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 int mdt_unlock_slaves(struct mdt_thread_info *mti,
 221                              struct mdt_object *obj,
 222                              struct ldlm_enqueue_info *einfo,
 223                              int decref)
 224 {
 225         union ldlm_policy_data *policy = &mti->mti_policy;
 226         struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
 227         struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
 228         int i;
 229
 230         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 231         LASSERT(slave_locks);
 232
 233         memset(policy, 0, sizeof(*policy));
 234         policy->l_inodebits.bits = einfo->ei_inodebits;
 235         mdt_lock_handle_init(lh);
 236         mdt_lock_reg_init(lh, einfo->ei_mode);
 237         for (i = 0; i < slave_locks->ha_count; i++) {
 238                 if (test_bit(i, (void *)slave_locks->ha_map))
 239                         lh->mlh_rreg_lh = slave_locks->ha_handles[i];
 240                 else
 241                         lh->mlh_reg_lh = slave_locks->ha_handles[i];
 242                 mdt_object_unlock(mti, NULL, lh, decref);
 243                 slave_locks->ha_handles[i].cookie = 0ull;
 244         }
 245
 246         return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
 247                                 policy);
 248 }
 249
 250 static inline int mdt_object_striped(struct mdt_thread_info *mti,
 251                                      struct mdt_object *obj)
 252 {
 253         struct lu_device *bottom_dev;
 254         struct lu_object *bottom_obj;
 255         int rc;
 256
 257         if (!S_ISDIR(obj->mot_header.loh_attr))
 258                 return 0;
 259
 260         /* getxattr from bottom obj to avoid reading in shard FIDs */
 261         bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom);
 262         bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev,
 263                                           mdt_object_fid(obj), NULL);
 264         if (IS_ERR(bottom_obj))
 265                 return PTR_ERR(bottom_obj);
 266
 267         rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL,
 268                           XATTR_NAME_LMV);
 269         lu_object_put(mti->mti_env, bottom_obj);
 270
 271         return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc;
 272 }
 273
 274 /**
 275  * Lock slave stripes if necessary, the lock handles of slave stripes
 276  * will be stored in einfo->ei_cbdata.
 277  **/
 278 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
 279                            enum ldlm_mode mode, __u64 ibits,
 280                            struct ldlm_enqueue_info *einfo)
 281 {
 282         union ldlm_policy_data *policy = &mti->mti_policy;
 283
 284         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 285
 286         einfo->ei_type = LDLM_IBITS;
 287         einfo->ei_mode = mode;
 288         einfo->ei_cb_bl = mdt_remote_blocking_ast;
 289         einfo->ei_cb_local_bl = mdt_blocking_ast;
 290         einfo->ei_cb_cp = ldlm_completion_ast;
 291         einfo->ei_enq_slave = 1;
 292         einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
 293         einfo->ei_inodebits = ibits;
 294         memset(policy, 0, sizeof(*policy));
 295         policy->l_inodebits.bits = ibits;
 296
 297         return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
 298                               policy);
 299 }
 300
 301 int mdt_reint_striped_lock(struct mdt_thread_info *info,
 302                            struct mdt_object *o,
 303                            struct mdt_lock_handle *lh,
 304                            __u64 ibits,
 305                            struct ldlm_enqueue_info *einfo,
 306                            bool cos_incompat)
 307 {
 308         int rc;
 309
 310         LASSERT(!mdt_object_remote(o));
 311
 312         memset(einfo, 0, sizeof(*einfo));
 313
 314         rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
 315         if (rc)
 316                 return rc;
 317
 318         rc = mdt_object_striped(info, o);
 319         if (rc != 1) {
 320                 if (rc < 0)
 321                         mdt_object_unlock(info, o, lh, rc);
 322                 return rc;
 323         }
 324
 325         rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
 326         if (rc) {
 327                 mdt_object_unlock(info, o, lh, rc);
 328                 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
 329                         rc = 0;
 330         }
 331
 332         return rc;
 333 }
 334
 335 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
 336                               struct mdt_object *o,
 337                               struct mdt_lock_handle *lh,
 338                               struct ldlm_enqueue_info *einfo, int decref)
 339 {
 340         if (einfo->ei_cbdata)
 341                 mdt_unlock_slaves(info, o, einfo, decref);
 342         mdt_object_unlock(info, o, lh, decref);
 343 }
 344
 345 /*
 346  * VBR: we save three versions in reply:
 347  * 0 - parent. Check that parent version is the same during replay.
 348  * 1 - name. Version of 'name' if file exists with the same name or
 349  * ENOENT_VERSION, it is needed because file may appear due to missed replays.
 350  * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
 351  * check.
 352  */
 353 static int mdt_create(struct mdt_thread_info *info)
 354 {
 355         struct mdt_device       *mdt = info->mti_mdt;
 356         struct mdt_object       *parent;
 357         struct mdt_object       *child;
 358         struct mdt_lock_handle  *lh;
 359         struct mdt_body         *repbody;
 360         struct md_attr          *ma = &info->mti_attr;
 361         struct mdt_reint_record *rr = &info->mti_rr;
 362         struct md_op_spec       *spec = &info->mti_spec;
 363         int rc;
 364         ENTRY;
 365
 366         DEBUG_REQ(D_INODE, mdt_info_req(info),
 367                   "Create ("DNAME"->"DFID") in "DFID,
 368                   PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
 369
 370         if (!fid_is_md_operative(rr->rr_fid1))
 371                 RETURN(-EPERM);
 372
 373         if (S_ISDIR(ma->ma_attr.la_mode) &&
 374             spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
 375                 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
 376                 struct lu_ucred *uc = mdt_ucred(info);
 377                 struct obd_export *exp = mdt_info_req(info)->rq_export;
 378
 379                 /* Only new clients can create remote dir( >= 2.4) and
 380                  * striped dir(>= 2.6), old client will return -ENOTSUPP */
 381                 if (!mdt_is_dne_client(exp))
 382                         RETURN(-ENOTSUPP);
 383
 384                 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
 385                         if (!mdt_is_striped_client(exp))
 386                                 RETURN(-ENOTSUPP);
 387
 388                         if (!mdt->mdt_enable_striped_dir)
 389                                 RETURN(-EPERM);
 390                 } else if (!mdt->mdt_enable_remote_dir) {
 391                         RETURN(-EPERM);
 392                 }
 393
 394                 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
 395                     (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) ==
 396                     LMV_HASH_TYPE_CRUSH)
 397                         RETURN(-EPROTO);
 398
 399                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 400                     uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 401                     mdt->mdt_enable_remote_dir_gid != -1)
 402                         RETURN(-EPERM);
 403         }
 404
 405         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 406
 407         parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 408         if (IS_ERR(parent))
 409                 RETURN(PTR_ERR(parent));
 410
 411         if (!mdt_object_exists(parent))
 412                 GOTO(put_parent, rc = -ENOENT);
 413
 414         /*
 415          * LU-10235: check if name exists locklessly first to avoid massive
 416          * lock recalls on existing directories.
 417          */
 418         rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
 419                                       &info->mti_tmp_fid1, 1);
 420         if (rc == 0)
 421                 GOTO(put_parent, rc = -EEXIST);
 422
 423         /* -ENOENT is expected here */
 424         if (rc != -ENOENT)
 425                 GOTO(put_parent, rc);
 426
 427         /* save version of file name for replay, it must be ENOENT here */
 428         mdt_enoent_version_save(info, 1);
 429
 430         OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
 431
 432         lh = &info->mti_lh[MDT_LH_PARENT];
 433         mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 434         rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
 435         if (rc)
 436                 GOTO(put_parent, rc);
 437
 438         if (!mdt_object_remote(parent)) {
 439                 rc = mdt_version_get_check_save(info, parent, 0);
 440                 if (rc)
 441                         GOTO(unlock_parent, rc);
 442         }
 443
 444         child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
 445         if (unlikely(IS_ERR(child)))
 446                 GOTO(unlock_parent, rc = PTR_ERR(child));
 447
 448         ma->ma_need = MA_INODE;
 449         ma->ma_valid = 0;
 450
 451         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 452                         OBD_FAIL_MDS_REINT_CREATE_WRITE);
 453
 454         /* Version of child will be updated on disk. */
 455         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 456         rc = mdt_version_get_check_save(info, child, 2);
 457         if (rc)
 458                 GOTO(put_child, rc);
 459
 460         /* Let lower layer know current lock mode. */
 461         info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
 462
 463         /*
 464          * Do not perform lookup sanity check. We know that name does
 465          * not exist.
 466          */
 467         info->mti_spec.sp_cr_lookup = 0;
 468         info->mti_spec.sp_feat = &dt_directory_features;
 469
 470         rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
 471                         mdt_object_child(child), &info->mti_spec, ma);
 472         if (rc == 0)
 473                 rc = mdt_attr_get_complex(info, child, ma);
 474
 475         if (rc < 0)
 476                 GOTO(put_child, rc);
 477
 478         /*
 479          * On DNE, we need to eliminate dependey between 'mkdir a' and
 480          * 'mkdir a/b' if b is a striped directory, to achieve this, two
 481          * things are done below:
 482          * 1. save child and slaves lock.
 483          * 2. if the child is a striped directory, relock parent so to
 484          *    compare against with COS locks to ensure parent was
 485          *    committed to disk.
