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  *
  31  * lustre/mdt/mdt_reint.c
  32  *
  33  * Lustre Metadata Target (mdt) reintegration routines
  34  *
  35  * Author: Peter Braam <braam@clusterfs.com>
  36  * Author: Andreas Dilger <adilger@clusterfs.com>
  37  * Author: Phil Schwan <phil@clusterfs.com>
  38  * Author: Huang Hua <huanghua@clusterfs.com>
  39  * Author: Yury Umanets <umka@clusterfs.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_MDS
  43
  44 #include <lprocfs_status.h>
  45 #include "mdt_internal.h"
  46 #include <lustre_lmv.h>
  47 #include <lustre_crypto.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
  85         ENTRY;
  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 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
 210                 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
 211                 if (likely(!IS_ERR(child))) {
 212                         mdt_obj_version_get(info, child, &info->mti_ver[idx]);
 213                         mdt_object_put(info->mti_env, child);
 214                 }
 215         }
 216         vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
 217         return vbrc ? vbrc : rc;
 218
 219 }
 220
 221 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
 222                              struct mdt_object *obj,
 223                              struct ldlm_enqueue_info *einfo,
 224                              int decref)
 225 {
 226         union ldlm_policy_data *policy = &mti->mti_policy;
 227         struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
 228         struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
 229         int i;
 230
 231         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 232         LASSERT(slave_locks);
 233
 234         memset(policy, 0, sizeof(*policy));
 235         policy->l_inodebits.bits = einfo->ei_inodebits;
 236         mdt_lock_handle_init(lh);
 237         mdt_lock_reg_init(lh, einfo->ei_mode);
 238         for (i = 0; i < slave_locks->ha_count; i++) {
 239                 if (test_bit(i, (void *)slave_locks->ha_map))
 240                         lh->mlh_rreg_lh = slave_locks->ha_handles[i];
 241                 else
 242                         lh->mlh_reg_lh = slave_locks->ha_handles[i];
 243                 mdt_object_unlock(mti, NULL, lh, decref);
 244                 slave_locks->ha_handles[i].cookie = 0ull;
 245         }
 246
 247         return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
 248                                 policy);
 249 }
 250
 251 static inline int mdt_object_striped(struct mdt_thread_info *mti,
 252                                      struct mdt_object *obj)
 253 {
 254         struct lu_device *bottom_dev;
 255         struct lu_object *bottom_obj;
 256         int rc;
 257
 258         if (!S_ISDIR(obj->mot_header.loh_attr))
 259                 return 0;
 260
 261         /* getxattr from bottom obj to avoid reading in shard FIDs */
 262         bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom);
 263         bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev,
 264                                           mdt_object_fid(obj), NULL);
 265         if (IS_ERR(bottom_obj))
 266                 return PTR_ERR(bottom_obj);
 267
 268         rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL,
 269                           XATTR_NAME_LMV);
 270         lu_object_put(mti->mti_env, bottom_obj);
 271
 272         return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc;
 273 }
 274
 275 /**
 276  * Lock slave stripes if necessary, the lock handles of slave stripes
 277  * will be stored in einfo->ei_cbdata.
 278  **/
 279 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
 280                            enum ldlm_mode mode, __u64 ibits,
 281                            struct ldlm_enqueue_info *einfo)
 282 {
 283         union ldlm_policy_data *policy = &mti->mti_policy;
 284
 285         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 286
 287         einfo->ei_type = LDLM_IBITS;
 288         einfo->ei_mode = mode;
 289         einfo->ei_cb_bl = mdt_remote_blocking_ast;
 290         einfo->ei_cb_local_bl = mdt_blocking_ast;
 291         einfo->ei_cb_cp = ldlm_completion_ast;
 292         einfo->ei_enq_slave = 1;
 293         einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
 294         einfo->ei_inodebits = ibits;
 295         memset(policy, 0, sizeof(*policy));
 296         policy->l_inodebits.bits = ibits;
 297
 298         return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
 299                               policy);
 300 }
 301
 302 int mdt_reint_striped_lock(struct mdt_thread_info *info,
 303                            struct mdt_object *o,
 304                            struct mdt_lock_handle *lh,
 305                            __u64 ibits,
 306                            struct ldlm_enqueue_info *einfo,
 307                            bool cos_incompat)
 308 {
 309         int rc;
 310
 311         LASSERT(!mdt_object_remote(o));
 312
 313         memset(einfo, 0, sizeof(*einfo));
 314
 315         rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
 316         if (rc)
 317                 return rc;
 318
 319         rc = mdt_object_striped(info, o);
 320         if (rc != 1) {
 321                 if (rc < 0)
 322                         mdt_object_unlock(info, o, lh, rc);
 323                 return rc;
 324         }
 325
 326         rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
 327         if (rc) {
 328                 mdt_object_unlock(info, o, lh, rc);
 329                 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
 330                         rc = 0;
 331         }
 332
 333         return rc;
 334 }
 335
 336 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
 337                               struct mdt_object *o,
 338                               struct mdt_lock_handle *lh,
 339                               struct ldlm_enqueue_info *einfo, int decref)
 340 {
 341         if (einfo->ei_cbdata)
 342                 mdt_unlock_slaves(info, o, einfo, decref);
 343         mdt_object_unlock(info, o, lh, decref);
 344 }
 345
 346 static int mdt_restripe(struct mdt_thread_info *info,
 347                         struct mdt_object *parent,
 348                         const struct lu_name *lname,
 349                         const struct lu_fid *tfid,
 350                         struct md_op_spec *spec,
 351                         struct md_attr *ma)
 352 {
 353         struct mdt_device *mdt = info->mti_mdt;
 354         struct lu_fid *fid = &info->mti_tmp_fid2;
 355         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 356         struct lmv_user_md *lum = spec->u.sp_ea.eadata;
 357         struct lmv_mds_md_v1 *lmv;
 358         struct mdt_object *child;
 359         struct mdt_lock_handle *lhp;
 360         struct mdt_lock_handle *lhc;
 361         struct mdt_body *repbody;
 362         int rc;
 363
 364         ENTRY;
 365         if (!mdt->mdt_enable_dir_restripe)
 366                 RETURN(-EPERM);
 367
 368         LASSERT(lum);
 369         lum->lum_hash_type |= cpu_to_le32(LMV_HASH_FLAG_FIXED);
 370
 371         rc = mdt_version_get_check_save(info, parent, 0);
 372         if (rc)
 373                 RETURN(rc);
 374
 375         lhp = &info->mti_lh[MDT_LH_PARENT];
 376         mdt_lock_pdo_init(lhp, LCK_PW, lname);
 377         rc = mdt_reint_object_lock(info, parent, lhp, MDS_INODELOCK_UPDATE,
 378                                    true);
 379         if (rc)
 380                 RETURN(rc);
 381
 382         rc = mdt_stripe_get(info, parent, ma, XATTR_NAME_LMV);
 383         if (rc)
 384                 GOTO(unlock_parent, rc);
 385
 386         if (ma->ma_valid & MA_LMV) {
 387                 /* don't allow restripe if parent dir layout is changing */
 388                 lmv = &ma->ma_lmv->lmv_md_v1;
 389                 if (!lmv_is_sane2(lmv))
 390                         GOTO(unlock_parent, rc = -EBADF);
 391
 392                 if (lmv_is_layout_changing(lmv))
 393                         GOTO(unlock_parent, rc = -EBUSY);
 394         }
 395
 396         fid_zero(fid);
 397         rc = mdt_lookup_version_check(info, parent, lname, fid, 1);
 398         if (rc)
 399                 GOTO(unlock_parent, rc);
 400
 401         child = mdt_object_find(info->mti_env, mdt, fid);
 402         if (IS_ERR(child))
 403                 GOTO(unlock_parent, rc = PTR_ERR(child));
 404
 405         if (!mdt_object_exists(child))
 406                 GOTO(out_child, rc = -ENOENT);
 407
 408         if (mdt_object_remote(child)) {
 409                 struct mdt_body *repbody;
 410
 411                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 412                 if (!repbody)
 413                         GOTO(out_child, rc = -EPROTO);
 414
 415                 repbody->mbo_fid1 = *fid;
 416                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 417                 GOTO(out_child, rc = -EREMOTE);
 418         }
 419
 420         if (!S_ISDIR(lu_object_attr(&child->mot_obj)))
 421                 GOTO(out_child, rc = -ENOTDIR);
 422
 423         rc = mdt_stripe_get(info, child, ma, XATTR_NAME_LMV);
 424         if (rc)
 425                 GOTO(out_child, rc);
 426
 427         /* race with migrate? */
 428         if ((ma->ma_valid & MA_LMV) &&
 429              lmv_is_migrating(&ma->ma_lmv->lmv_md_v1))
 430                 GOTO(out_child, rc = -EBUSY);
 431
 432         /* lock object */
 433         lhc = &info->mti_lh[MDT_LH_CHILD];
 434         mdt_lock_reg_init(lhc, LCK_EX);
 435
 436         /* enqueue object remote LOOKUP lock */
 437         if (mdt_object_remote(parent)) {
 438                 rc = mdt_remote_object_lock(info, parent, fid,
 439                                             &lhc->mlh_rreg_lh,
 440                                             lhc->mlh_rreg_mode,
 441                                             MDS_INODELOCK_LOOKUP, false);
 442                 if (rc != ELDLM_OK)
 443                         GOTO(out_child, rc);
 444         }
 445
 446         rc = mdt_reint_striped_lock(info, child, lhc, MDS_INODELOCK_FULL, einfo,
 447                                     true);
 448         if (rc)
 449                 GOTO(unlock_child, rc);
 450
 451         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 452         rc = mdt_version_get_check_save(info, child, 1);
 453         if (rc)
 454                 GOTO(unlock_child, rc);
 455
 456         spin_lock(&mdt->mdt_restriper.mdr_lock);
 457         if (child->mot_restriping) {
 458                 /* race? */
 459                 spin_unlock(&mdt->mdt_restriper.mdr_lock);
 460                 GOTO(unlock_child, rc = -EBUSY);
 461         }
 462         child->mot_restriping = 1;
 463         spin_unlock(&mdt->mdt_restriper.mdr_lock);
 464
 465         *fid = *tfid;
 466         rc = mdt_restripe_internal(info, parent, child, lname, fid, spec, ma);
 467         if (rc)
 468                 GOTO(restriping_clear, rc);
 469
 470         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 471         if (!repbody)
 472                 GOTO(restriping_clear, rc = -EPROTO);
 473
 474         mdt_pack_attr2body(info, repbody, &ma->ma_attr, fid);
 475         EXIT;
 476
 477 restriping_clear:
 478         child->mot_restriping = 0;
 479 unlock_child:
 480         mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 481 out_child:
 482         mdt_object_put(info->mti_env, child);
 483 unlock_parent:
 484         mdt_object_unlock(info, parent, lhp, rc);
 485
 486         return rc;
 487 }
 488
 489 /*
 490  * VBR: we save three versions in reply:
 491  * 0 - parent. Check that parent version is the same during replay.
 492  * 1 - name. Version of 'name' if file exists with the same name or
 493  * ENOENT_VERSION, it is needed because file may appear due to missed replays.
 494  * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
 495  * check.
 496  */
 497 static int mdt_create(struct mdt_thread_info *info)
 498 {
 499         struct mdt_device *mdt = info->mti_mdt;
 500         struct mdt_object *parent;
 501         struct mdt_object *child;
 502         struct mdt_lock_handle *lh;
 503         struct mdt_body *repbody;
 504         struct md_attr *ma = &info->mti_attr;
 505         struct mdt_reint_record *rr = &info->mti_rr;
 506         struct md_op_spec *spec = &info->mti_spec;
 507         bool restripe = false;
 508         int rc;
 509
 510         ENTRY;
 511         DEBUG_REQ(D_INODE, mdt_info_req(info),
 512                   "Create ("DNAME"->"DFID") in "DFID,
 513                   PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
 514
 515         if (!fid_is_md_operative(rr->rr_fid1))
 516                 RETURN(-EPERM);
 517
 518         if (S_ISDIR(ma->ma_attr.la_mode) &&
 519             spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
 520                 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
 521                 struct lu_ucred *uc = mdt_ucred(info);
 522                 struct obd_export *exp = mdt_info_req(info)->rq_export;
 523
 524                 /* Only new clients can create remote dir( >= 2.4) and
 525                  * striped dir(>= 2.6), old client will return -ENOTSUPP
 526                  */
 527                 if (!mdt_is_dne_client(exp))
 528                         RETURN(-ENOTSUPP);
 529
 530                 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
 531                         if (!mdt_is_striped_client(exp))
 532                                 RETURN(-ENOTSUPP);
 533
 534                         if (!mdt->mdt_enable_striped_dir)
 535                                 RETURN(-EPERM);
 536                 } else if (!mdt->mdt_enable_remote_dir) {
 537                         RETURN(-EPERM);
 538                 }
 539
 540                 if ((!(exp_connect_flags2(exp) & OBD_CONNECT2_CRUSH)) &&
 541                     (le32_to_cpu(lum->lum_hash_type) & LMV_HASH_TYPE_MASK) ==
 542                     LMV_HASH_TYPE_CRUSH)
 543                         RETURN(-EPROTO);
 544
 545                 if (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
 546                     uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 547                     mdt->mdt_enable_remote_dir_gid != -1)
 548                         RETURN(-EPERM);
 549
 550                 /* restripe if later found dir exists, MDS_OPEN_CREAT means
 551                  * this is create only, don't try restripe.
