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),
 367                   "Create ("DNAME"->"DFID") 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         OBD_RACE(OBD_FAIL_PTLRPC_RESEND_RACE);
 653
 654         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
 655         mo = mdt_object_find(info->mti_env, mdt, rr->rr_fid1);
 656         if (IS_ERR(mo))
 657                 GOTO(out, rc = PTR_ERR(mo));
 658
 659         if (!mdt_object_exists(mo))
 660                 GOTO(out_put, rc = -ENOENT);
 661
 662         if (mdt_object_remote(mo))
 663                 GOTO(out_put, rc = -EREMOTE);
 664
 665         ma->ma_enable_chprojid_gid = mdt->mdt_enable_chprojid_gid;
 666         /* revoke lease lock if size is going to be changed */
 667         if (unlikely(ma->ma_attr.la_valid & LA_SIZE &&
 668                      !(ma->ma_attr_flags & MDS_TRUNC_KEEP_LEASE) &&
 669                      atomic_read(&mo->mot_lease_count) > 0)) {
 670                 down_read(&mo->mot_open_sem);
 671
 672                 if (atomic_read(&mo->mot_lease_count) > 0) { /* lease exists */
 673                         lhc = &info->mti_lh[MDT_LH_LOCAL];
 674                         mdt_lock_reg_init(lhc, LCK_CW);
 675
 676                         rc = mdt_object_lock(info, mo, lhc, MDS_INODELOCK_OPEN);
 677                         if (rc != 0) {
 678                                 up_read(&mo->mot_open_sem);
 679                                 GOTO(out_put, rc);
 680                         }
 681
 682                         /* revoke lease lock */
 683                         mdt_object_unlock(info, mo, lhc, 1);
 684                 }
 685                 up_read(&mo->mot_open_sem);
 686         }
 687
 688         if (ma->ma_attr.la_valid & LA_SIZE || rr->rr_flags & MRF_OPEN_TRUNC) {
 689                 /* Check write access for the O_TRUNC case */
 690                 if (mdt_write_read(mo) < 0)
 691                         GOTO(out_put, rc = -ETXTBSY);
 692
 693                 /* LU-10286: compatibility check for FLR.
 694                  * Please check the comment in mdt_finish_open() for details */
 695                 if (!exp_connect_flr(info->mti_exp) ||
 696                     !exp_connect_overstriping(info->mti_exp)) {
 697                         rc = mdt_big_xattr_get(info, mo, XATTR_NAME_LOV);
 698                         if (rc < 0 && rc != -ENODATA)
 699                                 GOTO(out_put, rc);
 700
 701                         if (!exp_connect_flr(info->mti_exp)) {
 702                                 if (rc > 0 &&
 703                                     mdt_lmm_is_flr(info->mti_big_lmm))
 704                                         GOTO(out_put, rc = -EOPNOTSUPP);
 705                         }
 706
 707                         if (!exp_connect_overstriping(info->mti_exp)) {
 708                                 if (rc > 0 &&
 709                                     mdt_lmm_is_overstriping(info->mti_big_lmm))
 710                                         GOTO(out_put, rc = -EOPNOTSUPP);
 711                         }
 712                 }
 713
 714                 /* For truncate, the file size sent from client
 715                  * is believable, but the blocks are incorrect,
 716                  * which makes the block size in LSOM attribute
 717                  * inconsisent with the real block size.
 718                  */
 719                 rc = mdt_lsom_update(info, mo, true);
 720                 if (rc)
 721                         GOTO(out_put, rc);
 722         }
 723
 724         if ((ma->ma_valid & MA_INODE) && ma->ma_attr.la_valid) {
 725                 if (ma->ma_valid & MA_LOV)
 726                         GOTO(out_put, rc = -EPROTO);
 727
 728                 /* MDT supports FMD for regular files due to Data-on-MDT */
 729                 if (S_ISREG(lu_object_attr(&mo->mot_obj)) &&
 730                     ma->ma_attr.la_valid & (LA_ATIME | LA_MTIME | LA_CTIME))
 731                         tgt_fmd_update(info->mti_exp, mdt_object_fid(mo),
 732                                        req->rq_xid);
 733
 734                 rc = mdt_attr_set(info, mo, ma);
 735                 if (rc)
 736                         GOTO(out_put, rc);
 737         } else if ((ma->ma_valid & (MA_LOV | MA_LMV)) &&
 738                    (ma->ma_valid & MA_INODE)) {
 739                 struct lu_buf *buf = &info->mti_buf;
 740                 struct lu_ucred *uc = mdt_ucred(info);
 741                 struct mdt_lock_handle *lh;
 742                 const char *name;
 743                 __u64 lockpart = MDS_INODELOCK_XATTR;
 744
 745                 /* reject if either remote or striped dir is disabled */
 746                 if (ma->ma_valid & MA_LMV) {
 747                         if (!mdt->mdt_enable_remote_dir ||
 748                             !mdt->mdt_enable_striped_dir)
 749                                 GOTO(out_put, rc = -EPERM);
 750
 751                         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
 752                             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
 753                             mdt->mdt_enable_remote_dir_gid != -1)
 754                                 GOTO(out_put, rc = -EPERM);
 755                 }
 756
 757                 if (!S_ISDIR(lu_object_attr(&mo->mot_obj)))
 758                         GOTO(out_put, rc = -ENOTDIR);
 759
 760                 if (ma->ma_attr.la_valid != 0)
 761                         GOTO(out_put, rc = -EPROTO);
 762
 763                 if (ma->ma_valid & MA_LOV) {
 764                         buf->lb_buf = ma->ma_lmm;
 765                         buf->lb_len = ma->ma_lmm_size;
 766                         name = XATTR_NAME_LOV;
 767                 } else {
 768                         struct lmv_user_md *lmu = &ma->ma_lmv->lmv_user_md;
 769
 770                         buf->lb_buf = lmu;
 771                         buf->lb_len = ma->ma_lmv_size;
 772                         name = XATTR_NAME_DEFAULT_LMV;
 773                         /* force client to update dir default layout */
 774                         lockpart |= MDS_INODELOCK_LOOKUP;
 775                 }
 776
 777                 lh = &info->mti_lh[MDT_LH_PARENT];
 778                 mdt_lock_reg_init(lh, LCK_PW);
 779
 780                 rc = mdt_object_lock(info, mo, lh, lockpart);
 781                 if (rc != 0)
 782                         GOTO(out_put, rc);
 783
 784                 rc = mo_xattr_set(info->mti_env, mdt_object_child(mo), buf,
 785                                   name, 0);
 786
 787                 mdt_object_unlock(info, mo, lh, rc);
 788                 if (rc)
 789                         GOTO(out_put, rc);
 790         } else {
 791                 GOTO(out_put, rc = -EPROTO);
 792         }
 793
 794         /* If file data is modified, add the dirty flag */
 795         if (ma->ma_attr_flags & MDS_DATA_MODIFIED)
 796                 rc = mdt_add_dirty_flag(info, mo, ma);
 797
 798         ma->ma_need = MA_INODE;
 799         ma->ma_valid = 0;
 800         rc = mdt_attr_get_complex(info, mo, ma);
 801         if (rc != 0)
 802                 GOTO(out_put, rc);
 803
 804         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
 805
 806         EXIT;
 807 out_put:
 808         mdt_object_put(info->mti_env, mo);
 809 out:
 810         if (rc == 0)
 811                 mdt_counter_incr(req, LPROC_MDT_SETATTR);
 812
 813         mdt_client_compatibility(info);
 814         rc2 = mdt_fix_reply(info);
 815         if (rc == 0)
 816                 rc = rc2;
 817         return rc;
 818 }
 819
 820 static int mdt_reint_create(struct mdt_thread_info *info,
 821                             struct mdt_lock_handle *lhc)
 822 {
 823         struct ptlrpc_request   *req = mdt_info_req(info);
 824         int                     rc;
 825         ENTRY;
 826
 827         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_CREATE))
 828                 RETURN(err_serious(-ESTALE));
 829
 830         if (info->mti_dlm_req)
 831                 ldlm_request_cancel(mdt_info_req(info),
 832                                     info->mti_dlm_req, 0, LATF_SKIP);
 833
 834         if (!lu_name_is_valid(&info->mti_rr.rr_name))
 835                 RETURN(-EPROTO);
 836
 837         switch (info->mti_attr.ma_attr.la_mode & S_IFMT) {
 838         case S_IFDIR:
 839                 mdt_counter_incr(req, LPROC_MDT_MKDIR);
 840                 break;
 841         case S_IFREG:
 842         case S_IFLNK:
 843         case S_IFCHR:
 844         case S_IFBLK:
 845         case S_IFIFO:
 846         case S_IFSOCK:
 847                 /* Special file should stay on the same node as parent. */
 848                 mdt_counter_incr(req, LPROC_MDT_MKNOD);
 849                 break;
 850         default:
 851                 CERROR("%s: Unsupported mode %o\n",
 852                        mdt_obd_name(info->mti_mdt),
 853                        info->mti_attr.ma_attr.la_mode);
 854                 RETURN(err_serious(-EOPNOTSUPP));
 855         }
 856
 857         rc = mdt_create(info);
 858         RETURN(rc);
 859 }
 860
 861 /*
 862  * VBR: save parent version in reply and child version getting by its name.
 863  * Version of child is getting and checking during its lookup. If
 864  */
 865 static int mdt_reint_unlink(struct mdt_thread_info *info,
 866                             struct mdt_lock_handle *lhc)
 867 {
 868         struct mdt_reint_record *rr = &info->mti_rr;
 869         struct ptlrpc_request *req = mdt_info_req(info);
 870         struct md_attr *ma = &info->mti_attr;
 871         struct lu_fid *child_fid = &info->mti_tmp_fid1;
 872         struct mdt_object *mp;
 873         struct mdt_object *mc;
 874         struct mdt_lock_handle *parent_lh;
 875         struct mdt_lock_handle *child_lh;
 876         struct ldlm_enqueue_info *einfo = &info->mti_einfo[0];
 877         __u64 lock_ibits;
 878         bool cos_incompat = false;
 879         int no_name = 0;
 880         int rc;
 881
 882         ENTRY;
 883
 884         DEBUG_REQ(D_INODE, req, "unlink "DFID"/"DNAME"", PFID(rr->rr_fid1),
 885                   PNAME(&rr->rr_name));
 886
 887         if (info->mti_dlm_req)
 888                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
 889
 890         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_UNLINK))
 891                 RETURN(err_serious(-ENOENT));
 892
 893         if (!fid_is_md_operative(rr->rr_fid1))
 894                 RETURN(-EPERM);
 895
 896         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
 897         if (IS_ERR(mp))
 898                 RETURN(PTR_ERR(mp));
 899
 900         if (mdt_object_remote(mp)) {
 901                 cos_incompat = true;
 902         } else {
 903                 rc = mdt_version_get_check_save(info, mp, 0);
 904                 if (rc)
 905                         GOTO(put_parent, rc);
 906         }
 907
 908 relock:
 909         parent_lh = &info->mti_lh[MDT_LH_PARENT];
 910         mdt_lock_pdo_init(parent_lh, LCK_PW, &rr->rr_name);
 911         rc = mdt_reint_object_lock(info, mp, parent_lh, MDS_INODELOCK_UPDATE,
 912                                    cos_incompat);
 913         if (rc != 0)
 914                 GOTO(put_parent, rc);
 915
 916         /* lookup child object along with version checking */
 917         fid_zero(child_fid);
 918         rc = mdt_lookup_version_check(info, mp, &rr->rr_name, child_fid, 1);
 919         if (rc != 0) {
 920                 /* Name might not be able to find during resend of
 921                  * remote unlink, considering following case.
