lustre/mdt/mdt_reint.c

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