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, 2017, 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         struct lfsck_thread_info *info = lfsck_env_info(env);
 843         struct lu_buf *buf = &info->lti_buf;
 844         int size = sizeof(*lmv) + sizeof(struct lu_fid) * 2;
 845         struct lmv_foreign_md *lfm;
 846         int rc;
 847
 848         dt_read_lock(env, obj, 0);
 849         buf->lb_buf = lmv;
 850         buf->lb_len = sizeof(*lmv);
 851         rc = dt_xattr_get(env, obj, buf, XATTR_NAME_LMV);
 852         if (unlikely(rc == -ERANGE)) {
 853                 buf = &info->lti_big_buf;
 854                 /* For the in-migration directory, its LMV EA contains
 855                  * not only the LMV header, but also the FIDs for both
 856                  * source and target. So the LMV EA size is larger.
 857                  * Or may be this is a foreign LMV */
 858                 rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_LMV);
 859                 if (rc > sizeof(*lmv)) {
 860                         int rc1;
 861
 862                         lu_buf_check_and_alloc(buf, rc);
 863                         rc1 = dt_xattr_get(env, obj, buf, XATTR_NAME_LMV);
 864                         if (rc != rc1)
 865                                 rc = -EINVAL;
 866                 } else {
 867                         rc = -EINVAL;
 868                 }
 869         }
 870         dt_read_unlock(env, obj);
 871
 872         if (rc > 0 && rc > offsetof(typeof(*lfm), lfm_value) &&
 873             *((__u32 *)buf->lb_buf) == LMV_MAGIC_FOREIGN) {
 874                 __u32 value_len;
 875
 876                 lfm = buf->lb_buf;
 877                 value_len = le32_to_cpu(lfm->lfm_length);
 878                 CDEBUG(D_INFO,
 879                        "foreign LMV EA, magic %x, len %u, type %x, flags %x, for dir "DFID"\n",
 880                        le32_to_cpu(lfm->lfm_magic), value_len,
 881                        le32_to_cpu(lfm->lfm_type), le32_to_cpu(lfm->lfm_flags),
 882                        PFID(lfsck_dto2fid(obj)));
 883
 884                 if (rc != value_len + offsetof(typeof(*lfm), lfm_value))
 885                         CDEBUG(D_LFSCK,
 886                                "foreign LMV EA internal size %u does not match EA full size %d for dir "DFID"\n",
 887                                value_len, rc, PFID(lfsck_dto2fid(obj)));
 888
 889                 /* no further usage/decode of foreign LMV outside */
 890                 return -ENODATA;
 891         }
 892
 893         if (rc == size) {
 894                 rc = sizeof(*lmv);
 895                 memcpy(lmv, buf->lb_buf, rc);
 896         }
 897         if (rc != sizeof(*lmv))
 898                 return rc > 0 ? -EINVAL : rc;
 899
 900         lfsck_lmv_header_le_to_cpu(lmv, lmv);
 901         if ((lmv->lmv_magic == LMV_MAGIC &&
 902              !(lmv->lmv_hash_type & LMV_HASH_FLAG_MIGRATION)) ||
 903             lmv->lmv_magic == LMV_MAGIC_STRIPE)
 904                 return 0;
 905
 906         return -ENODATA;
 907 }
 908
 909 /**
 910  * Parse the shard's index from the given shard name.
 911  *
 912  * The valid shard name/type should be:
 913  * 1) The type must be S_IFDIR
 914  * 2) The name should be $FID:$index
 915  * 3) the index should within valid range.
 916  *
 917  * \param[in] env       pointer to the thread context
 918  * \param[in] name      the shard name
 919  * \param[in] namelen   the name length
 920  * \param[in] type      the entry's type
 921  * \param[in] fid       the entry's FID
 922  *
 923  * \retval              zero or positive number for the index from the name
 924  * \retval              negative error number on failure
 925  */
 926 int lfsck_shard_name_to_index(const struct lu_env *env, const char *name,
 927                               int namelen, __u16 type, const struct lu_fid *fid)
 928 {
 929         char    *name2  = lfsck_env_info(env)->lti_tmpbuf2;
 930         int      len;
 931         int      idx    = 0;
 932
 933         if (!S_ISDIR(type))
 934                 return -ENOTDIR;
 935
 936         LASSERT(name != name2);
 937
 938         len = snprintf(name2, sizeof(lfsck_env_info(env)->lti_tmpbuf2),
 939                        DFID":", PFID(fid));
 940         if (namelen < len + 1 || memcmp(name, name2, len) != 0)
 941                 return -EINVAL;
 942
 943         do {
 944                 if (!isdigit(name[len]))
 945                         return -EINVAL;
 946
 947                 idx = idx * 10 + name[len++] - '0';
 948         } while (len < namelen);
 949
 950         if (idx >= LFSCK_LMV_MAX_STRIPES)
 951                 return -EINVAL;
 952
 953         return idx;
 954 }
 955
 956 bool lfsck_is_valid_slave_name_entry(const struct lu_env *env,
 957                                      struct lfsck_lmv *llmv,
 958                                      const char *name, int namelen)
 959 {
 960         struct lmv_mds_md_v1    *lmv;
 961         int                      idx;
 962
 963         if (llmv == NULL || !llmv->ll_lmv_slave || !llmv->ll_lmv_verified)
 964                 return true;
 965
 966         lmv = &llmv->ll_lmv;
 967         idx = lmv_name_to_stripe_index(lmv->lmv_hash_type,
 968                                        lmv->lmv_stripe_count,
 969                                        name, namelen);
 970         if (unlikely(idx != lmv->lmv_master_mdt_index))
 971                 return false;
 972
 973         return true;
 974 }
 975
 976 /**
 977  * Check whether the given name is a valid entry under the @parent.
 978  *
 979  * If the @parent is a striped directory then the @child should one
 980  * shard of the striped directory, its name should be $FID:$index.
 981  *
 982  * If the @parent is a shard of a striped directory, then the name hash
 983  * should match the MDT, otherwise it is invalid.
 984  *
 985  * \param[in] env       pointer to the thread context
 986  * \param[in] parent    the parent directory
 987  * \param[in] child     the child object to be checked
 988  * \param[in] cname     the name for the @child in the parent directory
 989  *
 990  * \retval              positive number for invalid name entry
 991  * \retval              0 if the name is valid or uncertain
 992  * \retval              negative error number on failure
 993  */
 994 int lfsck_namespace_check_name(const struct lu_env *env,
 995                                struct dt_object *parent,
 996                                struct dt_object *child,
 997                                const struct lu_name *cname)
 998 {
 999         struct lmv_mds_md_v1    *lmv = &lfsck_env_info(env)->lti_lmv;
1000         int                      idx;
1001         int                      rc;
1002
1003         rc = lfsck_read_stripe_lmv(env, parent, lmv);
1004         if (rc != 0)
1005                 RETURN(rc == -ENODATA ? 0 : rc);
1006
1007         if (lmv->lmv_magic == LMV_MAGIC_STRIPE) {
1008                 if (!lfsck_is_valid_slave_lmv(lmv))
1009                         return 0;
1010
1011                 idx = lmv_name_to_stripe_index(lmv->lmv_hash_type,
1012                                                lmv->lmv_stripe_count,
1013                                                cname->ln_name,
1014                                                cname->ln_namelen);
1015                 if (unlikely(idx != lmv->lmv_master_mdt_index))
1016                         return 1;
1017         } else if (lfsck_shard_name_to_index(env, cname->ln_name,
1018                         cname->ln_namelen, lfsck_object_type(child),
1019                         lfsck_dto2fid(child)) < 0) {
1020                 return 1;
1021         }
1022
1023         return 0;
1024 }
1025
1026 /**
1027  * Update the object's LMV EA with the given @lmv.
