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         struct lu_device *bottom_dev;
 254         struct lu_object *bottom_obj;
 255         int rc;
 256
 257         if (!S_ISDIR(obj->mot_header.loh_attr))
 258                 return 0;
 259
 260         /* getxattr from bottom obj to avoid reading in shard FIDs */
 261         bottom_dev = dt2lu_dev(mti->mti_mdt->mdt_bottom);
 262         bottom_obj = lu_object_find_slice(mti->mti_env, bottom_dev,
 263                                           mdt_object_fid(obj), NULL);
 264         if (IS_ERR(bottom_obj))
 265                 return PTR_ERR(bottom_obj);
 266
 267         rc = dt_xattr_get(mti->mti_env, lu2dt(bottom_obj), &LU_BUF_NULL,
 268                           XATTR_NAME_LMV);
 269         lu_object_put(mti->mti_env, bottom_obj);
 270
 271         return (rc > 0) ? 1 : (rc == -ENODATA) ? 0 : rc;
 272 }
 273
 274 /**
 275  * Lock slave stripes if necessary, the lock handles of slave stripes
 276  * will be stored in einfo->ei_cbdata.
 277  **/
 278 static int mdt_lock_slaves(struct mdt_thread_info *mti, struct mdt_object *obj,
 279                            enum ldlm_mode mode, __u64 ibits,
 280                            struct ldlm_enqueue_info *einfo)
 281 {
 282         union ldlm_policy_data *policy = &mti->mti_policy;
 283
 284         LASSERT(S_ISDIR(obj->mot_header.loh_attr));
 285
 286         einfo->ei_type = LDLM_IBITS;
 287         einfo->ei_mode = mode;
 288         einfo->ei_cb_bl = mdt_remote_blocking_ast;
 289         einfo->ei_cb_local_bl = mdt_blocking_ast;
 290         einfo->ei_cb_cp = ldlm_completion_ast;
 291         einfo->ei_enq_slave = 1;
 292         einfo->ei_namespace = mti->mti_mdt->mdt_namespace;
 293         einfo->ei_inodebits = ibits;
 294         memset(policy, 0, sizeof(*policy));
 295         policy->l_inodebits.bits = ibits;
 296
 297         return mo_object_lock(mti->mti_env, mdt_object_child(obj), NULL, einfo,
 298                               policy);
 299 }
 300
 301 int mdt_reint_striped_lock(struct mdt_thread_info *info,
 302                            struct mdt_object *o,
 303                            struct mdt_lock_handle *lh,
 304                            __u64 ibits,
 305                            struct ldlm_enqueue_info *einfo,
 306                            bool cos_incompat)
 307 {
 308         int rc;
 309
 310         LASSERT(!mdt_object_remote(o));
 311
 312         memset(einfo, 0, sizeof(*einfo));
 313
 314         rc = mdt_reint_object_lock(info, o, lh, ibits, cos_incompat);
 315         if (rc)
 316                 return rc;
 317
 318         rc = mdt_object_striped(info, o);
 319         if (rc != 1) {
 320                 if (rc < 0)
 321                         mdt_object_unlock(info, o, lh, rc);
 322                 return rc;
 323         }
 324
 325         rc = mdt_lock_slaves(info, o, lh->mlh_reg_mode, ibits, einfo);
 326         if (rc) {
 327                 mdt_object_unlock(info, o, lh, rc);
 328                 if (rc == -EIO && OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_SLAVE_NAME))
 329                         rc = 0;
 330         }
 331
 332         return rc;
 333 }
 334
 335 void mdt_reint_striped_unlock(struct mdt_thread_info *info,
 336                               struct mdt_object *o,
 337                               struct mdt_lock_handle *lh,
 338                               struct ldlm_enqueue_info *einfo, int decref)
 339 {
 340         if (einfo->ei_cbdata)
 341                 mdt_unlock_slaves(info, o, einfo, decref);
 342         mdt_object_unlock(info, o, lh, decref);
 343 }
 344
 345 /*
 346  * VBR: we save three versions in reply:
 347  * 0 - parent. Check that parent version is the same during replay.
 348  * 1 - name. Version of 'name' if file exists with the same name or
 349  * ENOENT_VERSION, it is needed because file may appear due to missed replays.
 350  * 2 - child. Version of child by FID. Must be ENOENT. It is mostly sanity
 351  * check.
 352  */
 353 static int mdt_create(struct mdt_thread_info *info)
 354 {
 355         struct mdt_device       *mdt = info->mti_mdt;
 356         struct mdt_object       *parent;
 357         struct mdt_object       *child;
 358         struct mdt_lock_handle  *lh;
 359         struct mdt_body         *repbody;
 360         struct md_attr          *ma = &info->mti_attr;
 361         struct mdt_reint_record *rr = &info->mti_rr;
 362         struct md_op_spec       *spec = &info->mti_spec;
 363         int rc;
 364         ENTRY;
 365
 366         DEBUG_REQ(D_INODE, mdt_info_req(info), "Create  ("DNAME"->"DFID") "
 367                   "in "DFID,
 368                   PNAME(&rr->rr_name), PFID(rr->rr_fid2), PFID(rr->rr_fid1));
 369
 370         if (!fid_is_md_operative(rr->rr_fid1))
 371                 RETURN(-EPERM);
 372
 373         if (S_ISDIR(ma->ma_attr.la_mode) &&
 374             spec->u.sp_ea.eadata != NULL && spec->u.sp_ea.eadatalen != 0) {
 375                 const struct lmv_user_md *lum = spec->u.sp_ea.eadata;
 376                 struct lu_ucred *uc = mdt_ucred(info);
 377                 struct obd_export *exp = mdt_info_req(info)->rq_export;
 378
 379                 /* Only new clients can create remote dir( >= 2.4) and
 380                  * striped dir(>= 2.6), old client will return -ENOTSUPP */
 381                 if (!mdt_is_dne_client(exp))
 382                         RETURN(-ENOTSUPP);
 383
 384                 if (le32_to_cpu(lum->lum_stripe_count) > 1) {
 385                         if (!mdt_is_striped_client(exp))
 386                                 RETURN(-ENOTSUPP);
 387
 388                         if (!mdt->mdt_enable_striped_dir)
 389                                 RETURN(-EPERM);
 390                 } else if (!mdt->mdt_enable_remote_dir) {
 391                         RETURN(-EPERM);
 392                 }
 393
 394                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 395                     uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 396                     mdt->mdt_enable_remote_dir_gid != -1)
 397                         RETURN(-EPERM);
 398         }
 399
 400         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 401
 402         parent = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 403         if (IS_ERR(parent))
 404                 RETURN(PTR_ERR(parent));
 405
 406         if (!mdt_object_exists(parent))
 407                 GOTO(put_parent, rc = -ENOENT);
 408
 409         lh = &info->mti_lh[MDT_LH_PARENT];
 410         mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 411         rc = mdt_object_lock(info, parent, lh, MDS_INODELOCK_UPDATE);
 412         if (rc)
 413                 GOTO(put_parent, rc);
 414
 415         if (!mdt_object_remote(parent)) {
 416                 rc = mdt_version_get_check_save(info, parent, 0);
 417                 if (rc)
 418                         GOTO(unlock_parent, rc);
 419         }
 420
 421         /*
 422          * Check child name version during replay.
 423          * During create replay a file may exist with same name.
 424          */
 425         rc = mdt_lookup_version_check(info, parent, &rr->rr_name,
 426                                       &info->mti_tmp_fid1, 1);
 427         if (rc == 0)
 428                 GOTO(unlock_parent, rc = -EEXIST);
 429
 430         /* -ENOENT is expected here */
 431         if (rc != -ENOENT)
 432                 GOTO(unlock_parent, rc);
 433
 434         /* save version of file name for replay, it must be ENOENT here */
 435         mdt_enoent_version_save(info, 1);
 436
 437         child = mdt_object_new(info->mti_env, mdt, rr->rr_fid2);
 438         if (unlikely(IS_ERR(child)))
 439                 GOTO(unlock_parent, rc = PTR_ERR(child));
 440
 441         ma->ma_need = MA_INODE;
 442         ma->ma_valid = 0;
 443
 444         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 445                         OBD_FAIL_MDS_REINT_CREATE_WRITE);
 446
 447         /* Version of child will be updated on disk. */
 448         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(child));
 449         rc = mdt_version_get_check_save(info, child, 2);
 450         if (rc)
 451                 GOTO(put_child, rc);
 452
 453         /* Let lower layer know current lock mode. */
 454         info->mti_spec.sp_cr_mode = mdt_dlm_mode2mdl_mode(lh->mlh_pdo_mode);
 455
 456         /*
 457          * Do not perform lookup sanity check. We know that name does
 458          * not exist.
 459          */
 460         info->mti_spec.sp_cr_lookup = 0;
 461         info->mti_spec.sp_feat = &dt_directory_features;
 462
 463         rc = mdo_create(info->mti_env, mdt_object_child(parent), &rr->rr_name,
 464                         mdt_object_child(child), &info->mti_spec, ma);
 465         if (rc == 0)
 466                 rc = mdt_attr_get_complex(info, child, ma);
 467
 468         if (rc < 0)
 469                 GOTO(put_child, rc);
 470
 471         /*
 472          * On DNE, we need to eliminate dependey between 'mkdir a' and
 473          * 'mkdir a/b' if b is a striped directory, to achieve this, two
 474          * things are done below:
 475          * 1. save child and slaves lock.
 476          * 2. if the child is a striped directory, relock parent so to
 477          *    compare against with COS locks to ensure parent was
 478          *    committed to disk.