 486          */
 487         if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
 488                 struct mdt_lock_handle *lhc;
 489                 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 490                 bool cos_incompat;
 491
 492                 rc = mdt_object_striped(info, child);
 493                 if (rc < 0)
 494                         GOTO(put_child, rc);
 495
 496                 cos_incompat = rc;
 497                 if (cos_incompat) {
 498                         if (!mdt_object_remote(parent)) {
 499                                 mdt_object_unlock(info, parent, lh, 1);
 500                                 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 501                                 rc = mdt_reint_object_lock(info, parent, lh,
 502                                                            MDS_INODELOCK_UPDATE,
 503                                                            true);
 504                                 if (rc)
 505                                         GOTO(put_child, rc);
 506                         }
 507                 }
 508
 509                 lhc = &info->mti_lh[MDT_LH_CHILD];
 510                 mdt_lock_handle_init(lhc);
 511                 mdt_lock_reg_init(lhc, LCK_PW);
 512                 rc = mdt_reint_striped_lock(info, child, lhc,
 513                                             MDS_INODELOCK_UPDATE, einfo,
 514                                             cos_incompat);
 515                 if (rc)
 516                         GOTO(put_child, rc);
 517
 518                 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 519         }
 520
 521         /* Return fid & attr to client. */
 522         if (ma->ma_valid & MA_INODE)
 523                 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
 524                                    mdt_object_fid(child));
 525 put_child:
 526         mdt_object_put(info->mti_env, child);
 527 unlock_parent:
 528         mdt_object_unlock(info, parent, lh, rc);
 529 put_parent:
 530         mdt_object_put(info->mti_env, parent);
 531         RETURN(rc);
 532 }
 533
 534 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
 535                         struct md_attr *ma)
 536 {
 537         struct mdt_lock_handle  *lh;
 538         int do_vbr = ma->ma_attr.la_valid &
 539                         (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
 540         __u64 lockpart = MDS_INODELOCK_UPDATE;
 541         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 542         bool cos_incompat;
 543         int rc;
 544         ENTRY;
 545
 546         rc = mdt_object_striped(info, mo);
 547         if (rc < 0)
 548                 RETURN(rc);
 549
 550         cos_incompat = rc;
 551
 552         lh = &info->mti_lh[MDT_LH_PARENT];
 553         mdt_lock_reg_init(lh, LCK_PW);
 554
 555         /* Even though the new MDT will grant PERM lock to the old
 556          * client, but the old client will almost ignore that during
 557          * So it needs to revoke both LOOKUP and PERM lock here, so
 558          * both new and old client can cancel the dcache */
 559         if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
 560                 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
 561
 562         rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
 563                                     cos_incompat);
 564         if (rc != 0)
 565                 RETURN(rc);
 566
 567         /* all attrs are packed into mti_attr in unpack_setattr */
 568         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 569                        OBD_FAIL_MDS_REINT_SETATTR_WRITE);
 570
 571         /* VBR: update version if attr changed are important for recovery */
 572         if (do_vbr) {
 573                 /* update on-disk version of changed object */
 574                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
 575                 rc = mdt_version_get_check_save(info, mo, 0);
 576                 if (rc)
 577                         GOTO(out_unlock, rc);
 578         }
 579
 580         /* Ensure constant striping during chown(). See LU-2789. */
 581         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 582                 mutex_lock(&mo->mot_lov_mutex);
 583
 584         /* all attrs are packed into mti_attr in unpack_setattr */
 585         rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
 586
 587         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 588                 mutex_unlock(&mo->mot_lov_mutex);
 589
 590         if (rc != 0)
 591                 GOTO(out_unlock, rc);
 592         mdt_dom_obj_lvb_update(info->mti_env, mo, false);
 593         EXIT;
 594 out_unlock:
 595         mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
 596         return rc;
 597 }
 598
 599 /**
 600  * Check HSM flags and add HS_DIRTY flag if relevant.
 601  *
 602  * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
 603  * and is not RELEASED.
 604  */
 605 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
 606                         struct md_attr *ma)
 607 {
 608         struct lu_ucred *uc = mdt_ucred(info);
 609         cfs_cap_t cap_saved;
 610         int rc;
 611         ENTRY;
 612
 613         /* If the file was modified, add the dirty flag */
 614         ma->ma_need = MA_HSM;
 615         rc = mdt_attr_get_complex(info, mo, ma);
 616         if (rc) {
 617                 CERROR("file attribute read error for "DFID": %d.\n",
 618                         PFID(mdt_object_fid(mo)), rc);
 619                 RETURN(rc);
 620         }
 621
 622         /* If an up2date copy exists in the backend, add dirty flag */
 623         if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
 624             && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
 625                 ma->ma_hsm.mh_flags |= HS_DIRTY;
 626
 627                 /* Bump cap so that closes from non-owner writers can
 628                  * set the HSM state to dirty. */
 629                 cap_saved = uc->uc_cap;
 630                 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
 631                 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
 632                 uc->uc_cap = cap_saved;
 633                 if (rc)
 634                         CERROR("file attribute change error for "DFID": %d\n",
 635                                 PFID(mdt_object_fid(mo)), rc);
 636         }
 637
 638         RETURN(rc);
 639 }
 640
 641 static int mdt_reint_setattr(struct mdt_thread_info *info,
 642                              struct mdt_lock_handle *lhc)
 643 {
 644         struct mdt_device *mdt = info->mti_mdt;
 645         struct md_attr *ma = &info->mti_attr;
 646         struct mdt_reint_record *rr = &info->mti_rr;
 647         struct ptlrpc_request *req = mdt_info_req(info);
 648         struct mdt_object *mo;
 649         struct mdt_body *repbody;
 650         int rc, rc2;
 651         ENTRY;
 652
 653         DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
 654                   (unsigned int)ma->ma_attr.la_valid);
 655
 656         if (info->mti_dlm_req)
 657                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 658
 659         OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
 660
 661         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 662         mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
 663         if (IS_ERR(mo))
 664                 GOTO(out, rc = PTR_ERR(mo));
 665
 666         if (!mdt_object_exists(mo))
 667                 GOTO(out_put, rc = -ENOENT);
 668
 669         if (mdt_object_remote(mo))
 670                 GOTO(out_put, rc = -EREMOTE);
 671
 672         ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
 673         /* revoke lease lock if size is going to be changed */
 674         if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
 675                      !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
 676                      atomic_read(&mo->mot_lease_count) > 0)) {
 677                 down_read(&mo->mot_open_sem);
 678
 679                 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
 680                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 681                         mdt_lock_reg_init(lhc, LCK_CW);
 682
 683                         rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
 684                         if (rc != 0) {
 685                                 up_read(&mo->mot_open_sem);
 686                                 GOTO(out_put, rc);
 687                         }
 688
 689                         /* revoke lease lock */
 690                         mdt_object_unlock(info, mo, lhc, 1);
 691                 }
 692                 up_read(&mo->mot_open_sem);
 693         }
 694
 695         if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
 696                 /* Check write access for the O_TRUNC case */
 697                 if (mdt_write_read(mo) < 0)
 698                         GOTO(out_put, rc = -ETXTBSY);
 699
 700                 /* LU-10286: compatibility check for FLR.
 701                  * Please check the comment in mdt_finish_open() for details */
 702                 if (!exp_connect_flr(info->mti_exp) ||
 703                     !exp_connect_overstriping(info->mti_exp)) {
 704                         rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
 705                         if (rc < 0 && rc != -ENODATA)
 706                                 GOTO(out_put, rc);
 707
 708                         if (!exp_connect_flr(info->mti_exp)) {
 709                                 if (rc > 0 &&
 710                                     mdt_lmm_is_flr(info->mti_big_lmm))
 711                                         GOTO(out_put, rc = -EOPNOTSUPP);
 712                         }
 713
 714                         if (!exp_connect_overstriping(info->mti_exp)) {
 715                                 if (rc > 0 &&
 716                                     mdt_lmm_is_overstriping(info->mti_big_lmm))
 717                                         GOTO(out_put, rc = -EOPNOTSUPP);
 718                         }
 719                 }
 720
 721                 /* For truncate, the file size sent from client
 722                  * is believable, but the blocks are incorrect,
 723                  * which makes the block size in LSOM attribute
 724                  * inconsisent with the real block size.
 725                  */
 726                 rc = mdt_lsom_update(info, mo, true);
 727                 if (rc)
 728                         GOTO(out_put, rc);
 729         }
 730
 731         if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
 732                 if (ma->ma_valid & MA_LOV)
 733                         GOTO(out_put, rc = -EPROTO);
 734
 735                 /* MDT supports FMD for regular files due to Data-on-MDT */
 736                 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
 737                     ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
 738                         tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
 739                                        req->rq_xid);
 740
 741                 rc = mdt_attr_set(info, mo, ma);
 742                 if (rc)
 743                         GOTO(out_put, rc);
 744         } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
 745                    (ma->ma_valid & MA_INODE)) {
 746                 struct lu_buf *buf = &info->mti_buf;
 747                 struct lu_ucred *uc = mdt_ucred(info);
 748                 struct mdt_lock_handle *lh;
 749                 const char *name;
 750                 __u64 lockpart = MDS_INODELOCK_XATTR;
 751
 752                 /* reject if either remote or striped dir is disabled */
 753                 if (ma->ma_valid & MA_LMV) {
 754                         if (!mdt->mdt_enable_remote_dir ||
 755                             !mdt->mdt_enable_striped_dir)
 756                                 GOTO(out_put, rc = -EPERM);
 757
 758                         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 759                             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 760                             mdt->mdt_enable_remote_dir_gid != -1)
 761                                 GOTO(out_put, rc = -EPERM);
 762                 }
 763
 764                 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
 765                         GOTO(out_put, rc = -ENOTDIR);
 766
 767                 if (ma->ma_attr.la_valid != 0)
 768                         GOTO(out_put, rc = -EPROTO);
 769
 770                 if (ma->ma_valid & MA_LOV) {
 771                         buf->lb_buf = ma->ma_lmm;
 772                         buf->lb_len = ma->ma_lmm_size;
 773                         name = XATTR_NAME_LOV;
 774                 } else {
 775                         struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
 776
 777                         buf->lb_buf = lmu;
 778                         buf->lb_len = ma->ma_lmv_size;
 779                         name = XATTR_NAME_DEFAULT_LMV;
 780                         /* force client to update dir default layout */
 781                         lockpart |= MDS_INODELOCK_LOOKUP;
 782                 }
 783
 784                 lh = &info->mti_lh[MDT_LH_PARENT];
 785                 mdt_lock_reg_init(lh, LCK_PW);
 786
 787                 rc = mdt_object_lock(info, mo, lh, lockpart);
 788                 if (rc != 0)
 789                         GOTO(out_put, rc);
 790
 791                 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
 792                                   name, 0);
 793
 794                 mdt_object_unlock(info, mo, lh, rc);
 795                 if (rc)
 796                         GOTO(out_put, rc);
 797         } else {
 798                 GOTO(out_put, rc = -EPROTO);
 799         }
 800
 801         /* If file data is modified, add the dirty flag */
 802         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
 803                 rc = mdt_add_dirty_flag(info, mo, ma);
 804
 805         ma->ma_need = MA_INODE;
 806         ma->ma_valid = 0;
 807         rc = mdt_attr_get_complex(info, mo, ma);
 808         if (rc != 0)
 809                 GOTO(out_put, rc);
 810
 811         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
 812
 813         EXIT;
 814 out_put:
 815         mdt_object_put(info->mti_env, mo);
 816 out:
 817         if (rc == 0)
 818                 mdt_counter_incr(req, LPROC_MDT_SETATTR);
 819
 820         mdt_client_compatibility(info);
 821         rc2 = mdt_fix_reply(info);
 822         if (rc == 0)
 823                 rc = rc2;
 824         return rc;
 825 }
 826
 827 static int mdt_reint_create(struct mdt_thread_info *info,
 828                             struct mdt_lock_handle *lhc)
 829 {
 830         struct ptlrpc_request   *req = mdt_info_req(info);
 831         int                     rc;
 832         ENTRY;
 833
 834         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 835                 RETURN(err_serious(-ESTALE));
 836
 837         if (info->mti_dlm_req)
 838                 ldlm_request_cancel(mdt_info_req(info),
 839                                     info->mti_dlm_req, 0, LATF_SKIP);
 840
 841         if (!lu_name_is_valid(&info->mti_rr.rr_name))
 842                 RETURN(-EPROTO);
 843
 844         switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
 845         case S_IFDIR:
 846                 mdt_counter_incr(req, LPROC_MDT_MKDIR);
 847                 break;
 848         case S_IFREG:
 849         case S_IFLNK:
 850         case S_IFCHR:
 851         case S_IFBLK:
 852         case S_IFIFO:
 853         case S_IFSOCK:
 854                 /* Special file should stay on the same node as parent. */
 855                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
 856                 break;
 857         default:
 858                 CERROR("%s: Unsupported mode %o\n",
 859                        mdt_obd_name(info->mti_mdt),
 860                        info->mti_attr.ma_attr.la_mode);
 861                 RETURN(err_serious(-EOPNOTSUPP));
 862         }
 863
 864         rc = mdt_create(info);
 865         RETURN(rc);
 866 }
 867
 868 /*
 869  * VBR: save parent version in reply and child version getting by its name.
