attr_get_complex(info, mo, ma);
 802         if (rc != 0)
 803                 GOTO(out_put, rc);
 804
 805         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
 806
 807         EXIT;
 808 out_put:
 809         mdt_object_put(info->mti_env, mo);
 810 out:
 811         if (rc == 0)
 812                 mdt_counter_incr(req, LPROC_MDT_SETATTR);
 813
 814         mdt_client_compatibility(info);
 815         rc2 = mdt_fix_reply(info);
 816         if (rc == 0)
 817                 rc = rc2;
 818         return rc;
 819 }
 820
 821 static int mdt_reint_create(struct mdt_thread_info *info,
 822                             struct mdt_lock_handle *lhc)
 823 {
 824         struct ptlrpc_request   *req = mdt_info_req(info);
 825         int                     rc;
 826         ENTRY;
 827
 828         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 829                 RETURN(err_serious(-ESTALE));
 830
 831         if (info->mti_dlm_req)
 832                 ldlm_request_cancel(mdt_info_req(info),
 833                                     info->mti_dlm_req, 0, LATF_SKIP);
 834
 835         if (!lu_name_is_valid(&info->mti_rr.rr_name))
 836                 RETURN(-EPROTO);
 837
 838         switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
 839         case S_IFDIR:
 840                 mdt_counter_incr(req, LPROC_MDT_MKDIR);
 841                 break;
 842         case S_IFREG:
 843         case S_IFLNK:
 844         case S_IFCHR:
 845         case S_IFBLK:
 846         case S_IFIFO:
 847         case S_IFSOCK:
 848                 /* Special file should stay on the same node as parent. */
 849                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
 850                 break;
 851         default:
 852                 CERROR("%s: Unsupported mode %o\n",
 853                        mdt_obd_name(info->mti_mdt),
 854                        info->mti_attr.ma_attr.la_mode);
 855                 RETURN(err_serious(-EOPNOTSUPP));
 856         }
 857
 858         rc = mdt_create(info);
 859         RETURN(rc);
 860 }
 861
 862 /*
 863  * VBR: save parent version in reply and child version getting by its name.
 864  * Version of child is getting and checking during its lookup. If
 865  */
 866 static int mdt_reint_unlink(struct mdt_thread_info *info,
 867                             struct mdt_lock_handle *lhc)
 868 {
 869         struct mdt_reint_record *rr = &info->mti_rr;
 870         struct ptlrpc_request *req = mdt_info_req(info);
 871         struct md_attr *ma = &info->mti_attr;
 872         struct lu_fid *child_fid = &info->mti_tmp_fid1;
 873         struct mdt_object *mp;
 874         struct mdt_object *mc;
 875         struct mdt_lock_handle *parent_lh;
 876         struct mdt_lock_handle *child_lh;
 877         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 878         __u64 lock_ibits;
 879         bool cos_incompat = false, discard = false;
 880         int no_name = 0;
 881         int rc;
 882
 883         ENTRY;
 884
 885         DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
 886                   PNAME(&rr->rr_name));
 887
 888         if (info->mti_dlm_req)
 889                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 890
 891         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
 892                 RETURN(err_serious(-ENOENT));
 893
 894         if (!fid_is_md_operative(rr->rr_fid1))
 895                 RETURN(-EPERM);
 896
 897         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 898         if (IS_ERR(mp))
 899                 RETURN(PTR_ERR(mp));
 900
 901         if (mdt_object_remote(mp)) {
 902                 cos_incompat = true;
 903         } else {
 904                 rc = mdt_version_get_check_save(info, mp, 0);
 905                 if (rc)
 906                         GOTO(put_parent, rc);
 907         }
 908
 909 relock:
 910         parent_lh = &info->mti_lh[MDT_LH_PARENT];
 911         mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
 912         rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
 913                                    cos_incompat);
 914         if (rc != 0)
 915                 GOTO(put_parent, rc);
 916
 917         /* lookup child object along with version checking */
 918         fid_zero(child_fid);
 919         rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
 920         if (rc != 0) {
 921                 /* Name might not be able to find during resend of
 922                  * remote unlink, considering following case.
 923                  * dir_A is a remote directory, the name entry of
 924                  * dir_A is on MDT0, the directory is on MDT1,
 925                  *
 926                  * 1. client sends unlink req to MDT1.
 927                  * 2. MDT1 sends name delete update to MDT0.
 928                  * 3. name entry is being deleted in MDT0 synchronously.
 929                  * 4. MDT1 is restarted.
 930                  * 5. client resends unlink req to MDT1. So it can not
 931                  *    find the name entry on MDT0 anymore.
 932                  * In this case, MDT1 only needs to destory the local
 933                  * directory.
 934                  * */
 935                 if (mdt_object_remote(mp) && rc == -ENOENT &&
 936                     !fid_is_zero(rr->rr_fid2) &&
 937                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 938                         no_name = 1;
 939                         *child_fid = *rr->rr_fid2;
 940                  } else {
 941                         GOTO(unlock_parent, rc);
 942                  }
 943         }
 944
 945         if (!fid_is_md_operative(child_fid))
 946                 GOTO(unlock_parent, rc = -EPERM);
 947
 948         /* We will lock the child regardless it is local or remote. No harm. */
 949         mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
 950         if (IS_ERR(mc))
 951                 GOTO(unlock_parent, rc = PTR_ERR(mc));
 952
 953         if (!cos_incompat) {
 954                 rc = mdt_object_striped(info, mc);
 955                 if (rc < 0)
 956                         GOTO(unlock_parent, rc);
 957
 958                 cos_incompat = rc;
 959                 if (cos_incompat) {
 960                         mdt_object_put(info->mti_env, mc);
 961                         mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
 962                         goto relock;
 963                 }
 964         }
 965
 966         child_lh = &info->mti_lh[MDT_LH_CHILD];
 967         mdt_lock_reg_init(child_lh, LCK_EX);
 968         if (info->mti_spec.sp_rm_entry) {
 969                 struct lu_ucred *uc  = mdt_ucred(info);
 970
 971                 if (!mdt_is_dne_client(req->rq_export))
 972                         /* Return -ENOTSUPP for old client */
 973                         GOTO(put_child, rc = -ENOTSUPP);
 974
 975                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
 976                         GOTO(put_child, rc = -EPERM);
 977
 978                 ma->ma_need = MA_INODE;
 979                 ma->ma_valid = 0;
 980                 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
 981                                 NULL, &rr->rr_name, ma, no_name);
 982                 GOTO(put_child, rc);
 983         }
 984
 985         if (mdt_object_remote(mc)) {
 986                 struct mdt_body  *repbody;
 987
 988                 if (!fid_is_zero(rr->rr_fid2)) {
 989                         CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
 990                                mdt_obd_name(info->mti_mdt),
 991                                PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 992                         GOTO(put_child, rc = -ENOENT);
 993                 }
 994                 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
 995                        mdt_obd_name(info->mti_mdt),
 996                        PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 997
 998                 if (!mdt_is_dne_client(req->rq_export))
 999                         /* Return -ENOTSUPP for old client */
1000                         GOTO(put_child, rc = -ENOTSUPP);
1001
1002                 /* Revoke the LOOKUP lock of the remote object granted by
1003                  * this MDT. Since the unlink will happen on another MDT,
1004                  * it will release the LOOKUP lock right away. Then What
1005                  * would happen if another client try to grab the LOOKUP
1006                  * lock at the same time with unlink XXX */
1007                 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1008                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1009                 LASSERT(repbody != NULL);
1010                 repbody->mbo_fid1 = *mdt_object_fid(mc);
1011                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1012                 GOTO(unlock_child, rc = -EREMOTE);
1013         }
1014         /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1015          * this now because a running HSM restore on the child (unlink
1016          * victim) will hold the layout lock. See LU-4002. */
1017         lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1018         if (mdt_object_remote(mp)) {
1019                 /* Enqueue lookup lock from parent MDT */
1020                 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1021                                             &child_lh->mlh_rreg_lh,
1022                                             child_lh->mlh_rreg_mode,
1023                                             MDS_INODELOCK_LOOKUP, false);
1024                 if (rc != ELDLM_OK)
1025                         GOTO(put_child, rc);
1026
1027                 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1028         }
1029
1030         rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1031                                     cos_incompat);
1032         if (rc != 0)
1033                 GOTO(put_child, rc);
1034
1035         /*
1036          * Now we can only make sure we need MA_INODE, in mdd layer, will check
1037          * whether need MA_LOV and MA_COOKIE.