1028  *
1029  * \param[in] env       pointer to the thread context
1030  * \param[in] com       pointer to the lfsck component
1031  * \param[in] obj       pointer to the object which LMV EA will be updated
1032  * \param[in] lmv       pointer to buffer holding the new LMV EA
1033  * \param[in] locked    whether the caller has held ldlm lock on the @obj or not
1034  *
1035  * \retval              positive number for nothing to be done
1036  * \retval              zero if updated successfully
1037  * \retval              negative error number on failure
1038  */
1039 int lfsck_namespace_update_lmv(const struct lu_env *env,
1040                                struct lfsck_component *com,
1041                                struct dt_object *obj,
1042                                struct lmv_mds_md_v1 *lmv, bool locked)
1043 {
1044         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1045         struct lmv_mds_md_v1            *lmv4   = &info->lti_lmv4;
1046         struct lu_buf                   *buf    = &info->lti_buf;
1047         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1048         struct dt_device                *dev    = lfsck_obj2dev(obj);
1049         struct thandle                  *th     = NULL;
1050         struct lustre_handle             lh     = { 0 };
1051         int                              rc     = 0;
1052         int                              rc1    = 0;
1053         ENTRY;
1054
1055         LASSERT(lmv4 != lmv);
1056
1057         lfsck_lmv_header_cpu_to_le(lmv4, lmv);
1058         lfsck_buf_init(buf, lmv4, sizeof(*lmv4));
1059
1060         if (!locked) {
1061                 rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
1062                                       MDS_INODELOCK_UPDATE |
1063                                       MDS_INODELOCK_XATTR, LCK_EX);
1064                 if (rc != 0)
1065                         GOTO(log, rc);
1066         }
1067
1068         th = dt_trans_create(env, dev);
1069         if (IS_ERR(th))
1070                 GOTO(log, rc = PTR_ERR(th));
1071
1072         /* For remote updating LMV EA, there will be further LFSCK action on
1073          * remote MDT after the updating, so update the LMV EA synchronously. */
1074         if (dt_object_remote(obj))
1075                 th->th_sync = 1;
1076
1077         rc = dt_declare_xattr_set(env, obj, buf, XATTR_NAME_LMV, 0, th);
1078         if (rc != 0)
1079                 GOTO(stop, rc);
1080
1081         rc = dt_trans_start_local(env, dev, th);
1082         if (rc != 0)
1083                 GOTO(stop, rc);
1084
1085         dt_write_lock(env, obj, 0);
1086         if (unlikely(lfsck_is_dead_obj(obj)))
1087                 GOTO(unlock, rc = 1);
1088
1089         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1090                 GOTO(unlock, rc = 0);
1091
1092         rc = dt_xattr_set(env, obj, buf, XATTR_NAME_LMV, 0, th);
1093
1094         GOTO(unlock, rc);
1095
1096 unlock:
1097         dt_write_unlock(env, obj);
1098
1099 stop:
1100         rc1 = dt_trans_stop(env, dev, th);
1101         if (rc == 0)
1102                 rc = rc1;
1103
1104 log:
1105         lfsck_ibits_unlock(&lh, LCK_EX);
1106         CDEBUG(D_LFSCK, "%s: namespace LFSCK updated the %s LMV EA "
1107                "for the object "DFID": rc = %d\n",
1108                lfsck_lfsck2name(lfsck),
1109                lmv->lmv_magic == LMV_MAGIC ? "master" : "slave",
1110                PFID(lfsck_dto2fid(obj)), rc);
1111
1112         return rc;
1113 }
1114
1115 /**
1116  * Check whether allow to re-genereate the lost master LMV EA.
1117  *
1118  * If the master MDT-object of the striped directory lost its master LMV EA,
1119  * then before the LFSCK repaired the striped directory, some ones may have
1120  * created some objects (that are not normal shards of the striped directory)
1121  * under the master MDT-object. If such case happend, then the LFSCK cannot
1122  * re-generate the lost master LMV EA to keep those objects to be visible to
1123  * client.
1124  *
1125  * \param[in] env       pointer to the thread context
1126  * \param[in] com       pointer to the lfsck component
1127  * \param[in] obj       pointer to the master MDT-object to be checked
1128  * \param[in] cfid      the shard's FID used for verification
1129  * \param[in] cidx      the shard's index used for verification
1130  *
1131  * \retval              positive number if not allow to re-generate LMV EA
1132  * \retval              zero if allow to re-generate LMV EA
1133  * \retval              negative error number on failure
1134  */
1135 static int lfsck_allow_regenerate_master_lmv(const struct lu_env *env,
1136                                              struct lfsck_component *com,
1137                                              struct dt_object *obj,
1138                                              const struct lu_fid *cfid,
1139                                              __u32 cidx)
1140 {
1141         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1142         struct lu_fid                   *tfid   = &info->lti_fid3;
1143         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1144         struct lu_dirent                *ent    =
1145                         (struct lu_dirent *)info->lti_key;
1146         const struct dt_it_ops          *iops;
1147         struct dt_it                    *di;
1148         __u64                            cookie;
1149         __u32                            args;
1150         int                              rc;
1151         __u16                            type;
1152         ENTRY;
1153
1154         if (unlikely(!dt_try_as_dir(env, obj)))
1155                 RETURN(-ENOTDIR);
1156
1157         /* Check whether the shard and the master MDT-object matches or not. */
1158         snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf), DFID":%u",
1159                  PFID(cfid), cidx);
1160         rc = dt_lookup(env, obj, (struct dt_rec *)tfid,
1161                        (const struct dt_key *)info->lti_tmpbuf);
1162         if (rc != 0)
1163                 RETURN(rc);
1164
1165         if (!lu_fid_eq(tfid, cfid))
1166                 RETURN(-ENOENT);
1167
1168         args = lfsck->li_args_dir & ~(LUDA_VERIFY | LUDA_VERIFY_DRYRUN);
1169         iops = &obj->do_index_ops->dio_it;
1170         di = iops->init(env, obj, args);
1171         if (IS_ERR(di))
1172                 RETURN(PTR_ERR(di));
1173
1174         rc = iops->load(env, di, 0);
1175         if (rc == 0)
1176                 rc = iops->next(env, di);
1177         else if (rc > 0)
1178                 rc = 0;
1179
1180         if (rc != 0)
1181                 GOTO(out, rc);
1182
1183         do {
1184                 rc = iops->rec(env, di, (struct dt_rec *)ent, args);
1185                 if (rc == 0)
1186                         rc = lfsck_unpack_ent(ent, &cookie, &type);
1187
1188                 if (rc != 0)
1189                         GOTO(out, rc);
1190
1191                 /* skip dot and dotdot entries */
1192                 if (name_is_dot_or_dotdot(ent->lde_name, ent->lde_namelen))
1193                         goto next;
1194
1195                 /* If the subdir name does not match the shard name rule, then
1196                  * it is quite possible that it is NOT a shard, but created by
1197                  * someone after the master MDT-object lost the master LMV EA.
1198                  * But it is also possible that the subdir name entry crashed,
1199                  * under such double failure cases, the LFSCK cannot know how
1200                  * to repair the inconsistency. For data safe, the LFSCK will
1201                  * mark the master MDT-object as read-only. The administrator
1202                  * can fix the bad shard name manually, then run LFSCK again.
1203                  *
1204                  * XXX: If the subdir name matches the shard name rule, but it
1205                  *      is not a real shard of the striped directory, instead,
1206                  *      it was created by someone after the master MDT-object
1207                  *      lost the LMV EA, then re-generating the master LMV EA
1208                  *      will cause such subdir to be invisible to client, and
1209                  *      if its index occupies some lost shard index, then the
1210                  *      LFSCK will use it to replace the bad shard, and cause
1211                  *      the subdir (itself) to be invisible for ever. */
1212                 if (lfsck_shard_name_to_index(env, ent->lde_name,
1213                                 ent->lde_namelen, type, &ent->lde_fid) < 0)
1214                         GOTO(out, rc = 1);
1215
1216 next:
1217                 rc = iops->next(env, di);
1218         } while (rc == 0);
1219
1220         GOTO(out, rc = 0);
1221
1222 out:
1223         iops->put(env, di);
1224         iops->fini(env, di);
1225
1226         return rc;
1227 }
1228
1229 /**
1230  * Notify remote LFSCK instance that the object's LMV EA has been updated.