 870  * Version of child is getting and checking during its lookup. If
 871  */
 872 static int mdt_reint_unlink(struct mdt_thread_info *info,
 873                             struct mdt_lock_handle *lhc)
 874 {
 875         struct mdt_reint_record *rr = &info->mti_rr;
 876         struct ptlrpc_request *req = mdt_info_req(info);
 877         struct md_attr *ma = &info->mti_attr;
 878         struct lu_fid *child_fid = &info->mti_tmp_fid1;
 879         struct mdt_object *mp;
 880         struct mdt_object *mc;
 881         struct mdt_lock_handle *parent_lh;
 882         struct mdt_lock_handle *child_lh;
 883         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 884         __u64 lock_ibits;
 885         bool cos_incompat = false;
 886         int no_name = 0;
 887         int rc;
 888
 889         ENTRY;
 890
 891         DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
 892                   PNAME(&rr->rr_name));
 893
 894         if (info->mti_dlm_req)
 895                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 896
 897         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
 898                 RETURN(err_serious(-ENOENT));
 899
 900         if (!fid_is_md_operative(rr->rr_fid1))
 901                 RETURN(-EPERM);
 902
 903         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 904         if (IS_ERR(mp))
 905                 RETURN(PTR_ERR(mp));
 906
 907         if (mdt_object_remote(mp)) {
 908                 cos_incompat = true;
 909         } else {
 910                 rc = mdt_version_get_check_save(info, mp, 0);
 911                 if (rc)
 912                         GOTO(put_parent, rc);
 913         }
 914
 915 relock:
 916         parent_lh = &info->mti_lh[MDT_LH_PARENT];
 917         mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
 918         rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
 919                                    cos_incompat);
 920         if (rc != 0)
 921                 GOTO(put_parent, rc);
 922
 923         /* lookup child object along with version checking */
 924         fid_zero(child_fid);
 925         rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
 926         if (rc != 0) {
 927                 /* Name might not be able to find during resend of
 928                  * remote unlink, considering following case.
 929                  * dir_A is a remote directory, the name entry of
 930                  * dir_A is on MDT0, the directory is on MDT1,
 931                  *
 932                  * 1. client sends unlink req to MDT1.
 933                  * 2. MDT1 sends name delete update to MDT0.
 934                  * 3. name entry is being deleted in MDT0 synchronously.
 935                  * 4. MDT1 is restarted.
 936                  * 5. client resends unlink req to MDT1. So it can not
 937                  *    find the name entry on MDT0 anymore.
 938                  * In this case, MDT1 only needs to destory the local
 939                  * directory.
 940                  * */
 941                 if (mdt_object_remote(mp) && rc == -ENOENT &&
 942                     !fid_is_zero(rr->rr_fid2) &&
 943                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 944                         no_name = 1;
 945                         *child_fid = *rr->rr_fid2;
 946                  } else {
 947                         GOTO(unlock_parent, rc);
 948                  }
 949         }
 950
 951         if (!fid_is_md_operative(child_fid))
 952                 GOTO(unlock_parent, rc = -EPERM);
 953
 954         /* We will lock the child regardless it is local or remote. No harm. */
 955         mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
 956         if (IS_ERR(mc))
 957                 GOTO(unlock_parent, rc = PTR_ERR(mc));
 958
 959         if (!cos_incompat) {
 960                 rc = mdt_object_striped(info, mc);
 961                 if (rc < 0)
 962                         GOTO(put_child, rc);
 963
 964                 cos_incompat = rc;
 965                 if (cos_incompat) {
 966                         mdt_object_put(info->mti_env, mc);
 967                         mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
 968                         goto relock;
 969                 }
 970         }
 971
 972         child_lh = &info->mti_lh[MDT_LH_CHILD];
 973         mdt_lock_reg_init(child_lh, LCK_EX);
 974         if (info->mti_spec.sp_rm_entry) {
 975                 struct lu_ucred *uc  = mdt_ucred(info);
 976
 977                 if (!mdt_is_dne_client(req->rq_export))
 978                         /* Return -ENOTSUPP for old client */
 979                         GOTO(put_child, rc = -ENOTSUPP);
 980
 981                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
 982                         GOTO(put_child, rc = -EPERM);
 983
 984                 ma->ma_need = MA_INODE;
 985                 ma->ma_valid = 0;
 986                 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
 987                                 NULL, &rr->rr_name, ma, no_name);
 988                 GOTO(put_child, rc);
 989         }
 990
 991         if (mdt_object_remote(mc)) {
 992                 struct mdt_body  *repbody;
 993
 994                 if (!fid_is_zero(rr->rr_fid2)) {
 995                         CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
 996                                mdt_obd_name(info->mti_mdt),
 997                                PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 998                         GOTO(put_child, rc = -ENOENT);
 999                 }
1000                 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1001                        mdt_obd_name(info->mti_mdt),
1002                        PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1003
1004                 if (!mdt_is_dne_client(req->rq_export))
1005                         /* Return -ENOTSUPP for old client */
1006                         GOTO(put_child, rc = -ENOTSUPP);
1007
1008                 /* Revoke the LOOKUP lock of the remote object granted by
1009                  * this MDT. Since the unlink will happen on another MDT,
1010                  * it will release the LOOKUP lock right away. Then What
1011                  * would happen if another client try to grab the LOOKUP
1012                  * lock at the same time with unlink XXX */
1013                 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1014                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1015                 LASSERT(repbody != NULL);
1016                 repbody->mbo_fid1 = *mdt_object_fid(mc);
1017                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1018                 GOTO(unlock_child, rc = -EREMOTE);
1019         }
1020         /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1021          * this now because a running HSM restore on the child (unlink
1022          * victim) will hold the layout lock. See LU-4002. */
1023         lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1024         if (mdt_object_remote(mp)) {
1025                 /* Enqueue lookup lock from parent MDT */
1026                 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1027                                             &child_lh->mlh_rreg_lh,
1028                                             child_lh->mlh_rreg_mode,
1029                                             MDS_INODELOCK_LOOKUP, false);
1030                 if (rc != ELDLM_OK)
1031                         GOTO(put_child, rc);
1032
1033                 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1034         }
1035
1036         rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1037                                     cos_incompat);
1038         if (rc != 0)
1039                 GOTO(put_child, rc);
1040
1041         /*
1042          * Now we can only make sure we need MA_INODE, in mdd layer, will check
1043          * whether need MA_LOV and MA_COOKIE.
1044          */
1045         ma->ma_need = MA_INODE;
1046         ma->ma_valid = 0;
1047
1048         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1049                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1050         /* save version when object is locked */
1051         mdt_version_get_save(info, mc, 1);
1052
1053         mutex_lock(&mc->mot_lov_mutex);
1054
1055         rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1056                         mdt_object_child(mc), &rr->rr_name, ma, no_name);
1057
1058         mutex_unlock(&mc->mot_lov_mutex);
1059         if (rc != 0)
1060                 GOTO(unlock_child, rc);
1061
1062         if (!lu_object_is_dying(&mc->mot_header)) {
1063                 rc = mdt_attr_get_complex(info, mc, ma);
1064                 if (rc)
1065                         GOTO(out_stat, rc);
1066         } else if (mdt_dom_check_for_discard(info, mc)) {
1067                 mdt_dom_discard_data(info, mc);
1068         }
1069         mdt_handle_last_unlink(info, mc, ma);
1070
1071 out_stat:
1072         if (ma->ma_valid & MA_INODE) {
1073                 switch (ma->ma_attr.la_mode & S_IFMT) {
1074                 case S_IFDIR:
1075                         mdt_counter_incr(req, LPROC_MDT_RMDIR);
1076                         break;
1077                 case S_IFREG:
1078                 case S_IFLNK:
1079                 case S_IFCHR:
1080                 case S_IFBLK:
1081                 case S_IFIFO:
1082                 case S_IFSOCK:
1083                         mdt_counter_incr(req, LPROC_MDT_UNLINK);
1084                         break;
1085                 default:
1086                         LASSERTF(0, "bad file type %o unlinking\n",
1087                                 ma->ma_attr.la_mode);
1088                 }
1089         }
1090
1091         EXIT;
1092
1093 unlock_child:
1094         mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1095 put_child:
1096         mdt_object_put(info->mti_env, mc);
1097 unlock_parent:
1098         mdt_object_unlock(info, mp, parent_lh, rc);
1099 put_parent:
1100         mdt_object_put(info->mti_env, mp);
1101         CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
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 (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1130                 req->rq_no_reply = 1;
1131                 RETURN(err_serious(-ENOENT));
1132         }
1133
1134         if (info->mti_dlm_req)
1135                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1136
1137         /* Invalid case so return error immediately instead of
1138          * processing it */
1139         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1140                 RETURN(-EPERM);
1141
1142         if (!fid_is_md_operative(rr->rr_fid1) ||
1143             !fid_is_md_operative(rr->rr_fid2))
1144                 RETURN(-EPERM);
1145
1146         /* step 1: find target parent dir */
1147         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1148         if (IS_ERR(mp))
1149                 RETURN(PTR_ERR(mp));
1150
1151         rc = mdt_version_get_check_save(info, mp, 0);
1152         if (rc)
1153                 GOTO(put_parent, rc);
1154
1155         /* step 2: find source */
1156         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1157         if (IS_ERR(ms))
1158                 GOTO(put_parent, rc = PTR_ERR(ms));
1159
1160         if (!mdt_object_exists(ms)) {
1161                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1162                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1163                 GOTO(put_source, rc = -ENOENT);
1164         }
1165
1166         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1167
1168         lhp = &info->mti_lh[MDT_LH_PARENT];
1169         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1170         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1171                                    cos_incompat);
1172         if (rc != 0)
1173                 GOTO(put_source, rc);
1174
1175         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1176
1177         lhs = &info->mti_lh[MDT_LH_CHILD];
1178         mdt_lock_reg_init(lhs, LCK_EX);
1179         rc = mdt_reint_object_lock(info, ms, lhs,
1180                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1181                                    cos_incompat);
1182         if (rc != 0)
1183                 GOTO(unlock_parent, rc);
1184
1185         /* step 3: link it */
1186         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1187                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1188
1189         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1190         rc = mdt_version_get_check_save(info, ms, 1);
1191         if (rc)
1192                 GOTO(unlock_source, rc);
1193
1194         /** check target version by name during replay */
1195         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1196                                       &info->mti_tmp_fid1, 2);
1197         if (rc != 0 && rc != -ENOENT)
1198                 GOTO(unlock_source, rc);
1199         /* save version of file name for replay, it must be ENOENT here */
1200         if (!req_is_replay(mdt_info_req(info))) {
1201                 if (rc != -ENOENT) {
1202                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1203                                PNAME(&rr->rr_name));
1204                         GOTO(unlock_source, rc = -EEXIST);
1205                 }
1206                 info->mti_ver[2] = ENOENT_VERSION;
1207                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1208         }
1209
1210         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1211                       mdt_object_child(ms), &rr->rr_name, ma);
1212
1213         if (rc == 0)
1214                 mdt_counter_incr(req, LPROC_MDT_LINK);
1215
1216         EXIT;
1217 unlock_source:
1218         mdt_object_unlock(info, ms, lhs, rc);
1219 unlock_parent:
1220         mdt_object_unlock(info, mp, lhp, rc);
1221 put_source:
1222         mdt_object_put(info->mti_env, ms);
1223 put_parent:
1224         mdt_object_put(info->mti_env, mp);
1225         return rc;
1226 }
1227 /**
1228  * lock the part of the directory according to the hash of the name
1229  * (lh->mlh_pdo_hash) in parallel directory lock.
