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