lustre/lfsck/lfsck_striped_dir.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,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License version 2 for more details.  A copy is
  14  * included in the COPYING file that accompanied this code.
  15
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2014, 2015, Intel Corporation.
  24  */
  25 /*
  26  * lustre/lfsck/lfsck_striped_dir.c
  27  *
  28  * Author: Fan, Yong <fan.yong@intel.com>
  29  */
  30
  31 /*
  32  * About the verification for striped directory. Some rules and assumptions:
  33  *
  34  * 1) lmv_magic: The magic may be wrong. But it is almost impossible (1/2^32
  35  *    probability) that a master LMV EA claims as a slave LMV EA by wrong,
  36  *    so we can ignore such race case and the reverse case.
  37  *
  38  * 2) lmv_master_mdt_index: The master index can be self-verified by compared
  39  *    with the MDT index directly. The slave stripe index can be verified by
  40  *    compared with the file name. Although both the name entry and the LMV EA
  41  *    can be wrong, it is almost impossible that they hit the same bad data
  42  *    So if they match each other, then trust them. Similarly, for the shard,
  43  *    it stores index in both slave LMV EA and in linkEA, if the two copies
  44  *    match, then trust them.
  45  *
  46  * 3) lmv_hash_type: The valid hash type should be LMV_HASH_TYPE_ALL_CHARS or
  47  *    LMV_HASH_TYPE_FNV_1A_64. If the LFSCK instance on some slave finds that
  48  *    the name hash against the hash function does not match the MDT, then it
  49  *    will change the master LMV EA hash type as LMV_HASH_TYPE_UNKNOWN. With
  50  *    such hash type, the whole striped directory still can be accessed via
  51  *    lookup/readdir, and also support unlink, but cannot add new name entry.
  52  *
  53  * 3.1) If the master hash type is one of the valid values, then trust the
  54  *      master LMV EA. Because:
  55  *
  56  * 3.1.1) The master hash type is visible to the client and used by the client.
  57  *
  58  * 3.1.2) For a given name, different hash types may map the name entry to the
  59  *        same MDT. So simply checking one name entry or some name entries may
  60  *        cannot verify whether the hash type is correct or not.
  61  *
  62  * 3.1.3) Different shards can claim different hash types, it is not easy to
  63  *        distinguish which ones are correct. Even though the master is wrong,
  64  *        as the LFSCK processing, some LFSCK instance on other MDT may finds
  65  *        unmatched name hash, then it will change the master hash type to
  66  *        LMV_HASH_TYPE_UNKNOWN as described above. The worst case is euqal
  67  *        to the case without the LFSCK.
  68  *
  69  * 3.2) If the master hash type is invalid, nor LMV_HASH_TYPE_UNKNOWN, then
  70  *      trust the first shard with valid hash type (ALL_CHARS or FNV_1A_64).
  71  *      If the shard is also worng, means there are double failures, then as
  72  *      the LFSCK processing, other LFSCK instances on the other MDTs may
  73  *      find unmatched name hash, and then, the master hash type will be
  74  *      changed to LMV_HASH_TYPE_UNKNOWN as described in the 3).
  75  *
  76  * 3.3) If the master hash type is LMV_HASH_TYPE_UNKNOWN, then it is possible
  77  *      that some other LFSCK instance on other MDT found bad name hash, then
  78  *      changed the master hash type to LMV_HASH_TYPE_UNKNOWN as described in
  79  *      the 3). But it also maybe because of data corruption in master LMV EA.
  80  *      To make such two cases to be distinguishable, when the LFSCK changes
  81  *      the master hash type to LMV_HASH_TYPE_UNKNOWN, it will mark in the
  82  *      master LMV EA (new lmv flags LMV_HASH_FLAG_BAD_TYPE). Then subsequent
  83  *      LFSCK checking can distinguish them: for former case, turst the master
  84  *      LMV EA with nothing to be done; otherwise, trust the first shard with
  85  *      valid hash type (ALL_CHARS or FNV_1A_64) as the 3.2) does.
  86  *
  87  * 4) lmv_stripe_count: For a shard of a striped directory, if its index has
  88  *    been verified as the 2), then the stripe count must be larger than its
  89  *    index. For the master object, by scanning each shard's index, the LFSCK
  90  *    can know the highest index, and the stripe count must be larger than the
  91  *    known highest index. If the stipe count in the LMV EA matches above two
  92  *    rules, then it is may be trustable. If both the master claimed stripe
  93  *    count and the slave claimed stripe count match each own rule, but they
  94  *    are not the same, then trust the master. Because the stripe count in
  95  *    the master LMV EA is visible to client and used to distribute the name
  96  *    entry to some shard, but the slave LMV EA is only used for verification
  97  *    and invisible to client.
  98  *
  99  * 5) If the master LMV EA is lost, then there are two possible cases:
 100  *
 101  * 5.1) The slave claims slave LMV EA by wrong, means that the parent was not
 102  *      a striped directory, but its sub-directory has a wrong slave LMV EA.
 103  *      It is very very race case, similar as the 1), can be ignored.
 104  *
 105  * 5.2) The parent directory is a striped directory, but the master LMV EA
 106  *      is lost or crashed. Then the LFSCK needs to re-generate the master
 107  *      LMV EA: the lmv_master_mdt_index is from the MDT device index; the
 108  *      lmv_hash_type is from the first valid shard; the lmv_stripe_count
 109  *      will be calculated via scanning all the shards.
 110  *
 111  * 5.2.1) Before re-generating the master LMV EA, the LFSCK needs to check
 112  *        whether someone has created some file(s) under the master object
 113  *        after the master LMV EA disappear. If yes, the LFSCK will cannot
 114  *        re-generate the master LMV EA, otherwise, such new created files
 115  *        will be invisible to client. Under such case, the LFSCK will mark
 116  *        the master object as read only (without master LMV EA). Then all
 117  *        things under the master MDT-object, including those new created
 118  *        files and the shards themselves, will be visibile to client. And
 119  *        then the administrator can handle the bad striped directory with
 120  *        more human knowledge.
 121  *
 122  * 5.2.2) If someone created some special sub-directory under the master
 123  *        MDT-object with the same naming rule as shard name $FID:$index,
 124  *        as to the LFSCK cannot detect it before re-generating the master
 125  *        LMV EA, then such sub-directory itself will be invisible after
 126  *        the LFSCK re-generating the master LMV EA. The sub-items under
 127  *        such sub-directory are still visible to client. As the LFSCK
 128  *        processing, if such sub-directory cause some conflict with other
 129  *        normal shard, such as the index conflict, then the LFSCK will
 130  *        remove the master LMV EA and change the master MDT-object to
 131  *        read-only mode as the 5.2.1). But if there is no conflict, the
 132  *        LFSCK will regard such sub-directory as a striped shard that
 133  *        lost its slave LMV EA, and will re-generate slave LMV EA for it.
 134  *
 135  * 5.2.3) Anytime, if the LFSCK found some shards name/index conflict,
 136  *        and cannot make the distinguish which one is right, then it
 137  *        will remove the master LMV EA and change the MDT-object to
 138  *        read-only mode as the 5.2.2).
 139  */
 140
 141 #define DEBUG_SUBSYSTEM S_LFSCK
 142
 143 #include <lu_object.h>
 144 #include <dt_object.h>
 145 #include <md_object.h>
 146 #include <lustre_fid.h>
 147 #include <lustre_lib.h>
 148 #include <lustre_net.h>
 149 #include <lustre_lmv.h>
 150
 151 #include "lfsck_internal.h"
 152
 153 void lfsck_lmv_put(const struct lu_env *env, struct lfsck_lmv *llmv)
 154 {
 155         if (llmv != NULL && atomic_dec_and_test(&llmv->ll_ref)) {
 156                 if (llmv->ll_inline) {
 157                         struct lfsck_lmv_unit   *llu;
 158                         struct lfsck_instance   *lfsck;
 159
 160                         llu = list_entry(llmv, struct lfsck_lmv_unit, llu_lmv);
 161                         lfsck = llu->llu_lfsck;
 162
 163                         spin_lock(&lfsck->li_lock);
 164                         list_del(&llu->llu_link);
 165                         spin_unlock(&lfsck->li_lock);
 166
 167                         lfsck_object_put(env, llu->llu_obj);
 168
 169                         LASSERT(llmv->ll_lslr != NULL);
 170
 171                         OBD_FREE_LARGE(llmv->ll_lslr,
 172                                        sizeof(*llmv->ll_lslr) *
 173                                        llmv->ll_stripes_allocated);
 174                         OBD_FREE_PTR(llu);
 175                 } else {
 176                         if (llmv->ll_lslr != NULL)
 177                                 OBD_FREE_LARGE(llmv->ll_lslr,
 178                                         sizeof(*llmv->ll_lslr) *
 179                                         llmv->ll_stripes_allocated);
 180
 181                         OBD_FREE_PTR(llmv);
 182                 }
 183         }
 184 }
 185
 186 /**
 187  * Mark the specified directory as read-only by set LUSTRE_IMMUTABLE_FL.
 188  *
 189  * The caller has taken the ldlm lock on the @obj already.
 190  *
 191  * \param[in] env       pointer to the thread context
 192  * \param[in] com       pointer to the lfsck component
 193  * \param[in] obj       pointer to the object to be handled
 194  * \param[in] del_lmv   true if need to drop the LMV EA
 195  *
 196  * \retval              positive number if nothing to be done
 197  * \retval              zero for success
 198  * \retval              negative error number on failure
 199  */
 200 static int lfsck_disable_master_lmv(const struct lu_env *env,
 201                                     struct lfsck_component *com,
 202                                     struct dt_object *obj, bool del_lmv)
 203 {
 204         struct lfsck_thread_info        *info   = lfsck_env_info(env);
 205         struct lu_attr                  *la     = &info->lti_la;
 206         struct lfsck_instance           *lfsck  = com->lc_lfsck;
 207         struct dt_device                *dev    = lfsck_obj2dev(obj);
 208         struct thandle                  *th     = NULL;
 209         int                              rc     = 0;
 210         ENTRY;
 211
 212         th = dt_trans_create(env, dev);
 213         if (IS_ERR(th))
 214                 GOTO(log, rc = PTR_ERR(th));
 215
 216         if (del_lmv) {
 217                 rc = dt_declare_xattr_del(env, obj, XATTR_NAME_LMV, th);
 218                 if (rc != 0)
 219                         GOTO(stop, rc);
 220         }
 221
 222         la->la_valid = LA_FLAGS;
 223         rc = dt_declare_attr_set(env, obj, la, th);
 224         if (rc != 0)
 225                 GOTO(stop, rc);
 226
 227         rc = dt_trans_start_local(env, dev, th);
 228         if (rc != 0)
 229                 GOTO(stop, rc);
 230
 231         dt_write_lock(env, obj, 0);
 232         if (unlikely(lfsck_is_dead_obj(obj)))
 233                 GOTO(unlock, rc = 1);
 234
 235         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
 236                 GOTO(unlock, rc = 0);
 237
 238         if (del_lmv) {
 239                 rc = dt_xattr_del(env, obj, XATTR_NAME_LMV, th);
 240                 if (rc != 0)
 241                         GOTO(unlock, rc);
 242         }
 243
 244         rc = dt_attr_get(env, obj, la);
 245         if (rc == 0 && !(la->la_flags & LUSTRE_IMMUTABLE_FL)) {
 246                 la->la_valid = LA_FLAGS;
 247                 la->la_flags |= LUSTRE_IMMUTABLE_FL;
 248                 rc = dt_attr_set(env, obj, la, th);
 249         }
 250
 251         GOTO(unlock, rc);
 252
 253 unlock:
 254         dt_write_unlock(env, obj);
 255
 256 stop:
 257         dt_trans_stop(env, dev, th);
 258
 259 log:
 260         CDEBUG(D_LFSCK, "%s: namespace LFSCK set the master MDT-object of "
 261                "the striped directory "DFID" as read-only: rc = %d\n",
 262                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(obj)), rc);
 263
 264         if (rc <= 0) {
 265                 struct lfsck_namespace *ns = com->lc_file_ram;
 266
 267                 ns->ln_flags |= LF_INCONSISTENT;
 268                 if (rc == 0)
 269                         ns->ln_striped_dirs_disabled++;
 270         }
 271
 272         return rc;
 273 }
 274
 275 static inline bool lfsck_is_valid_slave_lmv(struct lmv_mds_md_v1 *lmv)
 276 {
 277         return lmv->lmv_stripe_count >= 1 &&
 278                lmv->lmv_stripe_count <= LFSCK_LMV_MAX_STRIPES &&
 279                lmv->lmv_stripe_count > lmv->lmv_master_mdt_index &&
 280                lmv_is_known_hash_type(lmv->lmv_hash_type);
 281 }
 282
 283 /**
 284  * Remove the striped directory's master LMV EA and mark it as read-only.
 285  *
 286  * Take ldlm lock on the striped directory before calling the
 287  * lfsck_disable_master_lmv().