1038          */
1039         ma->ma_need = MA_INODE;
1040         ma->ma_valid = 0;
1041
1042         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1043                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1044         /* save version when object is locked */
1045         mdt_version_get_save(info, mc, 1);
1046
1047         mutex_lock(&mc->mot_lov_mutex);
1048
1049         rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1050                         mdt_object_child(mc), &rr->rr_name, ma, no_name);
1051
1052         mutex_unlock(&mc->mot_lov_mutex);
1053         if (rc != 0)
1054                 GOTO(unlock_child, rc);
1055
1056         if (!lu_object_is_dying(&mc->mot_header)) {
1057                 rc = mdt_attr_get_complex(info, mc, ma);
1058                 if (rc)
1059                         GOTO(out_stat, rc);
1060         } else {
1061                 discard = mdt_dom_check_for_discard(info, mc);
1062         }
1063         mdt_handle_last_unlink(info, mc, ma);
1064
1065 out_stat:
1066         if (ma->ma_valid & MA_INODE) {
1067                 switch (ma->ma_attr.la_mode & S_IFMT) {
1068                 case S_IFDIR:
1069                         mdt_counter_incr(req, LPROC_MDT_RMDIR);
1070                         break;
1071                 case S_IFREG:
1072                 case S_IFLNK:
1073                 case S_IFCHR:
1074                 case S_IFBLK:
1075                 case S_IFIFO:
1076                 case S_IFSOCK:
1077                         mdt_counter_incr(req, LPROC_MDT_UNLINK);
1078                         break;
1079                 default:
1080                         LASSERTF(0, "bad file type %o unlinking\n",
1081                                 ma->ma_attr.la_mode);
1082                 }
1083         }
1084
1085         EXIT;
1086
1087 unlock_child:
1088         mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1089 put_child:
1090         mdt_object_put(info->mti_env, mc);
1091 unlock_parent:
1092         mdt_object_unlock(info, mp, parent_lh, rc);
1093 put_parent:
1094         mdt_object_put(info->mti_env, mp);
1095
1096         /* discard is just a PW DOM lock to drop the data on a client
1097          * no need to keep objects being get and locked, do that after all.
1098          */
1099         if (discard)
1100                 mdt_dom_discard_data(info, child_fid);
1101
1102         return rc;
1103 }
1104
1105 /*
1106  * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1107  * name.
1108  */
1109 static int mdt_reint_link(struct mdt_thread_info *info,
1110                           struct mdt_lock_handle *lhc)
1111 {
1112         struct mdt_reint_record *rr = &info->mti_rr;
1113         struct ptlrpc_request   *req = mdt_info_req(info);
1114         struct md_attr          *ma = &info->mti_attr;
1115         struct mdt_object       *ms;
1116         struct mdt_object       *mp;
1117         struct mdt_lock_handle  *lhs;
1118         struct mdt_lock_handle  *lhp;
1119         bool cos_incompat;
1120         int rc;
1121         ENTRY;
1122
1123         DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1124                   PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1125
1126         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1127                 RETURN(err_serious(-ENOENT));
1128
1129         if (info->mti_dlm_req)
1130                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1131
1132         /* Invalid case so return error immediately instead of
1133          * processing it */
1134         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1135                 RETURN(-EPERM);
1136
1137         if (!fid_is_md_operative(rr->rr_fid1) ||
1138             !fid_is_md_operative(rr->rr_fid2))
1139                 RETURN(-EPERM);
1140
1141         /* step 1: find target parent dir */
1142         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1143         if (IS_ERR(mp))
1144                 RETURN(PTR_ERR(mp));
1145
1146         rc = mdt_version_get_check_save(info, mp, 0);
1147         if (rc)
1148                 GOTO(put_parent, rc);
1149
1150         /* step 2: find source */
1151         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1152         if (IS_ERR(ms))
1153                 GOTO(put_parent, rc = PTR_ERR(ms));
1154
1155         if (!mdt_object_exists(ms)) {
1156                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1157                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1158                 GOTO(put_source, rc = -ENOENT);
1159         }
1160
1161         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1162
1163         lhp = &info->mti_lh[MDT_LH_PARENT];
1164         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1165         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1166                                    cos_incompat);
1167         if (rc != 0)
1168                 GOTO(put_source, rc);
1169
1170         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1171
1172         lhs = &info->mti_lh[MDT_LH_CHILD];
1173         mdt_lock_reg_init(lhs, LCK_EX);
1174         rc = mdt_reint_object_lock(info, ms, lhs,
1175                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1176                                    cos_incompat);
1177         if (rc != 0)
1178                 GOTO(unlock_parent, rc);
1179
1180         /* step 3: link it */
1181         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1182                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1183
1184         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1185         rc = mdt_version_get_check_save(info, ms, 1);
1186         if (rc)
1187                 GOTO(unlock_source, rc);
1188
1189         /** check target version by name during replay */
1190         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1191                                       &info->mti_tmp_fid1, 2);
1192         if (rc != 0 && rc != -ENOENT)
1193                 GOTO(unlock_source, rc);
1194         /* save version of file name for replay, it must be ENOENT here */
1195         if (!req_is_replay(mdt_info_req(info))) {
1196                 if (rc != -ENOENT) {
1197                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1198                                PNAME(&rr->rr_name));
1199                         GOTO(unlock_source, rc = -EEXIST);
1200                 }
1201                 info->mti_ver[2] = ENOENT_VERSION;
1202                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1203         }
1204
1205         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1206                       mdt_object_child(ms), &rr->rr_name, ma);
1207
1208         if (rc == 0)
1209                 mdt_counter_incr(req, LPROC_MDT_LINK);
1210
1211         EXIT;
1212 unlock_source:
1213         mdt_object_unlock(info, ms, lhs, rc);
1214 unlock_parent:
1215         mdt_object_unlock(info, mp, lhp, rc);
1216 put_source:
1217         mdt_object_put(info->mti_env, ms);
1218 put_parent:
1219         mdt_object_put(info->mti_env, mp);
1220         return rc;
1221 }
1222 /**
1223  * lock the part of the directory according to the hash of the name
1224  * (lh->mlh_pdo_hash) in parallel directory lock.
1225  */
1226 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1227                               struct mdt_lock_handle *lh,
1228                               struct mdt_object *obj, __u64 ibits,
1229                               bool cos_incompat)
1230 {
1231         struct ldlm_res_id *res = &info->mti_res_id;
1232         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1233         union ldlm_policy_data *policy = &info->mti_policy;
1234         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1235         int rc;
1236
1237         /*
1238          * Finish res_id initializing by name hash marking part of
1239          * directory which is taking modification.
1240          */
1241         LASSERT(lh->mlh_pdo_hash != 0);
1242         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1243         memset(policy, 0, sizeof(*policy));
1244         policy->l_inodebits.bits = ibits;
1245         if (cos_incompat &&
1246             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1247                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1248         /*
1249          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1250          * going to be sent to client. If it is - mdt_intent_policy() path will
1251          * fix it up and turn FL_LOCAL flag off.
1252          */
1253         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1254                           policy, res, dlmflags,
1255                           &info->mti_exp->exp_handle.h_cookie);
1256         return rc;
1257 }
1258
1259 /**
1260  * Get BFL lock for rename or migrate process.