 479          */
 480         if (mdt_slc_is_enabled(mdt) && S_ISDIR(ma->ma_attr.la_mode)) {
 481                 struct mdt_lock_handle *lhc;
 482                 struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 483                 bool cos_incompat;
 484
 485                 rc = mdt_object_striped(info, child);
 486                 if (rc < 0)
 487                         GOTO(put_child, rc);
 488
 489                 cos_incompat = rc;
 490                 if (cos_incompat) {
 491                         if (!mdt_object_remote(parent)) {
 492                                 mdt_object_unlock(info, parent, lh, 1);
 493                                 mdt_lock_pdo_init(lh, LCK_PW, &rr->rr_name);
 494                                 rc = mdt_reint_object_lock(info, parent, lh,
 495                                                            MDS_INODELOCK_UPDATE,
 496                                                            true);
 497                                 if (rc)
 498                                         GOTO(put_child, rc);
 499                         }
 500                 }
 501
 502                 lhc = &info->mti_lh[MDT_LH_CHILD];
 503                 mdt_lock_handle_init(lhc);
 504                 mdt_lock_reg_init(lhc, LCK_PW);
 505                 rc = mdt_reint_striped_lock(info, child, lhc,
 506                                             MDS_INODELOCK_UPDATE, einfo,
 507                                             cos_incompat);
 508                 if (rc)
 509                         GOTO(put_child, rc);
 510
 511                 mdt_reint_striped_unlock(info, child, lhc, einfo, rc);
 512         }
 513
 514         /* Return fid & attr to client. */
 515         if (ma->ma_valid & MA_INODE)
 516                 mdt_pack_attr2body(info, repbody, &ma->ma_attr,
 517                                    mdt_object_fid(child));
 518 put_child:
 519         mdt_object_put(info->mti_env, child);
 520 unlock_parent:
 521         mdt_object_unlock(info, parent, lh, rc);
 522 put_parent:
 523         mdt_object_put(info->mti_env, parent);
 524         RETURN(rc);
 525 }
 526
 527 static int mdt_attr_set(struct mdt_thread_info *info, struct mdt_object *mo,
 528                         struct md_attr *ma)
 529 {
 530         struct mdt_lock_handle  *lh;
 531         int do_vbr = ma->ma_attr.la_valid &
 532                         (LA_MODE | LA_UID | LA_GID | LA_PROJID | LA_FLAGS);
 533         __u64 lockpart = MDS_INODELOCK_UPDATE;
 534         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 535         bool cos_incompat;
 536         int rc;
 537         ENTRY;
 538
 539         rc = mdt_object_striped(info, mo);
 540         if (rc < 0)
 541                 RETURN(rc);
 542
 543         cos_incompat = rc;
 544
 545         lh = &info->mti_lh[MDT_LH_PARENT];
 546         mdt_lock_reg_init(lh, LCK_PW);
 547
 548         /* Even though the new MDT will grant PERM lock to the old
 549          * client, but the old client will almost ignore that during
 550          * So it needs to revoke both LOOKUP and PERM lock here, so
 551          * both new and old client can cancel the dcache */
 552         if (ma->ma_attr.la_valid & (LA_MODE|LA_UID|LA_GID))
 553                 lockpart |= MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
 554
 555         rc = mdt_reint_striped_lock(info, mo, lh, lockpart, einfo,
 556                                     cos_incompat);
 557         if (rc != 0)
 558                 RETURN(rc);
 559
 560         /* all attrs are packed into mti_attr in unpack_setattr */
 561         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
 562                        OBD_FAIL_MDS_REINT_SETATTR_WRITE);
 563
 564         /* VBR: update version if attr changed are important for recovery */
 565         if (do_vbr) {
 566                 /* update on-disk version of changed object */
 567                 tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mo));
 568                 rc = mdt_version_get_check_save(info, mo, 0);
 569                 if (rc)
 570                         GOTO(out_unlock, rc);
 571         }
 572
 573         /* Ensure constant striping during chown(). See LU-2789. */
 574         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 575                 mutex_lock(&mo->mot_lov_mutex);
 576
 577         /* all attrs are packed into mti_attr in unpack_setattr */
 578         rc = mo_attr_set(info->mti_env, mdt_object_child(mo), ma);
 579
 580         if (ma->ma_attr.la_valid & (LA_UID|LA_GID|LA_PROJID))
 581                 mutex_unlock(&mo->mot_lov_mutex);
 582
 583         if (rc != 0)
 584                 GOTO(out_unlock, rc);
 585         mdt_dom_obj_lvb_update(info->mti_env, mo, false);
 586         EXIT;
 587 out_unlock:
 588         mdt_reint_striped_unlock(info, mo, lh, einfo, rc);
 589         return rc;
 590 }
 591
 592 /**
 593  * Check HSM flags and add HS_DIRTY flag if relevant.
 594  *
 595  * A file could be set dirty only if it has a copy in the backend (HS_EXISTS)
 596  * and is not RELEASED.
 597  */
 598 int mdt_add_dirty_flag(struct mdt_thread_info *info, struct mdt_object *mo,
 599                         struct md_attr *ma)
 600 {
 601         struct lu_ucred *uc = mdt_ucred(info);
 602         cfs_cap_t cap_saved;
 603         int rc;
 604         ENTRY;
 605
 606         /* If the file was modified, add the dirty flag */
 607         ma->ma_need = MA_HSM;
 608         rc = mdt_attr_get_complex(info, mo, ma);
 609         if (rc) {
 610                 CERROR("file attribute read error for "DFID": %d.\n",
 611                         PFID(mdt_object_fid(mo)), rc);
 612                 RETURN(rc);
 613         }
 614
 615         /* If an up2date copy exists in the backend, add dirty flag */
 616         if ((ma->ma_valid & MA_HSM) && (ma->ma_hsm.mh_flags & HS_EXISTS)
 617             && !(ma->ma_hsm.mh_flags & (HS_DIRTY|HS_RELEASED))) {
 618                 ma->ma_hsm.mh_flags |= HS_DIRTY;
 619
 620                 /* Bump cap so that closes from non-owner writers can
 621                  * set the HSM state to dirty. */
 622                 cap_saved = uc->uc_cap;
 623                 uc->uc_cap |= MD_CAP_TO_MASK(CFS_CAP_FOWNER);
 624                 rc = mdt_hsm_attr_set(info, mo, &ma->ma_hsm);
 625                 uc->uc_cap = cap_saved;
 626                 if (rc)
 627                         CERROR("file attribute change error for "DFID": %d\n",
 628                                 PFID(mdt_object_fid(mo)), rc);
 629         }
 630
 631         RETURN(rc);
 632 }
 633
 634 static int mdt_reint_setattr(struct mdt_thread_info *info,
 635                              struct mdt_lock_handle *lhc)
 636 {
 637         struct mdt_device *mdt = info->mti_mdt;
 638         struct md_attr *ma = &info->mti_attr;
 639         struct mdt_reint_record *rr = &info->mti_rr;
 640         struct ptlrpc_request *req = mdt_info_req(info);
 641         struct mdt_object *mo;
 642         struct mdt_body *repbody;
 643         int rc, rc2;
 644         ENTRY;
 645
 646         DEBUG_REQ(D_INODE, req, "setattr "DFID" %x", PFID(rr->rr_fid1),
 647                   (unsigned int)ma->ma_attr.la_valid);
 648
 649         if (info->mti_dlm_req)
 650                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 651
 652         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 653         mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
 654         if (IS_ERR(mo))
 655                 GOTO(out, rc = PTR_ERR(mo));
 656
 657         if (!mdt_object_exists(mo))
 658                 GOTO(out_put, rc = -ENOENT);
 659
 660         if (mdt_object_remote(mo))
 661                 GOTO(out_put, rc = -EREMOTE);
 662
 663         /* revoke lease lock if size is going to be changed */
 664         if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
 665                      !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
 666                      atomic_read(&mo->mot_lease_count) > 0)) {
 667                 down_read(&mo->mot_open_sem);
 668
 669                 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
 670                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 671                         mdt_lock_reg_init(lhc, LCK_CW);
 672
 673                         rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
 674                         if (rc != 0) {
 675                                 up_read(&mo->mot_open_sem);
 676                                 GOTO(out_put, rc);
 677                         }
 678
 679                         /* revoke lease lock */
 680                         mdt_object_unlock(info, mo, lhc, 1);
 681                 }
 682                 up_read(&mo->mot_open_sem);
 683         }
 684
 685         if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
 686                 /* Check write access for the O_TRUNC case */
 687                 if (mdt_write_read(mo) < 0)
 688                         GOTO(out_put, rc = -ETXTBSY);
 689
 690                 /* LU-10286: compatibility check for FLR.
 691                  * Please check the comment in mdt_finish_open() for details */
 692                 if (!exp_connect_flr(info->mti_exp) ||
 693                     !exp_connect_overstriping(info->mti_exp)) {
 694                         rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
 695                         if (rc < 0 && rc != -ENODATA)
 696                                 GOTO(out_put, rc);
 697
 698                         if (!exp_connect_flr(info->mti_exp)) {
 699                                 if (rc > 0 &&
 700                                     mdt_lmm_is_flr(info->mti_big_lmm))
 701                                         GOTO(out_put, rc = -EOPNOTSUPP);
 702                         }
 703
 704                         if (!exp_connect_overstriping(info->mti_exp)) {
 705                                 if (rc > 0 &&
 706                                     mdt_lmm_is_overstriping(info->mti_big_lmm))
 707                                         GOTO(out_put, rc = -EOPNOTSUPP);
 708                         }
 709                 }
 710
 711                 /* For truncate, the file size sent from client
 712                  * is believable, but the blocks are incorrect,
 713                  * which makes the block size in LSOM attribute
 714                  * inconsisent with the real block size.
 715                  */
 716                 rc = mdt_lsom_update(info, mo, true);
 717                 if (rc)
 718                         GOTO(out_put, rc);
 719         }
 720
 721         if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
 722                 if (ma->ma_valid & MA_LOV)
 723                         GOTO(out_put, rc = -EPROTO);
 724
 725                 /* MDT supports FMD for regular files due to Data-on-MDT */
 726                 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
 727                     ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
 728                         tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
 729                                        req->rq_xid);
 730
 731                 rc = mdt_attr_set(info, mo, ma);
 732                 if (rc)
 733                         GOTO(out_put, rc);
 734         } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
 735                    (ma->ma_valid & MA_INODE)) {
 736                 struct lu_buf *buf = &info->mti_buf;
 737                 struct lu_ucred *uc = mdt_ucred(info);
 738                 struct mdt_lock_handle *lh;
 739                 const char *name;
 740                 __u64 lockpart = MDS_INODELOCK_XATTR;
 741
 742                 /* reject if either remote or striped dir is disabled */
 743                 if (ma->ma_valid & MA_LMV) {
 744                         if (!mdt->mdt_enable_remote_dir ||
 745                             !mdt->mdt_enable_striped_dir)
 746                                 GOTO(out_put, rc = -EPERM);
 747
 748                         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 749                             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 750                             mdt->mdt_enable_remote_dir_gid != -1)
 751                                 GOTO(out_put, rc = -EPERM);
 752                 }
 753
 754                 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
 755                         GOTO(out_put, rc = -ENOTDIR);
 756
 757                 if (ma->ma_attr.la_valid != 0)
 758                         GOTO(out_put, rc = -EPROTO);
 759
 760                 if (ma->ma_valid & MA_LOV) {
 761                         buf->lb_buf = ma->ma_lmm;
 762                         buf->lb_len = ma->ma_lmm_size;
 763                         name = XATTR_NAME_LOV;
 764                 } else {
 765                         struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
 766
 767                         buf->lb_buf = lmu;
 768                         buf->lb_len = ma->ma_lmv_size;
 769
 770                         if (le32_to_cpu(lmu->lum_hash_type) ==
 771                             LMV_HASH_TYPE_SPACE) {
 772                                 /*
 773                                  * only allow setting "space" hash type for
 774                                  * plain directory.