1230  */
1231 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1232                               struct mdt_lock_handle *lh,
1233                               struct mdt_object *obj, __u64 ibits,
1234                               bool cos_incompat)
1235 {
1236         struct ldlm_res_id *res = &info->mti_res_id;
1237         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1238         union ldlm_policy_data *policy = &info->mti_policy;
1239         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1240         int rc;
1241
1242         /*
1243          * Finish res_id initializing by name hash marking part of
1244          * directory which is taking modification.
1245          */
1246         LASSERT(lh->mlh_pdo_hash != 0);
1247         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1248         memset(policy, 0, sizeof(*policy));
1249         policy->l_inodebits.bits = ibits;
1250         if (cos_incompat &&
1251             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1252                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1253         /*
1254          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1255          * going to be sent to client. If it is - mdt_intent_policy() path will
1256          * fix it up and turn FL_LOCAL flag off.
1257          */
1258         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1259                           policy, res, dlmflags,
1260                           &info->mti_exp->exp_handle.h_cookie);
1261         return rc;
1262 }
1263
1264 /**
1265  * Get BFL lock for rename or migrate process.
1266  **/
1267 static int mdt_rename_lock(struct mdt_thread_info *info,
1268                            struct lustre_handle *lh)
1269 {
1270         int     rc;
1271         ENTRY;
1272
1273         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1274                 struct lu_fid *fid = &info->mti_tmp_fid1;
1275                 struct mdt_object *obj;
1276
1277                 /* XXX, right now, it has to use object API to
1278                  * enqueue lock cross MDT, so it will enqueue
1279                  * rename lock(with LUSTRE_BFL_FID) by root object */
1280                 lu_root_fid(fid);
1281                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1282                 if (IS_ERR(obj))
1283                         RETURN(PTR_ERR(obj));
1284
1285                 rc = mdt_remote_object_lock(info, obj,
1286                                             &LUSTRE_BFL_FID, lh,
1287                                             LCK_EX,
1288                                             MDS_INODELOCK_UPDATE, false);
1289                 mdt_object_put(info->mti_env, obj);
1290         } else {
1291                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1292                 union ldlm_policy_data *policy = &info->mti_policy;
1293                 struct ldlm_res_id *res_id = &info->mti_res_id;
1294                 __u64 flags = 0;
1295
1296                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1297                 memset(policy, 0, sizeof *policy);
1298                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1299                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1300                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1301                                             LDLM_IBITS, policy, LCK_EX, &flags,
1302                                             ldlm_blocking_ast,
1303                                             ldlm_completion_ast, NULL, NULL, 0,
1304                                             LVB_T_NONE,
1305                                             &info->mti_exp->exp_handle.h_cookie,
1306                                             lh);
1307                 RETURN(rc);
1308         }
1309         RETURN(rc);
1310 }
1311
1312 static void mdt_rename_unlock(struct lustre_handle *lh)
1313 {
1314         ENTRY;
1315         LASSERT(lustre_handle_is_used(lh));
1316         /* Cancel the single rename lock right away */
1317         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1318         EXIT;
1319 }
1320
1321 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1322                                                 const struct lu_fid *fid,
1323                                                 int idx)
1324 {
1325         struct mdt_object *dir;
1326         int rc;
1327
1328         ENTRY;
1329
1330         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1331         if (IS_ERR(dir))
1332                 RETURN(dir);
1333
1334         /* check early, the real version will be saved after locking */
1335         rc = mdt_version_get_check(info, dir, idx);
1336         if (rc)
1337                 GOTO(out_put, rc);
1338
1339         if (!mdt_object_exists(dir))
1340                 GOTO(out_put, rc = -ENOENT);
1341
1342         if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1343                 GOTO(out_put, rc = -ENOTDIR);
1344
1345         RETURN(dir);
1346 out_put:
1347         mdt_object_put(info->mti_env, dir);
1348         return ERR_PTR(rc);
1349 }
1350
1351 /*
1352  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1353  * herein we don't really check it, just do revoke.
1354  */
1355 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1356                                   struct mdt_object *pobj,
1357                                   struct mdt_object *obj)
1358 {
1359         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1360         int rc;
1361
1362         mdt_lock_handle_init(lh);
1363         mdt_lock_reg_init(lh, LCK_EX);
1364
1365         if (mdt_object_remote(pobj)) {
1366                 /* don't bother to check if pobj and obj are on the same MDT. */
1367                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1368                                             &lh->mlh_rreg_lh, LCK_EX,
1369                                             MDS_INODELOCK_LOOKUP, false);
1370         } else if (mdt_object_remote(obj)) {
1371                 struct ldlm_res_id *res = &info->mti_res_id;
1372                 union ldlm_policy_data *policy = &info->mti_policy;
1373                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1374                                  LDLM_FL_COS_INCOMPAT;
1375
1376                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1377                 memset(policy, 0, sizeof(*policy));
1378                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1379                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1380                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1381                                   dlmflags, NULL);
1382         } else {
1383                 /* do nothing if both are local */
1384                 return 0;
1385         }
1386
1387         if (rc != ELDLM_OK)
1388                 return rc;
1389
1390         /*
1391          * TODO, currently we don't save this lock because there is no place to
1392          * hold this lock handle, but to avoid race we need to save this lock.
1393          */
1394         mdt_object_unlock(info, NULL, lh, 1);
1395
1396         return 0;
1397 }
1398
1399 /*
1400  * operation may takes locks of linkea, or directory stripes, group them in
1401  * different list.
1402  */
1403 struct mdt_sub_lock {
1404         struct mdt_object      *msl_obj;
1405         struct mdt_lock_handle  msl_lh;
1406         struct list_head        msl_linkage;
1407 };
1408
1409 static void mdt_unlock_list(struct mdt_thread_info *info,
1410                             struct list_head *list, int decref)
1411 {
1412         struct mdt_sub_lock *msl;
1413         struct mdt_sub_lock *tmp;
1414
1415         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1416                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1417                 list_del(&msl->msl_linkage);
1418                 OBD_FREE_PTR(msl);
1419         }
1420 }
1421
1422 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1423                                              struct mdt_object *obj,
1424                                              struct mdt_lock_handle *lh,
1425                                              struct ldlm_enqueue_info *einfo,
1426                                              struct list_head *slave_locks,
1427                                              int decref)
1428 {
1429         if (mdt_object_remote(obj)) {
1430                 mdt_unlock_list(info, slave_locks, decref);
1431                 mdt_object_unlock(info, obj, lh, decref);
1432         } else {
1433                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1434         }
1435 }
1436
1437 /*
1438  * lock parents of links, and also check whether total locks don't exceed
1439  * RS_MAX_LOCKS.
1440  *
1441  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1442  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1443  * \retval      -ev negative errno upon error
1444  */
1445 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1446                                  struct mdt_object *pobj,
1447                                  const struct md_attr *ma,
1448                                  struct mdt_object *obj,
1449                                  struct mdt_lock_handle *lhp,
1450                                  struct ldlm_enqueue_info *peinfo,
1451                                  struct list_head *parent_slave_locks,
1452                                  struct list_head *link_locks)
1453 {
1454         struct mdt_device *mdt = info->mti_mdt;
1455         struct lu_buf *buf = &info->mti_big_buf;
1456         struct lu_name *lname = &info->mti_name;
1457         struct linkea_data ldata = { NULL };
1458         bool blocked = false;
1459         int local_lnkp_cnt = 0;
1460         int rc;
1461
1462         ENTRY;
1463
1464         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1465                 RETURN(0);
1466
1467         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1468         if (buf->lb_buf == NULL)
1469                 RETURN(-ENOMEM);
1470
1471         ldata.ld_buf = buf;
1472         rc = mdt_links_read(info, obj, &ldata);
1473         if (rc) {
1474                 if (rc == -ENOENT || rc == -ENODATA)
1475                         rc = 0;
1476                 RETURN(rc);
1477         }
1478
1479         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1480              linkea_next_entry(&ldata)) {
1481                 struct mdt_object *lnkp;
1482                 struct mdt_sub_lock *msl;
1483                 struct lu_fid fid;
1484                 __u64 ibits;
1485
1486                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1487                                     &fid);
1488
1489                 /* check if it's also linked to parent */
1490                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1491                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1492                                PFID(&fid), PNAME(lname));
1493                         /* in case link is remote object, revoke LOOKUP lock */
1494                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1495                         continue;
1496                 }
1497
1498                 lnkp = NULL;
1499
1500                 /* check if it's linked to a stripe of parent */
1501                 if (ma->ma_valid & MA_LMV) {
1502                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1503                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1504                         int j = 0;
1505
1506                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1507                                 fid_le_to_cpu(stripe_fid,
1508                                               &lmv->lmv_stripe_fids[j]);
1509                                 if (lu_fid_eq(stripe_fid, &fid)) {
1510                                         CDEBUG(D_INFO, "skip stripe "DFID
1511                                                ", reovke "DNAME"\n",
1512                                                PFID(&fid), PNAME(lname));
1513                                         lnkp = mdt_object_find(info->mti_env,
1514                                                                mdt, &fid);
1515                                         if (IS_ERR(lnkp))
1516                                                 GOTO(out, rc = PTR_ERR(lnkp));
1517                                         break;
1518                                 }
1519                         }
1520
1521                         if (lnkp) {
1522                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1523                                                                    obj);
1524                                 mdt_object_put(info->mti_env, lnkp);
1525                                 continue;
1526                         }
1527                 }
1528
1529                 /* Check if it's already locked */
1530                 list_for_each_entry(msl, link_locks, msl_linkage) {
1531                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1532                                 CDEBUG(D_INFO,
1533                                        DFID" was locked, revoke "DNAME"\n",
1534                                        PFID(&fid), PNAME(lname));
1535                                 lnkp = msl->msl_obj;
1536                                 break;
1537                         }
1538                 }
1539
1540                 if (lnkp) {
1541                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1542                         continue;
1543                 }
1544
1545                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1546                        PFID(&fid), PNAME(lname));
1547
1548                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1549                 if (IS_ERR(lnkp)) {
1550                         CWARN("%s: cannot find obj "DFID": %ld\n",
1551                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1552                         continue;
1553                 }
1554
1555                 if (!mdt_object_exists(lnkp)) {
1556                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1557                               PFID(&fid), PNAME(lname));
1558                         mdt_object_put(info->mti_env, lnkp);
1559                         continue;
1560                 }
1561
1562                 if (!mdt_object_remote(lnkp))
1563                         local_lnkp_cnt++;
1564
1565                 OBD_ALLOC_PTR(msl);
1566                 if (msl == NULL)
1567                         GOTO(out, rc = -ENOMEM);
1568
1569                 /*
1570                  * we can't follow parent-child lock order like other MD
1571                  * operations, use lock_try here to avoid deadlock, if the lock
1572                  * cannot be taken, drop all locks taken, revoke the blocked
1573                  * one, and continue processing the remaining entries, and in
1574                  * the end of the loop restart from beginning.