 288  *
 289  * \param[in] env       pointer to the thread context
 290  * \param[in] com       pointer to the lfsck component
 291  * \param[in] obj       pointer to the striped directory to be handled
 292  * \param[in] lnr       pointer to the namespace request that contains the
 293  *                      striped directory to be handled and other information
 294  *
 295  * \retval              positive number if nothing to be done
 296  * \retval              zero for success
 297  * \retval              negative error number on failure
 298  */
 299 static int lfsck_remove_lmv(const struct lu_env *env,
 300                             struct lfsck_component *com,
 301                             struct dt_object *obj,
 302                             struct lfsck_namespace_req *lnr)
 303 {
 304         struct lustre_handle     lh     = { 0 };
 305         int                      rc;
 306
 307         lnr->lnr_lmv->ll_ignore = 1;
 308         rc = lfsck_ibits_lock(env, com->lc_lfsck, obj, &lh,
 309                               MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
 310                               LCK_EX);
 311         if (rc == 0) {
 312                 rc = lfsck_disable_master_lmv(env, com, obj, true);
 313                 lfsck_ibits_unlock(&lh, LCK_EX);
 314         }
 315
 316         return rc;
 317 }
 318
 319 /**
 320  * Remove the name entry from the striped directory's master MDT-object.
 321  *
 322  * \param[in] env       pointer to the thread context
 323  * \param[in] com       pointer to the lfsck component
 324  * \param[in] dir       pointer to the striped directory
 325  * \param[in] fid       the shard's FID which name entry will be removed
 326  * \param[in] index     the shard's index which name entry will be removed
 327  *
 328  * \retval              positive number for repaired successfully
 329  * \retval              0 if nothing to be repaired
 330  * \retval              negative error number on failure
 331  */
 332 static int lfsck_remove_dirent(const struct lu_env *env,
 333                                struct lfsck_component *com,
 334                                struct dt_object *dir,
 335                                const struct lu_fid *fid, __u32 index)
 336 {
 337         struct lfsck_thread_info        *info = lfsck_env_info(env);
 338         struct dt_object                *obj;
 339         int                              rc;
 340
 341         snprintf(info->lti_tmpbuf2, sizeof(info->lti_tmpbuf2), DFID":%u",
 342                  PFID(fid), index);
 343         obj = lfsck_object_find_bottom(env, com->lc_lfsck, fid);
 344         if (IS_ERR(obj))
 345                 return PTR_ERR(obj);
 346
 347         rc = lfsck_namespace_repair_dirent(env, com, dir, obj,
 348                                         info->lti_tmpbuf2, info->lti_tmpbuf2,
 349                                         S_IFDIR, false, false);
 350         lfsck_object_put(env, obj);
 351         if (rc > 0) {
 352                 struct lfsck_namespace *ns = com->lc_file_ram;
 353
 354                 ns->ln_dirent_repaired++;
 355         }
 356
 357         return rc;
 358 }
 359
 360 /**
 361  * Remove old shard's name entry and refill the @lslr slot with new shard.
 362  *
 363  * Some old shard held the specified @lslr slot, but it is an invalid shard.
 364  * This function will remove the bad shard's name entry, and refill the @lslr
 365  * slot with the new shard.
 366  *
 367  * \param[in] env       pointer to the thread context
 368  * \param[in] com       pointer to the lfsck component
 369  * \param[in] dir       pointer to the striped directory to be handled
 370  * \param[in] lslr      pointer to lfsck_disable_master_lmv slot which content
 371  *                      will be replaced by the given information
 372  * \param[in] lnr       contain the shard's FID to be used to fill the
 373  *                      @lslr slot, it also records the known max filled index
 374  *                      and the known max stripe count
 375  * \param[in] lmv       contain the slave LMV EA to be used to fill the
 376  *                      @lslr slot
 377  * \param[in] index     the old shard's index in the striped directory
 378  * \param[in] flags     the new shard's flags in the @lslr slot
 379  *
 380  * \retval              zero for success
 381  * \retval              negative error number on failure
 382  */
 383 static int lfsck_replace_lmv(const struct lu_env *env,
 384                              struct lfsck_component *com,
 385                              struct dt_object *dir,
 386                              struct lfsck_slave_lmv_rec *lslr,
 387                              struct lfsck_namespace_req *lnr,
 388                              struct lmv_mds_md_v1 *lmv,
 389                              __u32 index, __u32 flags)
 390 {
 391         struct lfsck_lmv *llmv = lnr->lnr_lmv;
 392         int               rc;
 393
 394         rc = lfsck_remove_dirent(env, com, dir,
 395                                  &lslr->lslr_fid, index);
 396         if (rc < 0)
 397                 return rc;
 398
 399         lslr->lslr_fid = lnr->lnr_fid;
 400         lslr->lslr_flags = flags;
 401         lslr->lslr_stripe_count = lmv->lmv_stripe_count;
 402         lslr->lslr_index = lmv->lmv_master_mdt_index;
 403         lslr->lslr_hash_type = lmv->lmv_hash_type;
 404         if (flags == LSLF_NONE) {
 405                 if (llmv->ll_hash_type == LMV_HASH_TYPE_UNKNOWN &&
 406                     lmv_is_known_hash_type(lmv->lmv_hash_type))
 407                         llmv->ll_hash_type = lmv->lmv_hash_type;
 408
 409                 if (lslr->lslr_stripe_count <= LFSCK_LMV_MAX_STRIPES &&
 410                     llmv->ll_max_stripe_count < lslr->lslr_stripe_count)
 411                         llmv->ll_max_stripe_count = lslr->lslr_stripe_count;
 412         }
 413
 414         return 0;
 415 }
 416
 417 /**
 418  * Record the slave LMV EA in the lfsck_lmv::ll_lslr.
 419  *
 420  * If the lfsck_lmv::ll_lslr slot corresponding to the given @shard_idx is free,
 421  * then fill the slot with the given @lnr/@lmv/@flags directly (maybe need to
 422  * extend the lfsck_lmv::ll_lslr buffer).
 423  *
 424  * If the lfsck_lmv::ll_lslr slot corresponding to the given @shard_idx is taken
 425  * by other shard, then the LFSCK will try to resolve the conflict by checking
 426  * the two conflict shards' flags, and try other possible slot (if one of them
 427  * claims another possible @shard_idx).
 428  *
 429  * 1) If one of the two conflict shards can be recorded in another slot, then
 430  *    it is OK, go ahead. Otherwise,
 431  *
 432  * 2) If one of them is dangling name entry, then remove (one of) the dangling
 433  *    name entry (and replace related @lslr slot if needed). Otherwise,
 434  *
 435  * 3) If one of them has no slave LMV EA, then check whether the master LMV
 436  *    EA has ever been lost and re-generated (LMV_HASH_FLAG_LOST_LMV in the
 437  *    master LMV EA).
 438  *
 439  * 3.1) If yes, then it is possible that such object is not a real shard of
 440  *      the striped directory, instead, it was created by someone after the
 441  *      master LMV EA lost with the name that matches the shard naming rule.
 442  *      Then the LFSCK will remove the master LMV EA and mark the striped
 443  *      directory as read-only to allow those non-shard files to be visible
 444  *      to client.
 445  *
 446  * 3.2) If no, then remove (one of) the object what has no slave LMV EA.
 447  *
 448  * 4) If all above efforts cannot work, then the LFSCK cannot know how to
 449  *    recover the striped directory. To make the administrator can see the
 450  *    conflicts, the LFSCK will remove the master LMV EA and mark the striped
 451  *    directory as read-only.
 452  *
 453  * This function may be called recursively, to prevent overflow, we define
 454  * LFSCK_REC_LMV_MAX_DEPTH to restrict the recursive call depth.
 455  *
 456  * \param[in] env       pointer to the thread context
 457  * \param[in] com       pointer to the lfsck component
 458  * \param[in] dir       pointer to the striped directory to be handled
 459  * \param[in] lnr       contain the shard's FID to fill the @lslr slot,
 460  *                      it also records the known max filled index and
 461  *                      the known max stripe count
 462  * \param[in] lmv       pointer to the slave LMV EA to be recorded
 463  * \param[in] shard_idx the shard's index used for locating the @lslr slot,
 464  *                      it can be the index stored in the shard's name,
 465  *                      it also can be the index stored in the slave LMV EA
 466  *                      (for recursive case)
 467  * \param[in] flags     the shard's flags to be recorded in the @lslr slot
 468  *                      to indicate the shard status, such as whether has
 469  *                      slave LMV EA, whether dangling name entry, whether
 470  *                      the name entry and slave LMV EA unmatched, and ect
 471  * \param[in] flags2    when be called recursively, the @flags2 tells the
 472  *                      former conflict shard's flags in the @lslr slot.
 473  * \param[in,out] depth To prevent to be called recurisively too deep,
 474  *                      we define the max depth can be called recursively
 475  *                      (LFSCK_REC_LMV_MAX_DEPTH)
 476  *
 477  * \retval              zero for success
 478  * \retval              "-ERANGE" for invalid @shard_idx
 479  * \retval              "-EEXIST" for the required lslr slot has been
 480  *                      occupied by other shard
 481  * \retval              other negative error number on failure
 482  */
 483 static int lfsck_record_lmv(const struct lu_env *env,
 484                             struct lfsck_component *com,
 485                             struct dt_object *dir,
 486                             struct lfsck_namespace_req *lnr,
 487                             struct lmv_mds_md_v1 *lmv, __u32 shard_idx,
 488                             __u32 flags, __u32 flags2, __u32 *depth)
 489 {
 490         struct lfsck_instance      *lfsck = com->lc_lfsck;
 491         struct lfsck_lmv           *llmv  = lnr->lnr_lmv;
 492         const struct lu_fid        *fid   = &lnr->lnr_fid;
 493         struct lfsck_slave_lmv_rec *lslr;
 494         struct lfsck_rec_lmv_save  *lrls;
 495         int                         index = shard_idx;
 496         int                         rc    = 0;
 497         ENTRY;
 498
 499         CDEBUG(D_LFSCK, "%s: record slave LMV EA for the striped directory "
 500                DFID": shard = "DFID", index = %u, flags = %u, flags2 = %u, "
 501                "depth = %d\n", lfsck_lfsck2name(lfsck),
 502                PFID(lfsck_dto2fid(dir)), PFID(fid),
 503                index, flags, flags2, *depth);
 504
 505         if (index < 0 || index >= LFSCK_LMV_MAX_STRIPES)
 506                 RETURN(-ERANGE);
 507
 508         if (index >= llmv->ll_stripes_allocated) {
 509                 struct lfsck_slave_lmv_rec *new_lslr;
 510                 int new_stripes = index + 1;
 511                 size_t old_size = sizeof(*lslr) * llmv->ll_stripes_allocated;
 512
 513                 OBD_ALLOC_LARGE(new_lslr, sizeof(*new_lslr) * new_stripes);
 514                 if (new_lslr == NULL) {
 515                         llmv->ll_failed = 1;
 516
 517                         RETURN(-ENOMEM);
 518                 }
 519
 520                 memcpy(new_lslr, llmv->ll_lslr, old_size);
 521                 OBD_FREE_LARGE(llmv->ll_lslr, old_size);
 522                 llmv->ll_stripes_allocated = new_stripes;
 523                 llmv->ll_lslr = new_lslr;
 524         }
 525
 526         lslr = llmv->ll_lslr + index;
 527         if (unlikely(lu_fid_eq(&lslr->lslr_fid, fid)))
 528                 RETURN(0);
 529
 530         if (fid_is_zero(&lslr->lslr_fid)) {
 531                 lslr->lslr_fid = *fid;
 532                 lslr->lslr_stripe_count = lmv->lmv_stripe_count;
 533                 lslr->lslr_index = lmv->lmv_master_mdt_index;
 534                 lslr->lslr_hash_type = lmv->lmv_hash_type;
 535                 lslr->lslr_flags = flags;
 536                 llmv->ll_stripes_filled++;
 537                 if (flags == LSLF_NONE) {
 538                         if (llmv->ll_hash_type == LMV_HASH_TYPE_UNKNOWN &&
 539                             lmv_is_known_hash_type(lmv->lmv_hash_type))
 540                                 llmv->ll_hash_type = lmv->lmv_hash_type;
 541
 542                         if (lslr->lslr_stripe_count <= LFSCK_LMV_MAX_STRIPES &&
 543                             llmv->ll_max_stripe_count < lslr->lslr_stripe_count)
 544                                 llmv->ll_max_stripe_count =
 545                                                         lslr->lslr_stripe_count;
 546                 }
 547
 548                 if (llmv->ll_max_filled_off < index)
 549                         llmv->ll_max_filled_off = index;
 550
 551                 RETURN(0);
 552         }
 553
 554         (*depth)++;
 555         if (flags != LSLF_BAD_INDEX2)
 556                 LASSERTF(*depth == 1, "depth = %d\n", *depth);
 557
 558         /* Handle conflict cases. */
 559         switch (lslr->lslr_flags) {
 560         case LSLF_NONE:
 561         case LSLF_BAD_INDEX2:
 562                 /* The existing one is a normal valid object. */
 563                 switch (flags) {
 564                 case LSLF_NONE:
 565                         /* The two 'valid' name entries claims the same
 566                          * index, the LFSCK cannot distinguish which one
 567                          * is correct. Then remove the master LMV EA to
 568                          * make all shards to be visible to client, and
 569                          * mark the master MDT-object as read-only. The
 570                          * administrator can handle the conflict with
 571                          * more human knowledge. */
 572                         rc = lfsck_remove_lmv(env, com, dir, lnr);
 573                         break;
 574                 case LSLF_BAD_INDEX2:
 575                         GOTO(out, rc = -EEXIST);
 576                 case LSLF_NO_LMVEA:
 577
 578 no_lmvea:
 579                         if (llmv->ll_lmv.lmv_hash_type &
 580                             LMV_HASH_FLAG_LOST_LMV) {
 581                                 /* If the master LMV EA was re-generated
 582                                  * by the former LFSCK reparation, and
 583                                  * before such reparation, someone has
 584                                  * created the conflict object, but the
 585                                  * LFSCK did not detect such conflict,
 586                                  * then we have to remove the master
 587                                  * LMV EA and mark the master MDT-object
 588                                  * as read-only. The administrator can
 589                                  * handle the conflict with more human
 590                                  * knowledge. */
 591                                 rc = lfsck_remove_lmv(env, com, dir, lnr);
 592                         } else {
 593                                 /* Otherwise, remove the current name entry,
 594                                  * and add its FID in the LFSCK tracing file
 595                                  * for further processing. */
 596                                 rc = lfsck_namespace_trace_update(env, com, fid,
 597                                                 LNTF_CHECK_PARENT, true);
 598                                 if (rc == 0)
 599                                         rc = lfsck_remove_dirent(env, com, dir,
 600                                                                  fid, index);
 601                         }
 602
 603                         break;
 604                 case LSLF_DANGLING:
 605                         /* Remove the current dangling name entry. */
 606                         rc = lfsck_remove_dirent(env, com, dir, fid, index);
 607                         break;
 608                 case LSLF_BAD_INDEX1:
 609                         index = lmv->lmv_master_mdt_index;
 610                         lmv->lmv_master_mdt_index = shard_idx;
 611                         /* The name entry claims an index that is conflict
 612                          * with a valid existing name entry, then try the
 613                          * index in the lmv recursively. */
 614                         rc = lfsck_record_lmv(env, com, dir, lnr, lmv, index,
 615                                 LSLF_BAD_INDEX2, lslr->lslr_flags, depth);
 616                         lmv->lmv_master_mdt_index = index;
 617                         if (rc == -ERANGE || rc == -EEXIST)
 618                                 /* The index in the lmv is invalid or
 619                                  * also conflict with other. Then we do
 620                                  * not know how to resolve the conflict.