 552                  */
 553                 if (mdt->mdt_enable_dir_restripe &&
 554                     le32_to_cpu(lum->lum_stripe_offset) == LMV_OFFSET_DEFAULT &&
 555                     !(spec->sp_cr_flags & MDS_OPEN_CREAT))
 556                         restripe = true;
 557         }
 558
 559         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 560
 561         parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 562         if (IS_ERR(parent))
 563                 RETURN(PTR_ERR(parent));
 564
 565         if (!mdt_object_exists(parent))
 566                 GOTO(put_parent, rc = -ENOENT);
 567
 568         /*
 569          * LU-10235: check if name exists locklessly first to avoid massive
 570          * lock recalls on existing directories.
 571          */
 572         rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
 573                                       &info->mti_tmp_fid1, 1);
 574         if (rc == 0) {
 575                 if (!restripe)
 576                         GOTO(put_parent, rc = -EEXIST);
 577
 578                 rc = mdt_restripe(info, parent, &rr->rr_name, rr->rr_fid2, spec,
 579                                   ma);
 580         }
 581
 582         /* -ENOENT is expected here */
 583         if (rc != -ENOENT)
 584                 GOTO(put_parent, rc);
 585
 586         /* save version of file name for replay, it must be ENOENT here */
 587         mdt_enoent_version_save(info, 1);
 588
 589         OBD_RACE(OBD_FAIL_MDS_CREATE_RACE);
 590
 591         lh = &info->mti_lh[MDT_LH_PARENT];
 592         mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 593         rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
 594         if (rc)
 595                 GOTO(put_parent, rc);
 596
 597         if (!mdt_object_remote(parent)) {
 598                 rc = mdt_version_get_check_save(info, parent, 0);
 599                 if (rc)
 600                         GOTO(unlock_parent, rc);
 601         }
 602
 603         child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
 604         if (unlikely(IS_ERR(child)))
 605                 GOTO(unlock_parent, rc = PTR_ERR(child));
 606
 607         ma->ma_need = MA_INODE;
 608         ma->ma_valid = 0;
 609
 610         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 611                         OBD_FAIL_MDS_REINT_CREATE_WRITE);
 612
 613         /* Version of child will be updated on disk. */
 614         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 615         rc = mdt_version_get_check_save(info, child, 2);
 616         if (rc)
 617                 GOTO(put_child, rc);
 618
 619         /*
 620          * Do not perform lookup sanity check. We know that name does
 621          * not exist.
 622          */
 623         info->mti_spec.sp_cr_lookup = 0;
 624         info->mti_spec.sp_feat = &dt_directory_features;
 625
 626         rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
 627                         mdt_object_child(child), &info->mti_spec, ma);
 628         if (rc == 0)
 629                 rc = mdt_attr_get_complex(info, child, ma);
 630
 631         if (rc < 0)
 632                 GOTO(put_child, rc);
 633
 634         /*
 635          * On DNE, we need to eliminate dependey between 'mkdir a' and
 636          * 'mkdir a/b' if b is a striped directory, to achieve this, two
 637          * things are done below:
 638          * 1. save child and slaves lock.
 639          * 2. if the child is a striped directory, relock parent so to
 640          *    compare against with COS locks to ensure parent was
 641          *    committed to disk.
 642          */
 643         if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
 644                 struct mdt_lock_handle *lhc;
 645                 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 646                 bool cos_incompat;
 647
 648                 rc = mdt_object_striped(info, child);
 649                 if (rc < 0)
 650                         GOTO(put_child, rc);
 651
 652                 cos_incompat = rc;
 653                 if (cos_incompat) {
 654                         if (!mdt_object_remote(parent)) {
 655                                 mdt_object_unlock(info, parent, lh, 1);
 656                                 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 657                                 rc = mdt_reint_object_lock(info, parent, lh,
 658                                                            MDS_INODELOCK_UPDATE,
 659                                                            true);
 660                                 if (rc)
 661                                         GOTO(put_child, rc);
 662                         }
 663                 }
 664
 665                 lhc = &info->mti_lh[MDT_LH_CHILD];
 666                 mdt_lock_handle_init(lhc);
 667                 mdt_lock_reg_init(lhc, LCK_PW);
 668                 rc = mdt_reint_striped_lock(info, child, lhc,
 669                                             MDS_INODELOCK_UPDATE, einfo,
 670                                             cos_incompat);
 671                 if (rc)
 672                         GOTO(put_child, rc);
 673
 674                 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 675         }
 676
 677         /* Return fid & attr to client. */
 678         if (ma->ma_valid & MA_INODE)
 679                 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
 680                                    mdt_object_fid(child));
 681         EXIT;
 682 put_child:
 683         mdt_object_put(info->mti_env, child);
 684 unlock_parent:
 685         mdt_object_unlock(info, parent, lh, rc);
 686 put_parent:
 687         mdt_object_put(info->mti_env, parent);
 688         return rc;
 689 }
 690
 691 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
 692                         struct md_attr *ma)
 693 {
 694         struct mdt_lock_handle  *lh;
 695         int do_vbr = ma->ma_attr.la_valid &
 696                         (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
 697         __u64 lockpart = MDS_INODELOCK_UPDATE;
 698         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 699         bool cos_incompat;
 700         int rc;
 701
 702         ENTRY;
 703         rc = mdt_object_striped(info, mo);
 704         if (rc < 0)
 705                 RETURN(rc);
 706
 707         cos_incompat = rc;
 708
 709         lh = &info->mti_lh[MDT_LH_PARENT];
 710         mdt_lock_reg_init(lh, LCK_PW);
 711
 712         /* Even though the new MDT will grant PERM lock to the old
 713          * client, but the old client will almost ignore that during
 714          * So it needs to revoke both LOOKUP and PERM lock here, so
 715          * both new and old client can cancel the dcache
 716          */
 717         if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
 718                 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
 719
 720         rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
 721                                     cos_incompat);
 722         if (rc != 0)
 723                 RETURN(rc);
 724
 725         /* all attrs are packed into mti_attr in unpack_setattr */
 726         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 727                        OBD_FAIL_MDS_REINT_SETATTR_WRITE);
 728
 729         /* VBR: update version if attr changed are important for recovery */
 730         if (do_vbr) {
 731                 /* update on-disk version of changed object */
 732                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
 733                 rc = mdt_version_get_check_save(info, mo, 0);
 734                 if (rc)
 735                         GOTO(out_unlock, rc);
 736         }
 737
 738         /* Ensure constant striping during chown(). See LU-2789. */
 739         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 740                 mutex_lock(&mo->mot_lov_mutex);
 741
 742         /* all attrs are packed into mti_attr in unpack_setattr */
 743         rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
 744
 745         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 746                 mutex_unlock(&mo->mot_lov_mutex);
 747
 748         if (rc != 0)
 749                 GOTO(out_unlock, rc);
 750         mdt_dom_obj_lvb_update(info->mti_env, mo, NULL, false);
 751         EXIT;
 752 out_unlock:
 753         mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
 754         return rc;
 755 }
 756
 757 /**
 758  * Check HSM flags and add HS_DIRTY flag if relevant.
 759  *
 760  * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
 761  * and is not RELEASED.
 762  */
 763 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
 764                         struct md_attr *ma)
 765 {
 766         struct lu_ucred *uc = mdt_ucred(info);
 767         kernel_cap_t cap_saved;
 768         int rc;
 769
 770         ENTRY;
 771         /* If the file was modified, add the dirty flag */
 772         ma->ma_need = MA_HSM;
 773         rc = mdt_attr_get_complex(info, mo, ma);
 774         if (rc) {
 775                 CERROR("file attribute read error for "DFID": %d.\n",
 776                         PFID(mdt_object_fid(mo)), rc);
 777                 RETURN(rc);
 778         }
 779
 780         /* If an up2date copy exists in the backend, add dirty flag */
 781         if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
 782             && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
 783                 ma->ma_hsm.mh_flags |= HS_DIRTY;
 784
 785                 /* Bump cap so that closes from non-owner writers can
 786                  * set the HSM state to dirty.
 787                  */
 788                 cap_saved = uc->uc_cap;
 789                 cap_raise(uc->uc_cap, CAP_FOWNER);
 790                 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
 791                 uc->uc_cap = cap_saved;
 792                 if (rc)
 793                         CERROR("file attribute change error for "DFID": %d\n",
 794                                 PFID(mdt_object_fid(mo)), rc);
 795         }
 796
 797         RETURN(rc);
 798 }
 799
 800 static int mdt_reint_setattr(struct mdt_thread_info *info,
 801                              struct mdt_lock_handle *lhc)
 802 {
 803         struct mdt_device *mdt = info->mti_mdt;
 804         struct md_attr *ma = &info->mti_attr;
 805         struct mdt_reint_record *rr = &info->mti_rr;
 806         struct ptlrpc_request *req = mdt_info_req(info);
 807         struct mdt_object *mo;
 808         struct mdt_body *repbody;
 809         ktime_t kstart = ktime_get();
 810         int rc, rc2;
 811
 812         ENTRY;
 813         DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
 814                   (unsigned int)ma->ma_attr.la_valid);
 815
 816         if (info->mti_dlm_req)
 817                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 818
 819         OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
 820
 821         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 822         mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
 823         if (IS_ERR(mo))
 824                 GOTO(out, rc = PTR_ERR(mo));
 825
 826         if (!mdt_object_exists(mo))
 827                 GOTO(out_put, rc = -ENOENT);
 828
 829         if (mdt_object_remote(mo))
 830                 GOTO(out_put, rc = -EREMOTE);
 831
 832         ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
 833         /* revoke lease lock if size is going to be changed */
 834         if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
 835                      !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
 836                      atomic_read(&mo->mot_lease_count) > 0)) {
 837                 down_read(&mo->mot_open_sem);
 838
 839                 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
 840                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 841                         mdt_lock_reg_init(lhc, LCK_CW);
 842
 843                         rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
 844                         if (rc != 0) {
 845                                 up_read(&mo->mot_open_sem);
 846                                 GOTO(out_put, rc);
 847                         }
 848
 849                         /* revoke lease lock */
 850                         mdt_object_unlock(info, mo, lhc, 1);
 851                 }
 852                 up_read(&mo->mot_open_sem);
 853         }
 854
 855         if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
 856                 /* Check write access for the O_TRUNC case */
 857                 if (mdt_write_read(mo) < 0)
 858                         GOTO(out_put, rc = -ETXTBSY);
 859
 860                 /* LU-10286: compatibility check for FLR.
 861                  * Please check the comment in mdt_finish_open() for details
 862                  */
 863                 if (!exp_connect_flr(info->mti_exp) ||
 864                     !exp_connect_overstriping(info->mti_exp)) {
 865                         rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
 866                         if (rc < 0 && rc != -ENODATA)
 867                                 GOTO(out_put, rc);
 868
 869                         if (!exp_connect_flr(info->mti_exp)) {
 870                                 if (rc > 0 &&
 871                                     mdt_lmm_is_flr(info->mti_big_lmm))
 872                                         GOTO(out_put, rc = -EOPNOTSUPP);
 873                         }
 874
 875                         if (!exp_connect_overstriping(info->mti_exp)) {
 876                                 if (rc > 0 &&
 877                                     mdt_lmm_is_overstriping(info->mti_big_lmm))
 878                                         GOTO(out_put, rc = -EOPNOTSUPP);
 879                         }
 880                 }
 881
 882                 /* For truncate, the file size sent from client
 883                  * is believable, but the blocks are incorrect,
 884                  * which makes the block size in LSOM attribute
 885                  * inconsisent with the real block size.