1575                  */
1576                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1577                 ibits = 0;
1578                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1579                                          MDS_INODELOCK_UPDATE, true);
1580                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1581
1582                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1583                                PFID(&fid), PNAME(lname));
1584
1585                         mdt_unlock_list(info, link_locks, 1);
1586                         /* also unlock parent locks to avoid deadlock */
1587                         if (!blocked)
1588                                 mdt_migrate_object_unlock(info, pobj, lhp,
1589                                                           peinfo,
1590                                                           parent_slave_locks,
1591                                                           1);
1592
1593                         blocked = true;
1594
1595                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1596                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1597                                              MDS_INODELOCK_UPDATE);
1598                         if (rc) {
1599                                 mdt_object_put(info->mti_env, lnkp);
1600                                 OBD_FREE_PTR(msl);
1601                                 GOTO(out, rc);
1602                         }
1603
1604                         if (mdt_object_remote(lnkp)) {
1605                                 struct ldlm_lock *lock;
1606
1607                                 /*
1608                                  * for remote object, set lock cb_atomic,
1609                                  * so lock can be released in blocking_ast()
1610                                  * immediately, then the next lock_try will
1611                                  * have better chance of success.
1612                                  */
1613                                 lock = ldlm_handle2lock(
1614                                                 &msl->msl_lh.mlh_rreg_lh);
1615                                 LASSERT(lock != NULL);
1616                                 lock_res_and_lock(lock);
1617                                 ldlm_set_atomic_cb(lock);
1618                                 unlock_res_and_lock(lock);
1619                                 LDLM_LOCK_PUT(lock);
1620                         }
1621
1622                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1623                         OBD_FREE_PTR(msl);
1624                         continue;
1625                 }
1626
1627                 INIT_LIST_HEAD(&msl->msl_linkage);
1628                 msl->msl_obj = lnkp;
1629                 list_add_tail(&msl->msl_linkage, link_locks);
1630
1631                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1632         }
1633
1634         if (blocked)
1635                 GOTO(out, rc = -EBUSY);
1636
1637         EXIT;
1638 out:
1639         if (rc)
1640                 mdt_unlock_list(info, link_locks, rc);
1641         else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1642                 /*
1643                  * parent may have 3 local objects: master object and 2 stripes
1644                  * (if it's being migrated too); source may have 2 local
1645                  * objects: master and 1 stripe; target has 1 local object.
1646                  */
1647                 rc = 1;
1648         return rc;
1649 }
1650
1651 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1652                                   struct mdt_object *obj,
1653                                   const struct md_attr *ma,
1654                                   struct list_head *slave_locks)
1655 {
1656         struct mdt_device *mdt = info->mti_mdt;
1657         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1658         struct lu_fid *fid = &info->mti_tmp_fid1;
1659         struct mdt_object *slave;
1660         struct mdt_sub_lock *msl;
1661         int i;
1662         int rc;
1663
1664         ENTRY;
1665
1666         LASSERT(mdt_object_remote(obj));
1667         LASSERT(ma->ma_valid & MA_LMV);
1668         LASSERT(lmv);
1669
1670         if (!lmv_is_sane(lmv))
1671                 RETURN(-EINVAL);
1672
1673         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1674                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1675
1676                 if (!fid_is_sane(fid))
1677                         continue;
1678
1679                 slave = mdt_object_find(info->mti_env, mdt, fid);
1680                 if (IS_ERR(slave))
1681                         GOTO(out, rc = PTR_ERR(slave));
1682
1683                 OBD_ALLOC_PTR(msl);
1684                 if (!msl) {
1685                         mdt_object_put(info->mti_env, slave);
1686                         GOTO(out, rc = -ENOMEM);
1687                 }
1688
1689                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1690                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1691                                            MDS_INODELOCK_UPDATE, true);
1692                 if (rc) {
1693                         OBD_FREE_PTR(msl);
1694                         mdt_object_put(info->mti_env, slave);
1695                         GOTO(out, rc);
1696                 }
1697
1698                 INIT_LIST_HEAD(&msl->msl_linkage);
1699                 msl->msl_obj = slave;
1700                 list_add_tail(&msl->msl_linkage, slave_locks);
1701         }
1702         EXIT;
1703
1704 out:
1705         if (rc)
1706                 mdt_unlock_list(info, slave_locks, rc);
1707         return rc;
1708 }
1709
1710 /* lock parent and its stripes */
1711 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1712                                    struct mdt_object *obj,
1713                                    const struct md_attr *ma,
1714                                    struct mdt_lock_handle *lh,
1715                                    struct ldlm_enqueue_info *einfo,
1716                                    struct list_head *slave_locks)
1717 {
1718         int rc;
1719
1720         if (mdt_object_remote(obj)) {
1721                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1722                                             &lh->mlh_rreg_lh, LCK_PW,
1723                                             MDS_INODELOCK_UPDATE, false);
1724                 if (rc != ELDLM_OK)
1725                         return rc;
1726
1727                 /*
1728                  * if obj is remote and striped, lock its stripes explicitly
1729                  * because it's not striped in LOD layer on this MDT.
1730                  */
1731                 if (ma->ma_valid & MA_LMV) {
1732                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1733                         if (rc)
1734                                 mdt_object_unlock(info, obj, lh, rc);
1735                 }
1736         } else {
1737                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1738                                             einfo, true);
1739         }
1740
1741         return rc;
1742 }
1743
1744 /*
1745  * in migration, object may be remote, and we need take full lock of it and its
1746  * stripes if it's directory, besides, object may be a remote object on its
1747  * parent, revoke its LOOKUP lock on where its parent is located.
1748  */
1749 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1750                                    struct mdt_object *pobj,
1751                                    struct mdt_object *obj,
1752                                    struct mdt_lock_handle *lh,
1753                                    struct ldlm_enqueue_info *einfo,
1754                                    struct list_head *slave_locks)
1755 {
1756         int rc;
1757
1758         if (mdt_object_remote(obj)) {
1759                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1760                 if (rc)
1761                         return rc;
1762
1763                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1764                                             &lh->mlh_rreg_lh, LCK_EX,
1765                                             MDS_INODELOCK_FULL, false);
1766                 if (rc != ELDLM_OK)
1767                         return rc;
1768
1769                 /*
1770                  * if obj is remote and striped, lock its stripes explicitly
1771                  * because it's not striped in LOD layer on this MDT.
1772                  */
1773                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1774                         struct md_attr *ma = &info->mti_attr;
1775
1776                         ma->ma_lmv = info->mti_big_lmm;
1777                         ma->ma_lmv_size = info->mti_big_lmmsize;
1778                         ma->ma_valid = 0;
1779                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1780                         if (rc) {
1781                                 mdt_object_unlock(info, obj, lh, rc);
1782                                 return rc;
1783                         }
1784
1785                         if (ma->ma_valid & MA_LMV) {
1786                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1787                                                             slave_locks);
1788                                 if (rc)
1789                                         mdt_object_unlock(info, obj, lh, rc);
1790                         }
1791                 }
1792         } else {
1793                 if (mdt_object_remote(pobj)) {
1794                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1795                         if (rc)
1796                                 return rc;
1797                 }
1798
1799                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1800                                             einfo, true);
1801         }
1802
1803         return rc;
1804 }
1805
1806 /*
1807  * lookup source by name, if parent is striped directory, we need to find the
1808  * corresponding stripe where source is located, and then lookup there.
1809  *
1810  * besides, if parent is migrating too, and file is already in target stripe,
1811  * this should be a redo of 'lfs migrate' on client side.
1812  */
1813 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1814                               struct mdt_object *pobj,
1815                               const struct md_attr *ma,
1816                               const struct lu_name *lname,
1817                               struct mdt_object **spobj,
1818                               struct mdt_object **sobj)
1819 {
1820         const struct lu_env *env = info->mti_env;
1821         struct lu_fid *fid = &info->mti_tmp_fid1;
1822         struct mdt_object *stripe;
1823         int rc;
1824
1825         if (ma->ma_valid & MA_LMV) {
1826                 /* if parent is striped, lookup on corresponding stripe */
1827                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1828
1829                 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
1830                                                   lname->ln_namelen);
1831                 if (rc < 0)
1832                         return rc;
1833
1834                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1835
1836                 stripe = mdt_object_find(env, info->mti_mdt, fid);
1837                 if (IS_ERR(stripe))
1838                         return PTR_ERR(stripe);
1839
1840                 fid_zero(fid);
1841                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1842                                 &info->mti_spec);
1843                 if (rc == -ENOENT &&
1844                     (cpu_to_le32(lmv->lmv_hash_type) &
1845                      LMV_HASH_FLAG_LAYOUT_CHANGE)) {
1846                         /*
1847                          * if parent is migrating, and lookup child failed on
1848                          * source stripe, lookup again on target stripe, if it
1849                          * exists, it means previous migration was interrupted,
1850                          * and current file was migrated already.
