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         ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
 656         /* revoke lease lock if size is going to be changed */
 657         if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
 658                      !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
 659                      atomic_read(&mo->mot_lease_count) > 0)) {
 660                 down_read(&mo->mot_open_sem);
 661
 662                 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
 663                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 664                         mdt_lock_reg_init(lhc, LCK_CW);
 665
 666                         rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
 667                         if (rc != 0) {
 668                                 up_read(&mo->mot_open_sem);
 669                                 GOTO(out_put, rc);
 670                         }
 671
 672                         /* revoke lease lock */
 673                         mdt_object_unlock(info, mo, lhc, 1);
 674                 }
 675                 up_read(&mo->mot_open_sem);
 676         }
 677
 678         if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
 679                 /* Check write access for the O_TRUNC case */
 680                 if (mdt_write_read(mo) < 0)
 681                         GOTO(out_put, rc = -ETXTBSY);
 682
 683                 /* LU-10286: compatibility check for FLR.
 684                  * Please check the comment in mdt_finish_open() for details */
 685                 if (!exp_connect_flr(info->mti_exp)) {
 686                         rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
 687                         if (rc < 0 && rc != -ENODATA)
 688                                 GOTO(out_put, rc);
 689
 690                         if (rc > 0 && mdt_lmm_is_flr(info->mti_big_lmm))
 691                                 GOTO(out_put, rc = -EOPNOTSUPP);
 692                 }
 693
 694                 /* For truncate, the file size sent from client
 695                  * is believable, but the blocks are incorrect,
 696                  * which makes the block size in LSOM attribute
 697                  * inconsisent with the real block size.
 698                  */
 699                 rc = mdt_lsom_update(info, mo, true);
 700                 if (rc)
 701                         GOTO(out_put, rc);
 702         }
 703
 704         if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
 705                 if (ma->ma_valid & MA_LOV)
 706                         GOTO(out_put, rc = -EPROTO);
 707
 708                 /* MDT supports FMD for regular files due to Data-on-MDT */
 709                 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
 710                     ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
 711                         tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
 712                                        req->rq_xid);
 713
 714                 rc = mdt_attr_set(info, mo, ma);
 715                 if (rc)
 716                         GOTO(out_put, rc);
 717         } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
 718                    (ma->ma_valid & MA_INODE)) {
 719                 struct lu_buf *buf = &info->mti_buf;
 720                 struct lu_ucred *uc = mdt_ucred(info);
 721                 struct mdt_lock_handle *lh;
 722
 723                 /* reject if either remote or striped dir is disabled */
 724                 if (ma->ma_valid & MA_LMV) {
 725                         if (!mdt->mdt_enable_remote_dir ||
 726                             !mdt->mdt_enable_striped_dir)
 727                                 GOTO(out_put, rc = -EPERM);
 728
 729                         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 730                             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 731                             mdt->mdt_enable_remote_dir_gid != -1)
 732                                 GOTO(out_put, rc = -EPERM);
 733                 }
 734
 735                 if (ma->ma_attr.la_valid != 0)
 736                         GOTO(out_put, rc = -EPROTO);
 737
 738                 lh = &info->mti_lh[MDT_LH_PARENT];
 739                 mdt_lock_reg_init(lh, LCK_PW);
 740
 741                 rc = mdt_object_lock(info, mo, lh, MDS_INODELOCK_XATTR);
 742                 if (rc != 0)
 743                         GOTO(out_put, rc);
 744
 745                 if (ma->ma_valid & MA_LOV) {
 746                         buf->lb_buf = ma->ma_lmm;
 747                         buf->lb_len = ma->ma_lmm_size;
 748                 } else {
 749                         buf->lb_buf = ma->ma_lmv;
 750                         buf->lb_len = ma->ma_lmv_size;
 751                 }
 752                 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
 753                                   (ma->ma_valid & MA_LOV) ?
 754                                         XATTR_NAME_LOV : XATTR_NAME_DEFAULT_LMV,
 755                                   0);
 756
 757                 mdt_object_unlock(info, mo, lh, rc);
 758                 if (rc)
 759                         GOTO(out_put, rc);
 760         } else {
 761                 GOTO(out_put, rc = -EPROTO);
 762         }
 763
 764         /* If file data is modified, add the dirty flag */
 765         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
 766                 rc = mdt_add_dirty_flag(info, mo, ma);
 767
 768         ma->ma_need = MA_INODE;
 769         ma->ma_valid = 0;
 770         rc = mdt_attr_get_complex(info, mo, ma);
 771         if (rc != 0)
 772                 GOTO(out_put, rc);
 773
 774         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
 775
 776         EXIT;
 777 out_put:
 778         mdt_object_put(info->mti_env, mo);
 779 out:
 780         if (rc == 0)
 781                 mdt_counter_incr(req, LPROC_MDT_SETATTR);
 782
 783         mdt_client_compatibility(info);
 784         rc2 = mdt_fix_reply(info);
 785         if (rc == 0)
 786                 rc = rc2;
 787         return rc;
 788 }
 789
 790 static int mdt_reint_create(struct mdt_thread_info *info,
 791                             struct mdt_lock_handle *lhc)
 792 {
 793         struct ptlrpc_request   *req = mdt_info_req(info);
 794         int                     rc;
 795         ENTRY;
 796
 797         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 798                 RETURN(err_serious(-ESTALE));
 799
 800         if (info->mti_dlm_req)
 801                 ldlm_request_cancel(mdt_info_req(info),
 802                                     info->mti_dlm_req, 0, LATF_SKIP);
 803
 804         if (!lu_name_is_valid(&info->mti_rr.rr_name))
 805                 RETURN(-EPROTO);
 806
 807         switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
 808         case S_IFDIR:
 809                 mdt_counter_incr(req, LPROC_MDT_MKDIR);
 810                 break;
 811         case S_IFREG:
 812         case S_IFLNK:
 813         case S_IFCHR:
 814         case S_IFBLK:
 815         case S_IFIFO:
 816         case S_IFSOCK:
 817                 /* Special file should stay on the same node as parent. */
 818                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
 819                 break;
 820         default:
 821                 CERROR("%s: Unsupported mode %o\n",
 822                        mdt_obd_name(info->mti_mdt),
 823                        info->mti_attr.ma_attr.la_mode);
 824                 RETURN(err_serious(-EOPNOTSUPP));
 825         }
 826
 827         rc = mdt_create(info);
 828         RETURN(rc);
 829 }
 830
 831 /*
 832  * VBR: save parent version in reply and child version getting by its name.
 833  * Version of child is getting and checking during its lookup. If
 834  */
 835 static int mdt_reint_unlink(struct mdt_thread_info *info,
 836                             struct mdt_lock_handle *lhc)
 837 {
 838         struct mdt_reint_record *rr = &info->mti_rr;
 839         struct ptlrpc_request *req = mdt_info_req(info);
 840         struct md_attr *ma = &info->mti_attr;
 841         struct lu_fid *child_fid = &info->mti_tmp_fid1;
 842         struct mdt_object *mp;
 843         struct mdt_object *mc;
 844         struct mdt_lock_handle *parent_lh;
 845         struct mdt_lock_handle *child_lh;
 846         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 847         __u64 lock_ibits;
 848         bool cos_incompat = false;
 849         int no_name = 0;
 850         int rc;
 851
 852         ENTRY;
 853
 854         DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
 855                   PNAME(&rr->rr_name));
 856
 857         if (info->mti_dlm_req)
 858                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 859
 860         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
 861                 RETURN(err_serious(-ENOENT));
 862
 863         if (!fid_is_md_operative(rr->rr_fid1))
 864                 RETURN(-EPERM);
 865
 866         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 867         if (IS_ERR(mp))
 868                 RETURN(PTR_ERR(mp));
 869
 870         if (mdt_object_remote(mp)) {
 871                 cos_incompat = true;
 872         } else {
 873                 rc = mdt_version_get_check_save(info, mp, 0);
 874                 if (rc)
 875                         GOTO(put_parent, rc);
 876         }
 877
 878         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
 879         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
 880 relock:
 881         parent_lh = &info->mti_lh[MDT_LH_PARENT];
 882         mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
 883         rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
 884                                    cos_incompat);
 885         if (rc != 0)
 886                 GOTO(put_parent, rc);
 887
 888         /* lookup child object along with version checking */
 889         fid_zero(child_fid);
 890         rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
 891         if (rc != 0) {
 892                 /* Name might not be able to find during resend of
 893                  * remote unlink, considering following case.
 894                  * dir_A is a remote directory, the name entry of
 895                  * dir_A is on MDT0, the directory is on MDT1,
 896                  *
 897                  * 1. client sends unlink req to MDT1.
 898                  * 2. MDT1 sends name delete update to MDT0.
 899                  * 3. name entry is being deleted in MDT0 synchronously.
 900                  * 4. MDT1 is restarted.
 901                  * 5. client resends unlink req to MDT1. So it can not
 902                  *    find the name entry on MDT0 anymore.
 903                  * In this case, MDT1 only needs to destory the local
 904                  * directory.