 621                                  * We will handle it as handle the case
 622                                  * of 'LSLF_NONE' vs 'LSLF_NONE'. */
 623                                 rc = lfsck_remove_lmv(env, com, dir, lnr);
 624
 625                         break;
 626                 default:
 627                         break;
 628                 }
 629
 630                 break;
 631         case LSLF_NO_LMVEA:
 632                 /* The existing one has no slave LMV EA. */
 633                 switch (flags) {
 634                 case LSLF_NONE:
 635
 636 none:
 637                         if (llmv->ll_lmv.lmv_hash_type &
 638                             LMV_HASH_FLAG_LOST_LMV) {
 639                                 /* If the master LMV EA was re-generated
 640                                  * by the former LFSCK reparation, and
 641                                  * before such reparation, someone has
 642                                  * created the conflict object, but the
 643                                  * LFSCK did not detect such conflict,
 644                                  * then we have to remove the master
 645                                  * LMV EA and mark the master MDT-object
 646                                  * as read-only. The administrator can
 647                                  * handle the conflict with more human
 648                                  * knowledge. */
 649                                 rc = lfsck_remove_lmv(env, com, dir, lnr);
 650                         } else {
 651                                 lrls = &lfsck->li_rec_lmv_save[*depth - 1];
 652                                 lrls->lrls_fid = lslr->lslr_fid;
 653                                 /* Otherwise, remove the existing name entry,
 654                                  * and add its FID in the LFSCK tracing file
 655                                  * for further processing. Refill the slot
 656                                  * with current slave LMV EA. */
 657                                 rc = lfsck_namespace_trace_update(env,
 658                                                 com, &lrls->lrls_fid,
 659                                                 LNTF_CHECK_PARENT, true);
 660                                 if (rc == 0)
 661                                         rc = lfsck_replace_lmv(env, com, dir,
 662                                                 lslr, lnr, lmv, index, flags);
 663                         }
 664
 665                         break;
 666                 case LSLF_BAD_INDEX2:
 667                         if (flags2 >= lslr->lslr_flags)
 668                                 GOTO(out, rc = -EEXIST);
 669
 670                         goto none;
 671                 case LSLF_NO_LMVEA:
 672                         goto no_lmvea;
 673                 case LSLF_DANGLING:
 674                         /* Remove the current dangling name entry. */
 675                         rc = lfsck_remove_dirent(env, com, dir, fid, index);
 676                         break;
 677                 case LSLF_BAD_INDEX1:
 678                         index = lmv->lmv_master_mdt_index;
 679                         lmv->lmv_master_mdt_index = shard_idx;
 680                         /* The name entry claims an index that is conflict
 681                          * with a valid existing name entry, then try the
 682                          * index in the lmv recursively. */
 683                         rc = lfsck_record_lmv(env, com, dir, lnr, lmv, index,
 684                                 LSLF_BAD_INDEX2, lslr->lslr_flags, depth);
 685                         lmv->lmv_master_mdt_index = index;
 686                         if (rc == -ERANGE || rc == -EEXIST) {
 687                                 index = shard_idx;
 688                                 goto no_lmvea;
 689                         }
 690
 691                         break;
 692                 default:
 693                         break;
 694                 }
 695
 696                 break;
 697         case LSLF_DANGLING:
 698                 /* The existing one is a dangling name entry. */
 699                 switch (flags) {
 700                 case LSLF_NONE:
 701                 case LSLF_BAD_INDEX2:
 702                 case LSLF_NO_LMVEA:
 703                         /* Remove the existing dangling name entry.
 704                          * Refill the lslr slot with the given LMV. */
 705                         rc = lfsck_replace_lmv(env, com, dir, lslr, lnr,
 706                                                lmv, index, flags);
 707                         break;
 708                 case LSLF_DANGLING:
 709                         /* Two dangling name entries conflict,
 710                          * remove the current one. */
 711                         rc = lfsck_remove_dirent(env, com, dir, fid, index);
 712                         break;
 713                 case LSLF_BAD_INDEX1:
 714                         index = lmv->lmv_master_mdt_index;
 715                         lmv->lmv_master_mdt_index = shard_idx;
 716                         /* The name entry claims an index that is conflict
 717                          * with a valid existing name entry, then try the
 718                          * index in the lmv recursively. */
 719                         rc = lfsck_record_lmv(env, com, dir, lnr, lmv, index,
 720                                 LSLF_BAD_INDEX2, lslr->lslr_flags, depth);
 721                         lmv->lmv_master_mdt_index = index;
 722                         if (rc == -ERANGE || rc == -EEXIST)
 723                                 /* If the index in the lmv is invalid or
 724                                  * also conflict with other, then remove
 725                                  * the existing dangling name entry.
 726                                  * Refill the lslr slot with the given LMV. */
 727                                 rc = lfsck_replace_lmv(env, com, dir, lslr, lnr,
 728                                                        lmv, shard_idx, flags);
 729
 730                         break;
 731                 default:
 732                         break;
 733                 }
 734
 735                 break;
 736         case LSLF_BAD_INDEX1: {
 737                 if (*depth >= LFSCK_REC_LMV_MAX_DEPTH)
 738                         goto conflict;
 739
 740                 lrls = &lfsck->li_rec_lmv_save[*depth - 1];
 741                 lrls->lrls_fid = lnr->lnr_fid;
 742                 lrls->lrls_lmv = *lmv;
 743
 744                 lnr->lnr_fid = lslr->lslr_fid;
 745                 lmv->lmv_master_mdt_index = index;
 746                 lmv->lmv_stripe_count = lslr->lslr_stripe_count;
 747                 lmv->lmv_hash_type = lslr->lslr_hash_type;
 748                 index = lslr->lslr_index;
 749
 750                 /* The existing one has another possible slot,
 751                  * try it recursively. */
 752                 rc = lfsck_record_lmv(env, com, dir, lnr, lmv, index,
 753                                       LSLF_BAD_INDEX2, flags, depth);
 754                 *lmv = lrls->lrls_lmv;
 755                 lnr->lnr_fid = lrls->lrls_fid;
 756                 index = shard_idx;
 757                 if (rc != 0) {
 758                         if (rc == -ERANGE || rc == -EEXIST)
 759                                 goto conflict;
 760
 761                         break;
 762                 }
 763
 764                 lslr->lslr_fid = *fid;
 765                 lslr->lslr_flags = flags;
 766                 lslr->lslr_stripe_count = lmv->lmv_stripe_count;
 767                 lslr->lslr_index = lmv->lmv_master_mdt_index;
 768                 lslr->lslr_hash_type = lmv->lmv_hash_type;
 769                 if (flags == LSLF_NONE) {
 770                         if (llmv->ll_hash_type == LMV_HASH_TYPE_UNKNOWN &&
 771                             lmv_is_known_hash_type(lmv->lmv_hash_type))
 772                                 llmv->ll_hash_type = lmv->lmv_hash_type;
 773
 774                         if (lslr->lslr_stripe_count <= LFSCK_LMV_MAX_STRIPES &&
 775                             llmv->ll_max_stripe_count < lslr->lslr_stripe_count)
 776                                 llmv->ll_max_stripe_count =
 777                                                         lslr->lslr_stripe_count;
 778                 }
 779
 780                 break;
 781
 782 conflict:
 783                 switch (flags) {
 784                 case LSLF_NONE:
 785                         /* The two 'valid' name entries claims the same
 786                          * index, the LFSCK cannot distinguish which one
 787                          * is correct. Then remove the master LMV EA to
 788                          * make all shards to be visible to client, and
 789                          * mark the master MDT-object as read-only. The
 790                          * administrator can handle the conflict with
 791                          * more human knowledge. */
 792                         rc = lfsck_remove_lmv(env, com, dir, lnr);
 793                         break;
 794                 case LSLF_BAD_INDEX2:
 795                         GOTO(out, rc = -EEXIST);
 796                 case LSLF_NO_LMVEA:
 797                         goto no_lmvea;
 798                 case LSLF_DANGLING:
 799                         /* Remove the current dangling name entry. */
 800                         rc = lfsck_remove_dirent(env, com, dir, fid, index);
 801                         break;
 802                 case LSLF_BAD_INDEX1:
 803                         index = lmv->lmv_master_mdt_index;
 804                         lmv->lmv_master_mdt_index = shard_idx;
 805                         /* The name entry claims an index that is conflict
 806                          * with a valid existing name entry, then try the
 807                          * index in the lmv recursively. */
 808                         rc = lfsck_record_lmv(env, com, dir, lnr, lmv, index,
 809                                 LSLF_BAD_INDEX2, lslr->lslr_flags, depth);
 810                         lmv->lmv_master_mdt_index = index;
 811                         if (rc == -ERANGE || rc == -EEXIST)
 812                                 /* The index in the lmv is invalid or
 813                                  * also conflict with other. Then we do
 814                                  * not know how to resolve the conflict.
 815                                  * We will handle it as handle the case
 816                                  * of 'LSLF_NONE' vs 'LSLF_NONE'. */
 817                                 rc = lfsck_remove_lmv(env, com, dir, lnr);
 818
 819                         break;
 820                 }
 821
 822                 break;
 823         }
 824         default:
 825                 break;
 826         }
 827
 828         if (rc < 0)
 829                 llmv->ll_failed = 1;
 830
 831         GOTO(out, rc);
 832
 833 out:
 834         (*depth)--;
 835
 836         return rc > 0 ? 0 : rc;
 837 }
 838
 839 int lfsck_read_stripe_lmv(const struct lu_env *env, struct dt_object *obj,
 840                           struct lmv_mds_md_v1 *lmv)
 841 {
 842         int rc;
 843
 844         dt_read_lock(env, obj, 0);
 845         rc = dt_xattr_get(env, obj, lfsck_buf_get(env, lmv, sizeof(*lmv)),
 846                           XATTR_NAME_LMV);
 847         dt_read_unlock(env, obj);
 848         if (rc != sizeof(*lmv))
 849                 return rc > 0 ? -EINVAL : rc;
 850
 851         lfsck_lmv_header_le_to_cpu(lmv, lmv);
 852         if ((lmv->lmv_magic == LMV_MAGIC &&
 853              !(lmv->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) ||
 854             (lmv->lmv_magic == LMV_MAGIC_STRIPE &&
 855              !(lmv->lmv_hash_type & LMV_HASH_FLAG_DEAD)))
 856                 return 0;
 857
 858         return -ENODATA;
 859 }
 860
 861 /**
 862  * Parse the shard's index from the given shard name.
 863  *
 864  * The valid shard name/type should be:
 865  * 1) The type must be S_IFDIR
 866  * 2) The name should be $FID:$index
 867  * 3) the index should within valid range.