1851                          */
1852                         mdt_object_put(env, stripe);
1853
1854                         rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
1855                                                       lname->ln_namelen);
1856                         if (rc < 0)
1857                                 return rc;
1858
1859                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1860
1861                         stripe = mdt_object_find(env, info->mti_mdt, fid);
1862                         if (IS_ERR(stripe))
1863                                 return PTR_ERR(stripe);
1864
1865                         fid_zero(fid);
1866                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1867                                         fid, &info->mti_spec);
1868                         mdt_object_put(env, stripe);
1869                         return rc ?: -EALREADY;
1870                 } else if (rc) {
1871                         mdt_object_put(env, stripe);
1872                         return rc;
1873                 }
1874         } else {
1875                 fid_zero(fid);
1876                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1877                                 &info->mti_spec);
1878                 if (rc)
1879                         return rc;
1880
1881                 stripe = pobj;
1882                 mdt_object_get(env, stripe);
1883         }
1884
1885         *spobj = stripe;
1886
1887         *sobj = mdt_object_find(env, info->mti_mdt, fid);
1888         if (IS_ERR(*sobj)) {
1889                 mdt_object_put(env, stripe);
1890                 rc = PTR_ERR(*sobj);
1891                 *spobj = NULL;
1892                 *sobj = NULL;
1893         }
1894
1895         return rc;
1896 }
1897
1898 /* end lease and close file for regular file */
1899 static int mdd_migrate_close(struct mdt_thread_info *info,
1900                              struct mdt_object *obj)
1901 {
1902         struct close_data *data;
1903         struct mdt_body *repbody;
1904         struct ldlm_lock *lease;
1905         int rc;
1906         int rc2;
1907
1908         rc = -EPROTO;
1909         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1910                                       RCL_CLIENT) ||
1911             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1912                                       RCL_CLIENT))
1913                 goto close;
1914
1915         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1916         if (!data)
1917                 goto close;
1918
1919         rc = -ESTALE;
1920         lease = ldlm_handle2lock(&data->cd_handle);
1921         if (!lease)
1922                 goto close;
1923
1924         /* check if the lease was already canceled */
1925         lock_res_and_lock(lease);
1926         rc = ldlm_is_cancel(lease);
1927         unlock_res_and_lock(lease);
1928
1929         if (rc) {
1930                 rc = -EAGAIN;
1931                 LDLM_DEBUG(lease, DFID" lease broken",
1932                            PFID(mdt_object_fid(obj)));
1933         }
1934
1935         /*
1936          * cancel server side lease, client side counterpart should have been
1937          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1938          */
1939         ldlm_lock_cancel(lease);
1940         ldlm_reprocess_all(lease->l_resource, lease);
1941         LDLM_LOCK_PUT(lease);
1942
1943 close:
1944         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1945         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1946         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1947
1948         return rc ?: rc2;
1949 }
1950
1951 /*
1952  * migrate file in below steps:
1953  *  1. lock parent and its stripes
1954  *  2. lookup source by name
1955  *  3. lock parents of source links if source is not directory
1956  *  4. reject if source is in HSM
1957  *  5. take source open_sem and close file if source is regular file
1958  *  6. lock source and its stripes if it's directory
1959  *  7. lock target so subsequent change to it can trigger COS
1960  *  8. migrate file
1961  *  9. unlock above locks
1962  * 10. sync device if source has links
1963  */
1964 static int mdt_reint_migrate(struct mdt_thread_info *info,
1965                              struct mdt_lock_handle *unused)
1966 {
1967         const struct lu_env *env = info->mti_env;
1968         struct mdt_device *mdt = info->mti_mdt;
1969         struct ptlrpc_request *req = mdt_info_req(info);
1970         struct mdt_reint_record *rr = &info->mti_rr;
1971         struct lu_ucred *uc = mdt_ucred(info);
1972         struct md_attr *ma = &info->mti_attr;
1973         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1974         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1975         struct mdt_object *pobj;
1976         struct mdt_object *spobj = NULL;
1977         struct mdt_object *sobj = NULL;
1978         struct mdt_object *tobj;
1979         struct lustre_handle rename_lh = { 0 };
1980         struct mdt_lock_handle *lhp;
1981         struct mdt_lock_handle *lhs;
1982         struct mdt_lock_handle *lht;
1983         LIST_HEAD(parent_slave_locks);
1984         LIST_HEAD(child_slave_locks);
1985         LIST_HEAD(link_locks);
1986         int lock_retries = 5;
1987         bool open_sem_locked = false;
1988         bool do_sync = false;
1989         int rc;
1990         ENTRY;
1991
1992         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1993                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1994
1995         if (info->mti_dlm_req)
1996                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1997
1998         if (!fid_is_md_operative(rr->rr_fid1) ||
1999             !fid_is_md_operative(rr->rr_fid2))
2000                 RETURN(-EPERM);
2001
2002         /* don't allow migrate . or .. */
2003         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2004                 RETURN(-EBUSY);
2005
2006         if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2007                 RETURN(-EPERM);
2008
2009         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
2010             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2011             mdt->mdt_enable_remote_dir_gid != -1)
2012                 RETURN(-EPERM);
2013
2014         /*
2015          * Note: do not enqueue rename lock for replay request, because
2016          * if other MDT holds rename lock, but being blocked to wait for
2017          * this MDT to finish its recovery, and the failover MDT can not
2018          * get rename lock, which will cause deadlock.
2019          */
2020         if (!req_is_replay(req)) {
2021                 rc = mdt_rename_lock(info, &rename_lh);
2022                 if (rc != 0) {
2023                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2024                                mdt_obd_name(info->mti_mdt), rc);
2025                         RETURN(rc);
2026                 }
2027         }
2028
2029         /* pobj is master object of parent */
2030         pobj = mdt_parent_find_check(info, rr->rr_fid1, 0);
2031         if (IS_ERR(pobj))
2032                 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2033
2034         if (unlikely(!info->mti_big_lmm)) {
2035                 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
2036                 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
2037                 if (!info->mti_big_lmm)
2038                         GOTO(put_parent, rc = -ENOMEM);
2039         }
2040
2041         ma->ma_lmv = info->mti_big_lmm;
2042         ma->ma_lmv_size = info->mti_big_lmmsize;
2043         ma->ma_valid = 0;
2044         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2045         if (rc)
2046                 GOTO(put_parent, rc);
2047
2048 lock_parent:
2049         /* lock parent object */
2050         lhp = &info->mti_lh[MDT_LH_PARENT];
2051         mdt_lock_reg_init(lhp, LCK_PW);
2052         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2053                                      &parent_slave_locks);
2054         if (rc)
2055                 GOTO(put_parent, rc);
2056
2057         /*
2058          * spobj is the corresponding stripe against name if pobj is striped
2059          * directory, which is the real parent, and no need to lock, because
2060          * we've taken full lock of pobj.
2061          */
2062         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2063         if (rc)
2064                 GOTO(unlock_parent, rc);
2065
2066         /* lock parents of source links, and revoke LOOKUP lock of links */
2067         rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2068                                    &parent_slave_locks, &link_locks);
2069         if (rc == -EBUSY && lock_retries-- > 0) {
2070                 mdt_object_put(env, sobj);
2071                 mdt_object_put(env, spobj);
2072                 goto lock_parent;
2073         }
2074
2075         if (rc < 0)
2076                 GOTO(put_source, rc);
2077
2078         /*
2079          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2080          * with one request, if total lock count exceeds this limit, we will
2081          * drop all locks after migration, and synchronous device in the end.
2082          */
2083         do_sync = rc;
2084
2085         /* TODO: DoM migration is not supported yet */
2086         if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2087                 ma->ma_lmm = info->mti_big_lmm;
2088                 ma->ma_lmm_size = info->mti_big_lmmsize;
2089                 ma->ma_valid = 0;
2090                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2091                 if (rc)
2092                         GOTO(put_source, rc);
2093
2094                 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2095                         GOTO(put_source, rc = -EOPNOTSUPP);
2096         }
2097
2098         /* if migration HSM is allowed */
2099         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2100                 ma->ma_need = MA_HSM;
2101                 ma->ma_valid = 0;
2102                 rc = mdt_attr_get_complex(info, sobj, ma);
2103                 if (rc)
2104                         GOTO(unlock_links, rc);
2105
2106                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2107                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2108         }
2109
2110         /* end lease and close file for regular file */
2111         if (info->mti_spec.sp_migrate_close) {
2112                 /* try to hold open_sem so that nobody else can open the file */
2113                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2114                         /* close anyway */
2115                         mdd_migrate_close(info, sobj);
2116                         GOTO(unlock_links, rc = -EBUSY);
2117                 } else {
2118                         open_sem_locked = true;
2119                         rc = mdd_migrate_close(info, sobj);
2120                         if (rc)
2121                                 GOTO(unlock_open_sem, rc);
2122                 }
2123         }
2124
2125         /* lock source */
2126         lhs = &info->mti_lh[MDT_LH_OLD];
2127         mdt_lock_reg_init(lhs, LCK_EX);
2128         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2129                                      &child_slave_locks);
2130         if (rc)
2131                 GOTO(unlock_open_sem, rc);
2132
2133         /* lock target */
2134         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2135         if (IS_ERR(tobj))
2136                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2137
2138         lht = &info->mti_lh[MDT_LH_NEW];
2139         mdt_lock_reg_init(lht, LCK_EX);
2140         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2141         if (rc)
2142                 GOTO(put_target, rc);
2143
2144         /* Don't do lookup sanity check. We know name doesn't exist. */
2145         info->mti_spec.sp_cr_lookup = 0;
2146         info->mti_spec.sp_feat = &dt_directory_features;
2147
2148         rc = mdo_migrate(env, mdt_object_child(pobj),
2149                          mdt_object_child(sobj), &rr->rr_name,
2150                          mdt_object_child(tobj),
2151                          &info->mti_spec, ma);
2152         if (!rc)
2153                 mdt_counter_incr(req, LPROC_MDT_MIGRATE);
2154         EXIT;
2155
2156         mdt_object_unlock(info, tobj, lht, rc);
2157 put_target:
2158         mdt_object_put(env, tobj);
2159 unlock_source:
2160         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2161                                   &child_slave_locks, rc);
2162 unlock_open_sem:
2163         if (open_sem_locked)
2164                 up_write(&sobj->mot_open_sem);
2165 unlock_links:
2166         /* if we've got too many locks to save into RPC,
2167          * then just commit before the locks are released */
2168         if (!rc && do_sync)
2169                 mdt_device_sync(env, mdt);
2170         mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2171 put_source:
2172         mdt_object_put(env, sobj);
2173         mdt_object_put(env, spobj);
2174 unlock_parent:
2175         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2176                                   &parent_slave_locks, rc);
2177 put_parent:
2178         mdt_object_put(env, pobj);
2179 unlock_rename:
2180         if (lustre_handle_is_used(&rename_lh))
2181                 mdt_rename_unlock(&rename_lh);
2182
2183         return rc;
2184 }
2185
2186 static int mdt_object_lock_save(struct mdt_thread_info *info,
