lustre/lfsck/lfsck_namespace.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) 2013, 2014, Intel Corporation.
  24  */
  25 /*
  26  * lustre/lfsck/lfsck_namespace.c
  27  *
  28  * Author: Fan, Yong <fan.yong@intel.com>
  29  */
  30
  31 #define DEBUG_SUBSYSTEM S_LFSCK
  32
  33 #include <lustre/lustre_idl.h>
  34 #include <lu_object.h>
  35 #include <dt_object.h>
  36 #include <md_object.h>
  37 #include <lustre_fid.h>
  38 #include <lustre_lib.h>
  39 #include <lustre_net.h>
  40 #include <lustre/lustre_user.h>
  41
  42 #include "lfsck_internal.h"
  43
  44 #define LFSCK_NAMESPACE_MAGIC_V1        0xA0629D03
  45 #define LFSCK_NAMESPACE_MAGIC_V2        0xA0621A0B
  46
  47 /* For Lustre-2.x (x <= 6), the namespace LFSCK used LFSCK_NAMESPACE_MAGIC_V1
  48  * as the trace file magic. When downgrade to such old release, the old LFSCK
  49  * will not recognize the new LFSCK_NAMESPACE_MAGIC_V2 in the new trace file,
  50  * then it will reset the whole LFSCK, and will not cause start failure. The
  51  * similar case will happen when upgrade from such old release. */
  52 #define LFSCK_NAMESPACE_MAGIC           LFSCK_NAMESPACE_MAGIC_V2
  53
  54 enum lfsck_nameentry_check {
  55         LFSCK_NAMEENTRY_DEAD            = 1, /* The object has been unlinked. */
  56         LFSCK_NAMEENTRY_REMOVED         = 2, /* The entry has been removed. */
  57         LFSCK_NAMEENTRY_RECREATED       = 3, /* The entry has been recreated. */
  58 };
  59
  60 static struct lfsck_namespace_req *
  61 lfsck_namespace_assistant_req_init(struct lfsck_instance *lfsck,
  62                                    struct lu_dirent *ent, __u16 type)
  63 {
  64         struct lfsck_namespace_req *lnr;
  65         int                         size;
  66
  67         size = sizeof(*lnr) + (ent->lde_namelen & ~3) + 4;
  68         OBD_ALLOC(lnr, size);
  69         if (lnr == NULL)
  70                 return ERR_PTR(-ENOMEM);
  71
  72         INIT_LIST_HEAD(&lnr->lnr_lar.lar_list);
  73         lnr->lnr_lar.lar_fid = *lfsck_dto2fid(lfsck->li_obj_dir);
  74         lnr->lnr_lmv = lfsck_lmv_get(lfsck->li_lmv);
  75         lnr->lnr_fid = ent->lde_fid;
  76         lnr->lnr_oit_cookie = lfsck->li_pos_current.lp_oit_cookie;
  77         lnr->lnr_dir_cookie = ent->lde_hash;
  78         lnr->lnr_attr = ent->lde_attrs;
  79         lnr->lnr_size = size;
  80         lnr->lnr_type = type;
  81         lnr->lnr_namelen = ent->lde_namelen;
  82         memcpy(lnr->lnr_name, ent->lde_name, ent->lde_namelen);
  83
  84         return lnr;
  85 }
  86
  87 static void lfsck_namespace_assistant_req_fini(const struct lu_env *env,
  88                                                struct lfsck_assistant_req *lar)
  89 {
  90         struct lfsck_namespace_req *lnr =
  91                         container_of0(lar, struct lfsck_namespace_req, lnr_lar);
  92
  93         if (lnr->lnr_lmv != NULL)
  94                 lfsck_lmv_put(env, lnr->lnr_lmv);
  95
  96         OBD_FREE(lnr, lnr->lnr_size);
  97 }
  98
  99 static void lfsck_namespace_le_to_cpu(struct lfsck_namespace *dst,
 100                                       struct lfsck_namespace *src)
 101 {
 102         dst->ln_magic = le32_to_cpu(src->ln_magic);
 103         dst->ln_status = le32_to_cpu(src->ln_status);
 104         dst->ln_flags = le32_to_cpu(src->ln_flags);
 105         dst->ln_success_count = le32_to_cpu(src->ln_success_count);
 106         dst->ln_run_time_phase1 = le32_to_cpu(src->ln_run_time_phase1);
 107         dst->ln_run_time_phase2 = le32_to_cpu(src->ln_run_time_phase2);
 108         dst->ln_time_last_complete = le64_to_cpu(src->ln_time_last_complete);
 109         dst->ln_time_latest_start = le64_to_cpu(src->ln_time_latest_start);
 110         dst->ln_time_last_checkpoint =
 111                                 le64_to_cpu(src->ln_time_last_checkpoint);
 112         lfsck_position_le_to_cpu(&dst->ln_pos_latest_start,
 113                                  &src->ln_pos_latest_start);
 114         lfsck_position_le_to_cpu(&dst->ln_pos_last_checkpoint,
 115                                  &src->ln_pos_last_checkpoint);
 116         lfsck_position_le_to_cpu(&dst->ln_pos_first_inconsistent,
 117                                  &src->ln_pos_first_inconsistent);
 118         dst->ln_items_checked = le64_to_cpu(src->ln_items_checked);
 119         dst->ln_items_repaired = le64_to_cpu(src->ln_items_repaired);
 120         dst->ln_items_failed = le64_to_cpu(src->ln_items_failed);
 121         dst->ln_dirs_checked = le64_to_cpu(src->ln_dirs_checked);
 122         dst->ln_objs_checked_phase2 = le64_to_cpu(src->ln_objs_checked_phase2);
 123         dst->ln_objs_repaired_phase2 =
 124                                 le64_to_cpu(src->ln_objs_repaired_phase2);
 125         dst->ln_objs_failed_phase2 = le64_to_cpu(src->ln_objs_failed_phase2);
 126         dst->ln_objs_nlink_repaired = le64_to_cpu(src->ln_objs_nlink_repaired);
 127         fid_le_to_cpu(&dst->ln_fid_latest_scanned_phase2,
 128                       &src->ln_fid_latest_scanned_phase2);
 129         dst->ln_dirent_repaired = le64_to_cpu(src->ln_dirent_repaired);
 130         dst->ln_linkea_repaired = le64_to_cpu(src->ln_linkea_repaired);
 131         dst->ln_mul_linked_checked = le64_to_cpu(src->ln_mul_linked_checked);
 132         dst->ln_mul_linked_repaired = le64_to_cpu(src->ln_mul_linked_repaired);
 133         dst->ln_unknown_inconsistency =
 134                                 le64_to_cpu(src->ln_unknown_inconsistency);
 135         dst->ln_unmatched_pairs_repaired =
 136                                 le64_to_cpu(src->ln_unmatched_pairs_repaired);
 137         dst->ln_dangling_repaired = le64_to_cpu(src->ln_dangling_repaired);
 138         dst->ln_mul_ref_repaired = le64_to_cpu(src->ln_mul_ref_repaired);
 139         dst->ln_bad_type_repaired = le64_to_cpu(src->ln_bad_type_repaired);
 140         dst->ln_lost_dirent_repaired =
 141                                 le64_to_cpu(src->ln_lost_dirent_repaired);
 142         dst->ln_striped_dirs_scanned =
 143                                 le64_to_cpu(src->ln_striped_dirs_scanned);
 144         dst->ln_striped_dirs_repaired =
 145                                 le64_to_cpu(src->ln_striped_dirs_repaired);
 146         dst->ln_striped_dirs_failed =
 147                                 le64_to_cpu(src->ln_striped_dirs_failed);
 148         dst->ln_striped_dirs_disabled =
 149                                 le64_to_cpu(src->ln_striped_dirs_disabled);
 150         dst->ln_striped_dirs_skipped =
 151                                 le64_to_cpu(src->ln_striped_dirs_skipped);
 152         dst->ln_striped_shards_scanned =
 153                                 le64_to_cpu(src->ln_striped_shards_scanned);
 154         dst->ln_striped_shards_repaired =
 155                                 le64_to_cpu(src->ln_striped_shards_repaired);
 156         dst->ln_striped_shards_failed =
 157                                 le64_to_cpu(src->ln_striped_shards_failed);
 158         dst->ln_striped_shards_skipped =
 159                                 le64_to_cpu(src->ln_striped_shards_skipped);
 160         dst->ln_name_hash_repaired = le64_to_cpu(src->ln_name_hash_repaired);
 161         dst->ln_local_lpf_scanned = le64_to_cpu(src->ln_local_lpf_scanned);
 162         dst->ln_local_lpf_moved = le64_to_cpu(src->ln_local_lpf_moved);
 163         dst->ln_local_lpf_skipped = le64_to_cpu(src->ln_local_lpf_skipped);
 164         dst->ln_local_lpf_failed = le64_to_cpu(src->ln_local_lpf_failed);
 165         dst->ln_bitmap_size = le32_to_cpu(src->ln_bitmap_size);
 166 }
 167
 168 static void lfsck_namespace_cpu_to_le(struct lfsck_namespace *dst,
 169                                       struct lfsck_namespace *src)
 170 {
 171         dst->ln_magic = cpu_to_le32(src->ln_magic);
 172         dst->ln_status = cpu_to_le32(src->ln_status);
 173         dst->ln_flags = cpu_to_le32(src->ln_flags);
 174         dst->ln_success_count = cpu_to_le32(src->ln_success_count);
 175         dst->ln_run_time_phase1 = cpu_to_le32(src->ln_run_time_phase1);
 176         dst->ln_run_time_phase2 = cpu_to_le32(src->ln_run_time_phase2);
 177         dst->ln_time_last_complete = cpu_to_le64(src->ln_time_last_complete);
 178         dst->ln_time_latest_start = cpu_to_le64(src->ln_time_latest_start);
 179         dst->ln_time_last_checkpoint =
 180                                 cpu_to_le64(src->ln_time_last_checkpoint);
 181         lfsck_position_cpu_to_le(&dst->ln_pos_latest_start,
 182                                  &src->ln_pos_latest_start);
 183         lfsck_position_cpu_to_le(&dst->ln_pos_last_checkpoint,
 184                                  &src->ln_pos_last_checkpoint);
 185         lfsck_position_cpu_to_le(&dst->ln_pos_first_inconsistent,
 186                                  &src->ln_pos_first_inconsistent);
 187         dst->ln_items_checked = cpu_to_le64(src->ln_items_checked);
 188         dst->ln_items_repaired = cpu_to_le64(src->ln_items_repaired);
 189         dst->ln_items_failed = cpu_to_le64(src->ln_items_failed);
 190         dst->ln_dirs_checked = cpu_to_le64(src->ln_dirs_checked);
 191         dst->ln_objs_checked_phase2 = cpu_to_le64(src->ln_objs_checked_phase2);
 192         dst->ln_objs_repaired_phase2 =
 193                                 cpu_to_le64(src->ln_objs_repaired_phase2);
 194         dst->ln_objs_failed_phase2 = cpu_to_le64(src->ln_objs_failed_phase2);
 195         dst->ln_objs_nlink_repaired = cpu_to_le64(src->ln_objs_nlink_repaired);
 196         fid_cpu_to_le(&dst->ln_fid_latest_scanned_phase2,
 197                       &src->ln_fid_latest_scanned_phase2);
 198         dst->ln_dirent_repaired = cpu_to_le64(src->ln_dirent_repaired);
 199         dst->ln_linkea_repaired = cpu_to_le64(src->ln_linkea_repaired);
 200         dst->ln_mul_linked_checked = cpu_to_le64(src->ln_mul_linked_checked);
 201         dst->ln_mul_linked_repaired = cpu_to_le64(src->ln_mul_linked_repaired);
 202         dst->ln_unknown_inconsistency =
 203                                 cpu_to_le64(src->ln_unknown_inconsistency);
 204         dst->ln_unmatched_pairs_repaired =
 205                                 cpu_to_le64(src->ln_unmatched_pairs_repaired);
 206         dst->ln_dangling_repaired = cpu_to_le64(src->ln_dangling_repaired);
 207         dst->ln_mul_ref_repaired = cpu_to_le64(src->ln_mul_ref_repaired);
 208         dst->ln_bad_type_repaired = cpu_to_le64(src->ln_bad_type_repaired);
 209         dst->ln_lost_dirent_repaired =
 210                                 cpu_to_le64(src->ln_lost_dirent_repaired);
 211         dst->ln_striped_dirs_scanned =
 212                                 cpu_to_le64(src->ln_striped_dirs_scanned);
 213         dst->ln_striped_dirs_repaired =
 214                                 cpu_to_le64(src->ln_striped_dirs_repaired);
 215         dst->ln_striped_dirs_failed =
 216                                 cpu_to_le64(src->ln_striped_dirs_failed);
 217         dst->ln_striped_dirs_disabled =
 218                                 cpu_to_le64(src->ln_striped_dirs_disabled);
 219         dst->ln_striped_dirs_skipped =
 220                                 cpu_to_le64(src->ln_striped_dirs_skipped);
 221         dst->ln_striped_shards_scanned =
 222                                 cpu_to_le64(src->ln_striped_shards_scanned);
 223         dst->ln_striped_shards_repaired =
 224                                 cpu_to_le64(src->ln_striped_shards_repaired);
 225         dst->ln_striped_shards_failed =
 226                                 cpu_to_le64(src->ln_striped_shards_failed);
 227         dst->ln_striped_shards_skipped =
 228                                 cpu_to_le64(src->ln_striped_shards_skipped);
 229         dst->ln_name_hash_repaired = cpu_to_le64(src->ln_name_hash_repaired);
 230         dst->ln_local_lpf_scanned = cpu_to_le64(src->ln_local_lpf_scanned);
 231         dst->ln_local_lpf_moved = cpu_to_le64(src->ln_local_lpf_moved);
 232         dst->ln_local_lpf_skipped = cpu_to_le64(src->ln_local_lpf_skipped);
 233         dst->ln_local_lpf_failed = cpu_to_le64(src->ln_local_lpf_failed);
 234         dst->ln_bitmap_size = cpu_to_le32(src->ln_bitmap_size);
 235 }
 236
 237 static void lfsck_namespace_record_failure(const struct lu_env *env,
 238                                            struct lfsck_instance *lfsck,
 239                                            struct lfsck_namespace *ns)
 240 {
 241         struct lfsck_position pos;
 242
 243         ns->ln_items_failed++;
 244         lfsck_pos_fill(env, lfsck, &pos, false);
 245         if (lfsck_pos_is_zero(&ns->ln_pos_first_inconsistent) ||
 246             lfsck_pos_is_eq(&pos, &ns->ln_pos_first_inconsistent) < 0) {
 247                 ns->ln_pos_first_inconsistent = pos;
 248
 249                 CDEBUG(D_LFSCK, "%s: namespace LFSCK hit first non-repaired "
 250                        "inconsistency at the pos ["LPU64", "DFID", "LPX64"]\n",
 251                        lfsck_lfsck2name(lfsck),
 252                        ns->ln_pos_first_inconsistent.lp_oit_cookie,
 253                        PFID(&ns->ln_pos_first_inconsistent.lp_dir_parent),
 254                        ns->ln_pos_first_inconsistent.lp_dir_cookie);
 255         }
 256 }
 257
 258 /**
 259  * Load the MDT bitmap from the lfsck_namespace trace file.
