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         int rc;
 254
 255         if (!S_ISDIR(obj->mot_header.loh_attr))
 256                 return 0;
 257
 258         rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), &LU_BUF_NULL,
 259                           XATTR_NAME_LMV);
 260         if (rc <= 0)
 261                 return rc == -ENODATA ? 0 : rc;
 262
 263         return 1;
 264 }
 265
 266 /**
 267  * Lock slave stripes if necessary, the lock handles of slave stripes
 268  * will be stored in einfo->ei_cbdata.
 269  **/
 270 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
 271                            enum ldlm_mode mode, __u64 ibits,
 272                            struct ldlm_enqueue_info *einfo)
 273 {
 274         union ldlm_policy_data *policy = &mti->mti_policy;
 275
 276         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 277
 278         einfo->ei_type = LDLM_IBITS;
 279         einfo->ei_mode = mode;
 280         einfo->ei_cb_bl = mdt_remote_blocking_ast;
 281         einfo->ei_cb_local_bl = mdt_blocking_ast;
 282         einfo->ei_cb_cp = ldlm_completion_ast;
 283         einfo->ei_enq_slave = 1;
 284         einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
 285         einfo->ei_inodebits = ibits;
 286         memset(policy, 0, sizeof(*policy));
 287         policy->l_inodebits.bits = ibits;
 288
 289         return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
 290                               policy);
 291 }
 292
 293 int mdt_reint_striped_lock(struct mdt_thread_info *info,
 294                            struct mdt_object *o,
 295                            struct mdt_lock_handle *lh,
 296                            __u64 ibits,
 297                            struct ldlm_enqueue_info *einfo,
 298                            bool cos_incompat)
 299 {
 300         int rc;
 301
 302         LASSERT(!mdt_object_remote(o));
 303
 304         memset(einfo, 0, sizeof(*einfo));
 305
 306         rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
 307         if (rc)
 308                 return rc;
 309
 310         rc = mdt_object_striped(info, o);
 311         if (rc != 1) {
 312                 if (rc < 0)
 313                         mdt_object_unlock(info, o, lh, rc);
 314                 return rc;
 315         }
 316
 317         rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
 318         if (rc) {
 319                 mdt_object_unlock(info, o, lh, rc);
 320                 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
 321                         rc = 0;
 322         }
 323
 324         return rc;
 325 }
 326
 327 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
 328                               struct mdt_object *o,
 329                               struct mdt_lock_handle *lh,
 330                               struct ldlm_enqueue_info *einfo, int decref)
 331 {
 332         if (einfo->ei_cbdata)
 333                 mdt_unlock_slaves(info, o, einfo, decref);
 334         mdt_object_unlock(info, o, lh, decref);
 335 }
 336
 337 /*
 338  * VBR: we save three versions in reply:
 339  * 0 - parent. Check that parent version is the same during replay.
 340  * 1 - name. Version of 'name' if file exists with the same name or
 341  * ENOENT_VERSION, it is needed because file may appear due to missed replays.
 342  * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
 343  * check.
 344  */
 345 static int mdt_create(struct mdt_thread_info *info)
 346 {
 347         struct mdt_device       *mdt = info->mti_mdt;
 348         struct mdt_object       *parent;
 349         struct mdt_object       *child;
 350         struct mdt_lock_handle  *lh;
 351         struct mdt_body         *repbody;
 352         struct md_attr          *ma = &info->mti_attr;
 353         struct mdt_reint_record *rr = &info->mti_rr;
 354         struct md_op_spec       *spec = &info->mti_spec;
 355         int rc;
 356         ENTRY;
 357
 358         DEBUG_REQ(D_INODE, mdt_info_req(info), "Create  ("DNAME"->"DFID") "
 359                   "in "DFID,
 360                   PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
 361
 362         if (!fid_is_md_operative(rr->rr_fid1))
 363                 RETURN(-EPERM);
 364
 365         if (S_ISDIR(ma->ma_attr.la_mode) &&
 366             spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
 367                 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
 368                 struct lu_ucred *uc = mdt_ucred(info);
 369                 struct obd_export *exp = mdt_info_req(info)->rq_export;
 370
 371                 /* Only new clients can create remote dir( >= 2.4) and
 372                  * striped dir(>= 2.6), old client will return -ENOTSUPP */
 373                 if (!mdt_is_dne_client(exp))
 374                         RETURN(-ENOTSUPP);
 375
 376                 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
 377                         if (!mdt_is_striped_client(exp))
 378                                 RETURN(-ENOTSUPP);
 379
 380                         if (!mdt->mdt_enable_striped_dir)
 381                                 RETURN(-EPERM);
 382                 } else if (!mdt->mdt_enable_remote_dir) {
 383                         RETURN(-EPERM);
 384                 }
 385
 386                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 387                     uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 388                     mdt->mdt_enable_remote_dir_gid != -1)
 389                         RETURN(-EPERM);
 390         }
 391
 392         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 393
 394         parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 395         if (IS_ERR(parent))
 396                 RETURN(PTR_ERR(parent));
 397
 398         if (!mdt_object_exists(parent))
 399                 GOTO(put_parent, rc = -ENOENT);
 400
 401         lh = &info->mti_lh[MDT_LH_PARENT];
 402         mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 403         rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
 404         if (rc)
 405                 GOTO(put_parent, rc);
 406
 407         if (!mdt_object_remote(parent)) {
 408                 rc = mdt_version_get_check_save(info, parent, 0);
 409                 if (rc)
 410                         GOTO(unlock_parent, rc);
 411         }
 412
 413         /*
 414          * Check child name version during replay.
 415          * During create replay a file may exist with same name.
 416          */
 417         rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
 418                                       &info->mti_tmp_fid1, 1);
 419         if (rc == 0)
 420                 GOTO(unlock_parent, rc = -EEXIST);
 421
 422         /* -ENOENT is expected here */
 423         if (rc != -ENOENT)
 424                 GOTO(unlock_parent, rc);
 425
 426         /* save version of file name for replay, it must be ENOENT here */
 427         mdt_enoent_version_save(info, 1);
 428
 429         child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
 430         if (unlikely(IS_ERR(child)))
 431                 GOTO(unlock_parent, rc = PTR_ERR(child));
 432
 433         ma->ma_need = MA_INODE;
 434         ma->ma_valid = 0;
 435
 436         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 437                         OBD_FAIL_MDS_REINT_CREATE_WRITE);
 438
 439         /* Version of child will be updated on disk. */
 440         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 441         rc = mdt_version_get_check_save(info, child, 2);
 442         if (rc)
 443                 GOTO(put_child, rc);
 444
 445         /* Let lower layer know current lock mode. */
 446         info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
 447
 448         /*
 449          * Do not perform lookup sanity check. We know that name does
 450          * not exist.
 451          */
 452         info->mti_spec.sp_cr_lookup = 0;
 453         info->mti_spec.sp_feat = &dt_directory_features;
 454
 455         rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
 456                         mdt_object_child(child), &info->mti_spec, ma);
 457         if (rc == 0)
 458                 rc = mdt_attr_get_complex(info, child, ma);
 459
 460         if (rc < 0)
 461                 GOTO(put_child, rc);
 462
 463         /*
 464          * On DNE, we need to eliminate dependey between 'mkdir a' and
 465          * 'mkdir a/b' if b is a striped directory, to achieve this, two
 466          * things are done below:
 467          * 1. save child and slaves lock.
 468          * 2. if the child is a striped directory, relock parent so to
 469          *    compare against with COS locks to ensure parent was
 470          *    committed to disk.
 471          */
 472         if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
 473                 struct mdt_lock_handle *lhc;
 474                 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 475                 bool cos_incompat;
 476
 477                 rc = mdt_object_striped(info, child);
 478                 if (rc < 0)
 479                         GOTO(put_child, rc);
 480
 481                 cos_incompat = rc;
 482                 if (cos_incompat) {
 483                         if (!mdt_object_remote(parent)) {
 484                                 mdt_object_unlock(info, parent, lh, 1);
 485                                 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 486                                 rc = mdt_reint_object_lock(info, parent, lh,
 487                                                            MDS_INODELOCK_UPDATE,
 488                                                            true);
 489                                 if (rc)
 490                                         GOTO(put_child, rc);
 491                         }
 492                 }
 493
 494                 lhc = &info->mti_lh[MDT_LH_CHILD];
 495                 mdt_lock_handle_init(lhc);
 496                 mdt_lock_reg_init(lhc, LCK_PW);
 497                 rc = mdt_reint_striped_lock(info, child, lhc,
 498                                             MDS_INODELOCK_UPDATE, einfo,
 499                                             cos_incompat);
 500                 if (rc)
 501                         GOTO(put_child, rc);
 502
 503                 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 504         }
 505
 506         /* Return fid & attr to client. */
 507         if (ma->ma_valid & MA_INODE)
 508                 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
 509                                    mdt_object_fid(child));
 510 put_child:
 511         mdt_object_put(info->mti_env, child);
 512 unlock_parent:
 513         mdt_object_unlock(info, parent, lh, rc);
 514 put_parent:
 515         mdt_object_put(info->mti_env, parent);
 516         RETURN(rc);
 517 }
 518
 519 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
 520                         struct md_attr *ma)
 521 {
 522         struct mdt_lock_handle  *lh;
 523         int do_vbr = ma->ma_attr.la_valid &
 524                         (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
 525         __u64 lockpart = MDS_INODELOCK_UPDATE;
 526         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 527         bool cos_incompat;
 528         int rc;
 529         ENTRY;
 530
 531         rc = mdt_object_striped(info, mo);
 532         if (rc < 0)
 533                 RETURN(rc);
 534
 535         cos_incompat = rc;
 536
 537         lh = &info->mti_lh[MDT_LH_PARENT];
 538         mdt_lock_reg_init(lh, LCK_PW);
 539
 540         /* Even though the new MDT will grant PERM lock to the old
 541          * client, but the old client will almost ignore that during
 542          * So it needs to revoke both LOOKUP and PERM lock here, so
 543          * both new and old client can cancel the dcache */
 544         if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
 545                 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
 546
 547         rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
 548                                     cos_incompat);
 549         if (rc != 0)
 550                 RETURN(rc);
 551
 552         /* all attrs are packed into mti_attr in unpack_setattr */
 553         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 554                        OBD_FAIL_MDS_REINT_SETATTR_WRITE);
 555
 556         /* VBR: update version if attr changed are important for recovery */
 557         if (do_vbr) {
 558                 /* update on-disk version of changed object */
 559                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
 560                 rc = mdt_version_get_check_save(info, mo, 0);
 561                 if (rc)
 562                         GOTO(out_unlock, rc);
 563         }
 564
 565         /* Ensure constant striping during chown(). See LU-2789. */
 566         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 567                 mutex_lock(&mo->mot_lov_mutex);
 568
 569         /* all attrs are packed into mti_attr in unpack_setattr */
 570         rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
 571
 572         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 573                 mutex_unlock(&mo->mot_lov_mutex);
 574
 575         if (rc != 0)
 576                 GOTO(out_unlock, rc);
 577         mdt_dom_obj_lvb_update(info->mti_env, mo, false);
 578         EXIT;
 579 out_unlock:
 580         mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
 581         return rc;
 582 }
 583
 584 /**
 585  * Check HSM flags and add HS_DIRTY flag if relevant.
