lustre/mdt/mdt_reint.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/mdt/mdt_reint.c
  33  *
  34  * Lustre Metadata Target (mdt) reintegration routines
  35  *
  36  * Author: Peter Braam <braam@clusterfs.com>
  37  * Author: Andreas Dilger <adilger@clusterfs.com>
  38  * Author: Phil Schwan <phil@clusterfs.com>
  39  * Author: Huang Hua <huanghua@clusterfs.com>
  40  * Author: Yury Umanets <umka@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_MDS
  44
  45 #include <lprocfs_status.h>
  46 #include "mdt_internal.h"
  47 #include <lustre_lmv.h>
  48
  49 static inline void mdt_reint_init_ma(struct mdt_thread_info *info,
  50                                      struct md_attr *ma)
  51 {
  52         ma->ma_need = MA_INODE;
  53         ma->ma_valid = 0;
  54 }
  55
  56 /**
  57  * Get version of object by fid.
  58  *
  59  * Return real version or ENOENT_VERSION if object doesn't exist
  60  */
  61 static void mdt_obj_version_get(struct mdt_thread_info *info,
  62                                 struct mdt_object *o, __u64 *version)
  63 {
  64         LASSERT(o);
  65
  66         if (mdt_object_exists(o) && !mdt_object_remote(o) &&
  67             !fid_is_obf(mdt_object_fid(o)))
  68                 *version = dt_version_get(info->mti_env, mdt_obj2dt(o));
  69         else
  70                 *version = ENOENT_VERSION;
  71         CDEBUG(D_INODE, "FID "DFID" version is %#llx\n",
  72                PFID(mdt_object_fid(o)), *version);
  73 }
  74
  75 /**
  76  * Check version is correct.
  77  *
  78  * Should be called only during replay.
  79  */
  80 static int mdt_version_check(struct ptlrpc_request *req,
  81                              __u64 version, int idx)
  82 {
  83         __u64 *pre_ver = lustre_msg_get_versions(req->rq_reqmsg);
  84         ENTRY;
  85
  86         if (!exp_connect_vbr(req->rq_export))
  87                 RETURN(0);
  88
  89         LASSERT(req_is_replay(req));
  90         /** VBR: version is checked always because costs nothing */
  91         LASSERT(idx < PTLRPC_NUM_VERSIONS);
  92         /** Sanity check for malformed buffers */
  93         if (pre_ver == NULL) {
  94                 CERROR("No versions in request buffer\n");
  95                 spin_lock(&req->rq_export->exp_lock);
  96                 req->rq_export->exp_vbr_failed = 1;
  97                 spin_unlock(&req->rq_export->exp_lock);
  98                 RETURN(-EOVERFLOW);
  99         } else if (pre_ver[idx] != version) {
 100                 CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n",
 101                        pre_ver[idx], version);
 102                 spin_lock(&req->rq_export->exp_lock);
 103                 req->rq_export->exp_vbr_failed = 1;
 104                 spin_unlock(&req->rq_export->exp_lock);
 105                 RETURN(-EOVERFLOW);
 106         }
 107         RETURN(0);
 108 }
 109
 110 /**
 111  * Save pre-versions in reply.
 112  */
 113 static void mdt_version_save(struct ptlrpc_request *req, __u64 version,
 114                              int idx)
 115 {
 116         __u64 *reply_ver;
 117
 118         if (!exp_connect_vbr(req->rq_export))
 119                 return;
 120
 121         LASSERT(!req_is_replay(req));
 122         LASSERT(req->rq_repmsg != NULL);
 123         reply_ver = lustre_msg_get_versions(req->rq_repmsg);
 124         if (reply_ver)
 125                 reply_ver[idx] = version;
 126 }
 127
 128 /**
 129  * Save enoent version, it is needed when it is obvious that object doesn't
 130  * exist, e.g. child during create.
 131  */
 132 static void mdt_enoent_version_save(struct mdt_thread_info *info, int idx)
 133 {
 134         /* save version of file name for replay, it must be ENOENT here */
 135         if (!req_is_replay(mdt_info_req(info))) {
 136                 info->mti_ver[idx] = ENOENT_VERSION;
 137                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 138         }
 139 }
 140
 141 /**
 142  * Get version from disk and save in reply buffer.
 143  *
 144  * Versions are saved in reply only during normal operations not replays.
 145  */
 146 void mdt_version_get_save(struct mdt_thread_info *info,
 147                           struct mdt_object *mto, int idx)
 148 {
 149         /* don't save versions during replay */
 150         if (!req_is_replay(mdt_info_req(info))) {
 151                 mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 152                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 153         }
 154 }
 155
 156 /**
 157  * Get version from disk and check it, no save in reply.
 158  */
 159 int mdt_version_get_check(struct mdt_thread_info *info,
 160                           struct mdt_object *mto, int idx)
 161 {
 162         /* only check versions during replay */
 163         if (!req_is_replay(mdt_info_req(info)))
 164                 return 0;
 165
 166         mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 167         return mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
 168 }
 169
 170 /**
 171  * Get version from disk and check if recovery or just save.
 172  */
 173 int mdt_version_get_check_save(struct mdt_thread_info *info,
 174                                struct mdt_object *mto, int idx)
 175 {
 176         int rc = 0;
 177
 178         mdt_obj_version_get(info, mto, &info->mti_ver[idx]);
 179         if (req_is_replay(mdt_info_req(info)))
 180                 rc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx],
 181                                        idx);
 182         else
 183                 mdt_version_save(mdt_info_req(info), info->mti_ver[idx], idx);
 184         return rc;
 185 }
 186
 187 /**
 188  * Lookup with version checking.
 189  *
 190  * This checks version of 'name'. Many reint functions uses 'name' for child not
 191  * FID, therefore we need to get object by name and check its version.
 192  */
 193 static int mdt_lookup_version_check(struct mdt_thread_info *info,
 194                                     struct mdt_object *p,
 195                                     const struct lu_name *lname,
 196                                     struct lu_fid *fid, int idx)
 197 {
 198         int rc, vbrc;
 199
 200         rc = mdo_lookup(info->mti_env, mdt_object_child(p), lname, fid,
 201                         &info->mti_spec);
 202         /* Check version only during replay */
 203         if (!req_is_replay(mdt_info_req(info)))
 204                 return rc;
 205
 206         info->mti_ver[idx] = ENOENT_VERSION;
 207         if (rc == 0) {
 208                 struct mdt_object *child;
 209                 child = mdt_object_find(info->mti_env, info->mti_mdt, fid);
 210                 if (likely(!IS_ERR(child))) {
 211                         mdt_obj_version_get(info, child, &info->mti_ver[idx]);
 212                         mdt_object_put(info->mti_env, child);
 213                 }
 214         }
 215         vbrc = mdt_version_check(mdt_info_req(info), info->mti_ver[idx], idx);
 216         return vbrc ? vbrc : rc;
 217
 218 }
 219
 220 static int mdt_unlock_slaves(struct mdt_thread_info *mti,
 221                              struct mdt_object *obj,
 222                              struct ldlm_enqueue_info *einfo,
 223                              int decref)
 224 {
 225         union ldlm_policy_data *policy = &mti->mti_policy;
 226         struct mdt_lock_handle *lh = &mti->mti_lh[MDT_LH_LOCAL];
 227         struct lustre_handle_array *slave_locks = einfo->ei_cbdata;
 228         int i;
 229
 230         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 231         LASSERT(slave_locks);
 232
 233         memset(policy, 0, sizeof(*policy));
 234         policy->l_inodebits.bits = einfo->ei_inodebits;
 235         mdt_lock_handle_init(lh);
 236         mdt_lock_reg_init(lh, einfo->ei_mode);
 237         for (i = 0; i < slave_locks->ha_count; i++) {
 238                 if (test_bit(i, (void *)slave_locks->ha_map))
 239                         lh->mlh_rreg_lh = slave_locks->ha_handles[i];
 240                 else
 241                         lh->mlh_reg_lh = slave_locks->ha_handles[i];
 242                 mdt_object_unlock(mti, NULL, lh, decref);
 243                 slave_locks->ha_handles[i].cookie = 0ull;
 244         }
 245
 246         return mo_object_unlock(mti->mti_env, mdt_object_child(obj), einfo,
 247                                 policy);
 248 }
 249
 250 static inline int mdt_object_striped(struct mdt_thread_info *mti,
 251                                      struct mdt_object *obj)
 252 {
 253         int rc;
 254
 255         if (!S_ISDIR(obj->mot_header.loh_attr))
 256                 return 0;
 257
 258         rc = mo_xattr_get(mti->mti_env, mdt_object_child(obj), &LU_BUF_NULL,
 259                           XATTR_NAME_LMV);
 260         if (rc <= 0)
 261                 return rc == -ENODATA ? 0 : rc;
 262
 263         return 1;
 264 }
 265
 266 /**
 267  * Lock slave stripes if necessary, the lock handles of slave stripes
 268  * will be stored in einfo->ei_cbdata.
 269  **/
 270 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
 271                            enum ldlm_mode mode, __u64 ibits,
 272                            struct ldlm_enqueue_info *einfo)
 273 {
 274         union ldlm_policy_data *policy = &mti->mti_policy;
 275
 276         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 277
 278         einfo->ei_type = LDLM_IBITS;
 279         einfo->ei_mode = mode;
 280         einfo->ei_cb_bl = mdt_remote_blocking_ast;
 281         einfo->ei_cb_local_bl = mdt_blocking_ast;
 282         einfo->ei_cb_cp = ldlm_completion_ast;
 283         einfo->ei_enq_slave = 1;
 284         einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
 285         einfo->ei_inodebits = ibits;
 286         memset(policy, 0, sizeof(*policy));
 287         policy->l_inodebits.bits = ibits;
 288
 289         return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
 290                               policy);
 291 }
 292
 293 int mdt_reint_striped_lock(struct mdt_thread_info *info,
 294                            struct mdt_object *o,
 295                            struct mdt_lock_handle *lh,
 296                            __u64 ibits,
 297                            struct ldlm_enqueue_info *einfo,
 298                            bool cos_incompat)
 299 {
 300         int rc;
 301
 302         LASSERT(!mdt_object_remote(o));
 303
 304         memset(einfo, 0, sizeof(*einfo));
 305
 306         rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
 307         if (rc)
 308                 return rc;
 309
 310         rc = mdt_object_striped(info, o);
 311         if (rc != 1) {
 312                 if (rc < 0)
 313                         mdt_object_unlock(info, o, lh, rc);
 314                 return rc;
 315         }
 316
 317         rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
 318         if (rc) {
 319                 mdt_object_unlock(info, o, lh, rc);
 320                 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
 321                         rc = 0;
 322         }
 323
 324         return rc;
 325 }
 326
 327 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
 328                               struct mdt_object *o,
 329                               struct mdt_lock_handle *lh,
 330                               struct ldlm_enqueue_info *einfo, int decref)
 331 {
 332         if (einfo->ei_cbdata)
 333                 mdt_unlock_slaves(info, o, einfo, decref);
 334         mdt_object_unlock(info, o, lh, decref);
 335 }
 336
 337 /*
 338  * VBR: we save three versions in reply:
 339  * 0 - parent. Check that parent version is the same during replay.
 340  * 1 - name. Version of 'name' if file exists with the same name or
 341  * ENOENT_VERSION, it is needed because file may appear due to missed replays.
 342  * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
 343  * check.