 775                                  */
 776                                 rc = mdt_object_striped(info, mo);
 777                                 if (rc)
 778                                         GOTO(out_put,
 779                                              rc = (rc == 1) ? -EPERM : rc);
 780                         }
 781
 782                         name = XATTR_NAME_DEFAULT_LMV;
 783                         /* force client to update dir default layout */
 784                         lockpart |= MDS_INODELOCK_LOOKUP;
 785                 }
 786
 787                 lh = &info->mti_lh[MDT_LH_PARENT];
 788                 mdt_lock_reg_init(lh, LCK_PW);
 789
 790                 rc = mdt_object_lock(info, mo, lh, lockpart);
 791                 if (rc != 0)
 792                         GOTO(out_put, rc);
 793
 794                 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
 795                                   name, 0);
 796
 797                 mdt_object_unlock(info, mo, lh, rc);
 798                 if (rc)
 799                         GOTO(out_put, rc);
 800         } else {
 801                 GOTO(out_put, rc = -EPROTO);
 802         }
 803
 804         /* If file data is modified, add the dirty flag */
 805         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
 806                 rc = mdt_add_dirty_flag(info, mo, ma);
 807
 808         ma->ma_need = MA_INODE;
 809         ma->ma_valid = 0;
 810         rc = mdt_attr_get_complex(info, mo, ma);
 811         if (rc != 0)
 812                 GOTO(out_put, rc);
 813
 814         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
 815
 816         EXIT;
 817 out_put:
 818         mdt_object_put(info->mti_env, mo);
 819 out:
 820         if (rc == 0)
 821                 mdt_counter_incr(req, LPROC_MDT_SETATTR);
 822
 823         mdt_client_compatibility(info);
 824         rc2 = mdt_fix_reply(info);
 825         if (rc == 0)
 826                 rc = rc2;
 827         return rc;
 828 }
 829
 830 static int mdt_reint_create(struct mdt_thread_info *info,
 831                             struct mdt_lock_handle *lhc)
 832 {
 833         struct ptlrpc_request   *req = mdt_info_req(info);
 834         int                     rc;
 835         ENTRY;
 836
 837         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 838                 RETURN(err_serious(-ESTALE));
 839
 840         if (info->mti_dlm_req)
 841                 ldlm_request_cancel(mdt_info_req(info),
 842                                     info->mti_dlm_req, 0, LATF_SKIP);
 843
 844         if (!lu_name_is_valid(&info->mti_rr.rr_name))
 845                 RETURN(-EPROTO);
 846
 847         switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
 848         case S_IFDIR:
 849                 mdt_counter_incr(req, LPROC_MDT_MKDIR);
 850                 break;
 851         case S_IFREG:
 852         case S_IFLNK:
 853         case S_IFCHR:
 854         case S_IFBLK:
 855         case S_IFIFO:
 856         case S_IFSOCK:
 857                 /* Special file should stay on the same node as parent. */
 858                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
 859                 break;
 860         default:
 861                 CERROR("%s: Unsupported mode %o\n",
 862                        mdt_obd_name(info->mti_mdt),
 863                        info->mti_attr.ma_attr.la_mode);
 864                 RETURN(err_serious(-EOPNOTSUPP));
 865         }
 866
 867         rc = mdt_create(info);
 868         RETURN(rc);
 869 }
 870
 871 /*
 872  * VBR: save parent version in reply and child version getting by its name.
 873  * Version of child is getting and checking during its lookup. If
 874  */
 875 static int mdt_reint_unlink(struct mdt_thread_info *info,
 876                             struct mdt_lock_handle *lhc)
 877 {
 878         struct mdt_reint_record *rr = &info->mti_rr;
 879         struct ptlrpc_request *req = mdt_info_req(info);
 880         struct md_attr *ma = &info->mti_attr;
 881         struct lu_fid *child_fid = &info->mti_tmp_fid1;
 882         struct mdt_object *mp;
 883         struct mdt_object *mc;
 884         struct mdt_lock_handle *parent_lh;
 885         struct mdt_lock_handle *child_lh;
 886         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 887         __u64 lock_ibits;
 888         bool cos_incompat = false;
 889         int no_name = 0;
 890         int rc;
 891
 892         ENTRY;
 893
 894         DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
 895                   PNAME(&rr->rr_name));
 896
 897         if (info->mti_dlm_req)
 898                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 899
 900         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
 901                 RETURN(err_serious(-ENOENT));
 902
 903         if (!fid_is_md_operative(rr->rr_fid1))
 904                 RETURN(-EPERM);
 905
 906         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 907         if (IS_ERR(mp))
 908                 RETURN(PTR_ERR(mp));
 909
 910         if (mdt_object_remote(mp)) {
 911                 cos_incompat = true;
 912         } else {
 913                 rc = mdt_version_get_check_save(info, mp, 0);
 914                 if (rc)
 915                         GOTO(put_parent, rc);
 916         }
 917
 918 relock:
 919         parent_lh = &info->mti_lh[MDT_LH_PARENT];
 920         mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
 921         rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
 922                                    cos_incompat);
 923         if (rc != 0)
 924                 GOTO(put_parent, rc);
 925
 926         /* lookup child object along with version checking */
 927         fid_zero(child_fid);
 928         rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
 929         if (rc != 0) {
 930                 /* Name might not be able to find during resend of
 931                  * remote unlink, considering following case.
 932                  * dir_A is a remote directory, the name entry of
 933                  * dir_A is on MDT0, the directory is on MDT1,
 934                  *
 935                  * 1. client sends unlink req to MDT1.
 936                  * 2. MDT1 sends name delete update to MDT0.
 937                  * 3. name entry is being deleted in MDT0 synchronously.
 938                  * 4. MDT1 is restarted.
 939                  * 5. client resends unlink req to MDT1. So it can not
 940                  *    find the name entry on MDT0 anymore.
 941                  * In this case, MDT1 only needs to destory the local
 942                  * directory.
 943                  * */
 944                 if (mdt_object_remote(mp) && rc == -ENOENT &&
 945                     !fid_is_zero(rr->rr_fid2) &&
 946                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 947                         no_name = 1;
 948                         *child_fid = *rr->rr_fid2;
 949                  } else {
 950                         GOTO(unlock_parent, rc);
 951                  }
 952         }
 953
 954         if (!fid_is_md_operative(child_fid))
 955                 GOTO(unlock_parent, rc = -EPERM);
 956
 957         /* We will lock the child regardless it is local or remote. No harm. */
 958         mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
 959         if (IS_ERR(mc))
 960                 GOTO(unlock_parent, rc = PTR_ERR(mc));
 961
 962         if (!cos_incompat) {
 963                 rc = mdt_object_striped(info, mc);
 964                 if (rc < 0)
 965                         GOTO(unlock_parent, rc);
 966
 967                 cos_incompat = rc;
 968                 if (cos_incompat) {
 969                         mdt_object_put(info->mti_env, mc);
 970                         mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
 971                         goto relock;
 972                 }
 973         }
 974
 975         child_lh = &info->mti_lh[MDT_LH_CHILD];
 976         mdt_lock_reg_init(child_lh, LCK_EX);
 977         if (info->mti_spec.sp_rm_entry) {
 978                 struct lu_ucred *uc  = mdt_ucred(info);
 979
 980                 if (!mdt_is_dne_client(req->rq_export))
 981                         /* Return -ENOTSUPP for old client */
 982                         GOTO(put_child, rc = -ENOTSUPP);
 983
 984                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
 985                         GOTO(put_child, rc = -EPERM);
 986
 987                 ma->ma_need = MA_INODE;
 988                 ma->ma_valid = 0;
 989                 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
 990                                 NULL, &rr->rr_name, ma, no_name);
 991                 GOTO(put_child, rc);
 992         }
 993
 994         if (mdt_object_remote(mc)) {
 995                 struct mdt_body  *repbody;
 996
 997                 if (!fid_is_zero(rr->rr_fid2)) {
 998                         CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
 999                                mdt_obd_name(info->mti_mdt),
1000                                PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1001                         GOTO(put_child, rc = -ENOENT);
1002                 }
1003                 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
1004                        mdt_obd_name(info->mti_mdt),
1005                        PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
1006
1007                 if (!mdt_is_dne_client(req->rq_export))
1008                         /* Return -ENOTSUPP for old client */
1009                         GOTO(put_child, rc = -ENOTSUPP);
1010
1011                 /* Revoke the LOOKUP lock of the remote object granted by
1012                  * this MDT. Since the unlink will happen on another MDT,
1013                  * it will release the LOOKUP lock right away. Then What
1014                  * would happen if another client try to grab the LOOKUP
1015                  * lock at the same time with unlink XXX */
1016                 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1017                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1018                 LASSERT(repbody != NULL);
1019                 repbody->mbo_fid1 = *mdt_object_fid(mc);
1020                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1021                 GOTO(unlock_child, rc = -EREMOTE);
1022         }
1023         /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1024          * this now because a running HSM restore on the child (unlink
1025          * victim) will hold the layout lock. See LU-4002. */
1026         lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1027         if (mdt_object_remote(mp)) {
1028                 /* Enqueue lookup lock from parent MDT */
1029                 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1030                                             &child_lh->mlh_rreg_lh,
1031                                             child_lh->mlh_rreg_mode,
1032                                             MDS_INODELOCK_LOOKUP, false);
1033                 if (rc != ELDLM_OK)
1034                         GOTO(put_child, rc);
1035
1036                 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1037         }
1038
1039         rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1040                                     cos_incompat);
1041         if (rc != 0)
1042                 GOTO(put_child, rc);
1043
1044         /*
1045          * Now we can only make sure we need MA_INODE, in mdd layer, will check
1046          * whether need MA_LOV and MA_COOKIE.
1047          */
1048         ma->ma_need = MA_INODE;
1049         ma->ma_valid = 0;
1050
1051         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1052                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1053         /* save version when object is locked */
1054         mdt_version_get_save(info, mc, 1);
1055
1056         mutex_lock(&mc->mot_lov_mutex);
1057
1058         rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1059                         mdt_object_child(mc), &rr->rr_name, ma, no_name);
1060
1061         mutex_unlock(&mc->mot_lov_mutex);
1062         if (rc != 0)
1063                 GOTO(unlock_child, rc);
1064
1065         if (!lu_object_is_dying(&mc->mot_header)) {
1066                 rc = mdt_attr_get_complex(info, mc, ma);
1067                 if (rc)
1068                         GOTO(out_stat, rc);
1069         } else if (mdt_dom_check_for_discard(info, mc)) {
1070                 mdt_dom_discard_data(info, mc);
1071         }
1072         mdt_handle_last_unlink(info, mc, ma);
1073
1074 out_stat:
1075         if (ma->ma_valid & MA_INODE) {
1076                 switch (ma->ma_attr.la_mode & S_IFMT) {
1077                 case S_IFDIR:
1078                         mdt_counter_incr(req, LPROC_MDT_RMDIR);
1079                         break;
1080                 case S_IFREG:
1081                 case S_IFLNK:
1082                 case S_IFCHR:
1083                 case S_IFBLK:
1084                 case S_IFIFO:
1085                 case S_IFSOCK:
1086                         mdt_counter_incr(req, LPROC_MDT_UNLINK);
1087                         break;
1088                 default:
1089                         LASSERTF(0, "bad file type %o unlinking\n",
1090                                 ma->ma_attr.la_mode);
1091                 }
1092         }
1093
1094         EXIT;
1095
1096 unlock_child:
1097         mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1098 put_child:
1099         mdt_object_put(info->mti_env, mc);
1100 unlock_parent:
1101         mdt_object_unlock(info, mp, parent_lh, rc);
1102 put_parent:
1103         mdt_object_put(info->mti_env, mp);
1104         return rc;
1105 }
1106
1107 /*
1108  * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1109  * name.