1231  *
1232  * \param[in] env       pointer to the thread context
1233  * \param[in] com       pointer to the lfsck component
1234  * \param[in] obj       pointer to the object on which the LMV EA will be set
1235  * \param[in] event     indicate either master or slave LMV EA has been updated
1236  * \param[in] flags     indicate which element(s) in the LMV EA has been updated
1237  * \param[in] index     the MDT index on which the LFSCK instance to be notified
1238  *
1239  * \retval              positive number if nothing to be done
1240  * \retval              zero for success
1241  * \retval              negative error number on failure
1242  */
1243 static int lfsck_namespace_notify_lmv_remote(const struct lu_env *env,
1244                                              struct lfsck_component *com,
1245                                              struct dt_object *obj,
1246                                              __u32 event, __u32 flags,
1247                                              __u32 index)
1248 {
1249         struct lfsck_request            *lr     = &lfsck_env_info(env)->lti_lr;
1250         const struct lu_fid             *fid    = lfsck_dto2fid(obj);
1251         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1252         struct lfsck_tgt_desc           *ltd    = NULL;
1253         struct ptlrpc_request           *req    = NULL;
1254         int                              rc;
1255         ENTRY;
1256
1257         ltd = lfsck_tgt_get(&lfsck->li_mdt_descs, index);
1258         if (ltd == NULL)
1259                 GOTO(out, rc = -ENODEV);
1260
1261         req = ptlrpc_request_alloc(class_exp2cliimp(ltd->ltd_exp),
1262                                    &RQF_LFSCK_NOTIFY);
1263         if (req == NULL)
1264                 GOTO(out, rc = -ENOMEM);
1265
1266         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
1267         if (rc != 0) {
1268                 ptlrpc_request_free(req);
1269
1270                 GOTO(out, rc);
1271         }
1272
1273         lr = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
1274         memset(lr, 0, sizeof(*lr));
1275         lr->lr_event = event;
1276         lr->lr_index = lfsck_dev_idx(lfsck);
1277         lr->lr_active = LFSCK_TYPE_NAMESPACE;
1278         lr->lr_fid = *fid;
1279         lr->lr_flags = flags;
1280
1281         ptlrpc_request_set_replen(req);
1282         rc = ptlrpc_queue_wait(req);
1283         ptlrpc_req_finished(req);
1284
1285         GOTO(out, rc = (rc == -ENOENT ? 1 : rc));
1286
1287 out:
1288         CDEBUG(D_LFSCK, "%s: namespace LFSCK notify LMV EA updated for the "
1289                "object "DFID" on MDT %x remotely with event %u, flags %u: "
1290                "rc = %d\n", lfsck_lfsck2name(lfsck), PFID(fid), index,
1291                event, flags, rc);
1292
1293         if (ltd != NULL)
1294                 lfsck_tgt_put(ltd);
1295
1296         return rc;
1297 }
1298
1299 /**
1300  * Generate request for local LFSCK instance to rescan the striped directory.
1301  *
1302  * \param[in] env       pointer to the thread context
1303  * \param[in] com       pointer to the lfsck component
1304  * \param[in] obj       pointer to the striped directory to be rescanned
1305  *
1306  * \retval              positive number if nothing to be done
1307  * \retval              zero for success
1308  * \retval              negative error number on failure
1309  */
1310 int lfsck_namespace_notify_lmv_master_local(const struct lu_env *env,
1311                                             struct lfsck_component *com,
1312                                             struct dt_object *obj)
1313 {
1314         struct lfsck_instance      *lfsck = com->lc_lfsck;
1315         struct lfsck_namespace     *ns    = com->lc_file_ram;
1316         struct lmv_mds_md_v1       *lmv4  = &lfsck_env_info(env)->lti_lmv4;
1317         struct lfsck_lmv_unit      *llu;
1318         struct lfsck_lmv           *llmv;
1319         struct lfsck_slave_lmv_rec *lslr;
1320         int                         count = 0;
1321         int                         rc;
1322         ENTRY;
1323
1324         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1325                 RETURN(0);
1326
1327         rc = lfsck_read_stripe_lmv(env, obj, lmv4);
1328         if (rc != 0)
1329                 RETURN(rc);
1330
1331         OBD_ALLOC_PTR(llu);
1332         if (unlikely(llu == NULL))
1333                 RETURN(-ENOMEM);
1334
1335         if (lmv4->lmv_stripe_count < 1)
1336                 count = LFSCK_LMV_DEF_STRIPES;
1337         else if (lmv4->lmv_stripe_count > LFSCK_LMV_MAX_STRIPES)
1338                 count = LFSCK_LMV_MAX_STRIPES;
1339         else
1340                 count = lmv4->lmv_stripe_count;
1341
1342         OBD_ALLOC_LARGE(lslr, sizeof(struct lfsck_slave_lmv_rec) * count);
1343         if (lslr == NULL) {
1344                 OBD_FREE_PTR(llu);
1345
1346                 RETURN(-ENOMEM);
1347         }
1348
1349         INIT_LIST_HEAD(&llu->llu_link);
1350         llu->llu_lfsck = lfsck;
1351         llu->llu_obj = lfsck_object_get(obj);
1352         llmv = &llu->llu_lmv;
1353         llmv->ll_lmv_master = 1;
1354         llmv->ll_inline = 1;
1355         atomic_set(&llmv->ll_ref, 1);
1356         llmv->ll_stripes_allocated = count;
1357         llmv->ll_hash_type = LMV_HASH_TYPE_UNKNOWN;
1358         llmv->ll_lslr = lslr;
1359         llmv->ll_lmv = *lmv4;
1360
1361         down_write(&com->lc_sem);
1362         if (ns->ln_status != LS_SCANNING_PHASE1 &&
1363             ns->ln_status != LS_SCANNING_PHASE2) {
1364                 ns->ln_striped_dirs_skipped++;
1365                 up_write(&com->lc_sem);
1366                 lfsck_lmv_put(env, llmv);
1367         } else {
1368                 ns->ln_striped_dirs_repaired++;
1369                 llmv->ll_counted = 1;
1370                 spin_lock(&lfsck->li_lock);
1371                 list_add_tail(&llu->llu_link, &lfsck->li_list_lmv);
1372                 spin_unlock(&lfsck->li_lock);
1373                 up_write(&com->lc_sem);
1374         }
1375
1376         RETURN(0);
1377 }
1378
1379 /**
1380  * Set master LMV EA for the specified striped directory.
1381  *
1382  * First, if the master MDT-object of a striped directory lost its LMV EA,
1383  * then there may be some users have created some files under the master
1384  * MDT-object directly. Under such case, the LFSCK cannot re-generate LMV
1385  * EA for the master MDT-object, because we should keep the existing files
1386  * to be visible to client. Then the LFSCK will mark the striped directory
1387  * as read-only and keep it there to be handled by administrator manually.
1388  *
1389  * If nobody has created files under the master MDT-object of the striped
1390  * directory, then we will set the master LMV EA and generate a new rescan
1391  * (the striped directory) request that will be handled later by the LFSCK
1392  * instance on the MDT later.
1393  *
1394  * \param[in] env       pointer to the thread context
1395  * \param[in] com       pointer to the lfsck component
1396  * \param[in] obj       pointer to the object on which the LMV EA will be set
1397  * \param[in] lmv       pointer to the buffer holding the new LMV EA
1398  * \param[in] cfid      the shard's FID used for verification
1399  * \param[in] cidx      the shard's index used for verification
1400  * \param[in] flags     to indicate which element(s) in the LMV EA will be set
1401  *
1402  * \retval              positive number if nothing to be done
1403  * \retval              zero for success
1404  * \retval              negative error number on failure
1405  */
1406 static int lfsck_namespace_set_lmv_master(const struct lu_env *env,
1407                                           struct lfsck_component *com,
1408                                           struct dt_object *obj,
1409                                           struct lmv_mds_md_v1 *lmv,
1410                                           const struct lu_fid *cfid,
1411                                           __u32 cidx, __u32 flags)
1412 {
1413         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1414         struct lmv_mds_md_v1            *lmv3   = &info->lti_lmv3;
1415         struct lu_seq_range             *range  = &info->lti_range;
1416         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1417         struct seq_server_site          *ss     = lfsck_dev_site(lfsck);
1418         struct lustre_handle             lh     = { 0 };
1419         int                              pidx   = -1;
1420         int                              rc     = 0;
1421         ENTRY;
1422
1423         fld_range_set_mdt(range);
1424         rc = fld_server_lookup(env, ss->ss_server_fld,
1425                                fid_seq(lfsck_dto2fid(obj)), range);
1426         if (rc != 0)
1427                 GOTO(log, rc);
1428
1429         pidx = range->lsr_index;
1430         rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
1431                               MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1432                               LCK_EX);
1433         if (rc != 0)
1434                 GOTO(log, rc);
1435
1436         rc = lfsck_read_stripe_lmv(env, obj, lmv3);
1437         if (rc == -ENODATA) {
1438                 if (!(flags & LEF_SET_LMV_ALL))
1439                         GOTO(log, rc);
1440
1441                 *lmv3 = *lmv;
1442         } else if (rc == 0) {
1443                 if (flags & LEF_SET_LMV_ALL)
1444                         GOTO(log, rc = 1);
1445
1446                 if (flags & LEF_SET_LMV_HASH)
1447                         lmv3->lmv_hash_type = lmv->lmv_hash_type;
1448         } else {
1449                 GOTO(log, rc);
1450         }
1451
1452         lmv3->lmv_magic = LMV_MAGIC;
1453         lmv3->lmv_master_mdt_index = pidx;
1454         lmv3->lmv_layout_version++;
1455
1456         if (flags & LEF_SET_LMV_ALL) {
1457                 rc = lfsck_allow_regenerate_master_lmv(env, com, obj,
1458                                                        cfid, cidx);
1459                 if (rc > 0) {
1460                         rc = lfsck_disable_master_lmv(env, com, obj, false);
1461
1462                         GOTO(log, rc = (rc == 0 ? 1 : rc));
1463                 }
1464
1465                 if (rc < 0)
1466                         GOTO(log, rc);
1467
1468                 /* To indicate that the master has ever lost LMV EA. */
1469                 lmv3->lmv_hash_type |= LMV_HASH_FLAG_LOST_LMV;
1470         }
1471
1472         rc = lfsck_namespace_update_lmv(env, com, obj, lmv3, true);
1473         if (rc == 0 && flags & LEF_SET_LMV_ALL) {
1474                 if (dt_object_remote(obj))
1475                         rc = lfsck_namespace_notify_lmv_remote(env, com, obj,
1476                                                 LE_SET_LMV_MASTER, 0, pidx);
1477                 else
1478                         rc = lfsck_namespace_notify_lmv_master_local(env, com,
1479                                                                      obj);
1480         }
1481
1482         GOTO(log, rc);
1483
1484 log:
1485         lfsck_ibits_unlock(&lh, LCK_EX);
1486         CDEBUG(D_LFSCK, "%s: namespace LFSCK set master LMV EA for the object "
1487                DFID" on the %s MDT %d, flags %x: rc = %d\n",
1488                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(obj)),
1489                dt_object_remote(obj) ? "remote" : "local", pidx, flags, rc);
1490
1491         if (rc <= 0) {
1492                 struct lfsck_namespace *ns = com->lc_file_ram;
1493
1494                 ns->ln_flags |= LF_INCONSISTENT;
1495         }
1496
1497         return rc;
1498 }
1499
1500 /**
1501  * Repair the bad name hash.