 922                  * dir_A is a remote directory, the name entry of
 923                  * dir_A is on MDT0, the directory is on MDT1,
 924                  *
 925                  * 1. client sends unlink req to MDT1.
 926                  * 2. MDT1 sends name delete update to MDT0.
 927                  * 3. name entry is being deleted in MDT0 synchronously.
 928                  * 4. MDT1 is restarted.
 929                  * 5. client resends unlink req to MDT1. So it can not
 930                  *    find the name entry on MDT0 anymore.
 931                  * In this case, MDT1 only needs to destory the local
 932                  * directory.
 933                  * */
 934                 if (mdt_object_remote(mp) && rc == -ENOENT &&
 935                     !fid_is_zero(rr->rr_fid2) &&
 936                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) {
 937                         no_name = 1;
 938                         *child_fid = *rr->rr_fid2;
 939                  } else {
 940                         GOTO(unlock_parent, rc);
 941                  }
 942         }
 943
 944         if (!fid_is_md_operative(child_fid))
 945                 GOTO(unlock_parent, rc = -EPERM);
 946
 947         /* We will lock the child regardless it is local or remote. No harm. */
 948         mc = mdt_object_find(info->mti_env, info->mti_mdt, child_fid);
 949         if (IS_ERR(mc))
 950                 GOTO(unlock_parent, rc = PTR_ERR(mc));
 951
 952         if (!cos_incompat) {
 953                 rc = mdt_object_striped(info, mc);
 954                 if (rc < 0)
 955                         GOTO(put_child, rc);
 956
 957                 cos_incompat = rc;
 958                 if (cos_incompat) {
 959                         mdt_object_put(info->mti_env, mc);
 960                         mdt_object_unlock(info, mp, parent_lh, -EAGAIN);
 961                         goto relock;
 962                 }
 963         }
 964
 965         child_lh = &info->mti_lh[MDT_LH_CHILD];
 966         mdt_lock_reg_init(child_lh, LCK_EX);
 967         if (info->mti_spec.sp_rm_entry) {
 968                 struct lu_ucred *uc  = mdt_ucred(info);
 969
 970                 if (!mdt_is_dne_client(req->rq_export))
 971                         /* Return -ENOTSUPP for old client */
 972                         GOTO(put_child, rc = -ENOTSUPP);
 973
 974                 if (!md_capable(uc, CFS_CAP_SYS_ADMIN))
 975                         GOTO(put_child, rc = -EPERM);
 976
 977                 ma->ma_need = MA_INODE;
 978                 ma->ma_valid = 0;
 979                 rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
 980                                 NULL, &rr->rr_name, ma, no_name);
 981                 GOTO(put_child, rc);
 982         }
 983
 984         if (mdt_object_remote(mc)) {
 985                 struct mdt_body  *repbody;
 986
 987                 if (!fid_is_zero(rr->rr_fid2)) {
 988                         CDEBUG(D_INFO, "%s: name "DNAME" cannot find "DFID"\n",
 989                                mdt_obd_name(info->mti_mdt),
 990                                PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 991                         GOTO(put_child, rc = -ENOENT);
 992                 }
 993                 CDEBUG(D_INFO, "%s: name "DNAME": "DFID" is on another MDT\n",
 994                        mdt_obd_name(info->mti_mdt),
 995                        PNAME(&rr->rr_name), PFID(mdt_object_fid(mc)));
 996
 997                 if (!mdt_is_dne_client(req->rq_export))
 998                         /* Return -ENOTSUPP for old client */
 999                         GOTO(put_child, rc = -ENOTSUPP);
1000
1001                 /* Revoke the LOOKUP lock of the remote object granted by
1002                  * this MDT. Since the unlink will happen on another MDT,
1003                  * it will release the LOOKUP lock right away. Then What
1004                  * would happen if another client try to grab the LOOKUP
1005                  * lock at the same time with unlink XXX */
1006                 mdt_object_lock(info, mc, child_lh, MDS_INODELOCK_LOOKUP);
1007                 repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1008                 LASSERT(repbody != NULL);
1009                 repbody->mbo_fid1 = *mdt_object_fid(mc);
1010                 repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
1011                 GOTO(unlock_child, rc = -EREMOTE);
1012         }
1013         /* We used to acquire MDS_INODELOCK_FULL here but we can't do
1014          * this now because a running HSM restore on the child (unlink
1015          * victim) will hold the layout lock. See LU-4002. */
1016         lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE;
1017         if (mdt_object_remote(mp)) {
1018                 /* Enqueue lookup lock from parent MDT */
1019                 rc = mdt_remote_object_lock(info, mp, mdt_object_fid(mc),
1020                                             &child_lh->mlh_rreg_lh,
1021                                             child_lh->mlh_rreg_mode,
1022                                             MDS_INODELOCK_LOOKUP, false);
1023                 if (rc != ELDLM_OK)
1024                         GOTO(put_child, rc);
1025
1026                 lock_ibits &= ~MDS_INODELOCK_LOOKUP;
1027         }
1028
1029         rc = mdt_reint_striped_lock(info, mc, child_lh, lock_ibits, einfo,
1030                                     cos_incompat);
1031         if (rc != 0)
1032                 GOTO(put_child, rc);
1033
1034         /*
1035          * Now we can only make sure we need MA_INODE, in mdd layer, will check
1036          * whether need MA_LOV and MA_COOKIE.
1037          */
1038         ma->ma_need = MA_INODE;
1039         ma->ma_valid = 0;
1040
1041         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1042                        OBD_FAIL_MDS_REINT_UNLINK_WRITE);
1043         /* save version when object is locked */
1044         mdt_version_get_save(info, mc, 1);
1045
1046         mutex_lock(&mc->mot_lov_mutex);
1047
1048         rc = mdo_unlink(info->mti_env, mdt_object_child(mp),
1049                         mdt_object_child(mc), &rr->rr_name, ma, no_name);
1050
1051         mutex_unlock(&mc->mot_lov_mutex);
1052         if (rc != 0)
1053                 GOTO(unlock_child, rc);
1054
1055         if (!lu_object_is_dying(&mc->mot_header)) {
1056                 rc = mdt_attr_get_complex(info, mc, ma);
1057                 if (rc)
1058                         GOTO(out_stat, rc);
1059         } else if (mdt_dom_check_for_discard(info, mc)) {
1060                 mdt_dom_discard_data(info, mc);
1061         }
1062         mdt_handle_last_unlink(info, mc, ma);
1063
1064 out_stat:
1065         if (ma->ma_valid & MA_INODE) {
1066                 switch (ma->ma_attr.la_mode & S_IFMT) {
1067                 case S_IFDIR:
1068                         mdt_counter_incr(req, LPROC_MDT_RMDIR);
1069                         break;
1070                 case S_IFREG:
1071                 case S_IFLNK:
1072                 case S_IFCHR:
1073                 case S_IFBLK:
1074                 case S_IFIFO:
1075                 case S_IFSOCK:
1076                         mdt_counter_incr(req, LPROC_MDT_UNLINK);
1077                         break;
1078                 default:
1079                         LASSERTF(0, "bad file type %o unlinking\n",
1080                                 ma->ma_attr.la_mode);
1081                 }
1082         }
1083
1084         EXIT;
1085
1086 unlock_child:
1087         mdt_reint_striped_unlock(info, mc, child_lh, einfo, rc);
1088 put_child:
1089         mdt_object_put(info->mti_env, mc);
1090 unlock_parent:
1091         mdt_object_unlock(info, mp, parent_lh, rc);
1092 put_parent:
1093         mdt_object_put(info->mti_env, mp);
1094         CFS_RACE_WAKEUP(OBD_FAIL_OBD_ZERO_NLINK_RACE);
1095         return rc;
1096 }
1097
1098 /*
1099  * VBR: save versions in reply: 0 - parent; 1 - child by fid; 2 - target by
1100  * name.
1101  */
1102 static int mdt_reint_link(struct mdt_thread_info *info,
1103                           struct mdt_lock_handle *lhc)
1104 {
1105         struct mdt_reint_record *rr = &info->mti_rr;
1106         struct ptlrpc_request   *req = mdt_info_req(info);
1107         struct md_attr          *ma = &info->mti_attr;
1108         struct mdt_object       *ms;
1109         struct mdt_object       *mp;
1110         struct mdt_lock_handle  *lhs;
1111         struct mdt_lock_handle  *lhp;
1112         bool cos_incompat;
1113         int rc;
1114         ENTRY;
1115
1116         DEBUG_REQ(D_INODE, req, "link "DFID" to "DFID"/"DNAME,
1117                   PFID(rr->rr_fid1), PFID(rr->rr_fid2), PNAME(&rr->rr_name));
1118
1119         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_LINK))
1120                 RETURN(err_serious(-ENOENT));
1121
1122         if (OBD_FAIL_PRECHECK(OBD_FAIL_PTLRPC_RESEND_RACE)) {
1123                 req->rq_no_reply = 1;
1124                 RETURN(err_serious(-ENOENT));
1125         }
1126
1127         if (info->mti_dlm_req)
1128                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
1129
1130         /* Invalid case so return error immediately instead of
1131          * processing it */
1132         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2))
1133                 RETURN(-EPERM);
1134
1135         if (!fid_is_md_operative(rr->rr_fid1) ||
1136             !fid_is_md_operative(rr->rr_fid2))
1137                 RETURN(-EPERM);
1138
1139         /* step 1: find target parent dir */
1140         mp = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid2);
1141         if (IS_ERR(mp))
1142                 RETURN(PTR_ERR(mp));
1143
1144         rc = mdt_version_get_check_save(info, mp, 0);
1145         if (rc)
1146                 GOTO(put_parent, rc);
1147
1148         /* step 2: find source */
1149         ms = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
1150         if (IS_ERR(ms))
1151                 GOTO(put_parent, rc = PTR_ERR(ms));
1152
1153         if (!mdt_object_exists(ms)) {
1154                 CDEBUG(D_INFO, "%s: "DFID" does not exist.\n",
1155                        mdt_obd_name(info->mti_mdt), PFID(rr->rr_fid1));
1156                 GOTO(put_source, rc = -ENOENT);
1157         }
1158
1159         cos_incompat = (mdt_object_remote(mp) || mdt_object_remote(ms));
1160
1161         lhp = &info->mti_lh[MDT_LH_PARENT];
1162         mdt_lock_pdo_init(lhp, LCK_PW, &rr->rr_name);
1163         rc = mdt_reint_object_lock(info, mp, lhp, MDS_INODELOCK_UPDATE,
1164                                    cos_incompat);
1165         if (rc != 0)
1166                 GOTO(put_source, rc);
1167
1168         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
1169
1170         lhs = &info->mti_lh[MDT_LH_CHILD];
1171         mdt_lock_reg_init(lhs, LCK_EX);
1172         rc = mdt_reint_object_lock(info, ms, lhs,
1173                                    MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1174                                    cos_incompat);
1175         if (rc != 0)
1176                 GOTO(unlock_parent, rc);
1177
1178         /* step 3: link it */
1179         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
1180                         OBD_FAIL_MDS_REINT_LINK_WRITE);
1181
1182         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(ms));
1183         rc = mdt_version_get_check_save(info, ms, 1);
1184         if (rc)
1185                 GOTO(unlock_source, rc);
1186
1187         /** check target version by name during replay */
1188         rc = mdt_lookup_version_check(info, mp, &rr->rr_name,
1189                                       &info->mti_tmp_fid1, 2);
1190         if (rc != 0 && rc != -ENOENT)
1191                 GOTO(unlock_source, rc);
1192         /* save version of file name for replay, it must be ENOENT here */
1193         if (!req_is_replay(mdt_info_req(info))) {
1194                 if (rc != -ENOENT) {
1195                         CDEBUG(D_INFO, "link target "DNAME" existed!\n",
1196                                PNAME(&rr->rr_name));
1197                         GOTO(unlock_source, rc = -EEXIST);
1198                 }
1199                 info->mti_ver[2] = ENOENT_VERSION;
1200                 mdt_version_save(mdt_info_req(info), info->mti_ver[2], 2);
1201         }
1202
1203         rc = mdo_link(info->mti_env, mdt_object_child(mp),
1204                       mdt_object_child(ms), &rr->rr_name, ma);
1205
1206         if (rc == 0)
1207                 mdt_counter_incr(req, LPROC_MDT_LINK);
1208
1209         EXIT;
1210 unlock_source:
1211         mdt_object_unlock(info, ms, lhs, rc);
1212 unlock_parent:
1213         mdt_object_unlock(info, mp, lhp, rc);
1214 put_source:
1215         mdt_object_put(info->mti_env, ms);
1216 put_parent:
1217         mdt_object_put(info->mti_env, mp);
1218         return rc;
1219 }
1220 /**
1221  * lock the part of the directory according to the hash of the name
1222  * (lh->mlh_pdo_hash) in parallel directory lock.