 886                  */
 887                 rc = mdt_lsom_update(info, mo, true);
 888                 if (rc)
 889                         GOTO(out_put, rc);
 890         }
 891
 892         if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
 893                 if (ma->ma_valid & MA_LOV)
 894                         GOTO(out_put, rc = -EPROTO);
 895
 896                 /* MDT supports FMD for regular files due to Data-on-MDT */
 897                 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
 898                     ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
 899                         tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
 900                                        req->rq_xid);
 901
 902                         if (ma->ma_attr.la_valid & LA_MTIME) {
 903                                 rc = mdt_attr_get_pfid(info, mo, &ma->ma_pfid);
 904                                 if (!rc)
 905                                         ma->ma_valid |= MA_PFID;
 906                         }
 907                 }
 908
 909                 rc = mdt_attr_set(info, mo, ma);
 910                 if (rc)
 911                         GOTO(out_put, rc);
 912         } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
 913                    (ma->ma_valid & MA_INODE)) {
 914                 struct lu_buf *buf = &info->mti_buf;
 915                 struct lu_ucred *uc = mdt_ucred(info);
 916                 struct mdt_lock_handle *lh;
 917                 const char *name;
 918                 __u64 lockpart = MDS_INODELOCK_XATTR;
 919
 920                 /* reject if either remote or striped dir is disabled */
 921                 if (ma->ma_valid & MA_LMV) {
 922                         if (!mdt->mdt_enable_remote_dir ||
 923                             !mdt->mdt_enable_striped_dir)
 924                                 GOTO(out_put, rc = -EPERM);
 925
 926                         if (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
 927                             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 928                             mdt->mdt_enable_remote_dir_gid != -1)
 929                                 GOTO(out_put, rc = -EPERM);
 930                 }
 931
 932                 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
 933                         GOTO(out_put, rc = -ENOTDIR);
 934
 935                 if (ma->ma_attr.la_valid != 0)
 936                         GOTO(out_put, rc = -EPROTO);
 937
 938                 lh = &info->mti_lh[MDT_LH_PARENT];
 939                 mdt_lock_reg_init(lh, LCK_PW);
 940
 941                 if (ma->ma_valid & MA_LOV) {
 942                         buf->lb_buf = ma->ma_lmm;
 943                         buf->lb_len = ma->ma_lmm_size;
 944                         name = XATTR_NAME_LOV;
 945                 } else {
 946                         struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
 947                         struct lu_fid *pfid = &info->mti_tmp_fid1;
 948                         struct lu_name *pname = &info->mti_name;
 949                         const char dotdot[] = "..";
 950                         struct mdt_object *pobj;
 951
 952                         buf->lb_buf = lmu;
 953                         buf->lb_len = ma->ma_lmv_size;
 954                         name = XATTR_NAME_DEFAULT_LMV;
 955
 956                         if (fid_is_root(rr->rr_fid1)) {
 957                                 lockpart |= MDS_INODELOCK_LOOKUP;
 958                         } else {
 959                                 /* force client to update dir default layout */
 960                                 fid_zero(pfid);
 961                                 pname->ln_name = dotdot;
 962                                 pname->ln_namelen = sizeof(dotdot);
 963                                 rc = mdo_lookup(info->mti_env,
 964                                                 mdt_object_child(mo), pname,
 965                                                 pfid, NULL);
 966                                 if (rc)
 967                                         GOTO(out_put, rc);
 968
 969                                 pobj = mdt_object_find(info->mti_env, mdt,
 970                                                        pfid);
 971                                 if (IS_ERR(pobj))
 972                                         GOTO(out_put, rc = PTR_ERR(pobj));
 973
 974                                 if (mdt_object_remote(pobj))
 975                                         rc = mdt_remote_object_lock(info, pobj,
 976                                                 mdt_object_fid(mo),
 977                                                 &lh->mlh_rreg_lh, LCK_EX,
 978                                                 MDS_INODELOCK_LOOKUP, false);
 979                                 else
 980                                         lockpart |= MDS_INODELOCK_LOOKUP;
 981
 982                                 mdt_object_put(info->mti_env, pobj);
 983
 984                                 if (rc)
 985                                         GOTO(out_put, rc);
 986                         }
 987                 }
 988
 989                 rc = mdt_object_lock(info, mo, lh, lockpart);
 990                 if (rc != 0)
 991                         GOTO(out_put, rc);
 992
 993                 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
 994                                   name, 0);
 995
 996                 mdt_object_unlock(info, mo, lh, rc);
 997                 if (rc)
 998                         GOTO(out_put, rc);
 999         } else {
1000                 GOTO(out_put, rc = -EPROTO);
1001         }
1002
1003         /* If file data is modified, add the dirty flag */
1004         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
1005                 rc = mdt_add_dirty_flag(info, mo, ma);
1006
1007         ma->ma_need = MA_INODE;
1008         ma->ma_valid = 0;
1009         rc = mdt_attr_get_complex(info, mo, ma);
1010         if (rc != 0)
1011                 GOTO(out_put, rc);
1012
1013         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
1014
1015         EXIT;
1016 out_put:
1017         mdt_object_put(info->mti_env, mo);
1018 out:
1019         if (rc == 0)
1020                 mdt_counter_incr(req, LPROC_MDT_SETATTR,
1021                                  ktime_us_delta(ktime_get(), kstart));
1022
1023         mdt_client_compatibility(info);
1024         rc2 = mdt_fix_reply(info);
1025         if (rc == 0)
1026                 rc = rc2;
1027         return rc;
1028 }
1029
1030 static int mdt_reint_create(struct mdt_thread_info *info,
1031                             struct mdt_lock_handle *lhc)
1032 {
1033         struct ptlrpc_request   *req = mdt_info_req(info);
1034         ktime_t                 kstart = ktime_get();
1035         int                     rc;
1036
1037         ENTRY;
1038         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
1039                 RETURN(err_serious(-ESTALE));
1040
1041         if (info->mti_dlm_req)
1042                 ldlm_request_cancel(mdt_info_req(info),
1043                                     info->mti_dlm_req, 0, LATF_SKIP);
1044
1045         if (!lu_name_is_valid(&info->mti_rr.rr_name))
1046                 RETURN(-EPROTO);
1047
1048         switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
1049         case S_IFDIR:
1050         case S_IFREG:
1051         case S_IFLNK:
1052         case S_IFCHR:
1053         case S_IFBLK:
1054         case S_IFIFO:
1055         case S_IFSOCK:
1056                 break;
1057         default:
1058                 CERROR("%s: Unsupported mode %o\n",
1059                        mdt_obd_name(info->mti_mdt),
1060                        info->mti_attr.ma_attr.la_mode);
1061                 RETURN(err_serious(-EOPNOTSUPP));
1062         }
1063
1064         rc = mdt_create(info);
1065         if (rc == 0) {
1066                 if ((info->mti_attr.ma_attr.la_mode & S_IFMT) == S_IFDIR)
1067                         mdt_counter_incr(req, LPROC_MDT_MKDIR,
1068                                          ktime_us_delta(ktime_get(), kstart));
1069                 else
1070                         /* Special file should stay on the same node as parent*/
1071                         mdt_counter_incr(req, LPROC_MDT_MKNOD,
1072                                          ktime_us_delta(ktime_get(), kstart));
1073         }
1074
1075         RETURN(rc);
1076 }
1077
1078 /*
1079  * VBR: save parent version in reply and child version getting by its name.
1080  * Version of child is getting and checking during its lookup. If
1081  */
1082 static int mdt_reint_unlink(struct mdt_thread_info *info,
1083                             struct mdt_lock_handle *lhc)
1084 {
1085         struct mdt_reint_record *rr = &info->mti_rr;
1086         struct ptlrpc_request *req = mdt_info_req(info);
1087         struct md_attr *ma = &info->mti_attr;
1088         struct lu_fid *child_fid = &info->mti_tmp_fid1;
1089         struct mdt_object *mp;
1090         struct mdt_object *mc;
1091         struct mdt_lock_handle *parent_lh;
1092         struct mdt_lock_handle *child_lh;
1093         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
1094         __u64 lock_ibits;
1095         bool cos_incompat = false;
1096         int no_name = 0;
1097         ktime_t kstart = ktime_get();
1098         int rc;
1099
1100         ENTRY;
1101         DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
1102                   PNAME(&rr->rr_name));
1103
1104         if (info->mti_dlm_req)
1105                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1106
1107         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
1108                 RETURN(err_serious(-ENOENT));
1109
1110         if (!fid_is_md_operative(rr->rr_fid1))
1111                 RETURN(-EPERM);
1112
1113         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1114         if (IS_ERR(mp))
1115                 RETURN(PTR_ERR(mp));
1116
1117         if (mdt_object_remote(mp)) {
1118                 cos_incompat = true;
1119         } else {
1120                 rc = mdt_version_get_check_save(info, mp, 0);
1121                 if (rc)
1122                         GOTO(put_parent, rc);
1123         }
1124
1125         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
1126         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
1127 relock:
1128         parent_lh = &info->mti_lh[MDT_LH_PARENT];
1129         mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
1130         rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
1131                                    cos_incompat);
1132         if (rc != 0)
1133                 GOTO(put_parent, rc);
1134
1135         if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID) {
1136                 *child_fid = *rr->rr_fid2;
1137         } else {
1138                 /* lookup child object along with version checking */
1139                 fid_zero(child_fid);
1140                 rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid,
1141                                               1);
1142                 if (rc != 0) {
1143                         /* Name might not be able to find during resend of
1144                          * remote unlink, considering following case.
1145                          * dir_A is a remote directory, the name entry of
1146                          * dir_A is on MDT0, the directory is on MDT1,
1147                          *
1148                          * 1. client sends unlink req to MDT1.
1149                          * 2. MDT1 sends name delete update to MDT0.
1150                          * 3. name entry is being deleted in MDT0 synchronously.
1151                          * 4. MDT1 is restarted.
1152                          * 5. client resends unlink req to MDT1. So it can not
1153                          *    find the name entry on MDT0 anymore.
1154                          * In this case, MDT1 only needs to destory the local
1155                          * directory.
1156                          */
1157                         if (mdt_object_remote(mp) && rc == -ENOENT &&
1158                             !fid_is_zero(rr->rr_fid2) &&
1159                             lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
1160                                 no_name = 1;
1161                                 *child_fid = *rr->rr_fid2;
1162                         } else {
1163                                 GOTO(unlock_parent, rc);
1164                         }
1165                 }
1166         }
1167
1168         if (!fid_is_md_operative(child_fid))
1169                 GOTO(unlock_parent, rc = -EPERM);
1170
1171         /* We will lock the child regardless it is local or remote. No harm. */
1172         mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
1173         if (IS_ERR(mc))
1174                 GOTO(unlock_parent, rc = PTR_ERR(mc));
1175
1176         if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID) {
1177                 /* In this case, child fid is embedded in the request, and we do
1178                  * not have a proper name as rr_name contains an encoded
1179                  * hash. So find name that matches provided hash.
1180                  */
1181                 if (!find_name_matching_hash(info, &rr->rr_name,
1182                                              NULL, mc, false))
1183                         GOTO(put_child, rc = -ENOENT);
1184         }
1185
1186         if (!cos_incompat) {
1187                 rc = mdt_object_striped(info, mc);
1188                 if (rc < 0)
1189                         GOTO(put_child, rc);
1190
1191                 cos_incompat = rc;
1192                 if (cos_incompat) {
1193                         mdt_object_put(info->mti_env, mc);
1194                         mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
1195                         goto relock;
1196                 }
1197         }
1198
1199         child_lh = &info->mti_lh[MDT_LH_CHILD];
1200         mdt_lock_reg_init(child_lh, LCK_EX);
1201         if (info->mti_spec.sp_rm_entry) {
1202                 struct lu_ucred *uc  = mdt_ucred(info);
1203
1204                 if (!mdt_is_dne_client(req->rq_export))
1205                         /* Return -ENOTSUPP for old client */
1206                         GOTO(put_child, rc = -ENOTSUPP);
1207
1208                 if (!cap_raised(uc->uc_cap, CAP_SYS_ADMIN))
1209                         GOTO(put_child, rc = -EPERM);
1210
1211                 ma->ma_need = MA_INODE;
1212                 ma->ma_valid = 0;
1213                 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1214                                 NULL, &rr->rr_name, ma, no_name);
1215                 GOTO(put_child, rc);
1216         }
1217
1218         if (mdt_object_remote(mc)) {
1219                 struct mdt_body  *repbody;
1220
1221                 if (!fid_is_zero(rr->rr_fid2)) {
1222                         CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
1223                                mdt_obd_name(info->mti_mdt),
1224                                PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1225                         GOTO(put_child, rc = -ENOENT);
1226                 }
1227                 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1228                        mdt_obd_name(info->mti_mdt),
1229                        PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1230
1231                 if (!mdt_is_dne_client(req->rq_export))
1232                         /* Return -ENOTSUPP for old client */
1233                         GOTO(put_child, rc = -ENOTSUPP);
1234
1235                 /* Revoke the LOOKUP lock of the remote object granted by
1236                  * this MDT. Since the unlink will happen on another MDT,
1237                  * it will release the LOOKUP lock right away. Then What
1238                  * would happen if another client try to grab the LOOKUP
1239                  * lock at the same time with unlink XXX
1240                  */
1241                 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1242                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1243                 LASSERT(repbody != NULL);
1244                 repbody->mbo_fid1 = *mdt_object_fid(mc);
1245                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1246                 GOTO(unlock_child, rc = -EREMOTE);
1247         }
1248         /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1249          * this now because a running HSM restore on the child (unlink
1250          * victim) will hold the layout lock. See LU-4002.