 344  */
 345 static int mdt_create(struct mdt_thread_info *info)
 346 {
 347         struct mdt_device       *mdt = info->mti_mdt;
 348         struct mdt_object       *parent;
 349         struct mdt_object       *child;
 350         struct mdt_lock_handle  *lh;
 351         struct mdt_body         *repbody;
 352         struct md_attr          *ma = &info->mti_attr;
 353         struct mdt_reint_record *rr = &info->mti_rr;
 354         struct md_op_spec       *spec = &info->mti_spec;
 355         int rc;
 356         ENTRY;
 357
 358         DEBUG_REQ(D_INODE, mdt_info_req(info), "Create  ("DNAME"->"DFID") "
 359                   "in "DFID,
 360                   PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
 361
 362         if (!fid_is_md_operative(rr->rr_fid1))
 363                 RETURN(-EPERM);
 364
 365         if (S_ISDIR(ma->ma_attr.la_mode) &&
 366             spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
 367                 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
 368                 struct lu_ucred *uc = mdt_ucred(info);
 369                 struct obd_export *exp = mdt_info_req(info)->rq_export;
 370
 371                 /* Only new clients can create remote dir( >= 2.4) and
 372                  * striped dir(>= 2.6), old client will return -ENOTSUPP */
 373                 if (!mdt_is_dne_client(exp))
 374                         RETURN(-ENOTSUPP);
 375
 376                 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
 377                         if (!mdt_is_striped_client(exp))
 378                                 RETURN(-ENOTSUPP);
 379
 380                         if (!mdt->mdt_enable_striped_dir)
 381                                 RETURN(-EPERM);
 382                 } else if (!mdt->mdt_enable_remote_dir) {
 383                         RETURN(-EPERM);
 384                 }
 385
 386                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 387                     uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 388                     mdt->mdt_enable_remote_dir_gid != -1)
 389                         RETURN(-EPERM);
 390         }
 391
 392         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 393
 394         parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 395         if (IS_ERR(parent))
 396                 RETURN(PTR_ERR(parent));
 397
 398         if (!mdt_object_exists(parent))
 399                 GOTO(put_parent, rc = -ENOENT);
 400
 401         lh = &info->mti_lh[MDT_LH_PARENT];
 402         mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 403         rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
 404         if (rc)
 405                 GOTO(put_parent, rc);
 406
 407         if (!mdt_object_remote(parent)) {
 408                 rc = mdt_version_get_check_save(info, parent, 0);
 409                 if (rc)
 410                         GOTO(unlock_parent, rc);
 411         }
 412
 413         /*
 414          * Check child name version during replay.
 415          * During create replay a file may exist with same name.
 416          */
 417         rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
 418                                       &info->mti_tmp_fid1, 1);
 419         if (rc == 0)
 420                 GOTO(unlock_parent, rc = -EEXIST);
 421
 422         /* -ENOENT is expected here */
 423         if (rc != -ENOENT)
 424                 GOTO(unlock_parent, rc);
 425
 426         /* save version of file name for replay, it must be ENOENT here */
 427         mdt_enoent_version_save(info, 1);
 428
 429         child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
 430         if (unlikely(IS_ERR(child)))
 431                 GOTO(unlock_parent, rc = PTR_ERR(child));
 432
 433         ma->ma_need = MA_INODE;
 434         ma->ma_valid = 0;
 435
 436         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 437                         OBD_FAIL_MDS_REINT_CREATE_WRITE);
 438
 439         /* Version of child will be updated on disk. */
 440         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 441         rc = mdt_version_get_check_save(info, child, 2);
 442         if (rc)
 443                 GOTO(put_child, rc);
 444
 445         /* Let lower layer know current lock mode. */
 446         info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
 447
 448         /*
 449          * Do not perform lookup sanity check. We know that name does
 450          * not exist.
 451          */
 452         info->mti_spec.sp_cr_lookup = 0;
 453         info->mti_spec.sp_feat = &dt_directory_features;
 454
 455         rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
 456                         mdt_object_child(child), &info->mti_spec, ma);
 457         if (rc == 0)
 458                 rc = mdt_attr_get_complex(info, child, ma);
 459
 460         if (rc < 0)
 461                 GOTO(put_child, rc);
 462
 463         /*
 464          * On DNE, we need to eliminate dependey between 'mkdir a' and
 465          * 'mkdir a/b' if b is a striped directory, to achieve this, two
 466          * things are done below:
 467          * 1. save child and slaves lock.
 468          * 2. if the child is a striped directory, relock parent so to
 469          *    compare against with COS locks to ensure parent was
 470          *    committed to disk.
 471          */
 472         if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
 473                 struct mdt_lock_handle *lhc;
 474                 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 475                 bool cos_incompat;
 476
 477                 rc = mdt_object_striped(info, child);
 478                 if (rc < 0)
 479                         GOTO(put_child, rc);
 480
 481                 cos_incompat = rc;
 482                 if (cos_incompat) {
 483                         if (!mdt_object_remote(parent)) {
 484                                 mdt_object_unlock(info, parent, lh, 1);
 485                                 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 486                                 rc = mdt_reint_object_lock(info, parent, lh,
 487                                                            MDS_INODELOCK_UPDATE,
 488                                                            true);
 489                                 if (rc)
 490                                         GOTO(put_child, rc);
 491                         }
 492                 }
 493
 494                 lhc = &info->mti_lh[MDT_LH_CHILD];
 495                 mdt_lock_handle_init(lhc);
 496                 mdt_lock_reg_init(lhc, LCK_PW);
 497                 rc = mdt_reint_striped_lock(info, child, lhc,
 498                                             MDS_INODELOCK_UPDATE, einfo,
 499                                             cos_incompat);
 500                 if (rc)
 501                         GOTO(put_child, rc);
 502
 503                 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 504         }
 505
 506         /* Return fid & attr to client. */
 507         if (ma->ma_valid & MA_INODE)
 508                 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
 509                                    mdt_object_fid(child));
 510 put_child:
 511         mdt_object_put(info->mti_env, child);
 512 unlock_parent:
 513         mdt_object_unlock(info, parent, lh, rc);
 514 put_parent:
 515         mdt_object_put(info->mti_env, parent);
 516         RETURN(rc);
 517 }
 518
 519 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
 520                         struct md_attr *ma)
 521 {
 522         struct mdt_lock_handle  *lh;
 523         int do_vbr = ma->ma_attr.la_valid &
 524                         (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
 525         __u64 lockpart = MDS_INODELOCK_UPDATE;
 526         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 527         bool cos_incompat;
 528         int rc;
 529         ENTRY;
 530
 531         rc = mdt_object_striped(info, mo);
 532         if (rc < 0)
 533                 RETURN(rc);
 534
 535         cos_incompat = rc;
 536
 537         lh = &info->mti_lh[MDT_LH_PARENT];
 538         mdt_lock_reg_init(lh, LCK_PW);
 539
 540         /* Even though the new MDT will grant PERM lock to the old
 541          * client, but the old client will almost ignore that during
 542          * So it needs to revoke both LOOKUP and PERM lock here, so
 543          * both new and old client can cancel the dcache */
 544         if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
 545                 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
 546
 547         rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
 548                                     cos_incompat);
 549         if (rc != 0)
 550                 RETURN(rc);
 551
 552         /* all attrs are packed into mti_attr in unpack_setattr */
 553         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 554                        OBD_FAIL_MDS_REINT_SETATTR_WRITE);
 555
 556         /* VBR: update version if attr changed are important for recovery */
 557         if (do_vbr) {
 558                 /* update on-disk version of changed object */
 559                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
 560                 rc = mdt_version_get_check_save(info, mo, 0);
 561                 if (rc)
 562                         GOTO(out_unlock, rc);
 563         }
 564
 565         /* Ensure constant striping during chown(). See LU-2789. */
 566         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 567                 mutex_lock(&mo->mot_lov_mutex);
 568
 569         /* all attrs are packed into mti_attr in unpack_setattr */
 570         rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
 571
 572         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 573                 mutex_unlock(&mo->mot_lov_mutex);
 574
 575         if (rc != 0)
 576                 GOTO(out_unlock, rc);
 577         mdt_dom_obj_lvb_update(info->mti_env, mo, false);
 578         EXIT;
 579 out_unlock:
 580         mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
 581         return rc;
 582 }
 583
 584 /**
 585  * Check HSM flags and add HS_DIRTY flag if relevant.
 586  *
 587  * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
 588  * and is not RELEASED.
 589  */
 590 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
 591                         struct md_attr *ma)
 592 {
 593         struct lu_ucred *uc = mdt_ucred(info);
 594         cfs_cap_t cap_saved;
 595         int rc;
 596         ENTRY;
 597
 598         /* If the file was modified, add the dirty flag */
 599         ma->ma_need = MA_HSM;
 600         rc = mdt_attr_get_complex(info, mo, ma);
 601         if (rc) {
 602                 CERROR("file attribute read error for "DFID": %d.\n",
 603                         PFID(mdt_object_fid(mo)), rc);
 604                 RETURN(rc);
 605         }
 606
 607         /* If an up2date copy exists in the backend, add dirty flag */
 608         if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
 609             && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
 610                 ma->ma_hsm.mh_flags |= HS_DIRTY;
 611
 612                 /* Bump cap so that closes from non-owner writers can
 613                  * set the HSM state to dirty. */
 614                 cap_saved = uc->uc_cap;
 615                 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
 616                 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
 617                 uc->uc_cap = cap_saved;
 618                 if (rc)
 619                         CERROR("file attribute change error for "DFID": %d\n",
 620                                 PFID(mdt_object_fid(mo)), rc);
 621         }
 622
 623         RETURN(rc);
 624 }
 625
 626 static int mdt_reint_setattr(struct mdt_thread_info *info,
 627                              struct mdt_lock_handle *lhc)
 628 {
 629         struct mdt_device *mdt = info->mti_mdt;
 630         struct md_attr *ma = &info->mti_attr;
 631         struct mdt_reint_record *rr = &info->mti_rr;
 632         struct ptlrpc_request *req = mdt_info_req(info);
 633         struct mdt_object *mo;
 634         struct mdt_body *repbody;
 635         int rc, rc2;
 636         ENTRY;
 637
 638         DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
 639                   (unsigned int)ma->ma_attr.la_valid);
 640
 641         if (info->mti_dlm_req)
 642                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 643
 644         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 645         mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
 646         if (IS_ERR(mo))
 647                 GOTO(out, rc = PTR_ERR(mo));
 648
 649         if (!mdt_object_exists(mo))
 650                 GOTO(out_put, rc = -ENOENT);
 651
 652         if (mdt_object_remote(mo))
 653                 GOTO(out_put, rc = -EREMOTE);
 654
 655         /* revoke lease lock if size is going to be changed */
 656         if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
 657                      !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
 658                      atomic_read(&mo->mot_lease_count) > 0)) {
 659                 down_read(&mo->mot_open_sem);
 660
 661                 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
 662                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 663                         mdt_lock_reg_init(lhc, LCK_CW);
 664
 665                         rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
 666                         if (rc != 0) {
 667                                 up_read(&mo->mot_open_sem);
 668                                 GOTO(out_put, rc);
 669                         }
 670
 671                         /* revoke lease lock */
 672                         mdt_object_unlock(info, mo, lhc, 1);
 673                 }
 674                 up_read(&mo->mot_open_sem);
 675         }
 676
 677         if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
 678                 /* Check write access for the O_TRUNC case */
 679                 if (mdt_write_read(mo) < 0)
 680                         GOTO(out_put, rc = -ETXTBSY);
 681
 682                 /* LU-10286: compatibility check for FLR.