1223  */
1224 static int mdt_pdir_hash_lock(struct mdt_thread_info *info,
1225                               struct mdt_lock_handle *lh,
1226                               struct mdt_object *obj, __u64 ibits,
1227                               bool cos_incompat)
1228 {
1229         struct ldlm_res_id *res = &info->mti_res_id;
1230         struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1231         union ldlm_policy_data *policy = &info->mti_policy;
1232         __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1233         int rc;
1234
1235         /*
1236          * Finish res_id initializing by name hash marking part of
1237          * directory which is taking modification.
1238          */
1239         LASSERT(lh->mlh_pdo_hash != 0);
1240         fid_build_pdo_res_name(mdt_object_fid(obj), lh->mlh_pdo_hash, res);
1241         memset(policy, 0, sizeof(*policy));
1242         policy->l_inodebits.bits = ibits;
1243         if (cos_incompat &&
1244             (lh->mlh_reg_mode == LCK_PW || lh->mlh_reg_mode == LCK_EX))
1245                 dlmflags |= LDLM_FL_COS_INCOMPAT;
1246         /*
1247          * Use LDLM_FL_LOCAL_ONLY for this lock. We do not know yet if it is
1248          * going to be sent to client. If it is - mdt_intent_policy() path will
1249          * fix it up and turn FL_LOCAL flag off.
1250          */
1251         rc = mdt_fid_lock(info->mti_env, ns, &lh->mlh_reg_lh, lh->mlh_reg_mode,
1252                           policy, res, dlmflags,
1253                           &info->mti_exp->exp_handle.h_cookie);
1254         return rc;
1255 }
1256
1257 /**
1258  * Get BFL lock for rename or migrate process.
1259  **/
1260 static int mdt_rename_lock(struct mdt_thread_info *info,
1261                            struct lustre_handle *lh)
1262 {
1263         int     rc;
1264         ENTRY;
1265
1266         if (mdt_seq_site(info->mti_mdt)->ss_node_id != 0) {
1267                 struct lu_fid *fid = &info->mti_tmp_fid1;
1268                 struct mdt_object *obj;
1269
1270                 /* XXX, right now, it has to use object API to
1271                  * enqueue lock cross MDT, so it will enqueue
1272                  * rename lock(with LUSTRE_BFL_FID) by root object */
1273                 lu_root_fid(fid);
1274                 obj = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1275                 if (IS_ERR(obj))
1276                         RETURN(PTR_ERR(obj));
1277
1278                 rc = mdt_remote_object_lock(info, obj,
1279                                             &LUSTRE_BFL_FID, lh,
1280                                             LCK_EX,
1281                                             MDS_INODELOCK_UPDATE, false);
1282                 mdt_object_put(info->mti_env, obj);
1283         } else {
1284                 struct ldlm_namespace *ns = info->mti_mdt->mdt_namespace;
1285                 union ldlm_policy_data *policy = &info->mti_policy;
1286                 struct ldlm_res_id *res_id = &info->mti_res_id;
1287                 __u64 flags = 0;
1288
1289                 fid_build_reg_res_name(&LUSTRE_BFL_FID, res_id);
1290                 memset(policy, 0, sizeof *policy);
1291                 policy->l_inodebits.bits = MDS_INODELOCK_UPDATE;
1292                 flags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB;
1293                 rc = ldlm_cli_enqueue_local(info->mti_env, ns, res_id,
1294                                             LDLM_IBITS, policy, LCK_EX, &flags,
1295                                             ldlm_blocking_ast,
1296                                             ldlm_completion_ast, NULL, NULL, 0,
1297                                             LVB_T_NONE,
1298                                             &info->mti_exp->exp_handle.h_cookie,
1299                                             lh);
1300                 RETURN(rc);
1301         }
1302         RETURN(rc);
1303 }
1304
1305 static void mdt_rename_unlock(struct lustre_handle *lh)
1306 {
1307         ENTRY;
1308         LASSERT(lustre_handle_is_used(lh));
1309         /* Cancel the single rename lock right away */
1310         ldlm_lock_decref_and_cancel(lh, LCK_EX);
1311         EXIT;
1312 }
1313
1314 static struct mdt_object *mdt_parent_find_check(struct mdt_thread_info *info,
1315                                                 const struct lu_fid *fid,
1316                                                 int idx)
1317 {
1318         struct mdt_object *dir;
1319         int rc;
1320
1321         ENTRY;
1322
1323         dir = mdt_object_find(info->mti_env, info->mti_mdt, fid);
1324         if (IS_ERR(dir))
1325                 RETURN(dir);
1326
1327         /* check early, the real version will be saved after locking */
1328         rc = mdt_version_get_check(info, dir, idx);
1329         if (rc)
1330                 GOTO(out_put, rc);
1331
1332         if (!mdt_object_exists(dir))
1333                 GOTO(out_put, rc = -ENOENT);
1334
1335         if (!S_ISDIR(lu_object_attr(&dir->mot_obj)))
1336                 GOTO(out_put, rc = -ENOTDIR);
1337
1338         RETURN(dir);
1339 out_put:
1340         mdt_object_put(info->mti_env, dir);
1341         return ERR_PTR(rc);
1342 }
1343
1344 /*
1345  * in case obj is remote obj on its parent, revoke LOOKUP lock,
1346  * herein we don't really check it, just do revoke.
1347  */
1348 int mdt_revoke_remote_lookup_lock(struct mdt_thread_info *info,
1349                                   struct mdt_object *pobj,
1350                                   struct mdt_object *obj)
1351 {
1352         struct mdt_lock_handle *lh = &info->mti_lh[MDT_LH_LOCAL];
1353         int rc;
1354
1355         mdt_lock_handle_init(lh);
1356         mdt_lock_reg_init(lh, LCK_EX);
1357
1358         if (mdt_object_remote(pobj)) {
1359                 /* don't bother to check if pobj and obj are on the same MDT. */
1360                 rc = mdt_remote_object_lock(info, pobj, mdt_object_fid(obj),
1361                                             &lh->mlh_rreg_lh, LCK_EX,
1362                                             MDS_INODELOCK_LOOKUP, false);
1363         } else if (mdt_object_remote(obj)) {
1364                 struct ldlm_res_id *res = &info->mti_res_id;
1365                 union ldlm_policy_data *policy = &info->mti_policy;
1366                 __u64 dlmflags = LDLM_FL_LOCAL_ONLY | LDLM_FL_ATOMIC_CB |
1367                                  LDLM_FL_COS_INCOMPAT;
1368
1369                 fid_build_reg_res_name(mdt_object_fid(obj), res);
1370                 memset(policy, 0, sizeof(*policy));
1371                 policy->l_inodebits.bits = MDS_INODELOCK_LOOKUP;
1372                 rc = mdt_fid_lock(info->mti_env, info->mti_mdt->mdt_namespace,
1373                                   &lh->mlh_reg_lh, LCK_EX, policy, res,
1374                                   dlmflags, NULL);
1375         } else {
1376                 /* do nothing if both are local */
1377                 return 0;
1378         }
1379
1380         if (rc != ELDLM_OK)
1381                 return rc;
1382
1383         /*
1384          * TODO, currently we don't save this lock because there is no place to
1385          * hold this lock handle, but to avoid race we need to save this lock.
1386          */
1387         mdt_object_unlock(info, NULL, lh, 1);
1388
1389         return 0;
1390 }
1391
1392 /*
1393  * operation may takes locks of linkea, or directory stripes, group them in
1394  * different list.
1395  */
1396 struct mdt_sub_lock {
1397         struct mdt_object      *msl_obj;
1398         struct mdt_lock_handle  msl_lh;
1399         struct list_head        msl_linkage;
1400 };
1401
1402 static void mdt_unlock_list(struct mdt_thread_info *info,
1403                             struct list_head *list, int decref)
1404 {
1405         struct mdt_sub_lock *msl;
1406         struct mdt_sub_lock *tmp;
1407
1408         list_for_each_entry_safe(msl, tmp, list, msl_linkage) {
1409                 mdt_object_unlock_put(info, msl->msl_obj, &msl->msl_lh, decref);
1410                 list_del(&msl->msl_linkage);
1411                 OBD_FREE_PTR(msl);
1412         }
1413 }
1414
1415 static inline void mdt_migrate_object_unlock(struct mdt_thread_info *info,
1416                                              struct mdt_object *obj,
1417                                              struct mdt_lock_handle *lh,
1418                                              struct ldlm_enqueue_info *einfo,
1419                                              struct list_head *slave_locks,
1420                                              int decref)
1421 {
1422         if (mdt_object_remote(obj)) {
1423                 mdt_unlock_list(info, slave_locks, decref);
1424                 mdt_object_unlock(info, obj, lh, decref);
1425         } else {
1426                 mdt_reint_striped_unlock(info, obj, lh, einfo, decref);
1427         }
1428 }
1429
1430 /*
1431  * lock parents of links, and also check whether total locks don't exceed
1432  * RS_MAX_LOCKS.