2187                                 struct mdt_object *dir,
2188                                 struct mdt_lock_handle *lh,
2189                                 int idx, bool cos_incompat)
2190 {
2191         int rc;
2192
2193         /* we lock the target dir if it is local */
2194         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2195                                    cos_incompat);
2196         if (rc != 0)
2197                 return rc;
2198
2199         /* get and save correct version after locking */
2200         mdt_version_get_save(info, dir, idx);
2201         return 0;
2202 }
2203
2204 /*
2205  * determine lock order of sobj and tobj
2206  *
2207  * there are two situations we need to lock tobj before sobj:
2208  * 1. sobj is child of tobj
2209  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2210  *    larger than that of tobj
2211  *
2212  * \retval      1 lock tobj before sobj
2213  * \retval      0 lock sobj before tobj
2214  * \retval      -ev negative errno upon error
2215  */
2216 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2217                                            struct mdt_object *sobj,
2218                                            struct mdt_object *tobj)
2219 {
2220         struct md_attr *ma = &info->mti_attr;
2221         struct lu_fid *spfid = &info->mti_tmp_fid1;
2222         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2223         struct lmv_mds_md_v1 *lmv;
2224         __u32 sindex;
2225         __u32 tindex;
2226         int rc;
2227
2228         /* sobj and tobj are the same */
2229         if (sobj == tobj)
2230                 return 0;
2231
2232         if (fid_is_root(mdt_object_fid(sobj)))
2233                 return 0;
2234
2235         if (fid_is_root(mdt_object_fid(tobj)))
2236                 return 1;
2237
2238         /* check whether sobj is child of tobj */
2239         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2240                            mdt_object_fid(tobj));
2241         if (rc < 0)
2242                 return rc;
2243
2244         if (rc == 1)
2245                 return 1;
2246
2247         /* check whether sobj and tobj are children of the same parent */
2248         rc = mdt_attr_get_pfid(info, sobj, spfid);
2249         if (rc)
2250                 return rc;
2251
2252         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2253         if (rc)
2254                 return rc;
2255
2256         if (!lu_fid_eq(spfid, tpfid))
2257                 return 0;
2258
2259         /* check whether sobj and tobj are sibling stripes */
2260         ma->ma_need = MA_LMV;
2261         ma->ma_valid = 0;
2262         ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2263         ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2264         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2265         if (rc)
2266                 return rc;
2267
2268         if (!(ma->ma_valid & MA_LMV))
2269                 return 0;
2270
2271         lmv = &ma->ma_lmv->lmv_md_v1;
2272         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2273                 return 0;
2274         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2275
2276         ma->ma_valid = 0;
2277         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2278         if (rc)
2279                 return rc;
2280
2281         if (!(ma->ma_valid & MA_LMV))
2282                 return -ENODATA;
2283
2284         lmv = &ma->ma_lmv->lmv_md_v1;
2285         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2286                 return -EINVAL;
2287         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2288
2289         /* check stripe index of sobj and tobj */
2290         if (sindex == tindex)
2291                 return -EINVAL;
2292
2293         return sindex < tindex ? 0 : 1;
2294 }
2295
2296 /*
2297  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2298  * 2 - srcdir child; 3 - tgtdir child.
2299  * Update on disk version of srcdir child.
2300  */
2301 static int mdt_reint_rename(struct mdt_thread_info *info,
2302                             struct mdt_lock_handle *unused)
2303 {
2304         struct mdt_device *mdt = info->mti_mdt;
2305         struct mdt_reint_record *rr = &info->mti_rr;
2306         struct md_attr *ma = &info->mti_attr;
2307         struct ptlrpc_request *req = mdt_info_req(info);
2308         struct mdt_object *msrcdir = NULL;
2309         struct mdt_object *mtgtdir = NULL;
2310         struct mdt_object *mold;
2311         struct mdt_object *mnew = NULL;
2312         struct lustre_handle rename_lh = { 0 };
2313         struct mdt_lock_handle *lh_srcdirp;
2314         struct mdt_lock_handle *lh_tgtdirp;
2315         struct mdt_lock_handle *lh_oldp = NULL;
2316         struct mdt_lock_handle *lh_newp = NULL;
2317         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2318         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2319         __u64 lock_ibits;
2320         bool reverse = false, discard = false;
2321         bool cos_incompat;
2322         int rc;
2323         ENTRY;
2324
2325         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2326                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2327                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2328
2329         if (info->mti_dlm_req)
2330                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2331
2332         if (!fid_is_md_operative(rr->rr_fid1) ||
2333             !fid_is_md_operative(rr->rr_fid2))
2334                 RETURN(-EPERM);
2335
2336         /* find both parents. */
2337         msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2338         if (IS_ERR(msrcdir))
2339                 RETURN(PTR_ERR(msrcdir));
2340
2341         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2342
2343         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2344                 mtgtdir = msrcdir;
2345                 mdt_object_get(info->mti_env, mtgtdir);
2346         } else {
2347                 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2348                 if (IS_ERR(mtgtdir))
2349                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2350         }
2351
2352         /*
2353          * Note: do not enqueue rename lock for replay request, because
2354          * if other MDT holds rename lock, but being blocked to wait for
2355          * this MDT to finish its recovery, and the failover MDT can not
2356          * get rename lock, which will cause deadlock.
2357          */
2358         if (!req_is_replay(req)) {
2359                 /*
2360                  * Normally rename RPC is handled on the MDT with the target
2361                  * directory (if target exists, it's on the MDT with the
2362                  * target), if the source directory is remote, it's a hint that
2363                  * source is remote too (this may not be true, but it won't
2364                  * cause any issue), return -EXDEV early to avoid taking
2365                  * rename_lock.
2366                  */
2367                 if (!mdt->mdt_enable_remote_rename &&
2368                     mdt_object_remote(msrcdir))
2369                         GOTO(out_put_tgtdir, rc = -EXDEV);
2370
2371                 rc = mdt_rename_lock(info, &rename_lh);
2372                 if (rc != 0) {
2373                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2374                                mdt_obd_name(mdt), rc);
2375                         GOTO(out_put_tgtdir, rc);
2376                 }
2377         }
2378
2379         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2380         if (rc < 0)
2381                 GOTO(out_unlock_rename, rc);
2382
2383         reverse = rc;
2384
2385         /* source needs to be looked up after locking source parent, otherwise
2386          * this rename may race with unlink source, and cause rename hang, see
2387          * sanityn.sh 55b, so check parents first, if later we found source is
2388          * remote, relock parents. */
2389         cos_incompat = (mdt_object_remote(msrcdir) ||
2390                         mdt_object_remote(mtgtdir));
2391
2392         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2393
2394         /* lock parents in the proper order. */
2395         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2396         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2397
2398 relock:
2399         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2400         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2401
2402         if (reverse) {
2403                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2404                                           cos_incompat);
2405                 if (rc)
2406                         GOTO(out_unlock_rename, rc);
2407
2408                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2409
2410                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2411                                           cos_incompat);
2412                 if (rc != 0) {
2413                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2414                         GOTO(out_unlock_rename, rc);
2415                 }
2416         } else {
2417                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2418                                           cos_incompat);
2419                 if (rc)
2420                         GOTO(out_unlock_rename, rc);
2421
2422                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2423
2424                 if (mtgtdir != msrcdir) {
2425                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2426                                                   cos_incompat);
2427                 } else if (!mdt_object_remote(mtgtdir) &&
2428                            lh_srcdirp->mlh_pdo_hash !=
2429                            lh_tgtdirp->mlh_pdo_hash) {
2430                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2431                                                 MDS_INODELOCK_UPDATE,
2432                                                 cos_incompat);
2433                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2434                 }
2435                 if (rc != 0) {
2436                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2437                         GOTO(out_unlock_rename, rc);
2438                 }
2439         }
2440
2441         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2442         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2443
2444         /* find mold object. */
2445         fid_zero(old_fid);
2446         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2447         if (rc != 0)
2448                 GOTO(out_unlock_parents, rc);
2449
2450         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2451                 GOTO(out_unlock_parents, rc = -EINVAL);
2452
2453         if (!fid_is_md_operative(old_fid))
2454                 GOTO(out_unlock_parents, rc = -EPERM);
2455
2456         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2457         if (IS_ERR(mold))
2458                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2459
2460         if (!mdt_object_exists(mold)) {
2461                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2462                                 &mold->mot_obj,
2463                                 "object does not exist");
2464                 GOTO(out_put_old, rc = -ENOENT);
2465         }
2466
2467         if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2468                 GOTO(out_put_old, rc = -EXDEV);
2469
2470         /* Check if @mtgtdir is subdir of @mold, before locking child
2471          * to avoid reverse locking. */
2472         if (mtgtdir != msrcdir) {
2473                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2474                                    old_fid);
2475                 if (rc) {
2476                         if (rc == 1)
2477                                 rc = -EINVAL;
2478                         GOTO(out_put_old, rc);
2479                 }
2480         }
2481
2482         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2483         /* save version after locking */
2484         mdt_version_get_save(info, mold, 2);
2485
2486         if (!cos_incompat && mdt_object_remote(mold)) {
2487                 cos_incompat = true;
2488                 mdt_object_put(info->mti_env, mold);
2489                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2490                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2491                 goto relock;
2492         }
2493
2494         /* find mnew object:
2495          * mnew target object may not exist now
2496          * lookup with version checking */
2497         fid_zero(new_fid);
2498         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2499                                       3);
2500         if (rc == 0) {
2501                 /* the new_fid should have been filled at this moment */
2502                 if (lu_fid_eq(old_fid, new_fid))
2503                         GOTO(out_put_old, rc);
2504
2505                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2506                     lu_fid_eq(new_fid, rr->rr_fid2))
2507                         GOTO(out_put_old, rc = -EINVAL);
2508
2509                 if (!fid_is_md_operative(new_fid))
2510                         GOTO(out_put_old, rc = -EPERM);
2511
2512                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2513                 if (IS_ERR(mnew))
2514                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2515