 905                  * */
 906                 if (mdt_object_remote(mp) && rc == -ENOENT &&
 907                     !fid_is_zero(rr->rr_fid2) &&
 908                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 909                         no_name = 1;
 910                         *child_fid = *rr->rr_fid2;
 911                  } else {
 912                         GOTO(unlock_parent, rc);
 913                  }
 914         }
 915
 916         if (!fid_is_md_operative(child_fid))
 917                 GOTO(unlock_parent, rc = -EPERM);
 918
 919         /* We will lock the child regardless it is local or remote. No harm. */
 920         mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
 921         if (IS_ERR(mc))
 922                 GOTO(unlock_parent, rc = PTR_ERR(mc));
 923
 924         if (!cos_incompat) {
 925                 rc = mdt_object_striped(info, mc);
 926                 if (rc < 0)
 927                         GOTO(put_child, rc);
 928
 929                 cos_incompat = rc;
 930                 if (cos_incompat) {
 931                         mdt_object_put(info->mti_env, mc);
 932                         mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
 933                         goto relock;
 934                 }
 935         }
 936
 937         child_lh = &info->mti_lh[MDT_LH_CHILD];
 938         mdt_lock_reg_init(child_lh, LCK_EX);
 939         if (info->mti_spec.sp_rm_entry) {
 940                 struct lu_ucred *uc  = mdt_ucred(info);
 941
 942                 if (!mdt_is_dne_client(req->rq_export))
 943                         /* Return -ENOTSUPP for old client */
 944                         GOTO(put_child, rc = -ENOTSUPP);
 945
 946                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
 947                         GOTO(put_child, rc = -EPERM);
 948
 949                 ma->ma_need = MA_INODE;
 950                 ma->ma_valid = 0;
 951                 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
 952                                 NULL, &rr->rr_name, ma, no_name);
 953                 GOTO(put_child, rc);
 954         }
 955
 956         if (mdt_object_remote(mc)) {
 957                 struct mdt_body  *repbody;
 958
 959                 if (!fid_is_zero(rr->rr_fid2)) {
 960                         CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
 961                                mdt_obd_name(info->mti_mdt),
 962                                PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 963                         GOTO(put_child, rc = -ENOENT);
 964                 }
 965                 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
 966                        mdt_obd_name(info->mti_mdt),
 967                        PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 968
 969                 if (!mdt_is_dne_client(req->rq_export))
 970                         /* Return -ENOTSUPP for old client */
 971                         GOTO(put_child, rc = -ENOTSUPP);
 972
 973                 /* Revoke the LOOKUP lock of the remote object granted by
 974                  * this MDT. Since the unlink will happen on another MDT,
 975                  * it will release the LOOKUP lock right away. Then What
 976                  * would happen if another client try to grab the LOOKUP
 977                  * lock at the same time with unlink XXX */
 978                 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
 979                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 980                 LASSERT(repbody != NULL);
 981                 repbody->mbo_fid1 = *mdt_object_fid(mc);
 982                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
 983                 GOTO(unlock_child, rc = -EREMOTE);
 984         }
 985         /* We used to acquire MDS_INODELOCK_FULL here but we can't do
 986          * this now because a running HSM restore on the child (unlink
 987          * victim) will hold the layout lock. See LU-4002. */
 988         lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
 989         if (mdt_object_remote(mp)) {
 990                 /* Enqueue lookup lock from parent MDT */
 991                 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
 992                                             &child_lh->mlh_rreg_lh,
 993                                             child_lh->mlh_rreg_mode,
 994                                             MDS_INODELOCK_LOOKUP, false);
 995                 if (rc != ELDLM_OK)
 996                         GOTO(put_child, rc);
 997
 998                 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
 999         }
1000
1001         rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1002                                     cos_incompat);
1003         if (rc != 0)
1004                 GOTO(put_child, rc);
1005
1006         /*
1007          * Now we can only make sure we need MA_INODE, in mdd layer, will check
1008          * whether need MA_LOV and MA_COOKIE.
1009          */
1010         ma->ma_need = MA_INODE;
1011         ma->ma_valid = 0;
1012
1013         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1014                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1015         /* save version when object is locked */
1016         mdt_version_get_save(info, mc, 1);
1017
1018         mutex_lock(&mc->mot_lov_mutex);
1019
1020         rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1021                         mdt_object_child(mc), &rr->rr_name, ma, no_name);
1022
1023         mutex_unlock(&mc->mot_lov_mutex);
1024         if (rc != 0)
1025                 GOTO(unlock_child, rc);
1026
1027         if (!lu_object_is_dying(&mc->mot_header)) {
1028                 rc = mdt_attr_get_complex(info, mc, ma);
1029                 if (rc)
1030                         GOTO(out_stat, rc);
1031         } else if (mdt_dom_check_for_discard(info, mc)) {
1032                 mdt_dom_discard_data(info, mc);
1033         }
1034         mdt_handle_last_unlink(info, mc, ma);
1035
1036 out_stat:
1037         if (ma->ma_valid & MA_INODE) {
1038                 switch (ma->ma_attr.la_mode & S_IFMT) {
1039                 case S_IFDIR:
1040                         mdt_counter_incr(req, LPROC_MDT_RMDIR);
1041                         break;
1042                 case S_IFREG:
1043                 case S_IFLNK:
1044                 case S_IFCHR:
1045                 case S_IFBLK:
1046                 case S_IFIFO:
1047                 case S_IFSOCK:
1048                         mdt_counter_incr(req, LPROC_MDT_UNLINK);
1049                         break;
1050                 default:
1051                         LASSERTF(0, "bad file type %o unlinking\n",
1052                                 ma->ma_attr.la_mode);
1053                 }
1054         }
1055
1056         EXIT;
1057
1058 unlock_child:
1059         mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1060 put_child:
1061         mdt_object_put(info->mti_env, mc);
1062 unlock_parent:
1063         mdt_object_unlock(info, mp, parent_lh, rc);
1064 put_parent:
1065         mdt_object_put(info->mti_env, mp);
1066         CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1067         return rc;
1068 }
1069
1070 /*
1071  * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1072  * name.
1073  */
1074 static int mdt_reint_link(struct mdt_thread_info *info,
1075                           struct mdt_lock_handle *lhc)
1076 {
1077         struct mdt_reint_record *rr = &info->mti_rr;
1078         struct ptlrpc_request   *req = mdt_info_req(info);
1079         struct md_attr          *ma = &info->mti_attr;
1080         struct mdt_object       *ms;
1081         struct mdt_object       *mp;
1082         struct mdt_lock_handle  *lhs;
1083         struct mdt_lock_handle  *lhp;
1084         bool cos_incompat;
1085         int rc;
1086         ENTRY;
1087
1088         DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1089                   PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1090
1091         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1092                 RETURN(err_serious(-ENOENT));
1093
1094         if (info->mti_dlm_req)
1095                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1096
1097         /* Invalid case so return error immediately instead of
1098          * processing it */
1099         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1100                 RETURN(-EPERM);
1101
1102         if (!fid_is_md_operative(rr->rr_fid1) ||
1103             !fid_is_md_operative(rr->rr_fid2))
1104                 RETURN(-EPERM);
1105
1106         /* step 1: find target parent dir */
1107         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1108         if (IS_ERR(mp))
1109                 RETURN(PTR_ERR(mp));
1110
1111         rc = mdt_version_get_check_save(info, mp, 0);
1112         if (rc)
1113                 GOTO(put_parent, rc);
1114
1115         /* step 2: find source */
1116         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1117         if (IS_ERR(ms))
1118                 GOTO(put_parent, rc = PTR_ERR(ms));
1119
1120         if (!mdt_object_exists(ms)) {
1121                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1122                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1123                 GOTO(put_source, rc = -ENOENT);
1124         }
1125
1126         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1127
1128         lhp = &info->mti_lh[MDT_LH_PARENT];
1129         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1130         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1131                                    cos_incompat);
1132         if (rc != 0)
1133                 GOTO(put_source, rc);
1134
1135         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1136
1137         lhs = &info->mti_lh[MDT_LH_CHILD];
1138         mdt_lock_reg_init(lhs, LCK_EX);
1139         rc = mdt_reint_object_lock(info, ms, lhs,
1140                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1141                                    cos_incompat);
1142         if (rc != 0)
1143                 GOTO(unlock_parent, rc);
1144
1145         /* step 3: link it */
1146         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1147                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1148
1149         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1150         rc = mdt_version_get_check_save(info, ms, 1);
1151         if (rc)
1152                 GOTO(unlock_source, rc);
1153
1154         /** check target version by name during replay */
1155         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1156                                       &info->mti_tmp_fid1, 2);
1157         if (rc != 0 && rc != -ENOENT)
1158                 GOTO(unlock_source, rc);
1159         /* save version of file name for replay, it must be ENOENT here */
1160         if (!req_is_replay(mdt_info_req(info))) {
1161                 if (rc != -ENOENT) {
1162                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1163                                PNAME(&rr->rr_name));
1164                         GOTO(unlock_source, rc = -EEXIST);
1165                 }
1166                 info->mti_ver[2] = ENOENT_VERSION;
1167                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1168         }
1169
1170         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1171                       mdt_object_child(ms), &rr->rr_name, ma);
1172
1173         if (rc == 0)
1174                 mdt_counter_incr(req, LPROC_MDT_LINK);
1175
1176         EXIT;
1177 unlock_source:
1178         mdt_object_unlock(info, ms, lhs, rc);
1179 unlock_parent:
1180         mdt_object_unlock(info, mp, lhp, rc);
1181 put_source:
1182         mdt_object_put(info->mti_env, ms);
1183 put_parent:
1184         mdt_object_put(info->mti_env, mp);
1185         return rc;
1186 }
1187 /**
1188  * lock the part of the directory according to the hash of the name
1189  * (lh->mlh_pdo_hash) in parallel directory lock.