 586  *
 587  * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
 588  * and is not RELEASED.
 589  */
 590 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
 591                         struct md_attr *ma)
 592 {
 593         struct lu_ucred *uc = mdt_ucred(info);
 594         cfs_cap_t cap_saved;
 595         int rc;
 596         ENTRY;
 597
 598         /* If the file was modified, add the dirty flag */
 599         ma->ma_need = MA_HSM;
 600         rc = mdt_attr_get_complex(info, mo, ma);
 601         if (rc) {
 602                 CERROR("file attribute read error for "DFID": %d.\n",
 603                         PFID(mdt_object_fid(mo)), rc);
 604                 RETURN(rc);
 605         }
 606
 607         /* If an up2date copy exists in the backend, add dirty flag */
 608         if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
 609             && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
 610                 ma->ma_hsm.mh_flags |= HS_DIRTY;
 611
 612                 /* Bump cap so that closes from non-owner writers can
 613                  * set the HSM state to dirty. */
 614                 cap_saved = uc->uc_cap;
 615                 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
 616                 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
 617                 uc->uc_cap = cap_saved;
 618                 if (rc)
 619                         CERROR("file attribute change error for "DFID": %d\n",
 620                                 PFID(mdt_object_fid(mo)), rc);
 621         }
 622
 623         RETURN(rc);
 624 }
 625
 626 static int mdt_reint_setattr(struct mdt_thread_info *info,
 627                              struct mdt_lock_handle *lhc)
 628 {
 629         struct mdt_device *mdt = info->mti_mdt;
 630         struct md_attr *ma = &info->mti_attr;
 631         struct mdt_reint_record *rr = &info->mti_rr;
 632         struct ptlrpc_request *req = mdt_info_req(info);
 633         struct mdt_object *mo;
 634         struct mdt_body *repbody;
 635         int rc, rc2;
 636         ENTRY;
 637
 638         DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
 639                   (unsigned int)ma->ma_attr.la_valid);
 640
 641         if (info->mti_dlm_req)
 642                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 643
 644         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 645         mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
 646         if (IS_ERR(mo))
 647                 GOTO(out, rc = PTR_ERR(mo));
 648
 649         if (!mdt_object_exists(mo))
 650                 GOTO(out_put, rc = -ENOENT);
 651
 652         if (mdt_object_remote(mo))
 653                 GOTO(out_put, rc = -EREMOTE);
 654
 655         /* revoke lease lock if size is going to be changed */
 656         if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
 657                      !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
 658                      atomic_read(&mo->mot_lease_count) > 0)) {
 659                 down_read(&mo->mot_open_sem);
 660
 661                 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
 662                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 663                         mdt_lock_reg_init(lhc, LCK_CW);
 664
 665                         rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
 666                         if (rc != 0) {
 667                                 up_read(&mo->mot_open_sem);
 668                                 GOTO(out_put, rc);
 669                         }
 670
 671                         /* revoke lease lock */
 672                         mdt_object_unlock(info, mo, lhc, 1);
 673                 }
 674                 up_read(&mo->mot_open_sem);
 675         }
 676
 677         if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
 678                 /* Check write access for the O_TRUNC case */
 679                 if (mdt_write_read(mo) < 0)
 680                         GOTO(out_put, rc = -ETXTBSY);
 681
 682                 /* LU-10286: compatibility check for FLR.
 683                  * Please check the comment in mdt_finish_open() for details */
 684                 if (!exp_connect_flr(info->mti_exp)) {
 685                         rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
 686                         if (rc < 0 && rc != -ENODATA)
 687                                 GOTO(out_put, rc);
 688
 689                         if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm))
 690                                 GOTO(out_put, rc = -EOPNOTSUPP);
 691                 }
 692
 693                 /* For truncate, the file size sent from client
 694                  * is believable, but the blocks are incorrect,
 695                  * which makes the block size in LSOM attribute
 696                  * inconsisent with the real block size.
 697                  */
 698                 rc = mdt_lsom_update(info, mo, true);
 699                 if (rc)
 700                         GOTO(out_put, rc);
 701         }
 702
 703         if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
 704                 if (ma->ma_valid & MA_LOV)
 705                         GOTO(out_put, rc = -EPROTO);
 706
 707                 /* MDT supports FMD for regular files due to Data-on-MDT */
 708                 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
 709                     ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
 710                         tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
 711                                        req->rq_xid);
 712
 713                 rc = mdt_attr_set(info, mo, ma);
 714                 if (rc)
 715                         GOTO(out_put, rc);
 716         } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
 717                    (ma->ma_valid & MA_INODE)) {
 718                 struct lu_buf *buf = &info->mti_buf;
 719                 struct lu_ucred *uc = mdt_ucred(info);
 720                 struct mdt_lock_handle *lh;
 721
 722                 /* reject if either remote or striped dir is disabled */
 723                 if (ma->ma_valid & MA_LMV) {
 724                         if (!mdt->mdt_enable_remote_dir ||
 725                             !mdt->mdt_enable_striped_dir)
 726                                 GOTO(out_put, rc = -EPERM);
 727
 728                         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 729                             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 730                             mdt->mdt_enable_remote_dir_gid != -1)
 731                                 GOTO(out_put, rc = -EPERM);
 732                 }
 733
 734                 if (ma->ma_attr.la_valid != 0)
 735                         GOTO(out_put, rc = -EPROTO);
 736
 737                 lh = &info->mti_lh[MDT_LH_PARENT];
 738                 mdt_lock_reg_init(lh, LCK_PW);
 739
 740                 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
 741                 if (rc != 0)
 742                         GOTO(out_put, rc);
 743
 744                 if (ma->ma_valid & MA_LOV) {
 745                         buf->lb_buf = ma->ma_lmm;
 746                         buf->lb_len = ma->ma_lmm_size;
 747                 } else {
 748                         buf->lb_buf = ma->ma_lmv;
 749                         buf->lb_len = ma->ma_lmv_size;
 750                 }
 751                 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
 752                                   (ma->ma_valid & MA_LOV) ?
 753                                         XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
 754                                   0);
 755
 756                 mdt_object_unlock(info, mo, lh, rc);
 757                 if (rc)
 758                         GOTO(out_put, rc);
 759         } else {
 760                 GOTO(out_put, rc = -EPROTO);
 761         }
 762
 763         /* If file data is modified, add the dirty flag */
 764         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
 765                 rc = mdt_add_dirty_flag(info, mo, ma);
 766
 767         ma->ma_need = MA_INODE;
 768         ma->ma_valid = 0;
 769         rc = mdt_attr_get_complex(info, mo, ma);
 770         if (rc != 0)
 771                 GOTO(out_put, rc);
 772
 773         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
 774
 775         EXIT;
 776 out_put:
 777         mdt_object_put(info->mti_env, mo);
 778 out:
 779         if (rc == 0)
 780                 mdt_counter_incr(req, LPROC_MDT_SETATTR);
 781
 782         mdt_client_compatibility(info);
 783         rc2 = mdt_fix_reply(info);
 784         if (rc == 0)
 785                 rc = rc2;
 786         return rc;
 787 }
 788
 789 static int mdt_reint_create(struct mdt_thread_info *info,
 790                             struct mdt_lock_handle *lhc)
 791 {
 792         struct ptlrpc_request   *req = mdt_info_req(info);
 793         int                     rc;
 794         ENTRY;
 795
 796         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 797                 RETURN(err_serious(-ESTALE));
 798
 799         if (info->mti_dlm_req)
 800                 ldlm_request_cancel(mdt_info_req(info),
 801                                     info->mti_dlm_req, 0, LATF_SKIP);
 802
 803         if (!lu_name_is_valid(&info->mti_rr.rr_name))
 804                 RETURN(-EPROTO);
 805
 806         switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
 807         case S_IFDIR:
 808                 mdt_counter_incr(req, LPROC_MDT_MKDIR);
 809                 break;
 810         case S_IFREG:
 811         case S_IFLNK:
 812         case S_IFCHR:
 813         case S_IFBLK:
 814         case S_IFIFO:
 815         case S_IFSOCK:
 816                 /* Special file should stay on the same node as parent. */
 817                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
 818                 break;
 819         default:
 820                 CERROR("%s: Unsupported mode %o\n",
 821                        mdt_obd_name(info->mti_mdt),
 822                        info->mti_attr.ma_attr.la_mode);
 823                 RETURN(err_serious(-EOPNOTSUPP));
 824         }
 825
 826         rc = mdt_create(info);
 827         RETURN(rc);
 828 }
 829
 830 /*
 831  * VBR: save parent version in reply and child version getting by its name.
 832  * Version of child is getting and checking during its lookup. If
 833  */
 834 static int mdt_reint_unlink(struct mdt_thread_info *info,
 835                             struct mdt_lock_handle *lhc)
 836 {
 837         struct mdt_reint_record *rr = &info->mti_rr;
 838         struct ptlrpc_request *req = mdt_info_req(info);
 839         struct md_attr *ma = &info->mti_attr;
 840         struct lu_fid *child_fid = &info->mti_tmp_fid1;
 841         struct mdt_object *mp;
 842         struct mdt_object *mc;
 843         struct mdt_lock_handle *parent_lh;
 844         struct mdt_lock_handle *child_lh;
 845         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 846         __u64 lock_ibits;
 847         bool cos_incompat = false;
 848         int no_name = 0;
 849         int rc;
 850
 851         ENTRY;
 852
 853         DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
 854                   PNAME(&rr->rr_name));
 855
 856         if (info->mti_dlm_req)
 857                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 858
 859         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
 860                 RETURN(err_serious(-ENOENT));
 861
 862         if (!fid_is_md_operative(rr->rr_fid1))
 863                 RETURN(-EPERM);
 864
 865         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 866         if (IS_ERR(mp))
 867                 RETURN(PTR_ERR(mp));
 868
 869         if (mdt_object_remote(mp)) {
 870                 cos_incompat = true;
 871         } else {
 872                 rc = mdt_version_get_check_save(info, mp, 0);
 873                 if (rc)
 874                         GOTO(put_parent, rc);
 875         }
 876
 877 relock:
 878         parent_lh = &info->mti_lh[MDT_LH_PARENT];
 879         mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
 880         rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
 881                                    cos_incompat);
 882         if (rc != 0)
 883                 GOTO(put_parent, rc);
 884
 885         /* lookup child object along with version checking */
 886         fid_zero(child_fid);
 887         rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
 888         if (rc != 0) {
 889                 /* Name might not be able to find during resend of
 890                  * remote unlink, considering following case.