 868  *
 869  * \param[in] env       pointer to the thread context
 870  * \param[in] name      the shard name
 871  * \param[in] namelen   the name length
 872  * \param[in] type      the entry's type
 873  * \param[in] fid       the entry's FID
 874  *
 875  * \retval              zero or positive number for the index from the name
 876  * \retval              negative error number on failure
 877  */
 878 int lfsck_shard_name_to_index(const struct lu_env *env, const char *name,
 879                               int namelen, __u16 type, const struct lu_fid *fid)
 880 {
 881         char    *name2  = lfsck_env_info(env)->lti_tmpbuf2;
 882         int      len;
 883         int      idx    = 0;
 884
 885         if (!S_ISDIR(type))
 886                 return -ENOTDIR;
 887
 888         LASSERT(name != name2);
 889
 890         len = snprintf(name2, sizeof(lfsck_env_info(env)->lti_tmpbuf2),
 891                        DFID":", PFID(fid));
 892         if (namelen < len + 1 || memcmp(name, name2, len) != 0)
 893                 return -EINVAL;
 894
 895         do {
 896                 if (!isdigit(name[len]))
 897                         return -EINVAL;
 898
 899                 idx = idx * 10 + name[len++] - '0';
 900         } while (len < namelen);
 901
 902         if (idx >= LFSCK_LMV_MAX_STRIPES)
 903                 return -EINVAL;
 904
 905         return idx;
 906 }
 907
 908 bool lfsck_is_valid_slave_name_entry(const struct lu_env *env,
 909                                      struct lfsck_lmv *llmv,
 910                                      const char *name, int namelen)
 911 {
 912         struct lmv_mds_md_v1    *lmv;
 913         int                      idx;
 914
 915         if (llmv == NULL || !llmv->ll_lmv_slave || !llmv->ll_lmv_verified)
 916                 return true;
 917
 918         lmv = &llmv->ll_lmv;
 919         idx = lmv_name_to_stripe_index(lmv->lmv_hash_type,
 920                                        lmv->lmv_stripe_count,
 921                                        name, namelen);
 922         if (unlikely(idx != lmv->lmv_master_mdt_index))
 923                 return false;
 924
 925         return true;
 926 }
 927
 928 /**
 929  * Check whether the given name is a valid entry under the @parent.
 930  *
 931  * If the @parent is a striped directory then the @child should one
 932  * shard of the striped directory, its name should be $FID:$index.
 933  *
 934  * If the @parent is a shard of a striped directory, then the name hash
 935  * should match the MDT, otherwise it is invalid.
 936  *
 937  * \param[in] env       pointer to the thread context
 938  * \param[in] parent    the parent directory
 939  * \param[in] child     the child object to be checked
 940  * \param[in] cname     the name for the @child in the parent directory
 941  *
 942  * \retval              positive number for invalid name entry
 943  * \retval              0 if the name is valid or uncertain
 944  * \retval              negative error number on failure
 945  */
 946 int lfsck_namespace_check_name(const struct lu_env *env,
 947                                struct dt_object *parent,
 948                                struct dt_object *child,
 949                                const struct lu_name *cname)
 950 {
 951         struct lmv_mds_md_v1    *lmv = &lfsck_env_info(env)->lti_lmv;
 952         int                      idx;
 953         int                      rc;
 954
 955         rc = lfsck_read_stripe_lmv(env, parent, lmv);
 956         if (rc != 0)
 957                 RETURN(rc == -ENODATA ? 0 : rc);
 958
 959         if (lmv->lmv_magic == LMV_MAGIC_STRIPE) {
 960                 if (!lfsck_is_valid_slave_lmv(lmv))
 961                         return 0;
 962
 963                 idx = lmv_name_to_stripe_index(lmv->lmv_hash_type,
 964                                                lmv->lmv_stripe_count,
 965                                                cname->ln_name,
 966                                                cname->ln_namelen);
 967                 if (unlikely(idx != lmv->lmv_master_mdt_index))
 968                         return 1;
 969         } else if (lfsck_shard_name_to_index(env, cname->ln_name,
 970                         cname->ln_namelen, lfsck_object_type(child),
 971                         lfsck_dto2fid(child)) < 0) {
 972                 return 1;
 973         }
 974
 975         return 0;
 976 }
 977
 978 /**
 979  * Update the object's LMV EA with the given @lmv.
 980  *
 981  * \param[in] env       pointer to the thread context
 982  * \param[in] com       pointer to the lfsck component
 983  * \param[in] obj       pointer to the object which LMV EA will be updated
 984  * \param[in] lmv       pointer to buffer holding the new LMV EA
 985  * \param[in] locked    whether the caller has held ldlm lock on the @obj or not
 986  *
 987  * \retval              positive number for nothing to be done
 988  * \retval              zero if updated successfully
 989  * \retval              negative error number on failure
 990  */
 991 int lfsck_namespace_update_lmv(const struct lu_env *env,
 992                                struct lfsck_component *com,
 993                                struct dt_object *obj,
 994                                struct lmv_mds_md_v1 *lmv, bool locked)
 995 {
 996         struct lfsck_thread_info        *info   = lfsck_env_info(env);
 997         struct lmv_mds_md_v1            *lmv4   = &info->lti_lmv4;
 998         struct lu_buf                   *buf    = &info->lti_buf;
 999         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1000         struct dt_device                *dev    = lfsck_obj2dev(obj);
1001         struct thandle                  *th     = NULL;
1002         struct lustre_handle             lh     = { 0 };
1003         int                              rc     = 0;
1004         int                              rc1    = 0;
1005         ENTRY;
1006
1007         LASSERT(lmv4 != lmv);
1008
1009         lfsck_lmv_header_cpu_to_le(lmv4, lmv);
1010         lfsck_buf_init(buf, lmv4, sizeof(*lmv4));
1011
1012         if (!locked) {
1013                 rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
1014                                       MDS_INODELOCK_UPDATE |
1015                                       MDS_INODELOCK_XATTR, LCK_EX);
1016                 if (rc != 0)
1017                         GOTO(log, rc);
1018         }
1019
1020         th = dt_trans_create(env, dev);
1021         if (IS_ERR(th))
1022                 GOTO(log, rc = PTR_ERR(th));
1023
1024         /* For remote updating LMV EA, there will be further LFSCK action on
1025          * remote MDT after the updating, so update the LMV EA synchronously. */
1026         if (dt_object_remote(obj))
1027                 th->th_sync = 1;
1028
1029         rc = dt_declare_xattr_set(env, obj, buf, XATTR_NAME_LMV, 0, th);
1030         if (rc != 0)
1031                 GOTO(stop, rc);
1032
1033         rc = dt_trans_start_local(env, dev, th);
1034         if (rc != 0)
1035                 GOTO(stop, rc);
1036
1037         dt_write_lock(env, obj, 0);
1038         if (unlikely(lfsck_is_dead_obj(obj)))
1039                 GOTO(unlock, rc = 1);
1040
1041         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1042                 GOTO(unlock, rc = 0);
1043
1044         rc = dt_xattr_set(env, obj, buf, XATTR_NAME_LMV, 0, th);
1045
1046         GOTO(unlock, rc);
1047
1048 unlock:
1049         dt_write_unlock(env, obj);
1050
1051 stop:
1052         rc1 = dt_trans_stop(env, dev, th);
1053         if (rc == 0)
1054                 rc = rc1;
1055
1056 log:
1057         lfsck_ibits_unlock(&lh, LCK_EX);
1058         CDEBUG(D_LFSCK, "%s: namespace LFSCK updated the %s LMV EA "
1059                "for the object "DFID": rc = %d\n",
1060                lfsck_lfsck2name(lfsck),
1061                lmv->lmv_magic == LMV_MAGIC ? "master" : "slave",
1062                PFID(lfsck_dto2fid(obj)), rc);
1063
1064         return rc;
1065 }
1066
1067 /**
1068  * Check whether allow to re-genereate the lost master LMV EA.
1069  *
1070  * If the master MDT-object of the striped directory lost its master LMV EA,
1071  * then before the LFSCK repaired the striped directory, some ones may have
1072  * created some objects (that are not normal shards of the striped directory)
1073  * under the master MDT-object. If such case happend, then the LFSCK cannot
1074  * re-generate the lost master LMV EA to keep those objects to be visible to
1075  * client.
1076  *
1077  * \param[in] env       pointer to the thread context
1078  * \param[in] com       pointer to the lfsck component
1079  * \param[in] obj       pointer to the master MDT-object to be checked
1080  * \param[in] cfid      the shard's FID used for verification
1081  * \param[in] cidx      the shard's index used for verification
1082  *
1083  * \retval              positive number if not allow to re-generate LMV EA
1084  * \retval              zero if allow to re-generate LMV EA
1085  * \retval              negative error number on failure
1086  */
1087 static int lfsck_allow_regenerate_master_lmv(const struct lu_env *env,
1088                                              struct lfsck_component *com,
1089                                              struct dt_object *obj,
1090                                              const struct lu_fid *cfid,
1091                                              __u32 cidx)
1092 {
1093         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1094         struct lu_fid                   *tfid   = &info->lti_fid3;
1095         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1096         struct lu_dirent                *ent    =
1097                         (struct lu_dirent *)info->lti_key;
1098         const struct dt_it_ops          *iops;
1099         struct dt_it                    *di;
1100         __u64                            cookie;
1101         __u32                            args;
1102         int                              rc;
1103         __u16                            type;
1104         ENTRY;
1105
1106         if (unlikely(!dt_try_as_dir(env, obj)))
1107                 RETURN(-ENOTDIR);
1108
1109         /* Check whether the shard and the master MDT-object matches or not. */
1110         snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf), DFID":%u",
1111                  PFID(cfid), cidx);
1112         rc = dt_lookup(env, obj, (struct dt_rec *)tfid,
1113                        (const struct dt_key *)info->lti_tmpbuf);
1114         if (rc != 0)
1115                 RETURN(rc);
1116
1117         if (!lu_fid_eq(tfid, cfid))
1118                 RETURN(-ENOENT);
1119
1120         args = lfsck->li_args_dir & ~(LUDA_VERIFY | LUDA_VERIFY_DRYRUN);
1121         iops = &obj->do_index_ops->dio_it;
1122         di = iops->init(env, obj, args);
1123         if (IS_ERR(di))
1124                 RETURN(PTR_ERR(di));
1125
1126         rc = iops->load(env, di, 0);
1127         if (rc == 0)
1128                 rc = iops->next(env, di);
1129         else if (rc > 0)
1130                 rc = 0;
1131
1132         if (rc != 0)
1133                 GOTO(out, rc);
1134
1135         do {
1136                 rc = iops->rec(env, di, (struct dt_rec *)ent, args);
1137                 if (rc == 0)
1138                         rc = lfsck_unpack_ent(ent, &cookie, &type);
1139
1140                 if (rc != 0)
1141                         GOTO(out, rc);
1142
1143                 /* skip dot and dotdot entries */
1144                 if (name_is_dot_or_dotdot(ent->lde_name, ent->lde_namelen))
1145                         goto next;
1146
1147                 /* If the subdir name does not match the shard name rule, then
1148                  * it is quite possible that it is NOT a shard, but created by
1149                  * someone after the master MDT-object lost the master LMV EA.
1150                  * But it is also possible that the subdir name entry crashed,
1151                  * under such double failure cases, the LFSCK cannot know how
1152                  * to repair the inconsistency. For data safe, the LFSCK will
1153                  * mark the master MDT-object as read-only. The administrator
1154                  * can fix the bad shard name manually, then run LFSCK again.
1155                  *
1156                  * XXX: If the subdir name matches the shard name rule, but it
1157                  *      is not a real shard of the striped directory, instead,
1158                  *      it was created by someone after the master MDT-object
1159                  *      lost the LMV EA, then re-generating the master LMV EA
1160                  *      will cause such subdir to be invisible to client, and
1161                  *      if its index occupies some lost shard index, then the
1162                  *      LFSCK will use it to replace the bad shard, and cause
1163                  *      the subdir (itself) to be invisible for ever. */
1164                 if (lfsck_shard_name_to_index(env, ent->lde_name,
1165                                 ent->lde_namelen, type, &ent->lde_fid) < 0)
1166                         GOTO(out, rc = 1);
1167
1168 next:
1169                 rc = iops->next(env, di);
1170         } while (rc == 0);
1171
1172         GOTO(out, rc = 0);
1173
1174 out:
1175         iops->put(env, di);
1176         iops->fini(env, di);
1177
1178         return rc;
1179 }
1180
1181 /**
1182  * Notify remote LFSCK instance that the object's LMV EA has been updated.