1190  */
1191 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1192                               struct mdt_lock_handle *lh,
1193                               struct mdt_object *obj, __u64 ibits,
1194                               bool cos_incompat)
1195 {
1196         struct ldlm_res_id *res = &info->mti_res_id;
1197         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1198         union ldlm_policy_data *policy = &info->mti_policy;
1199         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1200         int rc;
1201
1202         /*
1203          * Finish res_id initializing by name hash marking part of
1204          * directory which is taking modification.
1205          */
1206         LASSERT(lh->mlh_pdo_hash != 0);
1207         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1208         memset(policy, 0, sizeof(*policy));
1209         policy->l_inodebits.bits = ibits;
1210         if (cos_incompat &&
1211             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1212                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1213         /*
1214          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1215          * going to be sent to client. If it is - mdt_intent_policy() path will
1216          * fix it up and turn FL_LOCAL flag off.
1217          */
1218         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1219                           policy, res, dlmflags,
1220                           &info->mti_exp->exp_handle.h_cookie);
1221         return rc;
1222 }
1223
1224 /**
1225  * Get BFL lock for rename or migrate process.
1226  **/
1227 static int mdt_rename_lock(struct mdt_thread_info *info,
1228                            struct lustre_handle *lh)
1229 {
1230         int     rc;
1231         ENTRY;
1232
1233         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1234                 struct lu_fid *fid = &info->mti_tmp_fid1;
1235                 struct mdt_object *obj;
1236
1237                 /* XXX, right now, it has to use object API to
1238                  * enqueue lock cross MDT, so it will enqueue
1239                  * rename lock(with LUSTRE_BFL_FID) by root object */
1240                 lu_root_fid(fid);
1241                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1242                 if (IS_ERR(obj))
1243                         RETURN(PTR_ERR(obj));
1244
1245                 rc = mdt_remote_object_lock(info, obj,
1246                                             &LUSTRE_BFL_FID, lh,
1247                                             LCK_EX,
1248                                             MDS_INODELOCK_UPDATE, false);
1249                 mdt_object_put(info->mti_env, obj);
1250         } else {
1251                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1252                 union ldlm_policy_data *policy = &info->mti_policy;
1253                 struct ldlm_res_id *res_id = &info->mti_res_id;
1254                 __u64 flags = 0;
1255
1256                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1257                 memset(policy, 0, sizeof *policy);
1258                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1259                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1260                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1261                                             LDLM_IBITS, policy, LCK_EX, &flags,
1262                                             ldlm_blocking_ast,
1263                                             ldlm_completion_ast, NULL, NULL, 0,
1264                                             LVB_T_NONE,
1265                                             &info->mti_exp->exp_handle.h_cookie,
1266                                             lh);
1267                 RETURN(rc);
1268         }
1269         RETURN(rc);
1270 }
1271
1272 static void mdt_rename_unlock(struct lustre_handle *lh)
1273 {
1274         ENTRY;
1275         LASSERT(lustre_handle_is_used(lh));
1276         /* Cancel the single rename lock right away */
1277         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1278         EXIT;
1279 }
1280
1281 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1282                                                 const struct lu_fid *fid,
1283                                                 int idx)
1284 {
1285         struct mdt_object *dir;
1286         int rc;
1287
1288         ENTRY;
1289
1290         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1291         if (IS_ERR(dir))
1292                 RETURN(dir);
1293
1294         /* check early, the real version will be saved after locking */
1295         rc = mdt_version_get_check(info, dir, idx);
1296         if (rc)
1297                 GOTO(out_put, rc);
1298
1299         if (!mdt_object_exists(dir))
1300                 GOTO(out_put, rc = -ENOENT);
1301
1302         if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1303                 GOTO(out_put, rc = -ENOTDIR);
1304
1305         RETURN(dir);
1306 out_put:
1307         mdt_object_put(info->mti_env, dir);
1308         return ERR_PTR(rc);
1309 }
1310
1311 /*
1312  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1313  * herein we don't really check it, just do revoke.
1314  */
1315 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1316                                   struct mdt_object *pobj,
1317                                   struct mdt_object *obj)
1318 {
1319         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1320         int rc;
1321
1322         mdt_lock_handle_init(lh);
1323         mdt_lock_reg_init(lh, LCK_EX);
1324
1325         if (mdt_object_remote(pobj)) {
1326                 /* don't bother to check if pobj and obj are on the same MDT. */
1327                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1328                                             &lh->mlh_rreg_lh, LCK_EX,
1329                                             MDS_INODELOCK_LOOKUP, false);
1330         } else if (mdt_object_remote(obj)) {
1331                 struct ldlm_res_id *res = &info->mti_res_id;
1332                 union ldlm_policy_data *policy = &info->mti_policy;
1333                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1334                                  LDLM_FL_COS_INCOMPAT;
1335
1336                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1337                 memset(policy, 0, sizeof(*policy));
1338                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1339                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1340                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1341                                   dlmflags, NULL);
1342         } else {
1343                 /* do nothing if both are local */
1344                 return 0;
1345         }
1346
1347         if (rc != ELDLM_OK)
1348                 return rc;
1349
1350         /*
1351          * TODO, currently we don't save this lock because there is no place to
1352          * hold this lock handle, but to avoid race we need to save this lock.
1353          */
1354         mdt_object_unlock(info, NULL, lh, 1);
1355
1356         return 0;
1357 }
1358
1359 /*
1360  * operation may takes locks of linkea, or directory stripes, group them in
1361  * different list.
1362  */
1363 struct mdt_sub_lock {
1364         struct mdt_object      *msl_obj;
1365         struct mdt_lock_handle  msl_lh;
1366         struct list_head        msl_linkage;
1367 };
1368
1369 static void mdt_unlock_list(struct mdt_thread_info *info,
1370                             struct list_head *list, int decref)
1371 {
1372         struct mdt_sub_lock *msl;
1373         struct mdt_sub_lock *tmp;
1374
1375         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1376                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1377                 list_del(&msl->msl_linkage);
1378                 OBD_FREE_PTR(msl);
1379         }
1380 }
1381
1382 /*
1383  * lock parents of links, and also check whether total locks don't exceed
1384  * RS_MAX_LOCKS.
1385  *
1386  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1387  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1388  * \retval      -ev negative errno upon error
1389  */
1390 static int mdt_lock_links(struct mdt_thread_info *info,
1391                           struct mdt_object *pobj,
1392                           const struct md_attr *ma,
1393                           struct mdt_object *obj,
1394                           struct list_head *link_locks)
1395 {
1396         struct mdt_device *mdt = info->mti_mdt;
1397         struct lu_buf *buf = &info->mti_big_buf;
1398         struct lu_name *lname = &info->mti_name;
1399         struct linkea_data ldata = { NULL };
1400         bool blocked = false;
1401         int retries = 5;
1402         int local_lnkp_cnt = 0;
1403         int rc;
1404
1405         ENTRY;
1406
1407         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1408                 RETURN(0);
1409
1410         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1411         if (buf->lb_buf == NULL)
1412                 RETURN(-ENOMEM);
1413
1414         ldata.ld_buf = buf;
1415         rc = mdt_links_read(info, obj, &ldata);
1416         if (rc) {
1417                 if (rc == -ENOENT || rc == -ENODATA)
1418                         rc = 0;
1419                 RETURN(rc);
1420         }
1421
1422 repeat:
1423         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1424              linkea_next_entry(&ldata)) {
1425                 struct mdt_object *lnkp;
1426                 struct mdt_sub_lock *msl;
1427                 struct lu_fid fid;
1428                 __u64 ibits;
1429
1430                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1431                                     &fid);
1432
1433                 /* check if it's also linked to parent */
1434                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1435                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1436                                PFID(&fid), PNAME(lname));
1437                         /* in case link is remote object, revoke LOOKUP lock */
1438                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1439                         continue;
1440                 }
1441
1442                 lnkp = NULL;
1443
1444                 /* check if it's linked to a stripe of parent */
1445                 if (ma->ma_valid & MA_LMV) {
1446                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1447                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1448                         int j = 0;
1449
1450                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1451                                 fid_le_to_cpu(stripe_fid,
1452                                               &lmv->lmv_stripe_fids[j]);
1453                                 if (lu_fid_eq(stripe_fid, &fid)) {
1454                                         CDEBUG(D_INFO, "skip stripe "DFID
1455                                                ", reovke "DNAME"\n",
1456                                                PFID(&fid), PNAME(lname));
1457                                         lnkp = mdt_object_find(info->mti_env,
1458                                                                mdt, &fid);
1459                                         if (IS_ERR(lnkp))
1460                                                 GOTO(out, rc = PTR_ERR(lnkp));
1461                                         break;
1462                                 }
1463                         }
1464
1465                         if (lnkp) {
1466                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1467                                                                    obj);
1468                                 mdt_object_put(info->mti_env, lnkp);
1469                                 continue;
1470                         }
1471                 }
1472
1473                 /* Check if it's already locked */
1474                 list_for_each_entry(msl, link_locks, msl_linkage) {
1475                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1476                                 CDEBUG(D_INFO,
1477                                        DFID" was locked, revoke "DNAME"\n",
1478                                        PFID(&fid), PNAME(lname));
1479                                 lnkp = msl->msl_obj;
1480                                 break;
1481                         }
1482                 }
1483
1484                 if (lnkp) {
1485                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1486                         continue;
1487                 }
1488
1489                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1490                        PFID(&fid), PNAME(lname));
1491
1492                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1493                 if (IS_ERR(lnkp)) {
1494                         CWARN("%s: cannot find obj "DFID": %ld\n",
1495                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1496                         continue;
1497                 }
1498
1499                 if (!mdt_object_exists(lnkp)) {
1500                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1501                               PFID(&fid), PNAME(lname));
1502                         mdt_object_put(info->mti_env, lnkp);
1503                         continue;
1504                 }
1505
1506                 if (!mdt_object_remote(lnkp))
1507                         local_lnkp_cnt++;
1508
1509                 OBD_ALLOC_PTR(msl);
1510                 if (msl == NULL)
1511                         GOTO(out, rc = -ENOMEM);
1512
1513                 /*
1514                  * we can't follow parent-child lock order like other MD
1515                  * operations, use lock_try here to avoid deadlock, if the lock
1516                  * cannot be taken, drop all locks taken, revoke the blocked
1517                  * one, and continue processing the remaining entries, and in
1518                  * the end of the loop restart from beginning.