1433  *
1434  * \retval      0 on success, and locks can be saved in ptlrpc_reply_stat
1435  * \retval      1 on success, but total lock count may exceed RS_MAX_LOCKS
1436  * \retval      -ev negative errno upon error
1437  */
1438 static int mdt_link_parents_lock(struct mdt_thread_info *info,
1439                                  struct mdt_object *pobj,
1440                                  const struct md_attr *ma,
1441                                  struct mdt_object *obj,
1442                                  struct mdt_lock_handle *lhp,
1443                                  struct ldlm_enqueue_info *peinfo,
1444                                  struct list_head *parent_slave_locks,
1445                                  struct list_head *link_locks)
1446 {
1447         struct mdt_device *mdt = info->mti_mdt;
1448         struct lu_buf *buf = &info->mti_big_buf;
1449         struct lu_name *lname = &info->mti_name;
1450         struct linkea_data ldata = { NULL };
1451         bool blocked = false;
1452         int local_lnkp_cnt = 0;
1453         int rc;
1454
1455         ENTRY;
1456
1457         if (S_ISDIR(lu_object_attr(&obj->mot_obj)))
1458                 RETURN(0);
1459
1460         buf = lu_buf_check_and_alloc(buf, MAX_LINKEA_SIZE);
1461         if (buf->lb_buf == NULL)
1462                 RETURN(-ENOMEM);
1463
1464         ldata.ld_buf = buf;
1465         rc = mdt_links_read(info, obj, &ldata);
1466         if (rc) {
1467                 if (rc == -ENOENT || rc == -ENODATA)
1468                         rc = 0;
1469                 RETURN(rc);
1470         }
1471
1472         for (linkea_first_entry(&ldata); ldata.ld_lee && !rc;
1473              linkea_next_entry(&ldata)) {
1474                 struct mdt_object *lnkp;
1475                 struct mdt_sub_lock *msl;
1476                 struct lu_fid fid;
1477                 __u64 ibits;
1478
1479                 linkea_entry_unpack(ldata.ld_lee, &ldata.ld_reclen, lname,
1480                                     &fid);
1481
1482                 /* check if it's also linked to parent */
1483                 if (lu_fid_eq(mdt_object_fid(pobj), &fid)) {
1484                         CDEBUG(D_INFO, "skip parent "DFID", reovke "DNAME"\n",
1485                                PFID(&fid), PNAME(lname));
1486                         /* in case link is remote object, revoke LOOKUP lock */
1487                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1488                         continue;
1489                 }
1490
1491                 lnkp = NULL;
1492
1493                 /* check if it's linked to a stripe of parent */
1494                 if (ma->ma_valid & MA_LMV) {
1495                         struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1496                         struct lu_fid *stripe_fid = &info->mti_tmp_fid1;
1497                         int j = 0;
1498
1499                         for (; j < le32_to_cpu(lmv->lmv_stripe_count); j++) {
1500                                 fid_le_to_cpu(stripe_fid,
1501                                               &lmv->lmv_stripe_fids[j]);
1502                                 if (lu_fid_eq(stripe_fid, &fid)) {
1503                                         CDEBUG(D_INFO, "skip stripe "DFID
1504                                                ", reovke "DNAME"\n",
1505                                                PFID(&fid), PNAME(lname));
1506                                         lnkp = mdt_object_find(info->mti_env,
1507                                                                mdt, &fid);
1508                                         if (IS_ERR(lnkp))
1509                                                 GOTO(out, rc = PTR_ERR(lnkp));
1510                                         break;
1511                                 }
1512                         }
1513
1514                         if (lnkp) {
1515                                 rc = mdt_revoke_remote_lookup_lock(info, lnkp,
1516                                                                    obj);
1517                                 mdt_object_put(info->mti_env, lnkp);
1518                                 continue;
1519                         }
1520                 }
1521
1522                 /* Check if it's already locked */
1523                 list_for_each_entry(msl, link_locks, msl_linkage) {
1524                         if (lu_fid_eq(mdt_object_fid(msl->msl_obj), &fid)) {
1525                                 CDEBUG(D_INFO,
1526                                        DFID" was locked, revoke "DNAME"\n",
1527                                        PFID(&fid), PNAME(lname));
1528                                 lnkp = msl->msl_obj;
1529                                 break;
1530                         }
1531                 }
1532
1533                 if (lnkp) {
1534                         rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1535                         continue;
1536                 }
1537
1538                 CDEBUG(D_INFO, "lock "DFID":"DNAME"\n",
1539                        PFID(&fid), PNAME(lname));
1540
1541                 lnkp = mdt_object_find(info->mti_env, mdt, &fid);
1542                 if (IS_ERR(lnkp)) {
1543                         CWARN("%s: cannot find obj "DFID": %ld\n",
1544                               mdt_obd_name(mdt), PFID(&fid), PTR_ERR(lnkp));
1545                         continue;
1546                 }
1547
1548                 if (!mdt_object_exists(lnkp)) {
1549                         CDEBUG(D_INFO, DFID" doesn't exist, skip "DNAME"\n",
1550                               PFID(&fid), PNAME(lname));
1551                         mdt_object_put(info->mti_env, lnkp);
1552                         continue;
1553                 }
1554
1555                 if (!mdt_object_remote(lnkp))
1556                         local_lnkp_cnt++;
1557
1558                 OBD_ALLOC_PTR(msl);
1559                 if (msl == NULL)
1560                         GOTO(out, rc = -ENOMEM);
1561
1562                 /*
1563                  * we can't follow parent-child lock order like other MD
1564                  * operations, use lock_try here to avoid deadlock, if the lock
1565                  * cannot be taken, drop all locks taken, revoke the blocked
1566                  * one, and continue processing the remaining entries, and in
1567                  * the end of the loop restart from beginning.
1568                  */
1569                 mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1570                 ibits = 0;
1571                 rc = mdt_object_lock_try(info, lnkp, &msl->msl_lh, &ibits,
1572                                          MDS_INODELOCK_UPDATE, true);
1573                 if (!(ibits & MDS_INODELOCK_UPDATE)) {
1574
1575                         CDEBUG(D_INFO, "busy lock on "DFID" "DNAME"\n",
1576                                PFID(&fid), PNAME(lname));
1577
1578                         mdt_unlock_list(info, link_locks, 1);
1579                         /* also unlock parent locks to avoid deadlock */
1580                         if (!blocked)
1581                                 mdt_migrate_object_unlock(info, pobj, lhp,
1582                                                           peinfo,
1583                                                           parent_slave_locks,
1584                                                           1);
1585
1586                         blocked = true;
1587
1588                         mdt_lock_pdo_init(&msl->msl_lh, LCK_PW, lname);
1589                         rc = mdt_object_lock(info, lnkp, &msl->msl_lh,
1590                                              MDS_INODELOCK_UPDATE);
1591                         if (rc) {
1592                                 mdt_object_put(info->mti_env, lnkp);
1593                                 OBD_FREE_PTR(msl);
1594                                 GOTO(out, rc);
1595                         }
1596
1597                         if (mdt_object_remote(lnkp)) {
1598                                 struct ldlm_lock *lock;
1599
1600                                 /*
1601                                  * for remote object, set lock cb_atomic,
1602                                  * so lock can be released in blocking_ast()
1603                                  * immediately, then the next lock_try will
1604                                  * have better chance of success.
1605                                  */
1606                                 lock = ldlm_handle2lock(
1607                                                 &msl->msl_lh.mlh_rreg_lh);
1608                                 LASSERT(lock != NULL);
1609                                 lock_res_and_lock(lock);
1610                                 ldlm_set_atomic_cb(lock);
1611                                 unlock_res_and_lock(lock);
1612                                 LDLM_LOCK_PUT(lock);
1613                         }
1614
1615                         mdt_object_unlock_put(info, lnkp, &msl->msl_lh, 1);
1616                         OBD_FREE_PTR(msl);
1617                         continue;
1618                 }
1619
1620                 INIT_LIST_HEAD(&msl->msl_linkage);
1621                 msl->msl_obj = lnkp;
1622                 list_add_tail(&msl->msl_linkage, link_locks);
1623
1624                 rc = mdt_revoke_remote_lookup_lock(info, lnkp, obj);
1625         }
1626
1627         if (blocked)
1628                 GOTO(out, rc = -EBUSY);
1629
1630         EXIT;
1631 out:
1632         if (rc)
1633                 mdt_unlock_list(info, link_locks, rc);
1634         else if (local_lnkp_cnt > RS_MAX_LOCKS - 6)
1635                 /*
1636                  * parent may have 3 local objects: master object and 2 stripes
1637                  * (if it's being migrated too); source may have 2 local
1638                  * objects: master and 1 stripe; target has 1 local object.
1639                  */
1640                 rc = 1;
1641         return rc;
1642 }
1643
1644 static int mdt_lock_remote_slaves(struct mdt_thread_info *info,
1645                                   struct mdt_object *obj,
1646                                   const struct md_attr *ma,
1647                                   struct list_head *slave_locks)
1648 {
1649         struct mdt_device *mdt = info->mti_mdt;
1650         const struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1651         struct lu_fid *fid = &info->mti_tmp_fid1;
1652         struct mdt_object *slave;
1653         struct mdt_sub_lock *msl;
1654         int i;
1655         int rc;
1656
1657         ENTRY;
1658
1659         LASSERT(mdt_object_remote(obj));
1660         LASSERT(ma->ma_valid & MA_LMV);
1661         LASSERT(lmv);
1662
1663         if (le32_to_cpu(lmv->lmv_magic) != LMV_MAGIC_V1)
1664                 RETURN(-EINVAL);
1665
1666         if (le32_to_cpu(lmv->lmv_stripe_count) < 1)
1667                 RETURN(0);
1668
1669         for (i = 0; i < le32_to_cpu(lmv->lmv_stripe_count); i++) {
1670                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[i]);
1671
1672                 if (!fid_is_sane(fid))
1673                         continue;
1674
1675                 slave = mdt_object_find(info->mti_env, mdt, fid);
1676                 if (IS_ERR(slave))
1677                         GOTO(out, rc = PTR_ERR(slave));
1678
1679                 OBD_ALLOC_PTR(msl);
1680                 if (!msl) {
1681                         mdt_object_put(info->mti_env, slave);
1682                         GOTO(out, rc = -ENOMEM);
1683                 }
1684
1685                 mdt_lock_reg_init(&msl->msl_lh, LCK_EX);
1686                 rc = mdt_reint_object_lock(info, slave, &msl->msl_lh,
1687                                            MDS_INODELOCK_UPDATE, true);
1688                 if (rc) {
1689                         OBD_FREE_PTR(msl);
1690                         mdt_object_put(info->mti_env, slave);
1691                         GOTO(out, rc);
1692                 }
1693
1694                 INIT_LIST_HEAD(&msl->msl_linkage);
1695                 msl->msl_obj = slave;
1696                 list_add_tail(&msl->msl_linkage, slave_locks);
1697         }
1698         EXIT;
1699
1700 out:
1701         if (rc)
1702                 mdt_unlock_list(info, slave_locks, rc);
1703         return rc;
1704 }
1705
1706 /* lock parent and its stripes */
1707 static int mdt_migrate_parent_lock(struct mdt_thread_info *info,
1708                                    struct mdt_object *obj,
1709                                    const struct md_attr *ma,
1710                                    struct mdt_lock_handle *lh,
1711                                    struct ldlm_enqueue_info *einfo,
1712                                    struct list_head *slave_locks)
1713 {
1714         int rc;
1715
1716         if (mdt_object_remote(obj)) {
1717                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1718                                             &lh->mlh_rreg_lh, LCK_PW,
1719                                             MDS_INODELOCK_UPDATE, false);
1720                 if (rc != ELDLM_OK)
1721                         return rc;
1722
1723                 /*
1724                  * if obj is remote and striped, lock its stripes explicitly
1725                  * because it's not striped in LOD layer on this MDT.
1726                  */
1727                 if (ma->ma_valid & MA_LMV) {
1728                         rc = mdt_lock_remote_slaves(info, obj, ma, slave_locks);
1729                         if (rc)
1730                                 mdt_object_unlock(info, obj, lh, rc);
1731                 }
1732         } else {
1733                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_UPDATE,
1734                                             einfo, true);
1735         }
1736
1737         return rc;
1738 }
1739
1740 /*
1741  * in migration, object may be remote, and we need take full lock of it and its
1742  * stripes if it's directory, besides, object may be a remote object on its
1743  * parent, revoke its LOOKUP lock on where its parent is located.