1183  *
1184  * \param[in] env       pointer to the thread context
1185  * \param[in] com       pointer to the lfsck component
1186  * \param[in] obj       pointer to the object on which the LMV EA will be set
1187  * \param[in] event     indicate either master or slave LMV EA has been updated
1188  * \param[in] flags     indicate which element(s) in the LMV EA has been updated
1189  * \param[in] index     the MDT index on which the LFSCK instance to be notified
1190  *
1191  * \retval              positive number if nothing to be done
1192  * \retval              zero for success
1193  * \retval              negative error number on failure
1194  */
1195 static int lfsck_namespace_notify_lmv_remote(const struct lu_env *env,
1196                                              struct lfsck_component *com,
1197                                              struct dt_object *obj,
1198                                              __u32 event, __u32 flags,
1199                                              __u32 index)
1200 {
1201         struct lfsck_request            *lr     = &lfsck_env_info(env)->lti_lr;
1202         const struct lu_fid             *fid    = lfsck_dto2fid(obj);
1203         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1204         struct lfsck_tgt_desc           *ltd    = NULL;
1205         struct ptlrpc_request           *req    = NULL;
1206         int                              rc;
1207         ENTRY;
1208
1209         ltd = lfsck_tgt_get(&lfsck->li_mdt_descs, index);
1210         if (ltd == NULL)
1211                 GOTO(out, rc = -ENODEV);
1212
1213         req = ptlrpc_request_alloc(class_exp2cliimp(ltd->ltd_exp),
1214                                    &RQF_LFSCK_NOTIFY);
1215         if (req == NULL)
1216                 GOTO(out, rc = -ENOMEM);
1217
1218         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
1219         if (rc != 0) {
1220                 ptlrpc_request_free(req);
1221
1222                 GOTO(out, rc);
1223         }
1224
1225         lr = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
1226         memset(lr, 0, sizeof(*lr));
1227         lr->lr_event = event;
1228         lr->lr_index = lfsck_dev_idx(lfsck);
1229         lr->lr_active = LFSCK_TYPE_NAMESPACE;
1230         lr->lr_fid = *fid;
1231         lr->lr_flags = flags;
1232
1233         ptlrpc_request_set_replen(req);
1234         rc = ptlrpc_queue_wait(req);
1235         ptlrpc_req_finished(req);
1236
1237         GOTO(out, rc = (rc == -ENOENT ? 1 : rc));
1238
1239 out:
1240         CDEBUG(D_LFSCK, "%s: namespace LFSCK notify LMV EA updated for the "
1241                "object "DFID" on MDT %x remotely with event %u, flags %u: "
1242                "rc = %d\n", lfsck_lfsck2name(lfsck), PFID(fid), index,
1243                event, flags, rc);
1244
1245         if (ltd != NULL)
1246                 lfsck_tgt_put(ltd);
1247
1248         return rc;
1249 }
1250
1251 /**
1252  * Generate request for local LFSCK instance to rescan the striped directory.
1253  *
1254  * \param[in] env       pointer to the thread context
1255  * \param[in] com       pointer to the lfsck component
1256  * \param[in] obj       pointer to the striped directory to be rescanned
1257  *
1258  * \retval              positive number if nothing to be done
1259  * \retval              zero for success
1260  * \retval              negative error number on failure
1261  */
1262 int lfsck_namespace_notify_lmv_master_local(const struct lu_env *env,
1263                                             struct lfsck_component *com,
1264                                             struct dt_object *obj)
1265 {
1266         struct lfsck_instance      *lfsck = com->lc_lfsck;
1267         struct lfsck_namespace     *ns    = com->lc_file_ram;
1268         struct lmv_mds_md_v1       *lmv4  = &lfsck_env_info(env)->lti_lmv4;
1269         struct lfsck_lmv_unit      *llu;
1270         struct lfsck_lmv           *llmv;
1271         struct lfsck_slave_lmv_rec *lslr;
1272         int                         count = 0;
1273         int                         rc;
1274         ENTRY;
1275
1276         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1277                 RETURN(0);
1278
1279         rc = lfsck_read_stripe_lmv(env, obj, lmv4);
1280         if (rc != 0)
1281                 RETURN(rc);
1282
1283         OBD_ALLOC_PTR(llu);
1284         if (unlikely(llu == NULL))
1285                 RETURN(-ENOMEM);
1286
1287         if (lmv4->lmv_stripe_count < 1)
1288                 count = LFSCK_LMV_DEF_STRIPES;
1289         else if (lmv4->lmv_stripe_count > LFSCK_LMV_MAX_STRIPES)
1290                 count = LFSCK_LMV_MAX_STRIPES;
1291         else
1292                 count = lmv4->lmv_stripe_count;
1293
1294         OBD_ALLOC_LARGE(lslr, sizeof(struct lfsck_slave_lmv_rec) * count);
1295         if (lslr == NULL) {
1296                 OBD_FREE_PTR(llu);
1297
1298                 RETURN(-ENOMEM);
1299         }
1300
1301         INIT_LIST_HEAD(&llu->llu_link);
1302         llu->llu_lfsck = lfsck;
1303         llu->llu_obj = lfsck_object_get(obj);
1304         llmv = &llu->llu_lmv;
1305         llmv->ll_lmv_master = 1;
1306         llmv->ll_inline = 1;
1307         atomic_set(&llmv->ll_ref, 1);
1308         llmv->ll_stripes_allocated = count;
1309         llmv->ll_hash_type = LMV_HASH_TYPE_UNKNOWN;
1310         llmv->ll_lslr = lslr;
1311         llmv->ll_lmv = *lmv4;
1312
1313         down_write(&com->lc_sem);
1314         if (ns->ln_status != LS_SCANNING_PHASE1 &&
1315             ns->ln_status != LS_SCANNING_PHASE2) {
1316                 ns->ln_striped_dirs_skipped++;
1317                 up_write(&com->lc_sem);
1318                 lfsck_lmv_put(env, llmv);
1319         } else {
1320                 ns->ln_striped_dirs_repaired++;
1321                 spin_lock(&lfsck->li_lock);
1322                 list_add_tail(&llu->llu_link, &lfsck->li_list_lmv);
1323                 spin_unlock(&lfsck->li_lock);
1324                 up_write(&com->lc_sem);
1325         }
1326
1327         RETURN(0);
1328 }
1329
1330 /**
1331  * Set master LMV EA for the specified striped directory.
1332  *
1333  * First, if the master MDT-object of a striped directory lost its LMV EA,
1334  * then there may be some users have created some files under the master
1335  * MDT-object directly. Under such case, the LFSCK cannot re-generate LMV
1336  * EA for the master MDT-object, because we should keep the existing files
1337  * to be visible to client. Then the LFSCK will mark the striped directory
1338  * as read-only and keep it there to be handled by administrator manually.
1339  *
1340  * If nobody has created files under the master MDT-object of the striped
1341  * directory, then we will set the master LMV EA and generate a new rescan
1342  * (the striped directory) request that will be handled later by the LFSCK
1343  * instance on the MDT later.
1344  *
1345  * \param[in] env       pointer to the thread context
1346  * \param[in] com       pointer to the lfsck component
1347  * \param[in] obj       pointer to the object on which the LMV EA will be set
1348  * \param[in] lmv       pointer to the buffer holding the new LMV EA
1349  * \param[in] cfid      the shard's FID used for verification
1350  * \param[in] cidx      the shard's index used for verification
1351  * \param[in] flags     to indicate which element(s) in the LMV EA will be set
1352  *
1353  * \retval              positive number if nothing to be done
1354  * \retval              zero for success
1355  * \retval              negative error number on failure
1356  */
1357 static int lfsck_namespace_set_lmv_master(const struct lu_env *env,
1358                                           struct lfsck_component *com,
1359                                           struct dt_object *obj,
1360                                           struct lmv_mds_md_v1 *lmv,
1361                                           const struct lu_fid *cfid,
1362                                           __u32 cidx, __u32 flags)
1363 {
1364         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1365         struct lmv_mds_md_v1            *lmv3   = &info->lti_lmv3;
1366         struct lu_seq_range             *range  = &info->lti_range;
1367         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1368         struct seq_server_site          *ss     = lfsck_dev_site(lfsck);
1369         struct lustre_handle             lh     = { 0 };
1370         int                              pidx   = -1;
1371         int                              rc     = 0;
1372         ENTRY;
1373
1374         fld_range_set_mdt(range);
1375         rc = fld_server_lookup(env, ss->ss_server_fld,
1376                                fid_seq(lfsck_dto2fid(obj)), range);
1377         if (rc != 0)
1378                 GOTO(log, rc);
1379
1380         pidx = range->lsr_index;
1381         rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
1382                               MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1383                               LCK_EX);
1384         if (rc != 0)
1385                 GOTO(log, rc);
1386
1387         rc = lfsck_read_stripe_lmv(env, obj, lmv3);
1388         if (rc == -ENODATA) {
1389                 if (!(flags & LEF_SET_LMV_ALL))
1390                         GOTO(log, rc);
1391
1392                 *lmv3 = *lmv;
1393         } else if (rc == 0) {
1394                 if (flags & LEF_SET_LMV_ALL)
1395                         GOTO(log, rc = 1);
1396
1397                 if (flags & LEF_SET_LMV_HASH)
1398                         lmv3->lmv_hash_type = lmv->lmv_hash_type;
1399         } else {
1400                 GOTO(log, rc);
1401         }
1402
1403         lmv3->lmv_magic = LMV_MAGIC;
1404         lmv3->lmv_master_mdt_index = pidx;
1405
1406         if (flags & LEF_SET_LMV_ALL) {
1407                 rc = lfsck_allow_regenerate_master_lmv(env, com, obj,
1408                                                        cfid, cidx);
1409                 if (rc > 0) {
1410                         rc = lfsck_disable_master_lmv(env, com, obj, false);
1411
1412                         GOTO(log, rc = (rc == 0 ? 1 : rc));
1413                 }
1414
1415                 if (rc < 0)
1416                         GOTO(log, rc);
1417
1418                 /* To indicate that the master has ever lost LMV EA. */
1419                 lmv3->lmv_hash_type |= LMV_HASH_FLAG_LOST_LMV;
1420         }
1421
1422         rc = lfsck_namespace_update_lmv(env, com, obj, lmv3, true);
1423         if (rc == 0 && flags & LEF_SET_LMV_ALL) {
1424                 if (dt_object_remote(obj))
1425                         rc = lfsck_namespace_notify_lmv_remote(env, com, obj,
1426                                                 LE_SET_LMV_MASTER, 0, pidx);
1427                 else
1428                         rc = lfsck_namespace_notify_lmv_master_local(env, com,
1429                                                                      obj);
1430         }
1431
1432         GOTO(log, rc);
1433
1434 log:
1435         lfsck_ibits_unlock(&lh, LCK_EX);
1436         CDEBUG(D_LFSCK, "%s: namespace LFSCK set master LMV EA for the object "
1437                DFID" on the %s MDT %d, flags %x: rc = %d\n",
1438                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(obj)),
1439                dt_object_remote(obj) ? "remote" : "local", pidx, flags, rc);
1440
1441         if (rc <= 0) {
1442                 struct lfsck_namespace *ns = com->lc_file_ram;
1443
1444                 ns->ln_flags |= LF_INCONSISTENT;
1445         }
1446
1447         return rc;
1448 }
1449
1450 /**
1451  * Repair the bad name hash.
1452  *
1453  * If the name hash of some name entry under the striped directory does not
1454  * match the shard of the striped directory, then the LFSCK will repair the
1455  * inconsistency. Ideally, the LFSCK should migrate the name entry from the
1456  * current MDT to the right MDT (another one), but before the async commit
1457  * finished, the LFSCK will change the striped directory's hash type as
1458  * LMV_HASH_TYPE_UNKNOWN and mark the lmv flags as LMV_HASH_FLAG_BAD_TYPE.
1459  *
1460  * \param[in] env       pointer to the thread context
1461  * \param[in] com       pointer to the lfsck component
1462  * \param[in] shard     pointer to the shard of the striped directory that
1463  *                      contains the bad name entry
1464  * \param[in] llmv      pointer to lfsck LMV EA structure
1465  * \param[in] name      the name of the bad name hash
1466  *
1467  * \retval              positive number if nothing to be done
1468  * \retval              zero for success
1469  * \retval              negative error number on failure
1470  */
1471 int lfsck_namespace_repair_bad_name_hash(const struct lu_env *env,
1472                                          struct lfsck_component *com,
1473                                          struct dt_object *shard,
1474                                          struct lfsck_lmv *llmv,
1475                                          const char *name)
1476 {
1477         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1478         struct lu_fid                   *pfid   = &info->lti_fid3;
1479         struct lmv_mds_md_v1            *lmv2   = &info->lti_lmv2;
1480         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1481         struct dt_object                *parent = NULL;
1482         int                              rc     = 0;
1483         ENTRY;
1484
1485         rc = dt_lookup(env, shard, (struct dt_rec *)pfid,
1486                        (const struct dt_key *)dotdot);
1487         if (rc != 0 || !fid_is_sane(pfid))
1488                 GOTO(log, rc);
1489
1490         parent = lfsck_object_find_bottom(env, lfsck, pfid);
1491         if (IS_ERR(parent))
1492                 GOTO(log, rc = PTR_ERR(parent));
1493
1494         if (unlikely(!dt_object_exists(parent)))
1495                 /* The parent object was previously accessed when verifying
1496                  * the slave LMV EA.  If this condition is true it is because
1497                  * the striped directory is being removed. */
1498                 GOTO(log, rc = 1);
1499
1500         *lmv2 = llmv->ll_lmv;
1501         lmv2->lmv_hash_type = LMV_HASH_TYPE_UNKNOWN | LMV_HASH_FLAG_BAD_TYPE;
1502         rc = lfsck_namespace_set_lmv_master(env, com, parent, lmv2,
1503                                             lfsck_dto2fid(shard),
1504                                             llmv->ll_lmv.lmv_master_mdt_index,
1505                                             LEF_SET_LMV_HASH);
1506
1507         GOTO(log, rc);
1508
1509 log:
1510         CDEBUG(D_LFSCK, "%s: namespace LFSCK assistant found bad name hash "
1511                "on the MDT %x, parent "DFID", name %s, shard_%x "DFID
1512                ": rc = %d\n",
1513                lfsck_lfsck2name(lfsck), lfsck_dev_idx(lfsck),
1514                PFID(pfid), name, llmv->ll_lmv.lmv_master_mdt_index,
1515                PFID(lfsck_dto2fid(shard)), rc);
1516
1517         if (parent != NULL && !IS_ERR(parent))
1518                 lfsck_object_put(env, parent);
1519
1520         return rc;
1521 }
1522
1523 /**
1524  * Scan the shard of a striped directory for name hash verification.