1519                  */
1520                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1521                 ibits = 0;
1522                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1523                                          MDS_INODELOCK_UPDATE, true);
1524                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1525                         blocked = true;
1526
1527                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME" retry %d\n",
1528                                PFID(&fid), PNAME(lname), retries);
1529
1530                         mdt_unlock_list(info, link_locks, 1);
1531
1532                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1533                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1534                                              MDS_INODELOCK_UPDATE);
1535                         if (rc) {
1536                                 mdt_object_put(info->mti_env, lnkp);
1537                                 OBD_FREE_PTR(msl);
1538                                 GOTO(out, rc);
1539                         }
1540
1541                         if (mdt_object_remote(lnkp)) {
1542                                 struct ldlm_lock *lock;
1543
1544                                 /*
1545                                  * for remote object, set lock cb_atomic,
1546                                  * so lock can be released in blocking_ast()
1547                                  * immediately, then the next lock_try will
1548                                  * have better chance of success.
1549                                  */
1550                                 lock = ldlm_handle2lock(
1551                                                 &msl->msl_lh.mlh_rreg_lh);
1552                                 LASSERT(lock != NULL);
1553                                 lock_res_and_lock(lock);
1554                                 ldlm_set_atomic_cb(lock);
1555                                 unlock_res_and_lock(lock);
1556                                 LDLM_LOCK_PUT(lock);
1557                         }
1558
1559                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1560                         OBD_FREE_PTR(msl);
1561                         continue;
1562                 }
1563
1564                 INIT_LIST_HEAD(&msl->msl_linkage);
1565                 msl->msl_obj = lnkp;
1566                 list_add_tail(&msl->msl_linkage, link_locks);
1567
1568                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1569         }
1570
1571         if (blocked) {
1572                 rc = -EBUSY;
1573                 if (--retries > 0) {
1574                         mdt_unlock_list(info, link_locks, rc);
1575                         blocked = false;
1576                         local_lnkp_cnt = 0;
1577                         goto repeat;
1578                 }
1579         }
1580
1581         EXIT;
1582 out:
1583         if (rc) {
1584                 mdt_unlock_list(info, link_locks, rc);
1585         } else if (local_lnkp_cnt > RS_MAX_LOCKS - 5) {
1586                 CDEBUG(D_INFO, "Too many links (%d), sync operations\n",
1587                        local_lnkp_cnt);
1588                 /*
1589                  * parent may have 3 local objects: master object and 2 stripes
1590                  * (if it's being migrated too); source may have 1 local object
1591                  * if regular file: master and 1 stripe; target has 1 local
1592                  * object.
1593                  * Note, if source is directory it may also have 2 local objects
1594                  * but can't has hardlinks, so consider only regular files here.
1595                  */
1596                 rc = 1;
1597         }
1598         return rc;
1599 }
1600
1601 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1602                                   struct mdt_object *obj,
1603                                   const struct md_attr *ma,
1604                                   struct list_head *slave_locks)
1605 {
1606         struct mdt_device *mdt = info->mti_mdt;
1607         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1608         struct lu_fid *fid = &info->mti_tmp_fid1;
1609         struct mdt_object *slave;
1610         struct mdt_sub_lock *msl;
1611         int i;
1612         int rc;
1613
1614         ENTRY;
1615
1616         LASSERT(mdt_object_remote(obj));
1617         LASSERT(ma->ma_valid & MA_LMV);
1618         LASSERT(lmv);
1619
1620         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1621                 RETURN(-EINVAL);
1622
1623         if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1624                 RETURN(0);
1625
1626         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1627                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1628
1629                 if (!fid_is_sane(fid))
1630                         continue;
1631
1632                 slave = mdt_object_find(info->mti_env, mdt, fid);
1633                 if (IS_ERR(slave))
1634                         GOTO(out, rc = PTR_ERR(slave));
1635
1636                 OBD_ALLOC_PTR(msl);
1637                 if (!msl) {
1638                         mdt_object_put(info->mti_env, slave);
1639                         GOTO(out, rc = -ENOMEM);
1640                 }
1641
1642                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1643                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1644                                            MDS_INODELOCK_UPDATE, true);
1645                 if (rc) {
1646                         OBD_FREE_PTR(msl);
1647                         mdt_object_put(info->mti_env, slave);
1648                         GOTO(out, rc);
1649                 }
1650
1651                 INIT_LIST_HEAD(&msl->msl_linkage);
1652                 msl->msl_obj = slave;
1653                 list_add_tail(&msl->msl_linkage, slave_locks);
1654         }
1655         EXIT;
1656
1657 out:
1658         if (rc)
1659                 mdt_unlock_list(info, slave_locks, rc);
1660         return rc;
1661 }
1662
1663 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1664                                              struct mdt_object *obj,
1665                                              struct mdt_lock_handle *lh,
1666                                              struct ldlm_enqueue_info *einfo,
1667                                              struct list_head *slave_locks,
1668                                              int decref)
1669 {
1670         if (mdt_object_remote(obj)) {
1671                 mdt_unlock_list(info, slave_locks, decref);
1672                 mdt_object_unlock(info, obj, lh, decref);
1673         } else {
1674                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1675         }
1676 }
1677
1678 /* lock parent and its stripes */
1679 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1680                                    struct mdt_object *obj,
1681                                    const struct md_attr *ma,
1682                                    struct mdt_lock_handle *lh,
1683                                    struct ldlm_enqueue_info *einfo,
1684                                    struct list_head *slave_locks)
1685 {
1686         int rc;
1687
1688         if (mdt_object_remote(obj)) {
1689                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1690                                             &lh->mlh_rreg_lh, LCK_PW,
1691                                             MDS_INODELOCK_UPDATE, false);
1692                 if (rc != ELDLM_OK)
1693                         return rc;
1694
1695                 /*
1696                  * if obj is remote and striped, lock its stripes explicitly
1697                  * because it's not striped in LOD layer on this MDT.
1698                  */
1699                 if (ma->ma_valid & MA_LMV) {
1700                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1701                         if (rc)
1702                                 mdt_object_unlock(info, obj, lh, rc);
1703                 }
1704         } else {
1705                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1706                                             einfo, true);
1707         }
1708
1709         return rc;
1710 }
1711
1712 /*
1713  * in migration, object may be remote, and we need take full lock of it and its
1714  * stripes if it's directory, besides, object may be a remote object on its
1715  * parent, revoke its LOOKUP lock on where its parent is located.
1716  */
1717 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1718                                    struct mdt_object *pobj,
1719                                    struct mdt_object *obj,
1720                                    struct mdt_lock_handle *lh,
1721                                    struct ldlm_enqueue_info *einfo,
1722                                    struct list_head *slave_locks)
1723 {
1724         int rc;
1725
1726         if (mdt_object_remote(obj)) {
1727                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1728                 if (rc)
1729                         return rc;
1730
1731                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1732                                             &lh->mlh_rreg_lh, LCK_EX,
1733                                             MDS_INODELOCK_FULL, false);
1734                 if (rc != ELDLM_OK)
1735                         return rc;
1736
1737                 /*
1738                  * if obj is remote and striped, lock its stripes explicitly
1739                  * because it's not striped in LOD layer on this MDT.