1744  */
1745 static int mdt_migrate_object_lock(struct mdt_thread_info *info,
1746                                    struct mdt_object *pobj,
1747                                    struct mdt_object *obj,
1748                                    struct mdt_lock_handle *lh,
1749                                    struct ldlm_enqueue_info *einfo,
1750                                    struct list_head *slave_locks)
1751 {
1752         int rc;
1753
1754         if (mdt_object_remote(obj)) {
1755                 rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1756                 if (rc)
1757                         return rc;
1758
1759                 rc = mdt_remote_object_lock(info, obj, mdt_object_fid(obj),
1760                                             &lh->mlh_rreg_lh, LCK_EX,
1761                                             MDS_INODELOCK_FULL, false);
1762                 if (rc != ELDLM_OK)
1763                         return rc;
1764
1765                 /*
1766                  * if obj is remote and striped, lock its stripes explicitly
1767                  * because it's not striped in LOD layer on this MDT.
1768                  */
1769                 if (S_ISDIR(lu_object_attr(&obj->mot_obj))) {
1770                         struct md_attr *ma = &info->mti_attr;
1771
1772                         ma->ma_lmv = info->mti_big_lmm;
1773                         ma->ma_lmv_size = info->mti_big_lmmsize;
1774                         ma->ma_valid = 0;
1775                         rc = mdt_stripe_get(info, obj, ma, XATTR_NAME_LMV);
1776                         if (rc) {
1777                                 mdt_object_unlock(info, obj, lh, rc);
1778                                 return rc;
1779                         }
1780
1781                         if (ma->ma_valid & MA_LMV) {
1782                                 rc = mdt_lock_remote_slaves(info, obj, ma,
1783                                                             slave_locks);
1784                                 if (rc)
1785                                         mdt_object_unlock(info, obj, lh, rc);
1786                         }
1787                 }
1788         } else {
1789                 if (mdt_object_remote(pobj)) {
1790                         rc = mdt_revoke_remote_lookup_lock(info, pobj, obj);
1791                         if (rc)
1792                                 return rc;
1793                 }
1794
1795                 rc = mdt_reint_striped_lock(info, obj, lh, MDS_INODELOCK_FULL,
1796                                             einfo, true);
1797         }
1798
1799         return rc;
1800 }
1801
1802 /*
1803  * lookup source by name, if parent is striped directory, we need to find the
1804  * corresponding stripe where source is located, and then lookup there.
1805  *
1806  * besides, if parent is migrating too, and file is already in target stripe,
1807  * this should be a redo of 'lfs migrate' on client side.
1808  */
1809 static int mdt_migrate_lookup(struct mdt_thread_info *info,
1810                               struct mdt_object *pobj,
1811                               const struct md_attr *ma,
1812                               const struct lu_name *lname,
1813                               struct mdt_object **spobj,
1814                               struct mdt_object **sobj)
1815 {
1816         const struct lu_env *env = info->mti_env;
1817         struct lu_fid *fid = &info->mti_tmp_fid1;
1818         struct mdt_object *stripe;
1819         int rc;
1820
1821         if (ma->ma_valid & MA_LMV) {
1822                 /* if parent is striped, lookup on corresponding stripe */
1823                 struct lmv_mds_md_v1 *lmv = &ma->ma_lmv->lmv_md_v1;
1824                 __u32 hash_type = le32_to_cpu(lmv->lmv_hash_type);
1825                 __u32 stripe_count = le32_to_cpu(lmv->lmv_stripe_count);
1826                 bool is_migrating = le32_to_cpu(lmv->lmv_hash_type) &
1827                                     LMV_HASH_FLAG_MIGRATION;
1828
1829                 if (is_migrating) {
1830                         hash_type = le32_to_cpu(lmv->lmv_migrate_hash);
1831                         stripe_count -= le32_to_cpu(lmv->lmv_migrate_offset);
1832                 }
1833
1834                 rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1835                                               lname->ln_name,
1836                                               lname->ln_namelen);
1837                 if (rc < 0)
1838                         return rc;
1839
1840                 if (le32_to_cpu(lmv->lmv_hash_type) & LMV_HASH_FLAG_MIGRATION)
1841                         rc += le32_to_cpu(lmv->lmv_migrate_offset);
1842
1843                 fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1844
1845                 stripe = mdt_object_find(env, info->mti_mdt, fid);
1846                 if (IS_ERR(stripe))
1847                         return PTR_ERR(stripe);
1848
1849                 fid_zero(fid);
1850                 rc = mdo_lookup(env, mdt_object_child(stripe), lname, fid,
1851                                 &info->mti_spec);
1852                 if (rc == -ENOENT && is_migrating) {
1853                         /*
1854                          * if parent is migrating, and lookup child failed on
1855                          * source stripe, lookup again on target stripe, if it
1856                          * exists, it means previous migration was interrupted,
1857                          * and current file was migrated already.
1858                          */
1859                         mdt_object_put(env, stripe);
1860
1861                         hash_type = le32_to_cpu(lmv->lmv_hash_type);
1862                         stripe_count = le32_to_cpu(lmv->lmv_migrate_offset);
1863
1864                         rc = lmv_name_to_stripe_index(hash_type, stripe_count,
1865                                                       lname->ln_name,
1866                                                       lname->ln_namelen);
1867                         if (rc < 0)
1868                                 return rc;
1869
1870                         fid_le_to_cpu(fid, &lmv->lmv_stripe_fids[rc]);
1871
1872                         stripe = mdt_object_find(env, info->mti_mdt, fid);
1873                         if (IS_ERR(stripe))
1874                                 return PTR_ERR(stripe);
1875
1876                         fid_zero(fid);
1877                         rc = mdo_lookup(env, mdt_object_child(stripe), lname,
1878                                         fid, &info->mti_spec);
1879                         mdt_object_put(env, stripe);
1880                         return rc ?: -EALREADY;
1881                 } else if (rc) {
1882                         mdt_object_put(env, stripe);
1883                         return rc;
1884                 }
1885         } else {
1886                 fid_zero(fid);
1887                 rc = mdo_lookup(env, mdt_object_child(pobj), lname, fid,
1888                                 &info->mti_spec);
1889                 if (rc)
1890                         return rc;
1891
1892                 stripe = pobj;
1893                 mdt_object_get(env, stripe);
1894         }
1895
1896         *spobj = stripe;
1897
1898         *sobj = mdt_object_find(env, info->mti_mdt, fid);
1899         if (IS_ERR(*sobj)) {
1900                 mdt_object_put(env, stripe);
1901                 rc = PTR_ERR(*sobj);
1902                 *spobj = NULL;
1903                 *sobj = NULL;
1904         }
1905
1906         return rc;
1907 }
1908
1909 /* end lease and close file for regular file */
1910 static int mdd_migrate_close(struct mdt_thread_info *info,
1911                              struct mdt_object *obj)
1912 {
1913         struct close_data *data;
1914         struct mdt_body *repbody;
1915         struct ldlm_lock *lease;
1916         int rc;
1917         int rc2;
1918
1919         rc = -EPROTO;
1920         if (!req_capsule_field_present(info->mti_pill, &RMF_MDT_EPOCH,
1921                                       RCL_CLIENT) ||
1922             !req_capsule_field_present(info->mti_pill, &RMF_CLOSE_DATA,
1923                                       RCL_CLIENT))
1924                 goto close;
1925
1926         data = req_capsule_client_get(info->mti_pill, &RMF_CLOSE_DATA);
1927         if (!data)
1928                 goto close;
1929
1930         rc = -ESTALE;
1931         lease = ldlm_handle2lock(&data->cd_handle);
1932         if (!lease)
1933                 goto close;
1934
1935         /* check if the lease was already canceled */
1936         lock_res_and_lock(lease);
1937         rc = ldlm_is_cancel(lease);
1938         unlock_res_and_lock(lease);
1939
1940         if (rc) {
1941                 rc = -EAGAIN;
1942                 LDLM_DEBUG(lease, DFID" lease broken",
1943                            PFID(mdt_object_fid(obj)));
1944         }
1945
1946         /*
1947          * cancel server side lease, client side counterpart should have been
1948          * cancelled, it's okay to cancel it now as we've held mot_open_sem.
1949          */
1950         ldlm_lock_cancel(lease);
1951         ldlm_reprocess_all(lease->l_resource, lease);
1952         LDLM_LOCK_PUT(lease);
1953
1954 close:
1955         rc2 = mdt_close_internal(info, mdt_info_req(info), NULL);
1956         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
1957         repbody->mbo_valid |= OBD_MD_CLOSE_INTENT_EXECED;
1958
1959         return rc ?: rc2;
1960 }
1961
1962 /*
1963  * migrate file in below steps:
1964  *  1. lock parent and its stripes
1965  *  2. lookup source by name
1966  *  3. lock parents of source links if source is not directory
1967  *  4. reject if source is in HSM
1968  *  5. take source open_sem and close file if source is regular file
1969  *  6. lock source and its stripes if it's directory
1970  *  7. lock target so subsequent change to it can trigger COS
1971  *  8. migrate file
1972  *  9. unlock above locks
1973  * 10. sync device if source has links
1974  */
1975 static int mdt_reint_migrate(struct mdt_thread_info *info,
1976                              struct mdt_lock_handle *unused)
1977 {
1978         const struct lu_env *env = info->mti_env;
1979         struct mdt_device *mdt = info->mti_mdt;
1980         struct ptlrpc_request *req = mdt_info_req(info);
1981         struct mdt_reint_record *rr = &info->mti_rr;
1982         struct lu_ucred *uc = mdt_ucred(info);
1983         struct md_attr *ma = &info->mti_attr;
1984         struct ldlm_enqueue_info *peinfo = &info->mti_einfo[0];
1985         struct ldlm_enqueue_info *seinfo = &info->mti_einfo[1];
1986         struct mdt_object *pobj;
1987         struct mdt_object *spobj = NULL;
1988         struct mdt_object *sobj = NULL;
1989         struct mdt_object *tobj;
1990         struct lustre_handle rename_lh = { 0 };
1991         struct mdt_lock_handle *lhp;
1992         struct mdt_lock_handle *lhs;
1993         struct mdt_lock_handle *lht;
1994         LIST_HEAD(parent_slave_locks);
1995         LIST_HEAD(child_slave_locks);
1996         LIST_HEAD(link_locks);
1997         int lock_retries = 5;
1998         bool open_sem_locked = false;
1999         bool do_sync = false;
2000         int rc;
2001         ENTRY;
2002
2003         CDEBUG(D_INODE, "migrate "DFID"/"DNAME" to "DFID"\n", PFID(rr->rr_fid1),
2004                PNAME(&rr->rr_name), PFID(rr->rr_fid2));
2005
2006         if (info->mti_dlm_req)
2007                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2008
2009         if (!fid_is_md_operative(rr->rr_fid1) ||
2010             !fid_is_md_operative(rr->rr_fid2))
2011                 RETURN(-EPERM);
2012
2013         /* don't allow migrate . or .. */
2014         if (lu_name_is_dot_or_dotdot(&rr->rr_name))
2015                 RETURN(-EBUSY);
2016
2017         if (!mdt->mdt_enable_remote_dir || !mdt->mdt_enable_dir_migration)
2018                 RETURN(-EPERM);
2019
2020         if (!md_capable(uc, CFS_CAP_SYS_ADMIN) &&
2021             uc->uc_gid != mdt->mdt_enable_remote_dir_gid &&
2022             mdt->mdt_enable_remote_dir_gid != -1)
2023                 RETURN(-EPERM);
2024
2025         /*
2026          * Note: do not enqueue rename lock for replay request, because
2027          * if other MDT holds rename lock, but being blocked to wait for
2028          * this MDT to finish its recovery, and the failover MDT can not
2029          * get rename lock, which will cause deadlock.