 260  *
 261  * \param[in] env       pointer to the thread context
 262  * \param[in] com       pointer to the lfsck component
 263  *
 264  * \retval              0 for success
 265  * \retval              negative error number on failure or data corruption
 266  */
 267 static int lfsck_namespace_load_bitmap(const struct lu_env *env,
 268                                        struct lfsck_component *com)
 269 {
 270         struct dt_object                *obj    = com->lc_obj;
 271         struct lfsck_assistant_data     *lad    = com->lc_data;
 272         struct lfsck_namespace          *ns     = com->lc_file_ram;
 273         cfs_bitmap_t                    *bitmap = lad->lad_bitmap;
 274         ssize_t                          size;
 275         __u32                            nbits;
 276         int                              rc;
 277         ENTRY;
 278
 279         if (com->lc_lfsck->li_mdt_descs.ltd_tgts_bitmap->size >
 280             ns->ln_bitmap_size)
 281                 nbits = com->lc_lfsck->li_mdt_descs.ltd_tgts_bitmap->size;
 282         else
 283                 nbits = ns->ln_bitmap_size;
 284
 285         if (unlikely(nbits < BITS_PER_LONG))
 286                 nbits = BITS_PER_LONG;
 287
 288         if (nbits > bitmap->size) {
 289                 __u32 new_bits = bitmap->size;
 290                 cfs_bitmap_t *new_bitmap;
 291
 292                 while (new_bits < nbits)
 293                         new_bits <<= 1;
 294
 295                 new_bitmap = CFS_ALLOCATE_BITMAP(new_bits);
 296                 if (new_bitmap == NULL)
 297                         RETURN(-ENOMEM);
 298
 299                 lad->lad_bitmap = new_bitmap;
 300                 CFS_FREE_BITMAP(bitmap);
 301                 bitmap = new_bitmap;
 302         }
 303
 304         if (ns->ln_bitmap_size == 0) {
 305                 lad->lad_incomplete = 0;
 306                 CFS_RESET_BITMAP(bitmap);
 307
 308                 RETURN(0);
 309         }
 310
 311         size = (ns->ln_bitmap_size + 7) >> 3;
 312         rc = dt_xattr_get(env, obj,
 313                           lfsck_buf_get(env, bitmap->data, size),
 314                           XATTR_NAME_LFSCK_BITMAP);
 315         if (rc != size)
 316                 RETURN(rc >= 0 ? -EINVAL : rc);
 317
 318         if (cfs_bitmap_check_empty(bitmap))
 319                 lad->lad_incomplete = 0;
 320         else
 321                 lad->lad_incomplete = 1;
 322
 323         RETURN(0);
 324 }
 325
 326 /**
 327  * Load namespace LFSCK statistics information from the trace file.
 328  *
 329  * For old release (Lustre-2.6 or older), the statistics information was
 330  * stored as XATTR_NAME_LFSCK_NAMESPACE_OLD EA. But in Lustre-2.7, we need
 331  * more statistics information. To avoid confusing old MDT when downgrade,
 332  * Lustre-2.7 stores the namespace LFSCK statistics information as new
 333  * XATTR_NAME_LFSCK_NAMESPACE EA.
 334  *
 335  * \param[in] env       pointer to the thread context
 336  * \param[in] com       pointer to the lfsck component
 337  *
 338  * \retval              0 for success
 339  * \retval              negative error number on failure
 340  */
 341 static int lfsck_namespace_load(const struct lu_env *env,
 342                                 struct lfsck_component *com)
 343 {
 344         int len = com->lc_file_size;
 345         int rc;
 346
 347         rc = dt_xattr_get(env, com->lc_obj,
 348                           lfsck_buf_get(env, com->lc_file_disk, len),
 349                           XATTR_NAME_LFSCK_NAMESPACE);
 350         if (rc == len) {
 351                 struct lfsck_namespace *ns = com->lc_file_ram;
 352
 353                 lfsck_namespace_le_to_cpu(ns,
 354                                 (struct lfsck_namespace *)com->lc_file_disk);
 355                 if (ns->ln_magic != LFSCK_NAMESPACE_MAGIC) {
 356                         CDEBUG(D_LFSCK, "%s: invalid lfsck_namespace magic "
 357                                "%#x != %#x\n", lfsck_lfsck2name(com->lc_lfsck),
 358                                ns->ln_magic, LFSCK_NAMESPACE_MAGIC);
 359                         rc = -ESTALE;
 360                 } else {
 361                         rc = 0;
 362                 }
 363         } else if (rc != -ENODATA) {
 364                 CDEBUG(D_LFSCK, "%s: fail to load lfsck_namespace, "
 365                        "expected = %d: rc = %d\n",
 366                        lfsck_lfsck2name(com->lc_lfsck), len, rc);
 367                 if (rc >= 0)
 368                         rc = -ESTALE;
 369         } else {
 370                 /* Check whether it is old trace file or not.
 371                  * If yes, it should be reset via returning -ESTALE. */
 372                 rc = dt_xattr_get(env, com->lc_obj,
 373                                   lfsck_buf_get(env, com->lc_file_disk, len),
 374                                   XATTR_NAME_LFSCK_NAMESPACE_OLD);
 375                 if (rc >= 0)
 376                         rc = -ESTALE;
 377         }
 378
 379         return rc;
 380 }
 381
 382 static int lfsck_namespace_store(const struct lu_env *env,
 383                                  struct lfsck_component *com, bool init)
 384 {
 385         struct dt_object                *obj    = com->lc_obj;
 386         struct lfsck_instance           *lfsck  = com->lc_lfsck;
 387         struct lfsck_namespace          *ns     = com->lc_file_ram;
 388         struct lfsck_assistant_data     *lad    = com->lc_data;
 389         struct dt_device                *dev    = lfsck_obj2dev(obj);
 390         cfs_bitmap_t                    *bitmap = NULL;
 391         struct thandle                  *handle;
 392         __u32                            nbits  = 0;
 393         int                              len    = com->lc_file_size;
 394         int                              rc;
 395 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 8, 53, 0)
 396         struct lu_buf            tbuf   = { &len, sizeof(len) };
 397 #endif
 398         ENTRY;
 399
 400         if (lad != NULL) {
 401                 bitmap = lad->lad_bitmap;
 402                 nbits = bitmap->size;
 403
 404                 LASSERT(nbits > 0);
 405                 LASSERTF((nbits & 7) == 0, "Invalid nbits %u\n", nbits);
 406         }
 407
 408         ns->ln_bitmap_size = nbits;
 409         lfsck_namespace_cpu_to_le((struct lfsck_namespace *)com->lc_file_disk,
 410                                   ns);
 411         handle = dt_trans_create(env, dev);
 412         if (IS_ERR(handle))
 413                 GOTO(log, rc = PTR_ERR(handle));
 414
 415         rc = dt_declare_xattr_set(env, obj,
 416                                   lfsck_buf_get(env, com->lc_file_disk, len),
 417                                   XATTR_NAME_LFSCK_NAMESPACE, 0, handle);
 418         if (rc != 0)
 419                 GOTO(out, rc);
 420
 421         if (bitmap != NULL) {
 422                 rc = dt_declare_xattr_set(env, obj,
 423                                 lfsck_buf_get(env, bitmap->data, nbits >> 3),
 424                                 XATTR_NAME_LFSCK_BITMAP, 0, handle);
 425                 if (rc != 0)
 426                         GOTO(out, rc);
 427         }
 428
 429 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 8, 53, 0)
 430         /* To be compatible with old Lustre-2.x MDT (x <= 6), generate dummy
 431          * XATTR_NAME_LFSCK_NAMESPACE_OLD EA, then when downgrade to Lustre-2.x,
 432          * the old LFSCK will find "invalid" XATTR_NAME_LFSCK_NAMESPACE_OLD EA,
 433          * then reset the namespace LFSCK trace file. */
 434         if (init) {
 435                 rc = dt_declare_xattr_set(env, obj, &tbuf,
 436                                           XATTR_NAME_LFSCK_NAMESPACE_OLD,
 437                                           LU_XATTR_CREATE, handle);
 438                 if (rc != 0)
 439                         GOTO(out, rc);
 440         }
 441 #endif
 442
 443         rc = dt_trans_start_local(env, dev, handle);
 444         if (rc != 0)
 445                 GOTO(out, rc);
 446
 447         rc = dt_xattr_set(env, obj,
 448                           lfsck_buf_get(env, com->lc_file_disk, len),
 449                           XATTR_NAME_LFSCK_NAMESPACE, 0, handle);
 450         if (rc == 0 && bitmap != NULL)
 451                 rc = dt_xattr_set(env, obj,
 452                                   lfsck_buf_get(env, bitmap->data, nbits >> 3),
 453                                   XATTR_NAME_LFSCK_BITMAP, 0, handle);
 454
 455 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 8, 53, 0)
 456         if (rc == 0 && init)
 457                 rc = dt_xattr_set(env, obj, &tbuf,
 458                                   XATTR_NAME_LFSCK_NAMESPACE_OLD,
 459                                   LU_XATTR_CREATE, handle);
 460 #endif
 461
 462         GOTO(out, rc);
 463
 464 out:
 465         dt_trans_stop(env, dev, handle);
 466
 467 log:
 468         if (rc != 0)
 469                 CDEBUG(D_LFSCK, "%s: fail to store lfsck_namespace: rc = %d\n",
 470                        lfsck_lfsck2name(lfsck), rc);
 471         return rc;
 472 }
 473
 474 static struct dt_object *
 475 lfsck_namespace_load_one_trace_file(const struct lu_env *env,
 476                                     struct lfsck_component *com,
 477                                     struct dt_object *parent,
 478                                     const char *name, bool reset)
 479 {
 480         struct lfsck_instance   *lfsck = com->lc_lfsck;
 481         struct dt_object        *obj;
 482         int                      rc;
 483
 484         if (reset) {
 485                 rc = local_object_unlink(env, lfsck->li_bottom, parent, name);
 486                 if (rc != 0 && rc != -ENOENT)
 487                         return ERR_PTR(rc);
 488         }
 489
 490         obj = local_index_find_or_create(env, lfsck->li_los, parent, name,
 491                                          S_IFREG | S_IRUGO | S_IWUSR,
 492                                          &dt_lfsck_features);
 493
 494         return obj;
 495 }
 496
 497 static int lfsck_namespace_load_sub_trace_files(const struct lu_env *env,
 498                                                 struct lfsck_component *com,
 499                                                 bool reset)
 500 {
 501         char                            *name = lfsck_env_info(env)->lti_key;
 502         struct lfsck_sub_trace_obj      *lsto;
 503         struct dt_object                *obj;
 504         int                              rc;
 505         int                              i;
 506
 507         for (i = 0, lsto = &com->lc_sub_trace_objs[0];
 508              i < LFSCK_STF_COUNT; i++, lsto++) {
 509                 snprintf(name, NAME_MAX, "%s_%02d", LFSCK_NAMESPACE, i);
 510                 if (lsto->lsto_obj != NULL) {
 511                         if (!reset)
 512                                 continue;
 513
 514                         lfsck_object_put(env, lsto->lsto_obj);
 515                         lsto->lsto_obj = NULL;
 516                 }
 517
 518                 obj = lfsck_namespace_load_one_trace_file(env, com,
 519                                 com->lc_lfsck->li_lfsck_dir, name, reset);
 520                 if (IS_ERR(obj))
 521                         return PTR_ERR(obj);
 522
 523                 lsto->lsto_obj = obj;
 524                 rc = obj->do_ops->do_index_try(env, obj, &dt_lfsck_features);
 525                 if (rc != 0)
 526                         return rc;
 527         }
 528
 529         return 0;
 530 }
 531
 532 static int lfsck_namespace_init(const struct lu_env *env,
 533                                 struct lfsck_component *com)
 534 {
 535         struct lfsck_namespace *ns = com->lc_file_ram;
 536         int rc;
 537
 538         memset(ns, 0, sizeof(*ns));
 539         ns->ln_magic = LFSCK_NAMESPACE_MAGIC;
 540         ns->ln_status = LS_INIT;
 541         down_write(&com->lc_sem);
 542         rc = lfsck_namespace_store(env, com, true);
 543         up_write(&com->lc_sem);
 544         if (rc == 0)
 545                 rc = lfsck_namespace_load_sub_trace_files(env, com, true);
 546
 547         return rc;
 548 }
 549
 550 /**
 551  * Update the namespace LFSCK trace file for the given @fid
 552  *
 553  * \param[in] env       pointer to the thread context
 554  * \param[in] com       pointer to the lfsck component
 555  * \param[in] fid       the fid which flags to be updated in the lfsck
 556  *                      trace file
 557  * \param[in] add       true if add new flags, otherwise remove flags
 558  *
 559  * \retval              0 for succeed or nothing to be done
 560  * \retval              negative error number on failure
 561  */
 562 int lfsck_namespace_trace_update(const struct lu_env *env,