1525  *
1526  * During the first-stage scanning, if the LFSCK cannot make sure whether
1527  * the shard of a stripe directory contains valid slave LMV EA or not, then
1528  * it will skip the name hash verification for this shard temporarily, and
1529  * record the shard's FID in the LFSCK tracing file. As the LFSCK processing,
1530  * the slave LMV EA may has been verified/fixed by LFSCK instance on master.
1531  * Then in the second-stage scanning, the shard will be re-scanned, and for
1532  * every name entry under the shard, the name hash will be verified, and for
1533  * unmatched name entry, the LFSCK will try to fix it.
1534  *
1535  * \param[in] env       pointer to the thread context
1536  * \param[in] com       pointer to the lfsck component
1537  * \param[in] child     pointer to the directory object to be handled
1538  *
1539  * \retval              positive number for scanning successfully
1540  * \retval              zero for the scanning is paused
1541  * \retval              negative error number on failure
1542  */
1543 int lfsck_namespace_scan_shard(const struct lu_env *env,
1544                                struct lfsck_component *com,
1545                                struct dt_object *child)
1546 {
1547         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1548         struct lmv_mds_md_v1            *lmv    = &info->lti_lmv;
1549         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1550         struct lfsck_namespace          *ns     = com->lc_file_ram;
1551         struct ptlrpc_thread            *thread = &lfsck->li_thread;
1552         struct lu_dirent                *ent    =
1553                         (struct lu_dirent *)info->lti_key;
1554         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
1555         struct lfsck_lmv                *llmv   = NULL;
1556         const struct dt_it_ops          *iops;
1557         struct dt_it                    *di;
1558         __u64                            cookie;
1559         __u32                            args;
1560         int                              rc;
1561         __u16                            type;
1562         ENTRY;
1563
1564         rc = lfsck_read_stripe_lmv(env, child, lmv);
1565         if (rc != 0)
1566                 RETURN(rc == -ENODATA ? 1 : rc);
1567
1568         if (lmv->lmv_magic != LMV_MAGIC_STRIPE)
1569                 RETURN(1);
1570
1571         if (unlikely(!dt_try_as_dir(env, child)))
1572                 RETURN(-ENOTDIR);
1573
1574         OBD_ALLOC_PTR(llmv);
1575         if (llmv == NULL)
1576                 RETURN(-ENOMEM);
1577
1578         llmv->ll_lmv_slave = 1;
1579         llmv->ll_lmv_verified = 1;
1580         llmv->ll_lmv = *lmv;
1581         atomic_set(&llmv->ll_ref, 1);
1582
1583         args = lfsck->li_args_dir & ~(LUDA_VERIFY | LUDA_VERIFY_DRYRUN);
1584         iops = &child->do_index_ops->dio_it;
1585         di = iops->init(env, child, args);
1586         if (IS_ERR(di))
1587                 GOTO(out, rc = PTR_ERR(di));
1588
1589         rc = iops->load(env, di, 0);
1590         if (rc == 0)
1591                 rc = iops->next(env, di);
1592         else if (rc > 0)
1593                 rc = 0;
1594
1595         while (rc == 0) {
1596                 if (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val) &&
1597                     unlikely(!thread_is_running(thread)))
1598                         GOTO(out, rc = 0);
1599
1600                 rc = iops->rec(env, di, (struct dt_rec *)ent, args);
1601                 if (rc == 0)
1602                         rc = lfsck_unpack_ent(ent, &cookie, &type);
1603
1604                 if (rc != 0) {
1605                         if (bk->lb_param & LPF_FAILOUT)
1606                                 GOTO(out, rc);
1607
1608                         goto next;
1609                 }
1610
1611                 /* skip dot and dotdot entries */
1612                 if (name_is_dot_or_dotdot(ent->lde_name, ent->lde_namelen))
1613                         goto next;
1614
1615                 if (!lfsck_is_valid_slave_name_entry(env, llmv, ent->lde_name,
1616                                                      ent->lde_namelen)) {
1617                         ns->ln_flags |= LF_INCONSISTENT;
1618                         rc = lfsck_namespace_repair_bad_name_hash(env, com,
1619                                                 child, llmv, ent->lde_name);
1620                         if (rc == 0)
1621                                 ns->ln_name_hash_repaired++;
1622                 }
1623
1624                 if (rc < 0 && bk->lb_param & LPF_FAILOUT)
1625                         GOTO(out, rc);
1626
1627                 /* Rate control. */
1628                 lfsck_control_speed(lfsck);
1629                 if (unlikely(!thread_is_running(thread)))
1630                         GOTO(out, rc = 0);
1631
1632                 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_FATAL2)) {
1633                         spin_lock(&lfsck->li_lock);
1634                         thread_set_flags(thread, SVC_STOPPING);
1635                         spin_unlock(&lfsck->li_lock);
1636
1637                         GOTO(out, rc = -EINVAL);
1638                 }
1639
1640 next:
1641                 rc = iops->next(env, di);
1642         }
1643
1644         GOTO(out, rc);
1645
1646 out:
1647         iops->put(env, di);
1648         iops->fini(env, di);
1649         lfsck_lmv_put(env, llmv);
1650
1651         return rc;
1652 }
1653
1654 /**
1655  * Verify the slave object's (of striped directory) LMV EA.
1656  *
1657  * For the slave object of a striped directory, before traversing the shard
1658  * the LFSCK will verify whether its slave LMV EA matches its parent's master
1659  * LMV EA or not.
1660  *
1661  * \param[in] env       pointer to the thread context
1662  * \param[in] com       pointer to the lfsck component
1663  * \param[in] obj       pointer to the object which LMV EA will be checked
1664  * \param[in] llmv      pointer to buffer holding the slave LMV EA
1665  *
1666  * \retval              positive number if nothing to be done
1667  * \retval              zero for success
1668  * \retval              negative error number on failure
1669  */
1670 int lfsck_namespace_verify_stripe_slave(const struct lu_env *env,
1671                                         struct lfsck_component *com,
1672                                         struct dt_object *obj,
1673                                         struct lfsck_lmv *llmv)
1674 {
1675         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1676         char                            *name   = info->lti_key;
1677         char                            *name2;
1678         struct lu_fid                   *pfid   = &info->lti_fid3;
1679         const struct lu_fid             *cfid   = lfsck_dto2fid(obj);
1680         struct lu_fid                    tfid;
1681         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1682         struct lmv_mds_md_v1            *clmv   = &llmv->ll_lmv;
1683         struct lmv_mds_md_v1            *plmv   = &info->lti_lmv;
1684         struct dt_object                *parent = NULL;
1685         int                              rc     = 0;
1686         ENTRY;
1687
1688         if (!lfsck_is_valid_slave_lmv(clmv)) {
1689                 rc = lfsck_namespace_trace_update(env, com, cfid,
1690                                         LNTF_UNCERTAIN_LMV, true);
1691
1692                 GOTO(out, rc);
1693         }
1694
1695         rc = dt_lookup(env, obj, (struct dt_rec *)pfid,
1696                        (const struct dt_key *)dotdot);
1697         if (rc != 0 || !fid_is_sane(pfid)) {
1698                 rc = lfsck_namespace_trace_update(env, com, cfid,
1699                                         LNTF_UNCERTAIN_LMV, true);
1700
1701                 GOTO(out, rc);
1702         }
1703
1704         parent = lfsck_object_find_bottom(env, lfsck, pfid);
1705         if (IS_ERR(parent)) {
1706                 rc = lfsck_namespace_trace_update(env, com, cfid,
1707                                         LNTF_UNCERTAIN_LMV, true);
1708
1709                 GOTO(out, rc);
1710         }
1711
1712         if (unlikely(!dt_object_exists(parent)))
1713                 GOTO(out, rc = 1);
1714
1715         if (unlikely(!dt_try_as_dir(env, parent)))
1716                 GOTO(out, rc = -ENOTDIR);
1717
1718         rc = lfsck_read_stripe_lmv(env, parent, plmv);
1719         if (rc != 0) {
1720                 int rc1;
1721
1722                 /* If the parent has no LMV EA, then it maybe because:
1723                  * 1) The parent lost the LMV EA.
1724                  * 2) The child claims a wrong (slave) LMV EA. */
1725                 if (rc == -ENODATA)
1726                         rc = lfsck_namespace_set_lmv_master(env, com, parent,
1727                                         clmv, cfid, clmv->lmv_master_mdt_index,
1728                                         LEF_SET_LMV_ALL);
1729                 else
1730                         rc = 0;
1731
1732                 rc1 = lfsck_namespace_trace_update(env, com, cfid,
1733                                                    LNTF_UNCERTAIN_LMV, true);
1734
1735                 GOTO(out, rc = (rc < 0 ? rc : rc1));
1736         }
1737
1738         /* Unmatched magic or stripe count. */
1739         if (unlikely(plmv->lmv_magic != LMV_MAGIC ||
1740                      plmv->lmv_stripe_count != clmv->lmv_stripe_count)) {
1741                 rc = lfsck_namespace_trace_update(env, com, cfid,
1742                                                   LNTF_UNCERTAIN_LMV, true);
1743
1744                 GOTO(out, rc);
1745         }
1746
1747         /* If the master hash type has been set as LMV_HASH_TYPE_UNKNOWN,
1748          * then the slave hash type is not important. */
1749         if ((plmv->lmv_hash_type & LMV_HASH_TYPE_MASK) ==
1750             LMV_HASH_TYPE_UNKNOWN &&
1751             plmv->lmv_hash_type & LMV_HASH_FLAG_BAD_TYPE)
1752                 GOTO(out, rc = 0);
1753
1754         /* Unmatched hash type. */
1755         if (unlikely((plmv->lmv_hash_type & LMV_HASH_TYPE_MASK) !=
1756                      (clmv->lmv_hash_type & LMV_HASH_TYPE_MASK))) {
1757                 rc = lfsck_namespace_trace_update(env, com, cfid,
1758                                                   LNTF_UNCERTAIN_LMV, true);
1759
1760                 GOTO(out, rc);
1761         }
1762
1763         snprintf(info->lti_tmpbuf2, sizeof(info->lti_tmpbuf2), DFID":%u",
1764                  PFID(cfid), clmv->lmv_master_mdt_index);
1765         name2 = info->lti_tmpbuf2;
1766
1767         rc = lfsck_links_get_first(env, obj, name, &tfid);
1768         if (rc == 0 && strcmp(name, name2) == 0 && lu_fid_eq(pfid, &tfid)) {
1769                 llmv->ll_lmv_verified = 1;
1770
1771                 GOTO(out, rc);
1772         }
1773
1774         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
1775                        (const struct dt_key *)name2);
1776         if (rc != 0 || !lu_fid_eq(cfid, &tfid))
1777                 rc = lfsck_namespace_trace_update(env, com, cfid,
1778                                                   LNTF_UNCERTAIN_LMV, true);
1779         else
1780                 llmv->ll_lmv_verified = 1;
1781
1782         GOTO(out, rc);
1783
1784 out:
1785         if (parent != NULL && !IS_ERR(parent))
1786                 lfsck_object_put(env, parent);
1787
1788         return rc;
1789 }
1790
1791 /**
1792  * Double scan the striped directory or the shard.
1793  *
1794  * All the shards' under the given striped directory or its shard have
1795  * been scanned, the LFSCK has got the global knownledge about the LMV
1796  * EA consistency.
1797  *
1798  * If the target is one shard of a striped directory, then only needs to
1799  * update related tracing file.
1800  *
1801  * If the target is the master MDT-object of a striped directory, then the
1802  * LFSCK will make the decision about whether the master LMV EA is invalid
1803  * or not, and repair it if inconsistenct; for every shard of the striped
1804  * directory, whether the slave LMV EA is invalid or not, and repair it if
1805  * inconsistent.