1502  *
1503  * If the name hash of some name entry under the striped directory does not
1504  * match the shard of the striped directory, then the LFSCK will repair the
1505  * inconsistency. Ideally, the LFSCK should migrate the name entry from the
1506  * current MDT to the right MDT (another one), but before the async commit
1507  * finished, the LFSCK will change the striped directory's hash type as
1508  * LMV_HASH_TYPE_UNKNOWN and mark the lmv flags as LMV_HASH_FLAG_BAD_TYPE.
1509  *
1510  * \param[in] env       pointer to the thread context
1511  * \param[in] com       pointer to the lfsck component
1512  * \param[in] shard     pointer to the shard of the striped directory that
1513  *                      contains the bad name entry
1514  * \param[in] llmv      pointer to lfsck LMV EA structure
1515  * \param[in] name      the name of the bad name hash
1516  *
1517  * \retval              positive number if nothing to be done
1518  * \retval              zero for success
1519  * \retval              negative error number on failure
1520  */
1521 int lfsck_namespace_repair_bad_name_hash(const struct lu_env *env,
1522                                          struct lfsck_component *com,
1523                                          struct dt_object *shard,
1524                                          struct lfsck_lmv *llmv,
1525                                          const char *name)
1526 {
1527         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1528         struct lu_fid                   *pfid   = &info->lti_fid3;
1529         struct lmv_mds_md_v1            *lmv2   = &info->lti_lmv2;
1530         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1531         struct dt_object                *parent = NULL;
1532         int                              rc     = 0;
1533         ENTRY;
1534
1535         rc = dt_lookup(env, shard, (struct dt_rec *)pfid,
1536                        (const struct dt_key *)dotdot);
1537         if (rc != 0 || !fid_is_sane(pfid))
1538                 GOTO(log, rc);
1539
1540         parent = lfsck_object_find_bottom(env, lfsck, pfid);
1541         if (IS_ERR(parent))
1542                 GOTO(log, rc = PTR_ERR(parent));
1543
1544         if (unlikely(!dt_object_exists(parent)))
1545                 /* The parent object was previously accessed when verifying
1546                  * the slave LMV EA.  If this condition is true it is because
1547                  * the striped directory is being removed. */
1548                 GOTO(log, rc = 1);
1549
1550         *lmv2 = llmv->ll_lmv;
1551         lmv2->lmv_hash_type = LMV_HASH_TYPE_UNKNOWN | LMV_HASH_FLAG_BAD_TYPE;
1552         rc = lfsck_namespace_set_lmv_master(env, com, parent, lmv2,
1553                                             lfsck_dto2fid(shard),
1554                                             llmv->ll_lmv.lmv_master_mdt_index,
1555                                             LEF_SET_LMV_HASH);
1556
1557         GOTO(log, rc);
1558
1559 log:
1560         CDEBUG(D_LFSCK, "%s: namespace LFSCK assistant found bad name hash "
1561                "on the MDT %x, parent "DFID", name %s, shard_%x "DFID
1562                ": rc = %d\n",
1563                lfsck_lfsck2name(lfsck), lfsck_dev_idx(lfsck),
1564                PFID(pfid), name, llmv->ll_lmv.lmv_master_mdt_index,
1565                PFID(lfsck_dto2fid(shard)), rc);
1566
1567         if (parent != NULL && !IS_ERR(parent))
1568                 lfsck_object_put(env, parent);
1569
1570         return rc;
1571 }
1572
1573 /**
1574  * Scan the shard of a striped directory for name hash verification.
1575  *
1576  * During the first-stage scanning, if the LFSCK cannot make sure whether
1577  * the shard of a stripe directory contains valid slave LMV EA or not, then
1578  * it will skip the name hash verification for this shard temporarily, and
1579  * record the shard's FID in the LFSCK tracing file. As the LFSCK processing,
1580  * the slave LMV EA may has been verified/fixed by LFSCK instance on master.
1581  * Then in the second-stage scanning, the shard will be re-scanned, and for
1582  * every name entry under the shard, the name hash will be verified, and for
1583  * unmatched name entry, the LFSCK will try to fix it.
1584  *
1585  * \param[in] env       pointer to the thread context
1586  * \param[in] com       pointer to the lfsck component
1587  * \param[in] child     pointer to the directory object to be handled
1588  *
1589  * \retval              positive number for scanning successfully
1590  * \retval              zero for the scanning is paused
1591  * \retval              negative error number on failure
1592  */
1593 int lfsck_namespace_scan_shard(const struct lu_env *env,
1594                                struct lfsck_component *com,
1595                                struct dt_object *child)
1596 {
1597         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1598         struct lmv_mds_md_v1            *lmv    = &info->lti_lmv;
1599         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1600         struct lfsck_namespace          *ns     = com->lc_file_ram;
1601         struct ptlrpc_thread            *thread = &lfsck->li_thread;
1602         struct lu_dirent                *ent    =
1603                         (struct lu_dirent *)info->lti_key;
1604         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
1605         struct lfsck_lmv                *llmv   = NULL;
1606         const struct dt_it_ops          *iops;
1607         struct dt_it                    *di;
1608         __u64                            cookie;
1609         __u32                            args;
1610         int                              rc;
1611         __u16                            type;
1612         ENTRY;
1613
1614         rc = lfsck_read_stripe_lmv(env, child, lmv);
1615         if (rc != 0)
1616                 RETURN(rc == -ENODATA ? 1 : rc);
1617
1618         if (lmv->lmv_magic != LMV_MAGIC_STRIPE)
1619                 RETURN(1);
1620
1621         if (unlikely(!dt_try_as_dir(env, child)))
1622                 RETURN(-ENOTDIR);
1623
1624         OBD_ALLOC_PTR(llmv);
1625         if (llmv == NULL)
1626                 RETURN(-ENOMEM);
1627
1628         llmv->ll_lmv_slave = 1;
1629         llmv->ll_lmv_verified = 1;
1630         llmv->ll_lmv = *lmv;
1631         atomic_set(&llmv->ll_ref, 1);
1632
1633         args = lfsck->li_args_dir & ~(LUDA_VERIFY | LUDA_VERIFY_DRYRUN);
1634         iops = &child->do_index_ops->dio_it;
1635         di = iops->init(env, child, args);
1636         if (IS_ERR(di))
1637                 GOTO(out, rc = PTR_ERR(di));
1638
1639         rc = iops->load(env, di, 0);
1640         if (rc == 0)
1641                 rc = iops->next(env, di);
1642         else if (rc > 0)
1643                 rc = 0;
1644
1645         while (rc == 0) {
1646                 if (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val) &&
1647                     unlikely(!thread_is_running(thread)))
1648                         GOTO(out, rc = 0);
1649
1650                 rc = iops->rec(env, di, (struct dt_rec *)ent, args);
1651                 if (rc == 0)
1652                         rc = lfsck_unpack_ent(ent, &cookie, &type);
1653
1654                 if (rc != 0) {
1655                         if (bk->lb_param & LPF_FAILOUT)
1656                                 GOTO(out, rc);
1657
1658                         goto next;
1659                 }
1660
1661                 /* skip dot and dotdot entries */
1662                 if (name_is_dot_or_dotdot(ent->lde_name, ent->lde_namelen))
1663                         goto next;
1664
1665                 if (!lfsck_is_valid_slave_name_entry(env, llmv, ent->lde_name,
1666                                                      ent->lde_namelen)) {
1667                         ns->ln_flags |= LF_INCONSISTENT;
1668                         rc = lfsck_namespace_repair_bad_name_hash(env, com,
1669                                                 child, llmv, ent->lde_name);
1670                         if (rc == 0)
1671                                 ns->ln_name_hash_repaired++;
1672                 }
1673
1674                 if (rc < 0 && bk->lb_param & LPF_FAILOUT)
1675                         GOTO(out, rc);
1676
1677                 /* Rate control. */
1678                 lfsck_control_speed(lfsck);
1679                 if (unlikely(!thread_is_running(thread)))
1680                         GOTO(out, rc = 0);
1681
1682                 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_FATAL2)) {
1683                         spin_lock(&lfsck->li_lock);
1684                         thread_set_flags(thread, SVC_STOPPING);
1685                         spin_unlock(&lfsck->li_lock);
1686
1687                         GOTO(out, rc = -EINVAL);
1688                 }
1689
1690 next:
1691                 rc = iops->next(env, di);
1692         }
1693
1694         GOTO(out, rc);
1695
1696 out:
1697         iops->put(env, di);
1698         iops->fini(env, di);
1699         lfsck_lmv_put(env, llmv);
1700
1701         return rc;
1702 }
1703
1704 /**
1705  * Verify the slave object's (of striped directory) LMV EA.