 891                  * dir_A is a remote directory, the name entry of
 892                  * dir_A is on MDT0, the directory is on MDT1,
 893                  *
 894                  * 1. client sends unlink req to MDT1.
 895                  * 2. MDT1 sends name delete update to MDT0.
 896                  * 3. name entry is being deleted in MDT0 synchronously.
 897                  * 4. MDT1 is restarted.
 898                  * 5. client resends unlink req to MDT1. So it can not
 899                  *    find the name entry on MDT0 anymore.
 900                  * In this case, MDT1 only needs to destory the local
 901                  * directory.
 902                  * */
 903                 if (mdt_object_remote(mp) && rc == -ENOENT &&
 904                     !fid_is_zero(rr->rr_fid2) &&
 905                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 906                         no_name = 1;
 907                         *child_fid = *rr->rr_fid2;
 908                  } else {
 909                         GOTO(unlock_parent, rc);
 910                  }
 911         }
 912
 913         if (!fid_is_md_operative(child_fid))
 914                 GOTO(unlock_parent, rc = -EPERM);
 915
 916         /* We will lock the child regardless it is local or remote. No harm. */
 917         mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
 918         if (IS_ERR(mc))
 919                 GOTO(unlock_parent, rc = PTR_ERR(mc));
 920
 921         if (!cos_incompat) {
 922                 rc = mdt_object_striped(info, mc);
 923                 if (rc < 0)
 924                         GOTO(unlock_parent, rc);
 925
 926                 cos_incompat = rc;
 927                 if (cos_incompat) {
 928                         mdt_object_put(info->mti_env, mc);
 929                         mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
 930                         goto relock;
 931                 }
 932         }
 933
 934         child_lh = &info->mti_lh[MDT_LH_CHILD];
 935         mdt_lock_reg_init(child_lh, LCK_EX);
 936         if (info->mti_spec.sp_rm_entry) {
 937                 struct lu_ucred *uc  = mdt_ucred(info);
 938
 939                 if (!mdt_is_dne_client(req->rq_export))
 940                         /* Return -ENOTSUPP for old client */
 941                         GOTO(put_child, rc = -ENOTSUPP);
 942
 943                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
 944                         GOTO(put_child, rc = -EPERM);
 945
 946                 ma->ma_need = MA_INODE;
 947                 ma->ma_valid = 0;
 948                 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
 949                                 NULL, &rr->rr_name, ma, no_name);
 950                 GOTO(put_child, rc);
 951         }
 952
 953         if (mdt_object_remote(mc)) {
 954                 struct mdt_body  *repbody;
 955
 956                 if (!fid_is_zero(rr->rr_fid2)) {
 957                         CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
 958                                mdt_obd_name(info->mti_mdt),
 959                                PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 960                         GOTO(put_child, rc = -ENOENT);
 961                 }
 962                 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
 963                        mdt_obd_name(info->mti_mdt),
 964                        PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 965
 966                 if (!mdt_is_dne_client(req->rq_export))
 967                         /* Return -ENOTSUPP for old client */
 968                         GOTO(put_child, rc = -ENOTSUPP);
 969
 970                 /* Revoke the LOOKUP lock of the remote object granted by
 971                  * this MDT. Since the unlink will happen on another MDT,
 972                  * it will release the LOOKUP lock right away. Then What
 973                  * would happen if another client try to grab the LOOKUP
 974                  * lock at the same time with unlink XXX */
 975                 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
 976                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 977                 LASSERT(repbody != NULL);
 978                 repbody->mbo_fid1 = *mdt_object_fid(mc);
 979                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 980                 GOTO(unlock_child, rc = -EREMOTE);
 981         }
 982         /* We used to acquire MDS_INODELOCK_FULL here but we can't do
 983          * this now because a running HSM restore on the child (unlink
 984          * victim) will hold the layout lock. See LU-4002. */
 985         lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
 986         if (mdt_object_remote(mp)) {
 987                 /* Enqueue lookup lock from parent MDT */
 988                 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
 989                                             &child_lh->mlh_rreg_lh,
 990                                             child_lh->mlh_rreg_mode,
 991                                             MDS_INODELOCK_LOOKUP, false);
 992                 if (rc != ELDLM_OK)
 993                         GOTO(put_child, rc);
 994
 995                 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
 996         }
 997
 998         rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
 999                                     cos_incompat);
1000         if (rc != 0)
1001                 GOTO(put_child, rc);
1002
1003         /*
1004          * Now we can only make sure we need MA_INODE, in mdd layer, will check
1005          * whether need MA_LOV and MA_COOKIE.
1006          */
1007         ma->ma_need = MA_INODE;
1008         ma->ma_valid = 0;
1009
1010         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1011                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1012         /* save version when object is locked */
1013         mdt_version_get_save(info, mc, 1);
1014
1015         mutex_lock(&mc->mot_lov_mutex);
1016
1017         rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1018                         mdt_object_child(mc), &rr->rr_name, ma, no_name);
1019
1020         mutex_unlock(&mc->mot_lov_mutex);
1021         if (rc != 0)
1022                 GOTO(unlock_child, rc);
1023
1024         if (!lu_object_is_dying(&mc->mot_header)) {
1025                 rc = mdt_attr_get_complex(info, mc, ma);
1026                 if (rc)
1027                         GOTO(out_stat, rc);
1028         } else if (mdt_dom_check_for_discard(info, mc)) {
1029                 mdt_dom_discard_data(info, mc);
1030         }
1031         mdt_handle_last_unlink(info, mc, ma);
1032
1033 out_stat:
1034         if (ma->ma_valid & MA_INODE) {
1035                 switch (ma->ma_attr.la_mode & S_IFMT) {
1036                 case S_IFDIR:
1037                         mdt_counter_incr(req, LPROC_MDT_RMDIR);
1038                         break;
1039                 case S_IFREG:
1040                 case S_IFLNK:
1041                 case S_IFCHR:
1042                 case S_IFBLK:
1043                 case S_IFIFO:
1044                 case S_IFSOCK:
1045                         mdt_counter_incr(req, LPROC_MDT_UNLINK);
1046                         break;
1047                 default:
1048                         LASSERTF(0, "bad file type %o unlinking\n",
1049                                 ma->ma_attr.la_mode);
1050                 }
1051         }
1052
1053         EXIT;
1054
1055 unlock_child:
1056         mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1057 put_child:
1058         mdt_object_put(info->mti_env, mc);
1059 unlock_parent:
1060         mdt_object_unlock(info, mp, parent_lh, rc);
1061 put_parent:
1062         mdt_object_put(info->mti_env, mp);
1063         return rc;
1064 }
1065
1066 /*
1067  * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1068  * name.
1069  */
1070 static int mdt_reint_link(struct mdt_thread_info *info,
1071                           struct mdt_lock_handle *lhc)
1072 {
1073         struct mdt_reint_record *rr = &info->mti_rr;
1074         struct ptlrpc_request   *req = mdt_info_req(info);
1075         struct md_attr          *ma = &info->mti_attr;
1076         struct mdt_object       *ms;
1077         struct mdt_object       *mp;
1078         struct mdt_lock_handle  *lhs;
1079         struct mdt_lock_handle  *lhp;
1080         bool cos_incompat;
1081         int rc;
1082         ENTRY;
1083
1084         DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1085                   PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1086
1087         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1088                 RETURN(err_serious(-ENOENT));
1089
1090         if (info->mti_dlm_req)
1091                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1092
1093         /* Invalid case so return error immediately instead of
1094          * processing it */
1095         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1096                 RETURN(-EPERM);
1097
1098         if (!fid_is_md_operative(rr->rr_fid1) ||
1099             !fid_is_md_operative(rr->rr_fid2))
1100                 RETURN(-EPERM);
1101
1102         /* step 1: find target parent dir */
1103         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1104         if (IS_ERR(mp))
1105                 RETURN(PTR_ERR(mp));
1106
1107         rc = mdt_version_get_check_save(info, mp, 0);
1108         if (rc)
1109                 GOTO(put_parent, rc);
1110
1111         /* step 2: find source */
1112         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1113         if (IS_ERR(ms))
1114                 GOTO(put_parent, rc = PTR_ERR(ms));
1115
1116         if (!mdt_object_exists(ms)) {
1117                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1118                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1119                 GOTO(put_source, rc = -ENOENT);
1120         }
1121
1122         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1123
1124         lhp = &info->mti_lh[MDT_LH_PARENT];
1125         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1126         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1127                                    cos_incompat);
1128         if (rc != 0)
1129                 GOTO(put_source, rc);
1130
1131         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1132
1133         lhs = &info->mti_lh[MDT_LH_CHILD];
1134         mdt_lock_reg_init(lhs, LCK_EX);
1135         rc = mdt_reint_object_lock(info, ms, lhs,
1136                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1137                                    cos_incompat);
1138         if (rc != 0)
1139                 GOTO(unlock_parent, rc);
1140
1141         /* step 3: link it */
1142         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1143                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1144
1145         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1146         rc = mdt_version_get_check_save(info, ms, 1);
1147         if (rc)
1148                 GOTO(unlock_source, rc);
1149
1150         /** check target version by name during replay */
1151         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1152                                       &info->mti_tmp_fid1, 2);
1153         if (rc != 0 && rc != -ENOENT)
1154                 GOTO(unlock_source, rc);
1155         /* save version of file name for replay, it must be ENOENT here */
1156         if (!req_is_replay(mdt_info_req(info))) {
1157                 if (rc != -ENOENT) {
1158                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1159                                PNAME(&rr->rr_name));
1160                         GOTO(unlock_source, rc = -EEXIST);
1161                 }
1162                 info->mti_ver[2] = ENOENT_VERSION;
1163                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1164         }
1165
1166         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1167                       mdt_object_child(ms), &rr->rr_name, ma);
1168
1169         if (rc == 0)
1170                 mdt_counter_incr(req, LPROC_MDT_LINK);
1171
1172         EXIT;
1173 unlock_source:
1174         mdt_object_unlock(info, ms, lhs, rc);
1175 unlock_parent:
1176         mdt_object_unlock(info, mp, lhp, rc);
1177 put_source:
1178         mdt_object_put(info->mti_env, ms);
1179 put_parent:
1180         mdt_object_put(info->mti_env, mp);
1181         return rc;
1182 }
1183 /**
1184  * lock the part of the directory according to the hash of the name
1185  * (lh->mlh_pdo_hash) in parallel directory lock.