 683                  * Please check the comment in mdt_finish_open() for details */
 684                 if (!exp_connect_flr(info->mti_exp)) {
 685                         rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
 686                         if (rc < 0 && rc != -ENODATA)
 687                                 GOTO(out_put, rc);
 688
 689                         if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm))
 690                                 GOTO(out_put, rc = -EOPNOTSUPP);
 691                 }
 692
 693                 /* For truncate, the file size sent from client
 694                  * is believable, but the blocks are incorrect,
 695                  * which makes the block size in LSOM attribute
 696                  * inconsisent with the real block size.
 697                  */
 698                 rc = mdt_lsom_update(info, mo, true);
 699                 if (rc)
 700                         GOTO(out_put, rc);
 701         }
 702
 703         if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
 704                 if (ma->ma_valid & MA_LOV)
 705                         GOTO(out_put, rc = -EPROTO);
 706
 707                 rc = mdt_attr_set(info, mo, ma);
 708                 if (rc)
 709                         GOTO(out_put, rc);
 710         } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
 711                    (ma->ma_valid & MA_INODE)) {
 712                 struct lu_buf *buf = &info->mti_buf;
 713                 struct lu_ucred *uc = mdt_ucred(info);
 714                 struct mdt_lock_handle *lh;
 715
 716                 /* reject if either remote or striped dir is disabled */
 717                 if (ma->ma_valid & MA_LMV) {
 718                         if (!mdt->mdt_enable_remote_dir ||
 719                             !mdt->mdt_enable_striped_dir)
 720                                 GOTO(out_put, rc = -EPERM);
 721
 722                         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 723                             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 724                             mdt->mdt_enable_remote_dir_gid != -1)
 725                                 GOTO(out_put, rc = -EPERM);
 726                 }
 727
 728                 if (ma->ma_attr.la_valid != 0)
 729                         GOTO(out_put, rc = -EPROTO);
 730
 731                 lh = &info->mti_lh[MDT_LH_PARENT];
 732                 mdt_lock_reg_init(lh, LCK_PW);
 733
 734                 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
 735                 if (rc != 0)
 736                         GOTO(out_put, rc);
 737
 738                 if (ma->ma_valid & MA_LOV) {
 739                         buf->lb_buf = ma->ma_lmm;
 740                         buf->lb_len = ma->ma_lmm_size;
 741                 } else {
 742                         buf->lb_buf = ma->ma_lmv;
 743                         buf->lb_len = ma->ma_lmv_size;
 744                 }
 745                 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
 746                                   (ma->ma_valid & MA_LOV) ?
 747                                         XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
 748                                   0);
 749
 750                 mdt_object_unlock(info, mo, lh, rc);
 751                 if (rc)
 752                         GOTO(out_put, rc);
 753         } else {
 754                 GOTO(out_put, rc = -EPROTO);
 755         }
 756
 757         /* If file data is modified, add the dirty flag */
 758         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
 759                 rc = mdt_add_dirty_flag(info, mo, ma);
 760
 761         ma->ma_need = MA_INODE;
 762         ma->ma_valid = 0;
 763         rc = mdt_attr_get_complex(info, mo, ma);
 764         if (rc != 0)
 765                 GOTO(out_put, rc);
 766
 767         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
 768
 769         EXIT;
 770 out_put:
 771         mdt_object_put(info->mti_env, mo);
 772 out:
 773         if (rc == 0)
 774                 mdt_counter_incr(req, LPROC_MDT_SETATTR);
 775
 776         mdt_client_compatibility(info);
 777         rc2 = mdt_fix_reply(info);
 778         if (rc == 0)
 779                 rc = rc2;
 780         return rc;
 781 }
 782
 783 static int mdt_reint_create(struct mdt_thread_info *info,
 784                             struct mdt_lock_handle *lhc)
 785 {
 786         struct ptlrpc_request   *req = mdt_info_req(info);
 787         int                     rc;
 788         ENTRY;
 789
 790         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 791                 RETURN(err_serious(-ESTALE));
 792
 793         if (info->mti_dlm_req)
 794                 ldlm_request_cancel(mdt_info_req(info),
 795                                     info->mti_dlm_req, 0, LATF_SKIP);
 796
 797         if (!lu_name_is_valid(&info->mti_rr.rr_name))
 798                 RETURN(-EPROTO);
 799
 800         switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
 801         case S_IFDIR:
 802                 mdt_counter_incr(req, LPROC_MDT_MKDIR);
 803                 break;
 804         case S_IFREG:
 805         case S_IFLNK:
 806         case S_IFCHR:
 807         case S_IFBLK:
 808         case S_IFIFO:
 809         case S_IFSOCK:
 810                 /* Special file should stay on the same node as parent. */
 811                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
 812                 break;
 813         default:
 814                 CERROR("%s: Unsupported mode %o\n",
 815                        mdt_obd_name(info->mti_mdt),
 816                        info->mti_attr.ma_attr.la_mode);
 817                 RETURN(err_serious(-EOPNOTSUPP));
 818         }
 819
 820         rc = mdt_create(info);
 821         RETURN(rc);
 822 }
 823
 824 /*
 825  * VBR: save parent version in reply and child version getting by its name.
 826  * Version of child is getting and checking during its lookup. If
 827  */
 828 static int mdt_reint_unlink(struct mdt_thread_info *info,
 829                             struct mdt_lock_handle *lhc)
 830 {
 831         struct mdt_reint_record *rr = &info->mti_rr;
 832         struct ptlrpc_request *req = mdt_info_req(info);
 833         struct md_attr *ma = &info->mti_attr;
 834         struct lu_fid *child_fid = &info->mti_tmp_fid1;
 835         struct mdt_object *mp;
 836         struct mdt_object *mc;
 837         struct mdt_lock_handle *parent_lh;
 838         struct mdt_lock_handle *child_lh;
 839         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 840         __u64 lock_ibits;
 841         bool cos_incompat = false, discard = false;
 842         int no_name = 0;
 843         int rc;
 844
 845         ENTRY;
 846
 847         DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
 848                   PNAME(&rr->rr_name));
 849
 850         if (info->mti_dlm_req)
 851                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 852
 853         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
 854                 RETURN(err_serious(-ENOENT));
 855
 856         if (!fid_is_md_operative(rr->rr_fid1))
 857                 RETURN(-EPERM);
 858
 859         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 860         if (IS_ERR(mp))
 861                 RETURN(PTR_ERR(mp));
 862
 863         if (mdt_object_remote(mp)) {
 864                 cos_incompat = true;
 865         } else {
 866                 rc = mdt_version_get_check_save(info, mp, 0);
 867                 if (rc)
 868                         GOTO(put_parent, rc);
 869         }
 870
 871 relock:
 872         parent_lh = &info->mti_lh[MDT_LH_PARENT];
 873         mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
 874         rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
 875                                    cos_incompat);
 876         if (rc != 0)
 877                 GOTO(put_parent, rc);
 878
 879         /* lookup child object along with version checking */
 880         fid_zero(child_fid);
 881         rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
 882         if (rc != 0) {
 883                 /* Name might not be able to find during resend of
 884                  * remote unlink, considering following case.
 885                  * dir_A is a remote directory, the name entry of
 886                  * dir_A is on MDT0, the directory is on MDT1,
 887                  *
 888                  * 1. client sends unlink req to MDT1.
 889                  * 2. MDT1 sends name delete update to MDT0.
 890                  * 3. name entry is being deleted in MDT0 synchronously.
 891                  * 4. MDT1 is restarted.
 892                  * 5. client resends unlink req to MDT1. So it can not
 893                  *    find the name entry on MDT0 anymore.
 894                  * In this case, MDT1 only needs to destory the local
 895                  * directory.
 896                  * */
 897                 if (mdt_object_remote(mp) && rc == -ENOENT &&
 898                     !fid_is_zero(rr->rr_fid2) &&
 899                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 900                         no_name = 1;
 901                         *child_fid = *rr->rr_fid2;
 902                  } else {
 903                         GOTO(unlock_parent, rc);
 904                  }
 905         }
 906
 907         if (!fid_is_md_operative(child_fid))
 908                 GOTO(unlock_parent, rc = -EPERM);
 909
 910         /* We will lock the child regardless it is local or remote. No harm. */
 911         mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
 912         if (IS_ERR(mc))
 913                 GOTO(unlock_parent, rc = PTR_ERR(mc));
 914
 915         if (!cos_incompat) {
 916                 rc = mdt_object_striped(info, mc);
 917                 if (rc < 0)
 918                         GOTO(unlock_parent, rc);
 919
 920                 cos_incompat = rc;
 921                 if (cos_incompat) {
 922                         mdt_object_put(info->mti_env, mc);
 923                         mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
 924                         goto relock;
 925                 }
 926         }
 927
 928         child_lh = &info->mti_lh[MDT_LH_CHILD];
 929         mdt_lock_reg_init(child_lh, LCK_EX);
 930         if (info->mti_spec.sp_rm_entry) {
 931                 struct lu_ucred *uc  = mdt_ucred(info);
 932
 933                 if (!mdt_is_dne_client(req->rq_export))
 934                         /* Return -ENOTSUPP for old client */
 935                         GOTO(put_child, rc = -ENOTSUPP);
 936
 937                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
 938                         GOTO(put_child, rc = -EPERM);
 939
 940                 ma->ma_need = MA_INODE;
 941                 ma->ma_valid = 0;
 942                 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
 943                                 NULL, &rr->rr_name, ma, no_name);
 944                 GOTO(put_child, rc);
 945         }
 946
 947         if (mdt_object_remote(mc)) {
 948                 struct mdt_body  *repbody;
 949
 950                 if (!fid_is_zero(rr->rr_fid2)) {
 951                         CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
 952                                mdt_obd_name(info->mti_mdt),
 953                                PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 954                         GOTO(put_child, rc = -ENOENT);
 955                 }
 956                 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
 957                        mdt_obd_name(info->mti_mdt),
 958                        PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 959
 960                 if (!mdt_is_dne_client(req->rq_export))
 961                         /* Return -ENOTSUPP for old client */
 962                         GOTO(put_child, rc = -ENOTSUPP);
 963
 964                 /* Revoke the LOOKUP lock of the remote object granted by
 965                  * this MDT. Since the unlink will happen on another MDT,
 966                  * it will release the LOOKUP lock right away. Then What
 967                  * would happen if another client try to grab the LOOKUP
 968                  * lock at the same time with unlink XXX */
 969                 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
 970                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 971                 LASSERT(repbody != NULL);
 972                 repbody->mbo_fid1 = *mdt_object_fid(mc);
 973                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 974                 GOTO(unlock_child, rc = -EREMOTE);
 975         }
 976         /* We used to acquire MDS_INODELOCK_FULL here but we can't do
 977          * this now because a running HSM restore on the child (unlink
 978          * victim) will hold the layout lock. See LU-4002. */
 979         lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
 980         if (mdt_object_remote(mp)) {
 981                 /* Enqueue lookup lock from parent MDT */
 982                 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
 983                                             &child_lh->mlh_rreg_lh,
 984                                             child_lh->mlh_rreg_mode,
 985                                             MDS_INODELOCK_LOOKUP, false);
 986                 if (rc != ELDLM_OK)
 987                         GOTO(put_child, rc);
 988
 989                 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
 990         }
 991
 992         rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
 993                                     cos_incompat);
 994         if (rc != 0)
 995                 GOTO(put_child, rc);
 996
 997         /*
 998          * Now we can only make sure we need MA_INODE, in mdd layer, will check
 999          * whether need MA_LOV and MA_COOKIE.