1261  **/
1262 static int mdt_rename_lock(struct mdt_thread_info *info,
1263                            struct lustre_handle *lh)
1264 {
1265         int     rc;
1266         ENTRY;
1267
1268         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1269                 struct lu_fid *fid = &info->mti_tmp_fid1;
1270                 struct mdt_object *obj;
1271
1272                 /* XXX, right now, it has to use object API to
1273                  * enqueue lock cross MDT, so it will enqueue
1274                  * rename lock(with LUSTRE_BFL_FID) by root object */
1275                 lu_root_fid(fid);
1276                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1277                 if (IS_ERR(obj))
1278                         RETURN(PTR_ERR(obj));
1279
1280                 rc = mdt_remote_object_lock(info, obj,
1281                                             &LUSTRE_BFL_FID, lh,
1282                                             LCK_EX,
1283                                             MDS_INODELOCK_UPDATE, false);
1284                 mdt_object_put(info->mti_env, obj);
1285         } else {
1286                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1287                 union ldlm_policy_data *policy = &info->mti_policy;
1288                 struct ldlm_res_id *res_id = &info->mti_res_id;
1289                 __u64 flags = 0;
1290
1291                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1292                 memset(policy, 0, sizeof *policy);
1293                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1294                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1295                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1296                                             LDLM_IBITS, policy, LCK_EX, &flags,
1297                                             ldlm_blocking_ast,
1298                                             ldlm_completion_ast, NULL, NULL, 0,
1299                                             LVB_T_NONE,
1300                                             &info->mti_exp->exp_handle.h_cookie,
1301                                             lh);
1302                 RETURN(rc);
1303         }
1304         RETURN(rc);
1305 }
1306
1307 static void mdt_rename_unlock(struct lustre_handle *lh)
1308 {
1309         ENTRY;
1310         LASSERT(lustre_handle_is_used(lh));
1311         /* Cancel the single rename lock right away */
1312         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1313         EXIT;
1314 }
1315
1316 static struct mdt_object *mdt_object_find_check(struct mdt_thread_info *info,
1317                                                 const struct lu_fid *fid,
1318                                                 int idx)
1319 {
1320         struct mdt_object *dir;
1321         int rc;
1322
1323         ENTRY;
1324
1325         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1326         if (IS_ERR(dir))
1327                 RETURN(dir);
1328
1329         /* check early, the real version will be saved after locking */
1330         rc = mdt_version_get_check(info, dir, idx);
1331         if (rc)
1332                 GOTO(out_put, rc);
1333
1334         RETURN(dir);
1335 out_put:
1336         mdt_object_put(info->mti_env, dir);
1337         return ERR_PTR(rc);
1338 }
1339
1340 /*
1341  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1342  * herein we don't really check it, just do revoke.
1343  */
1344 static int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1345                                          struct mdt_object *pobj,
1346                                          struct mdt_object *obj)
1347 {
1348         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1349         int rc;
1350
1351         mdt_lock_handle_init(lh);
1352         mdt_lock_reg_init(lh, LCK_EX);
1353
1354         if (mdt_object_remote(pobj)) {
1355                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1356                                             &lh->mlh_rreg_lh, LCK_EX,
1357                                             MDS_INODELOCK_LOOKUP, false);
1358         } else {
1359                 struct ldlm_res_id *res = &info->mti_res_id;
1360                 union ldlm_policy_data *policy = &info->mti_policy;
1361                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1362                                  LDLM_FL_COS_INCOMPAT;
1363
1364                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1365                 memset(policy, 0, sizeof(*policy));
1366                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1367                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1368                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1369                                   dlmflags, NULL);
1370         }
1371
1372         if (rc != ELDLM_OK)
1373                 return rc;
1374
1375         /*
1376          * TODO, currently we don't save this lock because there is no place to
1377          * hold this lock handle, but to avoid race we need to save this lock.
1378          */
1379         mdt_object_unlock(info, NULL, lh, 1);
1380
1381         return 0;
1382 }
1383
1384 /*
1385  * operation may takes locks of linkea, or directory stripes, group them in
1386  * different list.
1387  */
1388 struct mdt_sub_lock {
1389         struct mdt_object      *msl_obj;
1390         struct mdt_lock_handle  msl_lh;
1391         struct list_head        msl_linkage;
1392 };
1393
1394 static void mdt_unlock_list(struct mdt_thread_info *info,
1395                             struct list_head *list, int decref)
1396 {
1397         struct mdt_sub_lock *msl;
1398         struct mdt_sub_lock *tmp;
1399
1400         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1401                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1402                 list_del(&msl->msl_linkage);
1403                 OBD_FREE_PTR(msl);
1404         }
1405 }
1406
1407 /*
1408  * lock parents of links, and also check whether total locks don't exceed
1409  * RS_MAX_LOCKS.
1410  *
1411  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1412  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1413  * \retval      -ev negative errno upon error
1414  */
1415 static int mdt_lock_links(struct mdt_thread_info *info,
1416                           struct mdt_object *pobj,
1417                           const struct md_attr *ma,
1418                           struct mdt_object *obj,
1419                           struct list_head *link_locks)
1420 {
1421         struct mdt_device *mdt = info->mti_mdt;
1422         struct lu_buf *buf = &info->mti_big_buf;
1423         struct lu_name *lname = &info->mti_name;
1424         struct linkea_data ldata = { NULL };
1425         bool blocked = false;
1426         int retries = 5;
1427         int local_lnkp_cnt = 0;
1428         int rc;
1429
1430         ENTRY;
1431
1432         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1433                 RETURN(0);
1434
1435         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1436         if (buf->lb_buf == NULL)
1437                 RETURN(-ENOMEM);
1438
1439         ldata.ld_buf = buf;
1440         rc = mdt_links_read(info, obj, &ldata);
1441         if (rc) {
1442                 if (rc == -ENOENT || rc == -ENODATA)
1443                         rc = 0;
1444                 RETURN(rc);
1445         }
1446
1447 repeat:
1448         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1449              linkea_next_entry(&ldata)) {
1450                 struct mdt_object *lnkp;
1451                 struct mdt_sub_lock *msl;
1452                 struct lu_fid fid;
1453                 __u64 ibits;
1454
1455                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1456                                     &fid);
1457
1458                 /* check if it's also linked to parent */
1459                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1460                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1461                                PFID(&fid), PNAME(lname));
1462                         /* in case link is remote object, revoke LOOKUP lock */
1463                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1464                         continue;
1465                 }
1466
1467                 lnkp = NULL;
1468
1469                 /* check if it's linked to a stripe of parent */
1470                 if (ma->ma_valid & MA_LMV) {
1471                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1472                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1473                         int j = 0;
1474
1475                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1476                                 fid_le_to_cpu(stripe_fid,
1477                                               &lmv->lmv_stripe_fids[j]);
1478                                 if (lu_fid_eq(stripe_fid, &fid)) {
1479                                         CDEBUG(D_INFO, "skip stripe "DFID
1480                                                ", reovke "DNAME"\n",
1481                                                PFID(&fid), PNAME(lname));
1482                                         lnkp = mdt_object_find(info->mti_env,
1483                                                                mdt, &fid);
1484                                         if (IS_ERR(lnkp))
1485                                                 GOTO(out, rc = PTR_ERR(lnkp));
1486                                         break;
1487                                 }
1488                         }
1489
1490                         if (lnkp) {
1491                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1492                                                                    obj);
1493                                 mdt_object_put(info->mti_env, lnkp);
1494                                 continue;
1495                         }
1496                 }
1497
1498                 /* Check if it's already locked */
1499                 list_for_each_entry(msl, link_locks, msl_linkage) {
1500                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1501                                 CDEBUG(D_INFO,
1502                                        DFID" was locked, revoke "DNAME"\n",
1503                                        PFID(&fid), PNAME(lname));
1504                                 lnkp = msl->msl_obj;
1505                                 break;
1506                         }
1507                 }
1508
1509                 if (lnkp) {
1510                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1511                         continue;
1512                 }
1513
1514                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1515                        PFID(&fid), PNAME(lname));
1516
1517                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1518                 if (IS_ERR(lnkp)) {
1519                         CWARN("%s: cannot find obj "DFID": %ld\n",
1520                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1521                         continue;
1522                 }
1523
1524                 if (!mdt_object_exists(lnkp)) {
1525                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1526                               PFID(&fid), PNAME(lname));
1527                         mdt_object_put(info->mti_env, lnkp);
1528                         continue;
1529                 }
1530
1531                 if (!mdt_object_remote(lnkp))
1532                         local_lnkp_cnt++;
1533
1534                 OBD_ALLOC_PTR(msl);
1535                 if (msl == NULL)
1536                         GOTO(out, rc = -ENOMEM);
1537
1538                 /*
1539                  * we can't follow parent-child lock order like other MD
1540                  * operations, use lock_try here to avoid deadlock, if the lock
1541                  * cannot be taken, drop all locks taken, revoke the blocked
1542                  * one, and continue processing the remaining entries, and in
1543                  * the end of the loop restart from beginning.