1251          */
1252         lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1253         if (mdt_object_remote(mp)) {
1254                 /* Enqueue lookup lock from parent MDT */
1255                 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1256                                             &child_lh->mlh_rreg_lh,
1257                                             child_lh->mlh_rreg_mode,
1258                                             MDS_INODELOCK_LOOKUP, false);
1259                 if (rc != ELDLM_OK)
1260                         GOTO(put_child, rc);
1261
1262                 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1263         }
1264
1265         rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1266                                     cos_incompat);
1267         if (rc != 0)
1268                 GOTO(put_child, rc);
1269
1270         /*
1271          * Now we can only make sure we need MA_INODE, in mdd layer, will check
1272          * whether need MA_LOV and MA_COOKIE.
1273          */
1274         ma->ma_need = MA_INODE;
1275         ma->ma_valid = 0;
1276
1277         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1278                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1279         /* save version when object is locked */
1280         mdt_version_get_save(info, mc, 1);
1281
1282         mutex_lock(&mc->mot_lov_mutex);
1283
1284         rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1285                         mdt_object_child(mc), &rr->rr_name, ma, no_name);
1286
1287         mutex_unlock(&mc->mot_lov_mutex);
1288         if (rc != 0)
1289                 GOTO(unlock_child, rc);
1290
1291         if (!lu_object_is_dying(&mc->mot_header)) {
1292                 rc = mdt_attr_get_complex(info, mc, ma);
1293                 if (rc)
1294                         GOTO(out_stat, rc);
1295         } else if (mdt_dom_check_for_discard(info, mc)) {
1296                 mdt_dom_discard_data(info, mc);
1297         }
1298         mdt_handle_last_unlink(info, mc, ma);
1299
1300 out_stat:
1301         if (ma->ma_valid & MA_INODE) {
1302                 switch (ma->ma_attr.la_mode & S_IFMT) {
1303                 case S_IFDIR:
1304                         mdt_counter_incr(req, LPROC_MDT_RMDIR,
1305                                          ktime_us_delta(ktime_get(), kstart));
1306                         break;
1307                 case S_IFREG:
1308                 case S_IFLNK:
1309                 case S_IFCHR:
1310                 case S_IFBLK:
1311                 case S_IFIFO:
1312                 case S_IFSOCK:
1313                         mdt_counter_incr(req, LPROC_MDT_UNLINK,
1314                                          ktime_us_delta(ktime_get(), kstart));
1315                         break;
1316                 default:
1317                         LASSERTF(0, "bad file type %o unlinking\n",
1318                                 ma->ma_attr.la_mode);
1319                 }
1320         }
1321
1322         EXIT;
1323
1324 unlock_child:
1325         mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1326 put_child:
1327         if (info->mti_spec.sp_cr_flags & MDS_OP_WITH_FID &&
1328             info->mti_big_buf.lb_buf)
1329                 lu_buf_free(&info->mti_big_buf);
1330         mdt_object_put(info->mti_env, mc);
1331 unlock_parent:
1332         mdt_object_unlock(info, mp, parent_lh, rc);
1333 put_parent:
1334         mdt_object_put(info->mti_env, mp);
1335         CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1336         return rc;
1337 }
1338
1339 /*
1340  * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1341  * name.
1342  */
1343 static int mdt_reint_link(struct mdt_thread_info *info,
1344                           struct mdt_lock_handle *lhc)
1345 {
1346         struct mdt_reint_record *rr = &info->mti_rr;
1347         struct ptlrpc_request   *req = mdt_info_req(info);
1348         struct md_attr          *ma = &info->mti_attr;
1349         struct mdt_object       *ms;
1350         struct mdt_object       *mp;
1351         struct mdt_lock_handle  *lhs;
1352         struct mdt_lock_handle  *lhp;
1353         ktime_t kstart = ktime_get();
1354         bool cos_incompat;
1355         int rc;
1356
1357         ENTRY;
1358         DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1359                   PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1360
1361         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1362                 RETURN(err_serious(-ENOENT));
1363
1364         if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE) ||
1365             OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_ENQ_RESEND)) {
1366                 req->rq_no_reply = 1;
1367                 RETURN(err_serious(-ENOENT));
1368         }
1369
1370         if (info->mti_dlm_req)
1371                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1372
1373         /* Invalid case so return error immediately instead of
1374          * processing it
1375          */
1376         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1377                 RETURN(-EPERM);
1378
1379         if (!fid_is_md_operative(rr->rr_fid1) ||
1380             !fid_is_md_operative(rr->rr_fid2))
1381                 RETURN(-EPERM);
1382
1383         /* step 1: find target parent dir */
1384         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1385         if (IS_ERR(mp))
1386                 RETURN(PTR_ERR(mp));
1387
1388         rc = mdt_version_get_check_save(info, mp, 0);
1389         if (rc)
1390                 GOTO(put_parent, rc);
1391
1392         /* step 2: find source */
1393         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1394         if (IS_ERR(ms))
1395                 GOTO(put_parent, rc = PTR_ERR(ms));
1396
1397         if (!mdt_object_exists(ms)) {
1398                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1399                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1400                 GOTO(put_source, rc = -ENOENT);
1401         }
1402
1403         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1404
1405         OBD_RACE(OBD_FAIL_MDS_LINK_RENAME_RACE);
1406
1407         lhp = &info->mti_lh[MDT_LH_PARENT];
1408         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1409         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1410                                    cos_incompat);
1411         if (rc != 0)
1412                 GOTO(put_source, rc);
1413
1414         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1415
1416         lhs = &info->mti_lh[MDT_LH_CHILD];
1417         mdt_lock_reg_init(lhs, LCK_EX);
1418         rc = mdt_reint_object_lock(info, ms, lhs,
1419                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1420                                    cos_incompat);
1421         if (rc != 0)
1422                 GOTO(unlock_parent, rc);
1423
1424         /* step 3: link it */
1425         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1426                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1427
1428         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1429         rc = mdt_version_get_check_save(info, ms, 1);
1430         if (rc)
1431                 GOTO(unlock_source, rc);
1432
1433         /** check target version by name during replay */
1434         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1435                                       &info->mti_tmp_fid1, 2);
1436         if (rc != 0 && rc != -ENOENT)
1437                 GOTO(unlock_source, rc);
1438         /* save version of file name for replay, it must be ENOENT here */
1439         if (!req_is_replay(mdt_info_req(info))) {
1440                 if (rc != -ENOENT) {
1441                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1442                                PNAME(&rr->rr_name));
1443                         GOTO(unlock_source, rc = -EEXIST);
1444                 }
1445                 info->mti_ver[2] = ENOENT_VERSION;
1446                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1447         }
1448
1449         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1450                       mdt_object_child(ms), &rr->rr_name, ma);
1451
1452         if (rc == 0)
1453                 mdt_counter_incr(req, LPROC_MDT_LINK,
1454                                  ktime_us_delta(ktime_get(), kstart));
1455
1456         EXIT;
1457 unlock_source:
1458         mdt_object_unlock(info, ms, lhs, rc);
1459 unlock_parent:
1460         mdt_object_unlock(info, mp, lhp, rc);
1461 put_source:
1462         mdt_object_put(info->mti_env, ms);
1463 put_parent:
1464         mdt_object_put(info->mti_env, mp);
1465         return rc;
1466 }
1467 /**
1468  * lock the part of the directory according to the hash of the name
1469  * (lh->mlh_pdo_hash) in parallel directory lock.
1470  */
1471 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1472                               struct mdt_lock_handle *lh,
1473                               struct mdt_object *obj, __u64 ibits,
1474                               bool cos_incompat)
1475 {
1476         struct ldlm_res_id *res = &info->mti_res_id;
1477         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1478         union ldlm_policy_data *policy = &info->mti_policy;
1479         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1480         int rc;
1481
1482         /*
1483          * Finish res_id initializing by name hash marking part of
1484          * directory which is taking modification.
1485          */
1486         LASSERT(lh->mlh_pdo_hash != 0);
1487         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1488         memset(policy, 0, sizeof(*policy));
1489         policy->l_inodebits.bits = ibits;
1490         if (cos_incompat &&
1491             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1492                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1493         /*
1494          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1495          * going to be sent to client. If it is - mdt_intent_policy() path will
1496          * fix it up and turn FL_LOCAL flag off.
1497          */
1498         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1499                           policy, res, dlmflags,
1500                           &info->mti_exp->exp_handle.h_cookie);
1501         return rc;
1502 }
1503
1504 /**
1505  * Get BFL lock for rename or migrate process.
1506  **/
1507 static int mdt_rename_lock(struct mdt_thread_info *info,
1508                            struct lustre_handle *lh)
1509 {
1510         int     rc;
1511
1512         ENTRY;
1513         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1514                 struct lu_fid *fid = &info->mti_tmp_fid1;
1515                 struct mdt_object *obj;
1516
1517                 /* XXX, right now, it has to use object API to
1518                  * enqueue lock cross MDT, so it will enqueue
1519                  * rename lock(with LUSTRE_BFL_FID) by root object
1520                  */
1521                 lu_root_fid(fid);
1522                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1523                 if (IS_ERR(obj))
1524                         RETURN(PTR_ERR(obj));
1525
1526                 rc = mdt_remote_object_lock(info, obj,
1527                                             &LUSTRE_BFL_FID, lh,
1528                                             LCK_EX,
1529                                             MDS_INODELOCK_UPDATE, false);
1530                 mdt_object_put(info->mti_env, obj);
1531         } else {
1532                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1533                 union ldlm_policy_data *policy = &info->mti_policy;
1534                 struct ldlm_res_id *res_id = &info->mti_res_id;
1535                 __u64 flags = 0;
1536
1537                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1538                 memset(policy, 0, sizeof(*policy));
1539                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1540                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1541                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1542                                             LDLM_IBITS, policy, LCK_EX, &flags,
1543                                             ldlm_blocking_ast,
1544                                             ldlm_completion_ast, NULL, NULL, 0,
1545                                             LVB_T_NONE,
1546                                             &info->mti_exp->exp_handle.h_cookie,
1547                                             lh);
1548                 RETURN(rc);
1549         }
1550         RETURN(rc);
1551 }
1552
1553 static void mdt_rename_unlock(struct lustre_handle *lh)
1554 {
1555         ENTRY;
1556         LASSERT(lustre_handle_is_used(lh));
1557         /* Cancel the single rename lock right away */
1558         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1559         EXIT;
1560 }
1561
1562 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1563                                                 const struct lu_fid *fid,
1564                                                 int idx)
1565 {
1566         struct mdt_object *dir;
1567         int rc;
1568
1569         ENTRY;
1570         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1571         if (IS_ERR(dir))
1572                 RETURN(dir);
1573
1574         /* check early, the real version will be saved after locking */
1575         rc = mdt_version_get_check(info, dir, idx);
1576         if (rc)
1577                 GOTO(out_put, rc);
1578
1579         if (!mdt_object_exists(dir))
1580                 GOTO(out_put, rc = -ENOENT);
1581
1582         if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1583                 GOTO(out_put, rc = -ENOTDIR);
1584
1585         RETURN(dir);
1586 out_put:
1587         mdt_object_put(info->mti_env, dir);
1588         return ERR_PTR(rc);
1589 }
1590
1591 /*
1592  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1593  * herein we don't really check it, just do revoke.
1594  */
1595 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1596                                   struct mdt_object *pobj,
1597                                   struct mdt_object *obj)
1598 {
1599         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1600         int rc;
1601
1602         mdt_lock_handle_init(lh);
1603         mdt_lock_reg_init(lh, LCK_EX);
1604
1605         if (mdt_object_remote(pobj)) {
1606                 /* don't bother to check if pobj and obj are on the same MDT. */
1607                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1608                                             &lh->mlh_rreg_lh, LCK_EX,
1609                                             MDS_INODELOCK_LOOKUP, false);
1610         } else if (mdt_object_remote(obj)) {
1611                 struct ldlm_res_id *res = &info->mti_res_id;
1612                 union ldlm_policy_data *policy = &info->mti_policy;
1613                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1614                                  LDLM_FL_COS_INCOMPAT;
1615
1616                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1617                 memset(policy, 0, sizeof(*policy));
1618                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1619                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1620                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1621                                   dlmflags, NULL);
1622         } else {
1623                 /* do nothing if both are local */
1624                 return 0;
1625         }
1626
1627         if (rc != ELDLM_OK)
1628                 return rc;
1629
1630         /*
1631          * TODO, currently we don't save this lock because there is no place to
1632          * hold this lock handle, but to avoid race we need to save this lock.