2030          */
2031         if (!req_is_replay(req)) {
2032                 rc = mdt_rename_lock(info, &rename_lh);
2033                 if (rc != 0) {
2034                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2035                                mdt_obd_name(info->mti_mdt), rc);
2036                         RETURN(rc);
2037                 }
2038         }
2039
2040         /* pobj is master object of parent */
2041         pobj = mdt_parent_find_check(info, rr->rr_fid1, 0);
2042         if (IS_ERR(pobj))
2043                 GOTO(unlock_rename, rc = PTR_ERR(pobj));
2044
2045         if (unlikely(!info->mti_big_lmm)) {
2046                 info->mti_big_lmmsize = lmv_mds_md_size(64, LMV_MAGIC);
2047                 OBD_ALLOC(info->mti_big_lmm, info->mti_big_lmmsize);
2048                 if (!info->mti_big_lmm)
2049                         GOTO(put_parent, rc = -ENOMEM);
2050         }
2051
2052         ma->ma_lmv = info->mti_big_lmm;
2053         ma->ma_lmv_size = info->mti_big_lmmsize;
2054         ma->ma_valid = 0;
2055         rc = mdt_stripe_get(info, pobj, ma, XATTR_NAME_LMV);
2056         if (rc)
2057                 GOTO(put_parent, rc);
2058
2059 lock_parent:
2060         /* lock parent object */
2061         lhp = &info->mti_lh[MDT_LH_PARENT];
2062         mdt_lock_reg_init(lhp, LCK_PW);
2063         rc = mdt_migrate_parent_lock(info, pobj, ma, lhp, peinfo,
2064                                      &parent_slave_locks);
2065         if (rc)
2066                 GOTO(put_parent, rc);
2067
2068         /*
2069          * spobj is the corresponding stripe against name if pobj is striped
2070          * directory, which is the real parent, and no need to lock, because
2071          * we've taken full lock of pobj.
2072          */
2073         rc = mdt_migrate_lookup(info, pobj, ma, &rr->rr_name, &spobj, &sobj);
2074         if (rc)
2075                 GOTO(unlock_parent, rc);
2076
2077         /* lock parents of source links, and revoke LOOKUP lock of links */
2078         rc = mdt_link_parents_lock(info, pobj, ma, sobj, lhp, peinfo,
2079                                    &parent_slave_locks, &link_locks);
2080         if (rc == -EBUSY && lock_retries-- > 0) {
2081                 mdt_object_put(env, sobj);
2082                 mdt_object_put(env, spobj);
2083                 goto lock_parent;
2084         }
2085
2086         if (rc < 0)
2087                 GOTO(put_source, rc);
2088
2089         /*
2090          * RS_MAX_LOCKS is the limit of number of locks that can be saved along
2091          * with one request, if total lock count exceeds this limit, we will
2092          * drop all locks after migration, and synchronous device in the end.
2093          */
2094         do_sync = rc;
2095
2096         /* TODO: DoM migration is not supported yet */
2097         if (S_ISREG(lu_object_attr(&sobj->mot_obj))) {
2098                 ma->ma_lmm = info->mti_big_lmm;
2099                 ma->ma_lmm_size = info->mti_big_lmmsize;
2100                 ma->ma_valid = 0;
2101                 rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LOV);
2102                 if (rc)
2103                         GOTO(put_source, rc);
2104
2105                 if (ma->ma_valid & MA_LOV &&
2106                     mdt_lmm_dom_entry(ma->ma_lmm) != LMM_NO_DOM)
2107                         GOTO(put_source, rc = -EOPNOTSUPP);
2108         }
2109
2110         /* if migration HSM is allowed */
2111         if (!mdt->mdt_opts.mo_migrate_hsm_allowed) {
2112                 ma->ma_need = MA_HSM;
2113                 ma->ma_valid = 0;
2114                 rc = mdt_attr_get_complex(info, sobj, ma);
2115                 if (rc)
2116                         GOTO(unlock_links, rc);
2117
2118                 if ((ma->ma_valid & MA_HSM) && ma->ma_hsm.mh_flags != 0)
2119                         GOTO(unlock_links, rc = -EOPNOTSUPP);
2120         }
2121
2122         /* end lease and close file for regular file */
2123         if (info->mti_spec.sp_migrate_close) {
2124                 /* try to hold open_sem so that nobody else can open the file */
2125                 if (!down_write_trylock(&sobj->mot_open_sem)) {
2126                         /* close anyway */
2127                         mdd_migrate_close(info, sobj);
2128                         GOTO(unlock_links, rc = -EBUSY);
2129                 } else {
2130                         open_sem_locked = true;
2131                         rc = mdd_migrate_close(info, sobj);
2132                         if (rc)
2133                                 GOTO(unlock_open_sem, rc);
2134                 }
2135         }
2136
2137         /* lock source */
2138         lhs = &info->mti_lh[MDT_LH_OLD];
2139         mdt_lock_reg_init(lhs, LCK_EX);
2140         rc = mdt_migrate_object_lock(info, spobj, sobj, lhs, seinfo,
2141                                      &child_slave_locks);
2142         if (rc)
2143                 GOTO(unlock_open_sem, rc);
2144
2145         /* lock target */
2146         tobj = mdt_object_find(env, mdt, rr->rr_fid2);
2147         if (IS_ERR(tobj))
2148                 GOTO(unlock_source, rc = PTR_ERR(tobj));
2149
2150         lht = &info->mti_lh[MDT_LH_NEW];
2151         mdt_lock_reg_init(lht, LCK_EX);
2152         rc = mdt_reint_object_lock(info, tobj, lht, MDS_INODELOCK_FULL, true);
2153         if (rc)
2154                 GOTO(put_target, rc);
2155
2156         /* Don't do lookup sanity check. We know name doesn't exist. */
2157         info->mti_spec.sp_cr_lookup = 0;
2158         info->mti_spec.sp_feat = &dt_directory_features;
2159
2160         rc = mdo_migrate(env, mdt_object_child(pobj),
2161                          mdt_object_child(sobj), &rr->rr_name,
2162                          mdt_object_child(tobj), &info->mti_spec, ma);
2163         EXIT;
2164
2165         mdt_object_unlock(info, tobj, lht, rc);
2166 put_target:
2167         mdt_object_put(env, tobj);
2168 unlock_source:
2169         mdt_migrate_object_unlock(info, sobj, lhs, seinfo,
2170                                   &child_slave_locks, rc);
2171 unlock_open_sem:
2172         if (open_sem_locked)
2173                 up_write(&sobj->mot_open_sem);
2174 unlock_links:
2175         /* if we've got too many locks to save into RPC,
2176          * then just commit before the locks are released */
2177         if (!rc && do_sync)
2178                 mdt_device_sync(env, mdt);
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         return rc;
2193 }
2194
2195 static int mdt_object_lock_save(struct mdt_thread_info *info,
2196                                 struct mdt_object *dir,
2197                                 struct mdt_lock_handle *lh,
2198                                 int idx, bool cos_incompat)
2199 {
2200         int rc;
2201
2202         /* we lock the target dir if it is local */
2203         rc = mdt_reint_object_lock(info, dir, lh, MDS_INODELOCK_UPDATE,
2204                                    cos_incompat);
2205         if (rc != 0)
2206                 return rc;
2207
2208         /* get and save correct version after locking */
2209         mdt_version_get_save(info, dir, idx);
2210         return 0;
2211 }
2212
2213 /*
2214  * determine lock order of sobj and tobj
2215  *
2216  * there are two situations we need to lock tobj before sobj:
2217  * 1. sobj is child of tobj
2218  * 2. sobj and tobj are stripes of a directory, and stripe index of sobj is
2219  *    larger than that of tobj
2220  *
2221  * \retval      1 lock tobj before sobj
2222  * \retval      0 lock sobj before tobj
2223  * \retval      -ev negative errno upon error
2224  */
2225 static int mdt_rename_determine_lock_order(struct mdt_thread_info *info,
2226                                            struct mdt_object *sobj,
2227                                            struct mdt_object *tobj)
2228 {
2229         struct md_attr *ma = &info->mti_attr;
2230         struct lu_fid *spfid = &info->mti_tmp_fid1;
2231         struct lu_fid *tpfid = &info->mti_tmp_fid2;
2232         struct lmv_mds_md_v1 *lmv;
2233         __u32 sindex;
2234         __u32 tindex;
2235         int rc;
2236
2237         /* sobj and tobj are the same */
2238         if (sobj == tobj)
2239                 return 0;
2240
2241         if (fid_is_root(mdt_object_fid(sobj)))
2242                 return 0;
2243
2244         if (fid_is_root(mdt_object_fid(tobj)))
2245                 return 1;
2246
2247         /* check whether sobj is child of tobj */
2248         rc = mdo_is_subdir(info->mti_env, mdt_object_child(sobj),
2249                            mdt_object_fid(tobj));
2250         if (rc < 0)
2251                 return rc;
2252
2253         if (rc == 1)
2254                 return 1;
2255
2256         /* check whether sobj and tobj are children of the same parent */
2257         rc = mdt_attr_get_pfid(info, sobj, spfid);
2258         if (rc)
2259                 return rc;
2260
2261         rc = mdt_attr_get_pfid(info, tobj, tpfid);
2262         if (rc)
2263                 return rc;
2264
2265         if (!lu_fid_eq(spfid, tpfid))
2266                 return 0;
2267
2268         /* check whether sobj and tobj are sibling stripes */
2269         ma->ma_need = MA_LMV;
2270         ma->ma_valid = 0;
2271         ma->ma_lmv = (union lmv_mds_md *)info->mti_xattr_buf;
2272         ma->ma_lmv_size = sizeof(info->mti_xattr_buf);
2273         rc = mdt_stripe_get(info, sobj, ma, XATTR_NAME_LMV);
2274         if (rc)
2275                 return rc;
2276
2277         if (!(ma->ma_valid & MA_LMV))
2278                 return 0;
2279
2280         lmv = &ma->ma_lmv->lmv_md_v1;
2281         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2282                 return 0;
2283         sindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2284
2285         ma->ma_valid = 0;
2286         rc = mdt_stripe_get(info, tobj, ma, XATTR_NAME_LMV);
2287         if (rc)
2288                 return rc;
2289
2290         if (!(ma->ma_valid & MA_LMV))
2291                 return -ENODATA;
2292
2293         lmv = &ma->ma_lmv->lmv_md_v1;
2294         if (!(le32_to_cpu(lmv->lmv_magic) & LMV_MAGIC_STRIPE))
2295                 return -EINVAL;
2296         tindex = le32_to_cpu(lmv->lmv_master_mdt_index);
2297
2298         /* check stripe index of sobj and tobj */
2299         if (sindex == tindex)
2300                 return -EINVAL;
2301
2302         return sindex < tindex ? 0 : 1;
2303 }
2304
2305 /*
2306  * VBR: rename versions in reply: 0 - srcdir parent; 1 - tgtdir parent;
2307  * 2 - srcdir child; 3 - tgtdir child.
2308  * Update on disk version of srcdir child.