1186  */
1187 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1188                               struct mdt_lock_handle *lh,
1189                               struct mdt_object *obj, __u64 ibits,
1190                               bool cos_incompat)
1191 {
1192         struct ldlm_res_id *res = &info->mti_res_id;
1193         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1194         union ldlm_policy_data *policy = &info->mti_policy;
1195         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1196         int rc;
1197
1198         /*
1199          * Finish res_id initializing by name hash marking part of
1200          * directory which is taking modification.
1201          */
1202         LASSERT(lh->mlh_pdo_hash != 0);
1203         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1204         memset(policy, 0, sizeof(*policy));
1205         policy->l_inodebits.bits = ibits;
1206         if (cos_incompat &&
1207             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1208                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1209         /*
1210          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1211          * going to be sent to client. If it is - mdt_intent_policy() path will
1212          * fix it up and turn FL_LOCAL flag off.
1213          */
1214         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1215                           policy, res, dlmflags,
1216                           &info->mti_exp->exp_handle.h_cookie);
1217         return rc;
1218 }
1219
1220 /**
1221  * Get BFL lock for rename or migrate process.
1222  **/
1223 static int mdt_rename_lock(struct mdt_thread_info *info,
1224                            struct lustre_handle *lh)
1225 {
1226         int     rc;
1227         ENTRY;
1228
1229         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1230                 struct lu_fid *fid = &info->mti_tmp_fid1;
1231                 struct mdt_object *obj;
1232
1233                 /* XXX, right now, it has to use object API to
1234                  * enqueue lock cross MDT, so it will enqueue
1235                  * rename lock(with LUSTRE_BFL_FID) by root object */
1236                 lu_root_fid(fid);
1237                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1238                 if (IS_ERR(obj))
1239                         RETURN(PTR_ERR(obj));
1240
1241                 rc = mdt_remote_object_lock(info, obj,
1242                                             &LUSTRE_BFL_FID, lh,
1243                                             LCK_EX,
1244                                             MDS_INODELOCK_UPDATE, false);
1245                 mdt_object_put(info->mti_env, obj);
1246         } else {
1247                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1248                 union ldlm_policy_data *policy = &info->mti_policy;
1249                 struct ldlm_res_id *res_id = &info->mti_res_id;
1250                 __u64 flags = 0;
1251
1252                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1253                 memset(policy, 0, sizeof *policy);
1254                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1255                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1256                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1257                                             LDLM_IBITS, policy, LCK_EX, &flags,
1258                                             ldlm_blocking_ast,
1259                                             ldlm_completion_ast, NULL, NULL, 0,
1260                                             LVB_T_NONE,
1261                                             &info->mti_exp->exp_handle.h_cookie,
1262                                             lh);
1263                 RETURN(rc);
1264         }
1265         RETURN(rc);
1266 }
1267
1268 static void mdt_rename_unlock(struct lustre_handle *lh)
1269 {
1270         ENTRY;
1271         LASSERT(lustre_handle_is_used(lh));
1272         /* Cancel the single rename lock right away */
1273         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1274         EXIT;
1275 }
1276
1277 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1278                                                 const struct lu_fid *fid,
1279                                                 int idx)
1280 {
1281         struct mdt_object *dir;
1282         int rc;
1283
1284         ENTRY;
1285
1286         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1287         if (IS_ERR(dir))
1288                 RETURN(dir);
1289
1290         /* check early, the real version will be saved after locking */
1291         rc = mdt_version_get_check(info, dir, idx);
1292         if (rc)
1293                 GOTO(out_put, rc);
1294
1295         if (!mdt_object_exists(dir))
1296                 GOTO(out_put, rc = -ENOENT);
1297
1298         if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1299                 GOTO(out_put, rc = -ENOTDIR);
1300
1301         RETURN(dir);
1302 out_put:
1303         mdt_object_put(info->mti_env, dir);
1304         return ERR_PTR(rc);
1305 }
1306
1307 /*
1308  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1309  * herein we don't really check it, just do revoke.
1310  */
1311 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1312                                   struct mdt_object *pobj,
1313                                   struct mdt_object *obj)
1314 {
1315         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1316         int rc;
1317
1318         mdt_lock_handle_init(lh);
1319         mdt_lock_reg_init(lh, LCK_EX);
1320
1321         if (mdt_object_remote(pobj)) {
1322                 /* don't bother to check if pobj and obj are on the same MDT. */
1323                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1324                                             &lh->mlh_rreg_lh, LCK_EX,
1325                                             MDS_INODELOCK_LOOKUP, false);
1326         } else if (mdt_object_remote(obj)) {
1327                 struct ldlm_res_id *res = &info->mti_res_id;
1328                 union ldlm_policy_data *policy = &info->mti_policy;
1329                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1330                                  LDLM_FL_COS_INCOMPAT;
1331
1332                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1333                 memset(policy, 0, sizeof(*policy));
1334                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1335                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1336                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1337                                   dlmflags, NULL);
1338         } else {
1339                 /* do nothing if both are local */
1340                 return 0;
1341         }
1342
1343         if (rc != ELDLM_OK)
1344                 return rc;
1345
1346         /*
1347          * TODO, currently we don't save this lock because there is no place to
1348          * hold this lock handle, but to avoid race we need to save this lock.
1349          */
1350         mdt_object_unlock(info, NULL, lh, 1);
1351
1352         return 0;
1353 }
1354
1355 /*
1356  * operation may takes locks of linkea, or directory stripes, group them in
1357  * different list.
1358  */
1359 struct mdt_sub_lock {
1360         struct mdt_object      *msl_obj;
1361         struct mdt_lock_handle  msl_lh;
1362         struct list_head        msl_linkage;
1363 };
1364
1365 static void mdt_unlock_list(struct mdt_thread_info *info,
1366                             struct list_head *list, int decref)
1367 {
1368         struct mdt_sub_lock *msl;
1369         struct mdt_sub_lock *tmp;
1370
1371         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1372                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1373                 list_del(&msl->msl_linkage);
1374                 OBD_FREE_PTR(msl);
1375         }
1376 }
1377
1378 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1379                                              struct mdt_object *obj,
1380                                              struct mdt_lock_handle *lh,
1381                                              struct ldlm_enqueue_info *einfo,
1382                                              struct list_head *slave_locks,
1383                                              int decref)
1384 {
1385         if (mdt_object_remote(obj)) {
1386                 mdt_unlock_list(info, slave_locks, decref);
1387                 mdt_object_unlock(info, obj, lh, decref);
1388         } else {
1389                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1390         }
1391 }
1392
1393 /*
1394  * lock parents of links, and also check whether total locks don't exceed
1395  * RS_MAX_LOCKS.
1396  *
1397  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1398  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1399  * \retval      -ev negative errno upon error
1400  */
1401 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1402                                  struct mdt_object *pobj,
1403                                  const struct md_attr *ma,
1404                                  struct mdt_object *obj,
1405                                  struct mdt_lock_handle *lhp,
1406                                  struct ldlm_enqueue_info *peinfo,
1407                                  struct list_head *parent_slave_locks,
1408                                  struct list_head *link_locks)
1409 {
1410         struct mdt_device *mdt = info->mti_mdt;
1411         struct lu_buf *buf = &info->mti_big_buf;
1412         struct lu_name *lname = &info->mti_name;
1413         struct linkea_data ldata = { NULL };
1414         bool blocked = false;
1415         int local_lnkp_cnt = 0;
1416         int rc;
1417
1418         ENTRY;
1419
1420         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1421                 RETURN(0);
1422
1423         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1424         if (buf->lb_buf == NULL)
1425                 RETURN(-ENOMEM);
1426
1427         ldata.ld_buf = buf;
1428         rc = mdt_links_read(info, obj, &ldata);
1429         if (rc) {
1430                 if (rc == -ENOENT || rc == -ENODATA)
1431                         rc = 0;
1432                 RETURN(rc);
1433         }
1434
1435         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1436              linkea_next_entry(&ldata)) {
1437                 struct mdt_object *lnkp;
1438                 struct mdt_sub_lock *msl;
1439                 struct lu_fid fid;
1440                 __u64 ibits;
1441
1442                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1443                                     &fid);
1444
1445                 /* check if it's also linked to parent */
1446                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1447                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1448                                PFID(&fid), PNAME(lname));
1449                         /* in case link is remote object, revoke LOOKUP lock */
1450                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1451                         continue;
1452                 }
1453
1454                 lnkp = NULL;
1455
1456                 /* check if it's linked to a stripe of parent */
1457                 if (ma->ma_valid & MA_LMV) {
1458                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1459                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1460                         int j = 0;
1461
1462                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1463                                 fid_le_to_cpu(stripe_fid,
1464                                               &lmv->lmv_stripe_fids[j]);
1465                                 if (lu_fid_eq(stripe_fid, &fid)) {
1466                                         CDEBUG(D_INFO, "skip stripe "DFID
1467                                                ", reovke "DNAME"\n",
1468                                                PFID(&fid), PNAME(lname));
1469                                         lnkp = mdt_object_find(info->mti_env,
1470                                                                mdt, &fid);
1471                                         if (IS_ERR(lnkp))
1472                                                 GOTO(out, rc = PTR_ERR(lnkp));
1473                                         break;
1474                                 }
1475                         }
1476
1477                         if (lnkp) {
1478                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1479                                                                    obj);
1480                                 mdt_object_put(info->mti_env, lnkp);
1481                                 continue;
1482                         }
1483                 }
1484
1485                 /* Check if it's already locked */
1486                 list_for_each_entry(msl, link_locks, msl_linkage) {
1487                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1488                                 CDEBUG(D_INFO,
1489                                        DFID" was locked, revoke "DNAME"\n",
1490                                        PFID(&fid), PNAME(lname));
1491                                 lnkp = msl->msl_obj;
1492                                 break;
1493                         }
1494                 }
1495
1496                 if (lnkp) {
1497                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1498                         continue;
1499                 }
1500
1501                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1502                        PFID(&fid), PNAME(lname));
1503
1504                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1505                 if (IS_ERR(lnkp)) {
1506                         CWARN("%s: cannot find obj "DFID": %ld\n",
1507                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1508                         continue;
1509                 }
1510
1511                 if (!mdt_object_exists(lnkp)) {
1512                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1513                               PFID(&fid), PNAME(lname));
1514                         mdt_object_put(info->mti_env, lnkp);
1515                         continue;
1516                 }
1517
1518                 if (!mdt_object_remote(lnkp))
1519                         local_lnkp_cnt++;
1520
1521                 OBD_ALLOC_PTR(msl);
1522                 if (msl == NULL)
1523                         GOTO(out, rc = -ENOMEM);
1524
1525                 /*
1526                  * we can't follow parent-child lock order like other MD
1527                  * operations, use lock_try here to avoid deadlock, if the lock
1528                  * cannot be taken, drop all locks taken, revoke the blocked
1529                  * one, and continue processing the remaining entries, and in
1530                  * the end of the loop restart from beginning.