1110  */
1111 static int mdt_reint_link(struct mdt_thread_info *info,
1112                           struct mdt_lock_handle *lhc)
1113 {
1114         struct mdt_reint_record *rr = &info->mti_rr;
1115         struct ptlrpc_request   *req = mdt_info_req(info);
1116         struct md_attr          *ma = &info->mti_attr;
1117         struct mdt_object       *ms;
1118         struct mdt_object       *mp;
1119         struct mdt_lock_handle  *lhs;
1120         struct mdt_lock_handle  *lhp;
1121         bool cos_incompat;
1122         int rc;
1123         ENTRY;
1124
1125         DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1126                   PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1127
1128         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1129                 RETURN(err_serious(-ENOENT));
1130
1131         if (info->mti_dlm_req)
1132                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1133
1134         /* Invalid case so return error immediately instead of
1135          * processing it */
1136         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1137                 RETURN(-EPERM);
1138
1139         if (!fid_is_md_operative(rr->rr_fid1) ||
1140             !fid_is_md_operative(rr->rr_fid2))
1141                 RETURN(-EPERM);
1142
1143         /* step 1: find target parent dir */
1144         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1145         if (IS_ERR(mp))
1146                 RETURN(PTR_ERR(mp));
1147
1148         rc = mdt_version_get_check_save(info, mp, 0);
1149         if (rc)
1150                 GOTO(put_parent, rc);
1151
1152         /* step 2: find source */
1153         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1154         if (IS_ERR(ms))
1155                 GOTO(put_parent, rc = PTR_ERR(ms));
1156
1157         if (!mdt_object_exists(ms)) {
1158                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1159                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1160                 GOTO(put_source, rc = -ENOENT);
1161         }
1162
1163         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1164
1165         lhp = &info->mti_lh[MDT_LH_PARENT];
1166         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1167         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1168                                    cos_incompat);
1169         if (rc != 0)
1170                 GOTO(put_source, rc);
1171
1172         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1173
1174         lhs = &info->mti_lh[MDT_LH_CHILD];
1175         mdt_lock_reg_init(lhs, LCK_EX);
1176         rc = mdt_reint_object_lock(info, ms, lhs,
1177                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1178                                    cos_incompat);
1179         if (rc != 0)
1180                 GOTO(unlock_parent, rc);
1181
1182         /* step 3: link it */
1183         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1184                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1185
1186         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1187         rc = mdt_version_get_check_save(info, ms, 1);
1188         if (rc)
1189                 GOTO(unlock_source, rc);
1190
1191         /** check target version by name during replay */
1192         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1193                                       &info->mti_tmp_fid1, 2);
1194         if (rc != 0 && rc != -ENOENT)
1195                 GOTO(unlock_source, rc);
1196         /* save version of file name for replay, it must be ENOENT here */
1197         if (!req_is_replay(mdt_info_req(info))) {
1198                 if (rc != -ENOENT) {
1199                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1200                                PNAME(&rr->rr_name));
1201                         GOTO(unlock_source, rc = -EEXIST);
1202                 }
1203                 info->mti_ver[2] = ENOENT_VERSION;
1204                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1205         }
1206
1207         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1208                       mdt_object_child(ms), &rr->rr_name, ma);
1209
1210         if (rc == 0)
1211                 mdt_counter_incr(req, LPROC_MDT_LINK);
1212
1213         EXIT;
1214 unlock_source:
1215         mdt_object_unlock(info, ms, lhs, rc);
1216 unlock_parent:
1217         mdt_object_unlock(info, mp, lhp, rc);
1218 put_source:
1219         mdt_object_put(info->mti_env, ms);
1220 put_parent:
1221         mdt_object_put(info->mti_env, mp);
1222         return rc;
1223 }
1224 /**
1225  * lock the part of the directory according to the hash of the name
1226  * (lh->mlh_pdo_hash) in parallel directory lock.
1227  */
1228 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1229                               struct mdt_lock_handle *lh,
1230                               struct mdt_object *obj, __u64 ibits,
1231                               bool cos_incompat)
1232 {
1233         struct ldlm_res_id *res = &info->mti_res_id;
1234         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1235         union ldlm_policy_data *policy = &info->mti_policy;
1236         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1237         int rc;
1238
1239         /*
1240          * Finish res_id initializing by name hash marking part of
1241          * directory which is taking modification.
1242          */
1243         LASSERT(lh->mlh_pdo_hash != 0);
1244         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1245         memset(policy, 0, sizeof(*policy));
1246         policy->l_inodebits.bits = ibits;
1247         if (cos_incompat &&
1248             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1249                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1250         /*
1251          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1252          * going to be sent to client. If it is - mdt_intent_policy() path will
1253          * fix it up and turn FL_LOCAL flag off.
1254          */
1255         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1256                           policy, res, dlmflags,
1257                           &info->mti_exp->exp_handle.h_cookie);
1258         return rc;
1259 }
1260
1261 /**
1262  * Get BFL lock for rename or migrate process.
1263  **/
1264 static int mdt_rename_lock(struct mdt_thread_info *info,
1265                            struct lustre_handle *lh)
1266 {
1267         int     rc;
1268         ENTRY;
1269
1270         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1271                 struct lu_fid *fid = &info->mti_tmp_fid1;
1272                 struct mdt_object *obj;
1273
1274                 /* XXX, right now, it has to use object API to
1275                  * enqueue lock cross MDT, so it will enqueue
1276                  * rename lock(with LUSTRE_BFL_FID) by root object */
1277                 lu_root_fid(fid);
1278                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1279                 if (IS_ERR(obj))
1280                         RETURN(PTR_ERR(obj));
1281
1282                 rc = mdt_remote_object_lock(info, obj,
1283                                             &LUSTRE_BFL_FID, lh,
1284                                             LCK_EX,
1285                                             MDS_INODELOCK_UPDATE, false);
1286                 mdt_object_put(info->mti_env, obj);
1287         } else {
1288                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1289                 union ldlm_policy_data *policy = &info->mti_policy;
1290                 struct ldlm_res_id *res_id = &info->mti_res_id;
1291                 __u64 flags = 0;
1292
1293                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1294                 memset(policy, 0, sizeof *policy);
1295                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1296                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1297                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1298                                             LDLM_IBITS, policy, LCK_EX, &flags,
1299                                             ldlm_blocking_ast,
1300                                             ldlm_completion_ast, NULL, NULL, 0,
1301                                             LVB_T_NONE,
1302                                             &info->mti_exp->exp_handle.h_cookie,
1303                                             lh);
1304                 RETURN(rc);
1305         }
1306         RETURN(rc);
1307 }
1308
1309 static void mdt_rename_unlock(struct lustre_handle *lh)
1310 {
1311         ENTRY;
1312         LASSERT(lustre_handle_is_used(lh));
1313         /* Cancel the single rename lock right away */
1314         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1315         EXIT;
1316 }
1317
1318 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1319                                                 const struct lu_fid *fid,
1320                                                 int idx)
1321 {
1322         struct mdt_object *dir;
1323         int rc;
1324
1325         ENTRY;
1326
1327         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1328         if (IS_ERR(dir))
1329                 RETURN(dir);
1330
1331         /* check early, the real version will be saved after locking */
1332         rc = mdt_version_get_check(info, dir, idx);
1333         if (rc)
1334                 GOTO(out_put, rc);
1335
1336         if (!mdt_object_exists(dir))
1337                 GOTO(out_put, rc = -ENOENT);
1338
1339         if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1340                 GOTO(out_put, rc = -ENOTDIR);
1341
1342         RETURN(dir);
1343 out_put:
1344         mdt_object_put(info->mti_env, dir);
1345         return ERR_PTR(rc);
1346 }
1347
1348 /*
1349  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1350  * herein we don't really check it, just do revoke.
1351  */
1352 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1353                                   struct mdt_object *pobj,
1354                                   struct mdt_object *obj)
1355 {
1356         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1357         int rc;
1358
1359         mdt_lock_handle_init(lh);
1360         mdt_lock_reg_init(lh, LCK_EX);
1361
1362         if (mdt_object_remote(pobj)) {
1363                 /* don't bother to check if pobj and obj are on the same MDT. */
1364                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1365                                             &lh->mlh_rreg_lh, LCK_EX,
1366                                             MDS_INODELOCK_LOOKUP, false);
1367         } else if (mdt_object_remote(obj)) {
1368                 struct ldlm_res_id *res = &info->mti_res_id;
1369                 union ldlm_policy_data *policy = &info->mti_policy;
1370                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1371                                  LDLM_FL_COS_INCOMPAT;
1372
1373                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1374                 memset(policy, 0, sizeof(*policy));
1375                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1376                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1377                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1378                                   dlmflags, NULL);
1379         } else {
1380                 /* do nothing if both are local */
1381                 return 0;
1382         }
1383
1384         if (rc != ELDLM_OK)
1385                 return rc;
1386
1387         /*
1388          * TODO, currently we don't save this lock because there is no place to
1389          * hold this lock handle, but to avoid race we need to save this lock.
1390          */
1391         mdt_object_unlock(info, NULL, lh, 1);
1392
1393         return 0;
1394 }
1395
1396 /*
1397  * operation may takes locks of linkea, or directory stripes, group them in
1398  * different list.
1399  */
1400 struct mdt_sub_lock {
1401         struct mdt_object      *msl_obj;
1402         struct mdt_lock_handle  msl_lh;
1403         struct list_head        msl_linkage;
1404 };
1405
1406 static void mdt_unlock_list(struct mdt_thread_info *info,
1407                             struct list_head *list, int decref)
1408 {
1409         struct mdt_sub_lock *msl;
1410         struct mdt_sub_lock *tmp;
1411
1412         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1413                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1414                 list_del(&msl->msl_linkage);
1415                 OBD_FREE_PTR(msl);
1416         }
1417 }
1418
1419 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1420                                              struct mdt_object *obj,
1421                                              struct mdt_lock_handle *lh,
1422                                              struct ldlm_enqueue_info *einfo,
1423                                              struct list_head *slave_locks,
1424                                              int decref)
1425 {
1426         if (mdt_object_remote(obj)) {
1427                 mdt_unlock_list(info, slave_locks, decref);
1428                 mdt_object_unlock(info, obj, lh, decref);
1429         } else {
1430                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1431         }
1432 }
1433
1434 /*
1435  * lock parents of links, and also check whether total locks don't exceed
1436  * RS_MAX_LOCKS.