2309  */
2310 static int mdt_reint_rename(struct mdt_thread_info *info,
2311                             struct mdt_lock_handle *unused)
2312 {
2313         struct mdt_device *mdt = info->mti_mdt;
2314         struct mdt_reint_record *rr = &info->mti_rr;
2315         struct md_attr *ma = &info->mti_attr;
2316         struct ptlrpc_request *req = mdt_info_req(info);
2317         struct mdt_object *msrcdir = NULL;
2318         struct mdt_object *mtgtdir = NULL;
2319         struct mdt_object *mold;
2320         struct mdt_object *mnew = NULL;
2321         struct lustre_handle rename_lh = { 0 };
2322         struct mdt_lock_handle *lh_srcdirp;
2323         struct mdt_lock_handle *lh_tgtdirp;
2324         struct mdt_lock_handle *lh_oldp = NULL;
2325         struct mdt_lock_handle *lh_newp = NULL;
2326         struct lu_fid *old_fid = &info->mti_tmp_fid1;
2327         struct lu_fid *new_fid = &info->mti_tmp_fid2;
2328         __u64 lock_ibits;
2329         bool reverse = false, discard = false;
2330         bool cos_incompat;
2331         int rc;
2332         ENTRY;
2333
2334         DEBUG_REQ(D_INODE, req, "rename "DFID"/"DNAME" to "DFID"/"DNAME,
2335                   PFID(rr->rr_fid1), PNAME(&rr->rr_name),
2336                   PFID(rr->rr_fid2), PNAME(&rr->rr_tgt_name));
2337
2338         if (info->mti_dlm_req)
2339                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2340
2341         if (!fid_is_md_operative(rr->rr_fid1) ||
2342             !fid_is_md_operative(rr->rr_fid2))
2343                 RETURN(-EPERM);
2344
2345         /* find both parents. */
2346         msrcdir = mdt_parent_find_check(info, rr->rr_fid1, 0);
2347         if (IS_ERR(msrcdir))
2348                 RETURN(PTR_ERR(msrcdir));
2349
2350         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME3, 5);
2351
2352         if (lu_fid_eq(rr->rr_fid1, rr->rr_fid2)) {
2353                 mtgtdir = msrcdir;
2354                 mdt_object_get(info->mti_env, mtgtdir);
2355         } else {
2356                 mtgtdir = mdt_parent_find_check(info, rr->rr_fid2, 1);
2357                 if (IS_ERR(mtgtdir))
2358                         GOTO(out_put_srcdir, rc = PTR_ERR(mtgtdir));
2359         }
2360
2361         /*
2362          * Note: do not enqueue rename lock for replay request, because
2363          * if other MDT holds rename lock, but being blocked to wait for
2364          * this MDT to finish its recovery, and the failover MDT can not
2365          * get rename lock, which will cause deadlock.
2366          */
2367         if (!req_is_replay(req)) {
2368                 /*
2369                  * Normally rename RPC is handled on the MDT with the target
2370                  * directory (if target exists, it's on the MDT with the
2371                  * target), if the source directory is remote, it's a hint that
2372                  * source is remote too (this may not be true, but it won't
2373                  * cause any issue), return -EXDEV early to avoid taking
2374                  * rename_lock.
2375                  */
2376                 if (!mdt->mdt_enable_remote_rename &&
2377                     mdt_object_remote(msrcdir))
2378                         GOTO(out_put_tgtdir, rc = -EXDEV);
2379
2380                 rc = mdt_rename_lock(info, &rename_lh);
2381                 if (rc != 0) {
2382                         CERROR("%s: can't lock FS for rename: rc = %d\n",
2383                                mdt_obd_name(mdt), rc);
2384                         GOTO(out_put_tgtdir, rc);
2385                 }
2386         }
2387
2388         rc = mdt_rename_determine_lock_order(info, msrcdir, mtgtdir);
2389         if (rc < 0)
2390                 GOTO(out_unlock_rename, rc);
2391
2392         reverse = rc;
2393
2394         /* source needs to be looked up after locking source parent, otherwise
2395          * this rename may race with unlink source, and cause rename hang, see
2396          * sanityn.sh 55b, so check parents first, if later we found source is
2397          * remote, relock parents. */
2398         cos_incompat = (mdt_object_remote(msrcdir) ||
2399                         mdt_object_remote(mtgtdir));
2400
2401         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2402
2403         /* lock parents in the proper order. */
2404         lh_srcdirp = &info->mti_lh[MDT_LH_PARENT];
2405         lh_tgtdirp = &info->mti_lh[MDT_LH_CHILD];
2406
2407 relock:
2408         mdt_lock_pdo_init(lh_srcdirp, LCK_PW, &rr->rr_name);
2409         mdt_lock_pdo_init(lh_tgtdirp, LCK_PW, &rr->rr_tgt_name);
2410
2411         if (reverse) {
2412                 rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2413                                           cos_incompat);
2414                 if (rc)
2415                         GOTO(out_unlock_rename, rc);
2416
2417                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2418
2419                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2420                                           cos_incompat);
2421                 if (rc != 0) {
2422                         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2423                         GOTO(out_unlock_rename, rc);
2424                 }
2425         } else {
2426                 rc = mdt_object_lock_save(info, msrcdir, lh_srcdirp, 0,
2427                                           cos_incompat);
2428                 if (rc)
2429                         GOTO(out_unlock_rename, rc);
2430
2431                 OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME, 5);
2432
2433                 if (mtgtdir != msrcdir) {
2434                         rc = mdt_object_lock_save(info, mtgtdir, lh_tgtdirp, 1,
2435                                                   cos_incompat);
2436                 } else if (!mdt_object_remote(mtgtdir) &&
2437                            lh_srcdirp->mlh_pdo_hash !=
2438                            lh_tgtdirp->mlh_pdo_hash) {
2439                         rc = mdt_pdir_hash_lock(info, lh_tgtdirp, mtgtdir,
2440                                                 MDS_INODELOCK_UPDATE,
2441                                                 cos_incompat);
2442                         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_PDO_LOCK2, 10);
2443                 }
2444                 if (rc != 0) {
2445                         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2446                         GOTO(out_unlock_rename, rc);
2447                 }
2448         }
2449
2450         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME4, 5);
2451         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_RENAME2, 5);
2452
2453         /* find mold object. */
2454         fid_zero(old_fid);
2455         rc = mdt_lookup_version_check(info, msrcdir, &rr->rr_name, old_fid, 2);
2456         if (rc != 0)
2457                 GOTO(out_unlock_parents, rc);
2458
2459         if (lu_fid_eq(old_fid, rr->rr_fid1) || lu_fid_eq(old_fid, rr->rr_fid2))
2460                 GOTO(out_unlock_parents, rc = -EINVAL);
2461
2462         if (!fid_is_md_operative(old_fid))
2463                 GOTO(out_unlock_parents, rc = -EPERM);
2464
2465         mold = mdt_object_find(info->mti_env, info->mti_mdt, old_fid);
2466         if (IS_ERR(mold))
2467                 GOTO(out_unlock_parents, rc = PTR_ERR(mold));
2468
2469         if (!mdt_object_exists(mold)) {
2470                 LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2471                                 &mold->mot_obj,
2472                                 "object does not exist");
2473                 GOTO(out_put_old, rc = -ENOENT);
2474         }
2475
2476         if (mdt_object_remote(mold) && !mdt->mdt_enable_remote_rename)
2477                 GOTO(out_put_old, rc = -EXDEV);
2478
2479         /* Check if @mtgtdir is subdir of @mold, before locking child
2480          * to avoid reverse locking. */
2481         if (mtgtdir != msrcdir) {
2482                 rc = mdo_is_subdir(info->mti_env, mdt_object_child(mtgtdir),
2483                                    old_fid);
2484                 if (rc) {
2485                         if (rc == 1)
2486                                 rc = -EINVAL;
2487                         GOTO(out_put_old, rc);
2488                 }
2489         }
2490
2491         tgt_vbr_obj_set(info->mti_env, mdt_obj2dt(mold));
2492         /* save version after locking */
2493         mdt_version_get_save(info, mold, 2);
2494
2495         if (!cos_incompat && mdt_object_remote(mold)) {
2496                 cos_incompat = true;
2497                 mdt_object_put(info->mti_env, mold);
2498                 mdt_object_unlock(info, mtgtdir, lh_tgtdirp, -EAGAIN);
2499                 mdt_object_unlock(info, msrcdir, lh_srcdirp, -EAGAIN);
2500                 goto relock;
2501         }
2502
2503         /* find mnew object:
2504          * mnew target object may not exist now
2505          * lookup with version checking */
2506         fid_zero(new_fid);
2507         rc = mdt_lookup_version_check(info, mtgtdir, &rr->rr_tgt_name, new_fid,
2508                                       3);
2509         if (rc == 0) {
2510                 /* the new_fid should have been filled at this moment */
2511                 if (lu_fid_eq(old_fid, new_fid))
2512                         GOTO(out_put_old, rc);
2513
2514                 if (lu_fid_eq(new_fid, rr->rr_fid1) ||
2515                     lu_fid_eq(new_fid, rr->rr_fid2))
2516                         GOTO(out_put_old, rc = -EINVAL);
2517
2518                 if (!fid_is_md_operative(new_fid))
2519                         GOTO(out_put_old, rc = -EPERM);
2520
2521                 mnew = mdt_object_find(info->mti_env, info->mti_mdt, new_fid);
2522                 if (IS_ERR(mnew))
2523                         GOTO(out_put_old, rc = PTR_ERR(mnew));
2524
2525                 if (!mdt_object_exists(mnew)) {
2526                         LU_OBJECT_DEBUG(D_INODE, info->mti_env,
2527                                         &mnew->mot_obj,
2528                                         "object does not exist");
2529                         GOTO(out_put_new, rc = -ENOENT);
2530                 }
2531
2532                 if (mdt_object_remote(mnew)) {
2533                         struct mdt_body  *repbody;
2534
2535                         /* Always send rename req to the target child MDT */
2536                         repbody = req_capsule_server_get(info->mti_pill,
2537                                                          &RMF_MDT_BODY);
2538                         LASSERT(repbody != NULL);
2539                         repbody->mbo_fid1 = *new_fid;
2540                         repbody->mbo_valid |= (OBD_MD_FLID | OBD_MD_MDS);
2541                         GOTO(out_put_new, rc = -EXDEV);
2542                 }
2543                 /* Before locking the target dir, check we do not replace
2544                  * a dir with a non-dir, otherwise it may deadlock with
2545                  * link op which tries to create a link in this dir
2546                  * back to this non-dir. */
2547                 if (S_ISDIR(lu_object_attr(&mnew->mot_obj)) &&
2548                     !S_ISDIR(lu_object_attr(&mold->mot_obj)))
2549                         GOTO(out_put_new, rc = -EISDIR);
2550
2551                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2552                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2553                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2554                 if (mdt_object_remote(msrcdir)) {
2555                         /* Enqueue lookup lock from the parent MDT */
2556                         rc = mdt_remote_object_lock(info, msrcdir,
2557                                                     mdt_object_fid(mold),