1806  *
1807  * \param[in] env       pointer to the thread context
1808  * \param[in] com       pointer to the lfsck component
1809  * \param[in] lnr       pointer to the namespace request that contains the
1810  *                      striped directory or the shard
1811  *
1812  * \retval              zero for success
1813  * \retval              negative error number on failure
1814  */
1815 int lfsck_namespace_striped_dir_rescan(const struct lu_env *env,
1816                                        struct lfsck_component *com,
1817                                        struct lfsck_namespace_req *lnr)
1818 {
1819         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1820         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1821         struct lfsck_namespace          *ns     = com->lc_file_ram;
1822         struct lfsck_lmv                *llmv   = lnr->lnr_lmv;
1823         struct lmv_mds_md_v1            *lmv    = &llmv->ll_lmv;
1824         struct lmv_mds_md_v1            *lmv2   = &info->lti_lmv2;
1825         struct lfsck_assistant_object   *lso    = lnr->lnr_lar.lar_parent;
1826         const struct lu_fid             *pfid   = &lso->lso_fid;
1827         struct dt_object                *dir    = NULL;
1828         struct dt_object                *obj    = NULL;
1829         struct lu_seq_range             *range  = &info->lti_range;
1830         struct seq_server_site          *ss     = lfsck_dev_site(lfsck);
1831         __u32                            stripe_count;
1832         __u32                            hash_type;
1833         int                              rc     = 0;
1834         int                              i;
1835         ENTRY;
1836
1837         if (llmv->ll_lmv_slave) {
1838                 if (llmv->ll_lmv_verified) {
1839                         ns->ln_striped_shards_scanned++;
1840                         lfsck_namespace_trace_update(env, com, pfid,
1841                                         LNTF_UNCERTAIN_LMV |
1842                                         LNTF_RECHECK_NAME_HASH, false);
1843                 }
1844
1845                 RETURN(0);
1846         }
1847
1848         /* Either the striped directory has been disabled or only part of
1849          * the striped directory have been scanned. The LFSCK cannot repair
1850          * something based on incompleted knowledge. So skip it. */
1851         if (llmv->ll_ignore || llmv->ll_exit_value <= 0)
1852                 RETURN(0);
1853
1854         /* There ever been some failure, as to the LFSCK cannot know whether
1855          * it has got the global knowledge about the LMV EA consistency or not,
1856          * so it cannot make reparation about the incompleted knowledge. */
1857         if (llmv->ll_failed) {
1858                 ns->ln_striped_dirs_scanned++;
1859                 ns->ln_striped_dirs_failed++;
1860
1861                 RETURN(0);
1862         }
1863
1864         if (lmv->lmv_stripe_count > LFSCK_LMV_MAX_STRIPES)
1865                 stripe_count = max(llmv->ll_max_filled_off + 1,
1866                                    llmv->ll_max_stripe_count);
1867         else
1868                 stripe_count = max(llmv->ll_max_filled_off + 1,
1869                                    lmv->lmv_stripe_count);
1870
1871         if (lmv->lmv_stripe_count != stripe_count) {
1872                 lmv->lmv_stripe_count = stripe_count;
1873                 llmv->ll_lmv_updated = 1;
1874         }
1875
1876         if (!lmv_is_known_hash_type(lmv->lmv_hash_type) &&
1877             !(lmv->lmv_hash_type & LMV_HASH_FLAG_BAD_TYPE) &&
1878             lmv_is_known_hash_type(llmv->ll_hash_type)) {
1879                 hash_type = llmv->ll_hash_type & LMV_HASH_TYPE_MASK;
1880                 lmv->lmv_hash_type = llmv->ll_hash_type;
1881                 llmv->ll_lmv_updated = 1;
1882         } else {
1883                 hash_type = lmv->lmv_hash_type & LMV_HASH_TYPE_MASK;
1884                 if (!lmv_is_known_hash_type(hash_type))
1885                         hash_type = LMV_HASH_TYPE_UNKNOWN;
1886         }
1887
1888         if (llmv->ll_lmv_updated) {
1889                 if (dir == NULL) {
1890                         dir = lfsck_assistant_object_load(env, lfsck, lso);
1891                         if (IS_ERR(dir)) {
1892                                 rc = PTR_ERR(dir);
1893
1894                                 RETURN(rc == -ENOENT ? 0 : rc);
1895                         }
1896                 }
1897
1898                 lmv->lmv_layout_version++;
1899                 rc = lfsck_namespace_update_lmv(env, com, dir, lmv, false);
1900                 if (rc != 0)
1901                         RETURN(rc);
1902
1903                 ns->ln_striped_dirs_scanned++;
1904                 ns->ln_striped_dirs_repaired++;
1905         }
1906
1907         fld_range_set_mdt(range);
1908         for (i = 0; i <= llmv->ll_max_filled_off; i++) {
1909                 struct lfsck_slave_lmv_rec *lslr = llmv->ll_lslr + i;
1910                 const struct lu_fid *cfid = &lslr->lslr_fid;
1911                 const struct lu_name *cname;
1912                 struct linkea_data ldata = { NULL };
1913                 int len;
1914                 int rc1 = 0;
1915                 bool repair_linkea = false;
1916                 bool repair_lmvea = false;
1917                 bool rename = false;
1918                 bool create = false;
1919                 bool linkea_repaired = false;
1920                 bool lmvea_repaired = false;
1921                 bool rename_repaired = false;
1922                 bool create_repaired = false;
1923
1924                 /* LMV EA hole. */
1925                 if (fid_is_zero(cfid))
1926                         continue;
1927
1928                 len = snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf),
1929                                DFID":%u", PFID(cfid), i);
1930                 cname = lfsck_name_get_const(env, info->lti_tmpbuf, len);
1931                 memcpy(lnr->lnr_name, info->lti_tmpbuf, len);
1932
1933                 obj = lfsck_object_find_bottom(env, lfsck, cfid);
1934                 if (IS_ERR(obj)) {
1935                         if (dir == NULL) {
1936                                 dir = lfsck_assistant_object_load(env, lfsck,
1937                                                                   lso);
1938                                 if (IS_ERR(dir)) {
1939                                         if (PTR_ERR(dir) == -ENOENT)
1940                                                 RETURN(0);
1941
1942                                         dir = NULL;
1943                                 }
1944                         } else if (lfsck_is_dead_obj(dir)) {
1945                                 GOTO(out, rc = 0);
1946                         }
1947
1948                         rc1 = PTR_ERR(obj);
1949                         goto next;
1950                 }
1951
1952                 switch (lslr->lslr_flags) {
1953                 case LSLF_NONE:
1954                         if (llmv->ll_inline ||
1955                             lslr->lslr_stripe_count != stripe_count ||
1956                             (lslr->lslr_hash_type & LMV_HASH_TYPE_MASK) !=
1957                              hash_type)
1958                                 repair_lmvea = true;
1959                         break;
1960                 case LSLF_BAD_INDEX2:
1961                         /* The index in the slave LMV EA is right,
1962                          * the name entry should be updated. */
1963                         rename = true;
1964                         snprintf(info->lti_tmpbuf2, sizeof(info->lti_tmpbuf2),
1965                                  DFID":%u", PFID(cfid), lslr->lslr_index);
1966                         if (llmv->ll_inline ||
1967                             lslr->lslr_stripe_count != stripe_count ||
1968                             (lslr->lslr_hash_type & LMV_HASH_TYPE_MASK) !=
1969                              hash_type)
1970                                 repair_lmvea = true;
1971                         break;
1972                 case LSLF_BAD_INDEX1:
1973                         /* The index in the name entry is right,
1974                          * the slave LMV EA should be updated. */
1975                 case LSLF_NO_LMVEA:
1976                         repair_lmvea = true;
1977                         break;
1978                 case LSLF_DANGLING:
1979                         create = true;
1980                         goto repair;
1981                 default:
1982                         break;
1983                 }
1984
1985                 rc1 = lfsck_links_read_with_rec(env, obj, &ldata);
1986                 if (rc1 == -ENOENT) {
1987                         create = true;
1988                         goto repair;
1989                 }
1990
1991                 if (rc1 == -EINVAL || rc1 == -ENODATA) {
1992                         repair_linkea = true;
1993                         goto repair;
1994                 }
1995
1996                 if (rc1 != 0)
1997                         goto next;
1998
1999                 if (ldata.ld_leh->leh_reccount != 1) {
2000                         repair_linkea = true;
2001                         goto repair;
2002                 }
2003
2004                 rc1 = linkea_links_find(&ldata, cname, pfid);
2005                 if (rc1 != 0)
2006                         repair_linkea = true;
2007
2008 repair:
2009                 if (create) {
2010                         if (dir == NULL) {
2011                                 dir = lfsck_assistant_object_load(env, lfsck,
2012                                                                   lso);
2013                                 if (IS_ERR(dir)) {
2014                                         rc1 = PTR_ERR(dir);
2015
2016                                         if (rc1 == -ENOENT)
2017                                                 GOTO(out, rc = 0);
2018
2019                                         dir = NULL;
2020                                         goto next;
2021                                 }
2022                         }
2023
2024                         rc1 = lfsck_namespace_repair_dangling(env, com, dir,
2025                                                               obj, lnr);
2026                         if (rc1 >= 0) {
2027                                 create_repaired = true;
2028                                 if (rc == 0)
2029                                         ns->ln_dangling_repaired++;
2030                         }
2031                 }
2032
2033                 if (repair_lmvea) {
2034                         *lmv2 = *lmv;
2035                         lmv2->lmv_magic = LMV_MAGIC_STRIPE;
2036                         lmv2->lmv_stripe_count = stripe_count;
2037                         lmv2->lmv_master_mdt_index = i;
2038                         lmv2->lmv_hash_type = hash_type;
2039
2040                         rc1 = lfsck_namespace_update_lmv(env, com, obj,
2041                                                          lmv2, false);
2042                         if (rc1 < 0)
2043                                 goto next;
2044
2045                         if (dt_object_remote(obj)) {
2046                                 rc1 = fld_server_lookup(env, ss->ss_server_fld,
2047                                         fid_seq(lfsck_dto2fid(obj)), range);
2048                                 if (rc1 != 0)
2049                                         goto next;
2050
2051                                 rc1 = lfsck_namespace_notify_lmv_remote(env,
2052                                                 com, obj, LE_SET_LMV_SLAVE, 0,
2053                                                 range->lsr_index);
2054                         } else {
2055                                 ns->ln_striped_shards_repaired++;
2056                                 rc1 = lfsck_namespace_trace_update(env, com,
2057                                         cfid, LNTF_RECHECK_NAME_HASH, true);
2058                         }
2059
2060                         if (rc1 < 0)
2061                                 goto next;
2062
2063                         if (rc1 >= 0)
2064                                 lmvea_repaired = true;
2065                 } else if (llmv->ll_inline) {
2066                         if (dt_object_remote(obj)) {
2067                                 rc1 = fld_server_lookup(env, ss->ss_server_fld,
2068                                         fid_seq(lfsck_dto2fid(obj)), range);
2069                                 if (rc1 != 0)
2070                                         goto next;
2071
2072                                 /* The slave LMV EA on the remote shard is
2073                                  * correct, just notify the LFSCK instance
2074                                  * on such MDT to re-verify the name_hash. */
2075                                 rc1 = lfsck_namespace_notify_lmv_remote(env,
2076                                                 com, obj, LE_SET_LMV_SLAVE,
2077                                                 LEF_RECHECK_NAME_HASH,
2078                                                 range->lsr_index);
2079                         } else {
2080                                 rc1 = lfsck_namespace_trace_update(env, com,
2081                                         cfid, LNTF_RECHECK_NAME_HASH, true);
2082                         }
2083
2084                         if (rc1 < 0)
2085                                 goto next;
2086                 }
2087
2088                 if (rename) {
2089                         if (dir == NULL) {
2090                                 dir = lfsck_assistant_object_load(env, lfsck,
2091                                                                   lso);
2092                                 if (IS_ERR(dir)) {
2093                                         rc1 = PTR_ERR(dir);
2094
2095                                         if (rc1 == -ENOENT)
2096                                                 GOTO(out, rc = 0);
2097
2098                                         dir = NULL;
2099                                         goto next;
2100                                 }
2101                         }
2102
2103                         rc1 = lfsck_namespace_repair_dirent(env, com, dir, obj,
2104                                         info->lti_tmpbuf2, lnr->lnr_name,
2105                                         lnr->lnr_type, true, false);
2106                         if (rc1 >= 0) {
2107                                 rename_repaired = true;
2108                                 if (rc1 > 0) {
2109                                         ns->ln_dirent_repaired++;
2110                                         rc1 = lfsck_namespace_trace_update(env,
2111                                                 com, cfid,
2112                                                 LNTF_RECHECK_NAME_HASH, true);
2113                                 }
2114                         }
2115
2116                         if (rc1 < 0)
2117                                 goto next;
2118                 }
2119
2120                 if (repair_linkea) {
2121                         struct lustre_handle lh = { 0 };
2122
2123                         rc1 = linkea_links_new(&ldata, &info->lti_big_buf,
2124                                                cname, lfsck_dto2fid(dir));
2125                         if (rc1 != 0)
2126                                 goto next;
2127
2128                         if (dir == NULL) {
2129                                 dir = lfsck_assistant_object_load(env, lfsck,
2130                                                                   lso);
2131                                 if (IS_ERR(dir)) {
2132                                         rc1 = PTR_ERR(dir);
2133
2134                                         if (rc1 == -ENOENT)
2135                                                 GOTO(out, rc = 0);
2136
2137                                         dir = NULL;
2138                                         goto next;
2139                                 }
2140                         }
2141
2142                         rc1 = lfsck_ibits_lock(env, lfsck, obj, &lh,
2143                                                MDS_INODELOCK_UPDATE |
2144                                                MDS_INODELOCK_XATTR, LCK_EX);
2145                         if (rc1 != 0)
2146                                 goto next;
2147
2148                         rc1 = lfsck_namespace_rebuild_linkea(env, com, obj,
2149                                                              &ldata);
2150                         lfsck_ibits_unlock(&lh, LCK_EX);
2151                         if (rc1 >= 0) {
2152                                 linkea_repaired = true;
2153                                 if (rc1 > 0)
2154                                         ns->ln_linkea_repaired++;
2155                         }
2156                 }
2157
2158 next:
2159                 CDEBUG(D_LFSCK, "%s: namespace LFSCK repair the shard "
2160                       "%d "DFID" of the striped directory "DFID" with "
2161                       "dangling %s/%s, rename %s/%s, llinkea %s/%s, "
2162                       "repair_lmvea %s/%s: rc = %d\n", lfsck_lfsck2name(lfsck),
2163                       i, PFID(cfid), PFID(&lnr->lnr_fid),
2164                       create ? "yes" : "no", create_repaired ? "yes" : "no",
2165                       rename ? "yes" : "no", rename_repaired ? "yes" : "no",
2166                       repair_linkea ? "yes" : "no",
2167                       linkea_repaired ? "yes" : "no",
2168                       repair_lmvea ? "yes" : "no",
2169                       lmvea_repaired ? "yes" : "no", rc1);
2170
2171                 if (obj != NULL && !IS_ERR(obj)) {
2172                         lfsck_object_put(env, obj);
2173                         obj = NULL;
2174                 }
2175
2176                 if (rc1 < 0) {
2177                         rc = rc1;
2178                         ns->ln_striped_shards_failed++;
2179                 }
2180         }
2181
2182         GOTO(out, rc);
2183
2184 out:
2185         if (obj != NULL && !IS_ERR(obj))
2186                 lfsck_object_put(env, obj);
2187
2188         if (dir != NULL && !IS_ERR(dir))
2189                 lfsck_object_put(env, dir);
2190
2191         return rc;
2192 }
2193
2194 /**
2195  * Verify the shard's name entry under the striped directory.