1437  *
1438  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1439  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1440  * \retval      -ev negative errno upon error
1441  */
1442 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1443                                  struct mdt_object *pobj,
1444                                  const struct md_attr *ma,
1445                                  struct mdt_object *obj,
1446                                  struct mdt_lock_handle *lhp,
1447                                  struct ldlm_enqueue_info *peinfo,
1448                                  struct list_head *parent_slave_locks,
1449                                  struct list_head *link_locks)
1450 {
1451         struct mdt_device *mdt = info->mti_mdt;
1452         struct lu_buf *buf = &info->mti_big_buf;
1453         struct lu_name *lname = &info->mti_name;
1454         struct linkea_data ldata = { NULL };
1455         bool blocked = false;
1456         int local_lnkp_cnt = 0;
1457         int rc;
1458
1459         ENTRY;
1460
1461         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1462                 RETURN(0);
1463
1464         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1465         if (buf->lb_buf == NULL)
1466                 RETURN(-ENOMEM);
1467
1468         ldata.ld_buf = buf;
1469         rc = mdt_links_read(info, obj, &ldata);
1470         if (rc) {
1471                 if (rc == -ENOENT || rc == -ENODATA)
1472                         rc = 0;
1473                 RETURN(rc);
1474         }
1475
1476         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1477              linkea_next_entry(&ldata)) {
1478                 struct mdt_object *lnkp;
1479                 struct mdt_sub_lock *msl;
1480                 struct lu_fid fid;
1481                 __u64 ibits;
1482
1483                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1484                                     &fid);
1485
1486                 /* check if it's also linked to parent */
1487                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1488                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1489                                PFID(&fid), PNAME(lname));
1490                         /* in case link is remote object, revoke LOOKUP lock */
1491                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1492                         continue;
1493                 }
1494
1495                 lnkp = NULL;
1496
1497                 /* check if it's linked to a stripe of parent */
1498                 if (ma->ma_valid & MA_LMV) {
1499                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1500                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1501                         int j = 0;
1502
1503                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1504                                 fid_le_to_cpu(stripe_fid,
1505                                               &lmv->lmv_stripe_fids[j]);
1506                                 if (lu_fid_eq(stripe_fid, &fid)) {
1507                                         CDEBUG(D_INFO, "skip stripe "DFID
1508                                                ", reovke "DNAME"\n",
1509                                                PFID(&fid), PNAME(lname));
1510                                         lnkp = mdt_object_find(info->mti_env,
1511                                                                mdt, &fid);
1512                                         if (IS_ERR(lnkp))
1513                                                 GOTO(out, rc = PTR_ERR(lnkp));
1514                                         break;
1515                                 }
1516                         }
1517
1518                         if (lnkp) {
1519                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1520                                                                    obj);
1521                                 mdt_object_put(info->mti_env, lnkp);
1522                                 continue;
1523                         }
1524                 }
1525
1526                 /* Check if it's already locked */
1527                 list_for_each_entry(msl, link_locks, msl_linkage) {
1528                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1529                                 CDEBUG(D_INFO,
1530                                        DFID" was locked, revoke "DNAME"\n",
1531                                        PFID(&fid), PNAME(lname));
1532                                 lnkp = msl->msl_obj;
1533                                 break;
1534                         }
1535                 }
1536
1537                 if (lnkp) {
1538                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1539                         continue;
1540                 }
1541
1542                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1543                        PFID(&fid), PNAME(lname));
1544
1545                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1546                 if (IS_ERR(lnkp)) {
1547                         CWARN("%s: cannot find obj "DFID": %ld\n",
1548                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1549                         continue;
1550                 }
1551
1552                 if (!mdt_object_exists(lnkp)) {
1553                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1554                               PFID(&fid), PNAME(lname));
1555                         mdt_object_put(info->mti_env, lnkp);
1556                         continue;
1557                 }
1558
1559                 if (!mdt_object_remote(lnkp))
1560                         local_lnkp_cnt++;
1561
1562                 OBD_ALLOC_PTR(msl);
1563                 if (msl == NULL)
1564                         GOTO(out, rc = -ENOMEM);
1565
1566                 /*
1567                  * we can't follow parent-child lock order like other MD
1568                  * operations, use lock_try here to avoid deadlock, if the lock
1569                  * cannot be taken, drop all locks taken, revoke the blocked
1570                  * one, and continue processing the remaining entries, and in
1571                  * the end of the loop restart from beginning.
1572                  */
1573                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1574                 ibits = 0;
1575                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1576                                          MDS_INODELOCK_UPDATE, true);
1577                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1578
1579                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1580                                PFID(&fid), PNAME(lname));
1581
1582                         mdt_unlock_list(info, link_locks, 1);
1583                         /* also unlock parent locks to avoid deadlock */
1584                         if (!blocked)
1585                                 mdt_migrate_object_unlock(info, pobj, lhp,
1586                                                           peinfo,
1587                                                           parent_slave_locks,
1588                                                           1);
1589
1590                         blocked = true;
1591
1592                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1593                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1594                                              MDS_INODELOCK_UPDATE);
1595                         if (rc) {
1596                                 mdt_object_put(info->mti_env, lnkp);
1597                                 OBD_FREE_PTR(msl);
1598                                 GOTO(out, rc);
1599                         }
1600
1601                         if (mdt_object_remote(lnkp)) {
1602                                 struct ldlm_lock *lock;
1603
1604                                 /*
1605                                  * for remote object, set lock cb_atomic,
1606                                  * so lock can be released in blocking_ast()
1607                                  * immediately, then the next lock_try will
1608                                  * have better chance of success.
1609                                  */
1610                                 lock = ldlm_handle2lock(
1611                                                 &msl->msl_lh.mlh_rreg_lh);
1612                                 LASSERT(lock != NULL);
1613                                 lock_res_and_lock(lock);
1614                                 ldlm_set_atomic_cb(lock);
1615                                 unlock_res_and_lock(lock);
1616                                 LDLM_LOCK_PUT(lock);
1617                         }
1618
1619                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1620                         OBD_FREE_PTR(msl);
1621                         continue;
1622                 }
1623
1624                 INIT_LIST_HEAD(&msl->msl_linkage);
1625                 msl->msl_obj = lnkp;
1626                 list_add_tail(&msl->msl_linkage, link_locks);
1627
1628                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1629         }
1630
1631         if (blocked)
1632                 GOTO(out, rc = -EBUSY);
1633
1634         EXIT;
1635 out:
1636         if (rc)
1637                 mdt_unlock_list(info, link_locks, rc);
1638         else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1639                 /*
1640                  * parent may have 3 local objects: master object and 2 stripes
1641                  * (if it's being migrated too); source may have 2 local
1642                  * objects: master and 1 stripe; target has 1 local object.
1643                  */
1644                 rc = 1;
1645         return rc;
1646 }
1647
1648 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1649                                   struct mdt_object *obj,
1650                                   const struct md_attr *ma,
1651                                   struct list_head *slave_locks)
1652 {
1653         struct mdt_device *mdt = info->mti_mdt;
1654         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1655         struct lu_fid *fid = &info->mti_tmp_fid1;
1656         struct mdt_object *slave;
1657         struct mdt_sub_lock *msl;
1658         int i;
1659         int rc;
1660
1661         ENTRY;
1662
1663         LASSERT(mdt_object_remote(obj));
1664         LASSERT(ma->ma_valid & MA_LMV);
1665         LASSERT(lmv);
1666
1667         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1668                 RETURN(-EINVAL);
1669
1670         if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1671                 RETURN(0);
1672
1673         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1674                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1675
1676                 if (!fid_is_sane(fid))
1677                         continue;
1678
1679                 slave = mdt_object_find(info->mti_env, mdt, fid);
1680                 if (IS_ERR(slave))
1681                         GOTO(out, rc = PTR_ERR(slave));
1682
1683                 OBD_ALLOC_PTR(msl);
1684                 if (!msl) {
1685                         mdt_object_put(info->mti_env, slave);
1686                         GOTO(out, rc = -ENOMEM);
1687                 }
1688
1689                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1690                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1691                                            MDS_INODELOCK_UPDATE, true);
1692                 if (rc) {
1693                         OBD_FREE_PTR(msl);
1694                         mdt_object_put(info->mti_env, slave);
1695                         GOTO(out, rc);
1696                 }
1697
1698                 INIT_LIST_HEAD(&msl->msl_linkage);
1699                 msl->msl_obj = slave;
1700                 list_add_tail(&msl->msl_linkage, slave_locks);
1701         }
1702         EXIT;
1703
1704 out:
1705         if (rc)
1706                 mdt_unlock_list(info, slave_locks, rc);
1707         return rc;
1708 }
1709
1710 /* lock parent and its stripes */
1711 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1712                                    struct mdt_object *obj,
1713                                    const struct md_attr *ma,
1714                                    struct mdt_lock_handle *lh,
1715                                    struct ldlm_enqueue_info *einfo,
1716                                    struct list_head *slave_locks)
1717 {
1718         int rc;
1719
1720         if (mdt_object_remote(obj)) {
1721                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1722                                             &lh->mlh_rreg_lh, LCK_PW,
1723                                             MDS_INODELOCK_UPDATE, false);
1724                 if (rc != ELDLM_OK)
1725                         return rc;
1726
1727                 /*
1728                  * if obj is remote and striped, lock its stripes explicitly
1729                  * because it's not striped in LOD layer on this MDT.
1730                  */
1731                 if (ma->ma_valid & MA_LMV) {
1732                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1733                         if (rc)
1734                                 mdt_object_unlock(info, obj, lh, rc);
1735                 }
1736         } else {
1737                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1738                                             einfo, true);
1739         }
1740
1741         return rc;
1742 }
1743
1744 /*
1745  * in migration, object may be remote, and we need take full lock of it and its
1746  * stripes if it's directory, besides, object may be a remote object on its
1747  * parent, revoke its LOOKUP lock on where its parent is located.
1748  */
1749 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1750                                    struct mdt_object *pobj,
1751                                    struct mdt_object *obj,
1752                                    struct mdt_lock_handle *lh,
1753                                    struct ldlm_enqueue_info *einfo,
1754                                    struct list_head *slave_locks)
1755 {
1756         int rc;
1757
1758         if (mdt_object_remote(obj)) {
1759                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1760                 if (rc)
1761                         return rc;
1762
1763                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1764                                             &lh->mlh_rreg_lh, LCK_EX,
1765                                             MDS_INODELOCK_FULL, false);
1766                 if (rc != ELDLM_OK)
1767                         return rc;
1768
1769                 /*
1770                  * if obj is remote and striped, lock its stripes explicitly
1771                  * because it's not striped in LOD layer on this MDT.
1772                  */
1773                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1774                         struct md_attr *ma = &info->mti_attr;
1775
1776                         ma->ma_lmv = info->mti_big_lmm;
1777                         ma->ma_lmv_size = info->mti_big_lmmsize;
1778                         ma->ma_valid = 0;
1779                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1780                         if (rc) {
1781                                 mdt_object_unlock(info, obj, lh, rc);
1782                                 return rc;
1783                         }
1784
1785                         if (ma->ma_valid & MA_LMV) {
1786                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1787                                                             slave_locks);
1788                                 if (rc)
1789                                         mdt_object_unlock(info, obj, lh, rc);
1790                         }
1791                 }
1792         } else {
1793                 if (mdt_object_remote(pobj)) {
1794                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1795                         if (rc)
1796                                 return rc;
1797                 }
1798
1799                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1800                                             einfo, true);
1801         }
1802
1803         return rc;
1804 }
1805
1806 /*
1807  * lookup source by name, if parent is striped directory, we need to find the
1808  * corresponding stripe where source is located, and then lookup there.
1809  *
1810  * besides, if parent is migrating too, and file is already in target stripe,
1811  * this should be a redo of 'lfs migrate' on client side.