 563                                  struct lfsck_component *com,
 564                                  const struct lu_fid *fid,
 565                                  const __u8 flags, bool add)
 566 {
 567         struct lfsck_instance   *lfsck  = com->lc_lfsck;
 568         struct dt_object        *obj;
 569         struct lu_fid           *key    = &lfsck_env_info(env)->lti_fid3;
 570         struct dt_device        *dev;
 571         struct thandle          *th     = NULL;
 572         int                      idx;
 573         int                      rc     = 0;
 574         __u8                     old    = 0;
 575         __u8                     new    = 0;
 576         ENTRY;
 577
 578         LASSERT(flags != 0);
 579
 580         if (unlikely(!fid_is_sane(fid)))
 581                 RETURN(0);
 582
 583         idx = lfsck_sub_trace_file_fid2idx(fid);
 584         obj = com->lc_sub_trace_objs[idx].lsto_obj;
 585         dev = lfsck_obj2dev(obj);
 586         mutex_lock(&com->lc_sub_trace_objs[idx].lsto_mutex);
 587         fid_cpu_to_be(key, fid);
 588         rc = dt_lookup(env, obj, (struct dt_rec *)&old,
 589                        (const struct dt_key *)key);
 590         if (rc == -ENOENT) {
 591                 if (!add)
 592                         GOTO(unlock, rc = 0);
 593
 594                 old = 0;
 595                 new = flags;
 596         } else if (rc == 0) {
 597                 if (add) {
 598                         if ((old & flags) == flags)
 599                                 GOTO(unlock, rc = 0);
 600
 601                         new = old | flags;
 602                 } else {
 603                         if ((old & flags) == 0)
 604                                 GOTO(unlock, rc = 0);
 605
 606                         new = old & ~flags;
 607                 }
 608         } else {
 609                 GOTO(log, rc);
 610         }
 611
 612         th = dt_trans_create(env, dev);
 613         if (IS_ERR(th))
 614                 GOTO(log, rc = PTR_ERR(th));
 615
 616         if (old != 0) {
 617                 rc = dt_declare_delete(env, obj,
 618                                        (const struct dt_key *)key, th);
 619                 if (rc != 0)
 620                         GOTO(log, rc);
 621         }
 622
 623         if (new != 0) {
 624                 rc = dt_declare_insert(env, obj,
 625                                        (const struct dt_rec *)&new,
 626                                        (const struct dt_key *)key, th);
 627                 if (rc != 0)
 628                         GOTO(log, rc);
 629         }
 630
 631         rc = dt_trans_start_local(env, dev, th);
 632         if (rc != 0)
 633                 GOTO(log, rc);
 634
 635         if (old != 0) {
 636                 rc = dt_delete(env, obj, (const struct dt_key *)key, th);
 637                 if (rc != 0)
 638                         GOTO(log, rc);
 639         }
 640
 641         if (new != 0) {
 642                 rc = dt_insert(env, obj, (const struct dt_rec *)&new,
 643                                (const struct dt_key *)key, th, 1);
 644                 if (rc != 0)
 645                         GOTO(log, rc);
 646         }
 647
 648         GOTO(log, rc);
 649
 650 log:
 651         if (th != NULL && !IS_ERR(th))
 652                 dt_trans_stop(env, dev, th);
 653
 654         CDEBUG(D_LFSCK, "%s: namespace LFSCK %s flags for "DFID" in the "
 655                "trace file, flags %x, old %x, new %x: rc = %d\n",
 656                lfsck_lfsck2name(lfsck), add ? "add" : "del", PFID(fid),
 657                (__u32)flags, (__u32)old, (__u32)new, rc);
 658
 659 unlock:
 660         mutex_unlock(&com->lc_sub_trace_objs[idx].lsto_mutex);
 661
 662         return rc;
 663 }
 664
 665 int lfsck_namespace_check_exist(const struct lu_env *env,
 666                                 struct dt_object *dir,
 667                                 struct dt_object *obj, const char *name)
 668 {
 669         struct lu_fid    *fid = &lfsck_env_info(env)->lti_fid;
 670         int               rc;
 671         ENTRY;
 672
 673         if (unlikely(lfsck_is_dead_obj(obj)))
 674                 RETURN(LFSCK_NAMEENTRY_DEAD);
 675
 676         rc = dt_lookup(env, dir, (struct dt_rec *)fid,
 677                        (const struct dt_key *)name);
 678         if (rc == -ENOENT)
 679                 RETURN(LFSCK_NAMEENTRY_REMOVED);
 680
 681         if (rc < 0)
 682                 RETURN(rc);
 683
 684         if (!lu_fid_eq(fid, lfsck_dto2fid(obj)))
 685                 RETURN(LFSCK_NAMEENTRY_RECREATED);
 686
 687         RETURN(0);
 688 }
 689
 690 static int lfsck_declare_namespace_exec_dir(const struct lu_env *env,
 691                                             struct dt_object *obj,
 692                                             struct thandle *handle)
 693 {
 694         int rc;
 695
 696         /* For destroying all invalid linkEA entries. */
 697         rc = dt_declare_xattr_del(env, obj, XATTR_NAME_LINK, handle);
 698         if (rc != 0)
 699                 return rc;
 700
 701         /* For insert new linkEA entry. */
 702         rc = dt_declare_xattr_set(env, obj,
 703                         lfsck_buf_get_const(env, NULL, DEFAULT_LINKEA_SIZE),
 704                         XATTR_NAME_LINK, 0, handle);
 705         return rc;
 706 }
 707
 708 int __lfsck_links_read(const struct lu_env *env, struct dt_object *obj,
 709                        struct linkea_data *ldata)
 710 {
 711         int rc;
 712
 713         if (ldata->ld_buf->lb_buf == NULL)
 714                 return -ENOMEM;
 715
 716         if (!dt_object_exists(obj))
 717                 return -ENOENT;
 718
 719         rc = dt_xattr_get(env, obj, ldata->ld_buf, XATTR_NAME_LINK);
 720         if (rc == -ERANGE) {
 721                 /* Buf was too small, figure out what we need. */
 722                 rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_LINK);
 723                 if (rc <= 0)
 724                         return rc;
 725
 726                 lu_buf_realloc(ldata->ld_buf, rc);
 727                 if (ldata->ld_buf->lb_buf == NULL)
 728                         return -ENOMEM;
 729
 730                 rc = dt_xattr_get(env, obj, ldata->ld_buf, XATTR_NAME_LINK);
 731         }
 732
 733         if (rc > 0)
 734                 rc = linkea_init(ldata);
 735
 736         return rc;
 737 }
 738
 739 /**
 740  * Remove linkEA for the given object.
 741  *
 742  * The caller should take the ldlm lock before the calling.
 743  *
 744  * \param[in] env       pointer to the thread context
 745  * \param[in] com       pointer to the lfsck component
 746  * \param[in] obj       pointer to the dt_object to be handled
 747  *
 748  * \retval              0 for repaired cases
 749  * \retval              negative error number on failure
 750  */
 751 static int lfsck_namespace_links_remove(const struct lu_env *env,
 752                                         struct lfsck_component *com,
 753                                         struct dt_object *obj)
 754 {
 755         struct lfsck_instance           *lfsck  = com->lc_lfsck;
 756         struct dt_device                *dev    = lfsck_obj2dev(obj);
 757         struct thandle                  *th     = NULL;
 758         int                              rc     = 0;
 759         ENTRY;
 760
 761         LASSERT(dt_object_remote(obj) == 0);
 762
 763         th = dt_trans_create(env, dev);
 764         if (IS_ERR(th))
 765                 GOTO(log, rc = PTR_ERR(th));
 766
 767         rc = dt_declare_xattr_del(env, obj, XATTR_NAME_LINK, th);
 768         if (rc != 0)
 769                 GOTO(stop, rc);
 770
 771         rc = dt_trans_start_local(env, dev, th);
 772         if (rc != 0)
 773                 GOTO(stop, rc);
 774
 775         dt_write_lock(env, obj, 0);
 776         if (unlikely(lfsck_is_dead_obj(obj)))
 777                 GOTO(unlock, rc = -ENOENT);
 778
 779         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
 780                 GOTO(unlock, rc = 0);
 781
 782         rc = dt_xattr_del(env, obj, XATTR_NAME_LINK, th);
 783
 784         GOTO(unlock, rc);
 785
 786 unlock:
 787         dt_write_unlock(env, obj);
 788
 789 stop:
 790         dt_trans_stop(env, dev, th);
 791
 792 log:
 793         CDEBUG(D_LFSCK, "%s: namespace LFSCK remove invalid linkEA "
 794                "for the object "DFID": rc = %d\n",
 795                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(obj)), rc);
 796
 797         if (rc == 0) {
 798                 struct lfsck_namespace *ns = com->lc_file_ram;
 799
 800                 ns->ln_flags |= LF_INCONSISTENT;
 801         }
 802
 803         return rc;
 804 }
 805
 806 static int lfsck_links_write(const struct lu_env *env, struct dt_object *obj,
 807                              struct linkea_data *ldata, struct thandle *handle)
 808 {
 809         const struct lu_buf *buf = lfsck_buf_get_const(env,
 810                                                        ldata->ld_buf->lb_buf,
 811                                                        ldata->ld_leh->leh_len);
 812
 813         return dt_xattr_set(env, obj, buf, XATTR_NAME_LINK, 0, handle);
 814 }
 815
 816 static void lfsck_namespace_unpack_linkea_entry(struct linkea_data *ldata,
 817                                                 struct lu_name *cname,
 818                                                 struct lu_fid *pfid,
 819                                                 char *buf)
 820 {
 821         linkea_entry_unpack(ldata->ld_lee, &ldata->ld_reclen, cname, pfid);
 822         /* To guarantee the 'name' is terminated with '0'. */
 823         memcpy(buf, cname->ln_name, cname->ln_namelen);
 824         buf[cname->ln_namelen] = 0;
 825         cname->ln_name = buf;
 826 }
 827
 828 static int lfsck_namespace_filter_linkea_entry(struct linkea_data *ldata,
 829                                                struct lu_name *cname,
 830                                                struct lu_fid *pfid,
 831                                                bool remove)
 832 {
 833         struct link_ea_entry    *oldlee;
 834         int                      oldlen;
 835         int                      repeated = 0;
 836
 837         oldlee = ldata->ld_lee;
 838         oldlen = ldata->ld_reclen;
 839         linkea_next_entry(ldata);
 840         while (ldata->ld_lee != NULL) {
 841                 ldata->ld_reclen = (ldata->ld_lee->lee_reclen[0] << 8) |
 842                                    ldata->ld_lee->lee_reclen[1];
 843                 if (unlikely(ldata->ld_reclen == oldlen &&
 844                              memcmp(ldata->ld_lee, oldlee, oldlen) == 0)) {
 845                         repeated++;
 846                         if (!remove)
 847                                 break;
 848
 849                         linkea_del_buf(ldata, cname);
 850                 } else {
 851                         linkea_next_entry(ldata);
 852                 }
 853         }
 854         ldata->ld_lee = oldlee;
 855         ldata->ld_reclen = oldlen;
 856
 857         return repeated;
 858 }
 859
 860 /**
 861  * Insert orphan into .lustre/lost+found/MDTxxxx/ locally.
 862  *
 863  * Add the specified orphan MDT-object to the .lustre/lost+found/MDTxxxx/
 864  * with the given type to generate the name, the detailed rules for name
 865  * have been described as following.
 866  *
 867  * The function also generates the linkEA corresponding to the name entry
 868  * under the .lustre/lost+found/MDTxxxx/ for the orphan MDT-object.
 869  *
 870  * \param[in] env       pointer to the thread context
 871  * \param[in] com       pointer to the lfsck component
 872  * \param[in] orphan    pointer to the orphan MDT-object
 873  * \param[in] infix     additional information for the orphan name, such as
 874  *                      the FID for original
 875  * \param[in] type      the type for describing why the orphan MDT-object is
 876  *                      created. The rules are as following:
 877  *
 878  *  type "D":           The MDT-object is a directory, it may knows its parent
 879  *                      but because there is no valid linkEA, the LFSCK cannot
 880  *                      know where to put it back to the namespace.
 881  *  type "O":           The MDT-object has no linkEA, and there is no name
 882  *                      entry that references the MDT-object.
 883  *
 884  *  type "S":           The orphan MDT-object is a shard of a striped directory
 885  *
 886  * \see lfsck_layout_recreate_parent() for more types.
 887  *
 888  * The orphan name will be like:
 889  * ${FID}-${infix}-${type}-${conflict_version}
 890  *
 891  * \param[out] count    if some others inserted some linkEA entries by race,
 892  *                      then return the linkEA entries count.