1740                  */
1741                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1742                         struct md_attr *ma = &info->mti_attr;
1743
1744                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1745                         if (rc) {
1746                                 mdt_object_unlock(info, obj, lh, rc);
1747                                 return rc;
1748                         }
1749
1750                         if (ma->ma_valid & MA_LMV) {
1751                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1752                                                             slave_locks);
1753                                 if (rc)
1754                                         mdt_object_unlock(info, obj, lh, rc);
1755                         }
1756                 }
1757         } else {
1758                 if (mdt_object_remote(pobj)) {
1759                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1760                         if (rc)
1761                                 return rc;
1762                 }
1763
1764                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1765                                             einfo, true);
1766         }
1767
1768         return rc;
1769 }
1770
1771 /*
1772  * lookup source by name, if parent is striped directory, we need to find the
1773  * corresponding stripe where source is located, and then lookup there.
1774  *
1775  * besides, if parent is migrating too, and file is already in target stripe,
1776  * this should be a redo of 'lfs migrate' on client side.
1777  */
1778 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1779                               struct mdt_object *pobj,
1780                               const struct md_attr *ma,
1781                               const struct lu_name *lname,
1782                               struct mdt_object **spobj,
1783                               struct mdt_object **sobj)
1784 {
1785         const struct lu_env *env = info->mti_env;
1786         struct lu_fid *fid = &info->mti_tmp_fid1;
1787         struct mdt_object *stripe;
1788         int rc;
1789
1790         if (ma->ma_valid & MA_LMV) {
1791                 /* if parent is striped, lookup on corresponding stripe */
1792                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1793                 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1794                 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1795                 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1796                                     LMV_HASH_FLAG_MIGRATION;
1797
1798                 if (is_migrating) {
1799                         hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1800                         stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1801                 }
1802
1803                 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1804                                               lname->ln_name,
1805                                               lname->ln_namelen);
1806                 if (rc < 0)
1807                         return rc;
1808
1809                 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1810                         rc += le32_to_cpu(lmv->lmv_migrate_offset);
1811
1812                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1813
1814                 stripe = mdt_object_find(env, info->mti_mdt, fid);
1815                 if (IS_ERR(stripe))
1816                         return PTR_ERR(stripe);
1817
1818                 fid_zero(fid);
1819                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1820                                 &info->mti_spec);
1821                 if (rc == -ENOENT && is_migrating) {
1822                         /*
1823                          * if parent is migrating, and lookup child failed on
1824                          * source stripe, lookup again on target stripe, if it
1825                          * exists, it means previous migration was interrupted,
1826                          * and current file was migrated already.
1827                          */
1828                         mdt_object_put(env, stripe);
1829
1830                         hash_type = le32_to_cpu(lmv->lmv_hash_type);
1831                         stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1832
1833                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1834                                                       lname->ln_name,
1835                                                       lname->ln_namelen);
1836                         if (rc < 0)
1837                                 return rc;
1838
1839                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1840
1841                         stripe = mdt_object_find(env, info->mti_mdt, fid);
1842                         if (IS_ERR(stripe))
1843                                 return PTR_ERR(stripe);
1844
1845                         fid_zero(fid);
1846                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1847                                         fid, &info->mti_spec);
1848                         mdt_object_put(env, stripe);
1849                         return rc ?: -EALREADY;
1850                 } else if (rc) {
1851                         mdt_object_put(env, stripe);
1852                         return rc;
1853                 }
1854         } else {
1855                 fid_zero(fid);
1856                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1857                                 &info->mti_spec);
1858                 if (rc)
1859                         return rc;
1860
1861                 stripe = pobj;
1862                 mdt_object_get(env, stripe);
1863         }
1864
1865         *spobj = stripe;
1866
1867         *sobj = mdt_object_find(env, info->mti_mdt, fid);
1868         if (IS_ERR(*sobj)) {
1869                 mdt_object_put(env, stripe);
1870                 rc = PTR_ERR(*sobj);
1871                 *spobj = NULL;
1872                 *sobj = NULL;
1873         }
1874
1875         return rc;
1876 }
1877
1878 /* end lease and close file for regular file */
1879 static int mdd_migrate_close(struct mdt_thread_info *info,
1880                              struct mdt_object *obj)
1881 {
1882         struct close_data *data;
1883         struct mdt_body *repbody;
1884         struct ldlm_lock *lease;
1885         int rc;
1886         int rc2;
1887
1888         rc = -EPROTO;
1889         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1890                                       RCL_CLIENT) ||
1891             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1892                                       RCL_CLIENT))
1893                 goto close;
1894
1895         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1896         if (!data)
1897                 goto close;
1898
1899         rc = -ESTALE;
1900         lease = ldlm_handle2lock(&data->cd_handle);
1901         if (!lease)
1902                 goto close;
1903
1904         /* check if the lease was already canceled */
1905         lock_res_and_lock(lease);
1906         rc = ldlm_is_cancel(lease);
1907         unlock_res_and_lock(lease);
1908
1909         if (rc) {
1910                 rc = -EAGAIN;
1911                 LDLM_DEBUG(lease, DFID" lease broken",
1912                            PFID(mdt_object_fid(obj)));
1913         }
1914
1915         /*
1916          * cancel server side lease, client side counterpart should have been
1917          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1918          */
1919         ldlm_lock_cancel(lease);
1920         ldlm_reprocess_all(lease->l_resource, lease);
1921         LDLM_LOCK_PUT(lease);
1922
1923 close:
1924         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1925         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1926         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1927
1928         return rc ?: rc2;
1929 }
1930
1931 /*
1932  * migrate file in below steps:
1933  *  1. lock parent and its stripes
1934  *  2. lookup source by name
1935  *  3. lock parents of source links if source is not directory
1936  *  4. reject if source is in HSM
1937  *  5. take source open_sem and close file if source is regular file
1938  *  6. lock source and its stripes if it's directory
1939  *  7. lock target so subsequent change to it can trigger COS
1940  *  8. migrate file
1941  *  9. unlock above locks
1942  * 10. sync device if source has links
1943  */
1944 static int mdt_reint_migrate(struct mdt_thread_info *info,
1945                              struct mdt_lock_handle *unused)
1946 {
1947         const struct lu_env *env = info->mti_env;
1948         struct mdt_device *mdt = info->mti_mdt;
1949         struct ptlrpc_request *req = mdt_info_req(info);
1950         struct mdt_reint_record *rr = &info->mti_rr;
1951         struct lu_ucred *uc = mdt_ucred(info);
1952         struct md_attr *ma = &info->mti_attr;
1953         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1954         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1955         struct mdt_object *pobj;
1956         struct mdt_object *spobj = NULL;
1957         struct mdt_object *sobj = NULL;
1958         struct mdt_object *tobj;
1959         struct lustre_handle rename_lh = { 0 };
1960         struct mdt_lock_handle *lhp;
1961         struct mdt_lock_handle *lhs;
1962         struct mdt_lock_handle *lht;
1963         LIST_HEAD(parent_slave_locks);
1964         LIST_HEAD(child_slave_locks);
1965         LIST_HEAD(link_locks);
1966         bool open_sem_locked = false;
1967         bool do_sync = false;
1968         int rc;
1969         ENTRY;
1970
1971         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
1972                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
1973
1974         if (info->mti_dlm_req)
1975                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1976
1977         if (!fid_is_md_operative(rr->rr_fid1) ||
1978             !fid_is_md_operative(rr->rr_fid2))
1979                 RETURN(-EPERM);
1980
1981         /* don't allow migrate . or .. */
1982         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
1983                 RETURN(-EBUSY);
1984
1985         if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
1986                 RETURN(-EPERM);
1987
1988         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
1989             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
1990             mdt->mdt_enable_remote_dir_gid != -1)
1991                 RETURN(-EPERM);
1992
1993         /*
1994          * Note: do not enqueue rename lock for replay request, because
1995          * if other MDT holds rename lock, but being blocked to wait for
1996          * this MDT to finish its recovery, and the failover MDT can not
1997          * get rename lock, which will cause deadlock.
1998          */
1999         if (!req_is_replay(req)) {
2000                 rc = mdt_rename_lock(info, &rename_lh);
2001                 if (rc != 0) {
2002                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2003                                mdt_obd_name(info->mti_mdt), rc);
2004                         RETURN(rc);
2005                 }
2006         }
2007
2008         /* pobj is master object of parent */
2009         pobj = mdt_parent_find_check(info, rr->rr_fid1, 0);
2010         if (IS_ERR(pobj))
2011                 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2012
2013         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2014         if (rc)
2015                 GOTO(put_parent, rc);
2016
2017         /* lock parent object */
2018         lhp = &info->mti_lh[MDT_LH_PARENT];
2019         mdt_lock_reg_init(lhp, LCK_PW);
2020         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2021                                      &parent_slave_locks);
2022         if (rc)
2023                 GOTO(put_parent, rc);
2024
2025         /*
2026          * spobj is the corresponding stripe against name if pobj is striped
2027          * directory, which is the real parent, and no need to lock, because
2028          * we've taken full lock of pobj.
2029          */
2030         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2031         if (rc)
2032                 GOTO(unlock_parent, rc);
2033
2034         /* lock parents of source links, and revoke LOOKUP lock of links */
2035         rc = mdt_lock_links(info, pobj, ma, sobj, &link_locks);
2036         if (rc < 0)
2037                 GOTO(put_source, rc);
2038
2039         /*
2040          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2041          * with one request, if total lock count exceeds this limit, we will
2042          * drop all locks after migration, and synchronous device in the end.