2196  *
2197  * Before all shards of the striped directory scanned, the LFSCK cannot
2198  * know whether the master LMV EA is valid or not, and also cannot know
2199  * how to repair an invalid shard exactly. For example, the stripe index
2200  * stored in the shard's name does not match the stripe index stored in
2201  * the slave LMV EA, then the LFSCK cannot know which one is correct.
2202  * If the LFSCK just assumed one is correct, and fixed the other, then
2203  * as the LFSCK processing, it may find that the former reparation is
2204  * wrong and have to roll back. Unfortunately, if some applications saw
2205  * the changes and made further modification based on such changes, then
2206  * the roll back is almost impossible.
2207  *
2208  * To avoid above trouble, the LFSCK will scan the master object of the
2209  * striped directory twice, that is NOT the same as normal two-stages
2210  * scanning, the double scanning the striped directory will happen both
2211  * during the first-stage scanning:
2212  *
2213  * 1) When the striped directory is opened for scanning, the LFSCK will
2214  *    iterate each shard in turn, and records its slave LMV EA in the
2215  *    lfsck_lmv::ll_lslr. In this step, if the 'shard' (may be fake
2216  *    shard) name does not match the shard naming rule, for example, it
2217  *    does not contains the shard's FID, or not contains index, then we
2218  *    can remove the bad name entry directly. But if the name is valid,
2219  *    but the shard has no slave LMV EA or the slave LMV EA does not
2220  *    match its name, then we just record related information in the
2221  *    lfsck_lmv::ll_lslr in RAM.
2222  *
2223  * 2) When all the known shards have been scanned, then the engine will
2224  *    generate a dummy request (via lfsck_namespace_close_dir) to tell
2225  *    the assistant thread that all the known shards have been scanned.
2226  *    Since the assistant has got the global knowledge about the index
2227  *    conflict, stripe count, hash type, and so on. Then the assistant
2228  *    thread will scan the lfsck_lmv::ll_lslr, and for every shard in
2229  *    the record, check and repair inconsistency.
2230  *
2231  * Generally, the stripe directory has only several shards, and there
2232  * will NOT be a lof of striped directory. So double scanning striped
2233  * directory will not much affect the LFSCK performance.
2234  *
2235  * \param[in] env       pointer to the thread context
2236  * \param[in] com       pointer to the lfsck component
2237  * \param[in] lnr       pointer to the namespace request that contains the
2238  *                      shard's name, parent object, parent's LMV, and ect.
2239  *
2240  * \retval              zero for success
2241  * \retval              negative error number on failure
2242  */
2243 int lfsck_namespace_handle_striped_master(const struct lu_env *env,
2244                                           struct lfsck_component *com,
2245                                           struct lfsck_namespace_req *lnr)
2246 {
2247         struct lfsck_thread_info   *info        = lfsck_env_info(env);
2248         struct lmv_mds_md_v1       *lmv         = &info->lti_lmv;
2249         struct lfsck_instance      *lfsck       = com->lc_lfsck;
2250         struct lfsck_namespace     *ns          = com->lc_file_ram;
2251         struct lfsck_lmv           *llmv        = lnr->lnr_lmv;
2252         struct lfsck_assistant_object *lso      = lnr->lnr_lar.lar_parent;
2253         const struct lu_fid        *pfid        = &lso->lso_fid;
2254         struct dt_object           *dir;
2255         struct dt_object           *obj         = NULL;
2256         struct dt_device           *dev         = NULL;
2257         int                         shard_idx   = 0;
2258         int                         stripe      = 0;
2259         int                         rc          = 0;
2260         int                         depth       = 0;
2261         bool                        repaired    = false;
2262         enum lfsck_namespace_inconsistency_type type = LNIT_NONE;
2263         ENTRY;
2264
2265         if (unlikely(llmv->ll_ignore))
2266                 RETURN(0);
2267
2268         dir = lfsck_assistant_object_load(env, lfsck, lso);
2269         if (IS_ERR(dir)) {
2270                 rc = PTR_ERR(dir);
2271
2272                 RETURN(rc == -ENOENT ? 0 : rc);
2273         }
2274
2275         shard_idx = lfsck_find_mdt_idx_by_fid(env, lfsck, &lnr->lnr_fid);
2276         if (shard_idx < 0)
2277                 GOTO(fail_lmv, rc = shard_idx);
2278
2279         if (shard_idx == lfsck_dev_idx(lfsck)) {
2280                 if (unlikely(strcmp(lnr->lnr_name, dotdot) == 0))
2281                         GOTO(out, rc = 0);
2282
2283                 dev = lfsck->li_bottom;
2284         } else {
2285                 struct lfsck_tgt_desc *ltd;
2286
2287                 /* Usually, some local filesystem consistency verification
2288                  * tools can guarantee the local namespace tree consistenct.
2289                  * So the LFSCK will only verify the remote directory. */
2290                 if (unlikely(strcmp(lnr->lnr_name, dotdot) == 0)) {
2291                         rc = lfsck_namespace_trace_update(env, com, pfid,
2292                                                 LNTF_CHECK_PARENT, true);
2293
2294                         GOTO(out, rc);
2295                 }
2296
2297                 ltd = lfsck_ltd2tgt(&lfsck->li_mdt_descs, shard_idx);
2298                 if (unlikely(ltd == NULL)) {
2299                         CDEBUG(D_LFSCK, "%s: cannot talk with MDT %x which "
2300                                "did not join the namespace LFSCK\n",
2301                                lfsck_lfsck2name(lfsck), shard_idx);
2302                         lfsck_lad_set_bitmap(env, com, shard_idx);
2303
2304                         GOTO(fail_lmv, rc = -ENODEV);
2305                 }
2306
2307                 dev = ltd->ltd_tgt;
2308         }
2309
2310         obj = lfsck_object_find_by_dev(env, dev, &lnr->lnr_fid);
2311         if (IS_ERR(obj)) {
2312                 if (lfsck_is_dead_obj(dir))
2313                         RETURN(0);
2314
2315                 GOTO(fail_lmv, rc = PTR_ERR(obj));
2316         }
2317
2318         if (!dt_object_exists(obj)) {
2319                 stripe = lfsck_shard_name_to_index(env, lnr->lnr_name,
2320                                 lnr->lnr_namelen, lnr->lnr_type, &lnr->lnr_fid);
2321                 if (stripe < 0) {
2322                         type = LNIT_BAD_DIRENT;
2323
2324                         GOTO(out, rc = 0);
2325                 }
2326
2327 dangling:
2328                 rc = lfsck_namespace_check_exist(env, dir, obj, lnr->lnr_name);
2329                 if (rc == 0) {
2330                         memset(lmv, 0, sizeof(*lmv));
2331                         lmv->lmv_magic = LMV_MAGIC;
2332                         rc = lfsck_record_lmv(env, com, dir, lnr, lmv, stripe,
2333                                               LSLF_DANGLING, LSLF_NONE, &depth);
2334                 }
2335
2336                 GOTO(out, rc);
2337         }
2338
2339         stripe = lfsck_shard_name_to_index(env, lnr->lnr_name, lnr->lnr_namelen,
2340                                            lfsck_object_type(obj),
2341                                            &lnr->lnr_fid);
2342         if (stripe < 0) {
2343                 type = LNIT_BAD_DIRENT;
2344
2345                 GOTO(out, rc = 0);
2346         }
2347
2348         rc = lfsck_read_stripe_lmv(env, obj, lmv);
2349         if (unlikely(rc == -ENOENT))
2350                 /* It may happen when the remote object has been removed,
2351                  * but the local MDT does not aware of that. */
2352                 goto dangling;
2353
2354         if (rc == -ENODATA)
2355                 rc = lfsck_record_lmv(env, com, dir, lnr, lmv, stripe,
2356                                       LSLF_NO_LMVEA, LSLF_NONE, &depth);
2357         else if (rc == 0)
2358                 rc = lfsck_record_lmv(env, com, dir, lnr, lmv, stripe,
2359                                       lmv->lmv_master_mdt_index != stripe ?
2360                                       LSLF_BAD_INDEX1 : LSLF_NONE, LSLF_NONE,
2361                                       &depth);
2362
2363         GOTO(out, rc);
2364
2365 fail_lmv:
2366         llmv->ll_failed = 1;
2367
2368 out:
2369         if (rc >= 0 && type == LNIT_NONE && !S_ISDIR(lnr->lnr_type))
2370                 type = LNIT_BAD_TYPE;
2371
2372         switch (type) {
2373         case LNIT_BAD_TYPE:
2374                 rc = lfsck_namespace_repair_dirent(env, com, dir, obj,
2375                                                    lnr->lnr_name, lnr->lnr_name,
2376                                                    lnr->lnr_type, true, false);
2377                 if (rc > 0)
2378                         repaired = true;
2379                 break;
2380         case LNIT_BAD_DIRENT:
2381                 rc = lfsck_namespace_repair_dirent(env, com, dir, obj,
2382                                                    lnr->lnr_name, lnr->lnr_name,
2383                                                    lnr->lnr_type, false, false);
2384                 if (rc > 0)
2385                         repaired = true;
2386                 break;
2387         default:
2388                 break;
2389         }
2390
2391         if (rc < 0) {
2392                 CDEBUG(D_LFSCK, "%s: namespace LFSCK assistant fail to handle "
2393                        "the shard: "DFID", parent "DFID", name %.*s: rc = %d\n",
2394                        lfsck_lfsck2name(lfsck), PFID(&lnr->lnr_fid),
2395                        PFID(pfid), lnr->lnr_namelen, lnr->lnr_name, rc);
2396
2397                 if ((rc == -ENOTCONN || rc == -ESHUTDOWN || rc == -EREMCHG ||
2398                      rc == -ETIMEDOUT || rc == -EHOSTDOWN ||
2399                      rc == -EHOSTUNREACH || rc == -EINPROGRESS) &&
2400                     dev != NULL && dev != lfsck->li_bottom)
2401                         lfsck_lad_set_bitmap(env, com, shard_idx);
2402
2403                 if (!(lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT))
2404                         rc = 0;
2405         } else {
2406                 if (repaired) {
2407                         ns->ln_items_repaired++;
2408
2409                         switch (type) {
2410                         case LNIT_BAD_TYPE:
2411                                 ns->ln_bad_type_repaired++;
2412                                 break;
2413                         case LNIT_BAD_DIRENT:
2414                                 ns->ln_dirent_repaired++;
2415                                 break;
2416                         default:
2417                                 break;
2418                         }
2419                 }
2420
2421                 rc = 0;
2422         }
2423
2424         if (obj != NULL && !IS_ERR(obj))
2425                 lfsck_object_put(env, obj);
2426
2427         lfsck_object_put(env, dir);
2428
2429         return rc;
2430 }