 893  *
 894  * \retval              positive number for repaired cases
 895  * \retval              0 if needs to repair nothing
 896  * \retval              negative error number on failure
 897  */
 898 static int lfsck_namespace_insert_orphan(const struct lu_env *env,
 899                                          struct lfsck_component *com,
 900                                          struct dt_object *orphan,
 901                                          const char *infix, const char *type,
 902                                          int *count)
 903 {
 904         struct lfsck_thread_info        *info   = lfsck_env_info(env);
 905         struct lu_name                  *cname  = &info->lti_name;
 906         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
 907         struct lu_attr                  *la     = &info->lti_la3;
 908         const struct lu_fid             *cfid   = lfsck_dto2fid(orphan);
 909         const struct lu_fid             *pfid;
 910         struct lu_fid                    tfid;
 911         struct lfsck_instance           *lfsck  = com->lc_lfsck;
 912         struct dt_device                *dev    = lfsck_obj2dev(orphan);
 913         struct dt_object                *parent;
 914         struct thandle                  *th     = NULL;
 915         struct lfsck_lock_handle        *pllh   = &info->lti_llh;
 916         struct lustre_handle             clh    = { 0 };
 917         struct linkea_data               ldata  = { NULL };
 918         struct lu_buf                    linkea_buf;
 919         int                              namelen;
 920         int                              idx    = 0;
 921         int                              rc     = 0;
 922         bool                             exist  = false;
 923         ENTRY;
 924
 925         cname->ln_name = NULL;
 926         if (unlikely(lfsck->li_lpf_obj == NULL))
 927                 GOTO(log, rc = -ENXIO);
 928
 929         parent = lfsck->li_lpf_obj;
 930         pfid = lfsck_dto2fid(parent);
 931
 932 again:
 933         do {
 934                 namelen = snprintf(info->lti_key, NAME_MAX, DFID"%s-%s-%d",
 935                                    PFID(cfid), infix, type, idx++);
 936                 rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
 937                                (const struct dt_key *)info->lti_key);
 938                 if (rc != 0 && rc != -ENOENT)
 939                         GOTO(log, rc);
 940
 941                 if (unlikely(rc == 0 && lu_fid_eq(cfid, &tfid)))
 942                         exist = true;
 943         } while (rc == 0 && !exist);
 944
 945         rc = lfsck_lock(env, lfsck, parent, info->lti_key, pllh,
 946                         MDS_INODELOCK_UPDATE, LCK_PW);
 947         if (rc != 0)
 948                 GOTO(log, rc);
 949
 950         /* Re-check whether the name conflict with othrs after taken
 951          * the ldlm lock. */
 952         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
 953                        (const struct dt_key *)info->lti_key);
 954         if (rc == 0) {
 955                 if (!lu_fid_eq(cfid, &tfid)) {
 956                         exist = false;
 957                         lfsck_unlock(pllh);
 958                         goto again;
 959                 }
 960
 961                 exist = true;
 962         } else if (rc != -ENOENT) {
 963                 GOTO(log, rc);
 964         } else {
 965                 exist = false;
 966         }
 967
 968         cname->ln_name = info->lti_key;
 969         cname->ln_namelen = namelen;
 970         rc = linkea_data_new(&ldata, &info->lti_linkea_buf2);
 971         if (rc != 0)
 972                 GOTO(log, rc);
 973
 974         rc = linkea_add_buf(&ldata, cname, pfid);
 975         if (rc != 0)
 976                 GOTO(log, rc);
 977
 978         rc = lfsck_ibits_lock(env, lfsck, orphan, &clh,
 979                               MDS_INODELOCK_UPDATE | MDS_INODELOCK_LOOKUP |
 980                               MDS_INODELOCK_XATTR, LCK_EX);
 981         if (rc != 0)
 982                 GOTO(log, rc);
 983
 984         lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
 985                        ldata.ld_leh->leh_len);
 986         th = dt_trans_create(env, dev);
 987         if (IS_ERR(th))
 988                 GOTO(log, rc = PTR_ERR(th));
 989
 990         if (S_ISDIR(lfsck_object_type(orphan))) {
 991                 rc = dt_declare_delete(env, orphan,
 992                                        (const struct dt_key *)dotdot, th);
 993                 if (rc != 0)
 994                         GOTO(stop, rc);
 995
 996                 rec->rec_type = S_IFDIR;
 997                 rec->rec_fid = pfid;
 998                 rc = dt_declare_insert(env, orphan, (const struct dt_rec *)rec,
 999                                        (const struct dt_key *)dotdot, th);
1000                 if (rc != 0)
1001                         GOTO(stop, rc);
1002         }
1003
1004         rc = dt_declare_xattr_set(env, orphan, &linkea_buf,
1005                                   XATTR_NAME_LINK, 0, th);
1006         if (rc != 0)
1007                 GOTO(stop, rc);
1008
1009         if (!exist) {
1010                 rec->rec_type = lfsck_object_type(orphan) & S_IFMT;
1011                 rec->rec_fid = cfid;
1012                 rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
1013                                        (const struct dt_key *)cname->ln_name,
1014                                        th);
1015                 if (rc != 0)
1016                         GOTO(stop, rc);
1017
1018                 if (S_ISDIR(rec->rec_type)) {
1019                         rc = dt_declare_ref_add(env, parent, th);
1020                         if (rc != 0)
1021                                 GOTO(stop, rc);
1022                 }
1023         }
1024
1025         memset(la, 0, sizeof(*la));
1026         la->la_ctime = cfs_time_current_sec();
1027         la->la_valid = LA_CTIME;
1028         rc = dt_declare_attr_set(env, orphan, la, th);
1029         if (rc != 0)
1030                 GOTO(stop, rc);
1031
1032         rc = dt_trans_start_local(env, dev, th);
1033         if (rc != 0)
1034                 GOTO(stop, rc);
1035
1036         dt_write_lock(env, orphan, 0);
1037         rc = lfsck_links_read(env, orphan, &ldata);
1038         if (likely((rc == -ENODATA) || (rc == -EINVAL) ||
1039                    (rc == 0 && ldata.ld_leh->leh_reccount == 0))) {
1040                 if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1041                         GOTO(unlock, rc = 1);
1042
1043                 if (S_ISDIR(lfsck_object_type(orphan))) {
1044                         rc = dt_delete(env, orphan,
1045                                        (const struct dt_key *)dotdot, th);
1046                         if (rc != 0)
1047                                 GOTO(unlock, rc);
1048
1049                         rec->rec_type = S_IFDIR;
1050                         rec->rec_fid = pfid;
1051                         rc = dt_insert(env, orphan, (const struct dt_rec *)rec,
1052                                        (const struct dt_key *)dotdot, th, 1);
1053                         if (rc != 0)
1054                                 GOTO(unlock, rc);
1055                 }
1056
1057                 rc = dt_xattr_set(env, orphan, &linkea_buf, XATTR_NAME_LINK, 0,
1058                                   th);
1059         } else {
1060                 if (rc == 0 && count != NULL)
1061                         *count = ldata.ld_leh->leh_reccount;
1062
1063                 GOTO(unlock, rc);
1064         }
1065         dt_write_unlock(env, orphan);
1066
1067         if (rc == 0 && !exist) {
1068                 rec->rec_type = lfsck_object_type(orphan) & S_IFMT;
1069                 rec->rec_fid = cfid;
1070                 rc = dt_insert(env, parent, (const struct dt_rec *)rec,
1071                                (const struct dt_key *)cname->ln_name, th, 1);
1072                 if (rc == 0 && S_ISDIR(rec->rec_type)) {
1073                         dt_write_lock(env, parent, 0);
1074                         rc = dt_ref_add(env, parent, th);
1075                         dt_write_unlock(env, parent);
1076                 }
1077         }
1078
1079         if (rc == 0)
1080                 rc = dt_attr_set(env, orphan, la, th);
1081
1082         GOTO(stop, rc = (rc == 0 ? 1 : rc));
1083
1084 unlock:
1085         dt_write_unlock(env, orphan);
1086
1087 stop:
1088         dt_trans_stop(env, dev, th);
1089
1090 log:
1091         lfsck_ibits_unlock(&clh, LCK_EX);
1092         lfsck_unlock(pllh);
1093         CDEBUG(D_LFSCK, "%s: namespace LFSCK insert orphan for the "
1094                "object "DFID", name = %s: rc = %d\n",
1095                lfsck_lfsck2name(lfsck), PFID(cfid),
1096                cname->ln_name != NULL ? cname->ln_name : "<NULL>", rc);
1097
1098         if (rc != 0) {
1099                 struct lfsck_namespace *ns = com->lc_file_ram;
1100
1101                 ns->ln_flags |= LF_INCONSISTENT;
1102         }
1103
1104         return rc;
1105 }
1106
1107 /**
1108  * Add the specified name entry back to namespace.
1109  *
1110  * If there is a linkEA entry that back references a name entry under
1111  * some parent directory, but such parent directory does not have the
1112  * claimed name entry. On the other hand, the linkEA entries count is
1113  * not larger than the MDT-object's hard link count. Under such case,
1114  * it is quite possible that the name entry is lost. Then the LFSCK
1115  * should add the name entry back to the namespace.
1116  *
1117  * \param[in] env       pointer to the thread context
1118  * \param[in] com       pointer to the lfsck component
1119  * \param[in] parent    pointer to the directory under which the name entry
1120  *                      will be inserted into
1121  * \param[in] child     pointer to the object referenced by the name entry
1122  *                      that to be inserted into the parent
1123  * \param[in] name      the name for the child in the parent directory
1124  *
1125  * \retval              positive number for repaired cases
1126  * \retval              0 if nothing to be repaired
1127  * \retval              negative error number on failure
1128  */
1129 static int lfsck_namespace_insert_normal(const struct lu_env *env,
1130                                          struct lfsck_component *com,
1131                                          struct dt_object *parent,
1132                                          struct dt_object *child,
1133                                          const char *name)
1134 {
1135         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1136         struct lu_attr                  *la     = &info->lti_la;
1137         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
1138         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1139         /* The child and its name may be on different MDTs. */
1140         struct dt_device                *dev    = lfsck->li_next;
1141         struct dt_object                *pobj   = NULL;
1142         struct dt_object                *cobj   = NULL;
1143         struct thandle                  *th     = NULL;
1144         struct lfsck_lock_handle        *llh    = &info->lti_llh;
1145         int                              rc     = 0;
1146         ENTRY;
1147
1148         /* @parent/@child may be based on lfsck->li_bottom,
1149          * but here we need the object based on the lfsck->li_next. */
1150
1151         pobj = lfsck_object_locate(dev, parent);
1152         if (IS_ERR(pobj))
1153                 GOTO(log, rc = PTR_ERR(pobj));
1154
1155         if (unlikely(!dt_try_as_dir(env, pobj)))
1156                 GOTO(log, rc = -ENOTDIR);
1157
1158         cobj = lfsck_object_locate(dev, child);
1159         if (IS_ERR(cobj))
1160                 GOTO(log, rc = PTR_ERR(cobj));
1161
1162         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1163                 GOTO(log, rc = 1);
1164
1165         rc = lfsck_lock(env, lfsck, parent, name, llh,
1166                         MDS_INODELOCK_UPDATE, LCK_PW);
1167         if (rc != 0)
1168                 GOTO(log, rc);
1169
1170         th = dt_trans_create(env, dev);
1171         if (IS_ERR(th))
1172                 GOTO(unlock, rc = PTR_ERR(th));
1173
1174         rec->rec_type = lfsck_object_type(cobj) & S_IFMT;
1175         rec->rec_fid = lfsck_dto2fid(cobj);
1176         rc = dt_declare_insert(env, pobj, (const struct dt_rec *)rec,
1177                                (const struct dt_key *)name, th);
1178         if (rc != 0)
1179                 GOTO(stop, rc);
1180
1181         if (S_ISDIR(rec->rec_type)) {
1182                 rc = dt_declare_ref_add(env, pobj, th);
1183                 if (rc != 0)
1184                         GOTO(stop, rc);
1185         }
1186
1187         memset(la, 0, sizeof(*la));
1188         la->la_ctime = cfs_time_current_sec();
1189         la->la_valid = LA_CTIME;
1190         rc = dt_declare_attr_set(env, pobj, la, th);
1191         if (rc != 0)
1192                 GOTO(stop, rc);
1193
1194         rc = dt_declare_attr_set(env, cobj, la, th);
1195         if (rc != 0)
1196                 GOTO(stop, rc);
1197
1198         rc = dt_trans_start_local(env, dev, th);
1199         if (rc != 0)
1200                 GOTO(stop, rc);
1201
1202         rc = dt_insert(env, pobj, (const struct dt_rec *)rec,
1203                        (const struct dt_key *)name, th, 1);
1204         if (rc != 0)
1205                 GOTO(stop, rc);
1206
1207         if (S_ISDIR(rec->rec_type)) {
1208                 dt_write_lock(env, pobj, 0);
1209                 rc = dt_ref_add(env, pobj, th);
1210                 dt_write_unlock(env, pobj);
1211                 if (rc != 0)
1212                         GOTO(stop, rc);
1213         }
1214
1215         la->la_ctime = cfs_time_current_sec();
1216         rc = dt_attr_set(env, pobj, la, th);
1217         if (rc != 0)
1218                 GOTO(stop, rc);
1219
1220         rc = dt_attr_set(env, cobj, la, th);
1221
1222         GOTO(stop, rc = (rc == 0 ? 1 : rc));
1223
1224 stop:
1225         dt_trans_stop(env, dev, th);
1226
1227 unlock:
1228         lfsck_unlock(llh);
1229
1230 log:
1231         CDEBUG(D_LFSCK, "%s: namespace LFSCK insert object "DFID" with "
1232                "the name %s and type %o to the parent "DFID": rc = %d\n",
1233                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(child)), name,
1234                lfsck_object_type(child) & S_IFMT,
1235                PFID(lfsck_dto2fid(parent)), rc);
1236
1237         if (rc != 0) {
1238                 struct lfsck_namespace *ns = com->lc_file_ram;
1239
1240                 ns->ln_flags |= LF_INCONSISTENT;
1241                 if (rc > 0)
1242                         ns->ln_lost_dirent_repaired++;
1243         }
1244
1245         return rc;
1246 }
1247
1248 /**
1249  * Create the specified orphan directory.
1250  *
1251  * For the case that the parent MDT-object stored in some MDT-object's
1252  * linkEA entry is lost, the LFSCK will re-create the parent object as
1253  * an orphan and insert it into .lustre/lost+found/MDTxxxx/ directory
1254  * with the name ${FID}-P-${conflict_version}.