1812  */
1813 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1814                               struct mdt_object *pobj,
1815                               const struct md_attr *ma,
1816                               const struct lu_name *lname,
1817                               struct mdt_object **spobj,
1818                               struct mdt_object **sobj)
1819 {
1820         const struct lu_env *env = info->mti_env;
1821         struct lu_fid *fid = &info->mti_tmp_fid1;
1822         struct mdt_object *stripe;
1823         int rc;
1824
1825         if (ma->ma_valid & MA_LMV) {
1826                 /* if parent is striped, lookup on corresponding stripe */
1827                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1828                 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1829                 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1830                 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1831                                     LMV_HASH_FLAG_MIGRATION;
1832
1833                 if (is_migrating) {
1834                         hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1835                         stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1836                 }
1837
1838                 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1839                                               lname->ln_name,
1840                                               lname->ln_namelen);
1841                 if (rc < 0)
1842                         return rc;
1843
1844                 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1845                         rc += le32_to_cpu(lmv->lmv_migrate_offset);
1846
1847                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1848
1849                 stripe = mdt_object_find(env, info->mti_mdt, fid);
1850                 if (IS_ERR(stripe))
1851                         return PTR_ERR(stripe);
1852
1853                 fid_zero(fid);
1854                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1855                                 &info->mti_spec);
1856                 if (rc == -ENOENT && is_migrating) {
1857                         /*
1858                          * if parent is migrating, and lookup child failed on
1859                          * source stripe, lookup again on target stripe, if it
1860                          * exists, it means previous migration was interrupted,
1861                          * and current file was migrated already.
1862                          */
1863                         mdt_object_put(env, stripe);
1864
1865                         hash_type = le32_to_cpu(lmv->lmv_hash_type);
1866                         stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1867
1868                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1869                                                       lname->ln_name,
1870                                                       lname->ln_namelen);
1871                         if (rc < 0)
1872                                 return rc;
1873
1874                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1875
1876                         stripe = mdt_object_find(env, info->mti_mdt, fid);
1877                         if (IS_ERR(stripe))
1878                                 return PTR_ERR(stripe);
1879
1880                         fid_zero(fid);
1881                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1882                                         fid, &info->mti_spec);
1883                         mdt_object_put(env, stripe);
1884                         return rc ?: -EALREADY;
1885                 } else if (rc) {
1886                         mdt_object_put(env, stripe);
1887                         return rc;
1888                 }
1889         } else {
1890                 fid_zero(fid);
1891                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1892                                 &info->mti_spec);
1893                 if (rc)
1894                         return rc;
1895
1896                 stripe = pobj;
1897                 mdt_object_get(env, stripe);
1898         }
1899
1900         *spobj = stripe;
1901
1902         *sobj = mdt_object_find(env, info->mti_mdt, fid);
1903         if (IS_ERR(*sobj)) {
1904                 mdt_object_put(env, stripe);
1905                 rc = PTR_ERR(*sobj);
1906                 *spobj = NULL;
1907                 *sobj = NULL;
1908         }
1909
1910         return rc;
1911 }
1912
1913 /* end lease and close file for regular file */
1914 static int mdd_migrate_close(struct mdt_thread_info *info,
1915                              struct mdt_object *obj)
1916 {
1917         struct close_data *data;
1918         struct mdt_body *repbody;
1919         struct ldlm_lock *lease;
1920         int rc;
1921         int rc2;
1922
1923         rc = -EPROTO;
1924         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1925                                       RCL_CLIENT) ||
1926             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1927                                       RCL_CLIENT))
1928                 goto close;
1929
1930         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1931         if (!data)
1932                 goto close;
1933
1934         rc = -ESTALE;
1935         lease = ldlm_handle2lock(&data->cd_handle);
1936         if (!lease)
1937                 goto close;
1938
1939         /* check if the lease was already canceled */
1940         lock_res_and_lock(lease);
1941         rc = ldlm_is_cancel(lease);
1942         unlock_res_and_lock(lease);
1943
1944         if (rc) {
1945                 rc = -EAGAIN;
1946                 LDLM_DEBUG(lease, DFID" lease broken",
1947                            PFID(mdt_object_fid(obj)));
1948         }
1949
1950         /*
1951          * cancel server side lease, client side counterpart should have been
1952          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1953          */
1954         ldlm_lock_cancel(lease);
1955         ldlm_reprocess_all(lease->l_resource);
1956         LDLM_LOCK_PUT(lease);
1957
1958 close:
1959         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1960         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1961         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1962
1963         return rc ?: rc2;
1964 }
1965
1966 /*
1967  * migrate file in below steps:
1968  *  1. lock parent and its stripes
1969  *  2. lookup source by name
1970  *  3. lock parents of source links if source is not directory
1971  *  4. reject if source is in HSM
1972  *  5. take source open_sem and close file if source is regular file
1973  *  6. lock source and its stripes if it's directory
1974  *  7. lock target so subsequent change to it can trigger COS
1975  *  8. migrate file
1976  *  9. unlock above locks
1977  * 10. sync device if source has links
1978  */
1979 static int mdt_reint_migrate(struct mdt_thread_info *info,
1980                              struct mdt_lock_handle *unused)
1981 {
1982         const struct lu_env *env = info->mti_env;
1983         struct mdt_device *mdt = info->mti_mdt;
1984         struct ptlrpc_request *req = mdt_info_req(info);
1985         struct mdt_reint_record *rr = &info->mti_rr;
1986         struct lu_ucred *uc = mdt_ucred(info);
1987         struct md_attr *ma = &info->mti_attr;
1988         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1989         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1990         struct mdt_object *pobj;
1991         struct mdt_object *spobj = NULL;
1992         struct mdt_object *sobj = NULL;
1993         struct mdt_object *tobj;
1994         struct lustre_handle rename_lh = { 0 };
1995         struct mdt_lock_handle *lhp;
1996         struct mdt_lock_handle *lhs;
1997         struct mdt_lock_handle *lht;
1998         LIST_HEAD(parent_slave_locks);
1999         LIST_HEAD(child_slave_locks);
2000         LIST_HEAD(link_locks);
2001         int lock_retries = 5;
2002         bool open_sem_locked = false;
2003         bool do_sync = false;
2004         int rc;
2005         ENTRY;
2006
2007         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2008                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2009
2010         if (info->mti_dlm_req)
2011                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2012
2013         if (!fid_is_md_operative(rr->rr_fid1) ||
2014             !fid_is_md_operative(rr->rr_fid2))
2015                 RETURN(-EPERM);
2016
2017         /* don't allow migrate . or .. */
2018         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2019                 RETURN(-EBUSY);
2020
2021         if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2022                 RETURN(-EPERM);
2023
2024         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
2025             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2026             mdt->mdt_enable_remote_dir_gid != -1)
2027                 RETURN(-EPERM);
2028
2029         /*
2030          * Note: do not enqueue rename lock for replay request, because
2031          * if other MDT holds rename lock, but being blocked to wait for
2032          * this MDT to finish its recovery, and the failover MDT can not
2033          * get rename lock, which will cause deadlock.
2034          */
2035         if (!req_is_replay(req)) {
2036                 rc = mdt_rename_lock(info, &rename_lh);
2037                 if (rc != 0) {
2038                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2039                                mdt_obd_name(info->mti_mdt), rc);
2040                         RETURN(rc);
2041                 }
2042         }
2043
2044         /* pobj is master object of parent */
2045         pobj = mdt_parent_find_check(info, rr->rr_fid1, 0);
2046         if (IS_ERR(pobj))
2047                 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2048
2049         if (unlikely(!info->mti_big_lmm)) {
2050                 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
2051                 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
2052                 if (!info->mti_big_lmm)
2053                         GOTO(put_parent, rc = -ENOMEM);
2054         }
2055
2056         ma->ma_lmv = info->mti_big_lmm;
2057         ma->ma_lmv_size = info->mti_big_lmmsize;
2058         ma->ma_valid = 0;
2059         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2060         if (rc)
2061                 GOTO(put_parent, rc);
2062
2063 lock_parent:
2064         /* lock parent object */
2065         lhp = &info->mti_lh[MDT_LH_PARENT];
2066         mdt_lock_reg_init(lhp, LCK_PW);
2067         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2068                                      &parent_slave_locks);
2069         if (rc)
2070                 GOTO(put_parent, rc);
2071
2072         /*
2073          * spobj is the corresponding stripe against name if pobj is striped
2074          * directory, which is the real parent, and no need to lock, because
2075          * we've taken full lock of pobj.
2076          */
2077         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2078         if (rc)
2079                 GOTO(unlock_parent, rc);
2080
2081         /* lock parents of source links, and revoke LOOKUP lock of links */
2082         rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2083                                    &parent_slave_locks, &link_locks);
2084         if (rc == -EBUSY && lock_retries-- > 0) {
2085                 mdt_object_put(env, sobj);
2086                 mdt_object_put(env, spobj);
2087                 goto lock_parent;
2088         }
2089
2090         if (rc < 0)
2091                 GOTO(put_source, rc);
2092
2093         /*
2094          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2095          * with one request, if total lock count exceeds this limit, we will
2096          * drop all locks after migration, and synchronous device in the end.