1531                  */
1532                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1533                 ibits = 0;
1534                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1535                                          MDS_INODELOCK_UPDATE, true);
1536                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1537
1538                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1539                                PFID(&fid), PNAME(lname));
1540
1541                         mdt_unlock_list(info, link_locks, 1);
1542                         /* also unlock parent locks to avoid deadlock */
1543                         if (!blocked)
1544                                 mdt_migrate_object_unlock(info, pobj, lhp,
1545                                                           peinfo,
1546                                                           parent_slave_locks,
1547                                                           1);
1548
1549                         blocked = true;
1550
1551                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1552                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1553                                              MDS_INODELOCK_UPDATE);
1554                         if (rc) {
1555                                 mdt_object_put(info->mti_env, lnkp);
1556                                 OBD_FREE_PTR(msl);
1557                                 GOTO(out, rc);
1558                         }
1559
1560                         if (mdt_object_remote(lnkp)) {
1561                                 struct ldlm_lock *lock;
1562
1563                                 /*
1564                                  * for remote object, set lock cb_atomic,
1565                                  * so lock can be released in blocking_ast()
1566                                  * immediately, then the next lock_try will
1567                                  * have better chance of success.
1568                                  */
1569                                 lock = ldlm_handle2lock(
1570                                                 &msl->msl_lh.mlh_rreg_lh);
1571                                 LASSERT(lock != NULL);
1572                                 lock_res_and_lock(lock);
1573                                 ldlm_set_atomic_cb(lock);
1574                                 unlock_res_and_lock(lock);
1575                                 LDLM_LOCK_PUT(lock);
1576                         }
1577
1578                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1579                         OBD_FREE_PTR(msl);
1580                         continue;
1581                 }
1582
1583                 INIT_LIST_HEAD(&msl->msl_linkage);
1584                 msl->msl_obj = lnkp;
1585                 list_add_tail(&msl->msl_linkage, link_locks);
1586
1587                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1588         }
1589
1590         if (blocked)
1591                 GOTO(out, rc = -EBUSY);
1592
1593         EXIT;
1594 out:
1595         if (rc)
1596                 mdt_unlock_list(info, link_locks, rc);
1597         else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1598                 /*
1599                  * parent may have 3 local objects: master object and 2 stripes
1600                  * (if it's being migrated too); source may have 2 local
1601                  * objects: master and 1 stripe; target has 1 local object.
1602                  */
1603                 rc = 1;
1604         return rc;
1605 }
1606
1607 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1608                                   struct mdt_object *obj,
1609                                   const struct md_attr *ma,
1610                                   struct list_head *slave_locks)
1611 {
1612         struct mdt_device *mdt = info->mti_mdt;
1613         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1614         struct lu_fid *fid = &info->mti_tmp_fid1;
1615         struct mdt_object *slave;
1616         struct mdt_sub_lock *msl;
1617         int i;
1618         int rc;
1619
1620         ENTRY;
1621
1622         LASSERT(mdt_object_remote(obj));
1623         LASSERT(ma->ma_valid & MA_LMV);
1624         LASSERT(lmv);
1625
1626         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1627                 RETURN(-EINVAL);
1628
1629         if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1630                 RETURN(0);
1631
1632         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1633                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1634
1635                 if (!fid_is_sane(fid))
1636                         continue;
1637
1638                 slave = mdt_object_find(info->mti_env, mdt, fid);
1639                 if (IS_ERR(slave))
1640                         GOTO(out, rc = PTR_ERR(slave));
1641
1642                 OBD_ALLOC_PTR(msl);
1643                 if (!msl) {
1644                         mdt_object_put(info->mti_env, slave);
1645                         GOTO(out, rc = -ENOMEM);
1646                 }
1647
1648                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1649                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1650                                            MDS_INODELOCK_UPDATE, true);
1651                 if (rc) {
1652                         OBD_FREE_PTR(msl);
1653                         mdt_object_put(info->mti_env, slave);
1654                         GOTO(out, rc);
1655                 }
1656
1657                 INIT_LIST_HEAD(&msl->msl_linkage);
1658                 msl->msl_obj = slave;
1659                 list_add_tail(&msl->msl_linkage, slave_locks);
1660         }
1661         EXIT;
1662
1663 out:
1664         if (rc)
1665                 mdt_unlock_list(info, slave_locks, rc);
1666         return rc;
1667 }
1668
1669 /* lock parent and its stripes */
1670 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1671                                    struct mdt_object *obj,
1672                                    const struct md_attr *ma,
1673                                    struct mdt_lock_handle *lh,
1674                                    struct ldlm_enqueue_info *einfo,
1675                                    struct list_head *slave_locks)
1676 {
1677         int rc;
1678
1679         if (mdt_object_remote(obj)) {
1680                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1681                                             &lh->mlh_rreg_lh, LCK_PW,
1682                                             MDS_INODELOCK_UPDATE, false);
1683                 if (rc != ELDLM_OK)
1684                         return rc;
1685
1686                 /*
1687                  * if obj is remote and striped, lock its stripes explicitly
1688                  * because it's not striped in LOD layer on this MDT.
1689                  */
1690                 if (ma->ma_valid & MA_LMV) {
1691                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1692                         if (rc)
1693                                 mdt_object_unlock(info, obj, lh, rc);
1694                 }
1695         } else {
1696                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1697                                             einfo, true);
1698         }
1699
1700         return rc;
1701 }
1702
1703 /*
1704  * in migration, object may be remote, and we need take full lock of it and its
1705  * stripes if it's directory, besides, object may be a remote object on its
1706  * parent, revoke its LOOKUP lock on where its parent is located.
1707  */
1708 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1709                                    struct mdt_object *pobj,
1710                                    struct mdt_object *obj,
1711                                    struct mdt_lock_handle *lh,
1712                                    struct ldlm_enqueue_info *einfo,
1713                                    struct list_head *slave_locks)
1714 {
1715         int rc;
1716
1717         if (mdt_object_remote(obj)) {
1718                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1719                 if (rc)
1720                         return rc;
1721
1722                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1723                                             &lh->mlh_rreg_lh, LCK_EX,
1724                                             MDS_INODELOCK_FULL, false);
1725                 if (rc != ELDLM_OK)
1726                         return rc;
1727
1728                 /*
1729                  * if obj is remote and striped, lock its stripes explicitly
1730                  * because it's not striped in LOD layer on this MDT.
1731                  */
1732                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1733                         struct md_attr *ma = &info->mti_attr;
1734
1735                         ma->ma_lmv = info->mti_big_lmm;
1736                         ma->ma_lmv_size = info->mti_big_lmmsize;
1737                         ma->ma_valid = 0;
1738                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1739                         if (rc) {
1740                                 mdt_object_unlock(info, obj, lh, rc);
1741                                 return rc;
1742                         }
1743
1744                         if (ma->ma_valid & MA_LMV) {
1745                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1746                                                             slave_locks);
1747                                 if (rc)
1748                                         mdt_object_unlock(info, obj, lh, rc);
1749                         }
1750                 }
1751         } else {
1752                 if (mdt_object_remote(pobj)) {
1753                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1754                         if (rc)
1755                                 return rc;
1756                 }
1757
1758                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1759                                             einfo, true);
1760         }
1761
1762         return rc;
1763 }
1764
1765 /*
1766  * lookup source by name, if parent is striped directory, we need to find the
1767  * corresponding stripe where source is located, and then lookup there.
1768  *
1769  * besides, if parent is migrating too, and file is already in target stripe,
1770  * this should be a redo of 'lfs migrate' on client side.
1771  */
1772 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1773                               struct mdt_object *pobj,
1774                               const struct md_attr *ma,
1775                               const struct lu_name *lname,
1776                               struct mdt_object **spobj,
1777                               struct mdt_object **sobj)
1778 {
1779         const struct lu_env *env = info->mti_env;
1780         struct lu_fid *fid = &info->mti_tmp_fid1;
1781         struct mdt_object *stripe;
1782         int rc;
1783
1784         if (ma->ma_valid & MA_LMV) {
1785                 /* if parent is striped, lookup on corresponding stripe */
1786                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1787                 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1788                 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1789                 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1790                                     LMV_HASH_FLAG_MIGRATION;
1791
1792                 if (is_migrating) {
1793                         hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1794                         stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1795                 }
1796
1797                 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1798                                               lname->ln_name,
1799                                               lname->ln_namelen);
1800                 if (rc < 0)
1801                         return rc;
1802
1803                 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1804                         rc += le32_to_cpu(lmv->lmv_migrate_offset);
1805
1806                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1807
1808                 stripe = mdt_object_find(env, info->mti_mdt, fid);
1809                 if (IS_ERR(stripe))
1810                         return PTR_ERR(stripe);
1811
1812                 fid_zero(fid);
1813                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1814                                 &info->mti_spec);
1815                 if (rc == -ENOENT && is_migrating) {
1816                         /*
1817                          * if parent is migrating, and lookup child failed on
1818                          * source stripe, lookup again on target stripe, if it
1819                          * exists, it means previous migration was interrupted,
1820                          * and current file was migrated already.