1000          */
1001         ma->ma_need = MA_INODE;
1002         ma->ma_valid = 0;
1003
1004         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1005                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1006         /* save version when object is locked */
1007         mdt_version_get_save(info, mc, 1);
1008
1009         mutex_lock(&mc->mot_lov_mutex);
1010
1011         rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1012                         mdt_object_child(mc), &rr->rr_name, ma, no_name);
1013
1014         mutex_unlock(&mc->mot_lov_mutex);
1015         if (rc != 0)
1016                 GOTO(unlock_child, rc);
1017
1018         if (!lu_object_is_dying(&mc->mot_header)) {
1019                 rc = mdt_attr_get_complex(info, mc, ma);
1020                 if (rc)
1021                         GOTO(out_stat, rc);
1022         } else {
1023                 discard = mdt_dom_check_for_discard(info, mc);
1024         }
1025         mdt_handle_last_unlink(info, mc, ma);
1026
1027 out_stat:
1028         if (ma->ma_valid & MA_INODE) {
1029                 switch (ma->ma_attr.la_mode & S_IFMT) {
1030                 case S_IFDIR:
1031                         mdt_counter_incr(req, LPROC_MDT_RMDIR);
1032                         break;
1033                 case S_IFREG:
1034                 case S_IFLNK:
1035                 case S_IFCHR:
1036                 case S_IFBLK:
1037                 case S_IFIFO:
1038                 case S_IFSOCK:
1039                         mdt_counter_incr(req, LPROC_MDT_UNLINK);
1040                         break;
1041                 default:
1042                         LASSERTF(0, "bad file type %o unlinking\n",
1043                                 ma->ma_attr.la_mode);
1044                 }
1045         }
1046
1047         EXIT;
1048
1049 unlock_child:
1050         mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1051 put_child:
1052         mdt_object_put(info->mti_env, mc);
1053 unlock_parent:
1054         mdt_object_unlock(info, mp, parent_lh, rc);
1055 put_parent:
1056         mdt_object_put(info->mti_env, mp);
1057
1058         /* discard is just a PW DOM lock to drop the data on a client
1059          * no need to keep objects being get and locked, do that after all.
1060          */
1061         if (discard)
1062                 mdt_dom_discard_data(info, child_fid);
1063
1064         return rc;
1065 }
1066
1067 /*
1068  * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1069  * name.
1070  */
1071 static int mdt_reint_link(struct mdt_thread_info *info,
1072                           struct mdt_lock_handle *lhc)
1073 {
1074         struct mdt_reint_record *rr = &info->mti_rr;
1075         struct ptlrpc_request   *req = mdt_info_req(info);
1076         struct md_attr          *ma = &info->mti_attr;
1077         struct mdt_object       *ms;
1078         struct mdt_object       *mp;
1079         struct mdt_lock_handle  *lhs;
1080         struct mdt_lock_handle  *lhp;
1081         bool cos_incompat;
1082         int rc;
1083         ENTRY;
1084
1085         DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1086                   PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1087
1088         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1089                 RETURN(err_serious(-ENOENT));
1090
1091         if (info->mti_dlm_req)
1092                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1093
1094         /* Invalid case so return error immediately instead of
1095          * processing it */
1096         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1097                 RETURN(-EPERM);
1098
1099         if (!fid_is_md_operative(rr->rr_fid1) ||
1100             !fid_is_md_operative(rr->rr_fid2))
1101                 RETURN(-EPERM);
1102
1103         /* step 1: find target parent dir */
1104         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1105         if (IS_ERR(mp))
1106                 RETURN(PTR_ERR(mp));
1107
1108         rc = mdt_version_get_check_save(info, mp, 0);
1109         if (rc)
1110                 GOTO(put_parent, rc);
1111
1112         /* step 2: find source */
1113         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1114         if (IS_ERR(ms))
1115                 GOTO(put_parent, rc = PTR_ERR(ms));
1116
1117         if (!mdt_object_exists(ms)) {
1118                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1119                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1120                 GOTO(put_source, rc = -ENOENT);
1121         }
1122
1123         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1124
1125         lhp = &info->mti_lh[MDT_LH_PARENT];
1126         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1127         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1128                                    cos_incompat);
1129         if (rc != 0)
1130                 GOTO(put_source, rc);
1131
1132         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1133
1134         lhs = &info->mti_lh[MDT_LH_CHILD];
1135         mdt_lock_reg_init(lhs, LCK_EX);
1136         rc = mdt_reint_object_lock(info, ms, lhs,
1137                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1138                                    cos_incompat);
1139         if (rc != 0)
1140                 GOTO(unlock_parent, rc);
1141
1142         /* step 3: link it */
1143         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1144                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1145
1146         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1147         rc = mdt_version_get_check_save(info, ms, 1);
1148         if (rc)
1149                 GOTO(unlock_source, rc);
1150
1151         /** check target version by name during replay */
1152         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1153                                       &info->mti_tmp_fid1, 2);
1154         if (rc != 0 && rc != -ENOENT)
1155                 GOTO(unlock_source, rc);
1156         /* save version of file name for replay, it must be ENOENT here */
1157         if (!req_is_replay(mdt_info_req(info))) {
1158                 if (rc != -ENOENT) {
1159                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1160                                PNAME(&rr->rr_name));
1161                         GOTO(unlock_source, rc = -EEXIST);
1162                 }
1163                 info->mti_ver[2] = ENOENT_VERSION;
1164                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1165         }
1166
1167         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1168                       mdt_object_child(ms), &rr->rr_name, ma);
1169
1170         if (rc == 0)
1171                 mdt_counter_incr(req, LPROC_MDT_LINK);
1172
1173         EXIT;
1174 unlock_source:
1175         mdt_object_unlock(info, ms, lhs, rc);
1176 unlock_parent:
1177         mdt_object_unlock(info, mp, lhp, rc);
1178 put_source:
1179         mdt_object_put(info->mti_env, ms);
1180 put_parent:
1181         mdt_object_put(info->mti_env, mp);
1182         return rc;
1183 }
1184 /**
1185  * lock the part of the directory according to the hash of the name
1186  * (lh->mlh_pdo_hash) in parallel directory lock.
1187  */
1188 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1189                               struct mdt_lock_handle *lh,
1190                               struct mdt_object *obj, __u64 ibits,
1191                               bool cos_incompat)
1192 {
1193         struct ldlm_res_id *res = &info->mti_res_id;
1194         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1195         union ldlm_policy_data *policy = &info->mti_policy;
1196         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1197         int rc;
1198
1199         /*
1200          * Finish res_id initializing by name hash marking part of
1201          * directory which is taking modification.
1202          */
1203         LASSERT(lh->mlh_pdo_hash != 0);
1204         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1205         memset(policy, 0, sizeof(*policy));
1206         policy->l_inodebits.bits = ibits;
1207         if (cos_incompat &&
1208             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1209                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1210         /*
1211          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1212          * going to be sent to client. If it is - mdt_intent_policy() path will
1213          * fix it up and turn FL_LOCAL flag off.
1214          */
1215         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1216                           policy, res, dlmflags,
1217                           &info->mti_exp->exp_handle.h_cookie);
1218         return rc;
1219 }
1220
1221 /**
1222  * Get BFL lock for rename or migrate process.
1223  **/
1224 static int mdt_rename_lock(struct mdt_thread_info *info,
1225                            struct lustre_handle *lh)
1226 {
1227         int     rc;
1228         ENTRY;
1229
1230         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1231                 struct lu_fid *fid = &info->mti_tmp_fid1;
1232                 struct mdt_object *obj;
1233
1234                 /* XXX, right now, it has to use object API to
1235                  * enqueue lock cross MDT, so it will enqueue
1236                  * rename lock(with LUSTRE_BFL_FID) by root object */
1237                 lu_root_fid(fid);
1238                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1239                 if (IS_ERR(obj))
1240                         RETURN(PTR_ERR(obj));
1241
1242                 rc = mdt_remote_object_lock(info, obj,
1243                                             &LUSTRE_BFL_FID, lh,
1244                                             LCK_EX,
1245                                             MDS_INODELOCK_UPDATE, false);
1246                 mdt_object_put(info->mti_env, obj);
1247         } else {
1248                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1249                 union ldlm_policy_data *policy = &info->mti_policy;
1250                 struct ldlm_res_id *res_id = &info->mti_res_id;
1251                 __u64 flags = 0;
1252
1253                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1254                 memset(policy, 0, sizeof *policy);
1255                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1256                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1257                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1258                                             LDLM_IBITS, policy, LCK_EX, &flags,
1259                                             ldlm_blocking_ast,
1260                                             ldlm_completion_ast, NULL, NULL, 0,
1261                                             LVB_T_NONE,
1262                                             &info->mti_exp->exp_handle.h_cookie,
1263                                             lh);
1264                 RETURN(rc);
1265         }
1266         RETURN(rc);
1267 }
1268
1269 static void mdt_rename_unlock(struct lustre_handle *lh)
1270 {
1271         ENTRY;
1272         LASSERT(lustre_handle_is_used(lh));
1273         /* Cancel the single rename lock right away */
1274         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1275         EXIT;
1276 }
1277
1278 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1279                                                 const struct lu_fid *fid,
1280                                                 int idx)
1281 {
1282         struct mdt_object *dir;
1283         int rc;
1284
1285         ENTRY;
1286
1287         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1288         if (IS_ERR(dir))
1289                 RETURN(dir);
1290
1291         /* check early, the real version will be saved after locking */
1292         rc = mdt_version_get_check(info, dir, idx);
1293         if (rc)
1294                 GOTO(out_put, rc);
1295
1296         if (!mdt_object_exists(dir))
1297                 GOTO(out_put, rc = -ENOENT);
1298
1299         if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1300                 GOTO(out_put, rc = -ENOTDIR);
1301
1302         RETURN(dir);
1303 out_put:
1304         mdt_object_put(info->mti_env, dir);
1305         return ERR_PTR(rc);
1306 }
1307
1308 /*
1309  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1310  * herein we don't really check it, just do revoke.
1311  */
1312 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1313                                   struct mdt_object *pobj,
1314                                   struct mdt_object *obj)
1315 {
1316         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1317         int rc;
1318
1319         mdt_lock_handle_init(lh);
1320         mdt_lock_reg_init(lh, LCK_EX);
1321
1322         if (mdt_object_remote(pobj)) {
1323                 /* don't bother to check if pobj and obj are on the same MDT. */
1324                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1325                                             &lh->mlh_rreg_lh, LCK_EX,
1326                                             MDS_INODELOCK_LOOKUP, false);
1327         } else if (mdt_object_remote(obj)) {
1328                 struct ldlm_res_id *res = &info->mti_res_id;
1329                 union ldlm_policy_data *policy = &info->mti_policy;
1330                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1331                                  LDLM_FL_COS_INCOMPAT;
1332
1333                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1334                 memset(policy, 0, sizeof(*policy));
1335                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1336                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1337                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1338                                   dlmflags, NULL);
1339         } else {
1340                 /* do nothing if both are local */
1341                 return 0;
1342         }
1343
1344         if (rc != ELDLM_OK)
1345                 return rc;
1346
1347         /*
1348          * TODO, currently we don't save this lock because there is no place to
1349          * hold this lock handle, but to avoid race we need to save this lock.
1350          */
1351         mdt_object_unlock(info, NULL, lh, 1);
1352
1353         return 0;
1354 }
1355
1356 /*
1357  * operation may takes locks of linkea, or directory stripes, group them in
1358  * different list.