2097          */
2098         do_sync = rc;
2099
2100         /* TODO: DoM migration is not supported yet */
2101         if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2102                 ma->ma_lmm = info->mti_big_lmm;
2103                 ma->ma_lmm_size = info->mti_big_lmmsize;
2104                 ma->ma_valid = 0;
2105                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2106                 if (rc)
2107                         GOTO(put_source, rc);
2108
2109                 if (ma->ma_valid & MA_LOV &&
2110                     mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2111                         GOTO(put_source, rc = -EOPNOTSUPP);
2112         }
2113
2114         /* if migration HSM is allowed */
2115         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2116                 ma->ma_need = MA_HSM;
2117                 ma->ma_valid = 0;
2118                 rc = mdt_attr_get_complex(info, sobj, ma);
2119                 if (rc)
2120                         GOTO(unlock_links, rc);
2121
2122                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2123                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2124         }
2125
2126         /* end lease and close file for regular file */
2127         if (info->mti_spec.sp_migrate_close) {
2128                 /* try to hold open_sem so that nobody else can open the file */
2129                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2130                         /* close anyway */
2131                         mdd_migrate_close(info, sobj);
2132                         GOTO(unlock_links, rc = -EBUSY);
2133                 } else {
2134                         open_sem_locked = true;
2135                         rc = mdd_migrate_close(info, sobj);
2136                         if (rc)
2137                                 GOTO(unlock_open_sem, rc);
2138                 }
2139         }
2140
2141         /* lock source */
2142         lhs = &info->mti_lh[MDT_LH_OLD];
2143         mdt_lock_reg_init(lhs, LCK_EX);
2144         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2145                                      &child_slave_locks);
2146         if (rc)
2147                 GOTO(unlock_open_sem, rc);
2148
2149         /* lock target */
2150         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2151         if (IS_ERR(tobj))
2152                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2153
2154         lht = &info->mti_lh[MDT_LH_NEW];
2155         mdt_lock_reg_init(lht, LCK_EX);
2156         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2157         if (rc)
2158                 GOTO(put_target, rc);
2159
2160         /* Don't do lookup sanity check. We know name doesn't exist. */
2161         info->mti_spec.sp_cr_lookup = 0;
2162         info->mti_spec.sp_feat = &dt_directory_features;
2163
2164         rc = mdo_migrate(env, mdt_object_child(pobj),
2165                          mdt_object_child(sobj), &rr->rr_name,
2166                          mdt_object_child(tobj), &info->mti_spec, ma);
2167         EXIT;
2168
2169         mdt_object_unlock(info, tobj, lht, rc);
2170 put_target:
2171         mdt_object_put(env, tobj);
2172 unlock_source:
2173         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2174                                   &child_slave_locks, rc);
2175 unlock_open_sem:
2176         if (open_sem_locked)
2177                 up_write(&sobj->mot_open_sem);
2178 unlock_links:
2179         mdt_unlock_list(info, &link_locks, do_sync ? 1 : rc);
2180 put_source:
2181         mdt_object_put(env, sobj);
2182         mdt_object_put(env, spobj);
2183 unlock_parent:
2184         mdt_migrate_object_unlock(info, pobj, lhp, peinfo,
2185                                   &parent_slave_locks, rc);
2186 put_parent:
2187         mdt_object_put(env, pobj);
2188 unlock_rename:
2189         if (lustre_handle_is_used(&rename_lh))
2190                 mdt_rename_unlock(&rename_lh);
2191
2192         if (!rc && do_sync)
2193                 mdt_device_sync(env, mdt);
2194
2195         return rc;
2196 }
2197
2198 static int mdt_object_lock_save(struct mdt_thread_info *info,
2199                                 struct mdt_object *dir,
2200                                 struct mdt_lock_handle *lh,
2201                                 int idx, bool cos_incompat)
2202 {
2203         int rc;
2204
2205         /* we lock the target dir if it is local */
2206         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2207                                    cos_incompat);
2208         if (rc != 0)
2209                 return rc;
2210
2211         /* get and save correct version after locking */
2212         mdt_version_get_save(info, dir, idx);
2213         return 0;
2214 }
2215
2216 /*
2217  * determine lock order of sobj and tobj
2218  *
2219  * there are two situations we need to lock tobj before sobj:
2220  * 1. sobj is child of tobj
2221  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2222  *    larger than that of tobj
2223  *
2224  * \retval      1 lock tobj before sobj
2225  * \retval      0 lock sobj before tobj
2226  * \retval      -ev negative errno upon error
2227  */
2228 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2229                                            struct mdt_object *sobj,
2230                                            struct mdt_object *tobj)
2231 {
2232         struct md_attr *ma = &info->mti_attr;
2233         struct lu_fid *spfid = &info->mti_tmp_fid1;
2234         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2235         struct lmv_mds_md_v1 *lmv;
2236         __u32 sindex;
2237         __u32 tindex;
2238         int rc;
2239
2240         /* sobj and tobj are the same */
2241         if (sobj == tobj)
2242                 return 0;
2243
2244         if (fid_is_root(mdt_object_fid(sobj)))
2245                 return 0;
2246
2247         if (fid_is_root(mdt_object_fid(tobj)))
2248                 return 1;
2249
2250         /* check whether sobj is child of tobj */
2251         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2252                            mdt_object_fid(tobj));
2253         if (rc < 0)
2254                 return rc;
2255
2256         if (rc == 1)
2257                 return 1;
2258
2259         /* check whether sobj and tobj are children of the same parent */
2260         rc = mdt_attr_get_pfid(info, sobj, spfid);
2261         if (rc)
2262                 return rc;
2263
2264         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2265         if (rc)
2266                 return rc;
2267
2268         if (!lu_fid_eq(spfid, tpfid))
2269                 return 0;
2270
2271         /* check whether sobj and tobj are sibling stripes */
2272         ma->ma_need = MA_LMV;
2273         ma->ma_valid = 0;
2274         ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2275         ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2276         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2277         if (rc)
2278                 return rc;
2279
2280         if (!(ma->ma_valid & MA_LMV))
2281                 return 0;
2282
2283         lmv = &ma->ma_lmv->lmv_md_v1;
2284         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2285                 return 0;
2286         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2287
2288         ma->ma_valid = 0;
2289         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2290         if (rc)
2291                 return rc;
2292
2293         if (!(ma->ma_valid & MA_LMV))
2294                 return -ENODATA;
2295
2296         lmv = &ma->ma_lmv->lmv_md_v1;
2297         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2298                 return -EINVAL;
2299         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2300
2301         /* check stripe index of sobj and tobj */
2302         if (sindex == tindex)
2303                 return -EINVAL;
2304
2305         return sindex < tindex ? 0 : 1;
2306 }
2307
2308 /*
2309  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2310  * 2 - srcdir child; 3 - tgtdir child.
2311  * Update on disk version of srcdir child.
2312  */
2313 static int mdt_reint_rename(struct mdt_thread_info *info,
2314                             struct mdt_lock_handle *unused)
2315 {
2316         struct mdt_device *mdt = info->mti_mdt;
2317         struct mdt_reint_record *rr = &info->mti_rr;
2318         struct md_attr *ma = &info->mti_attr;
2319         struct ptlrpc_request *req = mdt_info_req(info);
2320         struct mdt_object *msrcdir = NULL;
2321         struct mdt_object *mtgtdir = NULL;
2322         struct mdt_object *mold;
2323         struct mdt_object *mnew = NULL;
2324         struct lustre_handle rename_lh = { 0 };
2325         struct mdt_lock_handle *lh_srcdirp;
2326         struct mdt_lock_handle *lh_tgtdirp;
2327         struct mdt_lock_handle *lh_oldp = NULL;
2328         struct mdt_lock_handle *lh_newp = NULL;
2329         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2330         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2331         __u64 lock_ibits;
2332         bool reverse = false, discard = false;
2333         bool cos_incompat;
2334         int rc;
2335         ENTRY;
2336
2337         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2338                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2339                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2340
2341         if (info->mti_dlm_req)
2342                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2343
2344         if (!fid_is_md_operative(rr->rr_fid1) ||
2345             !fid_is_md_operative(rr->rr_fid2))
2346                 RETURN(-EPERM);
2347
2348         /* find both parents. */
2349         msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2350         if (IS_ERR(msrcdir))
2351                 RETURN(PTR_ERR(msrcdir));
2352
2353         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2354
2355         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2356                 mtgtdir = msrcdir;
2357                 mdt_object_get(info->mti_env, mtgtdir);
2358         } else {
2359                 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2360                 if (IS_ERR(mtgtdir))
2361                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2362         }
2363
2364         /*
2365          * Note: do not enqueue rename lock for replay request, because
2366          * if other MDT holds rename lock, but being blocked to wait for
2367          * this MDT to finish its recovery, and the failover MDT can not
2368          * get rename lock, which will cause deadlock.
2369          */
2370         if (!req_is_replay(req)) {
2371                 /*
2372                  * Normally rename RPC is handled on the MDT with the target
2373                  * directory (if target exists, it's on the MDT with the
2374                  * target), if the source directory is remote, it's a hint that
2375                  * source is remote too (this may not be true, but it won't
2376                  * cause any issue), return -EXDEV early to avoid taking
2377                  * rename_lock.
2378                  */
2379                 if (!mdt->mdt_enable_remote_rename &&
2380                     mdt_object_remote(msrcdir))
2381                         GOTO(out_put_tgtdir, rc = -EXDEV);
2382
2383                 rc = mdt_rename_lock(info, &rename_lh);
2384                 if (rc != 0) {
2385                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2386                                mdt_obd_name(mdt), rc);
2387                         GOTO(out_put_tgtdir, rc);
2388                 }
2389         }
2390
2391         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2392         if (rc < 0)
2393                 GOTO(out_unlock_rename, rc);
2394
2395         reverse = rc;
2396
2397         /* source needs to be looked up after locking source parent, otherwise
2398          * this rename may race with unlink source, and cause rename hang, see
2399          * sanityn.sh 55b, so check parents first, if later we found source is
2400          * remote, relock parents. */
2401         cos_incompat = (mdt_object_remote(msrcdir) ||
2402                         mdt_object_remote(mtgtdir));
2403
2404         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2405
2406         /* lock parents in the proper order. */
2407         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2408         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2409
2410 relock:
2411         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2412         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2413
2414         if (reverse) {
2415                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2416                                           cos_incompat);
2417                 if (rc)
2418                         GOTO(out_unlock_rename, rc);
2419
2420                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2421
2422                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2423                                           cos_incompat);
2424                 if (rc != 0) {
2425                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2426                         GOTO(out_unlock_rename, rc);
2427                 }
2428         } else {
2429                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2430                                           cos_incompat);
2431                 if (rc)
2432                         GOTO(out_unlock_rename, rc);
2433
2434                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2435
2436                 if (mtgtdir != msrcdir) {
2437                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2438                                                   cos_incompat);
2439                 } else if (!mdt_object_remote(mtgtdir) &&
2440                            lh_srcdirp->mlh_pdo_hash !=
2441                            lh_tgtdirp->mlh_pdo_hash) {
2442                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2443                                                 MDS_INODELOCK_UPDATE,
2444                                                 cos_incompat);
2445                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2446                 }
2447                 if (rc != 0) {
2448                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2449                         GOTO(out_unlock_rename, rc);
2450                 }
2451         }
2452
2453         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2454         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2455
2456         /* find mold object. */
2457         fid_zero(old_fid);
2458         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2459         if (rc != 0)
2460                 GOTO(out_unlock_parents, rc);
2461
2462         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2463                 GOTO(out_unlock_parents, rc = -EINVAL);
2464
2465         if (!fid_is_md_operative(old_fid))
2466                 GOTO(out_unlock_parents, rc = -EPERM);
2467
2468         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2469         if (IS_ERR(mold))
2470                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2471
2472         if (!mdt_object_exists(mold)) {
2473                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2474                                 &mold->mot_obj,
2475                                 "object does not exist");
2476                 GOTO(out_put_old, rc = -ENOENT);
2477         }
2478
2479         if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2480                 GOTO(out_put_old, rc = -EXDEV);