1633          */
1634         mdt_object_unlock(info, NULL, lh, 1);
1635
1636         return 0;
1637 }
1638
1639 /*
1640  * operation may takes locks of linkea, or directory stripes, group them in
1641  * different list.
1642  */
1643 struct mdt_sub_lock {
1644         struct mdt_object *msl_obj;
1645         struct mdt_lock_handle msl_lh;
1646         struct list_head msl_linkage;
1647 };
1648
1649 static void mdt_unlock_list(struct mdt_thread_info *info,
1650                             struct list_head *list, int decref)
1651 {
1652         struct mdt_sub_lock *msl;
1653         struct mdt_sub_lock *tmp;
1654
1655         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1656                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1657                 list_del(&msl->msl_linkage);
1658                 OBD_FREE_PTR(msl);
1659         }
1660 }
1661
1662 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1663                                              struct mdt_object *obj,
1664                                              struct mdt_lock_handle *lh,
1665                                              struct ldlm_enqueue_info *einfo,
1666                                              struct list_head *slave_locks,
1667                                              int decref)
1668 {
1669         if (mdt_object_remote(obj)) {
1670                 mdt_unlock_list(info, slave_locks, decref);
1671                 mdt_object_unlock(info, obj, lh, decref);
1672         } else {
1673                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1674         }
1675 }
1676
1677 /*
1678  * lock parents of links, and also check whether total locks don't exceed
1679  * RS_MAX_LOCKS.
1680  *
1681  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1682  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1683  * \retval      -ev negative errno upon error
1684  */
1685 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1686                                  struct mdt_object *pobj,
1687                                  const struct md_attr *ma,
1688                                  struct mdt_object *obj,
1689                                  struct mdt_lock_handle *lhp,
1690                                  struct ldlm_enqueue_info *peinfo,
1691                                  struct list_head *parent_slave_locks,
1692                                  struct list_head *link_locks)
1693 {
1694         struct mdt_device *mdt = info->mti_mdt;
1695         struct lu_buf *buf = &info->mti_big_buf;
1696         struct lu_name *lname = &info->mti_name;
1697         struct linkea_data ldata = { NULL };
1698         bool blocked = false;
1699         int local_lnkp_cnt = 0;
1700         int rc;
1701
1702         ENTRY;
1703         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1704                 RETURN(0);
1705
1706         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1707         if (buf->lb_buf == NULL)
1708                 RETURN(-ENOMEM);
1709
1710         ldata.ld_buf = buf;
1711         rc = mdt_links_read(info, obj, &ldata);
1712         if (rc) {
1713                 if (rc == -ENOENT || rc == -ENODATA)
1714                         rc = 0;
1715                 RETURN(rc);
1716         }
1717
1718         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1719              linkea_next_entry(&ldata)) {
1720                 struct mdt_object *lnkp;
1721                 struct mdt_sub_lock *msl;
1722                 struct lu_fid fid;
1723                 __u64 ibits;
1724
1725                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1726                                     &fid);
1727
1728                 /* check if it's also linked to parent */
1729                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1730                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1731                                PFID(&fid), PNAME(lname));
1732                         /* in case link is remote object, revoke LOOKUP lock */
1733                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1734                         continue;
1735                 }
1736
1737                 lnkp = NULL;
1738
1739                 /* check if it's linked to a stripe of parent */
1740                 if (ma->ma_valid & MA_LMV) {
1741                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1742                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1743                         int j = 0;
1744
1745                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1746                                 fid_le_to_cpu(stripe_fid,
1747                                               &lmv->lmv_stripe_fids[j]);
1748                                 if (lu_fid_eq(stripe_fid, &fid)) {
1749                                         CDEBUG(D_INFO, "skip stripe "DFID
1750                                                ", reovke "DNAME"\n",
1751                                                PFID(&fid), PNAME(lname));
1752                                         lnkp = mdt_object_find(info->mti_env,
1753                                                                mdt, &fid);
1754                                         if (IS_ERR(lnkp))
1755                                                 GOTO(out, rc = PTR_ERR(lnkp));
1756                                         break;
1757                                 }
1758                         }
1759
1760                         if (lnkp) {
1761                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1762                                                                    obj);
1763                                 mdt_object_put(info->mti_env, lnkp);
1764                                 continue;
1765                         }
1766                 }
1767
1768                 /* Check if it's already locked */
1769                 list_for_each_entry(msl, link_locks, msl_linkage) {
1770                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1771                                 CDEBUG(D_INFO,
1772                                        DFID" was locked, revoke "DNAME"\n",
1773                                        PFID(&fid), PNAME(lname));
1774                                 lnkp = msl->msl_obj;
1775                                 break;
1776                         }
1777                 }
1778
1779                 if (lnkp) {
1780                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1781                         continue;
1782                 }
1783
1784                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1785                        PFID(&fid), PNAME(lname));
1786
1787                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1788                 if (IS_ERR(lnkp)) {
1789                         CWARN("%s: cannot find obj "DFID": %ld\n",
1790                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1791                         continue;
1792                 }
1793
1794                 if (!mdt_object_exists(lnkp)) {
1795                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1796                               PFID(&fid), PNAME(lname));
1797                         mdt_object_put(info->mti_env, lnkp);
1798                         continue;
1799                 }
1800
1801                 if (!mdt_object_remote(lnkp))
1802                         local_lnkp_cnt++;
1803
1804                 OBD_ALLOC_PTR(msl);
1805                 if (msl == NULL)
1806                         GOTO(out, rc = -ENOMEM);
1807
1808                 /*
1809                  * we can't follow parent-child lock order like other MD
1810                  * operations, use lock_try here to avoid deadlock, if the lock
1811                  * cannot be taken, drop all locks taken, revoke the blocked
1812                  * one, and continue processing the remaining entries, and in
1813                  * the end of the loop restart from beginning.
1814                  */
1815                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1816                 ibits = 0;
1817                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1818                                          MDS_INODELOCK_UPDATE, true);
1819                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1820
1821                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1822                                PFID(&fid), PNAME(lname));
1823
1824                         mdt_unlock_list(info, link_locks, 1);
1825                         /* also unlock parent locks to avoid deadlock */
1826                         if (!blocked)
1827                                 mdt_migrate_object_unlock(info, pobj, lhp,
1828                                                           peinfo,
1829                                                           parent_slave_locks,
1830                                                           1);
1831
1832                         blocked = true;
1833
1834                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1835                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1836                                              MDS_INODELOCK_UPDATE);
1837                         if (rc) {
1838                                 mdt_object_put(info->mti_env, lnkp);
1839                                 OBD_FREE_PTR(msl);
1840                                 GOTO(out, rc);
1841                         }
1842
1843                         if (mdt_object_remote(lnkp)) {
1844                                 struct ldlm_lock *lock;
1845
1846                                 /*
1847                                  * for remote object, set lock cb_atomic,
1848                                  * so lock can be released in blocking_ast()
1849                                  * immediately, then the next lock_try will
1850                                  * have better chance of success.
1851                                  */
1852                                 lock = ldlm_handle2lock(
1853                                                 &msl->msl_lh.mlh_rreg_lh);
1854                                 LASSERT(lock != NULL);
1855                                 lock_res_and_lock(lock);
1856                                 ldlm_set_atomic_cb(lock);
1857                                 unlock_res_and_lock(lock);
1858                                 LDLM_LOCK_PUT(lock);
1859                         }
1860
1861                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1862                         OBD_FREE_PTR(msl);
1863                         continue;
1864                 }
1865
1866                 INIT_LIST_HEAD(&msl->msl_linkage);
1867                 msl->msl_obj = lnkp;
1868                 list_add_tail(&msl->msl_linkage, link_locks);
1869
1870                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1871         }
1872
1873         if (blocked)
1874                 GOTO(out, rc = -EBUSY);
1875
1876         EXIT;
1877 out:
1878         if (rc) {
1879                 mdt_unlock_list(info, link_locks, rc);
1880         } else if (local_lnkp_cnt > RS_MAX_LOCKS - 5) {
1881                 CDEBUG(D_INFO, "Too many links (%d), sync operations\n",
1882                        local_lnkp_cnt);
1883                 /*
1884                  * parent may have 3 local objects: master object and 2 stripes
1885                  * (if it's being migrated too); source may have 1 local objects
1886                  * as regular file; target has 1 local object.
1887                  * Note, source may have 2 local locks if it is directory but it
1888                  * can't have hardlinks, so it is not considered here.
1889                  */
1890                 rc = 1;
1891         }
1892         return rc;
1893 }
1894
1895 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1896                                   struct mdt_object *obj,
1897                                   const struct md_attr *ma,
1898                                   struct list_head *slave_locks)
1899 {
1900         struct mdt_device *mdt = info->mti_mdt;
1901         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1902         struct lu_fid *fid = &info->mti_tmp_fid1;
1903         struct mdt_object *slave;
1904         struct mdt_sub_lock *msl;
1905         int i;
1906         int rc;
1907
1908         ENTRY;
1909         LASSERT(mdt_object_remote(obj));
1910         LASSERT(ma->ma_valid & MA_LMV);
1911         LASSERT(lmv);
1912
1913         if (!lmv_is_sane(lmv))
1914                 RETURN(-EINVAL);
1915
1916         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1917                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1918
1919                 if (!fid_is_sane(fid))
1920                         continue;
1921
1922                 slave = mdt_object_find(info->mti_env, mdt, fid);
1923                 if (IS_ERR(slave))
1924                         GOTO(out, rc = PTR_ERR(slave));
1925
1926                 OBD_ALLOC_PTR(msl);
1927                 if (!msl) {
1928                         mdt_object_put(info->mti_env, slave);
1929                         GOTO(out, rc = -ENOMEM);
1930                 }
1931
1932                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1933                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1934                                            MDS_INODELOCK_UPDATE, true);
1935                 if (rc) {
1936                         OBD_FREE_PTR(msl);
1937                         mdt_object_put(info->mti_env, slave);
1938                         GOTO(out, rc);
1939                 }
1940
1941                 INIT_LIST_HEAD(&msl->msl_linkage);
1942                 msl->msl_obj = slave;
1943                 list_add_tail(&msl->msl_linkage, slave_locks);
1944         }
1945         EXIT;
1946
1947 out:
1948         if (rc)
1949                 mdt_unlock_list(info, slave_locks, rc);
1950         return rc;
1951 }
1952
1953 /* lock parent and its stripes */
1954 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1955                                    struct mdt_object *obj,
1956                                    const struct md_attr *ma,
1957                                    struct mdt_lock_handle *lh,
1958                                    struct ldlm_enqueue_info *einfo,
1959                                    struct list_head *slave_locks)
1960 {
1961         int rc;
1962
1963         if (mdt_object_remote(obj)) {
1964                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1965                                             &lh->mlh_rreg_lh, LCK_PW,
1966                                             MDS_INODELOCK_UPDATE, false);
1967                 if (rc != ELDLM_OK)
1968                         return rc;
1969
1970                 /*
1971                  * if obj is remote and striped, lock its stripes explicitly
1972                  * because it's not striped in LOD layer on this MDT.
1973                  */
1974                 if (ma->ma_valid & MA_LMV) {
1975                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1976                         if (rc)
1977                                 mdt_object_unlock(info, obj, lh, rc);
1978                 }
1979         } else {
1980                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1981                                             einfo, true);
1982         }
1983
1984         return rc;
1985 }
1986
1987 /*
1988  * in migration, object may be remote, and we need take full lock of it and its
1989  * stripes if it's directory, besides, object may be a remote object on its
1990  * parent, revoke its LOOKUP lock on where its parent is located.
1991  */
1992 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1993                                    struct mdt_object *pobj,
1994                                    struct mdt_object *obj,
1995                                    struct mdt_lock_handle *lh,
1996                                    struct ldlm_enqueue_info *einfo,
1997                                    struct list_head *slave_locks)
1998 {
1999         int rc;
2000
2001         if (mdt_object_remote(obj)) {
2002                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
2003                 if (rc)
2004                         return rc;
2005
2006                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
2007                                             &lh->mlh_rreg_lh, LCK_EX,
2008                                             MDS_INODELOCK_FULL, false);
2009                 if (rc != ELDLM_OK)
2010                         return rc;
2011
2012                 /*
2013                  * if obj is remote and striped, lock its stripes explicitly
2014                  * because it's not striped in LOD layer on this MDT.
2015                  */
2016                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
2017                         struct md_attr *ma = &info->mti_attr;
2018
2019                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
2020                         if (rc) {
2021                                 mdt_object_unlock(info, obj, lh, rc);
2022                                 return rc;
2023                         }
2024
2025                         if (ma->ma_valid & MA_LMV) {
2026                                 rc = mdt_lock_remote_slaves(info, obj, ma,
2027                                                             slave_locks);
2028                                 if (rc)
2029                                         mdt_object_unlock(info, obj, lh, rc);
2030                         }
2031                 }
2032         } else {
2033                 if (mdt_object_remote(pobj)) {
2034                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
2035                         if (rc)
2036                                 return rc;
2037                 }
2038
2039                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
2040                                             einfo, true);
2041         }
2042
2043         return rc;
2044 }
2045
2046 /*
2047  * lookup source by name, if parent is striped directory, we need to find the
2048  * corresponding stripe where source is located, and then lookup there.