1359  */
1360 struct mdt_sub_lock {
1361         struct mdt_object      *msl_obj;
1362         struct mdt_lock_handle  msl_lh;
1363         struct list_head        msl_linkage;
1364 };
1365
1366 static void mdt_unlock_list(struct mdt_thread_info *info,
1367                             struct list_head *list, int decref)
1368 {
1369         struct mdt_sub_lock *msl;
1370         struct mdt_sub_lock *tmp;
1371
1372         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1373                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1374                 list_del(&msl->msl_linkage);
1375                 OBD_FREE_PTR(msl);
1376         }
1377 }
1378
1379 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1380                                              struct mdt_object *obj,
1381                                              struct mdt_lock_handle *lh,
1382                                              struct ldlm_enqueue_info *einfo,
1383                                              struct list_head *slave_locks,
1384                                              int decref)
1385 {
1386         if (mdt_object_remote(obj)) {
1387                 mdt_unlock_list(info, slave_locks, decref);
1388                 mdt_object_unlock(info, obj, lh, decref);
1389         } else {
1390                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1391         }
1392 }
1393
1394 /*
1395  * lock parents of links, and also check whether total locks don't exceed
1396  * RS_MAX_LOCKS.
1397  *
1398  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1399  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1400  * \retval      -ev negative errno upon error
1401  */
1402 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1403                                  struct mdt_object *pobj,
1404                                  const struct md_attr *ma,
1405                                  struct mdt_object *obj,
1406                                  struct mdt_lock_handle *lhp,
1407                                  struct ldlm_enqueue_info *peinfo,
1408                                  struct list_head *parent_slave_locks,
1409                                  struct list_head *link_locks)
1410 {
1411         struct mdt_device *mdt = info->mti_mdt;
1412         struct lu_buf *buf = &info->mti_big_buf;
1413         struct lu_name *lname = &info->mti_name;
1414         struct linkea_data ldata = { NULL };
1415         bool blocked = false;
1416         int local_lnkp_cnt = 0;
1417         int rc;
1418
1419         ENTRY;
1420
1421         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1422                 RETURN(0);
1423
1424         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1425         if (buf->lb_buf == NULL)
1426                 RETURN(-ENOMEM);
1427
1428         ldata.ld_buf = buf;
1429         rc = mdt_links_read(info, obj, &ldata);
1430         if (rc) {
1431                 if (rc == -ENOENT || rc == -ENODATA)
1432                         rc = 0;
1433                 RETURN(rc);
1434         }
1435
1436         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1437              linkea_next_entry(&ldata)) {
1438                 struct mdt_object *lnkp;
1439                 struct mdt_sub_lock *msl;
1440                 struct lu_fid fid;
1441                 __u64 ibits;
1442
1443                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1444                                     &fid);
1445
1446                 /* check if it's also linked to parent */
1447                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1448                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1449                                PFID(&fid), PNAME(lname));
1450                         /* in case link is remote object, revoke LOOKUP lock */
1451                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1452                         continue;
1453                 }
1454
1455                 lnkp = NULL;
1456
1457                 /* check if it's linked to a stripe of parent */
1458                 if (ma->ma_valid & MA_LMV) {
1459                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1460                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1461                         int j = 0;
1462
1463                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1464                                 fid_le_to_cpu(stripe_fid,
1465                                               &lmv->lmv_stripe_fids[j]);
1466                                 if (lu_fid_eq(stripe_fid, &fid)) {
1467                                         CDEBUG(D_INFO, "skip stripe "DFID
1468                                                ", reovke "DNAME"\n",
1469                                                PFID(&fid), PNAME(lname));
1470                                         lnkp = mdt_object_find(info->mti_env,
1471                                                                mdt, &fid);
1472                                         if (IS_ERR(lnkp))
1473                                                 GOTO(out, rc = PTR_ERR(lnkp));
1474                                         break;
1475                                 }
1476                         }
1477
1478                         if (lnkp) {
1479                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1480                                                                    obj);
1481                                 mdt_object_put(info->mti_env, lnkp);
1482                                 continue;
1483                         }
1484                 }
1485
1486                 /* Check if it's already locked */
1487                 list_for_each_entry(msl, link_locks, msl_linkage) {
1488                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1489                                 CDEBUG(D_INFO,
1490                                        DFID" was locked, revoke "DNAME"\n",
1491                                        PFID(&fid), PNAME(lname));
1492                                 lnkp = msl->msl_obj;
1493                                 break;
1494                         }
1495                 }
1496
1497                 if (lnkp) {
1498                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1499                         continue;
1500                 }
1501
1502                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1503                        PFID(&fid), PNAME(lname));
1504
1505                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1506                 if (IS_ERR(lnkp)) {
1507                         CWARN("%s: cannot find obj "DFID": %ld\n",
1508                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1509                         continue;
1510                 }
1511
1512                 if (!mdt_object_exists(lnkp)) {
1513                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1514                               PFID(&fid), PNAME(lname));
1515                         mdt_object_put(info->mti_env, lnkp);
1516                         continue;
1517                 }
1518
1519                 if (!mdt_object_remote(lnkp))
1520                         local_lnkp_cnt++;
1521
1522                 OBD_ALLOC_PTR(msl);
1523                 if (msl == NULL)
1524                         GOTO(out, rc = -ENOMEM);
1525
1526                 /*
1527                  * we can't follow parent-child lock order like other MD
1528                  * operations, use lock_try here to avoid deadlock, if the lock
1529                  * cannot be taken, drop all locks taken, revoke the blocked
1530                  * one, and continue processing the remaining entries, and in
1531                  * the end of the loop restart from beginning.
1532                  */
1533                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1534                 ibits = 0;
1535                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1536                                          MDS_INODELOCK_UPDATE, true);
1537                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1538
1539                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1540                                PFID(&fid), PNAME(lname));
1541
1542                         mdt_unlock_list(info, link_locks, 1);
1543                         /* also unlock parent locks to avoid deadlock */
1544                         if (!blocked)
1545                                 mdt_migrate_object_unlock(info, pobj, lhp,
1546                                                           peinfo,
1547                                                           parent_slave_locks,
1548                                                           1);
1549
1550                         blocked = true;
1551
1552                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1553                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1554                                              MDS_INODELOCK_UPDATE);
1555                         if (rc) {
1556                                 mdt_object_put(info->mti_env, lnkp);
1557                                 OBD_FREE_PTR(msl);
1558                                 GOTO(out, rc);
1559                         }
1560
1561                         if (mdt_object_remote(lnkp)) {
1562                                 struct ldlm_lock *lock;
1563
1564                                 /*
1565                                  * for remote object, set lock cb_atomic,
1566                                  * so lock can be released in blocking_ast()
1567                                  * immediately, then the next lock_try will
1568                                  * have better chance of success.
1569                                  */
1570                                 lock = ldlm_handle2lock(
1571                                                 &msl->msl_lh.mlh_rreg_lh);
1572                                 LASSERT(lock != NULL);
1573                                 lock_res_and_lock(lock);
1574                                 ldlm_set_atomic_cb(lock);
1575                                 unlock_res_and_lock(lock);
1576                                 LDLM_LOCK_PUT(lock);
1577                         }
1578
1579                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1580                         OBD_FREE_PTR(msl);
1581                         continue;
1582                 }
1583
1584                 INIT_LIST_HEAD(&msl->msl_linkage);
1585                 msl->msl_obj = lnkp;
1586                 list_add_tail(&msl->msl_linkage, link_locks);
1587
1588                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1589         }
1590
1591         if (blocked)
1592                 GOTO(out, rc = -EBUSY);
1593
1594         EXIT;
1595 out:
1596         if (rc)
1597                 mdt_unlock_list(info, link_locks, rc);
1598         else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1599                 /*
1600                  * parent may have 3 local objects: master object and 2 stripes
1601                  * (if it's being migrated too); source may have 2 local
1602                  * objects: master and 1 stripe; target has 1 local object.
1603                  */
1604                 rc = 1;
1605         return rc;
1606 }
1607
1608 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1609                                   struct mdt_object *obj,
1610                                   const struct md_attr *ma,
1611                                   struct list_head *slave_locks)
1612 {
1613         struct mdt_device *mdt = info->mti_mdt;
1614         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1615         struct lu_fid *fid = &info->mti_tmp_fid1;
1616         struct mdt_object *slave;
1617         struct mdt_sub_lock *msl;
1618         int i;
1619         int rc;
1620
1621         ENTRY;
1622
1623         LASSERT(mdt_object_remote(obj));
1624         LASSERT(ma->ma_valid & MA_LMV);
1625         LASSERT(lmv);
1626
1627         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1628                 RETURN(-EINVAL);
1629
1630         if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1631                 RETURN(0);
1632
1633         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1634                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1635
1636                 slave = mdt_object_find(info->mti_env, mdt, fid);
1637                 if (IS_ERR(slave))
1638                         GOTO(out, rc = PTR_ERR(slave));
1639
1640                 OBD_ALLOC_PTR(msl);
1641                 if (!msl) {
1642                         mdt_object_put(info->mti_env, slave);
1643                         GOTO(out, rc = -ENOMEM);
1644                 }
1645
1646                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1647                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1648                                            MDS_INODELOCK_UPDATE, true);
1649                 if (rc) {
1650                         OBD_FREE_PTR(msl);
1651                         mdt_object_put(info->mti_env, slave);
1652                         GOTO(out, rc);
1653                 }
1654
1655                 INIT_LIST_HEAD(&msl->msl_linkage);
1656                 msl->msl_obj = slave;
1657                 list_add_tail(&msl->msl_linkage, slave_locks);
1658         }
1659         EXIT;
1660
1661 out:
1662         if (rc)
1663                 mdt_unlock_list(info, slave_locks, rc);
1664         return rc;
1665 }
1666
1667 /* lock parent and its stripes */
1668 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1669                                    struct mdt_object *obj,
1670                                    const struct md_attr *ma,
1671                                    struct mdt_lock_handle *lh,
1672                                    struct ldlm_enqueue_info *einfo,
1673                                    struct list_head *slave_locks)
1674 {
1675         int rc;
1676
1677         if (mdt_object_remote(obj)) {
1678                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1679                                             &lh->mlh_rreg_lh, LCK_PW,
1680                                             MDS_INODELOCK_UPDATE, false);
1681                 if (rc != ELDLM_OK)
1682                         return rc;
1683
1684                 /*
1685                  * if obj is remote and striped, lock its stripes explicitly
1686                  * because it's not striped in LOD layer on this MDT.
1687                  */
1688                 if (ma->ma_valid & MA_LMV) {
1689                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1690                         if (rc)
1691                                 mdt_object_unlock(info, obj, lh, rc);
1692                 }
1693         } else {
1694                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1695                                             einfo, true);
1696         }
1697
1698         return rc;
1699 }
1700
1701 /*
1702  * in migration, object may be remote, and we need take full lock of it and its
1703  * stripes if it's directory, besides, object may be a remote object on its
1704  * parent, revoke its LOOKUP lock on where its parent is located.
1705  */
1706 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1707                                    struct mdt_object *pobj,
1708                                    struct mdt_object *obj,
1709                                    struct mdt_lock_handle *lh,
1710                                    struct ldlm_enqueue_info *einfo,
1711                                    struct list_head *slave_locks)
1712 {
1713         int rc;
1714
1715         if (mdt_object_remote(obj)) {
1716                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1717                 if (rc)
1718                         return rc;
1719
1720                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1721                                             &lh->mlh_rreg_lh, LCK_EX,
1722                                             MDS_INODELOCK_FULL, false);
1723                 if (rc != ELDLM_OK)
1724                         return rc;
1725
1726                 /*
1727                  * if obj is remote and striped, lock its stripes explicitly
1728                  * because it's not striped in LOD layer on this MDT.