2558                                                     &lh_oldp->mlh_rreg_lh,
2559                                                     lh_oldp->mlh_rreg_mode,
2560                                                     MDS_INODELOCK_LOOKUP,
2561                                                     false);
2562                         if (rc != ELDLM_OK)
2563                                 GOTO(out_put_new, rc);
2564
2565                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2566                 }
2567
2568                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2569                                            cos_incompat);
2570                 if (rc != 0)
2571                         GOTO(out_unlock_old, rc);
2572
2573                 /* Check if @msrcdir is subdir of @mnew, before locking child
2574                  * to avoid reverse locking. */
2575                 if (mtgtdir != msrcdir) {
2576                         rc = mdo_is_subdir(info->mti_env,
2577                                            mdt_object_child(msrcdir), new_fid);
2578                         if (rc) {
2579                                 if (rc == 1)
2580                                         rc = -EINVAL;
2581                                 GOTO(out_unlock_old, rc);
2582                         }
2583                 }
2584
2585                 /* We used to acquire MDS_INODELOCK_FULL here but we
2586                  * can't do this now because a running HSM restore on
2587                  * the rename onto victim will hold the layout
2588                  * lock. See LU-4002. */
2589
2590                 lh_newp = &info->mti_lh[MDT_LH_NEW];
2591                 mdt_lock_reg_init(lh_newp, LCK_EX);
2592                 rc = mdt_reint_object_lock(info, mnew, lh_newp,
2593                                            MDS_INODELOCK_LOOKUP |
2594                                            MDS_INODELOCK_UPDATE,
2595                                            cos_incompat);
2596                 if (rc != 0)
2597                         GOTO(out_unlock_old, rc);
2598
2599                 /* get and save version after locking */
2600                 mdt_version_get_save(info, mnew, 3);
2601         } else if (rc != -EREMOTE && rc != -ENOENT) {
2602                 GOTO(out_put_old, rc);
2603         } else {
2604                 lh_oldp = &info->mti_lh[MDT_LH_OLD];
2605                 mdt_lock_reg_init(lh_oldp, LCK_EX);
2606                 lock_ibits = MDS_INODELOCK_LOOKUP | MDS_INODELOCK_XATTR;
2607                 if (mdt_object_remote(msrcdir)) {
2608                         /* Enqueue lookup lock from the parent MDT */
2609                         rc = mdt_remote_object_lock(info, msrcdir,
2610                                                     mdt_object_fid(mold),
2611                                                     &lh_oldp->mlh_rreg_lh,
2612                                                     lh_oldp->mlh_rreg_mode,
2613                                                     MDS_INODELOCK_LOOKUP,
2614                                                     false);
2615                         if (rc != ELDLM_OK)
2616                                 GOTO(out_put_old, rc);
2617
2618                         lock_ibits &= ~MDS_INODELOCK_LOOKUP;
2619                 }
2620
2621                 rc = mdt_reint_object_lock(info, mold, lh_oldp, lock_ibits,
2622                                            cos_incompat);
2623                 if (rc != 0)
2624                         GOTO(out_unlock_old, rc);
2625
2626                 mdt_enoent_version_save(info, 3);
2627         }
2628
2629         /* step 5: rename it */
2630         mdt_reint_init_ma(info, ma);
2631
2632         mdt_fail_write(info->mti_env, info->mti_mdt->mdt_bottom,
2633                        OBD_FAIL_MDS_REINT_RENAME_WRITE);
2634
2635         if (mnew != NULL)
2636                 mutex_lock(&mnew->mot_lov_mutex);
2637
2638         rc = mdo_rename(info->mti_env, mdt_object_child(msrcdir),
2639                         mdt_object_child(mtgtdir), old_fid, &rr->rr_name,
2640                         mnew != NULL ? mdt_object_child(mnew) : NULL,
2641                         &rr->rr_tgt_name, ma);
2642
2643         if (mnew != NULL)
2644                 mutex_unlock(&mnew->mot_lov_mutex);
2645
2646         /* handle last link of tgt object */
2647         if (rc == 0) {
2648                 mdt_counter_incr(req, LPROC_MDT_RENAME);
2649                 if (mnew) {
2650                         mdt_handle_last_unlink(info, mnew, ma);
2651                         discard = mdt_dom_check_for_discard(info, mnew);
2652                 }
2653                 mdt_rename_counter_tally(info, info->mti_mdt, req,
2654                                          msrcdir, mtgtdir);
2655         }
2656
2657         EXIT;
2658         if (mnew != NULL)
2659                 mdt_object_unlock(info, mnew, lh_newp, rc);
2660 out_unlock_old:
2661         mdt_object_unlock(info, mold, lh_oldp, rc);
2662 out_put_new:
2663         if (mnew && !discard)
2664                 mdt_object_put(info->mti_env, mnew);
2665 out_put_old:
2666         mdt_object_put(info->mti_env, mold);
2667 out_unlock_parents:
2668         mdt_object_unlock(info, mtgtdir, lh_tgtdirp, rc);
2669         mdt_object_unlock(info, msrcdir, lh_srcdirp, rc);
2670 out_unlock_rename:
2671         if (lustre_handle_is_used(&rename_lh))
2672                 mdt_rename_unlock(&rename_lh);
2673 out_put_tgtdir:
2674         mdt_object_put(info->mti_env, mtgtdir);
2675 out_put_srcdir:
2676         mdt_object_put(info->mti_env, msrcdir);
2677
2678         /* The DoM discard can be done right in the place above where it is
2679          * assigned, meanwhile it is done here after rename unlock due to
2680          * compatibility with old clients, for them the discard blocks
2681          * the main thread until completion. Check LU-11359 for details.
2682          */
2683         if (discard) {
2684                 mdt_dom_discard_data(info, mnew);
2685                 mdt_object_put(info->mti_env, mnew);
2686         }
2687         return rc;
2688 }
2689
2690 static int mdt_reint_resync(struct mdt_thread_info *info,
2691                             struct mdt_lock_handle *lhc)
2692 {
2693         struct mdt_reint_record *rr = &info->mti_rr;
2694         struct ptlrpc_request *req = mdt_info_req(info);
2695         struct md_attr *ma = &info->mti_attr;
2696         struct mdt_object *mo;
2697         struct ldlm_lock *lease;
2698         struct mdt_body *repbody;
2699         struct md_layout_change layout = { .mlc_mirror_id = rr->rr_mirror_id };
2700         bool lease_broken;
2701         int rc, rc2;
2702
2703         ENTRY;
2704
2705         DEBUG_REQ(D_INODE, req, DFID", FLR file resync", PFID(rr->rr_fid1));
2706
2707         if (info->mti_dlm_req)
2708                 ldlm_request_cancel(req, info->mti_dlm_req, 0, LATF_SKIP);
2709
2710         mo = mdt_object_find(info->mti_env, info->mti_mdt, rr->rr_fid1);
2711         if (IS_ERR(mo))
2712                 GOTO(out, rc = PTR_ERR(mo));
2713
2714         if (!mdt_object_exists(mo))
2715                 GOTO(out_obj, rc = -ENOENT);
2716
2717         if (!S_ISREG(lu_object_attr(&mo->mot_obj)))
2718                 GOTO(out_obj, rc = -EINVAL);
2719
2720         if (mdt_object_remote(mo))
2721                 GOTO(out_obj, rc = -EREMOTE);
2722
2723         lease = ldlm_handle2lock(rr->rr_lease_handle);
2724         if (lease == NULL)
2725                 GOTO(out_obj, rc = -ESTALE);
2726
2727         /* It's really necessary to grab open_sem and check if the lease lock
2728          * has been lost. There would exist a concurrent writer coming in and
2729          * generating some dirty data in memory cache, the writeback would fail
2730          * after the layout version is increased by MDS_REINT_RESYNC RPC. */
2731         if (!down_write_trylock(&mo->mot_open_sem))
2732                 GOTO(out_put_lease, rc = -EBUSY);
2733
2734         lock_res_and_lock(lease);
2735         lease_broken = ldlm_is_cancel(lease);
2736         unlock_res_and_lock(lease);
2737         if (lease_broken)
2738                 GOTO(out_unlock, rc = -EBUSY);
2739
2740         /* the file has yet opened by anyone else after we took the lease. */
2741         layout.mlc_opc = MD_LAYOUT_RESYNC;
2742         lhc = &info->mti_lh[MDT_LH_LOCAL];
2743         rc = mdt_layout_change(info, mo, lhc, &layout);
2744         if (rc)
2745                 GOTO(out_unlock, rc);
2746
2747         mdt_object_unlock(info, mo, lhc, 0);
2748
2749         ma->ma_need = MA_INODE;
2750         ma->ma_valid = 0;
2751         rc = mdt_attr_get_complex(info, mo, ma);
2752         if (rc != 0)
2753                 GOTO(out_unlock, rc);
2754
2755         repbody = req_capsule_server_get(info->mti_pill, &RMF_MDT_BODY);
2756         mdt_pack_attr2body(info, repbody, &ma->ma_attr, mdt_object_fid(mo));
2757
2758         EXIT;
2759 out_unlock:
2760         up_write(&mo->mot_open_sem);
2761 out_put_lease:
2762         LDLM_LOCK_PUT(lease);
2763 out_obj:
2764         mdt_object_put(info->mti_env, mo);
2765 out:
2766         mdt_client_compatibility(info);
2767         rc2 = mdt_fix_reply(info);
2768         if (rc == 0)
2769                 rc = rc2;
2770         return rc;
2771 }
2772
2773 struct mdt_reinter {
2774         int (*mr_handler)(struct mdt_thread_info *, struct mdt_lock_handle *);
2775         enum lprocfs_extra_opc mr_extra_opc;
2776 };
2777
2778 static const struct mdt_reinter mdt_reinters[] = {
2779         [REINT_SETATTR] = {
2780                 .mr_handler = &mdt_reint_setattr,
2781                 .mr_extra_opc = MDS_REINT_SETATTR,
2782         },
2783         [REINT_CREATE] = {
2784                 .mr_handler = &mdt_reint_create,
2785                 .mr_extra_opc = MDS_REINT_CREATE,
2786         },
2787         [REINT_LINK] = {
2788                 .mr_handler = &mdt_reint_link,
2789                 .mr_extra_opc = MDS_REINT_LINK,
2790         },
2791         [REINT_UNLINK] = {
2792                 .mr_handler = &mdt_reint_unlink,
2793                 .mr_extra_opc = MDS_REINT_UNLINK,
2794         },
2795         [REINT_RENAME] = {
2796                 .mr_handler = &mdt_reint_rename,
2797                 .mr_extra_opc = MDS_REINT_RENAME,
2798         },
2799         [REINT_OPEN] = {
2800                 .mr_handler = &mdt_reint_open,
2801                 .mr_extra_opc = MDS_REINT_OPEN,
2802         },
2803         [REINT_SETXATTR] = {
2804                 .mr_handler = &mdt_reint_setxattr,
2805                 .mr_extra_opc = MDS_REINT_SETXATTR,
2806         },
2807         [REINT_RMENTRY] = {
2808                 .mr_handler = &mdt_reint_unlink,
2809                 .mr_extra_opc = MDS_REINT_UNLINK,
2810         },
2811         [REINT_MIGRATE] = {
2812                 .mr_handler = &mdt_reint_migrate,
2813                 .mr_extra_opc = MDS_REINT_RENAME,
2814         },
2815         [REINT_RESYNC] = {
2816                 .mr_handler = &mdt_reint_resync,
2817                 .mr_extra_opc = MDS_REINT_RESYNC,
2818         },
2819 };
2820
2821 int mdt_reint_rec(struct mdt_thread_info *info,
2822                   struct mdt_lock_handle *lhc)
2823 {
2824         const struct mdt_reinter *mr;
2825         int rc;
2826         ENTRY;
2827
2828         if (!(info->mti_rr.rr_opcode < ARRAY_SIZE(mdt_reinters)))
2829                 RETURN(-EPROTO);
2830
2831         mr = &mdt_reinters[info->mti_rr.rr_opcode];
2832         if (mr->mr_handler == NULL)
2833                 RETURN(-EPROTO);
2834
2835         rc = (*mr->mr_handler)(info, lhc);
2836
2837         lprocfs_counter_incr(ptlrpc_req2svc(mdt_info_req(info))->srv_stats,
2838                              PTLRPC_LAST_CNTR + mr->mr_extra_opc);
2839
2840         RETURN(rc);
2841 }