2516                 if (!mdt_object_exists(mnew)) {
2517                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2518                                         &mnew->mot_obj,
2519                                         "object does not exist");
2520                         GOTO(out_put_new, rc = -ENOENT);
2521                 }
2522
2523                 if (mdt_object_remote(mnew)) {
2524                         struct mdt_body  *repbody;
2525
2526                         /* Always send rename req to the target child MDT */
2527                         repbody = req_capsule_server_get(info->mti_pill,
2528                                                          &RMF_MDT_BODY);
2529                         LASSERT(repbody != NULL);
2530                         repbody->mbo_fid1 = *new_fid;
2531                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2532                         GOTO(out_put_new, rc = -EXDEV);
2533                 }
2534                 /* Before locking the target dir, check we do not replace
2535                  * a dir with a non-dir, otherwise it may deadlock with
2536                  * link op which tries to create a link in this dir
2537                  * back to this non-dir. */
2538                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2539                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2540                         GOTO(out_put_new, rc = -EISDIR);
2541
2542                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2543                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2544                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2545                 if (mdt_object_remote(msrcdir)) {
2546                         /* Enqueue lookup lock from the parent MDT */
2547                         rc = mdt_remote_object_lock(info, msrcdir,
2548                                                     mdt_object_fid(mold),
2549                                                     &lh_oldp->mlh_rreg_lh,
2550                                                     lh_oldp->mlh_rreg_mode,
2551                                                     MDS_INODELOCK_LOOKUP,
2552                                                     false);
2553                         if (rc != ELDLM_OK)
2554                                 GOTO(out_put_new, rc);
2555
2556                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2557                 }
2558
2559                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2560                                            cos_incompat);
2561                 if (rc != 0)
2562                         GOTO(out_unlock_old, rc);
2563
2564                 /* Check if @msrcdir is subdir of @mnew, before locking child
2565                  * to avoid reverse locking. */
2566                 if (mtgtdir != msrcdir) {
2567                         rc = mdo_is_subdir(info->mti_env,
2568                                            mdt_object_child(msrcdir), new_fid);
2569                         if (rc) {
2570                                 if (rc == 1)
2571                                         rc = -EINVAL;
2572                                 GOTO(out_unlock_old, rc);
2573                         }
2574                 }
2575
2576                 /* We used to acquire MDS_INODELOCK_FULL here but we
2577                  * can't do this now because a running HSM restore on
2578                  * the rename onto victim will hold the layout
2579                  * lock. See LU-4002. */
2580
2581                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2582                 mdt_lock_reg_init(lh_newp, LCK_EX);
2583                 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2584                                            MDS_INODELOCK_LOOKUP |
2585                                            MDS_INODELOCK_UPDATE,
2586                                            cos_incompat);
2587                 if (rc != 0)
2588                         GOTO(out_unlock_old, rc);
2589
2590                 /* get and save version after locking */
2591                 mdt_version_get_save(info, mnew, 3);
2592         } else if (rc != -EREMOTE && rc != -ENOENT) {
2593                 GOTO(out_put_old, rc);
2594         } else {
2595                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2596                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2597                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2598                 if (mdt_object_remote(msrcdir)) {
2599                         /* Enqueue lookup lock from the parent MDT */
2600                         rc = mdt_remote_object_lock(info, msrcdir,
2601                                                     mdt_object_fid(mold),
2602                                                     &lh_oldp->mlh_rreg_lh,
2603                                                     lh_oldp->mlh_rreg_mode,
2604                                                     MDS_INODELOCK_LOOKUP,
2605                                                     false);
2606                         if (rc != ELDLM_OK)
2607                                 GOTO(out_put_old, rc);
2608
2609                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2610                 }
2611
2612                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2613                                            cos_incompat);
2614                 if (rc != 0)
2615                         GOTO(out_unlock_old, rc);
2616
2617                 mdt_enoent_version_save(info, 3);
2618         }
2619
2620         /* step 5: rename it */
2621         mdt_reint_init_ma(info, ma);
2622
2623         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2624                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2625
2626         if (mnew != NULL)
2627                 mutex_lock(&mnew->mot_lov_mutex);
2628
2629         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2630                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2631                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2632                         &rr->rr_tgt_name, ma);
2633
2634         if (mnew != NULL)
2635                 mutex_unlock(&mnew->mot_lov_mutex);
2636
2637         /* handle last link of tgt object */
2638         if (rc == 0) {
2639                 mdt_counter_incr(req, LPROC_MDT_RENAME);
2640                 if (mnew) {
2641                         mdt_handle_last_unlink(info, mnew, ma);
2642                         discard = mdt_dom_check_for_discard(info, mnew);
2643                 }
2644                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2645                                          msrcdir, mtgtdir);
2646         }
2647
2648         EXIT;
2649         if (mnew != NULL)
2650                 mdt_object_unlock(info, mnew, lh_newp, rc);
2651 out_unlock_old:
2652         mdt_object_unlock(info, mold, lh_oldp, rc);
2653 out_put_new:
2654         if (mnew && !discard)
2655                 mdt_object_put(info->mti_env, mnew);
2656 out_put_old:
2657         mdt_object_put(info->mti_env, mold);
2658 out_unlock_parents:
2659         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2660         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2661 out_unlock_rename:
2662         if (lustre_handle_is_used(&rename_lh))
2663                 mdt_rename_unlock(&rename_lh);
2664 out_put_tgtdir:
2665         mdt_object_put(info->mti_env, mtgtdir);
2666 out_put_srcdir:
2667         mdt_object_put(info->mti_env, msrcdir);
2668
2669         /* The DoM discard can be done right in the place above where it is
2670          * assigned, meanwhile it is done here after rename unlock due to
2671          * compatibility with old clients, for them the discard blocks
2672          * the main thread until completion. Check LU-11359 for details.
2673          */
2674         if (discard) {
2675                 mdt_dom_discard_data(info, mnew);
2676                 mdt_object_put(info->mti_env, mnew);
2677         }
2678         return rc;
2679 }
2680
2681 static int mdt_reint_resync(struct mdt_thread_info *info,
2682                             struct mdt_lock_handle *lhc)
2683 {
2684         struct mdt_reint_record *rr = &info->mti_rr;
2685         struct ptlrpc_request *req = mdt_info_req(info);
2686         struct md_attr *ma = &info->mti_attr;
2687         struct mdt_object *mo;
2688         struct ldlm_lock *lease;
2689         struct mdt_body *repbody;
2690         struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2691         bool lease_broken;
2692         int rc, rc2;
2693
2694         ENTRY;
2695
2696         DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2697
2698         if (info->mti_dlm_req)
2699                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2700
2701         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2702         if (IS_ERR(mo))
2703                 GOTO(out, rc = PTR_ERR(mo));
2704
2705         if (!mdt_object_exists(mo))
2706                 GOTO(out_obj, rc = -ENOENT);
2707
2708         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2709                 GOTO(out_obj, rc = -EINVAL);
2710
2711         if (mdt_object_remote(mo))
2712                 GOTO(out_obj, rc = -EREMOTE);
2713
2714         lease = ldlm_handle2lock(rr->rr_lease_handle);
2715         if (lease == NULL)
2716                 GOTO(out_obj, rc = -ESTALE);
2717
2718         /* It's really necessary to grab open_sem and check if the lease lock
2719          * has been lost. There would exist a concurrent writer coming in and
2720          * generating some dirty data in memory cache, the writeback would fail
2721          * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2722         if (!down_write_trylock(&mo->mot_open_sem))
2723                 GOTO(out_put_lease, rc = -EBUSY);
2724
2725         lock_res_and_lock(lease);
2726         lease_broken = ldlm_is_cancel(lease);
2727         unlock_res_and_lock(lease);
2728         if (lease_broken)
2729                 GOTO(out_unlock, rc = -EBUSY);
2730
2731         /* the file has yet opened by anyone else after we took the lease. */
2732         layout.mlc_opc = MD_LAYOUT_RESYNC;
2733         lhc = &info->mti_lh[MDT_LH_LOCAL];
2734         rc = mdt_layout_change(info, mo, lhc, &layout);
2735         if (rc)
2736                 GOTO(out_unlock, rc);
2737
2738         mdt_object_unlock(info, mo, lhc, 0);
2739
2740         ma->ma_need = MA_INODE;
2741         ma->ma_valid = 0;
2742         rc = mdt_attr_get_complex(info, mo, ma);
2743         if (rc != 0)
2744                 GOTO(out_unlock, rc);
2745
2746         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2747         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2748
2749         EXIT;
2750 out_unlock:
2751         up_write(&mo->mot_open_sem);
2752 out_put_lease:
2753         LDLM_LOCK_PUT(lease);
2754 out_obj:
2755         mdt_object_put(info->mti_env, mo);
2756 out:
2757         mdt_client_compatibility(info);
2758         rc2 = mdt_fix_reply(info);
2759         if (rc == 0)
2760                 rc = rc2;
2761         return rc;
2762 }
2763
2764 struct mdt_reinter {
2765         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2766         enum lprocfs_extra_opc mr_extra_opc;
2767 };
2768
2769 static const struct mdt_reinter mdt_reinters[] = {
2770         [REINT_SETATTR] = {
2771                 .mr_handler = &mdt_reint_setattr,
2772                 .mr_extra_opc = MDS_REINT_SETATTR,
2773         },
2774         [REINT_CREATE] = {
2775                 .mr_handler = &mdt_reint_create,
2776                 .mr_extra_opc = MDS_REINT_CREATE,
2777         },
2778         [REINT_LINK] = {
2779                 .mr_handler = &mdt_reint_link,
2780                 .mr_extra_opc = MDS_REINT_LINK,
2781         },
2782         [REINT_UNLINK] = {
2783                 .mr_handler = &mdt_reint_unlink,
2784                 .mr_extra_opc = MDS_REINT_UNLINK,
2785         },
2786         [REINT_RENAME] = {
2787                 .mr_handler = &mdt_reint_rename,
2788                 .mr_extra_opc = MDS_REINT_RENAME,
2789         },
2790         [REINT_OPEN] = {
2791                 .mr_handler = &mdt_reint_open,
2792                 .mr_extra_opc = MDS_REINT_OPEN,
2793         },
2794         [REINT_SETXATTR] = {
2795                 .mr_handler = &mdt_reint_setxattr,
2796                 .mr_extra_opc = MDS_REINT_SETXATTR,
2797         },
2798         [REINT_RMENTRY] = {
2799                 .mr_handler = &mdt_reint_unlink,
2800                 .mr_extra_opc = MDS_REINT_UNLINK,
2801         },
2802         [REINT_MIGRATE] = {
2803                 .mr_handler = &mdt_reint_migrate,
2804                 .mr_extra_opc = MDS_REINT_RENAME,
2805         },
2806         [REINT_RESYNC] = {
2807                 .mr_handler = &mdt_reint_resync,
2808                 .mr_extra_opc = MDS_REINT_RESYNC,
2809         },
2810 };
2811
2812 int mdt_reint_rec(struct mdt_thread_info *info,
2813                   struct mdt_lock_handle *lhc)
2814 {
2815         const struct mdt_reinter *mr;
2816         int rc;
2817         ENTRY;
2818
2819         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2820                 RETURN(-EPROTO);
2821
2822         mr = &mdt_reinters[info->mti_rr.rr_opcode];
2823         if (mr->mr_handler == NULL)
2824                 RETURN(-EPROTO);
2825
2826         rc = (*mr->mr_handler)(info, lhc);
2827
2828         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2829                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
2830
2831         RETURN(rc);
2832 }