2481
2482         /* Check if @mtgtdir is subdir of @mold, before locking child
2483          * to avoid reverse locking. */
2484         if (mtgtdir != msrcdir) {
2485                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2486                                    old_fid);
2487                 if (rc) {
2488                         if (rc == 1)
2489                                 rc = -EINVAL;
2490                         GOTO(out_put_old, rc);
2491                 }
2492         }
2493
2494         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2495         /* save version after locking */
2496         mdt_version_get_save(info, mold, 2);
2497
2498         if (!cos_incompat && mdt_object_remote(mold)) {
2499                 cos_incompat = true;
2500                 mdt_object_put(info->mti_env, mold);
2501                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2502                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2503                 goto relock;
2504         }
2505
2506         /* find mnew object:
2507          * mnew target object may not exist now
2508          * lookup with version checking */
2509         fid_zero(new_fid);
2510         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2511                                       3);
2512         if (rc == 0) {
2513                 /* the new_fid should have been filled at this moment */
2514                 if (lu_fid_eq(old_fid, new_fid))
2515                         GOTO(out_put_old, rc);
2516
2517                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2518                     lu_fid_eq(new_fid, rr->rr_fid2))
2519                         GOTO(out_put_old, rc = -EINVAL);
2520
2521                 if (!fid_is_md_operative(new_fid))
2522                         GOTO(out_put_old, rc = -EPERM);
2523
2524                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2525                 if (IS_ERR(mnew))
2526                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2527
2528                 if (!mdt_object_exists(mnew)) {
2529                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2530                                         &mnew->mot_obj,
2531                                         "object does not exist");
2532                         GOTO(out_put_new, rc = -ENOENT);
2533                 }
2534
2535                 if (mdt_object_remote(mnew)) {
2536                         struct mdt_body  *repbody;
2537
2538                         /* Always send rename req to the target child MDT */
2539                         repbody = req_capsule_server_get(info->mti_pill,
2540                                                          &RMF_MDT_BODY);
2541                         LASSERT(repbody != NULL);
2542                         repbody->mbo_fid1 = *new_fid;
2543                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2544                         GOTO(out_put_new, rc = -EXDEV);
2545                 }
2546                 /* Before locking the target dir, check we do not replace
2547                  * a dir with a non-dir, otherwise it may deadlock with
2548                  * link op which tries to create a link in this dir
2549                  * back to this non-dir. */
2550                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2551                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2552                         GOTO(out_put_new, rc = -EISDIR);
2553
2554                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2555                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2556                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2557                 if (mdt_object_remote(msrcdir)) {
2558                         /* Enqueue lookup lock from the parent MDT */
2559                         rc = mdt_remote_object_lock(info, msrcdir,
2560                                                     mdt_object_fid(mold),
2561                                                     &lh_oldp->mlh_rreg_lh,
2562                                                     lh_oldp->mlh_rreg_mode,
2563                                                     MDS_INODELOCK_LOOKUP,
2564                                                     false);
2565                         if (rc != ELDLM_OK)
2566                                 GOTO(out_put_new, rc);
2567
2568                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2569                 }
2570
2571                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2572                                            cos_incompat);
2573                 if (rc != 0)
2574                         GOTO(out_unlock_old, rc);
2575
2576                 /* Check if @msrcdir is subdir of @mnew, before locking child
2577                  * to avoid reverse locking. */
2578                 if (mtgtdir != msrcdir) {
2579                         rc = mdo_is_subdir(info->mti_env,
2580                                            mdt_object_child(msrcdir), new_fid);
2581                         if (rc) {
2582                                 if (rc == 1)
2583                                         rc = -EINVAL;
2584                                 GOTO(out_unlock_old, rc);
2585                         }
2586                 }
2587
2588                 /* We used to acquire MDS_INODELOCK_FULL here but we
2589                  * can't do this now because a running HSM restore on
2590                  * the rename onto victim will hold the layout
2591                  * lock. See LU-4002. */
2592
2593                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2594                 mdt_lock_reg_init(lh_newp, LCK_EX);
2595                 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2596                                            MDS_INODELOCK_LOOKUP |
2597                                            MDS_INODELOCK_UPDATE,
2598                                            cos_incompat);
2599                 if (rc != 0)
2600                         GOTO(out_unlock_old, rc);
2601
2602                 /* get and save version after locking */
2603                 mdt_version_get_save(info, mnew, 3);
2604         } else if (rc != -EREMOTE && rc != -ENOENT) {
2605                 GOTO(out_put_old, rc);
2606         } else {
2607                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2608                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2609                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2610                 if (mdt_object_remote(msrcdir)) {
2611                         /* Enqueue lookup lock from the parent MDT */
2612                         rc = mdt_remote_object_lock(info, msrcdir,
2613                                                     mdt_object_fid(mold),
2614                                                     &lh_oldp->mlh_rreg_lh,
2615                                                     lh_oldp->mlh_rreg_mode,
2616                                                     MDS_INODELOCK_LOOKUP,
2617                                                     false);
2618                         if (rc != ELDLM_OK)
2619                                 GOTO(out_put_old, rc);
2620
2621                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2622                 }
2623
2624                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2625                                            cos_incompat);
2626                 if (rc != 0)
2627                         GOTO(out_unlock_old, rc);
2628
2629                 mdt_enoent_version_save(info, 3);
2630         }
2631
2632         /* step 5: rename it */
2633         mdt_reint_init_ma(info, ma);
2634
2635         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2636                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2637
2638         if (mnew != NULL)
2639                 mutex_lock(&mnew->mot_lov_mutex);
2640
2641         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2642                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2643                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2644                         &rr->rr_tgt_name, ma);
2645
2646         if (mnew != NULL)
2647                 mutex_unlock(&mnew->mot_lov_mutex);
2648
2649         /* handle last link of tgt object */
2650         if (rc == 0) {
2651                 mdt_counter_incr(req, LPROC_MDT_RENAME);
2652                 if (mnew) {
2653                         mdt_handle_last_unlink(info, mnew, ma);
2654                         discard = mdt_dom_check_for_discard(info, mnew);
2655                 }
2656                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2657                                          msrcdir, mtgtdir);
2658         }
2659
2660         EXIT;
2661         if (mnew != NULL)
2662                 mdt_object_unlock(info, mnew, lh_newp, rc);
2663 out_unlock_old:
2664         mdt_object_unlock(info, mold, lh_oldp, rc);
2665 out_put_new:
2666         if (mnew && !discard)
2667                 mdt_object_put(info->mti_env, mnew);
2668 out_put_old:
2669         mdt_object_put(info->mti_env, mold);
2670 out_unlock_parents:
2671         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2672         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2673 out_unlock_rename:
2674         if (lustre_handle_is_used(&rename_lh))
2675                 mdt_rename_unlock(&rename_lh);
2676 out_put_tgtdir:
2677         mdt_object_put(info->mti_env, mtgtdir);
2678 out_put_srcdir:
2679         mdt_object_put(info->mti_env, msrcdir);
2680
2681         /* The DoM discard can be done right in the place above where it is
2682          * assigned, meanwhile it is done here after rename unlock due to
2683          * compatibility with old clients, for them the discard blocks
2684          * the main thread until completion. Check LU-11359 for details.
2685          */
2686         if (discard) {
2687                 mdt_dom_discard_data(info, mnew);
2688                 mdt_object_put(info->mti_env, mnew);
2689         }
2690         return rc;
2691 }
2692
2693 static int mdt_reint_resync(struct mdt_thread_info *info,
2694                             struct mdt_lock_handle *lhc)
2695 {
2696         struct mdt_reint_record *rr = &info->mti_rr;
2697         struct ptlrpc_request   *req = mdt_info_req(info);
2698         struct md_attr          *ma = &info->mti_attr;
2699         struct mdt_object       *mo;
2700         struct ldlm_lock        *lease;
2701         struct mdt_body         *repbody;
2702         struct md_layout_change  layout = { .mlc_mirror_id = rr->rr_mirror_id };
2703         bool                     lease_broken;
2704         int                      rc, rc2;
2705         ENTRY;
2706
2707         DEBUG_REQ(D_INODE, req, DFID": FLR file resync\n", PFID(rr->rr_fid1));
2708
2709         if (info->mti_dlm_req)
2710                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2711
2712         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2713         if (IS_ERR(mo))
2714                 GOTO(out, rc = PTR_ERR(mo));
2715
2716         if (!mdt_object_exists(mo))
2717                 GOTO(out_obj, rc = -ENOENT);
2718
2719         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2720                 GOTO(out_obj, rc = -EINVAL);
2721
2722         if (mdt_object_remote(mo))
2723                 GOTO(out_obj, rc = -EREMOTE);
2724
2725         lease = ldlm_handle2lock(rr->rr_lease_handle);
2726         if (lease == NULL)
2727                 GOTO(out_obj, rc = -ESTALE);
2728
2729         /* It's really necessary to grab open_sem and check if the lease lock
2730          * has been lost. There would exist a concurrent writer coming in and
2731          * generating some dirty data in memory cache, the writeback would fail
2732          * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2733         if (!down_write_trylock(&mo->mot_open_sem))
2734                 GOTO(out_put_lease, rc = -EBUSY);
2735
2736         lock_res_and_lock(lease);
2737         lease_broken = ldlm_is_cancel(lease);
2738         unlock_res_and_lock(lease);
2739         if (lease_broken)
2740                 GOTO(out_unlock, rc = -EBUSY);
2741
2742         /* the file has yet opened by anyone else after we took the lease. */
2743         layout.mlc_opc = MD_LAYOUT_RESYNC;
2744         rc = mdt_layout_change(info, mo, &layout);
2745         if (rc)
2746                 GOTO(out_unlock, rc);
2747
2748         ma->ma_need = MA_INODE;
2749         ma->ma_valid = 0;
2750         rc = mdt_attr_get_complex(info, mo, ma);
2751         if (rc != 0)
2752                 GOTO(out_unlock, rc);
2753
2754         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2755         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2756
2757         EXIT;
2758 out_unlock:
2759         up_write(&mo->mot_open_sem);
2760 out_put_lease:
2761         LDLM_LOCK_PUT(lease);
2762 out_obj:
2763         mdt_object_put(info->mti_env, mo);
2764 out:
2765         mdt_client_compatibility(info);
2766         rc2 = mdt_fix_reply(info);
2767         if (rc == 0)
2768                 rc = rc2;
2769         return rc;
2770 }
2771
2772 struct mdt_reinter {
2773         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2774         enum lprocfs_extra_opc mr_extra_opc;
2775 };
2776
2777 static const struct mdt_reinter mdt_reinters[] = {
2778         [REINT_SETATTR] = {
2779                 .mr_handler = &mdt_reint_setattr,
2780                 .mr_extra_opc = MDS_REINT_SETATTR,
2781         },
2782         [REINT_CREATE] = {
2783                 .mr_handler = &mdt_reint_create,
2784                 .mr_extra_opc = MDS_REINT_CREATE,
2785         },
2786         [REINT_LINK] = {
2787                 .mr_handler = &mdt_reint_link,
2788                 .mr_extra_opc = MDS_REINT_LINK,
2789         },
2790         [REINT_UNLINK] = {
2791                 .mr_handler = &mdt_reint_unlink,
2792                 .mr_extra_opc = MDS_REINT_UNLINK,
2793         },
2794         [REINT_RENAME] = {
2795                 .mr_handler = &mdt_reint_rename,
2796                 .mr_extra_opc = MDS_REINT_RENAME,
2797         },
2798         [REINT_OPEN] = {
2799                 .mr_handler = &mdt_reint_open,
2800                 .mr_extra_opc = MDS_REINT_OPEN,
2801         },
2802         [REINT_SETXATTR] = {
2803                 .mr_handler = &mdt_reint_setxattr,
2804                 .mr_extra_opc = MDS_REINT_SETXATTR,
2805         },
2806         [REINT_RMENTRY] = {
2807                 .mr_handler = &mdt_reint_unlink,
2808                 .mr_extra_opc = MDS_REINT_UNLINK,
2809         },
2810         [REINT_MIGRATE] = {
2811                 .mr_handler = &mdt_reint_migrate,
2812                 .mr_extra_opc = MDS_REINT_RENAME,
2813         },
2814         [REINT_RESYNC] = {
2815                 .mr_handler = &mdt_reint_resync,
2816                 .mr_extra_opc = MDS_REINT_RESYNC,
2817         },
2818 };
2819
2820 int mdt_reint_rec(struct mdt_thread_info *info,
2821                   struct mdt_lock_handle *lhc)
2822 {
2823         const struct mdt_reinter *mr;
2824         int rc;
2825         ENTRY;
2826
2827         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2828                 RETURN(-EPROTO);
2829
2830         mr = &mdt_reinters[info->mti_rr.rr_opcode];
2831         if (mr->mr_handler == NULL)
2832                 RETURN(-EPROTO);
2833
2834         rc = (*mr->mr_handler)(info, lhc);
2835
2836         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2837                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
2838
2839         RETURN(rc);
2840 }