1544                  */
1545                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1546                 ibits = 0;
1547                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1548                                          MDS_INODELOCK_UPDATE, true);
1549                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1550                         blocked = true;
1551
1552                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME" retry %d\n",
1553                                PFID(&fid), PNAME(lname), retries);
1554
1555                         mdt_unlock_list(info, link_locks, 1);
1556
1557                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1558                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1559                                              MDS_INODELOCK_UPDATE);
1560                         if (rc) {
1561                                 mdt_object_put(info->mti_env, lnkp);
1562                                 OBD_FREE_PTR(msl);
1563                                 GOTO(out, rc);
1564                         }
1565
1566                         if (mdt_object_remote(lnkp)) {
1567                                 struct ldlm_lock *lock;
1568
1569                                 /*
1570                                  * for remote object, set lock cb_atomic,
1571                                  * so lock can be released in blocking_ast()
1572                                  * immediately, then the next lock_try will
1573                                  * have better chance of success.
1574                                  */
1575                                 lock = ldlm_handle2lock(
1576                                                 &msl->msl_lh.mlh_rreg_lh);
1577                                 LASSERT(lock != NULL);
1578                                 lock_res_and_lock(lock);
1579                                 ldlm_set_atomic_cb(lock);
1580                                 unlock_res_and_lock(lock);
1581                                 LDLM_LOCK_PUT(lock);
1582                         }
1583
1584                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1585                         OBD_FREE_PTR(msl);
1586                         continue;
1587                 }
1588
1589                 INIT_LIST_HEAD(&msl->msl_linkage);
1590                 msl->msl_obj = lnkp;
1591                 list_add_tail(&msl->msl_linkage, link_locks);
1592
1593                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1594         }
1595
1596         if (blocked) {
1597                 rc = -EBUSY;
1598                 if (--retries > 0) {
1599                         mdt_unlock_list(info, link_locks, rc);
1600                         blocked = false;
1601                         local_lnkp_cnt = 0;
1602                         goto repeat;
1603                 }
1604         }
1605
1606         EXIT;
1607 out:
1608         if (rc)
1609                 mdt_unlock_list(info, link_locks, rc);
1610         else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1611                 /*
1612                  * parent may have 3 local objects: master object and 2 stripes
1613                  * (if it's being migrated too); source may have 2 local
1614                  * objects: master and 1 stripe; target has 1 local object.
1615                  */
1616                 rc = 1;
1617         return rc;
1618 }
1619
1620 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1621                                   struct mdt_object *obj,
1622                                   const struct md_attr *ma,
1623                                   struct list_head *slave_locks)
1624 {
1625         struct mdt_device *mdt = info->mti_mdt;
1626         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1627         struct lu_fid *fid = &info->mti_tmp_fid1;
1628         struct mdt_object *slave;
1629         struct mdt_sub_lock *msl;
1630         int i;
1631         int rc;
1632
1633         ENTRY;
1634
1635         LASSERT(mdt_object_remote(obj));
1636         LASSERT(ma->ma_valid & MA_LMV);
1637         LASSERT(lmv);
1638
1639         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1640                 RETURN(-EINVAL);
1641
1642         if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1643                 RETURN(0);
1644
1645         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1646                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1647
1648                 slave = mdt_object_find(info->mti_env, mdt, fid);
1649                 if (IS_ERR(slave))
1650                         GOTO(out, rc = PTR_ERR(slave));
1651
1652                 OBD_ALLOC_PTR(msl);
1653                 if (!msl) {
1654                         mdt_object_put(info->mti_env, slave);
1655                         GOTO(out, rc = -ENOMEM);
1656                 }
1657
1658                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1659                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1660                                            MDS_INODELOCK_UPDATE, true);
1661                 if (rc) {
1662                         OBD_FREE_PTR(msl);
1663                         mdt_object_put(info->mti_env, slave);
1664                         GOTO(out, rc);
1665                 }
1666
1667                 INIT_LIST_HEAD(&msl->msl_linkage);
1668                 msl->msl_obj = slave;
1669                 list_add_tail(&msl->msl_linkage, slave_locks);
1670
1671         }
1672         EXIT;
1673
1674 out:
1675         if (rc)
1676                 mdt_unlock_list(info, slave_locks, rc);
1677         return rc;
1678 }
1679
1680 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1681                                              struct mdt_object *obj,
1682                                              struct mdt_lock_handle *lh,
1683                                              struct ldlm_enqueue_info *einfo,
1684                                              struct list_head *slave_locks,
1685                                              int decref)
1686 {
1687         if (mdt_object_remote(obj)) {
1688                 mdt_unlock_list(info, slave_locks, decref);
1689                 mdt_object_unlock(info, obj, lh, decref);
1690         } else {
1691                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1692         }
1693 }
1694
1695 /* lock parent and its stripes */
1696 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1697                                    struct mdt_object *obj,
1698                                    const struct md_attr *ma,
1699                                    struct mdt_lock_handle *lh,
1700                                    struct ldlm_enqueue_info *einfo,
1701                                    struct list_head *slave_locks)
1702 {
1703         int rc;
1704
1705         if (mdt_object_remote(obj)) {
1706                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1707                                             &lh->mlh_rreg_lh, LCK_PW,
1708                                             MDS_INODELOCK_UPDATE, false);
1709                 if (rc != ELDLM_OK)
1710                         return rc;
1711
1712                 /*
1713                  * if obj is remote and striped, lock its stripes explicitly
1714                  * because it's not striped in LOD layer on this MDT.
1715                  */
1716                 if (ma->ma_valid & MA_LMV) {
1717                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1718                         if (rc)
1719                                 mdt_object_unlock(info, obj, lh, rc);
1720                 }
1721         } else {
1722                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1723                                             einfo, true);
1724         }
1725
1726         return rc;
1727 }
1728
1729 /*
1730  * in migration, object may be remote, and we need take full lock of it and its
1731  * stripes if it's directory, besides, object may be a remote object on its
1732  * parent, revoke its LOOKUP lock on where its parent is located.
1733  */
1734 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1735                                    struct mdt_object *pobj,
1736                                    struct mdt_object *obj,
1737                                    struct mdt_lock_handle *lh,
1738                                    struct ldlm_enqueue_info *einfo,
1739                                    struct list_head *slave_locks)
1740 {
1741         int rc;
1742
1743         if (mdt_object_remote(obj)) {
1744                 /* don't bother to check if pobj and obj are on the same MDT. */
1745                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1746                 if (rc)
1747                         return rc;
1748
1749                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1750                                             &lh->mlh_rreg_lh, LCK_EX,
1751                                             MDS_INODELOCK_FULL, false);
1752                 if (rc != ELDLM_OK)
1753                         return rc;
1754
1755                 /*
1756                  * if obj is remote and striped, lock its stripes explicitly
1757                  * because it's not striped in LOD layer on this MDT.
1758                  */
1759                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1760                         struct md_attr *ma = &info->mti_attr;
1761
1762                         ma->ma_lmv = info->mti_big_lmm;
1763                         ma->ma_lmv_size = info->mti_big_lmmsize;
1764                         ma->ma_valid = 0;
1765                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1766                         if (rc) {
1767                                 mdt_object_unlock(info, obj, lh, rc);
1768                                 return rc;
1769                         }
1770
1771                         if (ma->ma_valid & MA_LMV) {
1772                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1773                                                             slave_locks);
1774                                 if (rc)
1775                                         mdt_object_unlock(info, obj, lh, rc);
1776                         }
1777                 }
1778         } else {
1779                 if (mdt_object_remote(pobj)) {
1780                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1781                         if (rc)
1782                                 return rc;
1783                 }
1784
1785                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1786                                             einfo, true);
1787         }
1788
1789         return rc;
1790 }
1791
1792 /*
1793  * lookup source by name, if parent is striped directory, we need to find the
1794  * corresponding stripe where source is located, and then lookup there.
1795  *
1796  * besides, if parent is migrating too, and file is already in target stripe,
1797  * this should be a redo of 'lfs migrate' on client side.