1821                          */
1822                         mdt_object_put(env, stripe);
1823
1824                         hash_type = le32_to_cpu(lmv->lmv_hash_type);
1825                         stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1826
1827                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1828                                                       lname->ln_name,
1829                                                       lname->ln_namelen);
1830                         if (rc < 0)
1831                                 return rc;
1832
1833                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1834
1835                         stripe = mdt_object_find(env, info->mti_mdt, fid);
1836                         if (IS_ERR(stripe))
1837                                 return PTR_ERR(stripe);
1838
1839                         fid_zero(fid);
1840                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1841                                         fid, &info->mti_spec);
1842                         mdt_object_put(env, stripe);
1843                         return rc ?: -EALREADY;
1844                 } else if (rc) {
1845                         mdt_object_put(env, stripe);
1846                         return rc;
1847                 }
1848         } else {
1849                 fid_zero(fid);
1850                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1851                                 &info->mti_spec);
1852                 if (rc)
1853                         return rc;
1854
1855                 stripe = pobj;
1856                 mdt_object_get(env, stripe);
1857         }
1858
1859         *spobj = stripe;
1860
1861         *sobj = mdt_object_find(env, info->mti_mdt, fid);
1862         if (IS_ERR(*sobj)) {
1863                 mdt_object_put(env, stripe);
1864                 rc = PTR_ERR(*sobj);
1865                 *spobj = NULL;
1866                 *sobj = NULL;
1867         }
1868
1869         return rc;
1870 }
1871
1872 /* end lease and close file for regular file */
1873 static int mdd_migrate_close(struct mdt_thread_info *info,
1874                              struct mdt_object *obj)
1875 {
1876         struct close_data *data;
1877         struct mdt_body *repbody;
1878         struct ldlm_lock *lease;
1879         int rc;
1880         int rc2;
1881
1882         rc = -EPROTO;
1883         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1884                                       RCL_CLIENT) ||
1885             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1886                                       RCL_CLIENT))
1887                 goto close;
1888
1889         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1890         if (!data)
1891                 goto close;
1892
1893         rc = -ESTALE;
1894         lease = ldlm_handle2lock(&data->cd_handle);
1895         if (!lease)
1896                 goto close;
1897
1898         /* check if the lease was already canceled */
1899         lock_res_and_lock(lease);
1900         rc = ldlm_is_cancel(lease);
1901         unlock_res_and_lock(lease);
1902
1903         if (rc) {
1904                 rc = -EAGAIN;
1905                 LDLM_DEBUG(lease, DFID" lease broken",
1906                            PFID(mdt_object_fid(obj)));
1907         }
1908
1909         /*
1910          * cancel server side lease, client side counterpart should have been
1911          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1912          */
1913         ldlm_lock_cancel(lease);
1914         ldlm_reprocess_all(lease->l_resource);
1915         LDLM_LOCK_PUT(lease);
1916
1917 close:
1918         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1919         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1920         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1921
1922         return rc ?: rc2;
1923 }
1924
1925 /*
1926  * migrate file in below steps:
1927  *  1. lock parent and its stripes
1928  *  2. lookup source by name
1929  *  3. lock parents of source links if source is not directory
1930  *  4. reject if source is in HSM
1931  *  5. take source open_sem and close file if source is regular file
1932  *  6. lock source and its stripes if it's directory
1933  *  7. lock target so subsequent change to it can trigger COS
1934  *  8. migrate file
1935  *  9. unlock above locks
1936  * 10. sync device if source has links
1937  */
1938 static int mdt_reint_migrate(struct mdt_thread_info *info,
1939                              struct mdt_lock_handle *unused)
1940 {
1941         const struct lu_env *env = info->mti_env;
1942         struct mdt_device *mdt = info->mti_mdt;
1943         struct ptlrpc_request *req = mdt_info_req(info);
1944         struct mdt_reint_record *rr = &info->mti_rr;
1945         struct lu_ucred *uc = mdt_ucred(info);
1946         struct md_attr *ma = &info->mti_attr;
1947         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1948         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1949         struct mdt_object *pobj;
1950         struct mdt_object *spobj = NULL;
1951         struct mdt_object *sobj = NULL;
1952         struct mdt_object *tobj;
1953         struct lustre_handle rename_lh = { 0 };
1954         struct mdt_lock_handle *lhp;
1955         struct mdt_lock_handle *lhs;
1956         struct mdt_lock_handle *lht;
1957         LIST_HEAD(parent_slave_locks);
1958         LIST_HEAD(child_slave_locks);
1959         LIST_HEAD(link_locks);
1960         int lock_retries = 5;
1961         bool open_sem_locked = false;
1962         bool do_sync = false;
1963         int rc;
1964         ENTRY;
1965
1966         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1967                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1968
1969         if (info->mti_dlm_req)
1970                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1971
1972         if (!fid_is_md_operative(rr->rr_fid1) ||
1973             !fid_is_md_operative(rr->rr_fid2))
1974                 RETURN(-EPERM);
1975
1976         /* don't allow migrate . or .. */
1977         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
1978                 RETURN(-EBUSY);
1979
1980         if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
1981                 RETURN(-EPERM);
1982
1983         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
1984             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
1985             mdt->mdt_enable_remote_dir_gid != -1)
1986                 RETURN(-EPERM);
1987
1988         /*
1989          * Note: do not enqueue rename lock for replay request, because
1990          * if other MDT holds rename lock, but being blocked to wait for
1991          * this MDT to finish its recovery, and the failover MDT can not
1992          * get rename lock, which will cause deadlock.
1993          */
1994         if (!req_is_replay(req)) {
1995                 rc = mdt_rename_lock(info, &rename_lh);
1996                 if (rc != 0) {
1997                         CERROR("%s: can't lock FS for rename: rc = %d\n",
1998                                mdt_obd_name(info->mti_mdt), rc);
1999                         RETURN(rc);
2000                 }
2001         }
2002
2003         /* pobj is master object of parent */
2004         pobj = mdt_parent_find_check(info, rr->rr_fid1, 0);
2005         if (IS_ERR(pobj))
2006                 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2007
2008         if (unlikely(!info->mti_big_lmm)) {
2009                 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
2010                 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
2011                 if (!info->mti_big_lmm)
2012                         GOTO(put_parent, rc = -ENOMEM);
2013         }
2014
2015         ma->ma_lmv = info->mti_big_lmm;
2016         ma->ma_lmv_size = info->mti_big_lmmsize;
2017         ma->ma_valid = 0;
2018         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2019         if (rc)
2020                 GOTO(put_parent, rc);
2021
2022 lock_parent:
2023         /* lock parent object */
2024         lhp = &info->mti_lh[MDT_LH_PARENT];
2025         mdt_lock_reg_init(lhp, LCK_PW);
2026         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2027                                      &parent_slave_locks);
2028         if (rc)
2029                 GOTO(put_parent, rc);
2030
2031         /*
2032          * spobj is the corresponding stripe against name if pobj is striped
2033          * directory, which is the real parent, and no need to lock, because
2034          * we've taken full lock of pobj.
2035          */
2036         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2037         if (rc)
2038                 GOTO(unlock_parent, rc);
2039
2040         /* lock parents of source links, and revoke LOOKUP lock of links */
2041         rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2042                                    &parent_slave_locks, &link_locks);
2043         if (rc == -EBUSY && lock_retries-- > 0) {
2044                 mdt_object_put(env, sobj);
2045                 mdt_object_put(env, spobj);
2046                 goto lock_parent;
2047         }
2048
2049         if (rc < 0)
2050                 GOTO(put_source, rc);
2051
2052         /*
2053          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2054          * with one request, if total lock count exceeds this limit, we will
2055          * drop all locks after migration, and synchronous device in the end.
2056          */
2057         do_sync = rc;
2058
2059         /* TODO: DoM migration is not supported yet */
2060         if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2061                 ma->ma_lmm = info->mti_big_lmm;
2062                 ma->ma_lmm_size = info->mti_big_lmmsize;
2063                 ma->ma_valid = 0;
2064                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2065                 if (rc)
2066                         GOTO(put_source, rc);
2067
2068                 if (ma->ma_valid & MA_LOV &&
2069                     mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2070                         GOTO(put_source, rc = -EOPNOTSUPP);
2071         }
2072
2073         /* if migration HSM is allowed */
2074         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2075                 ma->ma_need = MA_HSM;
2076                 ma->ma_valid = 0;
2077                 rc = mdt_attr_get_complex(info, sobj, ma);
2078                 if (rc)
2079                         GOTO(unlock_links, rc);
2080
2081                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2082                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2083         }
2084
2085         /* end lease and close file for regular file */
2086         if (info->mti_spec.sp_migrate_close) {
2087                 /* try to hold open_sem so that nobody else can open the file */
2088                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2089                         /* close anyway */
2090                         mdd_migrate_close(info, sobj);
2091                         GOTO(unlock_links, rc = -EBUSY);
2092                 } else {
2093                         open_sem_locked = true;
2094                         rc = mdd_migrate_close(info, sobj);
2095                         if (rc)
2096                                 GOTO(unlock_open_sem, rc);
2097                 }
2098         }
2099
2100         /* lock source */
2101         lhs = &info->mti_lh[MDT_LH_OLD];
2102         mdt_lock_reg_init(lhs, LCK_EX);
2103         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2104                                      &child_slave_locks);
2105         if (rc)
2106                 GOTO(unlock_open_sem, rc);
2107
2108         /* lock target */
2109         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2110         if (IS_ERR(tobj))
2111                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2112
2113         lht = &info->mti_lh[MDT_LH_NEW];
2114         mdt_lock_reg_init(lht, LCK_EX);
2115         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2116         if (rc)
2117                 GOTO(put_target, rc);
2118
2119         /* Don't do lookup sanity check. We know name doesn't exist. */
2120         info->mti_spec.sp_cr_lookup = 0;
2121         info->mti_spec.sp_feat = &dt_directory_features;
2122
2123         rc = mdo_migrate(env, mdt_object_child(pobj),
2124                          mdt_object_child(sobj), &rr->rr_name,
2125                          mdt_object_child(tobj), &info->mti_spec, ma);
2126         EXIT;
2127
2128         mdt_object_unlock(info, tobj, lht, rc);
2129 put_target:
2130         mdt_object_put(env, tobj);
2131 unlock_source:
2132         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2133                                   &child_slave_locks, rc);
2134 unlock_open_sem:
2135         if (open_sem_locked)
2136                 up_write(&sobj->mot_open_sem);
2137 unlock_links:
2138         mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2139 put_source:
2140         mdt_object_put(env, sobj);
2141         mdt_object_put(env, spobj);
2142 unlock_parent:
2143         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2144                                   &parent_slave_locks, rc);
2145 put_parent:
2146         mdt_object_put(env, pobj);
2147 unlock_rename:
2148         if (lustre_handle_is_used(&rename_lh))
2149                 mdt_rename_unlock(&rename_lh);
2150
2151         if (!rc && do_sync)
2152                 mdt_device_sync(env, mdt);
2153
2154         return rc;
2155 }
2156
2157 static int mdt_object_lock_save(struct mdt_thread_info *info,
2158                                 struct mdt_object *dir,
2159                                 struct mdt_lock_handle *lh,
2160                                 int idx, bool cos_incompat)
2161 {
2162         int rc;
2163
2164         /* we lock the target dir if it is local */
2165         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2166                                    cos_incompat);
2167         if (rc != 0)
2168                 return rc;
2169
2170         /* get and save correct version after locking */
2171         mdt_version_get_save(info, dir, idx);
2172         return 0;
2173 }
2174
2175 /*
2176  * determine lock order of sobj and tobj
2177  *
2178  * there are two situations we need to lock tobj before sobj:
2179  * 1. sobj is child of tobj
2180  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2181  *    larger than that of tobj
2182  *
2183  * \retval      1 lock tobj before sobj
2184  * \retval      0 lock sobj before tobj
2185  * \retval      -ev negative errno upon error
2186  */
2187 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2188                                            struct mdt_object *sobj,
2189                                            struct mdt_object *tobj)
2190 {
2191         struct md_attr *ma = &info->mti_attr;
2192         struct lu_fid *spfid = &info->mti_tmp_fid1;
2193         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2194         struct lmv_mds_md_v1 *lmv;
2195         __u32 sindex;
2196         __u32 tindex;
2197         int rc;
2198
2199         /* sobj and tobj are the same */
2200         if (sobj == tobj)
2201                 return 0;
2202
2203         if (fid_is_root(mdt_object_fid(sobj)))
2204                 return 0;
2205
2206         if (fid_is_root(mdt_object_fid(tobj)))
2207                 return 1;
2208
2209         /* check whether sobj is child of tobj */
2210         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2211                            mdt_object_fid(tobj));
2212         if (rc < 0)
2213                 return rc;
2214
2215         if (rc == 1)
2216                 return 1;
2217
2218         /* check whether sobj and tobj are children of the same parent */
2219         rc = mdt_attr_get_pfid(info, sobj, spfid);
2220         if (rc)
2221                 return rc;
2222
2223         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2224         if (rc)
2225                 return rc;
2226
2227         if (!lu_fid_eq(spfid, tpfid))
2228                 return 0;
2229
2230         /* check whether sobj and tobj are sibling stripes */
2231         ma->ma_need = MA_LMV;
2232         ma->ma_valid = 0;
2233         ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2234         ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2235         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2236         if (rc)
2237                 return rc;
2238
2239         if (!(ma->ma_valid & MA_LMV))
2240                 return 0;
2241
2242         lmv = &ma->ma_lmv->lmv_md_v1;
2243         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2244                 return 0;
2245         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2246
2247         ma->ma_valid = 0;
2248         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2249         if (rc)
2250                 return rc;
2251
2252         if (!(ma->ma_valid & MA_LMV))
2253                 return -ENODATA;
2254
2255         lmv = &ma->ma_lmv->lmv_md_v1;
2256         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2257                 return -EINVAL;
2258         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2259
2260         /* check stripe index of sobj and tobj */
2261         if (sindex == tindex)
2262                 return -EINVAL;
2263
2264         return sindex < tindex ? 0 : 1;
2265 }
2266
2267 /*
2268  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2269  * 2 - srcdir child; 3 - tgtdir child.