2049  *
2050  * besides, if parent is migrating too, and file is already in target stripe,
2051  * this should be a redo of 'lfs migrate' on client side.
2052  */
2053 static int mdt_migrate_lookup(struct mdt_thread_info *info,
2054                               struct mdt_object *pobj,
2055                               const struct md_attr *ma,
2056                               const struct lu_name *lname,
2057                               struct mdt_object **spobj,
2058                               struct mdt_object **sobj)
2059 {
2060         const struct lu_env *env = info->mti_env;
2061         struct lu_fid *fid = &info->mti_tmp_fid1;
2062         struct mdt_object *stripe;
2063         int rc;
2064
2065         if (ma->ma_valid & MA_LMV) {
2066                 /* if parent is striped, lookup on corresponding stripe */
2067                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
2068
2069                 if (!lmv_is_sane(lmv))
2070                         return -EBADF;
2071
2072                 rc = lmv_name_to_stripe_index_old(lmv, lname->ln_name,
2073                                                   lname->ln_namelen);
2074                 if (rc < 0)
2075                         return rc;
2076
2077                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2078
2079                 stripe = mdt_object_find(env, info->mti_mdt, fid);
2080                 if (IS_ERR(stripe))
2081                         return PTR_ERR(stripe);
2082
2083                 fid_zero(fid);
2084                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
2085                                 &info->mti_spec);
2086                 if (rc == -ENOENT && lmv_is_layout_changing(lmv)) {
2087                         /*
2088                          * if parent layout is changeing, and lookup child
2089                          * failed on source stripe, lookup again on target
2090                          * stripe, if it exists, it means previous migration
2091                          * was interrupted, and current file was migrated
2092                          * already.
2093                          */
2094                         mdt_object_put(env, stripe);
2095
2096                         rc = lmv_name_to_stripe_index(lmv, lname->ln_name,
2097                                                       lname->ln_namelen);
2098                         if (rc < 0)
2099                                 return rc;
2100
2101                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
2102
2103                         stripe = mdt_object_find(env, info->mti_mdt, fid);
2104                         if (IS_ERR(stripe))
2105                                 return PTR_ERR(stripe);
2106
2107                         fid_zero(fid);
2108                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
2109                                         fid, &info->mti_spec);
2110                         mdt_object_put(env, stripe);
2111                         return rc ?: -EALREADY;
2112                 } else if (rc) {
2113                         mdt_object_put(env, stripe);
2114                         return rc;
2115                 }
2116         } else {
2117                 fid_zero(fid);
2118                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
2119                                 &info->mti_spec);
2120                 if (rc)
2121                         return rc;
2122
2123                 stripe = pobj;
2124                 mdt_object_get(env, stripe);
2125         }
2126
2127         *spobj = stripe;
2128
2129         *sobj = mdt_object_find(env, info->mti_mdt, fid);
2130         if (IS_ERR(*sobj)) {
2131                 mdt_object_put(env, stripe);
2132                 rc = PTR_ERR(*sobj);
2133                 *spobj = NULL;
2134                 *sobj = NULL;
2135         }
2136
2137         return rc;
2138 }
2139
2140 /* end lease and close file for regular file */
2141 static int mdd_migrate_close(struct mdt_thread_info *info,
2142                              struct mdt_object *obj)
2143 {
2144         struct close_data *data;
2145         struct mdt_body *repbody;
2146         struct ldlm_lock *lease;
2147         int rc;
2148         int rc2;
2149
2150         rc = -EPROTO;
2151         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
2152                                       RCL_CLIENT) ||
2153             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
2154                                       RCL_CLIENT))
2155                 goto close;
2156
2157         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
2158         if (!data)
2159                 goto close;
2160
2161         rc = -ESTALE;
2162         lease = ldlm_handle2lock(&data->cd_handle);
2163         if (!lease)
2164                 goto close;
2165
2166         /* check if the lease was already canceled */
2167         lock_res_and_lock(lease);
2168         rc = ldlm_is_cancel(lease);
2169         unlock_res_and_lock(lease);
2170
2171         if (rc) {
2172                 rc = -EAGAIN;
2173                 LDLM_DEBUG(lease, DFID" lease broken",
2174                            PFID(mdt_object_fid(obj)));
2175         }
2176
2177         /*
2178          * cancel server side lease, client side counterpart should have been
2179          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
2180          */
2181         ldlm_lock_cancel(lease);
2182         ldlm_reprocess_all(lease->l_resource, lease);
2183         LDLM_LOCK_PUT(lease);
2184
2185 close:
2186         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
2187         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2188         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
2189
2190         return rc ?: rc2;
2191 }
2192
2193 /*
2194  * migrate file in below steps:
2195  *  1. lock parent and its stripes
2196  *  2. lookup source by name
2197  *  3. lock parents of source links if source is not directory
2198  *  4. reject if source is in HSM
2199  *  5. take source open_sem and close file if source is regular file
2200  *  6. lock source and its stripes if it's directory
2201  *  7. lock target so subsequent change to it can trigger COS
2202  *  8. migrate file
2203  *  9. unlock above locks
2204  * 10. sync device if source has links
2205  */
2206 int mdt_reint_migrate(struct mdt_thread_info *info,
2207                       struct mdt_lock_handle *unused)
2208 {
2209         const struct lu_env *env = info->mti_env;
2210         struct mdt_device *mdt = info->mti_mdt;
2211         struct ptlrpc_request *req = mdt_info_req(info);
2212         struct mdt_reint_record *rr = &info->mti_rr;
2213         struct lu_ucred *uc = mdt_ucred(info);
2214         struct md_attr *ma = &info->mti_attr;
2215         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
2216         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
2217         struct mdt_object *pobj;
2218         struct mdt_object *spobj = NULL;
2219         struct mdt_object *sobj = NULL;
2220         struct mdt_object *tobj;
2221         struct lustre_handle rename_lh = { 0 };
2222         struct mdt_lock_handle *lhp;
2223         struct mdt_lock_handle *lhs;
2224         struct mdt_lock_handle *lht;
2225         LIST_HEAD(parent_slave_locks);
2226         LIST_HEAD(child_slave_locks);
2227         LIST_HEAD(link_locks);
2228         int lock_retries = 5;
2229         bool open_sem_locked = false;
2230         bool do_sync = false;
2231         int rc;
2232
2233         ENTRY;
2234         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2235                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2236
2237         if (info->mti_dlm_req)
2238                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2239
2240         if (!fid_is_md_operative(rr->rr_fid1) ||
2241             !fid_is_md_operative(rr->rr_fid2))
2242                 RETURN(-EPERM);
2243
2244         /* don't allow migrate . or .. */
2245         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2246                 RETURN(-EBUSY);
2247
2248         if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2249                 RETURN(-EPERM);
2250
2251         if (uc && !cap_raised(uc->uc_cap, CAP_SYS_ADMIN) &&
2252             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2253             mdt->mdt_enable_remote_dir_gid != -1)
2254                 RETURN(-EPERM);
2255
2256         /*
2257          * Note: do not enqueue rename lock for replay request, because
2258          * if other MDT holds rename lock, but being blocked to wait for
2259          * this MDT to finish its recovery, and the failover MDT can not
2260          * get rename lock, which will cause deadlock.
2261          *
2262          * req is NULL if this is called by directory auto-split.
2263          */
2264         if (req && !req_is_replay(req)) {
2265                 rc = mdt_rename_lock(info, &rename_lh);
2266                 if (rc != 0) {
2267                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2268                                mdt_obd_name(info->mti_mdt), rc);
2269                         RETURN(rc);
2270                 }
2271         }
2272
2273         /* pobj is master object of parent */
2274         pobj = mdt_object_find(env, mdt, rr->rr_fid1);
2275         if (IS_ERR(pobj))
2276                 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2277
2278         if (req) {
2279                 rc = mdt_version_get_check(info, pobj, 0);
2280                 if (rc)
2281                         GOTO(put_parent, rc);
2282         }
2283
2284         if (!mdt_object_exists(pobj))
2285                 GOTO(put_parent, rc = -ENOENT);
2286
2287         if (!S_ISDIR(lu_object_attr(&pobj->mot_obj)))
2288                 GOTO(put_parent, rc = -ENOTDIR);
2289
2290         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2291         if (rc)
2292                 GOTO(put_parent, rc);
2293
2294 lock_parent:
2295         /* lock parent object */
2296         lhp = &info->mti_lh[MDT_LH_PARENT];
2297         mdt_lock_reg_init(lhp, LCK_PW);
2298         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2299                                      &parent_slave_locks);
2300         if (rc)
2301                 GOTO(put_parent, rc);
2302
2303         /*
2304          * spobj is the corresponding stripe against name if pobj is striped
2305          * directory, which is the real parent, and no need to lock, because
2306          * we've taken full lock of pobj.
2307          */
2308         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2309         if (rc)
2310                 GOTO(unlock_parent, rc);
2311
2312         /* lock parents of source links, and revoke LOOKUP lock of links */
2313         rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2314                                    &parent_slave_locks, &link_locks);
2315         if (rc == -EBUSY && lock_retries-- > 0) {
2316                 mdt_object_put(env, sobj);
2317                 mdt_object_put(env, spobj);
2318                 goto lock_parent;
2319         }
2320
2321         if (rc < 0)
2322                 GOTO(put_source, rc);
2323
2324         /*
2325          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2326          * with one request, if total lock count exceeds this limit, we will
2327          * drop all locks after migration, and synchronous device in the end.