1729                  */
1730                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1731                         struct md_attr *ma = &info->mti_attr;
1732
1733                         ma->ma_lmv = info->mti_big_lmm;
1734                         ma->ma_lmv_size = info->mti_big_lmmsize;
1735                         ma->ma_valid = 0;
1736                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1737                         if (rc) {
1738                                 mdt_object_unlock(info, obj, lh, rc);
1739                                 return rc;
1740                         }
1741
1742                         if (ma->ma_valid & MA_LMV) {
1743                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1744                                                             slave_locks);
1745                                 if (rc)
1746                                         mdt_object_unlock(info, obj, lh, rc);
1747                         }
1748                 }
1749         } else {
1750                 if (mdt_object_remote(pobj)) {
1751                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1752                         if (rc)
1753                                 return rc;
1754                 }
1755
1756                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1757                                             einfo, true);
1758         }
1759
1760         return rc;
1761 }
1762
1763 /*
1764  * lookup source by name, if parent is striped directory, we need to find the
1765  * corresponding stripe where source is located, and then lookup there.
1766  *
1767  * besides, if parent is migrating too, and file is already in target stripe,
1768  * this should be a redo of 'lfs migrate' on client side.
1769  */
1770 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1771                               struct mdt_object *pobj,
1772                               const struct md_attr *ma,
1773                               const struct lu_name *lname,
1774                               struct mdt_object **spobj,
1775                               struct mdt_object **sobj)
1776 {
1777         const struct lu_env *env = info->mti_env;
1778         struct lu_fid *fid = &info->mti_tmp_fid1;
1779         struct mdt_object *stripe;
1780         int rc;
1781
1782         if (ma->ma_valid & MA_LMV) {
1783                 /* if parent is striped, lookup on corresponding stripe */
1784                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1785                 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1786                 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1787                 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1788                                     LMV_HASH_FLAG_MIGRATION;
1789
1790                 if (is_migrating) {
1791                         hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1792                         stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1793                 }
1794
1795                 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1796                                               lname->ln_name,
1797                                               lname->ln_namelen);
1798                 if (rc < 0)
1799                         return rc;
1800
1801                 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1802                         rc += le32_to_cpu(lmv->lmv_migrate_offset);
1803
1804                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1805
1806                 stripe = mdt_object_find(env, info->mti_mdt, fid);
1807                 if (IS_ERR(stripe))
1808                         return PTR_ERR(stripe);
1809
1810                 fid_zero(fid);
1811                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1812                                 &info->mti_spec);
1813                 if (rc == -ENOENT && is_migrating) {
1814                         /*
1815                          * if parent is migrating, and lookup child failed on
1816                          * source stripe, lookup again on target stripe, if it
1817                          * exists, it means previous migration was interrupted,
1818                          * and current file was migrated already.
1819                          */
1820                         mdt_object_put(env, stripe);
1821
1822                         hash_type = le32_to_cpu(lmv->lmv_hash_type);
1823                         stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1824
1825                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1826                                                       lname->ln_name,
1827                                                       lname->ln_namelen);
1828                         if (rc < 0)
1829                                 return rc;
1830
1831                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1832
1833                         stripe = mdt_object_find(env, info->mti_mdt, fid);
1834                         if (IS_ERR(stripe))
1835                                 return PTR_ERR(stripe);
1836
1837                         fid_zero(fid);
1838                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1839                                         fid, &info->mti_spec);
1840                         mdt_object_put(env, stripe);
1841                         return rc ?: -EALREADY;
1842                 } else if (rc) {
1843                         mdt_object_put(env, stripe);
1844                         return rc;
1845                 }
1846         } else {
1847                 fid_zero(fid);
1848                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1849                                 &info->mti_spec);
1850                 if (rc)
1851                         return rc;
1852
1853                 stripe = pobj;
1854                 mdt_object_get(env, stripe);
1855         }
1856
1857         *spobj = stripe;
1858
1859         *sobj = mdt_object_find(env, info->mti_mdt, fid);
1860         if (IS_ERR(*sobj)) {
1861                 mdt_object_put(env, stripe);
1862                 rc = PTR_ERR(*sobj);
1863                 *spobj = NULL;
1864                 *sobj = NULL;
1865         }
1866
1867         return rc;
1868 }
1869
1870 /* end lease and close file for regular file */
1871 static int mdd_migrate_close(struct mdt_thread_info *info,
1872                              struct mdt_object *obj)
1873 {
1874         struct close_data *data;
1875         struct mdt_body *repbody;
1876         struct ldlm_lock *lease;
1877         int rc;
1878         int rc2;
1879
1880         rc = -EPROTO;
1881         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1882                                       RCL_CLIENT) ||
1883             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1884                                       RCL_CLIENT))
1885                 goto close;
1886
1887         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1888         if (!data)
1889                 goto close;
1890
1891         rc = -ESTALE;
1892         lease = ldlm_handle2lock(&data->cd_handle);
1893         if (!lease)
1894                 goto close;
1895
1896         /* check if the lease was already canceled */
1897         lock_res_and_lock(lease);
1898         rc = ldlm_is_cancel(lease);
1899         unlock_res_and_lock(lease);
1900
1901         if (rc) {
1902                 rc = -EAGAIN;
1903                 LDLM_DEBUG(lease, DFID" lease broken",
1904                            PFID(mdt_object_fid(obj)));
1905         }
1906
1907         /*
1908          * cancel server side lease, client side counterpart should have been
1909          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1910          */
1911         ldlm_lock_cancel(lease);
1912         ldlm_reprocess_all(lease->l_resource);
1913         LDLM_LOCK_PUT(lease);
1914
1915 close:
1916         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1917         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1918         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1919
1920         return rc ?: rc2;
1921 }
1922
1923 /*
1924  * migrate file in below steps:
1925  *  1. lock parent and its stripes
1926  *  2. lookup source by name
1927  *  3. lock parents of source links if source is not directory
1928  *  4. reject if source is in HSM
1929  *  5. take source open_sem and close file if source is regular file
1930  *  6. lock source and its stripes if it's directory
1931  *  7. lock target so subsequent change to it can trigger COS
1932  *  8. migrate file
1933  *  9. unlock above locks
1934  * 10. sync device if source has links
1935  */
1936 static int mdt_reint_migrate_internal(struct mdt_thread_info *info)
1937 {
1938         const struct lu_env *env = info->mti_env;
1939         struct mdt_device *mdt = info->mti_mdt;
1940         struct mdt_reint_record *rr = &info->mti_rr;
1941         struct lu_ucred *uc = mdt_ucred(info);
1942         struct md_attr *ma = &info->mti_attr;
1943         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1944         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1945         struct mdt_object *pobj;
1946         struct mdt_object *spobj = NULL;
1947         struct mdt_object *sobj = NULL;
1948         struct mdt_object *tobj;
1949         struct mdt_lock_handle *lhp;
1950         struct mdt_lock_handle *lhs;
1951         struct mdt_lock_handle *lht;
1952         LIST_HEAD(parent_slave_locks);
1953         LIST_HEAD(child_slave_locks);
1954         LIST_HEAD(link_locks);
1955         int lock_retries = 5;
1956         bool open_sem_locked = false;
1957         bool do_sync = false;
1958         int rc;
1959         ENTRY;
1960
1961         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1962                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1963
1964         /* don't allow migrate . or .. */
1965         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
1966                 RETURN(-EBUSY);
1967
1968         if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
1969                 RETURN(-EPERM);
1970
1971         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
1972             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
1973             mdt->mdt_enable_remote_dir_gid != -1)
1974                 RETURN(-EPERM);
1975
1976         /* pobj is master object of parent */
1977         pobj = mdt_parent_find_check(info, rr->rr_fid1, 0);
1978         if (IS_ERR(pobj))
1979                 RETURN(PTR_ERR(pobj));
1980
1981         if (unlikely(!info->mti_big_lmm)) {
1982                 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
1983                 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
1984                 if (!info->mti_big_lmm)
1985                         GOTO(put_parent, rc = -ENOMEM);
1986         }
1987
1988         ma->ma_lmv = info->mti_big_lmm;
1989         ma->ma_lmv_size = info->mti_big_lmmsize;
1990         ma->ma_valid = 0;
1991         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
1992         if (rc)
1993                 GOTO(put_parent, rc);
1994
1995 lock_parent:
1996         /* lock parent object */
1997         lhp = &info->mti_lh[MDT_LH_PARENT];
1998         mdt_lock_reg_init(lhp, LCK_PW);
1999         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2000                                      &parent_slave_locks);
2001         if (rc)
2002                 GOTO(put_parent, rc);
2003
2004         /*
2005          * spobj is the corresponding stripe against name if pobj is striped
2006          * directory, which is the real parent, and no need to lock, because
2007          * we've taken full lock of pobj.
2008          */
2009         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2010         if (rc)
2011                 GOTO(unlock_parent, rc);
2012
2013         /* lock parents of source links, and revoke LOOKUP lock of links */
2014         rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2015                                    &parent_slave_locks, &link_locks);
2016         if (rc == -EBUSY && lock_retries-- > 0) {
2017                 mdt_object_put(env, sobj);
2018                 mdt_object_put(env, spobj);
2019                 goto lock_parent;
2020         }
2021
2022         if (rc < 0)
2023                 GOTO(put_source, rc);
2024
2025         /*
2026          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2027          * with one request, if total lock count exceeds this limit, we will
2028          * drop all locks after migration, and synchronous device in the end.