1798  */
1799 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1800                               struct mdt_object *pobj,
1801                               const struct md_attr *ma,
1802                               const struct lu_name *lname,
1803                               struct mdt_object **spobj,
1804                               struct mdt_object **sobj)
1805 {
1806         const struct lu_env *env = info->mti_env;
1807         struct lu_fid *fid = &info->mti_tmp_fid1;
1808         struct mdt_object *stripe;
1809         int rc;
1810
1811         if (ma->ma_valid & MA_LMV) {
1812                 /* if parent is striped, lookup on corresponding stripe */
1813                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1814                 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1815                 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1816                 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1817                                     LMV_HASH_FLAG_MIGRATION;
1818
1819                 if (is_migrating) {
1820                         hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1821                         stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1822                 }
1823
1824                 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1825                                               lname->ln_name,
1826                                               lname->ln_namelen);
1827                 if (rc < 0)
1828                         return rc;
1829
1830                 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1831                         rc += le32_to_cpu(lmv->lmv_migrate_offset);
1832
1833                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1834
1835                 stripe = mdt_object_find(env, info->mti_mdt, fid);
1836                 if (IS_ERR(stripe))
1837                         return PTR_ERR(stripe);
1838
1839                 fid_zero(fid);
1840                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1841                                 &info->mti_spec);
1842                 if (rc == -ENOENT && is_migrating) {
1843                         /*
1844                          * if parent is migrating, and lookup child failed on
1845                          * source stripe, lookup again on target stripe, if it
1846                          * exists, it means previous migration was interrupted,
1847                          * and current file was migrated already.
1848                          */
1849                         mdt_object_put(env, stripe);
1850
1851                         hash_type = le32_to_cpu(lmv->lmv_hash_type);
1852                         stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1853
1854                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1855                                                       lname->ln_name,
1856                                                       lname->ln_namelen);
1857                         if (rc < 0)
1858                                 return rc;
1859
1860                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1861
1862                         stripe = mdt_object_find(env, info->mti_mdt, fid);
1863                         if (IS_ERR(stripe))
1864                                 return PTR_ERR(stripe);
1865
1866                         fid_zero(fid);
1867                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1868                                         fid, &info->mti_spec);
1869                         mdt_object_put(env, stripe);
1870                         return rc ?: -EALREADY;
1871                 } else if (rc) {
1872                         mdt_object_put(env, stripe);
1873                         return rc;
1874                 }
1875         } else {
1876                 fid_zero(fid);
1877                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1878                                 &info->mti_spec);
1879                 if (rc)
1880                         return rc;
1881
1882                 stripe = pobj;
1883                 mdt_object_get(env, stripe);
1884         }
1885
1886         *spobj = stripe;
1887
1888         *sobj = mdt_object_find(env, info->mti_mdt, fid);
1889         if (IS_ERR(*sobj)) {
1890                 mdt_object_put(env, stripe);
1891                 rc = PTR_ERR(*sobj);
1892                 *spobj = NULL;
1893                 *sobj = NULL;
1894         }
1895
1896         return rc;
1897 }
1898
1899 /* end lease and close file for regular file */
1900 static int mdd_migrate_close(struct mdt_thread_info *info,
1901                              struct mdt_object *obj)
1902 {
1903         struct close_data *data;
1904         struct mdt_body *repbody;
1905         struct ldlm_lock *lease;
1906         int rc;
1907         int rc2;
1908
1909         rc = -EPROTO;
1910         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1911                                       RCL_CLIENT) ||
1912             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1913                                       RCL_CLIENT))
1914                 goto close;
1915
1916         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1917         if (!data)
1918                 goto close;
1919
1920         rc = -ESTALE;
1921         lease = ldlm_handle2lock(&data->cd_handle);
1922         if (!lease)
1923                 goto close;
1924
1925         /* check if the lease was already canceled */
1926         lock_res_and_lock(lease);
1927         rc = ldlm_is_cancel(lease);
1928         unlock_res_and_lock(lease);
1929
1930         if (rc) {
1931                 rc = -EAGAIN;
1932                 LDLM_DEBUG(lease, DFID" lease broken",
1933                            PFID(mdt_object_fid(obj)));
1934         }
1935
1936         /*
1937          * cancel server side lease, client side counterpart should have been
1938          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1939          */
1940         ldlm_lock_cancel(lease);
1941         ldlm_reprocess_all(lease->l_resource);
1942         LDLM_LOCK_PUT(lease);
1943
1944 close:
1945         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1946         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1947         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1948
1949         return rc ?: rc2;
1950 }
1951
1952 /*
1953  * migrate file in below steps:
1954  *  1. lock parent and its stripes
1955  *  2. lookup source by name
1956  *  3. lock parents of source links if source is not directory
1957  *  4. reject if source is in HSM
1958  *  5. take source open_sem and close file if source is regular file
1959  *  6. lock source and its stripes if it's directory
1960  *  7. lock target so subsequent change to it can trigger COS
1961  *  8. migrate file
1962  *  9. unlock above locks
1963  * 10. sync device if source has links
1964  */
1965 static int mdt_reint_migrate_internal(struct mdt_thread_info *info)
1966 {
1967         const struct lu_env *env = info->mti_env;
1968         struct mdt_device *mdt = info->mti_mdt;
1969         struct mdt_reint_record *rr = &info->mti_rr;
1970         struct md_attr *ma = &info->mti_attr;
1971         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1972         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1973         struct mdt_object *pobj;
1974         struct mdt_object *spobj = NULL;
1975         struct mdt_object *sobj = NULL;
1976         struct mdt_object *tobj;
1977         struct mdt_lock_handle *lhp;
1978         struct mdt_lock_handle *lhs;
1979         struct mdt_lock_handle *lht;
1980         LIST_HEAD(parent_slave_locks);
1981         LIST_HEAD(child_slave_locks);
1982         LIST_HEAD(link_locks);
1983         bool open_sem_locked = false;
1984         bool do_sync = false;
1985         int rc;
1986         ENTRY;
1987
1988         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1989                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1990
1991         /* don't allow migrate . or .. */
1992         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
1993                 RETURN(-EBUSY);
1994
1995         rc = mdt_remote_dir_permission(info);
1996         if (rc)
1997                 RETURN(rc);
1998
1999         /* pobj is master object of parent */
2000         pobj = mdt_object_find_check(info, rr->rr_fid1, 0);
2001         if (IS_ERR(pobj))
2002                 RETURN(PTR_ERR(pobj));
2003
2004         if (unlikely(!info->mti_big_lmm)) {
2005                 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
2006                 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
2007                 if (!info->mti_big_lmm)
2008                         GOTO(put_parent, rc = -ENOMEM);
2009         }
2010
2011         ma->ma_lmv = info->mti_big_lmm;
2012         ma->ma_lmv_size = info->mti_big_lmmsize;
2013         ma->ma_valid = 0;
2014         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2015         if (rc)
2016                 GOTO(put_parent, rc);
2017
2018         /* lock parent object */
2019         lhp = &info->mti_lh[MDT_LH_PARENT];
2020         mdt_lock_reg_init(lhp, LCK_PW);
2021         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2022                                      &parent_slave_locks);
2023         if (rc)
2024                 GOTO(put_parent, rc);
2025
2026         /*
2027          * spobj is the corresponding stripe against name if pobj is striped
2028          * directory, which is the real parent, and no need to lock, because
2029          * we've taken full lock of pobj.
2030          */
2031         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2032         if (rc)
2033                 GOTO(unlock_parent, rc);
2034
2035         /* lock parents of source links, and revoke LOOKUP lock of links */
2036         rc = mdt_lock_links(info, pobj, ma, sobj, &link_locks);
2037         if (rc < 0)
2038                 GOTO(put_source, rc);
2039
2040         /*
2041          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2042          * with one request, if total lock count exceeds this limit, we will
2043          * drop all locks after migration, and synchronous device in the end.