1255  *
1256  * \param[in] env       pointer to the thread context
1257  * \param[in] com       pointer to the lfsck component
1258  * \param[in] orphan    pointer to the orphan MDT-object to be created
1259  * \param[in] lmv       pointer to master LMV EA that will be set to the orphan
1260  *
1261  * \retval              positive number for repaired cases
1262  * \retval              negative error number on failure
1263  */
1264 static int lfsck_namespace_create_orphan_dir(const struct lu_env *env,
1265                                              struct lfsck_component *com,
1266                                              struct dt_object *orphan,
1267                                              struct lmv_mds_md_v1 *lmv)
1268 {
1269         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1270         struct lu_attr                  *la     = &info->lti_la;
1271         struct dt_allocation_hint       *hint   = &info->lti_hint;
1272         struct dt_object_format         *dof    = &info->lti_dof;
1273         struct lu_name                  *cname  = &info->lti_name2;
1274         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
1275         struct lmv_mds_md_v1            *lmv2   = &info->lti_lmv2;
1276         const struct lu_fid             *cfid   = lfsck_dto2fid(orphan);
1277         struct lu_fid                    tfid;
1278         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1279         struct lfsck_namespace          *ns     = com->lc_file_ram;
1280         struct dt_device                *dev    = lfsck_obj2dev(orphan);
1281         struct dt_object                *parent = NULL;
1282         struct thandle                  *th     = NULL;
1283         struct lfsck_lock_handle        *llh    = &info->lti_llh;
1284         struct linkea_data               ldata  = { NULL };
1285         struct lu_buf                    linkea_buf;
1286         struct lu_buf                    lmv_buf;
1287         char                             name[32];
1288         int                              namelen;
1289         int                              idx    = 0;
1290         int                              rc     = 0;
1291         ENTRY;
1292
1293         LASSERT(!dt_object_exists(orphan));
1294
1295         cname->ln_name = NULL;
1296         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1297                 GOTO(log, rc = 1);
1298
1299         if (dt_object_remote(orphan)) {
1300                 LASSERT(lfsck->li_lpf_root_obj != NULL);
1301
1302                 idx = lfsck_find_mdt_idx_by_fid(env, lfsck, cfid);
1303                 if (idx < 0)
1304                         GOTO(log, rc = idx);
1305
1306                 snprintf(name, 8, "MDT%04x", idx);
1307                 rc = dt_lookup(env, lfsck->li_lpf_root_obj,
1308                                (struct dt_rec *)&tfid,
1309                                (const struct dt_key *)name);
1310                 if (rc != 0)
1311                         GOTO(log, rc = (rc == -ENOENT ? -ENXIO : rc));
1312
1313                 parent = lfsck_object_find_bottom(env, lfsck, &tfid);
1314                 if (IS_ERR(parent))
1315                         GOTO(log, rc = PTR_ERR(parent));
1316
1317                 if (unlikely(!dt_try_as_dir(env, parent)))
1318                         GOTO(log, rc = -ENOTDIR);
1319         } else {
1320                 if (unlikely(lfsck->li_lpf_obj == NULL))
1321                         GOTO(log, rc = -ENXIO);
1322
1323                 parent = lfsck->li_lpf_obj;
1324         }
1325
1326         dev = lfsck_find_dev_by_fid(env, lfsck, cfid);
1327         if (IS_ERR(dev))
1328                 GOTO(log, rc = PTR_ERR(dev));
1329
1330         idx = 0;
1331
1332 again:
1333         do {
1334                 namelen = snprintf(name, 31, DFID"-P-%d",
1335                                    PFID(cfid), idx++);
1336                 rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
1337                                (const struct dt_key *)name);
1338                 if (rc != 0 && rc != -ENOENT)
1339                         GOTO(log, rc);
1340         } while (rc == 0);
1341
1342         rc = lfsck_lock(env, lfsck, parent, name, llh,
1343                         MDS_INODELOCK_UPDATE, LCK_PW);
1344         if (rc != 0)
1345                 GOTO(log, rc);
1346
1347         /* Re-check whether the name conflict with othrs after taken
1348          * the ldlm lock. */
1349         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
1350                        (const struct dt_key *)name);
1351         if (unlikely(rc == 0)) {
1352                 lfsck_unlock(llh);
1353                 goto again;
1354         }
1355
1356         if (rc != -ENOENT)
1357                 GOTO(unlock1, rc);
1358
1359         cname->ln_name = name;
1360         cname->ln_namelen = namelen;
1361
1362         memset(la, 0, sizeof(*la));
1363         la->la_mode = S_IFDIR | 0700;
1364         la->la_valid = LA_TYPE | LA_MODE | LA_UID | LA_GID |
1365                        LA_ATIME | LA_MTIME | LA_CTIME;
1366
1367         orphan->do_ops->do_ah_init(env, hint, parent, orphan,
1368                                    la->la_mode & S_IFMT);
1369
1370         memset(dof, 0, sizeof(*dof));
1371         dof->dof_type = dt_mode_to_dft(S_IFDIR);
1372
1373         rc = linkea_data_new(&ldata, &info->lti_linkea_buf2);
1374         if (rc != 0)
1375                 GOTO(unlock1, rc);
1376
1377         rc = linkea_add_buf(&ldata, cname, lfsck_dto2fid(parent));
1378         if (rc != 0)
1379                 GOTO(unlock1, rc);
1380
1381         th = dt_trans_create(env, dev);
1382         if (IS_ERR(th))
1383                 GOTO(unlock1, rc = PTR_ERR(th));
1384
1385         /* Sync the remote transaction to guarantee that the subsequent
1386          * lock against the @orphan can find the @orphan in time. */
1387         if (dt_object_remote(orphan))
1388                 th->th_sync = 1;
1389
1390         rc = dt_declare_create(env, orphan, la, hint, dof, th);
1391         if (rc != 0)
1392                 GOTO(stop, rc);
1393
1394         if (unlikely(!dt_try_as_dir(env, orphan)))
1395                 GOTO(stop, rc = -ENOTDIR);
1396
1397         rec->rec_type = S_IFDIR;
1398         rec->rec_fid = cfid;
1399         rc = dt_declare_insert(env, orphan, (const struct dt_rec *)rec,
1400                                (const struct dt_key *)dot, th);
1401         if (rc != 0)
1402                 GOTO(stop, rc);
1403
1404         rec->rec_fid = lfsck_dto2fid(parent);
1405         rc = dt_declare_insert(env, orphan, (const struct dt_rec *)rec,
1406                                (const struct dt_key *)dotdot, th);
1407         if (rc == 0)
1408                 rc = dt_declare_ref_add(env, orphan, th);
1409
1410         if (rc != 0)
1411                 GOTO(stop, rc);
1412
1413         rc = dt_declare_ref_add(env, orphan, th);
1414         if (rc != 0)
1415                 GOTO(stop, rc);
1416
1417         if (lmv != NULL) {
1418                 lmv->lmv_magic = LMV_MAGIC;
1419                 lmv->lmv_master_mdt_index = lfsck_dev_idx(lfsck);
1420                 lfsck_lmv_header_cpu_to_le(lmv2, lmv);
1421                 lfsck_buf_init(&lmv_buf, lmv2, sizeof(*lmv2));
1422                 rc = dt_declare_xattr_set(env, orphan, &lmv_buf,
1423                                           XATTR_NAME_LMV, 0, th);
1424                 if (rc != 0)
1425                         GOTO(stop, rc);
1426         }
1427
1428         lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
1429                        ldata.ld_leh->leh_len);
1430         rc = dt_declare_xattr_set(env, orphan, &linkea_buf,
1431                                   XATTR_NAME_LINK, 0, th);
1432         if (rc != 0)
1433                 GOTO(stop, rc);
1434
1435         rec->rec_fid = cfid;
1436         rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
1437                                (const struct dt_key *)name, th);
1438         if (rc == 0)
1439                 rc = dt_declare_ref_add(env, parent, th);
1440
1441         if (rc != 0)
1442                 GOTO(stop, rc);
1443
1444         rc = dt_trans_start_local(env, dev, th);
1445         if (rc != 0)
1446                 GOTO(stop, rc);
1447
1448         dt_write_lock(env, orphan, 0);
1449         rc = dt_create(env, orphan, la, hint, dof, th);
1450         if (rc != 0)
1451                 GOTO(unlock2, rc);
1452
1453         rec->rec_fid = cfid;
1454         rc = dt_insert(env, orphan, (const struct dt_rec *)rec,
1455                        (const struct dt_key *)dot, th, 1);
1456         if (rc != 0)
1457                 GOTO(unlock2, rc);
1458
1459         rec->rec_fid = lfsck_dto2fid(parent);
1460         rc = dt_insert(env, orphan, (const struct dt_rec *)rec,
1461                        (const struct dt_key *)dotdot, th, 1);
1462         if (rc != 0)
1463                 GOTO(unlock2, rc);
1464
1465         rc = dt_ref_add(env, orphan, th);
1466         if (rc != 0)
1467                 GOTO(unlock2, rc);
1468
1469         if (lmv != NULL) {
1470                 rc = dt_xattr_set(env, orphan, &lmv_buf, XATTR_NAME_LMV, 0, th);
1471                 if (rc != 0)
1472                         GOTO(unlock2, rc);
1473         }
1474
1475         rc = dt_xattr_set(env, orphan, &linkea_buf,
1476                           XATTR_NAME_LINK, 0, th);
1477         dt_write_unlock(env, orphan);
1478         if (rc != 0)
1479                 GOTO(stop, rc);
1480
1481         rec->rec_fid = cfid;
1482         rc = dt_insert(env, parent, (const struct dt_rec *)rec,
1483                        (const struct dt_key *)name, th, 1);
1484         if (rc == 0) {
1485                 dt_write_lock(env, parent, 0);
1486                 rc = dt_ref_add(env, parent, th);
1487                 dt_write_unlock(env, parent);
1488         }
1489
1490         GOTO(stop, rc = (rc == 0 ? 1 : rc));
1491
1492 unlock2:
1493         dt_write_unlock(env, orphan);
1494
1495 stop:
1496         dt_trans_stop(env, dev, th);
1497
1498 unlock1:
1499         lfsck_unlock(llh);
1500
1501 log:
1502         CDEBUG(D_LFSCK, "%s: namespace LFSCK create orphan dir for "
1503                "the object "DFID", name = %s: rc = %d\n",
1504                lfsck_lfsck2name(lfsck), PFID(cfid),
1505                cname->ln_name != NULL ? cname->ln_name : "<NULL>", rc);
1506
1507         if (parent != NULL && !IS_ERR(parent) && parent != lfsck->li_lpf_obj)
1508                 lfsck_object_put(env, parent);
1509
1510         if (rc != 0)
1511                 ns->ln_flags |= LF_INCONSISTENT;
1512
1513         return rc;
1514 }
1515
1516 /**
1517  * Remove the specified entry from the linkEA.
1518  *
1519  * Locate the linkEA entry with the given @cname and @pfid, then
1520  * remove this entry or the other entries those are repeated with
1521  * this entry.
1522  *
1523  * \param[in] env       pointer to the thread context
1524  * \param[in] com       pointer to the lfsck component
1525  * \param[in] obj       pointer to the dt_object to be handled
1526  * \param[in,out]ldata  pointer to the buffer that holds the linkEA
1527  * \param[in] cname     the name for the child in the parent directory
1528  * \param[in] pfid      the parent directory's FID for the linkEA
1529  * \param[in] next      if true, then remove the first found linkEA
1530  *                      entry, and move the ldata->ld_lee to next entry
1531  *
1532  * \retval              positive number for repaired cases
1533  * \retval              0 if nothing to be repaired
1534  * \retval              negative error number on failure
1535  */
1536 static int lfsck_namespace_shrink_linkea(const struct lu_env *env,
1537                                          struct lfsck_component *com,
1538                                          struct dt_object *obj,
1539                                          struct linkea_data *ldata,
1540                                          struct lu_name *cname,
1541                                          struct lu_fid *pfid,
1542                                          bool next)
1543 {
1544         struct lfsck_instance           *lfsck     = com->lc_lfsck;
1545         struct dt_device                *dev       = lfsck_obj2dev(obj);
1546         struct lfsck_bookmark           *bk        = &lfsck->li_bookmark_ram;
1547         struct thandle                  *th        = NULL;
1548         struct lustre_handle             lh        = { 0 };
1549         struct linkea_data               ldata_new = { NULL };
1550         struct lu_buf                    linkea_buf;
1551         int                              rc        = 0;
1552         ENTRY;
1553
1554         rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
1555                               MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1556                               LCK_EX);
1557         if (rc != 0)
1558                 GOTO(log, rc);
1559
1560         if (next)
1561                 linkea_del_buf(ldata, cname);
1562         else
1563                 lfsck_namespace_filter_linkea_entry(ldata, cname, pfid,
1564                                                     true);
1565         lfsck_buf_init(&linkea_buf, ldata->ld_buf->lb_buf,
1566                        ldata->ld_leh->leh_len);
1567
1568 again:
1569         th = dt_trans_create(env, dev);
1570         if (IS_ERR(th))
1571                 GOTO(unlock1, rc = PTR_ERR(th));
1572
1573         rc = dt_declare_xattr_set(env, obj, &linkea_buf,
1574                                   XATTR_NAME_LINK, 0, th);
1575         if (rc != 0)
1576                 GOTO(stop, rc);
1577
1578         rc = dt_trans_start_local(env, dev, th);
1579         if (rc != 0)
1580                 GOTO(stop, rc);
1581
1582         dt_write_lock(env, obj, 0);
1583         if (unlikely(lfsck_is_dead_obj(obj)))
1584                 GOTO(unlock2, rc = -ENOENT);
1585
1586         rc = lfsck_links_read2(env, obj, &ldata_new);
1587         if (rc != 0)
1588                 GOTO(unlock2, rc);
1589
1590         /* The specified linkEA entry has been removed by race. */
1591         rc = linkea_links_find(&ldata_new, cname, pfid);
1592         if (rc != 0)
1593                 GOTO(unlock2, rc = 0);
1594
1595         if (bk->lb_param & LPF_DRYRUN)
1596                 GOTO(unlock2, rc = 1);
1597
1598         if (next)
1599                 linkea_del_buf(&ldata_new, cname);
1600         else
1601                 lfsck_namespace_filter_linkea_entry(&ldata_new, cname, pfid,
1602                                                     true);
1603
1604         if (linkea_buf.lb_len < ldata_new.ld_leh->leh_len) {
1605                 dt_write_unlock(env, obj);
1606                 dt_trans_stop(env, dev, th);
1607                 lfsck_buf_init(&linkea_buf, ldata_new.ld_buf->lb_buf,
1608                                ldata_new.ld_leh->leh_len);
1609                 goto again;
1610         }
1611
1612         lfsck_buf_init(&linkea_buf, ldata_new.ld_buf->lb_buf,
1613                        ldata_new.ld_leh->leh_len);
1614         rc = dt_xattr_set(env, obj, &linkea_buf, XATTR_NAME_LINK, 0, th);
1615
1616         GOTO(unlock2, rc = (rc == 0 ? 1 : rc));
1617
1618 unlock2:
1619         dt_write_unlock(env, obj);
1620
1621 stop:
1622         dt_trans_stop(env, dev, th);
1623
1624 unlock1:
1625         lfsck_ibits_unlock(&lh, LCK_EX);
1626
1627 log:
1628         CDEBUG(D_LFSCK, "%s: namespace LFSCK remove %s linkEA entry "
1629                "for the object: "DFID", parent "DFID", name %.*s\n",
1630                lfsck_lfsck2name(lfsck), next ? "invalid" : "redundant",
1631                PFID(lfsck_dto2fid(obj)), PFID(pfid), cname->ln_namelen,
1632                cname->ln_name);
1633
1634         if (rc != 0) {
1635                 struct lfsck_namespace *ns = com->lc_file_ram;
1636
1637                 ns->ln_flags |= LF_INCONSISTENT;
1638         }
1639
1640         return rc;
1641 }
1642
1643 /**
1644  * Conditionally remove the specified entry from the linkEA.
1645  *
1646  * Take the parent lock firstly, then check whether the specified
1647  * name entry exists or not: if yes, do nothing; otherwise, call
1648  * lfsck_namespace_shrink_linkea() to remove the linkea entry.
1649  *
1650  * \param[in] env       pointer to the thread context
1651  * \param[in] com       pointer to the lfsck component
1652  * \param[in] parent    pointer to the parent directory
1653  * \param[in] child     pointer to the child object that holds the linkEA
1654  * \param[in,out]ldata  pointer to the buffer that holds the linkEA
1655  * \param[in] cname     the name for the child in the parent directory
1656  * \param[in] pfid      the parent directory's FID for the linkEA
1657  *
1658  * \retval              positive number for repaired cases
1659  * \retval              0 if nothing to be repaired
1660  * \retval              negative error number on failure
1661  */
1662 static int lfsck_namespace_shrink_linkea_cond(const struct lu_env *env,
1663                                               struct lfsck_component *com,
1664                                               struct dt_object *parent,
1665                                               struct dt_object *child,
1666                                               struct linkea_data *ldata,
1667                                               struct lu_name *cname,
1668                                               struct lu_fid *pfid)
1669 {
1670         struct lfsck_thread_info *info  = lfsck_env_info(env);
1671         struct lu_fid            *cfid  = &info->lti_fid3;
1672         struct lfsck_lock_handle *llh   = &info->lti_llh;
1673         int                       rc;
1674         ENTRY;
1675
1676         rc = lfsck_lock(env, com->lc_lfsck, parent, cname->ln_name, llh,
1677                         MDS_INODELOCK_UPDATE, LCK_PR);
1678         if (rc != 0)
1679                 RETURN(rc);
1680
1681         dt_read_lock(env, parent, 0);
1682         if (unlikely(lfsck_is_dead_obj(parent))) {
1683                 dt_read_unlock(env, parent);
1684                 lfsck_unlock(llh);
1685                 rc = lfsck_namespace_shrink_linkea(env, com, child, ldata,
1686                                                    cname, pfid, true);
1687
1688                 RETURN(rc);
1689         }
1690
1691         rc = dt_lookup(env, parent, (struct dt_rec *)cfid,
1692                        (const struct dt_key *)cname->ln_name);
1693         dt_read_unlock(env, parent);
1694
1695         /* It is safe to release the ldlm lock, because when the logic come
1696          * here, we have got all the needed information above whether the
1697          * linkEA entry is valid or not. It is not important that others
1698          * may add new linkEA entry after the ldlm lock released. If other
1699          * has removed the specified linkEA entry by race, then it is OK,
1700          * because the subsequent lfsck_namespace_shrink_linkea() can handle
1701          * such case. */
1702         lfsck_unlock(llh);
1703         if (rc == -ENOENT) {
1704                 rc = lfsck_namespace_shrink_linkea(env, com, child, ldata,
1705                                                    cname, pfid, true);
1706
1707                 RETURN(rc);
1708         }
1709
1710         if (rc != 0)
1711                 RETURN(rc);
1712
1713         /* The LFSCK just found some internal status of cross-MDTs
1714          * create operation. That is normal. */
1715         if (lu_fid_eq(cfid, lfsck_dto2fid(child))) {
1716                 linkea_next_entry(ldata);
1717
1718                 RETURN(0);
1719         }
1720
1721         rc = lfsck_namespace_shrink_linkea(env, com, child, ldata, cname,
1722                                            pfid, true);
1723
1724         RETURN(rc);
1725 }
1726
1727 /**
1728  * Conditionally replace name entry in the parent.