2043          */
2044         do_sync = rc;
2045
2046         /* TODO: DoM migration is not supported yet */
2047         if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2048                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2049                 if (rc)
2050                         GOTO(put_source, rc);
2051
2052                 if (ma->ma_valid & MA_LOV &&
2053                     mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2054                         GOTO(put_source, rc = -EOPNOTSUPP);
2055         }
2056
2057         /* if migration HSM is allowed */
2058         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2059                 ma->ma_need = MA_HSM;
2060                 ma->ma_valid = 0;
2061                 rc = mdt_attr_get_complex(info, sobj, ma);
2062                 if (rc)
2063                         GOTO(unlock_links, rc);
2064
2065                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2066                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2067         }
2068
2069         /* end lease and close file for regular file */
2070         if (info->mti_spec.sp_migrate_close) {
2071                 /* try to hold open_sem so that nobody else can open the file */
2072                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2073                         /* close anyway */
2074                         mdd_migrate_close(info, sobj);
2075                         GOTO(unlock_links, rc = -EBUSY);
2076                 } else {
2077                         open_sem_locked = true;
2078                         rc = mdd_migrate_close(info, sobj);
2079                         if (rc)
2080                                 GOTO(unlock_open_sem, rc);
2081                 }
2082         }
2083
2084         /* lock source */
2085         lhs = &info->mti_lh[MDT_LH_OLD];
2086         mdt_lock_reg_init(lhs, LCK_EX);
2087         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2088                                      &child_slave_locks);
2089         if (rc)
2090                 GOTO(unlock_open_sem, rc);
2091
2092         /* lock target */
2093         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2094         if (IS_ERR(tobj))
2095                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2096
2097         lht = &info->mti_lh[MDT_LH_NEW];
2098         mdt_lock_reg_init(lht, LCK_EX);
2099         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2100         if (rc)
2101                 GOTO(put_target, rc);
2102
2103         /* Don't do lookup sanity check. We know name doesn't exist. */
2104         info->mti_spec.sp_cr_lookup = 0;
2105         info->mti_spec.sp_feat = &dt_directory_features;
2106
2107         rc = mdo_migrate(env, mdt_object_child(pobj),
2108                          mdt_object_child(sobj), &rr->rr_name,
2109                          mdt_object_child(tobj), &info->mti_spec, ma);
2110         EXIT;
2111
2112         mdt_object_unlock(info, tobj, lht, rc || do_sync);
2113 put_target:
2114         mdt_object_put(env, tobj);
2115 unlock_source:
2116         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2117                                   &child_slave_locks, rc || do_sync);
2118 unlock_open_sem:
2119         if (open_sem_locked)
2120                 up_write(&sobj->mot_open_sem);
2121 unlock_links:
2122         mdt_unlock_list(info, &link_locks, rc || do_sync);
2123 put_source:
2124         mdt_object_put(env, sobj);
2125         mdt_object_put(env, spobj);
2126 unlock_parent:
2127         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2128                                   &parent_slave_locks, rc || do_sync);
2129 put_parent:
2130         mdt_object_put(env, pobj);
2131 unlock_rename:
2132         if (lustre_handle_is_used(&rename_lh))
2133                 mdt_rename_unlock(&rename_lh);
2134
2135         if (!rc && do_sync)
2136                 mdt_device_sync(env, mdt);
2137
2138         return rc;
2139 }
2140
2141 static int mdt_object_lock_save(struct mdt_thread_info *info,
2142                                 struct mdt_object *dir,
2143                                 struct mdt_lock_handle *lh,
2144                                 int idx, bool cos_incompat)
2145 {
2146         int rc;
2147
2148         /* we lock the target dir if it is local */
2149         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2150                                    cos_incompat);
2151         if (rc != 0)
2152                 return rc;
2153
2154         /* get and save correct version after locking */
2155         mdt_version_get_save(info, dir, idx);
2156         return 0;
2157 }
2158
2159 /*
2160  * determine lock order of sobj and tobj
2161  *
2162  * there are two situations we need to lock tobj before sobj:
2163  * 1. sobj is child of tobj
2164  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2165  *    larger than that of tobj
2166  *
2167  * \retval      1 lock tobj before sobj
2168  * \retval      0 lock sobj before tobj
2169  * \retval      -ev negative errno upon error
2170  */
2171 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2172                                            struct mdt_object *sobj,
2173                                            struct mdt_object *tobj)
2174 {
2175         struct md_attr *ma = &info->mti_attr;
2176         struct lu_fid *spfid = &info->mti_tmp_fid1;
2177         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2178         struct lmv_mds_md_v1 *lmv;
2179         __u32 sindex;
2180         __u32 tindex;
2181         int rc;
2182
2183         /* sobj and tobj are the same */
2184         if (sobj == tobj)
2185                 return 0;
2186
2187         if (fid_is_root(mdt_object_fid(sobj)))
2188                 return 0;
2189
2190         if (fid_is_root(mdt_object_fid(tobj)))
2191                 return 1;
2192
2193         /* check whether sobj is child of tobj */
2194         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2195                            mdt_object_fid(tobj));
2196         if (rc < 0)
2197                 return rc;
2198
2199         if (rc == 1)
2200                 return 1;
2201
2202         /* check whether sobj and tobj are children of the same parent */
2203         rc = mdt_attr_get_pfid(info, sobj, spfid);
2204         if (rc)
2205                 return rc;
2206
2207         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2208         if (rc)
2209                 return rc;
2210
2211         if (!lu_fid_eq(spfid, tpfid))
2212                 return 0;
2213
2214         /* check whether sobj and tobj are sibling stripes */
2215         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2216         if (rc)
2217                 return rc;
2218
2219         if (!(ma->ma_valid & MA_LMV))
2220                 return 0;
2221
2222         lmv = &ma->ma_lmv->lmv_md_v1;
2223         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2224                 return 0;
2225         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2226
2227         ma->ma_valid = 0;
2228         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2229         if (rc)
2230                 return rc;
2231
2232         if (!(ma->ma_valid & MA_LMV))
2233                 return -ENODATA;
2234
2235         lmv = &ma->ma_lmv->lmv_md_v1;
2236         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2237                 return -EINVAL;
2238         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2239
2240         /* check stripe index of sobj and tobj */
2241         if (sindex == tindex)
2242                 return -EINVAL;
2243
2244         return sindex < tindex ? 0 : 1;
2245 }
2246
2247 /*
2248  * lock rename source object.
2249  *
2250  * Both source and source parent may be remote, and source may be a remote
2251  * object on source parent, to avoid overriding lock handle, store remote
2252  * LOOKUP lock separately in @lhr.
2253  *
2254  * \retval      0 on success
2255  * \retval      -ev negative errno upon error
2256  */
2257 static int mdt_rename_source_lock(struct mdt_thread_info *info,
2258                                   struct mdt_object *parent,
2259                                   struct mdt_object *child,
2260                                   struct mdt_lock_handle *lhc,
2261                                   struct mdt_lock_handle *lhr,
2262                                   __u64 ibits,
2263                                   bool cos_incompat)
2264 {
2265         int rc;
2266
2267         rc = mdt_is_remote_object(info, parent, child);
2268         if (rc < 0)
2269                 return rc;
2270
2271         if (rc) {
2272                 /* enqueue remote LOOKUP lock from the parent MDT */
2273                 __u64 rmt_ibits = MDS_INODELOCK_LOOKUP;
2274
2275                 if (mdt_object_remote(parent)) {
2276                         rc = mdt_remote_object_lock(info, parent,
2277                                                     mdt_object_fid(child),
2278                                                     &lhr->mlh_rreg_lh,
2279                                                     lhr->mlh_rreg_mode,
2280                                                     rmt_ibits, false);
2281                         if (rc != ELDLM_OK)
2282                                 return rc;
2283                 } else {
2284                         LASSERT(mdt_object_remote(child));
2285                         rc = mdt_object_local_lock(info, child, lhr,
2286                                                    &rmt_ibits, 0, true);
2287                         if (rc < 0)
2288                                 return rc;
2289                 }
2290
2291                 ibits &= ~MDS_INODELOCK_LOOKUP;
2292         }
2293
2294         if (mdt_object_remote(child)) {
2295                 rc = mdt_remote_object_lock(info, child, mdt_object_fid(child),
2296                                             &lhc->mlh_rreg_lh,
2297                                             lhc->mlh_rreg_mode,
2298                                             ibits, false);
2299                 if (rc == ELDLM_OK)
2300                         rc = 0;
2301         } else {
2302                 rc = mdt_reint_object_lock(info, child, lhc, ibits,
2303                                            cos_incompat);
2304         }
2305
2306         if (!rc)
2307                 mdt_object_unlock(info, child, lhr, rc);
2308
2309         return rc;
2310 }
2311
2312 /*
2313  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2314  * 2 - srcdir child; 3 - tgtdir child.
2315  * Update on disk version of srcdir child.