2044          */
2045         do_sync = rc;
2046
2047         /* TODO: DoM migration is not supported yet */
2048         if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2049                 ma->ma_lmm = info->mti_big_lmm;
2050                 ma->ma_lmm_size = info->mti_big_lmmsize;
2051                 ma->ma_valid = 0;
2052                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2053                 if (rc)
2054                         GOTO(put_source, rc);
2055
2056                 if (ma->ma_valid & MA_LOV &&
2057                     mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2058                         GOTO(put_source, rc = -EOPNOTSUPP);
2059         }
2060
2061         /* if migration HSM is allowed */
2062         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2063                 ma->ma_need = MA_HSM;
2064                 ma->ma_valid = 0;
2065                 rc = mdt_attr_get_complex(info, sobj, ma);
2066                 if (rc)
2067                         GOTO(unlock_links, rc);
2068
2069                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2070                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2071         }
2072
2073         /* end lease and close file for regular file */
2074         if (info->mti_spec.sp_migrate_close) {
2075                 /* try to hold open_sem so that nobody else can open the file */
2076                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2077                         /* close anyway */
2078                         mdd_migrate_close(info, sobj);
2079                         GOTO(unlock_links, rc = -EBUSY);
2080                 } else {
2081                         open_sem_locked = true;
2082                         rc = mdd_migrate_close(info, sobj);
2083                         if (rc)
2084                                 GOTO(unlock_open_sem, rc);
2085                 }
2086         }
2087
2088         /* lock source */
2089         lhs = &info->mti_lh[MDT_LH_OLD];
2090         mdt_lock_reg_init(lhs, LCK_EX);
2091         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2092                                      &child_slave_locks);
2093         if (rc)
2094                 GOTO(unlock_open_sem, rc);
2095
2096         /* lock target */
2097         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2098         if (IS_ERR(tobj))
2099                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2100
2101         lht = &info->mti_lh[MDT_LH_NEW];
2102         mdt_lock_reg_init(lht, LCK_EX);
2103         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2104         if (rc)
2105                 GOTO(put_target, rc);
2106
2107         /* Don't do lookup sanity check. We know name doesn't exist. */
2108         info->mti_spec.sp_cr_lookup = 0;
2109         info->mti_spec.sp_feat = &dt_directory_features;
2110
2111         rc = mdo_migrate(env, mdt_object_child(pobj),
2112                          mdt_object_child(sobj), &rr->rr_name,
2113                          mdt_object_child(tobj), &info->mti_spec, ma);
2114         EXIT;
2115
2116         mdt_object_unlock(info, tobj, lht, rc);
2117 put_target:
2118         mdt_object_put(env, tobj);
2119 unlock_source:
2120         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2121                                   &child_slave_locks, rc);
2122 unlock_open_sem:
2123         if (open_sem_locked)
2124                 up_write(&sobj->mot_open_sem);
2125 unlock_links:
2126         mdt_unlock_list(info, &link_locks, rc);
2127 put_source:
2128         mdt_object_put(env, sobj);
2129         mdt_object_put(env, spobj);
2130 unlock_parent:
2131         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2132                                   &parent_slave_locks, rc);
2133 put_parent:
2134         mdt_object_put(env, pobj);
2135
2136         if (!rc && do_sync)
2137                 mdt_device_sync(env, mdt);
2138
2139         return rc;
2140 }
2141
2142 static int mdt_object_lock_save(struct mdt_thread_info *info,
2143                                 struct mdt_object *dir,
2144                                 struct mdt_lock_handle *lh,
2145                                 int idx, bool cos_incompat)
2146 {
2147         int rc;
2148
2149         /* we lock the target dir if it is local */
2150         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2151                                    cos_incompat);
2152         if (rc != 0)
2153                 return rc;
2154
2155         /* get and save correct version after locking */
2156         mdt_version_get_save(info, dir, idx);
2157         return 0;
2158 }
2159
2160 /*
2161  * determine lock order of sobj and tobj
2162  *
2163  * there are two situations we need to lock tobj before sobj:
2164  * 1. sobj is child of tobj
2165  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2166  *    larger than that of tobj
2167  *
2168  * \retval      1 lock tobj before sobj
2169  * \retval      0 lock sobj before tobj
2170  * \retval      -ev negative errno upon error
2171  */
2172 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2173                                            struct mdt_object *sobj,
2174                                            struct mdt_object *tobj)
2175 {
2176         struct md_attr *ma = &info->mti_attr;
2177         struct lu_fid *spfid = &info->mti_tmp_fid1;
2178         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2179         struct lmv_mds_md_v1 *lmv;
2180         __u32 sindex;
2181         __u32 tindex;
2182         int rc;
2183
2184         /* sobj and tobj are the same */
2185         if (sobj == tobj)
2186                 return 0;
2187
2188         if (fid_is_root(mdt_object_fid(sobj)))
2189                 return 0;
2190
2191         if (fid_is_root(mdt_object_fid(tobj)))
2192                 return 1;
2193
2194         /* check whether sobj is child of tobj */
2195         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2196                            mdt_object_fid(tobj));
2197         if (rc < 0)
2198                 return rc;
2199
2200         if (rc == 1)
2201                 return 1;
2202
2203         /* check whether sobj and tobj are children of the same parent */
2204         rc = mdt_attr_get_pfid(info, sobj, spfid);
2205         if (rc)
2206                 return rc;
2207
2208         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2209         if (rc)
2210                 return rc;
2211
2212         if (!lu_fid_eq(spfid, tpfid))
2213                 return 0;
2214
2215         /* check whether sobj and tobj are sibling stripes */
2216         ma->ma_need = MA_LMV;
2217         ma->ma_valid = 0;
2218         ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2219         ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2220         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2221         if (rc)
2222                 return rc;
2223
2224         if (!(ma->ma_valid & MA_LMV))
2225                 return 0;
2226
2227         lmv = &ma->ma_lmv->lmv_md_v1;
2228         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2229                 return 0;
2230         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2231
2232         ma->ma_valid = 0;
2233         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2234         if (rc)
2235                 return rc;
2236
2237         if (!(ma->ma_valid & MA_LMV))
2238                 return -ENODATA;
2239
2240         lmv = &ma->ma_lmv->lmv_md_v1;
2241         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2242                 return -EINVAL;
2243         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2244
2245         /* check stripe index of sobj and tobj */
2246         if (sindex == tindex)
2247                 return -EINVAL;
2248
2249         return sindex < tindex ? 0 : 1;
2250 }
2251
2252 /*
2253  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2254  * 2 - srcdir child; 3 - tgtdir child.
2255  * Update on disk version of srcdir child.
2256  */
2257 /**
2258  * For DNE phase I, only these renames are allowed
2259  *      mv src_p/src_c tgt_p/tgt_c
2260  * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
2261  * 2. src_p and tgt_p are same directory, and tgt_c does not
2262  *    exists. In this case, all of modification will happen
2263  *    in the MDT where ithesource parent is, only one remote
2264  *    update is needed, i.e. set c_time/m_time on the child.
2265  *    And tgt_c will be still in the same MDT as the original
2266  *    src_c.