1729  *
1730  * As required, the LFSCK may re-create the lost MDT-object for dangling
1731  * name entry, but such repairing may be wrong because of bad FID in the
1732  * name entry. As the LFSCK processing, the real MDT-object may be found,
1733  * then the LFSCK should check whether the former re-created MDT-object
1734  * has been modified or not, if not, then destroy it and update the name
1735  * entry in the parent to reference the real MDT-object.
1736  *
1737  * \param[in] env       pointer to the thread context
1738  * \param[in] com       pointer to the lfsck component
1739  * \param[in] parent    pointer to the parent directory
1740  * \param[in] child     pointer to the MDT-object that may be the real
1741  *                      MDT-object corresponding to the name entry in parent
1742  * \param[in] cfid      the current FID in the name entry
1743  * \param[in] cname     contains the name of the child in the parent directory
1744  *
1745  * \retval              positive number for repaired cases
1746  * \retval              0 if nothing to be repaired
1747  * \retval              negative error number on failure
1748  */
1749 static int lfsck_namespace_replace_cond(const struct lu_env *env,
1750                                         struct lfsck_component *com,
1751                                         struct dt_object *parent,
1752                                         struct dt_object *child,
1753                                         const struct lu_fid *cfid,
1754                                         const struct lu_name *cname)
1755 {
1756         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1757         struct lu_attr                  *la     = &info->lti_la;
1758         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
1759         struct lu_fid                    tfid;
1760         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1761         /* The child and its name may be on different MDTs. */
1762         struct dt_device                *dev    = lfsck->li_next;
1763         const char                      *name   = cname->ln_name;
1764         struct dt_object                *pobj   = NULL;
1765         struct dt_object                *cobj   = NULL;
1766         struct lfsck_lock_handle        *pllh   = &info->lti_llh;
1767         struct lustre_handle             clh    = { 0 };
1768         struct linkea_data               ldata  = { NULL };
1769         struct thandle                  *th     = NULL;
1770         bool                             exist  = true;
1771         int                              rc     = 0;
1772         ENTRY;
1773
1774         /* @parent/@child may be based on lfsck->li_bottom,
1775          * but here we need the object based on the lfsck->li_next. */
1776
1777         pobj = lfsck_object_locate(dev, parent);
1778         if (IS_ERR(pobj))
1779                 GOTO(log, rc = PTR_ERR(pobj));
1780
1781         if (unlikely(!dt_try_as_dir(env, pobj)))
1782                 GOTO(log, rc = -ENOTDIR);
1783
1784         rc = lfsck_lock(env, lfsck, parent, name, pllh,
1785                         MDS_INODELOCK_UPDATE, LCK_PW);
1786         if (rc != 0)
1787                 GOTO(log, rc);
1788
1789         if (!fid_is_sane(cfid)) {
1790                 exist = false;
1791                 goto replace;
1792         }
1793
1794         cobj = lfsck_object_find_by_dev(env, dev, cfid);
1795         if (IS_ERR(cobj)) {
1796                 rc = PTR_ERR(cobj);
1797                 if (rc == -ENOENT) {
1798                         exist = false;
1799                         goto replace;
1800                 }
1801
1802                 GOTO(log, rc);
1803         }
1804
1805         if (!dt_object_exists(cobj)) {
1806                 exist = false;
1807                 goto replace;
1808         }
1809
1810         rc = dt_lookup(env, pobj, (struct dt_rec *)&tfid,
1811                        (const struct dt_key *)name);
1812         if (rc == -ENOENT) {
1813                 exist = false;
1814                 goto replace;
1815         }
1816
1817         if (rc != 0)
1818                 GOTO(log, rc);
1819
1820         /* Someone changed the name entry, cannot replace it. */
1821         if (!lu_fid_eq(cfid, &tfid))
1822                 GOTO(log, rc = 0);
1823
1824         /* lock the object to be destroyed. */
1825         rc = lfsck_ibits_lock(env, lfsck, cobj, &clh,
1826                               MDS_INODELOCK_UPDATE |
1827                               MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1828                               LCK_EX);
1829         if (rc != 0)
1830                 GOTO(log, rc);
1831
1832         if (unlikely(lfsck_is_dead_obj(cobj))) {
1833                 exist = false;
1834                 goto replace;
1835         }
1836
1837         rc = dt_attr_get(env, cobj, la);
1838         if (rc != 0)
1839                 GOTO(log, rc);
1840
1841         /* The object has been modified by other(s), or it is not created by
1842          * LFSCK, the two cases are indistinguishable. So cannot replace it. */
1843         if (la->la_ctime != 0)
1844                 GOTO(log, rc);
1845
1846         if (S_ISREG(la->la_mode)) {
1847                 rc = dt_xattr_get(env, cobj, &LU_BUF_NULL, XATTR_NAME_LOV);
1848                 /* If someone has created related OST-object(s),
1849                  * then keep it. */
1850                 if ((rc > 0) || (rc < 0 && rc != -ENODATA))
1851                         GOTO(log, rc = (rc > 0 ? 0 : rc));
1852         }
1853
1854 replace:
1855         dt_read_lock(env, child, 0);
1856         rc = lfsck_links_read2(env, child, &ldata);
1857         dt_read_unlock(env, child);
1858
1859         /* Someone changed the child, no need to replace. */
1860         if (rc == -ENODATA)
1861                 GOTO(log, rc = 0);
1862
1863         if (rc != 0)
1864                 GOTO(log, rc);
1865
1866         rc = linkea_links_find(&ldata, cname, lfsck_dto2fid(pobj));
1867         /* Someone moved the child, no need to replace. */
1868         if (rc != 0)
1869                 GOTO(log, rc = 0);
1870
1871         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1872                 GOTO(log, rc = 1);
1873
1874         th = dt_trans_create(env, dev);
1875         if (IS_ERR(th))
1876                 GOTO(log, rc = PTR_ERR(th));
1877
1878         if (exist) {
1879                 rc = dt_declare_destroy(env, cobj, th);
1880                 if (rc != 0)
1881                         GOTO(stop, rc);
1882         }
1883
1884         rc = dt_declare_delete(env, pobj, (const struct dt_key *)name, th);
1885         if (rc != 0)
1886                 GOTO(stop, rc);
1887
1888         rec->rec_type = S_IFDIR;
1889         rec->rec_fid = lfsck_dto2fid(child);
1890         rc = dt_declare_insert(env, pobj, (const struct dt_rec *)rec,
1891                                (const struct dt_key *)name, th);
1892         if (rc != 0)
1893                 GOTO(stop, rc);
1894
1895         rc = dt_trans_start_local(env, dev, th);
1896         if (rc != 0)
1897                 GOTO(stop, rc);
1898
1899         if (exist) {
1900                 rc = dt_destroy(env, cobj, th);
1901                 if (rc != 0)
1902                         GOTO(stop, rc);
1903         }
1904
1905         /* The old name entry maybe not exist. */
1906         rc = dt_delete(env, pobj, (const struct dt_key *)name, th);
1907         if (rc != 0 && rc != -ENOENT)
1908                 GOTO(stop, rc);
1909
1910         rc = dt_insert(env, pobj, (const struct dt_rec *)rec,
1911                        (const struct dt_key *)name, th, 1);
1912
1913         GOTO(stop, rc = (rc == 0 ? 1 : rc));
1914
1915 stop:
1916         dt_trans_stop(env, dev, th);
1917
1918 log:
1919         lfsck_ibits_unlock(&clh, LCK_EX);
1920         lfsck_unlock(pllh);
1921
1922         if (cobj != NULL && !IS_ERR(cobj))
1923                 lfsck_object_put(env, cobj);
1924
1925         CDEBUG(D_LFSCK, "%s: namespace LFSCK conditionally destroy the "
1926                "object "DFID" because of conflict with the object "DFID
1927                " under the parent "DFID" with name %s: rc = %d\n",
1928                lfsck_lfsck2name(lfsck), PFID(cfid),
1929                PFID(lfsck_dto2fid(child)), PFID(lfsck_dto2fid(parent)),
1930                name, rc);
1931
1932         return rc;
1933 }
1934
1935 /**
1936  * Overwrite the linkEA for the object with the given ldata.
1937  *
1938  * The caller should take the ldlm lock before the calling.
1939  *
1940  * \param[in] env       pointer to the thread context
1941  * \param[in] com       pointer to the lfsck component
1942  * \param[in] obj       pointer to the dt_object to be handled
1943  * \param[in] ldata     pointer to the new linkEA data
1944  *
1945  * \retval              positive number for repaired cases
1946  * \retval              0 if nothing to be repaired
1947  * \retval              negative error number on failure
1948  */
1949 int lfsck_namespace_rebuild_linkea(const struct lu_env *env,
1950                                    struct lfsck_component *com,
1951                                    struct dt_object *obj,
1952                                    struct linkea_data *ldata)
1953 {
1954         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1955         struct dt_device                *dev    = lfsck_obj2dev(obj);
1956         struct thandle                  *th     = NULL;
1957         struct lu_buf                    linkea_buf;
1958         int                              rc     = 0;
1959         ENTRY;
1960
1961         th = dt_trans_create(env, dev);
1962         if (IS_ERR(th))
1963                 GOTO(log, rc = PTR_ERR(th));
1964
1965         lfsck_buf_init(&linkea_buf, ldata->ld_buf->lb_buf,
1966                        ldata->ld_leh->leh_len);
1967         rc = dt_declare_xattr_set(env, obj, &linkea_buf,
1968                                   XATTR_NAME_LINK, 0, th);
1969         if (rc != 0)
1970                 GOTO(stop, rc);
1971
1972         rc = dt_trans_start_local(env, dev, th);
1973         if (rc != 0)
1974                 GOTO(stop, rc);
1975
1976         dt_write_lock(env, obj, 0);
1977         if (unlikely(lfsck_is_dead_obj(obj)))
1978                 GOTO(unlock, rc = 0);
1979
1980         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1981                 GOTO(unlock, rc = 1);
1982
1983         rc = dt_xattr_set(env, obj, &linkea_buf,
1984                           XATTR_NAME_LINK, 0, th);
1985
1986         GOTO(unlock, rc = (rc == 0 ? 1 : rc));
1987
1988 unlock:
1989         dt_write_unlock(env, obj);
1990
1991 stop:
1992         dt_trans_stop(env, dev, th);
1993
1994 log:
1995         CDEBUG(D_LFSCK, "%s: namespace LFSCK rebuild linkEA for the "
1996                "object "DFID": rc = %d\n",
1997                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(obj)), rc);
1998
1999         if (rc != 0) {
2000                 struct lfsck_namespace *ns = com->lc_file_ram;
2001
2002                 ns->ln_flags |= LF_INCONSISTENT;
2003         }
2004
2005         return rc;
2006 }
2007
2008 /**
2009  * Repair invalid name entry.
2010  *
2011  * If the name entry contains invalid information, such as bad file type
2012  * or (and) corrupted object FID, then either remove the name entry or
2013  * udpate the name entry with the given (right) information.
2014  *
2015  * \param[in] env       pointer to the thread context
2016  * \param[in] com       pointer to the lfsck component
2017  * \param[in] parent    pointer to the parent directory
2018  * \param[in] child     pointer to the object referenced by the name entry
2019  * \param[in] name      the old name of the child under the parent directory
2020  * \param[in] name2     the new name of the child under the parent directory
2021  * \param[in] type      the type claimed by the name entry
2022  * \param[in] update    update the name entry if true; otherwise, remove it
2023  * \param[in] dec       decrease the parent nlink count if true
2024  *
2025  * \retval              positive number for repaired successfully
2026  * \retval              0 if nothing to be repaired
2027  * \retval              negative error number on failure
2028  */
2029 int lfsck_namespace_repair_dirent(const struct lu_env *env,
2030                                   struct lfsck_component *com,
2031                                   struct dt_object *parent,
2032                                   struct dt_object *child,
2033                                   const char *name, const char *name2,
2034                                   __u16 type, bool update, bool dec)
2035 {
2036         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2037         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
2038         const struct lu_fid             *cfid   = lfsck_dto2fid(child);
2039         struct lu_fid                    tfid;
2040         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2041         struct dt_object                *dto;
2042         struct dt_device                *dev    = lfsck->li_next;
2043         struct thandle                  *th     = NULL;
2044         struct lfsck_lock_handle        *llh    = &info->lti_llh;
2045         struct lustre_handle             lh     = { 0 };
2046         int                              rc     = 0;
2047         ENTRY;
2048
2049         if (unlikely(!dt_try_as_dir(env, parent)))
2050                 GOTO(log, rc = -ENOTDIR);
2051
2052         if (!update || strcmp(name, name2) == 0)
2053                 rc = lfsck_lock(env, lfsck, parent, name, llh,
2054                                 MDS_INODELOCK_UPDATE, LCK_PW);
2055         else
2056                 rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
2057                                       MDS_INODELOCK_UPDATE, LCK_PW);
2058         if (rc != 0)
2059                 GOTO(log, rc);
2060
2061         th = dt_trans_create(env, dev);
2062         if (IS_ERR(th))
2063                 GOTO(unlock1, rc = PTR_ERR(th));
2064
2065         dto = dt_object_locate(parent, th->th_dev);
2066         rc = dt_declare_delete(env, dto, (const struct dt_key *)name, th);
2067         if (rc != 0)
2068                 GOTO(stop, rc);
2069
2070         if (update) {
2071                 rec->rec_type = lfsck_object_type(child) & S_IFMT;
2072                 rec->rec_fid = cfid;
2073                 rc = dt_declare_insert(env, dto,
2074                                        (const struct dt_rec *)rec,
2075                                        (const struct dt_key *)name2, th);
2076                 if (rc != 0)
2077                         GOTO(stop, rc);
2078         }
2079
2080         if (dec) {
2081                 rc = dt_declare_ref_del(env, dto, th);
2082                 if (rc != 0)
2083                         GOTO(stop, rc);
2084         }
2085
2086         rc = dt_trans_start_local(env, dev, th);
2087         if (rc != 0)
2088                 GOTO(stop, rc);
2089
2090
2091         dt_write_lock(env, dto, 0);
2092         rc = dt_lookup(env, dto, (struct dt_rec *)&tfid,
2093                        (const struct dt_key *)name);
2094         /* Someone has removed the bad name entry by race. */
2095         if (rc == -ENOENT)
2096                 GOTO(unlock2, rc = 0);
2097
2098         if (rc != 0)
2099                 GOTO(unlock2, rc);
2100
2101         /* Someone has removed the bad name entry and reused it for other
2102          * object by race. */
2103         if (!lu_fid_eq(&tfid, cfid))
2104                 GOTO(unlock2, rc = 0);
2105
2106         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
2107                 GOTO(unlock2, rc = 1);
2108
2109         rc = dt_delete(env, dto, (const struct dt_key *)name, th);
2110         if (rc != 0)
2111                 GOTO(unlock2, rc);
2112
2113         if (update) {
2114                 rc = dt_insert(env, dto,
2115                                (const struct dt_rec *)rec,
2116                                (const struct dt_key *)name2, th, 1);
2117                 if (rc != 0)
2118                         GOTO(unlock2, rc);
2119         }
2120
2121         if (dec) {
2122                 rc = dt_ref_del(env, dto, th);
2123                 if (rc != 0)
2124                         GOTO(unlock2, rc);
2125         }
2126
2127         GOTO(unlock2, rc = (rc == 0 ? 1 : rc));
2128
2129 unlock2:
2130         dt_write_unlock(env, parent);
2131
2132 stop:
2133         dt_trans_stop(env, dev, th);
2134
2135         /* We are not sure whether the child will become orphan or not.