1706  *
1707  * For the slave object of a striped directory, before traversing the shard
1708  * the LFSCK will verify whether its slave LMV EA matches its parent's master
1709  * LMV EA or not.
1710  *
1711  * \param[in] env       pointer to the thread context
1712  * \param[in] com       pointer to the lfsck component
1713  * \param[in] obj       pointer to the object which LMV EA will be checked
1714  * \param[in] llmv      pointer to buffer holding the slave LMV EA
1715  *
1716  * \retval              positive number if nothing to be done
1717  * \retval              zero for success
1718  * \retval              negative error number on failure
1719  */
1720 int lfsck_namespace_verify_stripe_slave(const struct lu_env *env,
1721                                         struct lfsck_component *com,
1722                                         struct dt_object *obj,
1723                                         struct lfsck_lmv *llmv)
1724 {
1725         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1726         char                            *name   = info->lti_key;
1727         char                            *name2;
1728         struct lu_fid                   *pfid   = &info->lti_fid3;
1729         const struct lu_fid             *cfid   = lfsck_dto2fid(obj);
1730         struct lu_fid                    tfid;
1731         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1732         struct lmv_mds_md_v1            *clmv   = &llmv->ll_lmv;
1733         struct lmv_mds_md_v1            *plmv   = &info->lti_lmv;
1734         struct dt_object                *parent = NULL;
1735         int                              rc     = 0;
1736         ENTRY;
1737
1738         if (!lfsck_is_valid_slave_lmv(clmv)) {
1739                 rc = lfsck_namespace_trace_update(env, com, cfid,
1740                                         LNTF_UNCERTAIN_LMV, true);
1741
1742                 GOTO(out, rc);
1743         }
1744
1745         rc = dt_lookup(env, obj, (struct dt_rec *)pfid,
1746                        (const struct dt_key *)dotdot);
1747         if (rc != 0 || !fid_is_sane(pfid)) {
1748                 rc = lfsck_namespace_trace_update(env, com, cfid,
1749                                         LNTF_UNCERTAIN_LMV, true);
1750
1751                 GOTO(out, rc);
1752         }
1753
1754         CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_ENGINE_DELAY, cfs_fail_val);
1755
1756         parent = lfsck_object_find_bottom(env, lfsck, pfid);
1757         if (IS_ERR(parent)) {
1758                 rc = lfsck_namespace_trace_update(env, com, cfid,
1759                                         LNTF_UNCERTAIN_LMV, true);
1760
1761                 GOTO(out, rc);
1762         }
1763
1764         if (unlikely(!dt_object_exists(parent)))
1765                 GOTO(out, rc = 1);
1766
1767         if (unlikely(!dt_try_as_dir(env, parent)))
1768                 GOTO(out, rc = -ENOTDIR);
1769
1770         rc = lfsck_read_stripe_lmv(env, parent, plmv);
1771         if (rc != 0) {
1772                 int rc1;
1773
1774                 /* If the parent has no LMV EA, then it maybe because:
1775                  * 1) The parent lost the LMV EA.
1776                  * 2) The child claims a wrong (slave) LMV EA. */
1777                 if (rc == -ENODATA)
1778                         rc = lfsck_namespace_set_lmv_master(env, com, parent,
1779                                         clmv, cfid, clmv->lmv_master_mdt_index,
1780                                         LEF_SET_LMV_ALL);
1781                 else
1782                         rc = 0;
1783
1784                 rc1 = lfsck_namespace_trace_update(env, com, cfid,
1785                                                    LNTF_UNCERTAIN_LMV, true);
1786
1787                 GOTO(out, rc = (rc < 0 ? rc : rc1));
1788         }
1789
1790         /* Unmatched magic or stripe count. */
1791         if (unlikely(plmv->lmv_magic != LMV_MAGIC ||
1792                      plmv->lmv_stripe_count != clmv->lmv_stripe_count)) {
1793                 rc = lfsck_namespace_trace_update(env, com, cfid,
1794                                                   LNTF_UNCERTAIN_LMV, true);
1795
1796                 GOTO(out, rc);
1797         }
1798
1799         /* If the master hash type has been set as LMV_HASH_TYPE_UNKNOWN,
1800          * then the slave hash type is not important. */
1801         if ((plmv->lmv_hash_type & LMV_HASH_TYPE_MASK) ==
1802             LMV_HASH_TYPE_UNKNOWN &&
1803             plmv->lmv_hash_type & LMV_HASH_FLAG_BAD_TYPE)
1804                 GOTO(out, rc = 0);
1805
1806         /* Unmatched hash type. */
1807         if (unlikely((plmv->lmv_hash_type & LMV_HASH_TYPE_MASK) !=
1808                      (clmv->lmv_hash_type & LMV_HASH_TYPE_MASK))) {
1809                 rc = lfsck_namespace_trace_update(env, com, cfid,
1810                                                   LNTF_UNCERTAIN_LMV, true);
1811
1812                 GOTO(out, rc);
1813         }
1814
1815         snprintf(info->lti_tmpbuf2, sizeof(info->lti_tmpbuf2), DFID":%u",
1816                  PFID(cfid), clmv->lmv_master_mdt_index);
1817         name2 = info->lti_tmpbuf2;
1818
1819         rc = lfsck_links_get_first(env, obj, name, &tfid);
1820         if (rc == 0 && strcmp(name, name2) == 0 && lu_fid_eq(pfid, &tfid)) {
1821                 llmv->ll_lmv_verified = 1;
1822
1823                 GOTO(out, rc);
1824         }
1825
1826         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
1827                        (const struct dt_key *)name2);
1828         if (rc != 0 || !lu_fid_eq(cfid, &tfid))
1829                 rc = lfsck_namespace_trace_update(env, com, cfid,
1830                                                   LNTF_UNCERTAIN_LMV, true);
1831         else
1832                 llmv->ll_lmv_verified = 1;
1833
1834         GOTO(out, rc);
1835
1836 out:
1837         if (parent != NULL && !IS_ERR(parent))
1838                 lfsck_object_put(env, parent);
1839
1840         return rc;
1841 }
1842
1843 /**
1844  * Double scan the striped directory or the shard.
1845  *
1846  * All the shards' under the given striped directory or its shard have
1847  * been scanned, the LFSCK has got the global knownledge about the LMV
1848  * EA consistency.
1849  *
1850  * If the target is one shard of a striped directory, then only needs to
1851  * update related tracing file.
1852  *
1853  * If the target is the master MDT-object of a striped directory, then the
1854  * LFSCK will make the decision about whether the master LMV EA is invalid
1855  * or not, and repair it if inconsistenct; for every shard of the striped
1856  * directory, whether the slave LMV EA is invalid or not, and repair it if
1857  * inconsistent.