2029          */
2030         do_sync = rc;
2031
2032         /* TODO: DoM migration is not supported yet */
2033         if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2034                 ma->ma_lmm = info->mti_big_lmm;
2035                 ma->ma_lmm_size = info->mti_big_lmmsize;
2036                 ma->ma_valid = 0;
2037                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2038                 if (rc)
2039                         GOTO(put_source, rc);
2040
2041                 if (ma->ma_valid & MA_LOV &&
2042                     mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2043                         GOTO(put_source, rc = -EOPNOTSUPP);
2044         }
2045
2046         /* if migration HSM is allowed */
2047         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2048                 ma->ma_need = MA_HSM;
2049                 ma->ma_valid = 0;
2050                 rc = mdt_attr_get_complex(info, sobj, ma);
2051                 if (rc)
2052                         GOTO(unlock_links, rc);
2053
2054                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2055                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2056         }
2057
2058         /* end lease and close file for regular file */
2059         if (info->mti_spec.sp_migrate_close) {
2060                 /* try to hold open_sem so that nobody else can open the file */
2061                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2062                         /* close anyway */
2063                         mdd_migrate_close(info, sobj);
2064                         GOTO(unlock_links, rc = -EBUSY);
2065                 } else {
2066                         open_sem_locked = true;
2067                         rc = mdd_migrate_close(info, sobj);
2068                         if (rc)
2069                                 GOTO(unlock_open_sem, rc);
2070                 }
2071         }
2072
2073         /* lock source */
2074         lhs = &info->mti_lh[MDT_LH_OLD];
2075         mdt_lock_reg_init(lhs, LCK_EX);
2076         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2077                                      &child_slave_locks);
2078         if (rc)
2079                 GOTO(unlock_open_sem, rc);
2080
2081         /* lock target */
2082         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2083         if (IS_ERR(tobj))
2084                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2085
2086         lht = &info->mti_lh[MDT_LH_NEW];
2087         mdt_lock_reg_init(lht, LCK_EX);
2088         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2089         if (rc)
2090                 GOTO(put_target, rc);
2091
2092         /* Don't do lookup sanity check. We know name doesn't exist. */
2093         info->mti_spec.sp_cr_lookup = 0;
2094         info->mti_spec.sp_feat = &dt_directory_features;
2095
2096         rc = mdo_migrate(env, mdt_object_child(pobj),
2097                          mdt_object_child(sobj), &rr->rr_name,
2098                          mdt_object_child(tobj), &info->mti_spec, ma);
2099         EXIT;
2100
2101         mdt_object_unlock(info, tobj, lht, rc);
2102 put_target:
2103         mdt_object_put(env, tobj);
2104 unlock_source:
2105         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2106                                   &child_slave_locks, rc);
2107 unlock_open_sem:
2108         if (open_sem_locked)
2109                 up_write(&sobj->mot_open_sem);
2110 unlock_links:
2111         mdt_unlock_list(info, &link_locks, do_sync ?: rc);
2112 put_source:
2113         mdt_object_put(env, sobj);
2114         mdt_object_put(env, spobj);
2115 unlock_parent:
2116         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2117                                   &parent_slave_locks, rc);
2118 put_parent:
2119         mdt_object_put(env, pobj);
2120
2121         if (!rc && do_sync)
2122                 mdt_device_sync(env, mdt);
2123
2124         return rc;
2125 }
2126
2127 static int mdt_object_lock_save(struct mdt_thread_info *info,
2128                                 struct mdt_object *dir,
2129                                 struct mdt_lock_handle *lh,
2130                                 int idx, bool cos_incompat)
2131 {
2132         int rc;
2133
2134         /* we lock the target dir if it is local */
2135         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2136                                    cos_incompat);
2137         if (rc != 0)
2138                 return rc;
2139
2140         /* get and save correct version after locking */
2141         mdt_version_get_save(info, dir, idx);
2142         return 0;
2143 }
2144
2145 /*
2146  * determine lock order of sobj and tobj
2147  *
2148  * there are two situations we need to lock tobj before sobj:
2149  * 1. sobj is child of tobj
2150  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2151  *    larger than that of tobj
2152  *
2153  * \retval      1 lock tobj before sobj
2154  * \retval      0 lock sobj before tobj
2155  * \retval      -ev negative errno upon error
2156  */
2157 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2158                                            struct mdt_object *sobj,
2159                                            struct mdt_object *tobj)
2160 {
2161         struct md_attr *ma = &info->mti_attr;
2162         struct lu_fid *spfid = &info->mti_tmp_fid1;
2163         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2164         struct lmv_mds_md_v1 *lmv;
2165         __u32 sindex;
2166         __u32 tindex;
2167         int rc;
2168
2169         /* sobj and tobj are the same */
2170         if (sobj == tobj)
2171                 return 0;
2172
2173         if (fid_is_root(mdt_object_fid(sobj)))
2174                 return 0;
2175
2176         if (fid_is_root(mdt_object_fid(tobj)))
2177                 return 1;
2178
2179         /* check whether sobj is child of tobj */
2180         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2181                            mdt_object_fid(tobj));
2182         if (rc < 0)
2183                 return rc;
2184
2185         if (rc == 1)
2186                 return 1;
2187
2188         /* check whether sobj and tobj are children of the same parent */
2189         rc = mdt_attr_get_pfid(info, sobj, spfid);
2190         if (rc)
2191                 return rc;
2192
2193         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2194         if (rc)
2195                 return rc;
2196
2197         if (!lu_fid_eq(spfid, tpfid))
2198                 return 0;
2199
2200         /* check whether sobj and tobj are sibling stripes */
2201         ma->ma_need = MA_LMV;
2202         ma->ma_valid = 0;
2203         ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2204         ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2205         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2206         if (rc)
2207                 return rc;
2208
2209         if (!(ma->ma_valid & MA_LMV))
2210                 return 0;
2211
2212         lmv = &ma->ma_lmv->lmv_md_v1;
2213         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2214                 return 0;
2215         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2216
2217         ma->ma_valid = 0;
2218         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2219         if (rc)
2220                 return rc;
2221
2222         if (!(ma->ma_valid & MA_LMV))
2223                 return -ENODATA;
2224
2225         lmv = &ma->ma_lmv->lmv_md_v1;
2226         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2227                 return -EINVAL;
2228         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2229
2230         /* check stripe index of sobj and tobj */
2231         if (sindex == tindex)
2232                 return -EINVAL;
2233
2234         return sindex < tindex ? 0 : 1;
2235 }
2236
2237 /*
2238  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2239  * 2 - srcdir child; 3 - tgtdir child.
2240  * Update on disk version of srcdir child.
2241  */
2242 /**
2243  * For DNE phase I, only these renames are allowed
2244  *      mv src_p/src_c tgt_p/tgt_c
2245  * 1. src_p/src_c/tgt_p/tgt_c are in the same MDT.
2246  * 2. src_p and tgt_p are same directory, and tgt_c does not
2247  *    exists. In this case, all of modification will happen
2248  *    in the MDT where ithesource parent is, only one remote
2249  *    update is needed, i.e. set c_time/m_time on the child.
2250  *    And tgt_c will be still in the same MDT as the original
2251  *    src_c.
2252  */
2253 static int mdt_reint_rename_internal(struct mdt_thread_info *info,
2254                                      struct mdt_lock_handle *lhc)
2255 {
2256         struct mdt_reint_record *rr = &info->mti_rr;
2257         struct md_attr *ma = &info->mti_attr;
2258         struct ptlrpc_request *req = mdt_info_req(info);
2259         struct mdt_object *msrcdir = NULL;
2260         struct mdt_object *mtgtdir = NULL;
2261         struct mdt_object *mold;
2262         struct mdt_object *mnew = NULL;
2263         struct mdt_lock_handle *lh_srcdirp;
2264         struct mdt_lock_handle *lh_tgtdirp;
2265         struct mdt_lock_handle *lh_oldp = NULL;
2266         struct mdt_lock_handle *lh_newp = NULL;
2267         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2268         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2269         __u64 lock_ibits;
2270         bool reverse = false;
2271         bool cos_incompat, discard = false;
2272         int rc;
2273         ENTRY;
2274
2275         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2276                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2277                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2278
2279         /* find both parents. */
2280         msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2281         if (IS_ERR(msrcdir))
2282                 RETURN(PTR_ERR(msrcdir));
2283
2284         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2285
2286         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2287                 mtgtdir = msrcdir;
2288                 mdt_object_get(info->mti_env, mtgtdir);
2289         } else {
2290                 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2291                 if (IS_ERR(mtgtdir))
2292                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2293         }
2294
2295         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2296         if (rc < 0)
2297                 GOTO(out_put_tgtdir, rc);
2298
2299         reverse = rc;
2300
2301         /* source needs to be looked up after locking source parent, otherwise
2302          * this rename may race with unlink source, and cause rename hang, see
2303          * sanityn.sh 55b, so check parents first, if later we found source is
2304          * remote, relock parents. */
2305         cos_incompat = (mdt_object_remote(msrcdir) ||
2306                         mdt_object_remote(mtgtdir));
2307
2308         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2309
2310         /* lock parents in the proper order. */
2311         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2312         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2313
2314 relock:
2315         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2316         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2317
2318         if (reverse) {
2319                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2320                                           cos_incompat);
2321                 if (rc)
2322                         GOTO(out_put_tgtdir, rc);
2323
2324                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2325
2326                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2327                                           cos_incompat);
2328                 if (rc != 0) {
2329                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2330                         GOTO(out_put_tgtdir, rc);
2331                 }
2332         } else {
2333                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2334                                           cos_incompat);
2335                 if (rc)
2336                         GOTO(out_put_tgtdir, rc);
2337
2338                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2339
2340                 if (mtgtdir != msrcdir) {
2341                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2342                                                   cos_incompat);
2343                 } else if (lh_srcdirp->mlh_pdo_hash !=
2344                            lh_tgtdirp->mlh_pdo_hash) {
2345                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2346                                                 MDS_INODELOCK_UPDATE,
2347                                                 cos_incompat);
2348                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2349                 }
2350                 if (rc != 0) {
2351                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2352                         GOTO(out_put_tgtdir, rc);
2353                 }
2354         }
2355
2356         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2357         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2358
2359         /* find mold object. */
2360         fid_zero(old_fid);
2361         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2362         if (rc != 0)
2363                 GOTO(out_unlock_parents, rc);
2364
2365         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2366                 GOTO(out_unlock_parents, rc = -EINVAL);
2367
2368         if (!fid_is_md_operative(old_fid))
2369                 GOTO(out_unlock_parents, rc = -EPERM);
2370
2371         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2372         if (IS_ERR(mold))
2373                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2374
2375         if (!mdt_object_exists(mold)) {
2376                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2377                                 &mold->mot_obj,
2378                                 "object does not exist");
2379                 GOTO(out_put_old, rc = -ENOENT);
2380         }
2381
2382         /* Check if @mtgtdir is subdir of @mold, before locking child
2383          * to avoid reverse locking. */
2384         if (mtgtdir != msrcdir) {
2385                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2386                                    old_fid);
2387                 if (rc) {
2388                         if (rc == 1)
2389                                 rc = -EINVAL;
2390                         GOTO(out_put_old, rc);
2391                 }
2392         }
2393
2394         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2395         /* save version after locking */
2396         mdt_version_get_save(info, mold, 2);
2397
2398         if (!cos_incompat && mdt_object_remote(mold)) {
2399                 cos_incompat = true;
2400                 mdt_object_put(info->mti_env, mold);
2401                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2402                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2403                 goto relock;
2404         }
2405
2406         /* find mnew object:
2407          * mnew target object may not exist now
2408          * lookup with version checking */
2409         fid_zero(new_fid);
2410         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2411                                       3);
2412         if (rc == 0) {
2413                 /* the new_fid should have been filled at this moment */
2414                 if (lu_fid_eq(old_fid, new_fid))
2415                         GOTO(out_put_old, rc);
2416