2136          * Record it in the LFSCK trace file for further checking in
2137          * the second-stage scanning. */
2138         if (!update && !dec && rc == 0)
2139                 lfsck_namespace_trace_update(env, com, cfid,
2140                                              LNTF_CHECK_LINKEA, true);
2141
2142 unlock1:
2143         /* It is harmless even if unlock the unused lock_handle */
2144         lfsck_ibits_unlock(&lh, LCK_PW);
2145         lfsck_unlock(llh);
2146
2147 log:
2148         CDEBUG(D_LFSCK, "%s: namespace LFSCK assistant found bad name "
2149                "entry for: parent "DFID", child "DFID", name %s, type "
2150                "in name entry %o, type claimed by child %o. repair it "
2151                "by %s with new name2 %s: rc = %d\n", lfsck_lfsck2name(lfsck),
2152                PFID(lfsck_dto2fid(parent)), PFID(lfsck_dto2fid(child)),
2153                name, type, update ? lfsck_object_type(child) : 0,
2154                update ? "updating" : "removing", name2, rc);
2155
2156         if (rc != 0) {
2157                 struct lfsck_namespace *ns = com->lc_file_ram;
2158
2159                 ns->ln_flags |= LF_INCONSISTENT;
2160         }
2161
2162         return rc;
2163 }
2164
2165 /**
2166  * Update the ".." name entry for the given object.
2167  *
2168  * The object's ".." is corrupted, this function will update the ".." name
2169  * entry with the given pfid, and the linkEA with the given ldata.
2170  *
2171  * The caller should take the ldlm lock before the calling.
2172  *
2173  * \param[in] env       pointer to the thread context
2174  * \param[in] com       pointer to the lfsck component
2175  * \param[in] obj       pointer to the dt_object to be handled
2176  * \param[in] pfid      the new fid for the object's ".." name entry
2177  * \param[in] cname     the name for the @obj in the parent directory
2178  *
2179  * \retval              positive number for repaired cases
2180  * \retval              0 if nothing to be repaired
2181  * \retval              negative error number on failure
2182  */
2183 static int lfsck_namespace_repair_unmatched_pairs(const struct lu_env *env,
2184                                                   struct lfsck_component *com,
2185                                                   struct dt_object *obj,
2186                                                   const struct lu_fid *pfid,
2187                                                   struct lu_name *cname)
2188 {
2189         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2190         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
2191         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2192         struct dt_device                *dev    = lfsck_obj2dev(obj);
2193         struct thandle                  *th     = NULL;
2194         struct linkea_data               ldata  = { NULL };
2195         struct lu_buf                    linkea_buf;
2196         int                              rc     = 0;
2197         ENTRY;
2198
2199         LASSERT(!dt_object_remote(obj));
2200         LASSERT(S_ISDIR(lfsck_object_type(obj)));
2201
2202         rc = linkea_data_new(&ldata, &info->lti_big_buf);
2203         if (rc != 0)
2204                 GOTO(log, rc);
2205
2206         rc = linkea_add_buf(&ldata, cname, pfid);
2207         if (rc != 0)
2208                 GOTO(log, rc);
2209
2210         lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
2211                        ldata.ld_leh->leh_len);
2212
2213         th = dt_trans_create(env, dev);
2214         if (IS_ERR(th))
2215                 GOTO(log, rc = PTR_ERR(th));
2216
2217         rc = dt_declare_delete(env, obj, (const struct dt_key *)dotdot, th);
2218         if (rc != 0)
2219                 GOTO(stop, rc);
2220
2221         rec->rec_type = S_IFDIR;
2222         rec->rec_fid = pfid;
2223         rc = dt_declare_insert(env, obj, (const struct dt_rec *)rec,
2224                                (const struct dt_key *)dotdot, th);
2225         if (rc != 0)
2226                 GOTO(stop, rc);
2227
2228         rc = dt_declare_xattr_set(env, obj, &linkea_buf,
2229                                   XATTR_NAME_LINK, 0, th);
2230         if (rc != 0)
2231                 GOTO(stop, rc);
2232
2233         rc = dt_trans_start_local(env, dev, th);
2234         if (rc != 0)
2235                 GOTO(stop, rc);
2236
2237         dt_write_lock(env, obj, 0);
2238         if (unlikely(lfsck_is_dead_obj(obj)))
2239                 GOTO(unlock, rc = 0);
2240
2241         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
2242                 GOTO(unlock, rc = 1);
2243
2244         /* The old ".." name entry maybe not exist. */
2245         dt_delete(env, obj, (const struct dt_key *)dotdot, th);
2246
2247         rc = dt_insert(env, obj, (const struct dt_rec *)rec,
2248                        (const struct dt_key *)dotdot, th, 1);
2249         if (rc != 0)
2250                 GOTO(unlock, rc);
2251
2252         rc = dt_xattr_set(env, obj, &linkea_buf,
2253                           XATTR_NAME_LINK, 0, th);
2254
2255         GOTO(unlock, rc = (rc == 0 ? 1 : rc));
2256
2257 unlock:
2258         dt_write_unlock(env, obj);
2259
2260 stop:
2261         dt_trans_stop(env, dev, th);
2262
2263 log:
2264         CDEBUG(D_LFSCK, "%s: namespace LFSCK rebuild dotdot name entry for "
2265                "the object "DFID", new parent "DFID": rc = %d\n",
2266                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(obj)),
2267                PFID(pfid), rc);
2268
2269         if (rc != 0) {
2270                 struct lfsck_namespace *ns = com->lc_file_ram;
2271
2272                 ns->ln_flags |= LF_INCONSISTENT;
2273         }
2274
2275         return rc;
2276 }
2277
2278 /**
2279  * Handle orphan @obj during Double Scan Directory.
2280  *
2281  * Remove the @obj's current (invalid) linkEA entries, and insert
2282  * it in the directory .lustre/lost+found/MDTxxxx/ with the name:
2283  * ${FID}-${PFID}-D-${conflict_version}
2284  *
2285  * The caller should take the ldlm lock before the calling.
2286  *
2287  * \param[in] env       pointer to the thread context
2288  * \param[in] com       pointer to the lfsck component
2289  * \param[in] obj       pointer to the orphan object to be handled
2290  * \param[in] pfid      the new fid for the object's ".." name entry
2291  * \param[in,out] lh    ldlm lock handler for the given @obj
2292  * \param[out] type     to tell the caller what the inconsistency is
2293  *
2294  * \retval              positive number for repaired cases
2295  * \retval              0 if nothing to be repaired
2296  * \retval              negative error number on failure
2297  */
2298 static int
2299 lfsck_namespace_dsd_orphan(const struct lu_env *env,
2300                            struct lfsck_component *com,
2301                            struct dt_object *obj,
2302                            const struct lu_fid *pfid,
2303                            struct lustre_handle *lh,
2304                            enum lfsck_namespace_inconsistency_type *type)
2305 {
2306         struct lfsck_thread_info *info = lfsck_env_info(env);
2307         struct lfsck_namespace   *ns   = com->lc_file_ram;
2308         int                       rc;
2309         ENTRY;
2310
2311         /* Remove the unrecognized linkEA. */
2312         rc = lfsck_namespace_links_remove(env, com, obj);
2313         lfsck_ibits_unlock(lh, LCK_EX);
2314         if (rc < 0 && rc != -ENODATA)
2315                 RETURN(rc);
2316
2317         *type = LNIT_MUL_REF;
2318
2319         /* If the LFSCK is marked as LF_INCOMPLETE, then means some MDT has
2320          * ever tried to verify some remote MDT-object that resides on this
2321          * MDT, but this MDT failed to respond such request. So means there
2322          * may be some remote name entry on other MDT that references this
2323          * object with another name, so we cannot know whether this linkEA
2324          * is valid or not. So keep it there and maybe resolved when next
2325          * LFSCK run. */
2326         if (ns->ln_flags & LF_INCOMPLETE)
2327                 RETURN(0);
2328
2329         /* The unique linkEA is invalid, even if the ".." name entry may be
2330          * valid, we still cannot know via which name entry this directory
2331          * will be referenced. Then handle it as pure orphan. */
2332         snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf),
2333                  "-"DFID, PFID(pfid));
2334         rc = lfsck_namespace_insert_orphan(env, com, obj,
2335                                            info->lti_tmpbuf, "D", NULL);
2336
2337         RETURN(rc);
2338 }
2339
2340 /**
2341  * Double Scan Directory object for single linkEA entry case.
2342  *
2343  * The given @child has unique linkEA entry. If the linkEA entry is valid,
2344  * then check whether the name is in the namespace or not, if not, add the
2345  * missing name entry back to namespace. If the linkEA entry is invalid,
2346  * then remove it and insert the @child in the .lustre/lost+found/MDTxxxx/
2347  * as an orphan.
2348  *
2349  * \param[in] env       pointer to the thread context
2350  * \param[in] com       pointer to the lfsck component
2351  * \param[in] child     pointer to the directory to be double scanned
2352  * \param[in] pfid      the FID corresponding to the ".." entry
2353  * \param[in] ldata     pointer to the linkEA data for the given @child
2354  * \param[in,out] lh    ldlm lock handler for the given @child
2355  * \param[out] type     to tell the caller what the inconsistency is
2356  * \param[in] retry     if found inconsistency, but the caller does not hold
2357  *                      ldlm lock on the @child, then set @retry as true
2358  *
2359  * \retval              positive number for repaired cases
2360  * \retval              0 if nothing to be repaired
2361  * \retval              negative error number on failure
2362  */
2363 static int
2364 lfsck_namespace_dsd_single(const struct lu_env *env,
2365                            struct lfsck_component *com,
2366                            struct dt_object *child,
2367                            const struct lu_fid *pfid,
2368                            struct linkea_data *ldata,
2369                            struct lustre_handle *lh,
2370                            enum lfsck_namespace_inconsistency_type *type,
2371                            bool *retry)
2372 {
2373         struct lfsck_thread_info *info          = lfsck_env_info(env);
2374         struct lu_name           *cname         = &info->lti_name;
2375         const struct lu_fid      *cfid          = lfsck_dto2fid(child);
2376         struct lu_fid             tfid;
2377         struct lfsck_namespace   *ns            = com->lc_file_ram;
2378         struct lfsck_instance    *lfsck         = com->lc_lfsck;
2379         struct dt_object         *parent        = NULL;
2380         struct lmv_mds_md_v1     *lmv;
2381         int                       rc            = 0;
2382         ENTRY;
2383
2384         lfsck_namespace_unpack_linkea_entry(ldata, cname, &tfid, info->lti_key);
2385         /* The unique linkEA entry with bad parent will be handled as orphan. */
2386         if (!fid_is_sane(&tfid)) {
2387                 if (!lustre_handle_is_used(lh) && retry != NULL)
2388                         *retry = true;
2389                 else
2390                         rc = lfsck_namespace_dsd_orphan(env, com, child,
2391                                                         pfid, lh, type);
2392
2393                 GOTO(out, rc);
2394         }
2395
2396         parent = lfsck_object_find_bottom(env, lfsck, &tfid);
2397         if (IS_ERR(parent))
2398                 GOTO(out, rc = PTR_ERR(parent));
2399
2400         /* We trust the unique linkEA entry in spite of whether it matches the
2401          * ".." name entry or not. Because even if the linkEA entry is wrong
2402          * and the ".." name entry is right, we still cannot know via which
2403          * name entry the child will be referenced, since all known entries
2404          * have been verified during the first-stage scanning. */
2405         if (!dt_object_exists(parent)) {
2406                 /* If the LFSCK is marked as LF_INCOMPLETE, then means some MDT
2407                  * has ever tried to verify some remote MDT-object that resides
2408                  * on this MDT, but this MDT failed to respond such request. So
2409                  * means there may be some remote name entry on other MDT that
2410                  * references this object with another name, so we cannot know
2411                  * whether this linkEA is valid or not. So keep it there and
2412                  * maybe resolved when next LFSCK run. */
2413                 if (ns->ln_flags & LF_INCOMPLETE)
2414                         GOTO(out, rc = 0);
2415
2416                 if (!lustre_handle_is_used(lh) && retry != NULL) {
2417                         *retry = true;
2418
2419                         GOTO(out, rc = 0);
2420                 }
2421
2422                 lfsck_ibits_unlock(lh, LCK_EX);
2423
2424 lost_parent:
2425                 lmv = &info->lti_lmv;
2426                 rc = lfsck_read_stripe_lmv(env, child, lmv);
2427                 if (rc != 0 && rc != -ENODATA)
2428                         GOTO(out, rc);
2429
2430                 if (rc == -ENODATA || lmv->lmv_magic != LMV_MAGIC_STRIPE) {
2431                         lmv = NULL;
2432                 } else if (lfsck_shard_name_to_index(env,
2433                                         cname->ln_name, cname->ln_namelen,
2434                                         S_IFDIR, cfid) < 0) {
2435                         /* It is an invalid name entry, we
2436                          * cannot trust the parent also. */
2437                         rc = lfsck_namespace_shrink_linkea(env, com, child,
2438                                                 ldata, cname, &tfid, true);
2439                         if (rc < 0)
2440                                 GOTO(out, rc);
2441
2442                         snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf),
2443                                  "-"DFID, PFID(pfid));
2444                         rc = lfsck_namespace_insert_orphan(env, com, child,
2445                                                 info->lti_tmpbuf, "S", NULL);
2446
2447                         GOTO(out, rc);
2448                 }
2449
2450                 /* Create the lost parent as an orphan. */
2451                 rc = lfsck_namespace_create_orphan_dir(env, com, parent, lmv);
2452                 if (rc >= 0) {
2453                         /* Add the missing name entry to the parent. */
2454                         rc = lfsck_namespace_insert_normal(env, com, parent,
2455                                                         child, cname->ln_name);
2456                         if (unlikely(rc == -EEXIST)) {
2457                                 /* Unfortunately, someone reused the name
2458                                  * under the parent by race. So we have
2459                                  * to remove the linkEA entry from
2460                                  * current child object. It means that the
2461                                  * LFSCK cannot recover the system
2462                                  * totally back to its original status,
2463                                  * but it is necessary to make the
2464                                  * current system to be consistent. */
2465                                 rc = lfsck_namespace_shrink_linkea(env,
2466                                                 com, child, ldata,
2467                                                 cname, &tfid, true);
2468                                 if (rc >= 0) {
2469                                         snprintf(info->lti_tmpbuf,
2470                                                  sizeof(info->lti_tmpbuf),
2471                                                  "-"DFID, PFID(pfid));
2472                                         rc = lfsck_namespace_insert_orphan(env,
2473                                                 com, child, info->lti_tmpbuf,
2474                                                 "D", NULL);
2475                                 }
2476                         }
2477                 }
2478
2479                 GOTO(out, rc);
2480         }
2481
2482         /* The unique linkEA entry with bad parent will be handled as orphan. */
2483         if (unlikely(!dt_try_as_dir(env, parent))) {
2484                 if (!lustre_handle_is_used(lh) && retry != NULL)
2485                         *retry = true;
2486                 else
2487                         rc = lfsck_namespace_dsd_orphan(env, com, child,
2488                                                         pfid, lh, type);
2489
2490                 GOTO(out, rc);
2491         }
2492
2493         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
2494                        (const struct dt_key *)cname->ln_name);
2495         if (rc == -ENOENT) {
2496                 /* If the LFSCK is marked as LF_INCOMPLETE, then means some MDT
2497                  * has ever tried to verify some remote MDT-object that resides
2498                  * on this MDT, but this MDT failed to respond such request. So
2499                  * means there may be some remote name entry on other MDT that
2500                  * references this object with another name, so we cannot know
2501                  * whether this linkEA is valid or not. So keep it there and
2502                  * maybe resolved when next LFSCK run. */
2503                 if (ns->ln_flags & LF_INCOMPLETE)
2504                         GOTO(out, rc = 0);
2505
2506                 if (!lustre_handle_is_used(lh) && retry != NULL) {
2507                         *retry = true;
2508
2509                         GOTO(out, rc = 0);
2510                 }
2511
2512                 lfsck_ibits_unlock(lh, LCK_EX);
2513                 rc = lfsck_namespace_check_name(env, parent, child, cname);
2514                 if (rc == -ENOENT)
2515                         goto lost_parent;
2516
2517                 if (rc < 0)
2518                         GOTO(out, rc);
2519
2520                 /* It is an invalid name entry, drop it. */
2521                 if (unlikely(rc > 0)) {
2522                         rc = lfsck_namespace_shrink_linkea(env, com, child,
2523                                                 ldata, cname, &tfid, true);
2524                         if (rc >= 0) {
2525                                 snprintf(info->lti_tmpbuf,
2526                                          sizeof(info->lti_tmpbuf),
2527                                          "-"DFID, PFID(pfid));
2528                                 rc = lfsck_namespace_insert_orphan(env, com,
2529                                         child, info->lti_tmpbuf, "D", NULL);
2530                         }
2531
2532                         GOTO(out, rc);
2533                 }
2534
2535                 /* Add the missing name entry back to the namespace. */
2536                 rc = lfsck_namespace_insert_normal(env, com, parent, child,
2537                                                    cname->ln_name);
2538                 if (unlikely(rc == -ESTALE))
2539                         /* It may happen when the remote object has been
2540                          * removed, but the local MDT is not aware of that. */
2541                         goto lost_parent;
2542
2543                 if (unlikely(rc == -EEXIST)) {
2544                         /* Unfortunately, someone reused the name under the
2545                          * parent by race. So we have to remove the linkEA
2546                          * entry from current child object. It means that the
2547                          * LFSCK cannot recover the system totally back to
2548                          * its original status, but it is necessary to make
2549                          * the current system to be consistent.