2328          */
2329         do_sync = rc;
2330
2331         /* TODO: DoM migration is not supported, migrate dirent only */
2332         if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2333                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2334                 if (rc)
2335                         GOTO(unlock_links, rc);
2336
2337                 if (ma->ma_valid & MA_LOV && mdt_lmm_dom_stripesize(ma->ma_lmm))
2338                         info->mti_spec.sp_migrate_nsonly = 1;
2339         } else if (S_ISDIR(lu_object_attr(&sobj->mot_obj))) {
2340                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2341                 if (rc)
2342                         GOTO(unlock_links, rc);
2343
2344                 /* race with restripe/auto-split? */
2345                 if ((ma->ma_valid & MA_LMV) &&
2346                     lmv_is_restriping(&ma->ma_lmv->lmv_md_v1))
2347                         GOTO(unlock_links, rc = -EBUSY);
2348         }
2349
2350         /* if migration HSM is allowed */
2351         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2352                 ma->ma_need = MA_HSM;
2353                 ma->ma_valid = 0;
2354                 rc = mdt_attr_get_complex(info, sobj, ma);
2355                 if (rc)
2356                         GOTO(unlock_links, rc);
2357
2358                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2359                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2360         }
2361
2362         /* end lease and close file for regular file */
2363         if (info->mti_spec.sp_migrate_close) {
2364                 /* try to hold open_sem so that nobody else can open the file */
2365                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2366                         /* close anyway */
2367                         mdd_migrate_close(info, sobj);
2368                         GOTO(unlock_links, rc = -EBUSY);
2369                 } else {
2370                         open_sem_locked = true;
2371                         rc = mdd_migrate_close(info, sobj);
2372                         if (rc)
2373                                 GOTO(unlock_open_sem, rc);
2374                 }
2375         }
2376
2377         /* lock source */
2378         lhs = &info->mti_lh[MDT_LH_OLD];
2379         mdt_lock_reg_init(lhs, LCK_EX);
2380         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2381                                      &child_slave_locks);
2382         if (rc)
2383                 GOTO(unlock_open_sem, rc);
2384
2385         /* lock target */
2386         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2387         if (IS_ERR(tobj))
2388                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2389
2390         lht = &info->mti_lh[MDT_LH_NEW];
2391         mdt_lock_reg_init(lht, LCK_EX);
2392         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2393         if (rc)
2394                 GOTO(put_target, rc);
2395
2396         /* Don't do lookup sanity check. We know name doesn't exist. */
2397         info->mti_spec.sp_cr_lookup = 0;
2398         info->mti_spec.sp_feat = &dt_directory_features;
2399
2400         rc = mdo_migrate(env, mdt_object_child(pobj),
2401                          mdt_object_child(sobj), &rr->rr_name,
2402                          mdt_object_child(tobj),
2403                          &info->mti_spec, ma);
2404         if (!rc)
2405                 lprocfs_counter_incr(mdt->mdt_lu_dev.ld_obd->obd_md_stats,
2406                                      LPROC_MDT_MIGRATE + LPROC_MD_LAST_OPC);
2407         EXIT;
2408
2409         mdt_object_unlock(info, tobj, lht, rc);
2410 put_target:
2411         mdt_object_put(env, tobj);
2412 unlock_source:
2413         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2414                                   &child_slave_locks, rc);
2415 unlock_open_sem:
2416         if (open_sem_locked)
2417                 up_write(&sobj->mot_open_sem);
2418 unlock_links:
2419         /* if we've got too many locks to save into RPC,
2420          * then just commit before the locks are released
2421          */
2422         if (!rc && do_sync)
2423                 mdt_device_sync(env, mdt);
2424         mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2425 put_source:
2426         mdt_object_put(env, sobj);
2427         mdt_object_put(env, spobj);
2428 unlock_parent:
2429         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2430                                   &parent_slave_locks, rc);
2431 put_parent:
2432         mdt_object_put(env, pobj);
2433 unlock_rename:
2434         if (lustre_handle_is_used(&rename_lh))
2435                 mdt_rename_unlock(&rename_lh);
2436
2437         return rc;
2438 }
2439
2440 static int mdt_object_lock_save(struct mdt_thread_info *info,
2441                                 struct mdt_object *dir,
2442                                 struct mdt_lock_handle *lh,
2443                                 int idx, bool cos_incompat)
2444 {
2445         int rc;
2446
2447         /* we lock the target dir if it is local */
2448         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2449                                    cos_incompat);
2450         if (rc != 0)
2451                 return rc;
2452
2453         /* get and save correct version after locking */
2454         mdt_version_get_save(info, dir, idx);
2455         return 0;
2456 }
2457
2458 /*
2459  * determine lock order of sobj and tobj
2460  *
2461  * there are two situations we need to lock tobj before sobj:
2462  * 1. sobj is child of tobj
2463  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2464  *    larger than that of tobj
2465  *
2466  * \retval      1 lock tobj before sobj
2467  * \retval      0 lock sobj before tobj
2468  * \retval      -ev negative errno upon error
2469  */
2470 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2471                                            struct mdt_object *sobj,
2472                                            struct mdt_object *tobj)
2473 {
2474         struct md_attr *ma = &info->mti_attr;
2475         struct lu_fid *spfid = &info->mti_tmp_fid1;
2476         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2477         struct lmv_mds_md_v1 *lmv;
2478         __u32 sindex;
2479         __u32 tindex;
2480         int rc;
2481
2482         /* sobj and tobj are the same */
2483         if (sobj == tobj)
2484                 return 0;
2485
2486         if (fid_is_root(mdt_object_fid(sobj)))
2487                 return 0;
2488
2489         if (fid_is_root(mdt_object_fid(tobj)))
2490                 return 1;
2491
2492         /* check whether sobj is child of tobj */
2493         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2494                            mdt_object_fid(tobj));
2495         if (rc < 0)
2496                 return rc;
2497
2498         if (rc == 1)
2499                 return 1;
2500
2501         /* check whether sobj and tobj are children of the same parent */
2502         rc = mdt_attr_get_pfid(info, sobj, spfid);
2503         if (rc)
2504                 return rc;
2505
2506         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2507         if (rc)
2508                 return rc;
2509
2510         if (!lu_fid_eq(spfid, tpfid))
2511                 return 0;
2512
2513         /* check whether sobj and tobj are sibling stripes */
2514         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2515         if (rc)
2516                 return rc;
2517
2518         if (!(ma->ma_valid & MA_LMV))
2519                 return 0;
2520
2521         lmv = &ma->ma_lmv->lmv_md_v1;
2522         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2523                 return 0;
2524         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2525
2526         ma->ma_valid = 0;
2527         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2528         if (rc)
2529                 return rc;
2530
2531         if (!(ma->ma_valid & MA_LMV))
2532                 return -ENODATA;
2533
2534         lmv = &ma->ma_lmv->lmv_md_v1;
2535         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2536                 return -EINVAL;
2537         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2538
2539         /* check stripe index of sobj and tobj */
2540         if (sindex == tindex)
2541                 return -EINVAL;
2542
2543         return sindex < tindex ? 0 : 1;
2544 }
2545
2546 /*
2547  * lock rename source object.
2548  *
2549  * Both source and source parent may be remote, and source may be a remote
2550  * object on source parent, to avoid overriding lock handle, store remote
2551  * LOOKUP lock separately in @lhr.
2552  *
2553  * \retval      0 on success
2554  * \retval      -ev negative errno upon error
2555  */
2556 static int mdt_rename_source_lock(struct mdt_thread_info *info,
2557                                   struct mdt_object *parent,
2558                                   struct mdt_object *child,
2559                                   struct mdt_lock_handle *lhc,
2560                                   struct mdt_lock_handle *lhr,
2561                                   __u64 ibits,
2562                                   bool cos_incompat)
2563 {
2564         int rc;
2565
2566         rc = mdt_is_remote_object(info, parent, child);
2567         if (rc < 0)
2568                 return rc;
2569
2570         if (rc) {
2571                 /* enqueue remote LOOKUP lock from the parent MDT */
2572                 __u64 rmt_ibits = MDS_INODELOCK_LOOKUP;
2573
2574                 if (mdt_object_remote(parent)) {
2575                         rc = mdt_remote_object_lock(info, parent,
2576                                                     mdt_object_fid(child),
2577                                                     &lhr->mlh_rreg_lh,
2578                                                     lhr->mlh_rreg_mode,
2579                                                     rmt_ibits, false);
2580                         if (rc != ELDLM_OK)
2581                                 return rc;
2582                 } else {
2583                         LASSERT(mdt_object_remote(child));
2584                         rc = mdt_object_local_lock(info, child, lhr,
2585                                                    &rmt_ibits, 0, true);
2586                         if (rc < 0)
2587                                 return rc;
2588                 }
2589
2590                 ibits &= ~MDS_INODELOCK_LOOKUP;
2591         }
2592
2593         if (mdt_object_remote(child)) {
2594                 rc = mdt_remote_object_lock(info, child, mdt_object_fid(child),
2595                                             &lhc->mlh_rreg_lh,
2596                                             lhc->mlh_rreg_mode,
2597                                             ibits, false);
2598                 if (rc == ELDLM_OK)
2599                         rc = 0;
2600         } else {
2601                 rc = mdt_reint_object_lock(info, child, lhc, ibits,
2602                                            cos_incompat);
2603         }
2604
2605         if (!rc)
2606                 mdt_object_unlock(info, child, lhr, rc);
2607
2608         return rc;
2609 }
2610
2611 /*
2612  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2613  * 2 - srcdir child; 3 - tgtdir child.
2614  * Update on disk version of srcdir child.
2615  */
2616 static int mdt_reint_rename(struct mdt_thread_info *info,
2617                             struct mdt_lock_handle *unused)
2618 {
2619         struct mdt_device *mdt = info->mti_mdt;
2620         struct mdt_reint_record *rr = &info->mti_rr;
2621         struct md_attr *ma = &info->mti_attr;
2622         struct ptlrpc_request *req = mdt_info_req(info);
2623         struct mdt_object *msrcdir = NULL;
2624         struct mdt_object *mtgtdir = NULL;
2625         struct mdt_object *mold;
2626         struct mdt_object *mnew = NULL;
2627         struct lustre_handle rename_lh = { 0 };
2628         struct mdt_lock_handle *lh_srcdirp;
2629         struct mdt_lock_handle *lh_tgtdirp;
2630         struct mdt_lock_handle *lh_oldp = NULL;
2631         struct mdt_lock_handle *lh_rmt = NULL;
2632         struct mdt_lock_handle *lh_newp = NULL;
2633         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2634         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2635         __u64 lock_ibits;
2636         bool reverse = false, discard = false;
2637         bool cos_incompat;
2638         ktime_t kstart = ktime_get();
2639         int rc;
2640
2641         ENTRY;
2642         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2643                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2644                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2645
2646         if (info->mti_dlm_req)
2647                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2648
2649         if (!fid_is_md_operative(rr->rr_fid1) ||
2650             !fid_is_md_operative(rr->rr_fid2))
2651                 RETURN(-EPERM);
2652
2653         /* find both parents. */
2654         msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2655         if (IS_ERR(msrcdir))
2656                 RETURN(PTR_ERR(msrcdir));
2657
2658         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2659
2660         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2661                 mtgtdir = msrcdir;
2662                 mdt_object_get(info->mti_env, mtgtdir);
2663         } else {
2664                 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2665                 if (IS_ERR(mtgtdir))
2666                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2667         }
2668
2669         /*
2670          * Note: do not enqueue rename lock for replay request, because
2671          * if other MDT holds rename lock, but being blocked to wait for
2672          * this MDT to finish its recovery, and the failover MDT can not
2673          * get rename lock, which will cause deadlock.
2674          */
2675         if (!req_is_replay(req)) {
2676                 /*
2677                  * Normally rename RPC is handled on the MDT with the target
2678                  * directory (if target exists, it's on the MDT with the
2679                  * target), if the source directory is remote, it's a hint that
2680                  * source is remote too (this may not be true, but it won't
2681                  * cause any issue), return -EXDEV early to avoid taking
2682                  * rename_lock.
2683                  */
2684                 if (!mdt->mdt_enable_remote_rename &&
2685                     mdt_object_remote(msrcdir))
2686                         GOTO(out_put_tgtdir, rc = -EXDEV);
2687
2688                 /* This might be further relaxed in the future for regular file
2689                  * renames in different source and target parents. Start with
2690                  * only same-directory renames for simplicity and because this
2691                  * is by far the most the common use case.
2692                  */
2693                 if (msrcdir != mtgtdir) {
2694                         rc = mdt_rename_lock(info, &rename_lh);
2695                         if (rc != 0) {
2696                                 CERROR("%s: cannot lock for rename: rc = %d\n",
2697                                        mdt_obd_name(mdt), rc);
2698                                 GOTO(out_put_tgtdir, rc);
2699                         }
2700                 } else {
2701                         CDEBUG(D_INFO, "%s: samedir rename "DFID"/"DNAME"\n",
2702                                mdt_obd_name(mdt), PFID(rr->rr_fid1),
2703                                PNAME(&rr->rr_name));
2704                 }
2705         }
2706
2707         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2708         if (rc < 0)
2709                 GOTO(out_unlock_rename, rc);
2710
2711         reverse = rc;
2712
2713         /* source needs to be looked up after locking source parent, otherwise
2714          * this rename may race with unlink source, and cause rename hang, see
2715          * sanityn.sh 55b, so check parents first, if later we found source is
2716          * remote, relock parents.
2717          */
2718         cos_incompat = (mdt_object_remote(msrcdir) ||
2719                         mdt_object_remote(mtgtdir));
2720
2721         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2722
2723         /* lock parents in the proper order. */
2724         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2725         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2726
2727         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
2728         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
2729 relock:
2730         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2731         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2732
2733         if (reverse) {
2734                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2735                                           cos_incompat);
2736                 if (rc)
2737                         GOTO(out_unlock_rename, rc);
2738
2739                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2740
2741                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2742                                           cos_incompat);
2743                 if (rc != 0) {
2744                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2745                         GOTO(out_unlock_rename, rc);
2746                 }
2747         } else {
2748                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2749                                           cos_incompat);
2750                 if (rc)
2751                         GOTO(out_unlock_rename, rc);
2752
2753                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2754
2755                 if (mtgtdir != msrcdir) {
2756                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2757                                                   cos_incompat);
2758                 } else if (!mdt_object_remote(mtgtdir) &&
2759                            lh_srcdirp->mlh_pdo_hash !=
2760                            lh_tgtdirp->mlh_pdo_hash) {
2761                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2762                                                 MDS_INODELOCK_UPDATE,
2763                                                 cos_incompat);
2764                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2765                 }
2766                 if (rc != 0) {
2767                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2768                         GOTO(out_unlock_rename, rc);
2769                 }
2770         }
2771
2772         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2773         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2774
2775         /* find mold object. */
2776         fid_zero(old_fid);
2777         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2778         if (rc != 0)
2779                 GOTO(out_unlock_parents, rc);
2780
2781         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2782                 GOTO(out_unlock_parents, rc = -EINVAL);
2783
2784         if (!fid_is_md_operative(old_fid))
2785                 GOTO(out_unlock_parents, rc = -EPERM);
2786
2787         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2788         if (IS_ERR(mold))
2789                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2790
2791         if (!mdt_object_exists(mold)) {
2792                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2793                                 &mold->mot_obj,
2794                                 "object does not exist");
2795                 GOTO(out_put_old, rc = -ENOENT);
2796         }
2797
2798         if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2799                 GOTO(out_put_old, rc = -EXDEV);
2800
2801         /* Check if @mtgtdir is subdir of @mold, before locking child
2802          * to avoid reverse locking.