1858  *
1859  * \param[in] env       pointer to the thread context
1860  * \param[in] com       pointer to the lfsck component
1861  * \param[in] lnr       pointer to the namespace request that contains the
1862  *                      striped directory or the shard
1863  *
1864  * \retval              zero for success
1865  * \retval              negative error number on failure
1866  */
1867 int lfsck_namespace_striped_dir_rescan(const struct lu_env *env,
1868                                        struct lfsck_component *com,
1869                                        struct lfsck_namespace_req *lnr)
1870 {
1871         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1872         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1873         struct lfsck_namespace          *ns     = com->lc_file_ram;
1874         struct lfsck_lmv                *llmv   = lnr->lnr_lmv;
1875         struct lmv_mds_md_v1            *lmv    = &llmv->ll_lmv;
1876         struct lmv_mds_md_v1            *lmv2   = &info->lti_lmv2;
1877         struct lfsck_assistant_object   *lso    = lnr->lnr_lar.lar_parent;
1878         const struct lu_fid             *pfid   = &lso->lso_fid;
1879         struct dt_object                *dir    = NULL;
1880         struct dt_object                *obj    = NULL;
1881         struct lu_seq_range             *range  = &info->lti_range;
1882         struct seq_server_site          *ss     = lfsck_dev_site(lfsck);
1883         __u32                            stripe_count;
1884         __u32                            hash_type;
1885         int                              rc     = 0;
1886         int                              i;
1887         ENTRY;
1888
1889         if (llmv->ll_lmv_slave) {
1890                 if (llmv->ll_lmv_verified) {
1891                         ns->ln_striped_shards_scanned++;
1892                         lfsck_namespace_trace_update(env, com, pfid,
1893                                         LNTF_UNCERTAIN_LMV |
1894                                         LNTF_RECHECK_NAME_HASH, false);
1895                 }
1896
1897                 RETURN(0);
1898         }
1899
1900         /* Either the striped directory has been disabled or only part of
1901          * the striped directory have been scanned. The LFSCK cannot repair
1902          * something based on incompleted knowledge. So skip it. */
1903         if (llmv->ll_ignore || llmv->ll_exit_value <= 0)
1904                 RETURN(0);
1905
1906         /* There ever been some failure, as to the LFSCK cannot know whether
1907          * it has got the global knowledge about the LMV EA consistency or not,
1908          * so it cannot make reparation about the incompleted knowledge. */
1909         if (llmv->ll_failed) {
1910                 ns->ln_striped_dirs_scanned++;
1911                 ns->ln_striped_dirs_failed++;
1912
1913                 RETURN(0);
1914         }
1915
1916         if (lmv->lmv_stripe_count > LFSCK_LMV_MAX_STRIPES)
1917                 stripe_count = max(llmv->ll_max_filled_off + 1,
1918                                    llmv->ll_max_stripe_count);
1919         else
1920                 stripe_count = max(llmv->ll_max_filled_off + 1,
1921                                    lmv->lmv_stripe_count);
1922
1923         if (lmv->lmv_stripe_count != stripe_count) {
1924                 lmv->lmv_stripe_count = stripe_count;
1925                 llmv->ll_lmv_updated = 1;
1926         }
1927
1928         if (!lmv_is_known_hash_type(lmv->lmv_hash_type) &&
1929             !(lmv->lmv_hash_type & LMV_HASH_FLAG_BAD_TYPE) &&
1930             lmv_is_known_hash_type(llmv->ll_hash_type)) {
1931                 hash_type = llmv->ll_hash_type & LMV_HASH_TYPE_MASK;
1932                 lmv->lmv_hash_type = llmv->ll_hash_type;
1933                 llmv->ll_lmv_updated = 1;
1934         } else {
1935                 hash_type = lmv->lmv_hash_type & LMV_HASH_TYPE_MASK;
1936                 if (!lmv_is_known_hash_type(hash_type))
1937                         hash_type = LMV_HASH_TYPE_UNKNOWN;
1938         }
1939
1940         if (llmv->ll_lmv_updated) {
1941                 if (dir == NULL) {
1942                         dir = lfsck_assistant_object_load(env, lfsck, lso);
1943                         if (IS_ERR(dir)) {
1944                                 rc = PTR_ERR(dir);
1945
1946                                 RETURN(rc == -ENOENT ? 0 : rc);
1947                         }
1948                 }
1949
1950                 lmv->lmv_layout_version++;
1951                 rc = lfsck_namespace_update_lmv(env, com, dir, lmv, false);
1952                 if (rc != 0)
1953                         RETURN(rc);
1954
1955                 ns->ln_striped_dirs_scanned++;
1956                 if (!llmv->ll_counted)
1957                         ns->ln_striped_dirs_repaired++;
1958         }
1959
1960         fld_range_set_mdt(range);
1961         for (i = 0; i <= llmv->ll_max_filled_off; i++) {
1962                 struct lfsck_slave_lmv_rec *lslr = llmv->ll_lslr + i;
1963                 const struct lu_fid *cfid = &lslr->lslr_fid;
1964                 const struct lu_name *cname;
1965                 struct linkea_data ldata = { NULL };
1966                 int rc1 = 0;
1967                 bool repair_linkea = false;
1968                 bool repair_lmvea = false;
1969                 bool rename = false;
1970                 bool create = false;
1971                 bool linkea_repaired = false;
1972                 bool lmvea_repaired = false;
1973                 bool rename_repaired = false;
1974                 bool create_repaired = false;
1975
1976                 /* LMV EA hole. */
1977                 if (fid_is_zero(cfid))
1978                         continue;
1979
1980                 lnr->lnr_fid = *cfid;
1981                 lnr->lnr_namelen = snprintf(lnr->lnr_name,
1982                                             lnr->lnr_size - sizeof(*lnr),
1983                                             DFID":%u", PFID(cfid), i);
1984                 cname = lfsck_name_get_const(env, lnr->lnr_name,
1985                                              lnr->lnr_namelen);
1986                 obj = lfsck_object_find_bottom(env, lfsck, cfid);
1987                 if (IS_ERR(obj)) {
1988                         if (dir == NULL) {
1989                                 dir = lfsck_assistant_object_load(env, lfsck,
1990                                                                   lso);
1991                                 if (IS_ERR(dir)) {
1992                                         if (PTR_ERR(dir) == -ENOENT)
1993                                                 RETURN(0);
1994
1995                                         dir = NULL;
1996                                 }
1997                         } else if (lfsck_is_dead_obj(dir)) {
1998                                 GOTO(out, rc = 0);
1999                         }
2000
2001                         rc1 = PTR_ERR(obj);
2002                         goto next;
2003                 }
2004
2005                 switch (lslr->lslr_flags) {
2006                 case LSLF_NONE:
2007                         if (llmv->ll_inline ||
2008                             lslr->lslr_stripe_count != stripe_count ||
2009                             (lslr->lslr_hash_type & LMV_HASH_TYPE_MASK) !=
2010                              hash_type)
2011                                 repair_lmvea = true;
2012                         break;
2013                 case LSLF_BAD_INDEX2:
2014                         /* The index in the slave LMV EA is right,
2015                          * the name entry should be updated. */
2016                         rename = true;
2017                         snprintf(info->lti_tmpbuf2, sizeof(info->lti_tmpbuf2),
2018                                  DFID":%u", PFID(cfid), lslr->lslr_index);
2019                         if (llmv->ll_inline ||
2020                             lslr->lslr_stripe_count != stripe_count ||
2021                             (lslr->lslr_hash_type & LMV_HASH_TYPE_MASK) !=
2022                              hash_type)
2023                                 repair_lmvea = true;
2024                         break;
2025                 case LSLF_BAD_INDEX1:
2026                         /* The index in the name entry is right,
2027                          * the slave LMV EA should be updated. */
2028                 case LSLF_NO_LMVEA:
2029                         repair_lmvea = true;
2030                         break;
2031                 case LSLF_DANGLING:
2032                         create = true;
2033                         goto repair;
2034                 default:
2035                         break;
2036                 }
2037
2038                 rc1 = lfsck_links_read_with_rec(env, obj, &ldata);
2039                 if (rc1 == -ENOENT) {
2040                         create = true;
2041                         goto repair;
2042                 }
2043
2044                 if (rc1 == -EINVAL || rc1 == -ENODATA) {
2045                         repair_linkea = true;
2046                         goto repair;
2047                 }
2048
2049                 if (rc1 != 0)
2050                         goto next;
2051
2052                 if (ldata.ld_leh->leh_reccount != 1) {
2053                         repair_linkea = true;
2054                         goto repair;
2055                 }
2056
2057                 rc1 = linkea_links_find(&ldata, cname, pfid);
2058                 if (rc1 != 0)
2059                         repair_linkea = true;
2060
2061 repair:
2062                 if (create) {
2063                         if (dir == NULL) {
2064                                 dir = lfsck_assistant_object_load(env, lfsck,
2065                                                                   lso);
2066                                 if (IS_ERR(dir)) {
2067                                         rc1 = PTR_ERR(dir);
2068
2069                                         if (rc1 == -ENOENT)
2070                                                 GOTO(out, rc = 0);
2071
2072                                         dir = NULL;
2073                                         goto next;
2074                                 }
2075                         }
2076
2077                         rc1 = lfsck_namespace_repair_dangling(env, com, dir,
2078                                                               obj, lnr);