2267  */
2268 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
2269                                      struct mdt_lock_handle *lhc)
2270 {
2271         struct mdt_reint_record *rr = &info->mti_rr;
2272         struct md_attr *ma = &info->mti_attr;
2273         struct ptlrpc_request *req = mdt_info_req(info);
2274         struct mdt_object *msrcdir = NULL;
2275         struct mdt_object *mtgtdir = NULL;
2276         struct mdt_object *mold;
2277         struct mdt_object *mnew = NULL;
2278         struct mdt_lock_handle *lh_srcdirp;
2279         struct mdt_lock_handle *lh_tgtdirp;
2280         struct mdt_lock_handle *lh_oldp = NULL;
2281         struct mdt_lock_handle *lh_newp = NULL;
2282         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2283         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2284         __u64 lock_ibits;
2285         bool reverse = false;
2286         bool cos_incompat, discard = false;
2287         int rc;
2288         ENTRY;
2289
2290         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2291                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2292                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2293
2294         /* find both parents. */
2295         msrcdir = mdt_object_find_check(info, rr->rr_fid1, 0);
2296         if (IS_ERR(msrcdir))
2297                 RETURN(PTR_ERR(msrcdir));
2298
2299         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2300
2301         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2302                 mtgtdir = msrcdir;
2303                 mdt_object_get(info->mti_env, mtgtdir);
2304         } else {
2305                 mtgtdir = mdt_object_find_check(info, rr->rr_fid2, 1);
2306                 if (IS_ERR(mtgtdir))
2307                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2308         }
2309
2310         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2311         if (rc < 0)
2312                 GOTO(out_put_tgtdir, rc);
2313
2314         reverse = rc;
2315
2316         /* source needs to be looked up after locking source parent, otherwise
2317          * this rename may race with unlink source, and cause rename hang, see
2318          * sanityn.sh 55b, so check parents first, if later we found source is
2319          * remote, relock parents. */
2320         cos_incompat = (mdt_object_remote(msrcdir) ||
2321                         mdt_object_remote(mtgtdir));
2322
2323         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2324
2325         /* lock parents in the proper order. */
2326         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2327         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2328
2329 relock:
2330         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2331         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2332
2333         if (reverse) {
2334                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2335                                           cos_incompat);
2336                 if (rc)
2337                         GOTO(out_put_tgtdir, rc);
2338
2339                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2340
2341                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2342                                           cos_incompat);
2343                 if (rc != 0) {
2344                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2345                         GOTO(out_put_tgtdir, rc);
2346                 }
2347         } else {
2348                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2349                                           cos_incompat);
2350                 if (rc)
2351                         GOTO(out_put_tgtdir, rc);
2352
2353                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2354
2355                 if (mtgtdir != msrcdir) {
2356                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2357                                                   cos_incompat);
2358                 } else if (lh_srcdirp->mlh_pdo_hash !=
2359                            lh_tgtdirp->mlh_pdo_hash) {
2360                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2361                                                 MDS_INODELOCK_UPDATE,
2362                                                 cos_incompat);
2363                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2364                 }
2365                 if (rc != 0) {
2366                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2367                         GOTO(out_put_tgtdir, rc);
2368                 }
2369         }
2370
2371         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2372         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2373
2374         /* find mold object. */
2375         fid_zero(old_fid);
2376         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2377         if (rc != 0)
2378                 GOTO(out_unlock_parents, rc);
2379
2380         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2381                 GOTO(out_unlock_parents, rc = -EINVAL);
2382
2383         if (!fid_is_md_operative(old_fid))
2384                 GOTO(out_unlock_parents, rc = -EPERM);
2385
2386         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2387         if (IS_ERR(mold))
2388                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2389
2390         if (!mdt_object_exists(mold)) {
2391                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2392                                 &mold->mot_obj,
2393                                 "object does not exist");
2394                 GOTO(out_put_old, rc = -ENOENT);
2395         }
2396
2397         /* Check if @mtgtdir is subdir of @mold, before locking child
2398          * to avoid reverse locking. */
2399         if (mtgtdir != msrcdir) {
2400                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2401                                    old_fid);
2402                 if (rc) {
2403                         if (rc == 1)
2404                                 rc = -EINVAL;
2405                         GOTO(out_put_old, rc);
2406                 }
2407         }
2408
2409         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2410         /* save version after locking */
2411         mdt_version_get_save(info, mold, 2);
2412
2413         if (!cos_incompat && mdt_object_remote(mold)) {
2414                 cos_incompat = true;
2415                 mdt_object_put(info->mti_env, mold);
2416                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2417                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2418                 goto relock;
2419         }
2420
2421         /* find mnew object:
2422          * mnew target object may not exist now
2423          * lookup with version checking */
2424         fid_zero(new_fid);
2425         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2426                                       3);
2427         if (rc == 0) {
2428                 /* the new_fid should have been filled at this moment */
2429                 if (lu_fid_eq(old_fid, new_fid))
2430                         GOTO(out_put_old, rc);
2431
2432                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2433                     lu_fid_eq(new_fid, rr->rr_fid2))
2434                         GOTO(out_put_old, rc = -EINVAL);
2435
2436                 if (!fid_is_md_operative(new_fid))
2437                         GOTO(out_put_old, rc = -EPERM);
2438
2439                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2440                 if (IS_ERR(mnew))
2441                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2442
2443                 if (!mdt_object_exists(mnew)) {
2444                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2445                                         &mnew->mot_obj,
2446                                         "object does not exist");
2447                         GOTO(out_put_new, rc = -ENOENT);
2448                 }
2449
2450                 if (mdt_object_remote(mnew)) {
2451                         struct mdt_body  *repbody;
2452
2453                         /* Always send rename req to the target child MDT */
2454                         repbody = req_capsule_server_get(info->mti_pill,
2455                                                          &RMF_MDT_BODY);
2456                         LASSERT(repbody != NULL);
2457                         repbody->mbo_fid1 = *new_fid;
2458                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2459                         GOTO(out_put_new, rc = -EXDEV);
2460                 }
2461                 /* Before locking the target dir, check we do not replace
2462                  * a dir with a non-dir, otherwise it may deadlock with
2463                  * link op which tries to create a link in this dir
2464                  * back to this non-dir. */
2465                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2466                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2467                         GOTO(out_put_new, rc = -EISDIR);
2468
2469                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2470                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2471                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2472                 if (mdt_object_remote(msrcdir)) {
2473                         /* Enqueue lookup lock from the parent MDT */
2474                         rc = mdt_remote_object_lock(info, msrcdir,
2475                                                     mdt_object_fid(mold),
2476                                                     &lh_oldp->mlh_rreg_lh,
2477                                                     lh_oldp->mlh_rreg_mode,
2478                                                     MDS_INODELOCK_LOOKUP,
2479                                                     false);
2480                         if (rc != ELDLM_OK)
2481                                 GOTO(out_put_new, rc);
2482
2483                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2484                 }
2485
2486                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2487                                            cos_incompat);
2488                 if (rc != 0)
2489                         GOTO(out_unlock_old, rc);
2490
2491                 /* Check if @msrcdir is subdir of @mnew, before locking child
2492                  * to avoid reverse locking. */
2493                 if (mtgtdir != msrcdir) {
2494                         rc = mdo_is_subdir(info->mti_env,
2495                                            mdt_object_child(msrcdir), new_fid);
2496                         if (rc) {
2497                                 if (rc == 1)
2498                                         rc = -EINVAL;
2499                                 GOTO(out_unlock_old, rc);
2500                         }
2501                 }
2502
2503                 /* We used to acquire MDS_INODELOCK_FULL here but we
2504                  * can't do this now because a running HSM restore on
2505                  * the rename onto victim will hold the layout
2506                  * lock. See LU-4002. */
2507
2508                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2509                 mdt_lock_reg_init(lh_newp, LCK_EX);
2510                 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2511                                            MDS_INODELOCK_LOOKUP |
2512                                            MDS_INODELOCK_UPDATE,
2513                                            cos_incompat);
2514                 if (rc != 0)
2515                         GOTO(out_unlock_old, rc);
2516
2517                 /* get and save version after locking */
2518                 mdt_version_get_save(info, mnew, 3);
2519         } else if (rc != -EREMOTE && rc != -ENOENT) {
2520                 GOTO(out_put_old, rc);
2521         } else {
2522                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2523                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2524                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2525                 if (mdt_object_remote(msrcdir)) {
2526                         /* Enqueue lookup lock from the parent MDT */
2527                         rc = mdt_remote_object_lock(info, msrcdir,
2528                                                     mdt_object_fid(mold),
2529                                                     &lh_oldp->mlh_rreg_lh,
2530                                                     lh_oldp->mlh_rreg_mode,
2531                                                     MDS_INODELOCK_LOOKUP,
2532                                                     false);
2533                         if (rc != ELDLM_OK)
2534                                 GOTO(out_put_old, rc);
2535
2536                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2537                 }
2538
2539                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2540                                            cos_incompat);
2541                 if (rc != 0)
2542                         GOTO(out_unlock_old, rc);
2543
2544                 mdt_enoent_version_save(info, 3);
2545         }
2546
2547         /* step 5: rename it */
2548         mdt_reint_init_ma(info, ma);
2549
2550         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2551                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2552
2553         if (mnew != NULL)
2554                 mutex_lock(&mnew->mot_lov_mutex);
2555
2556         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2557                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2558                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2559                         &rr->rr_tgt_name, ma);
2560
2561         if (mnew != NULL)
2562                 mutex_unlock(&mnew->mot_lov_mutex);
2563
2564         /* handle last link of tgt object */
2565         if (rc == 0) {
2566                 mdt_counter_incr(req, LPROC_MDT_RENAME);
2567                 if (mnew) {
2568                         mdt_handle_last_unlink(info, mnew, ma);
2569                         discard = mdt_dom_check_for_discard(info, mnew);
2570                 }
2571
2572                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2573                                          msrcdir, mtgtdir);
2574         }
2575
2576         EXIT;
2577         if (mnew != NULL)
2578                 mdt_object_unlock(info, mnew, lh_newp, rc);
2579 out_unlock_old:
2580         mdt_object_unlock(info, mold, lh_oldp, rc);
2581 out_put_new:
2582         if (mnew != NULL)
2583                 mdt_object_put(info->mti_env, mnew);
2584 out_put_old:
2585         mdt_object_put(info->mti_env, mold);
2586 out_unlock_parents:
2587         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2588         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2589 out_put_tgtdir:
2590         mdt_object_put(info->mti_env, mtgtdir);
2591 out_put_srcdir:
2592         mdt_object_put(info->mti_env, msrcdir);
2593
2594         /* If 'discard' is set then new_fid must exits.