2270  * Update on disk version of srcdir child.
2271  */
2272 static int mdt_reint_rename(struct mdt_thread_info *info,
2273                             struct mdt_lock_handle *unused)
2274 {
2275         struct mdt_device *mdt = info->mti_mdt;
2276         struct mdt_reint_record *rr = &info->mti_rr;
2277         struct md_attr *ma = &info->mti_attr;
2278         struct ptlrpc_request *req = mdt_info_req(info);
2279         struct mdt_object *msrcdir = NULL;
2280         struct mdt_object *mtgtdir = NULL;
2281         struct mdt_object *mold;
2282         struct mdt_object *mnew = NULL;
2283         struct lustre_handle rename_lh = { 0 };
2284         struct mdt_lock_handle *lh_srcdirp;
2285         struct mdt_lock_handle *lh_tgtdirp;
2286         struct mdt_lock_handle *lh_oldp = NULL;
2287         struct mdt_lock_handle *lh_newp = NULL;
2288         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2289         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2290         __u64 lock_ibits;
2291         bool reverse = false, discard = false;
2292         bool cos_incompat;
2293         int rc;
2294         ENTRY;
2295
2296         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2297                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2298                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2299
2300         if (info->mti_dlm_req)
2301                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2302
2303         if (!fid_is_md_operative(rr->rr_fid1) ||
2304             !fid_is_md_operative(rr->rr_fid2))
2305                 RETURN(-EPERM);
2306
2307         /* find both parents. */
2308         msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2309         if (IS_ERR(msrcdir))
2310                 RETURN(PTR_ERR(msrcdir));
2311
2312         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2313
2314         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2315                 mtgtdir = msrcdir;
2316                 mdt_object_get(info->mti_env, mtgtdir);
2317         } else {
2318                 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2319                 if (IS_ERR(mtgtdir))
2320                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2321         }
2322
2323         /*
2324          * Note: do not enqueue rename lock for replay request, because
2325          * if other MDT holds rename lock, but being blocked to wait for
2326          * this MDT to finish its recovery, and the failover MDT can not
2327          * get rename lock, which will cause deadlock.
2328          */
2329         if (!req_is_replay(req)) {
2330                 /*
2331                  * Normally rename RPC is handled on the MDT with the target
2332                  * directory (if target exists, it's on the MDT with the
2333                  * target), if the source directory is remote, it's a hint that
2334                  * source is remote too (this may not be true, but it won't
2335                  * cause any issue), return -EXDEV early to avoid taking
2336                  * rename_lock.
2337                  */
2338                 if (!mdt->mdt_enable_remote_rename &&
2339                     mdt_object_remote(msrcdir))
2340                         GOTO(out_put_tgtdir, rc = -EXDEV);
2341
2342                 rc = mdt_rename_lock(info, &rename_lh);
2343                 if (rc != 0) {
2344                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2345                                mdt_obd_name(mdt), rc);
2346                         GOTO(out_put_tgtdir, rc);
2347                 }
2348         }
2349
2350         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2351         if (rc < 0)
2352                 GOTO(out_unlock_rename, rc);
2353
2354         reverse = rc;
2355
2356         /* source needs to be looked up after locking source parent, otherwise
2357          * this rename may race with unlink source, and cause rename hang, see
2358          * sanityn.sh 55b, so check parents first, if later we found source is
2359          * remote, relock parents. */
2360         cos_incompat = (mdt_object_remote(msrcdir) ||
2361                         mdt_object_remote(mtgtdir));
2362
2363         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2364
2365         /* lock parents in the proper order. */
2366         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2367         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2368
2369 relock:
2370         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2371         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2372
2373         if (reverse) {
2374                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2375                                           cos_incompat);
2376                 if (rc)
2377                         GOTO(out_unlock_rename, rc);
2378
2379                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2380
2381                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2382                                           cos_incompat);
2383                 if (rc != 0) {
2384                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2385                         GOTO(out_unlock_rename, rc);
2386                 }
2387         } else {
2388                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2389                                           cos_incompat);
2390                 if (rc)
2391                         GOTO(out_unlock_rename, rc);
2392
2393                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2394
2395                 if (mtgtdir != msrcdir) {
2396                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2397                                                   cos_incompat);
2398                 } else if (!mdt_object_remote(mtgtdir) &&
2399                            lh_srcdirp->mlh_pdo_hash !=
2400                            lh_tgtdirp->mlh_pdo_hash) {
2401                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2402                                                 MDS_INODELOCK_UPDATE,
2403                                                 cos_incompat);
2404                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2405                 }
2406                 if (rc != 0) {
2407                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2408                         GOTO(out_unlock_rename, rc);
2409                 }
2410         }
2411
2412         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2413         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2414
2415         /* find mold object. */
2416         fid_zero(old_fid);
2417         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2418         if (rc != 0)
2419                 GOTO(out_unlock_parents, rc);
2420
2421         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2422                 GOTO(out_unlock_parents, rc = -EINVAL);
2423
2424         if (!fid_is_md_operative(old_fid))
2425                 GOTO(out_unlock_parents, rc = -EPERM);
2426
2427         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2428         if (IS_ERR(mold))
2429                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2430
2431         if (!mdt_object_exists(mold)) {
2432                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2433                                 &mold->mot_obj,
2434                                 "object does not exist");
2435                 GOTO(out_put_old, rc = -ENOENT);
2436         }
2437
2438         if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2439                 GOTO(out_put_old, rc = -EXDEV);
2440
2441         /* Check if @mtgtdir is subdir of @mold, before locking child
2442          * to avoid reverse locking. */
2443         if (mtgtdir != msrcdir) {
2444                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2445                                    old_fid);
2446                 if (rc) {
2447                         if (rc == 1)
2448                                 rc = -EINVAL;
2449                         GOTO(out_put_old, rc);
2450                 }
2451         }
2452
2453         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2454         /* save version after locking */
2455         mdt_version_get_save(info, mold, 2);
2456
2457         if (!cos_incompat && mdt_object_remote(mold)) {
2458                 cos_incompat = true;
2459                 mdt_object_put(info->mti_env, mold);
2460                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2461                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2462                 goto relock;
2463         }
2464
2465         /* find mnew object:
2466          * mnew target object may not exist now
2467          * lookup with version checking */
2468         fid_zero(new_fid);
2469         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2470                                       3);
2471         if (rc == 0) {
2472                 /* the new_fid should have been filled at this moment */
2473                 if (lu_fid_eq(old_fid, new_fid))
2474                         GOTO(out_put_old, rc);
2475
2476                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2477                     lu_fid_eq(new_fid, rr->rr_fid2))
2478                         GOTO(out_put_old, rc = -EINVAL);
2479
2480                 if (!fid_is_md_operative(new_fid))
2481                         GOTO(out_put_old, rc = -EPERM);
2482
2483                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2484                 if (IS_ERR(mnew))
2485                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2486
2487                 if (!mdt_object_exists(mnew)) {
2488                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2489                                         &mnew->mot_obj,
2490                                         "object does not exist");
2491                         GOTO(out_put_new, rc = -ENOENT);
2492                 }
2493
2494                 if (mdt_object_remote(mnew)) {
2495                         struct mdt_body  *repbody;
2496
2497                         /* Always send rename req to the target child MDT */
2498                         repbody = req_capsule_server_get(info->mti_pill,
2499                                                          &RMF_MDT_BODY);
2500                         LASSERT(repbody != NULL);
2501                         repbody->mbo_fid1 = *new_fid;
2502                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2503                         GOTO(out_put_new, rc = -EXDEV);
2504                 }
2505                 /* Before locking the target dir, check we do not replace
2506                  * a dir with a non-dir, otherwise it may deadlock with
2507                  * link op which tries to create a link in this dir
2508                  * back to this non-dir. */
2509                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2510                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2511                         GOTO(out_put_new, rc = -EISDIR);
2512
2513                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2514                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2515                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2516                 if (mdt_object_remote(msrcdir)) {
2517                         /* Enqueue lookup lock from the parent MDT */
2518                         rc = mdt_remote_object_lock(info, msrcdir,
2519                                                     mdt_object_fid(mold),