2079                         if (rc1 >= 0) {
2080                                 create_repaired = true;
2081                                 if (rc == 0)
2082                                         ns->ln_dangling_repaired++;
2083                         }
2084                 }
2085
2086                 if (repair_lmvea) {
2087                         *lmv2 = *lmv;
2088                         lmv2->lmv_magic = LMV_MAGIC_STRIPE;
2089                         lmv2->lmv_stripe_count = stripe_count;
2090                         lmv2->lmv_master_mdt_index = i;
2091                         lmv2->lmv_hash_type = hash_type;
2092
2093                         rc1 = lfsck_namespace_update_lmv(env, com, obj,
2094                                                          lmv2, false);
2095                         if (rc1 < 0)
2096                                 goto next;
2097
2098                         if (dt_object_remote(obj)) {
2099                                 rc1 = fld_server_lookup(env, ss->ss_server_fld,
2100                                         fid_seq(lfsck_dto2fid(obj)), range);
2101                                 if (rc1 != 0)
2102                                         goto next;
2103
2104                                 rc1 = lfsck_namespace_notify_lmv_remote(env,
2105                                                 com, obj, LE_SET_LMV_SLAVE, 0,
2106                                                 range->lsr_index);
2107                         } else {
2108                                 ns->ln_striped_shards_repaired++;
2109                                 rc1 = lfsck_namespace_trace_update(env, com,
2110                                         cfid, LNTF_RECHECK_NAME_HASH, true);
2111                         }
2112
2113                         if (rc1 < 0)
2114                                 goto next;
2115
2116                         if (rc1 >= 0)
2117                                 lmvea_repaired = true;
2118                 } else if (llmv->ll_inline) {
2119                         if (dt_object_remote(obj)) {
2120                                 rc1 = fld_server_lookup(env, ss->ss_server_fld,
2121                                         fid_seq(lfsck_dto2fid(obj)), range);
2122                                 if (rc1 != 0)
2123                                         goto next;
2124
2125                                 /* The slave LMV EA on the remote shard is
2126                                  * correct, just notify the LFSCK instance
2127                                  * on such MDT to re-verify the name_hash. */
2128                                 rc1 = lfsck_namespace_notify_lmv_remote(env,
2129                                                 com, obj, LE_SET_LMV_SLAVE,
2130                                                 LEF_RECHECK_NAME_HASH,
2131                                                 range->lsr_index);
2132                         } else {
2133                                 rc1 = lfsck_namespace_trace_update(env, com,
2134                                         cfid, LNTF_RECHECK_NAME_HASH, true);
2135                         }
2136
2137                         if (rc1 < 0)
2138                                 goto next;
2139                 }
2140
2141                 if (rename) {
2142                         if (dir == NULL) {
2143                                 dir = lfsck_assistant_object_load(env, lfsck,
2144                                                                   lso);
2145                                 if (IS_ERR(dir)) {
2146                                         rc1 = PTR_ERR(dir);
2147
2148                                         if (rc1 == -ENOENT)
2149                                                 GOTO(out, rc = 0);
2150
2151                                         dir = NULL;
2152                                         goto next;
2153                                 }
2154                         }
2155
2156                         rc1 = lfsck_namespace_repair_dirent(env, com, dir, obj,
2157                                         info->lti_tmpbuf2, lnr->lnr_name,
2158                                         lnr->lnr_type, true, false);
2159                         if (rc1 >= 0) {
2160                                 rename_repaired = true;
2161                                 if (rc1 > 0) {
2162                                         ns->ln_dirent_repaired++;
2163                                         rc1 = lfsck_namespace_trace_update(env,
2164                                                 com, cfid,
2165                                                 LNTF_RECHECK_NAME_HASH, true);
2166                                 }
2167                         }
2168
2169                         if (rc1 < 0)
2170                                 goto next;
2171                 }
2172
2173                 if (repair_linkea) {
2174                         struct lustre_handle lh = { 0 };
2175
2176                         rc1 = linkea_links_new(&ldata, &info->lti_big_buf,
2177                                                cname, lfsck_dto2fid(dir));
2178                         if (rc1 != 0)
2179                                 goto next;
2180
2181                         if (dir == NULL) {
2182                                 dir = lfsck_assistant_object_load(env, lfsck,
2183                                                                   lso);
2184                                 if (IS_ERR(dir)) {
2185                                         rc1 = PTR_ERR(dir);
2186
2187                                         if (rc1 == -ENOENT)
2188                                                 GOTO(out, rc = 0);
2189
2190                                         dir = NULL;
2191                                         goto next;
2192                                 }
2193                         }
2194
2195                         rc1 = lfsck_ibits_lock(env, lfsck, obj, &lh,
2196                                                MDS_INODELOCK_UPDATE |
2197                                                MDS_INODELOCK_XATTR, LCK_EX);
2198                         if (rc1 != 0)
2199                                 goto next;
2200
2201                         rc1 = lfsck_namespace_rebuild_linkea(env, com, obj,
2202                                                              &ldata);
2203                         lfsck_ibits_unlock(&lh, LCK_EX);
2204                         if (rc1 >= 0) {
2205                                 linkea_repaired = true;
2206                                 if (rc1 > 0)
2207                                         ns->ln_linkea_repaired++;
2208                         }
2209                 }
2210
2211 next:
2212                 CDEBUG(D_LFSCK, "%s: namespace LFSCK repair the shard "
2213                       "%d "DFID" of the striped directory "DFID" with "
2214                       "dangling %s/%s, rename %s/%s, llinkea %s/%s, "
2215                       "repair_lmvea %s/%s: rc = %d\n", lfsck_lfsck2name(lfsck),
2216                       i, PFID(cfid), PFID(pfid),
2217                       create ? "yes" : "no", create_repaired ? "yes" : "no",
2218                       rename ? "yes" : "no", rename_repaired ? "yes" : "no",
2219                       repair_linkea ? "yes" : "no",
2220                       linkea_repaired ? "yes" : "no",
2221                       repair_lmvea ? "yes" : "no",
2222                       lmvea_repaired ? "yes" : "no", rc1);
2223
2224                 if (obj != NULL && !IS_ERR(obj)) {
2225                         lfsck_object_put(env, obj);
2226                         obj = NULL;
2227                 }
2228
2229                 if (rc1 < 0) {
2230                         rc = rc1;
2231                         ns->ln_striped_shards_failed++;
2232                 }
2233         }
2234
2235         GOTO(out, rc);
2236
2237 out:
2238         if (obj != NULL && !IS_ERR(obj))
2239                 lfsck_object_put(env, obj);
2240
2241         if (dir != NULL && !IS_ERR(dir))
2242                 lfsck_object_put(env, dir);
2243
2244         return rc;
2245 }
2246
2247 /**
2248  * Verify the shard's name entry under the striped directory.
2249  *
2250  * Before all shards of the striped directory scanned, the LFSCK cannot
2251  * know whether the master LMV EA is valid or not, and also cannot know
2252  * how to repair an invalid shard exactly. For example, the stripe index
2253  * stored in the shard's name does not match the stripe index stored in
2254  * the slave LMV EA, then the LFSCK cannot know which one is correct.
2255  * If the LFSCK just assumed one is correct, and fixed the other, then
2256  * as the LFSCK processing, it may find that the former reparation is
2257  * wrong and have to roll back. Unfortunately, if some applications saw
2258  * the changes and made further modification based on such changes, then
2259  * the roll back is almost impossible.
2260  *
2261  * To avoid above trouble, the LFSCK will scan the master object of the
2262  * striped directory twice, that is NOT the same as normal two-stages
2263  * scanning, the double scanning the striped directory will happen both
2264  * during the first-stage scanning:
2265  *
2266  * 1) When the striped directory is opened for scanning, the LFSCK will
2267  *    iterate each shard in turn, and records its slave LMV EA in the
2268  *    lfsck_lmv::ll_lslr. In this step, if the 'shard' (may be fake
2269  *    shard) name does not match the shard naming rule, for example, it
2270  *    does not contains the shard's FID, or not contains index, then we
2271  *    can remove the bad name entry directly. But if the name is valid,
2272  *    but the shard has no slave LMV EA or the slave LMV EA does not
2273  *    match its name, then we just record related information in the
2274  *    lfsck_lmv::ll_lslr in RAM.
2275  *
2276  * 2) When all the known shards have been scanned, then the engine will
2277  *    generate a dummy request (via lfsck_namespace_close_dir) to tell
2278  *    the assistant thread that all the known shards have been scanned.
2279  *    Since the assistant has got the global knowledge about the index
2280  *    conflict, stripe count, hash type, and so on. Then the assistant
2281  *    thread will scan the lfsck_lmv::ll_lslr, and for every shard in
2282  *    the record, check and repair inconsistency.