2803          */
2804         if (mtgtdir != msrcdir) {
2805                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2806                                    old_fid);
2807                 if (rc) {
2808                         if (rc == 1)
2809                                 rc = -EINVAL;
2810                         GOTO(out_put_old, rc);
2811                 }
2812         }
2813
2814         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2815         /* save version after locking */
2816         mdt_version_get_save(info, mold, 2);
2817
2818         if (!cos_incompat && mdt_object_remote(mold)) {
2819                 cos_incompat = true;
2820                 mdt_object_put(info->mti_env, mold);
2821                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2822                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2823                 goto relock;
2824         }
2825
2826         /* find mnew object:
2827          * mnew target object may not exist now
2828          * lookup with version checking
2829          */
2830         fid_zero(new_fid);
2831         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2832                                       3);
2833         if (rc == 0) {
2834                 /* the new_fid should have been filled at this moment */
2835                 if (lu_fid_eq(old_fid, new_fid))
2836                         GOTO(out_put_old, rc);
2837
2838                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2839                     lu_fid_eq(new_fid, rr->rr_fid2))
2840                         GOTO(out_put_old, rc = -EINVAL);
2841
2842                 if (!fid_is_md_operative(new_fid))
2843                         GOTO(out_put_old, rc = -EPERM);
2844
2845                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2846                 if (IS_ERR(mnew))
2847                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2848
2849                 if (!mdt_object_exists(mnew)) {
2850                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2851                                         &mnew->mot_obj,
2852                                         "object does not exist");
2853                         GOTO(out_put_new, rc = -ENOENT);
2854                 }
2855
2856                 if (mdt_object_remote(mnew)) {
2857                         struct mdt_body  *repbody;
2858
2859                         /* Always send rename req to the target child MDT */
2860                         repbody = req_capsule_server_get(info->mti_pill,
2861                                                          &RMF_MDT_BODY);
2862                         LASSERT(repbody != NULL);
2863                         repbody->mbo_fid1 = *new_fid;
2864                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2865                         GOTO(out_put_new, rc = -EXDEV);
2866                 }
2867                 /* Before locking the target dir, check we do not replace
2868                  * a dir with a non-dir, otherwise it may deadlock with
2869                  * link op which tries to create a link in this dir
2870                  * back to this non-dir.
2871                  */
2872                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2873                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2874                         GOTO(out_put_new, rc = -EISDIR);
2875
2876                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2877                 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2878                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2879                 mdt_lock_reg_init(lh_rmt, LCK_EX);
2880                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2881                 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2882                                             lh_rmt, lock_ibits, cos_incompat);
2883                 if (rc < 0)
2884                         GOTO(out_put_new, rc);
2885
2886                 /* Check if @msrcdir is subdir of @mnew, before locking child
2887                  * to avoid reverse locking.
2888                  */
2889                 if (mtgtdir != msrcdir) {
2890                         rc = mdo_is_subdir(info->mti_env,
2891                                            mdt_object_child(msrcdir), new_fid);
2892                         if (rc) {
2893                                 if (rc == 1)
2894                                         rc = -EINVAL;
2895                                 GOTO(out_unlock_old, rc);
2896                         }
2897                 }
2898
2899                 /* We used to acquire MDS_INODELOCK_FULL here but we
2900                  * can't do this now because a running HSM restore on
2901                  * the rename onto victim will hold the layout
2902                  * lock. See LU-4002.
2903                  */
2904
2905                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2906                 mdt_lock_reg_init(lh_newp, LCK_EX);
2907                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
2908                 if (mdt_object_remote(mtgtdir)) {
2909                         rc = mdt_remote_object_lock(info, mtgtdir,
2910                                                     mdt_object_fid(mnew),
2911                                                     &lh_newp->mlh_rreg_lh,
2912                                                     lh_newp->mlh_rreg_mode,
2913                                                     MDS_INODELOCK_LOOKUP,
2914                                                     false);
2915                         if (rc != ELDLM_OK)
2916                                 GOTO(out_unlock_old, rc);
2917
2918                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2919                 }
2920                 rc = mdt_reint_object_lock(info, mnew, lh_newp, lock_ibits,
2921                                            cos_incompat);
2922                 if (rc != 0)
2923                         GOTO(out_unlock_new, rc);
2924
2925                 /* get and save version after locking */
2926                 mdt_version_get_save(info, mnew, 3);
2927         } else if (rc != -ENOENT) {
2928                 GOTO(out_put_old, rc);
2929         } else {
2930                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2931                 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2932                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2933                 mdt_lock_reg_init(lh_rmt, LCK_EX);
2934                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2935                 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2936                                             lh_rmt, lock_ibits, cos_incompat);
2937                 if (rc != 0)
2938                         GOTO(out_put_old, rc);
2939
2940                 mdt_enoent_version_save(info, 3);
2941         }
2942
2943         /* step 5: rename it */
2944         mdt_reint_init_ma(info, ma);
2945
2946         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2947                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2948
2949         if (mnew != NULL)
2950                 mutex_lock(&mnew->mot_lov_mutex);
2951
2952         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2953                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2954                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2955                         &rr->rr_tgt_name, ma);
2956
2957         if (mnew != NULL)
2958                 mutex_unlock(&mnew->mot_lov_mutex);
2959
2960         /* handle last link of tgt object */
2961         if (rc == 0) {
2962                 mdt_counter_incr(req, LPROC_MDT_RENAME,
2963                                  ktime_us_delta(ktime_get(), kstart));
2964                 if (mnew) {
2965                         mdt_handle_last_unlink(info, mnew, ma);
2966                         discard = mdt_dom_check_for_discard(info, mnew);
2967                 }
2968                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2969                                          msrcdir, mtgtdir,
2970                                          ktime_us_delta(ktime_get(), kstart));
2971         }
2972
2973         EXIT;
2974 out_unlock_new:
2975         if (mnew != NULL)
2976                 mdt_object_unlock(info, mnew, lh_newp, rc);
2977 out_unlock_old:
2978         mdt_object_unlock(info, NULL, lh_rmt, rc);
2979         mdt_object_unlock(info, mold, lh_oldp, rc);
2980 out_put_new:
2981         if (mnew && !discard)
2982                 mdt_object_put(info->mti_env, mnew);
2983 out_put_old:
2984         mdt_object_put(info->mti_env, mold);
2985 out_unlock_parents:
2986         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2987         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2988 out_unlock_rename:
2989         if (lustre_handle_is_used(&rename_lh))
2990                 mdt_rename_unlock(&rename_lh);
2991 out_put_tgtdir:
2992         mdt_object_put(info->mti_env, mtgtdir);
2993 out_put_srcdir:
2994         mdt_object_put(info->mti_env, msrcdir);
2995
2996         /* The DoM discard can be done right in the place above where it is
2997          * assigned, meanwhile it is done here after rename unlock due to
2998          * compatibility with old clients, for them the discard blocks
2999          * the main thread until completion. Check LU-11359 for details.
3000          */
3001         if (discard) {
3002                 mdt_dom_discard_data(info, mnew);
3003                 mdt_object_put(info->mti_env, mnew);
3004         }
3005         OBD_RACE(OBD_FAIL_MDS_LINK_RENAME_RACE);
3006         return rc;
3007 }
3008
3009 static int mdt_reint_resync(struct mdt_thread_info *info,
3010                             struct mdt_lock_handle *lhc)
3011 {
3012         struct mdt_reint_record *rr = &info->mti_rr;
3013         struct ptlrpc_request *req = mdt_info_req(info);
3014         struct md_attr *ma = &info->mti_attr;
3015         struct mdt_object *mo;
3016         struct ldlm_lock *lease;
3017         struct mdt_body *repbody;
3018         struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
3019         bool lease_broken;
3020         int rc, rc2;
3021
3022         ENTRY;
3023         DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
3024
3025         if (info->mti_dlm_req)
3026                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
3027
3028         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
3029         if (IS_ERR(mo))
3030                 GOTO(out, rc = PTR_ERR(mo));
3031
3032         if (!mdt_object_exists(mo))
3033                 GOTO(out_obj, rc = -ENOENT);
3034
3035         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
3036                 GOTO(out_obj, rc = -EINVAL);
3037
3038         if (mdt_object_remote(mo))
3039                 GOTO(out_obj, rc = -EREMOTE);
3040
3041         lease = ldlm_handle2lock(rr->rr_lease_handle);
3042         if (lease == NULL)
3043                 GOTO(out_obj, rc = -ESTALE);
3044
3045         /* It's really necessary to grab open_sem and check if the lease lock
3046          * has been lost. There would exist a concurrent writer coming in and
3047          * generating some dirty data in memory cache, the writeback would fail
3048          * after the layout version is increased by MDS_REINT_RESYNC RPC.
3049          */
3050         if (!down_write_trylock(&mo->mot_open_sem))
3051                 GOTO(out_put_lease, rc = -EBUSY);
3052
3053         lock_res_and_lock(lease);
3054         lease_broken = ldlm_is_cancel(lease);
3055         unlock_res_and_lock(lease);
3056         if (lease_broken)
3057                 GOTO(out_unlock, rc = -EBUSY);
3058
3059         /* the file has yet opened by anyone else after we took the lease. */
3060         layout.mlc_opc = MD_LAYOUT_RESYNC;
3061         lhc = &info->mti_lh[MDT_LH_LOCAL];
3062         rc = mdt_layout_change(info, mo, lhc, &layout);
3063         if (rc)
3064                 GOTO(out_unlock, rc);
3065
3066         mdt_object_unlock(info, mo, lhc, 0);
3067
3068         ma->ma_need = MA_INODE;
3069         ma->ma_valid = 0;
3070         rc = mdt_attr_get_complex(info, mo, ma);
3071         if (rc != 0)
3072                 GOTO(out_unlock, rc);
3073
3074         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
3075         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
3076
3077         EXIT;
3078 out_unlock:
3079         up_write(&mo->mot_open_sem);
3080 out_put_lease:
3081         LDLM_LOCK_PUT(lease);
3082 out_obj:
3083         mdt_object_put(info->mti_env, mo);
3084 out:
3085         mdt_client_compatibility(info);
3086         rc2 = mdt_fix_reply(info);
3087         if (rc == 0)
3088                 rc = rc2;
3089         return rc;
3090 }
3091
3092 struct mdt_reinter {
3093         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
3094         enum lprocfs_extra_opc mr_extra_opc;
3095 };
3096
3097 static const struct mdt_reinter mdt_reinters[] = {
3098         [REINT_SETATTR] = {
3099                 .mr_handler = &mdt_reint_setattr,
3100                 .mr_extra_opc = MDS_REINT_SETATTR,
3101         },
3102         [REINT_CREATE] = {
3103                 .mr_handler = &mdt_reint_create,
3104                 .mr_extra_opc = MDS_REINT_CREATE,
3105         },
3106         [REINT_LINK] = {
3107                 .mr_handler = &mdt_reint_link,
3108                 .mr_extra_opc = MDS_REINT_LINK,
3109         },
3110         [REINT_UNLINK] = {
3111                 .mr_handler = &mdt_reint_unlink,
3112                 .mr_extra_opc = MDS_REINT_UNLINK,
3113         },
3114         [REINT_RENAME] = {
3115                 .mr_handler = &mdt_reint_rename,
3116                 .mr_extra_opc = MDS_REINT_RENAME,
3117         },
3118         [REINT_OPEN] = {
3119                 .mr_handler = &mdt_reint_open,
3120                 .mr_extra_opc = MDS_REINT_OPEN,
3121         },
3122         [REINT_SETXATTR] = {
3123                 .mr_handler = &mdt_reint_setxattr,
3124                 .mr_extra_opc = MDS_REINT_SETXATTR,
3125         },
3126         [REINT_RMENTRY] = {
3127                 .mr_handler = &mdt_reint_unlink,
3128                 .mr_extra_opc = MDS_REINT_UNLINK,
3129         },
3130         [REINT_MIGRATE] = {
3131                 .mr_handler = &mdt_reint_migrate,
3132                 .mr_extra_opc = MDS_REINT_RENAME,
3133         },
3134         [REINT_RESYNC] = {
3135                 .mr_handler = &mdt_reint_resync,
3136                 .mr_extra_opc = MDS_REINT_RESYNC,
3137         },
3138 };
3139
3140 int mdt_reint_rec(struct mdt_thread_info *info,
3141                   struct mdt_lock_handle *lhc)
3142 {
3143         const struct mdt_reinter *mr;
3144         int rc;
3145
3146         ENTRY;
3147         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
3148                 RETURN(-EPROTO);
3149
3150         mr = &mdt_reinters[info->mti_rr.rr_opcode];
3151         if (mr->mr_handler == NULL)
3152                 RETURN(-EPROTO);
3153
3154         rc = (*mr->mr_handler)(info, lhc);
3155
3156         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
3157                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
3158
3159         RETURN(rc);
3160 }