2417                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2418                     lu_fid_eq(new_fid, rr->rr_fid2))
2419                         GOTO(out_put_old, rc = -EINVAL);
2420
2421                 if (!fid_is_md_operative(new_fid))
2422                         GOTO(out_put_old, rc = -EPERM);
2423
2424                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2425                 if (IS_ERR(mnew))
2426                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2427
2428                 if (!mdt_object_exists(mnew)) {
2429                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2430                                         &mnew->mot_obj,
2431                                         "object does not exist");
2432                         GOTO(out_put_new, rc = -ENOENT);
2433                 }
2434
2435                 if (mdt_object_remote(mnew)) {
2436                         struct mdt_body  *repbody;
2437
2438                         /* Always send rename req to the target child MDT */
2439                         repbody = req_capsule_server_get(info->mti_pill,
2440                                                          &RMF_MDT_BODY);
2441                         LASSERT(repbody != NULL);
2442                         repbody->mbo_fid1 = *new_fid;
2443                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2444                         GOTO(out_put_new, rc = -EXDEV);
2445                 }
2446                 /* Before locking the target dir, check we do not replace
2447                  * a dir with a non-dir, otherwise it may deadlock with
2448                  * link op which tries to create a link in this dir
2449                  * back to this non-dir. */
2450                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2451                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2452                         GOTO(out_put_new, rc = -EISDIR);
2453
2454                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2455                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2456                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2457                 if (mdt_object_remote(msrcdir)) {
2458                         /* Enqueue lookup lock from the parent MDT */
2459                         rc = mdt_remote_object_lock(info, msrcdir,
2460                                                     mdt_object_fid(mold),
2461                                                     &lh_oldp->mlh_rreg_lh,
2462                                                     lh_oldp->mlh_rreg_mode,
2463                                                     MDS_INODELOCK_LOOKUP,
2464                                                     false);
2465                         if (rc != ELDLM_OK)
2466                                 GOTO(out_put_new, rc);
2467
2468                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2469                 }
2470
2471                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2472                                            cos_incompat);
2473                 if (rc != 0)
2474                         GOTO(out_unlock_old, rc);
2475
2476                 /* Check if @msrcdir is subdir of @mnew, before locking child
2477                  * to avoid reverse locking. */
2478                 if (mtgtdir != msrcdir) {
2479                         rc = mdo_is_subdir(info->mti_env,
2480                                            mdt_object_child(msrcdir), new_fid);
2481                         if (rc) {
2482                                 if (rc == 1)
2483                                         rc = -EINVAL;
2484                                 GOTO(out_unlock_old, rc);
2485                         }
2486                 }
2487
2488                 /* We used to acquire MDS_INODELOCK_FULL here but we
2489                  * can't do this now because a running HSM restore on
2490                  * the rename onto victim will hold the layout
2491                  * lock. See LU-4002. */
2492
2493                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2494                 mdt_lock_reg_init(lh_newp, LCK_EX);
2495                 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2496                                            MDS_INODELOCK_LOOKUP |
2497                                            MDS_INODELOCK_UPDATE,
2498                                            cos_incompat);
2499                 if (rc != 0)
2500                         GOTO(out_unlock_old, rc);
2501
2502                 /* get and save version after locking */
2503                 mdt_version_get_save(info, mnew, 3);
2504         } else if (rc != -EREMOTE && rc != -ENOENT) {
2505                 GOTO(out_put_old, rc);
2506         } else {
2507                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2508                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2509                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2510                 if (mdt_object_remote(msrcdir)) {
2511                         /* Enqueue lookup lock from the parent MDT */
2512                         rc = mdt_remote_object_lock(info, msrcdir,
2513                                                     mdt_object_fid(mold),
2514                                                     &lh_oldp->mlh_rreg_lh,
2515                                                     lh_oldp->mlh_rreg_mode,
2516                                                     MDS_INODELOCK_LOOKUP,
2517                                                     false);
2518                         if (rc != ELDLM_OK)
2519                                 GOTO(out_put_old, rc);
2520
2521                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2522                 }
2523
2524                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2525                                            cos_incompat);
2526                 if (rc != 0)
2527                         GOTO(out_unlock_old, rc);
2528
2529                 mdt_enoent_version_save(info, 3);
2530         }
2531
2532         /* step 5: rename it */
2533         mdt_reint_init_ma(info, ma);
2534
2535         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2536                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2537
2538         if (mnew != NULL)
2539                 mutex_lock(&mnew->mot_lov_mutex);
2540
2541         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2542                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2543                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2544                         &rr->rr_tgt_name, ma);
2545
2546         if (mnew != NULL)
2547                 mutex_unlock(&mnew->mot_lov_mutex);
2548
2549         /* handle last link of tgt object */
2550         if (rc == 0) {
2551                 mdt_counter_incr(req, LPROC_MDT_RENAME);
2552                 if (mnew) {
2553                         mdt_handle_last_unlink(info, mnew, ma);
2554                         discard = mdt_dom_check_for_discard(info, mnew);
2555                 }
2556
2557                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2558                                          msrcdir, mtgtdir);
2559         }
2560
2561         EXIT;
2562         if (mnew != NULL)
2563                 mdt_object_unlock(info, mnew, lh_newp, rc);
2564 out_unlock_old:
2565         mdt_object_unlock(info, mold, lh_oldp, rc);
2566 out_put_new:
2567         if (mnew != NULL)
2568                 mdt_object_put(info->mti_env, mnew);
2569 out_put_old:
2570         mdt_object_put(info->mti_env, mold);
2571 out_unlock_parents:
2572         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2573         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2574 out_put_tgtdir:
2575         mdt_object_put(info->mti_env, mtgtdir);
2576 out_put_srcdir:
2577         mdt_object_put(info->mti_env, msrcdir);
2578
2579         /* If 'discard' is set then new_fid must exits.
2580          * DOM data discard need neither object nor lock,
2581          * so do this at the end.
2582          */
2583         if (discard)
2584                 mdt_dom_discard_data(info, new_fid);
2585
2586         return rc;
2587 }
2588
2589 static int mdt_reint_rename_or_migrate(struct mdt_thread_info *info,
2590                                        struct mdt_lock_handle *lhc, bool rename)
2591 {
2592         struct mdt_reint_record *rr = &info->mti_rr;
2593         struct ptlrpc_request   *req = mdt_info_req(info);
2594         struct lustre_handle    rename_lh = { 0 };
2595         int                     rc;
2596         ENTRY;
2597
2598         if (info->mti_dlm_req)
2599                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2600
2601         if (!fid_is_md_operative(rr->rr_fid1) ||
2602             !fid_is_md_operative(rr->rr_fid2))
2603                 RETURN(-EPERM);
2604
2605         /* Note: do not enqueue rename lock for replay request, because
2606          * if other MDT holds rename lock, but being blocked to wait for
2607          * this MDT to finish its recovery, and the failover MDT can not
2608          * get rename lock, which will cause deadlock. */
2609         if (!req_is_replay(req)) {
2610                 rc = mdt_rename_lock(info, &rename_lh);
2611                 if (rc != 0) {
2612                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2613                                mdt_obd_name(info->mti_mdt), rc);
2614                         RETURN(rc);
2615                 }
2616         }
2617
2618         if (rename)
2619                 rc = mdt_reint_rename_internal(info, lhc);
2620         else
2621                 rc = mdt_reint_migrate_internal(info);
2622
2623         if (lustre_handle_is_used(&rename_lh))
2624                 mdt_rename_unlock(&rename_lh);
2625
2626         RETURN(rc);
2627 }
2628
2629 static int mdt_reint_rename(struct mdt_thread_info *info,
2630                             struct mdt_lock_handle *lhc)
2631 {
2632         return mdt_reint_rename_or_migrate(info, lhc, true);
2633 }
2634
2635 static int mdt_reint_migrate(struct mdt_thread_info *info,
2636                             struct mdt_lock_handle *lhc)
2637 {
2638         return mdt_reint_rename_or_migrate(info, lhc, false);
2639 }
2640
2641 static int mdt_reint_resync(struct mdt_thread_info *info,
2642                             struct mdt_lock_handle *lhc)
2643 {
2644         struct mdt_reint_record *rr = &info->mti_rr;
2645         struct ptlrpc_request   *req = mdt_info_req(info);
2646         struct md_attr          *ma = &info->mti_attr;
2647         struct mdt_object       *mo;
2648         struct ldlm_lock        *lease;
2649         struct mdt_body         *repbody;
2650         struct md_layout_change  layout = { .mlc_mirror_id = rr->rr_mirror_id };
2651         bool                     lease_broken;
2652         int                      rc, rc2;
2653         ENTRY;
2654
2655         DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2656
2657         if (info->mti_dlm_req)
2658                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2659
2660         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2661         if (IS_ERR(mo))
2662                 GOTO(out, rc = PTR_ERR(mo));
2663
2664         if (!mdt_object_exists(mo))
2665                 GOTO(out_obj, rc = -ENOENT);
2666
2667         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2668                 GOTO(out_obj, rc = -EINVAL);
2669
2670         if (mdt_object_remote(mo))
2671                 GOTO(out_obj, rc = -EREMOTE);
2672
2673         lease = ldlm_handle2lock(rr->rr_lease_handle);
2674         if (lease == NULL)
2675                 GOTO(out_obj, rc = -ESTALE);
2676
2677         /* It's really necessary to grab open_sem and check if the lease lock
2678          * has been lost. There would exist a concurrent writer coming in and
2679          * generating some dirty data in memory cache, the writeback would fail
2680          * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2681         if (!down_write_trylock(&mo->mot_open_sem))
2682                 GOTO(out_put_lease, rc = -EBUSY);
2683
2684         lock_res_and_lock(lease);
2685         lease_broken = ldlm_is_cancel(lease);
2686         unlock_res_and_lock(lease);
2687         if (lease_broken)
2688                 GOTO(out_unlock, rc = -EBUSY);
2689
2690         /* the file has yet opened by anyone else after we took the lease. */
2691         layout.mlc_opc = MD_LAYOUT_RESYNC;
2692         rc = mdt_layout_change(info, mo, &layout);
2693         if (rc)
2694                 GOTO(out_unlock, rc);
2695
2696         ma->ma_need = MA_INODE;
2697         ma->ma_valid = 0;
2698         rc = mdt_attr_get_complex(info, mo, ma);
2699         if (rc != 0)
2700                 GOTO(out_unlock, rc);
2701
2702         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2703         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2704
2705         EXIT;
2706 out_unlock:
2707         up_write(&mo->mot_open_sem);
2708 out_put_lease:
2709         LDLM_LOCK_PUT(lease);
2710 out_obj:
2711         mdt_object_put(info->mti_env, mo);
2712 out:
2713         mdt_client_compatibility(info);
2714         rc2 = mdt_fix_reply(info);
2715         if (rc == 0)
2716                 rc = rc2;
2717         return rc;
2718 }
2719
2720 struct mdt_reinter {
2721         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2722         enum lprocfs_extra_opc mr_extra_opc;
2723 };
2724
2725 static const struct mdt_reinter mdt_reinters[] = {
2726         [REINT_SETATTR] = {
2727                 .mr_handler = &mdt_reint_setattr,
2728                 .mr_extra_opc = MDS_REINT_SETATTR,
2729         },
2730         [REINT_CREATE] = {
2731                 .mr_handler = &mdt_reint_create,
2732                 .mr_extra_opc = MDS_REINT_CREATE,
2733         },
2734         [REINT_LINK] = {
2735                 .mr_handler = &mdt_reint_link,
2736                 .mr_extra_opc = MDS_REINT_LINK,
2737         },
2738         [REINT_UNLINK] = {
2739                 .mr_handler = &mdt_reint_unlink,
2740                 .mr_extra_opc = MDS_REINT_UNLINK,
2741         },
2742         [REINT_RENAME] = {
2743                 .mr_handler = &mdt_reint_rename,
2744                 .mr_extra_opc = MDS_REINT_RENAME,
2745         },
2746         [REINT_OPEN] = {
2747                 .mr_handler = &mdt_reint_open,
2748                 .mr_extra_opc = MDS_REINT_OPEN,
2749         },
2750         [REINT_SETXATTR] = {
2751                 .mr_handler = &mdt_reint_setxattr,
2752                 .mr_extra_opc = MDS_REINT_SETXATTR,
2753         },
2754         [REINT_RMENTRY] = {
2755                 .mr_handler = &mdt_reint_unlink,
2756                 .mr_extra_opc = MDS_REINT_UNLINK,
2757         },
2758         [REINT_MIGRATE] = {
2759                 .mr_handler = &mdt_reint_migrate,
2760                 .mr_extra_opc = MDS_REINT_RENAME,
2761         },
2762         [REINT_RESYNC] = {
2763                 .mr_handler = &mdt_reint_resync,
2764                 .mr_extra_opc = MDS_REINT_RESYNC,
2765         },
2766 };
2767
2768 int mdt_reint_rec(struct mdt_thread_info *info,
2769                   struct mdt_lock_handle *lhc)
2770 {
2771         const struct mdt_reinter *mr;
2772         int rc;
2773         ENTRY;
2774
2775         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2776                 RETURN(-EPROTO);
2777
2778         mr = &mdt_reinters[info->mti_rr.rr_opcode];
2779         if (mr->mr_handler == NULL)
2780                 RETURN(-EPROTO);
2781
2782         rc = (*mr->mr_handler)(info, lhc);
2783
2784         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2785                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
2786
2787         RETURN(rc);
2788 }