2316  */
2317 static int mdt_reint_rename(struct mdt_thread_info *info,
2318                             struct mdt_lock_handle *unused)
2319 {
2320         struct mdt_device *mdt = info->mti_mdt;
2321         struct mdt_reint_record *rr = &info->mti_rr;
2322         struct md_attr *ma = &info->mti_attr;
2323         struct ptlrpc_request *req = mdt_info_req(info);
2324         struct mdt_object *msrcdir = NULL;
2325         struct mdt_object *mtgtdir = NULL;
2326         struct mdt_object *mold;
2327         struct mdt_object *mnew = NULL;
2328         struct lustre_handle rename_lh = { 0 };
2329         struct mdt_lock_handle *lh_srcdirp;
2330         struct mdt_lock_handle *lh_tgtdirp;
2331         struct mdt_lock_handle *lh_oldp = NULL;
2332         struct mdt_lock_handle *lh_rmt = NULL;
2333         struct mdt_lock_handle *lh_newp = NULL;
2334         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2335         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2336         __u64 lock_ibits;
2337         bool reverse = false, discard = false;
2338         bool cos_incompat;
2339         int rc;
2340         ENTRY;
2341
2342         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2343                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2344                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2345
2346         if (info->mti_dlm_req)
2347                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2348
2349         if (!fid_is_md_operative(rr->rr_fid1) ||
2350             !fid_is_md_operative(rr->rr_fid2))
2351                 RETURN(-EPERM);
2352
2353         /* find both parents. */
2354         msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2355         if (IS_ERR(msrcdir))
2356                 RETURN(PTR_ERR(msrcdir));
2357
2358         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2359
2360         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2361                 mtgtdir = msrcdir;
2362                 mdt_object_get(info->mti_env, mtgtdir);
2363         } else {
2364                 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2365                 if (IS_ERR(mtgtdir))
2366                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2367         }
2368
2369         /*
2370          * Note: do not enqueue rename lock for replay request, because
2371          * if other MDT holds rename lock, but being blocked to wait for
2372          * this MDT to finish its recovery, and the failover MDT can not
2373          * get rename lock, which will cause deadlock.
2374          */
2375         if (!req_is_replay(req)) {
2376                 /*
2377                  * Normally rename RPC is handled on the MDT with the target
2378                  * directory (if target exists, it's on the MDT with the
2379                  * target), if the source directory is remote, it's a hint that
2380                  * source is remote too (this may not be true, but it won't
2381                  * cause any issue), return -EXDEV early to avoid taking
2382                  * rename_lock.
2383                  */
2384                 if (!mdt->mdt_enable_remote_rename &&
2385                     mdt_object_remote(msrcdir))
2386                         GOTO(out_put_tgtdir, rc = -EXDEV);
2387
2388                 rc = mdt_rename_lock(info, &rename_lh);
2389                 if (rc != 0) {
2390                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2391                                mdt_obd_name(mdt), rc);
2392                         GOTO(out_put_tgtdir, rc);
2393                 }
2394         }
2395
2396         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2397         if (rc < 0)
2398                 GOTO(out_unlock_rename, rc);
2399
2400         reverse = rc;
2401
2402         /* source needs to be looked up after locking source parent, otherwise
2403          * this rename may race with unlink source, and cause rename hang, see
2404          * sanityn.sh 55b, so check parents first, if later we found source is
2405          * remote, relock parents. */
2406         cos_incompat = (mdt_object_remote(msrcdir) ||
2407                         mdt_object_remote(mtgtdir));
2408
2409         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2410
2411         /* lock parents in the proper order. */
2412         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2413         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2414
2415         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN);
2416         OBD_RACE(OBD_FAIL_MDS_REINT_OPEN2);
2417 relock:
2418         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2419         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2420
2421         if (reverse) {
2422                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2423                                           cos_incompat);
2424                 if (rc)
2425                         GOTO(out_unlock_rename, rc);
2426
2427                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2428
2429                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2430                                           cos_incompat);
2431                 if (rc != 0) {
2432                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2433                         GOTO(out_unlock_rename, rc);
2434                 }
2435         } else {
2436                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2437                                           cos_incompat);
2438                 if (rc)
2439                         GOTO(out_unlock_rename, rc);
2440
2441                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2442
2443                 if (mtgtdir != msrcdir) {
2444                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2445                                                   cos_incompat);
2446                 } else if (!mdt_object_remote(mtgtdir) &&
2447                            lh_srcdirp->mlh_pdo_hash !=
2448                            lh_tgtdirp->mlh_pdo_hash) {
2449                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2450                                                 MDS_INODELOCK_UPDATE,
2451                                                 cos_incompat);
2452                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2453                 }
2454                 if (rc != 0) {
2455                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2456                         GOTO(out_unlock_rename, rc);
2457                 }
2458         }
2459
2460         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2461         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2462
2463         /* find mold object. */
2464         fid_zero(old_fid);
2465         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2466         if (rc != 0)
2467                 GOTO(out_unlock_parents, rc);
2468
2469         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2470                 GOTO(out_unlock_parents, rc = -EINVAL);
2471
2472         if (!fid_is_md_operative(old_fid))
2473                 GOTO(out_unlock_parents, rc = -EPERM);
2474
2475         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2476         if (IS_ERR(mold))
2477                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2478
2479         if (!mdt_object_exists(mold)) {
2480                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2481                                 &mold->mot_obj,
2482                                 "object does not exist");
2483                 GOTO(out_put_old, rc = -ENOENT);
2484         }
2485
2486         if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2487                 GOTO(out_put_old, rc = -EXDEV);
2488
2489         /* Check if @mtgtdir is subdir of @mold, before locking child
2490          * to avoid reverse locking. */
2491         if (mtgtdir != msrcdir) {
2492                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2493                                    old_fid);
2494                 if (rc) {
2495                         if (rc == 1)
2496                                 rc = -EINVAL;
2497                         GOTO(out_put_old, rc);
2498                 }
2499         }
2500
2501         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2502         /* save version after locking */
2503         mdt_version_get_save(info, mold, 2);
2504
2505         if (!cos_incompat && mdt_object_remote(mold)) {
2506                 cos_incompat = true;
2507                 mdt_object_put(info->mti_env, mold);
2508                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2509                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2510                 goto relock;
2511         }
2512
2513         /* find mnew object:
2514          * mnew target object may not exist now
2515          * lookup with version checking */
2516         fid_zero(new_fid);
2517         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2518                                       3);
2519         if (rc == 0) {
2520                 /* the new_fid should have been filled at this moment */
2521                 if (lu_fid_eq(old_fid, new_fid))
2522                         GOTO(out_put_old, rc);
2523
2524                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2525                     lu_fid_eq(new_fid, rr->rr_fid2))
2526                         GOTO(out_put_old, rc = -EINVAL);
2527
2528                 if (!fid_is_md_operative(new_fid))
2529                         GOTO(out_put_old, rc = -EPERM);
2530
2531                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2532                 if (IS_ERR(mnew))
2533                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2534
2535                 if (!mdt_object_exists(mnew)) {
2536                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2537                                         &mnew->mot_obj,
2538                                         "object does not exist");
2539                         GOTO(out_put_new, rc = -ENOENT);
2540                 }
2541
2542                 if (mdt_object_remote(mnew)) {
2543                         struct mdt_body  *repbody;
2544
2545                         /* Always send rename req to the target child MDT */
2546                         repbody = req_capsule_server_get(info->mti_pill,
2547                                                          &RMF_MDT_BODY);
2548                         LASSERT(repbody != NULL);
2549                         repbody->mbo_fid1 = *new_fid;
2550                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2551                         GOTO(out_put_new, rc = -EXDEV);
2552                 }
2553                 /* Before locking the target dir, check we do not replace
2554                  * a dir with a non-dir, otherwise it may deadlock with
2555                  * link op which tries to create a link in this dir
2556                  * back to this non-dir. */
2557                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2558                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2559                         GOTO(out_put_new, rc = -EISDIR);
2560
2561                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2562                 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2563                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2564                 mdt_lock_reg_init(lh_rmt, LCK_EX);
