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