2595          * DOM data discard need neither object nor lock,
2596          * so do this at the end.
2597          */
2598         if (discard)
2599                 mdt_dom_discard_data(info, new_fid);
2600
2601         return rc;
2602 }
2603
2604 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2605                                        struct mdt_lock_handle *lhc, bool rename)
2606 {
2607         struct mdt_reint_record *rr = &info->mti_rr;
2608         struct ptlrpc_request   *req = mdt_info_req(info);
2609         struct lustre_handle    rename_lh = { 0 };
2610         int                     rc;
2611         ENTRY;
2612
2613         if (info->mti_dlm_req)
2614                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2615
2616         if (!fid_is_md_operative(rr->rr_fid1) ||
2617             !fid_is_md_operative(rr->rr_fid2))
2618                 RETURN(-EPERM);
2619
2620         /* Note: do not enqueue rename lock for replay request, because
2621          * if other MDT holds rename lock, but being blocked to wait for
2622          * this MDT to finish its recovery, and the failover MDT can not
2623          * get rename lock, which will cause deadlock. */
2624         if (!req_is_replay(req)) {
2625                 rc = mdt_rename_lock(info, &rename_lh);
2626                 if (rc != 0) {
2627                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2628                                mdt_obd_name(info->mti_mdt), rc);
2629                         RETURN(rc);
2630                 }
2631         }
2632
2633         if (rename)
2634                 rc = mdt_reint_rename_internal(info, lhc);
2635         else
2636                 rc = mdt_reint_migrate_internal(info);
2637
2638         if (lustre_handle_is_used(&rename_lh))
2639                 mdt_rename_unlock(&rename_lh);
2640
2641         RETURN(rc);
2642 }
2643
2644 static int mdt_reint_rename(struct mdt_thread_info *info,
2645                             struct mdt_lock_handle *lhc)
2646 {
2647         return mdt_reint_rename_or_migrate(info, lhc, true);
2648 }
2649
2650 static int mdt_reint_migrate(struct mdt_thread_info *info,
2651                             struct mdt_lock_handle *lhc)
2652 {
2653         return mdt_reint_rename_or_migrate(info, lhc, false);
2654 }
2655
2656 static int mdt_reint_resync(struct mdt_thread_info *info,
2657                             struct mdt_lock_handle *lhc)
2658 {
2659         struct mdt_reint_record *rr = &info->mti_rr;
2660         struct ptlrpc_request   *req = mdt_info_req(info);
2661         struct md_attr          *ma = &info->mti_attr;
2662         struct mdt_object       *mo;
2663         struct ldlm_lock        *lease;
2664         struct mdt_body         *repbody;
2665         struct md_layout_change  layout = { .mlc_mirror_id = rr->rr_mirror_id };
2666         bool                     lease_broken;
2667         int                      rc, rc2;
2668         ENTRY;
2669
2670         DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2671
2672         if (info->mti_dlm_req)
2673                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2674
2675         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2676         if (IS_ERR(mo))
2677                 GOTO(out, rc = PTR_ERR(mo));
2678
2679         if (!mdt_object_exists(mo))
2680                 GOTO(out_obj, rc = -ENOENT);
2681
2682         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2683                 GOTO(out_obj, rc = -EINVAL);
2684
2685         if (mdt_object_remote(mo))
2686                 GOTO(out_obj, rc = -EREMOTE);
2687
2688         lease = ldlm_handle2lock(rr->rr_lease_handle);
2689         if (lease == NULL)
2690                 GOTO(out_obj, rc = -ESTALE);
2691
2692         /* It's really necessary to grab open_sem and check if the lease lock
2693          * has been lost. There would exist a concurrent writer coming in and
2694          * generating some dirty data in memory cache, the writeback would fail
2695          * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2696         if (!down_write_trylock(&mo->mot_open_sem))
2697                 GOTO(out_put_lease, rc = -EBUSY);
2698
2699         lock_res_and_lock(lease);
2700         lease_broken = ldlm_is_cancel(lease);
2701         unlock_res_and_lock(lease);
2702         if (lease_broken)
2703                 GOTO(out_unlock, rc = -EBUSY);
2704
2705         /* the file has yet opened by anyone else after we took the lease. */
2706         layout.mlc_opc = MD_LAYOUT_RESYNC;
2707         rc = mdt_layout_change(info, mo, &layout);
2708         if (rc)
2709                 GOTO(out_unlock, rc);
2710
2711         ma->ma_need = MA_INODE;
2712         ma->ma_valid = 0;
2713         rc = mdt_attr_get_complex(info, mo, ma);
2714         if (rc != 0)
2715                 GOTO(out_unlock, rc);
2716
2717         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2718         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2719
2720         EXIT;
2721 out_unlock:
2722         up_write(&mo->mot_open_sem);
2723 out_put_lease:
2724         LDLM_LOCK_PUT(lease);
2725 out_obj:
2726         mdt_object_put(info->mti_env, mo);
2727 out:
2728         mdt_client_compatibility(info);
2729         rc2 = mdt_fix_reply(info);
2730         if (rc == 0)
2731                 rc = rc2;
2732         return rc;
2733 }
2734
2735 struct mdt_reinter {
2736         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2737         enum lprocfs_extra_opc mr_extra_opc;
2738 };
2739
2740 static const struct mdt_reinter mdt_reinters[] = {
2741         [REINT_SETATTR] = {
2742                 .mr_handler = &mdt_reint_setattr,
2743                 .mr_extra_opc = MDS_REINT_SETATTR,
2744         },
2745         [REINT_CREATE] = {
2746                 .mr_handler = &mdt_reint_create,
2747                 .mr_extra_opc = MDS_REINT_CREATE,
2748         },
2749         [REINT_LINK] = {
2750                 .mr_handler = &mdt_reint_link,
2751                 .mr_extra_opc = MDS_REINT_LINK,
2752         },
2753         [REINT_UNLINK] = {
2754                 .mr_handler = &mdt_reint_unlink,
2755                 .mr_extra_opc = MDS_REINT_UNLINK,
2756         },
2757         [REINT_RENAME] = {
2758                 .mr_handler = &mdt_reint_rename,
2759                 .mr_extra_opc = MDS_REINT_RENAME,
2760         },
2761         [REINT_OPEN] = {
2762                 .mr_handler = &mdt_reint_open,
2763                 .mr_extra_opc = MDS_REINT_OPEN,
2764         },
2765         [REINT_SETXATTR] = {
2766                 .mr_handler = &mdt_reint_setxattr,
2767                 .mr_extra_opc = MDS_REINT_SETXATTR,
2768         },
2769         [REINT_RMENTRY] = {
2770                 .mr_handler = &mdt_reint_unlink,
2771                 .mr_extra_opc = MDS_REINT_UNLINK,
2772         },
2773         [REINT_MIGRATE] = {
2774                 .mr_handler = &mdt_reint_migrate,
2775                 .mr_extra_opc = MDS_REINT_RENAME,
2776         },
2777         [REINT_RESYNC] = {
2778                 .mr_handler = &mdt_reint_resync,
2779                 .mr_extra_opc = MDS_REINT_RESYNC,
2780         },
2781 };
2782
2783 int mdt_reint_rec(struct mdt_thread_info *info,
2784                   struct mdt_lock_handle *lhc)
2785 {
2786         const struct mdt_reinter *mr;
2787         int rc;
2788         ENTRY;
2789
2790         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2791                 RETURN(-EPROTO);
2792
2793         mr = &mdt_reinters[info->mti_rr.rr_opcode];
2794         if (mr->mr_handler == NULL)
2795                 RETURN(-EPROTO);
2796
2797         rc = (*mr->mr_handler)(info, lhc);
2798
2799         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2800                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
2801
2802         RETURN(rc);
2803 }