2565                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2566                 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2567                                             lh_rmt, lock_ibits, cos_incompat);
2568                 if (rc < 0)
2569                         GOTO(out_put_new, rc);
2570
2571                 /* Check if @msrcdir is subdir of @mnew, before locking child
2572                  * to avoid reverse locking. */
2573                 if (mtgtdir != msrcdir) {
2574                         rc = mdo_is_subdir(info->mti_env,
2575                                            mdt_object_child(msrcdir), new_fid);
2576                         if (rc) {
2577                                 if (rc == 1)
2578                                         rc = -EINVAL;
2579                                 GOTO(out_unlock_old, rc);
2580                         }
2581                 }
2582
2583                 /* We used to acquire MDS_INODELOCK_FULL here but we
2584                  * can't do this now because a running HSM restore on
2585                  * the rename onto victim will hold the layout
2586                  * lock. See LU-4002. */
2587
2588                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2589                 mdt_lock_reg_init(lh_newp, LCK_EX);
2590                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
2591                 if (mdt_object_remote(mtgtdir)) {
2592                         rc = mdt_remote_object_lock(info, mtgtdir,
2593                                                     mdt_object_fid(mnew),
2594                                                     &lh_newp->mlh_rreg_lh,
2595                                                     lh_newp->mlh_rreg_mode,
2596                                                     MDS_INODELOCK_LOOKUP,
2597                                                     false);
2598                         if (rc != ELDLM_OK)
2599                                 GOTO(out_unlock_old, rc);
2600
2601                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2602                 }
2603                 rc = mdt_reint_object_lock(info, mnew, lh_newp, lock_ibits,
2604                                            cos_incompat);
2605                 if (rc != 0)
2606                         GOTO(out_unlock_new, rc);
2607
2608                 /* get and save version after locking */
2609                 mdt_version_get_save(info, mnew, 3);
2610         } else if (rc != -ENOENT) {
2611                 GOTO(out_put_old, rc);
2612         } else {
2613                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2614                 lh_rmt = &info->mti_lh[MDT_LH_RMT];
2615                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2616                 mdt_lock_reg_init(lh_rmt, LCK_EX);
2617                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2618                 rc = mdt_rename_source_lock(info, msrcdir, mold, lh_oldp,
2619                                             lh_rmt, lock_ibits, cos_incompat);
2620                 if (rc != 0)
2621                         GOTO(out_put_old, rc);
2622
2623                 mdt_enoent_version_save(info, 3);
2624         }
2625
2626         /* step 5: rename it */
2627         mdt_reint_init_ma(info, ma);
2628
2629         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2630                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2631
2632         if (mnew != NULL)
2633                 mutex_lock(&mnew->mot_lov_mutex);
2634
2635         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2636                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2637                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2638                         &rr->rr_tgt_name, ma);
2639
2640         if (mnew != NULL)
2641                 mutex_unlock(&mnew->mot_lov_mutex);
2642
2643         /* handle last link of tgt object */
2644         if (rc == 0) {
2645                 mdt_counter_incr(req, LPROC_MDT_RENAME);
2646                 if (mnew) {
2647                         mdt_handle_last_unlink(info, mnew, ma);
2648                         discard = mdt_dom_check_for_discard(info, mnew);
2649                 }
2650                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2651                                          msrcdir, mtgtdir);
2652         }
2653
2654         EXIT;
2655 out_unlock_new:
2656         if (mnew != NULL)
2657                 mdt_object_unlock(info, mnew, lh_newp, rc);
2658 out_unlock_old:
2659         mdt_object_unlock(info, NULL, lh_rmt, rc);
2660         mdt_object_unlock(info, mold, lh_oldp, rc);
2661 out_put_new:
2662         if (mnew && !discard)
2663                 mdt_object_put(info->mti_env, mnew);
2664 out_put_old:
2665         mdt_object_put(info->mti_env, mold);
2666 out_unlock_parents:
2667         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2668         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2669 out_unlock_rename:
2670         if (lustre_handle_is_used(&rename_lh))
2671                 mdt_rename_unlock(&rename_lh);
2672 out_put_tgtdir:
2673         mdt_object_put(info->mti_env, mtgtdir);
2674 out_put_srcdir:
2675         mdt_object_put(info->mti_env, msrcdir);
2676
2677         /* The DoM discard can be done right in the place above where it is
2678          * assigned, meanwhile it is done here after rename unlock due to
2679          * compatibility with old clients, for them the discard blocks
2680          * the main thread until completion. Check LU-11359 for details.
2681          */
2682         if (discard) {
2683                 mdt_dom_discard_data(info, mnew);
2684                 mdt_object_put(info->mti_env, mnew);
2685         }
2686         return rc;
2687 }
2688
2689 static int mdt_reint_resync(struct mdt_thread_info *info,
2690                             struct mdt_lock_handle *lhc)
2691 {
2692         struct mdt_reint_record *rr = &info->mti_rr;
2693         struct ptlrpc_request *req = mdt_info_req(info);
2694         struct md_attr *ma = &info->mti_attr;
2695         struct mdt_object *mo;
2696         struct ldlm_lock *lease;
2697         struct mdt_body *repbody;
2698         struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2699         bool lease_broken;
2700         int rc, rc2;
2701
2702         ENTRY;
2703
2704         DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2705
2706         if (info->mti_dlm_req)
2707                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2708
2709         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2710         if (IS_ERR(mo))
2711                 GOTO(out, rc = PTR_ERR(mo));
2712
2713         if (!mdt_object_exists(mo))
2714                 GOTO(out_obj, rc = -ENOENT);
2715
2716         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2717                 GOTO(out_obj, rc = -EINVAL);
2718
2719         if (mdt_object_remote(mo))
2720                 GOTO(out_obj, rc = -EREMOTE);
2721
2722         lease = ldlm_handle2lock(rr->rr_lease_handle);
2723         if (lease == NULL)
2724                 GOTO(out_obj, rc = -ESTALE);
2725
2726         /* It's really necessary to grab open_sem and check if the lease lock
2727          * has been lost. There would exist a concurrent writer coming in and
2728          * generating some dirty data in memory cache, the writeback would fail
2729          * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2730         if (!down_write_trylock(&mo->mot_open_sem))
2731                 GOTO(out_put_lease, rc = -EBUSY);
2732
2733         lock_res_and_lock(lease);
2734         lease_broken = ldlm_is_cancel(lease);
2735         unlock_res_and_lock(lease);
2736         if (lease_broken)
2737                 GOTO(out_unlock, rc = -EBUSY);
2738
2739         /* the file has yet opened by anyone else after we took the lease. */
2740         layout.mlc_opc = MD_LAYOUT_RESYNC;
2741         lhc = &info->mti_lh[MDT_LH_LOCAL];
2742         rc = mdt_layout_change(info, mo, lhc, &layout);
2743         if (rc)
2744                 GOTO(out_unlock, rc);
2745
2746         mdt_object_unlock(info, mo, lhc, 0);
2747
2748         ma->ma_need = MA_INODE;
2749         ma->ma_valid = 0;
2750         rc = mdt_attr_get_complex(info, mo, ma);
2751         if (rc != 0)
2752                 GOTO(out_unlock, rc);
2753
2754         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2755         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2756
2757         EXIT;
2758 out_unlock:
2759         up_write(&mo->mot_open_sem);
2760 out_put_lease:
2761         LDLM_LOCK_PUT(lease);
2762 out_obj:
2763         mdt_object_put(info->mti_env, mo);
2764 out:
2765         mdt_client_compatibility(info);
2766         rc2 = mdt_fix_reply(info);
2767         if (rc == 0)
2768                 rc = rc2;
2769         return rc;
2770 }
2771
2772 struct mdt_reinter {
2773         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2774         enum lprocfs_extra_opc mr_extra_opc;
2775 };
2776
2777 static const struct mdt_reinter mdt_reinters[] = {
2778         [REINT_SETATTR] = {
2779                 .mr_handler = &mdt_reint_setattr,
2780                 .mr_extra_opc = MDS_REINT_SETATTR,
2781         },
2782         [REINT_CREATE] = {
2783                 .mr_handler = &mdt_reint_create,
2784                 .mr_extra_opc = MDS_REINT_CREATE,
2785         },
2786         [REINT_LINK] = {
2787                 .mr_handler = &mdt_reint_link,
2788                 .mr_extra_opc = MDS_REINT_LINK,
2789         },
2790         [REINT_UNLINK] = {
2791                 .mr_handler = &mdt_reint_unlink,
2792                 .mr_extra_opc = MDS_REINT_UNLINK,
2793         },
2794         [REINT_RENAME] = {
2795                 .mr_handler = &mdt_reint_rename,
2796                 .mr_extra_opc = MDS_REINT_RENAME,
2797         },
2798         [REINT_OPEN] = {
2799                 .mr_handler = &mdt_reint_open,
2800                 .mr_extra_opc = MDS_REINT_OPEN,
2801         },
2802         [REINT_SETXATTR] = {
2803                 .mr_handler = &mdt_reint_setxattr,
2804                 .mr_extra_opc = MDS_REINT_SETXATTR,
2805         },
2806         [REINT_RMENTRY] = {
2807                 .mr_handler = &mdt_reint_unlink,
2808                 .mr_extra_opc = MDS_REINT_UNLINK,
2809         },
2810         [REINT_MIGRATE] = {
2811                 .mr_handler = &mdt_reint_migrate,
2812                 .mr_extra_opc = MDS_REINT_RENAME,
2813         },
2814         [REINT_RESYNC] = {
2815                 .mr_handler = &mdt_reint_resync,
2816                 .mr_extra_opc = MDS_REINT_RESYNC,
2817         },
2818 };
2819
2820 int mdt_reint_rec(struct mdt_thread_info *info,
2821                   struct mdt_lock_handle *lhc)
2822 {
2823         const struct mdt_reinter *mr;
2824         int rc;
2825         ENTRY;
2826
2827         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2828                 RETURN(-EPROTO);
2829
2830         mr = &mdt_reinters[info->mti_rr.rr_opcode];
2831         if (mr->mr_handler == NULL)
2832                 RETURN(-EPROTO);
2833
2834         rc = (*mr->mr_handler)(info, lhc);
2835
2836         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2837                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
2838
2839         RETURN(rc);
2840 }