2520                                                     &lh_oldp->mlh_rreg_lh,
2521                                                     lh_oldp->mlh_rreg_mode,
2522                                                     MDS_INODELOCK_LOOKUP,
2523                                                     false);
2524                         if (rc != ELDLM_OK)
2525                                 GOTO(out_put_new, rc);
2526
2527                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2528                 }
2529
2530                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2531                                            cos_incompat);
2532                 if (rc != 0)
2533                         GOTO(out_unlock_old, rc);
2534
2535                 /* Check if @msrcdir is subdir of @mnew, before locking child
2536                  * to avoid reverse locking. */
2537                 if (mtgtdir != msrcdir) {
2538                         rc = mdo_is_subdir(info->mti_env,
2539                                            mdt_object_child(msrcdir), new_fid);
2540                         if (rc) {
2541                                 if (rc == 1)
2542                                         rc = -EINVAL;
2543                                 GOTO(out_unlock_old, rc);
2544                         }
2545                 }
2546
2547                 /* We used to acquire MDS_INODELOCK_FULL here but we
2548                  * can't do this now because a running HSM restore on
2549                  * the rename onto victim will hold the layout
2550                  * lock. See LU-4002. */
2551
2552                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2553                 mdt_lock_reg_init(lh_newp, LCK_EX);
2554                 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2555                                            MDS_INODELOCK_LOOKUP |
2556                                            MDS_INODELOCK_UPDATE,
2557                                            cos_incompat);
2558                 if (rc != 0)
2559                         GOTO(out_unlock_old, rc);
2560
2561                 /* get and save version after locking */
2562                 mdt_version_get_save(info, mnew, 3);
2563         } else if (rc != -EREMOTE && rc != -ENOENT) {
2564                 GOTO(out_put_old, rc);
2565         } else {
2566                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2567                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2568                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2569                 if (mdt_object_remote(msrcdir)) {
2570                         /* Enqueue lookup lock from the parent MDT */
2571                         rc = mdt_remote_object_lock(info, msrcdir,
2572                                                     mdt_object_fid(mold),
2573                                                     &lh_oldp->mlh_rreg_lh,
2574                                                     lh_oldp->mlh_rreg_mode,
2575                                                     MDS_INODELOCK_LOOKUP,
2576                                                     false);
2577                         if (rc != ELDLM_OK)
2578                                 GOTO(out_put_old, rc);
2579
2580                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2581                 }
2582
2583                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2584                                            cos_incompat);
2585                 if (rc != 0)
2586                         GOTO(out_unlock_old, rc);
2587
2588                 mdt_enoent_version_save(info, 3);
2589         }
2590
2591         /* step 5: rename it */
2592         mdt_reint_init_ma(info, ma);
2593
2594         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2595                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2596
2597         if (mnew != NULL)
2598                 mutex_lock(&mnew->mot_lov_mutex);
2599
2600         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2601                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2602                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2603                         &rr->rr_tgt_name, ma);
2604
2605         if (mnew != NULL)
2606                 mutex_unlock(&mnew->mot_lov_mutex);
2607
2608         /* handle last link of tgt object */
2609         if (rc == 0) {
2610                 mdt_counter_incr(req, LPROC_MDT_RENAME);
2611                 if (mnew) {
2612                         mdt_handle_last_unlink(info, mnew, ma);
2613                         discard = mdt_dom_check_for_discard(info, mnew);
2614                 }
2615                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2616                                          msrcdir, mtgtdir);
2617         }
2618
2619         EXIT;
2620         if (mnew != NULL)
2621                 mdt_object_unlock(info, mnew, lh_newp, rc);
2622 out_unlock_old:
2623         mdt_object_unlock(info, mold, lh_oldp, rc);
2624 out_put_new:
2625         if (mnew && !discard)
2626                 mdt_object_put(info->mti_env, mnew);
2627 out_put_old:
2628         mdt_object_put(info->mti_env, mold);
2629 out_unlock_parents:
2630         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2631         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2632 out_unlock_rename:
2633         if (lustre_handle_is_used(&rename_lh))
2634                 mdt_rename_unlock(&rename_lh);
2635 out_put_tgtdir:
2636         mdt_object_put(info->mti_env, mtgtdir);
2637 out_put_srcdir:
2638         mdt_object_put(info->mti_env, msrcdir);
2639
2640         /* The DoM discard can be done right in the place above where it is
2641          * assigned, meanwhile it is done here after rename unlock due to
2642          * compatibility with old clients, for them the discard blocks
2643          * the main thread until completion. Check LU-11359 for details.
2644          */
2645         if (discard) {
2646                 mdt_dom_discard_data(info, mnew);
2647                 mdt_object_put(info->mti_env, mnew);
2648         }
2649         return rc;
2650 }
2651
2652 static int mdt_reint_resync(struct mdt_thread_info *info,
2653                             struct mdt_lock_handle *lhc)
2654 {
2655         struct mdt_reint_record *rr = &info->mti_rr;
2656         struct ptlrpc_request   *req = mdt_info_req(info);
2657         struct md_attr          *ma = &info->mti_attr;
2658         struct mdt_object       *mo;
2659         struct ldlm_lock        *lease;
2660         struct mdt_body         *repbody;
2661         struct md_layout_change  layout = { .mlc_mirror_id = rr->rr_mirror_id };
2662         bool                     lease_broken;
2663         int                      rc, rc2;
2664         ENTRY;
2665
2666         DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2667
2668         if (info->mti_dlm_req)
2669                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2670
2671         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2672         if (IS_ERR(mo))
2673                 GOTO(out, rc = PTR_ERR(mo));
2674
2675         if (!mdt_object_exists(mo))
2676                 GOTO(out_obj, rc = -ENOENT);
2677
2678         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2679                 GOTO(out_obj, rc = -EINVAL);
2680
2681         if (mdt_object_remote(mo))
2682                 GOTO(out_obj, rc = -EREMOTE);
2683
2684         lease = ldlm_handle2lock(rr->rr_lease_handle);
2685         if (lease == NULL)
2686                 GOTO(out_obj, rc = -ESTALE);
2687
2688         /* It's really necessary to grab open_sem and check if the lease lock
2689          * has been lost. There would exist a concurrent writer coming in and
2690          * generating some dirty data in memory cache, the writeback would fail
2691          * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2692         if (!down_write_trylock(&mo->mot_open_sem))
2693                 GOTO(out_put_lease, rc = -EBUSY);
2694
2695         lock_res_and_lock(lease);
2696         lease_broken = ldlm_is_cancel(lease);
2697         unlock_res_and_lock(lease);
2698         if (lease_broken)
2699                 GOTO(out_unlock, rc = -EBUSY);
2700
2701         /* the file has yet opened by anyone else after we took the lease. */
2702         layout.mlc_opc = MD_LAYOUT_RESYNC;
2703         rc = mdt_layout_change(info, mo, &layout);
2704         if (rc)
2705                 GOTO(out_unlock, rc);
2706
2707         ma->ma_need = MA_INODE;
2708         ma->ma_valid = 0;
2709         rc = mdt_attr_get_complex(info, mo, ma);
2710         if (rc != 0)
2711                 GOTO(out_unlock, rc);
2712
2713         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2714         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2715
2716         EXIT;
2717 out_unlock:
2718         up_write(&mo->mot_open_sem);
2719 out_put_lease:
2720         LDLM_LOCK_PUT(lease);
2721 out_obj:
2722         mdt_object_put(info->mti_env, mo);
2723 out:
2724         mdt_client_compatibility(info);
2725         rc2 = mdt_fix_reply(info);
2726         if (rc == 0)
2727                 rc = rc2;
2728         return rc;
2729 }
2730
2731 struct mdt_reinter {
2732         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2733         enum lprocfs_extra_opc mr_extra_opc;
2734 };
2735
2736 static const struct mdt_reinter mdt_reinters[] = {
2737         [REINT_SETATTR] = {
2738                 .mr_handler = &mdt_reint_setattr,
2739                 .mr_extra_opc = MDS_REINT_SETATTR,
2740         },
2741         [REINT_CREATE] = {
2742                 .mr_handler = &mdt_reint_create,
2743                 .mr_extra_opc = MDS_REINT_CREATE,
2744         },
2745         [REINT_LINK] = {
2746                 .mr_handler = &mdt_reint_link,
2747                 .mr_extra_opc = MDS_REINT_LINK,
2748         },
2749         [REINT_UNLINK] = {
2750                 .mr_handler = &mdt_reint_unlink,
2751                 .mr_extra_opc = MDS_REINT_UNLINK,
2752         },
2753         [REINT_RENAME] = {
2754                 .mr_handler = &mdt_reint_rename,
2755                 .mr_extra_opc = MDS_REINT_RENAME,
2756         },
2757         [REINT_OPEN] = {
2758                 .mr_handler = &mdt_reint_open,
2759                 .mr_extra_opc = MDS_REINT_OPEN,
2760         },
2761         [REINT_SETXATTR] = {
2762                 .mr_handler = &mdt_reint_setxattr,
2763                 .mr_extra_opc = MDS_REINT_SETXATTR,
2764         },
2765         [REINT_RMENTRY] = {
2766                 .mr_handler = &mdt_reint_unlink,
2767                 .mr_extra_opc = MDS_REINT_UNLINK,
2768         },
2769         [REINT_MIGRATE] = {
2770                 .mr_handler = &mdt_reint_migrate,
2771                 .mr_extra_opc = MDS_REINT_RENAME,
2772         },
2773         [REINT_RESYNC] = {
2774                 .mr_handler = &mdt_reint_resync,
2775                 .mr_extra_opc = MDS_REINT_RESYNC,
2776         },
2777 };
2778
2779 int mdt_reint_rec(struct mdt_thread_info *info,
2780                   struct mdt_lock_handle *lhc)
2781 {
2782         const struct mdt_reinter *mr;
2783         int rc;
2784         ENTRY;
2785
2786         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2787                 RETURN(-EPROTO);
2788
2789         mr = &mdt_reinters[info->mti_rr.rr_opcode];
2790         if (mr->mr_handler == NULL)
2791                 RETURN(-EPROTO);
2792
2793         rc = (*mr->mr_handler)(info, lhc);
2794
2795         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2796                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
2797
2798         RETURN(rc);
2799 }