2283  *
2284  * Generally, the stripe directory has only several shards, and there
2285  * will NOT be a lof of striped directory. So double scanning striped
2286  * directory will not much affect the LFSCK performance.
2287  *
2288  * \param[in] env       pointer to the thread context
2289  * \param[in] com       pointer to the lfsck component
2290  * \param[in] lnr       pointer to the namespace request that contains the
2291  *                      shard's name, parent object, parent's LMV, and ect.
2292  *
2293  * \retval              zero for success
2294  * \retval              negative error number on failure
2295  */
2296 int lfsck_namespace_handle_striped_master(const struct lu_env *env,
2297                                           struct lfsck_component *com,
2298                                           struct lfsck_namespace_req *lnr)
2299 {
2300         struct lfsck_thread_info   *info        = lfsck_env_info(env);
2301         struct lmv_mds_md_v1       *lmv         = &info->lti_lmv;
2302         struct lfsck_instance      *lfsck       = com->lc_lfsck;
2303         struct lfsck_namespace     *ns          = com->lc_file_ram;
2304         struct lfsck_lmv           *llmv        = lnr->lnr_lmv;
2305         struct lfsck_assistant_object *lso      = lnr->lnr_lar.lar_parent;
2306         const struct lu_fid        *pfid        = &lso->lso_fid;
2307         struct dt_object           *dir;
2308         struct dt_object           *obj         = NULL;
2309         struct dt_device           *dev         = NULL;
2310         int                         shard_idx   = 0;
2311         int                         stripe      = 0;
2312         int                         rc          = 0;
2313         int                         depth       = 0;
2314         bool                        repaired    = false;
2315         enum lfsck_namespace_inconsistency_type type = LNIT_NONE;
2316         ENTRY;
2317
2318         if (unlikely(llmv->ll_ignore))
2319                 RETURN(0);
2320
2321         dir = lfsck_assistant_object_load(env, lfsck, lso);
2322         if (IS_ERR(dir)) {
2323                 rc = PTR_ERR(dir);
2324
2325                 RETURN(rc == -ENOENT ? 0 : rc);
2326         }
2327
2328         shard_idx = lfsck_find_mdt_idx_by_fid(env, lfsck, &lnr->lnr_fid);
2329         if (shard_idx < 0)
2330                 GOTO(fail_lmv, rc = shard_idx);
2331
2332         if (shard_idx == lfsck_dev_idx(lfsck)) {
2333                 if (unlikely(strcmp(lnr->lnr_name, dotdot) == 0))
2334                         GOTO(out, rc = 0);
2335
2336                 dev = lfsck->li_bottom;
2337         } else {
2338                 struct lfsck_tgt_desc *ltd;
2339
2340                 /* Usually, some local filesystem consistency verification
2341                  * tools can guarantee the local namespace tree consistenct.
2342                  * So the LFSCK will only verify the remote directory. */
2343                 if (unlikely(strcmp(lnr->lnr_name, dotdot) == 0)) {
2344                         rc = lfsck_namespace_trace_update(env, com, pfid,
2345                                                 LNTF_CHECK_PARENT, true);
2346
2347                         GOTO(out, rc);
2348                 }
2349
2350                 ltd = lfsck_ltd2tgt(&lfsck->li_mdt_descs, shard_idx);
2351                 if (unlikely(ltd == NULL)) {
2352                         CDEBUG(D_LFSCK, "%s: cannot talk with MDT %x which "
2353                                "did not join the namespace LFSCK\n",
2354                                lfsck_lfsck2name(lfsck), shard_idx);
2355                         lfsck_lad_set_bitmap(env, com, shard_idx);
2356
2357                         GOTO(fail_lmv, rc = -ENODEV);
2358                 }
2359
2360                 dev = ltd->ltd_tgt;
2361         }
2362
2363         obj = lfsck_object_find_by_dev(env, dev, &lnr->lnr_fid);
2364         if (IS_ERR(obj)) {
2365                 if (lfsck_is_dead_obj(dir))
2366                         RETURN(0);
2367
2368                 GOTO(fail_lmv, rc = PTR_ERR(obj));
2369         }
2370
2371         if (!dt_object_exists(obj)) {
2372                 stripe = lfsck_shard_name_to_index(env, lnr->lnr_name,
2373                                 lnr->lnr_namelen, lnr->lnr_type, &lnr->lnr_fid);
2374                 if (stripe < 0) {
2375                         type = LNIT_BAD_DIRENT;
2376
2377                         GOTO(out, rc = 0);
2378                 }
2379
2380 dangling:
2381                 rc = lfsck_namespace_check_exist(env, dir, obj, lnr->lnr_name);
2382                 if (rc == 0) {
2383                         memset(lmv, 0, sizeof(*lmv));
2384                         lmv->lmv_magic = LMV_MAGIC;
2385                         rc = lfsck_record_lmv(env, com, dir, lnr, lmv, stripe,
2386                                               LSLF_DANGLING, LSLF_NONE, &depth);
2387                 }
2388
2389                 GOTO(out, rc);
2390         }
2391
2392         stripe = lfsck_shard_name_to_index(env, lnr->lnr_name, lnr->lnr_namelen,
2393                                            lfsck_object_type(obj),
2394                                            &lnr->lnr_fid);
2395         if (stripe < 0) {
2396                 type = LNIT_BAD_DIRENT;
2397
2398                 GOTO(out, rc = 0);
2399         }
2400
2401         rc = lfsck_read_stripe_lmv(env, obj, lmv);
2402         if (unlikely(rc == -ENOENT))
2403                 /* It may happen when the remote object has been removed,
2404                  * but the local MDT does not aware of that. */
2405                 goto dangling;
2406
2407         if (rc == -ENODATA)
2408                 rc = lfsck_record_lmv(env, com, dir, lnr, lmv, stripe,
2409                                       LSLF_NO_LMVEA, LSLF_NONE, &depth);
2410         else if (rc == 0)
2411                 rc = lfsck_record_lmv(env, com, dir, lnr, lmv, stripe,
2412                                       lmv->lmv_master_mdt_index != stripe ?
2413                                       LSLF_BAD_INDEX1 : LSLF_NONE, LSLF_NONE,
2414                                       &depth);
2415
2416         GOTO(out, rc);
2417
2418 fail_lmv:
2419         llmv->ll_failed = 1;
2420
2421 out:
2422         if (rc >= 0 && type == LNIT_NONE && !S_ISDIR(lnr->lnr_type))
2423                 type = LNIT_BAD_TYPE;
2424
2425         switch (type) {
2426         case LNIT_BAD_TYPE:
2427                 rc = lfsck_namespace_repair_dirent(env, com, dir, obj,
2428                                                    lnr->lnr_name, lnr->lnr_name,
2429                                                    lnr->lnr_type, true, false);
2430                 if (rc > 0)
2431                         repaired = true;
2432                 break;
2433         case LNIT_BAD_DIRENT:
2434                 rc = lfsck_namespace_repair_dirent(env, com, dir, obj,
2435                                                    lnr->lnr_name, lnr->lnr_name,
2436                                                    lnr->lnr_type, false, false);
2437                 if (rc > 0)
2438                         repaired = true;
2439                 break;
2440         default:
2441                 break;
2442         }
2443
2444         if (rc < 0) {
2445                 CDEBUG(D_LFSCK, "%s: namespace LFSCK assistant fail to handle "
2446                        "the shard: "DFID", parent "DFID", name %.*s: rc = %d\n",
2447                        lfsck_lfsck2name(lfsck), PFID(&lnr->lnr_fid),
2448                        PFID(pfid), lnr->lnr_namelen, lnr->lnr_name, rc);
2449
2450                 if ((rc == -ENOTCONN || rc == -ESHUTDOWN || rc == -EREMCHG ||
2451                      rc == -ETIMEDOUT || rc == -EHOSTDOWN ||
2452                      rc == -EHOSTUNREACH || rc == -EINPROGRESS) &&
2453                     dev != NULL && dev != lfsck->li_bottom)
2454                         lfsck_lad_set_bitmap(env, com, shard_idx);
2455
2456                 if (!(lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT))
2457                         rc = 0;
2458         } else {
2459                 if (repaired) {
2460                         ns->ln_items_repaired++;
2461
2462                         switch (type) {
2463                         case LNIT_BAD_TYPE:
2464                                 ns->ln_bad_type_repaired++;
2465                                 break;
2466                         case LNIT_BAD_DIRENT:
2467                                 ns->ln_dirent_repaired++;
2468                                 break;
2469                         default:
2470                                 break;
2471                         }
2472                 }
2473
2474                 rc = 0;
2475         }
2476
2477         if (obj != NULL && !IS_ERR(obj))
2478                 lfsck_object_put(env, obj);
2479
2480         lfsck_object_put(env, dir);
2481
2482         return rc;
2483 }