2550                          *
2551                          * It also may be because of the LFSCK found some
2552                          * internal status of create operation. Under such
2553                          * case, nothing to be done. */
2554                         rc = lfsck_namespace_shrink_linkea_cond(env, com,
2555                                         parent, child, ldata, cname, &tfid);
2556                         if (rc >= 0) {
2557                                 snprintf(info->lti_tmpbuf,
2558                                          sizeof(info->lti_tmpbuf),
2559                                          "-"DFID, PFID(pfid));
2560                                 rc = lfsck_namespace_insert_orphan(env, com,
2561                                         child, info->lti_tmpbuf, "D", NULL);
2562                         }
2563                 }
2564
2565                 GOTO(out, rc);
2566         }
2567
2568         if (rc != 0)
2569                 GOTO(out, rc);
2570
2571         if (!lu_fid_eq(&tfid, cfid)) {
2572                 if (!lustre_handle_is_used(lh) && retry != NULL) {
2573                         *retry = true;
2574
2575                         GOTO(out, rc = 0);
2576                 }
2577
2578                 lfsck_ibits_unlock(lh, LCK_EX);
2579                 /* The name entry references another MDT-object that
2580                  * may be created by the LFSCK for repairing dangling
2581                  * name entry. Try to replace it. */
2582                 rc = lfsck_namespace_replace_cond(env, com, parent, child,
2583                                                   &tfid, cname);
2584                 if (rc == 0)
2585                         rc = lfsck_namespace_dsd_orphan(env, com, child,
2586                                                         pfid, lh, type);
2587
2588                 GOTO(out, rc);
2589         }
2590
2591         if (fid_is_zero(pfid))
2592                 GOTO(out, rc = 0);
2593
2594         /* The ".." name entry is wrong, update it. */
2595         if (!lu_fid_eq(pfid, lfsck_dto2fid(parent))) {
2596                 if (!lustre_handle_is_used(lh) && retry != NULL) {
2597                         *retry = true;
2598
2599                         GOTO(out, rc = 0);
2600                 }
2601
2602                 *type = LNIT_UNMATCHED_PAIRS;
2603                 rc = lfsck_namespace_repair_unmatched_pairs(env, com, child,
2604                                                 lfsck_dto2fid(parent), cname);
2605         }
2606
2607         GOTO(out, rc);
2608
2609 out:
2610         if (parent != NULL && !IS_ERR(parent))
2611                 lfsck_object_put(env, parent);
2612
2613         return rc;
2614 }
2615
2616 /**
2617  * Double Scan Directory object for multiple linkEA entries case.
2618  *
2619  * The given @child has multiple linkEA entries. There is at most one linkEA
2620  * entry will be valid, all the others will be removed. Firstly, the function
2621  * will try to find out the linkEA entry for which the name entry exists under
2622  * the given parent (@pfid). If there is no linkEA entry that matches the given
2623  * ".." name entry, then tries to find out the first linkEA entry that both the
2624  * parent and the name entry exist to rebuild a new ".." name entry.
2625  *
2626  * \param[in] env       pointer to the thread context
2627  * \param[in] com       pointer to the lfsck component
2628  * \param[in] child     pointer to the directory to be double scanned
2629  * \param[in] pfid      the FID corresponding to the ".." entry
2630  * \param[in] ldata     pointer to the linkEA data for the given @child
2631  * \param[in,out] lh    ldlm lock handler for the given @child
2632  * \param[out] type     to tell the caller what the inconsistency is
2633  * \param[in] lpf       true if the ".." entry is under lost+found/MDTxxxx/
2634  *
2635  * \retval              positive number for repaired cases
2636  * \retval              0 if nothing to be repaired
2637  * \retval              negative error number on failure
2638  */
2639 static int
2640 lfsck_namespace_dsd_multiple(const struct lu_env *env,
2641                              struct lfsck_component *com,
2642                              struct dt_object *child,
2643                              const struct lu_fid *pfid,
2644                              struct linkea_data *ldata,
2645                              struct lustre_handle *lh,
2646                              enum lfsck_namespace_inconsistency_type *type,
2647                              bool lpf)
2648 {
2649         struct lfsck_thread_info *info          = lfsck_env_info(env);
2650         struct lu_name           *cname         = &info->lti_name;
2651         const struct lu_fid      *cfid          = lfsck_dto2fid(child);
2652         struct lu_fid            *pfid2         = &info->lti_fid3;
2653         struct lu_fid             tfid;
2654         struct lfsck_namespace   *ns            = com->lc_file_ram;
2655         struct lfsck_instance    *lfsck         = com->lc_lfsck;
2656         struct lfsck_bookmark    *bk            = &lfsck->li_bookmark_ram;
2657         struct dt_object         *parent        = NULL;
2658         struct linkea_data        ldata_new     = { NULL };
2659         int                       dirent_count  = 0;
2660         int                       linkea_count  = 0;
2661         int                       rc            = 0;
2662         bool                      once          = true;
2663         ENTRY;
2664
2665 again:
2666         while (ldata->ld_lee != NULL) {
2667                 lfsck_namespace_unpack_linkea_entry(ldata, cname, &tfid,
2668                                                     info->lti_key);
2669                 /* Drop repeated linkEA entries. */
2670                 lfsck_namespace_filter_linkea_entry(ldata, cname, &tfid, true);
2671                 /* Drop invalid linkEA entry. */
2672                 if (!fid_is_sane(&tfid)) {
2673                         linkea_del_buf(ldata, cname);
2674                         linkea_count++;
2675                         continue;
2676                 }
2677
2678                 /* If current dotdot is the .lustre/lost+found/MDTxxxx/,
2679                  * then it is possible that: the directry object has ever
2680                  * been lost, but its name entry was there. In the former
2681                  * LFSCK run, during the first-stage scanning, the LFSCK
2682                  * found the dangling name entry, but it did not recreate
2683                  * the lost object, and when moved to the second-stage
2684                  * scanning, some children objects of the lost directory
2685                  * object were found, then the LFSCK recreated such lost
2686                  * directory object as an orphan.
2687                  *
2688                  * When the LFSCK runs again, if the dangling name is still
2689                  * there, the LFSCK should move the orphan directory object
2690                  * back to the normal namespace. */
2691                 if (!lpf && !lu_fid_eq(pfid, &tfid) && once) {
2692                         linkea_next_entry(ldata);
2693                         continue;
2694                 }
2695
2696                 parent = lfsck_object_find_bottom(env, lfsck, &tfid);
2697                 if (IS_ERR(parent))
2698                         RETURN(PTR_ERR(parent));
2699
2700                 if (!dt_object_exists(parent)) {
2701                         lfsck_object_put(env, parent);
2702                         if (ldata->ld_leh->leh_reccount > 1) {
2703                                 /* If it is NOT the last linkEA entry, then
2704                                  * there is still other chance to make the
2705                                  * child to be visible via other parent, then
2706                                  * remove this linkEA entry. */
2707                                 linkea_del_buf(ldata, cname);
2708                                 linkea_count++;
2709                                 continue;
2710                         }
2711
2712                         break;
2713                 }
2714
2715                 /* The linkEA entry with bad parent will be removed. */
2716                 if (unlikely(!dt_try_as_dir(env, parent))) {
2717                         lfsck_object_put(env, parent);
2718                         linkea_del_buf(ldata, cname);
2719                         linkea_count++;
2720                         continue;
2721                 }
2722
2723                 rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
2724                                (const struct dt_key *)cname->ln_name);
2725                 *pfid2 = *lfsck_dto2fid(parent);
2726                 if (rc == -ENOENT) {
2727                         lfsck_object_put(env, parent);
2728                         linkea_next_entry(ldata);
2729                         continue;
2730                 }
2731
2732                 if (rc != 0) {
2733                         lfsck_object_put(env, parent);
2734
2735                         RETURN(rc);
2736                 }
2737
2738                 if (lu_fid_eq(&tfid, cfid)) {
2739                         lfsck_object_put(env, parent);
2740                         if (!lu_fid_eq(pfid, pfid2)) {
2741                                 *type = LNIT_UNMATCHED_PAIRS;
2742                                 rc = lfsck_namespace_repair_unmatched_pairs(env,
2743                                                 com, child, pfid2, cname);
2744
2745                                 RETURN(rc);
2746                         }
2747
2748 rebuild:
2749                         /* It is the most common case that we find the
2750                          * name entry corresponding to the linkEA entry
2751                          * that matches the ".." name entry. */
2752                         rc = linkea_data_new(&ldata_new, &info->lti_big_buf);
2753                         if (rc != 0)
2754                                 RETURN(rc);
2755
2756                         rc = linkea_add_buf(&ldata_new, cname, pfid2);
2757                         if (rc != 0)
2758                                 RETURN(rc);
2759
2760                         rc = lfsck_namespace_rebuild_linkea(env, com, child,
2761                                                             &ldata_new);
2762                         if (rc < 0)
2763                                 RETURN(rc);
2764
2765                         linkea_del_buf(ldata, cname);
2766                         linkea_count++;
2767                         linkea_first_entry(ldata);
2768                         /* There may be some invalid dangling name entries under
2769                          * other parent directories, remove all of them. */
2770                         while (ldata->ld_lee != NULL) {
2771                                 lfsck_namespace_unpack_linkea_entry(ldata,
2772                                                 cname, &tfid, info->lti_key);
2773                                 if (!fid_is_sane(&tfid))
2774                                         goto next;
2775
2776                                 parent = lfsck_object_find_bottom(env, lfsck,
2777                                                                   &tfid);
2778                                 if (IS_ERR(parent)) {
2779                                         rc = PTR_ERR(parent);
2780                                         if (rc != -ENOENT &&
2781                                             bk->lb_param & LPF_FAILOUT)
2782                                                 RETURN(rc);
2783
2784                                         goto next;
2785                                 }
2786
2787                                 if (!dt_object_exists(parent)) {
2788                                         lfsck_object_put(env, parent);
2789                                         goto next;
2790                                 }
2791
2792                                 rc = lfsck_namespace_repair_dirent(env, com,
2793                                         parent, child, cname->ln_name,
2794                                         cname->ln_name, S_IFDIR, false, true);
2795                                 lfsck_object_put(env, parent);
2796                                 if (rc < 0) {
2797                                         if (bk->lb_param & LPF_FAILOUT)
2798                                                 RETURN(rc);
2799
2800                                         goto next;
2801                                 }
2802
2803                                 dirent_count += rc;
2804
2805 next:
2806                                 linkea_del_buf(ldata, cname);
2807                         }
2808
2809                         ns->ln_dirent_repaired += dirent_count;
2810
2811                         RETURN(rc);
2812                 }
2813
2814                 lfsck_ibits_unlock(lh, LCK_EX);
2815                 /* The name entry references another MDT-object that may be
2816                  * created by the LFSCK for repairing dangling name entry.
2817                  * Try to replace it. */
2818                 rc = lfsck_namespace_replace_cond(env, com, parent, child,
2819                                                   &tfid, cname);
2820                 lfsck_object_put(env, parent);
2821                 if (rc < 0)
2822                         RETURN(rc);
2823
2824                 if (rc > 0)
2825                         goto rebuild;
2826
2827                 linkea_del_buf(ldata, cname);
2828         }
2829
2830         linkea_first_entry(ldata);
2831         if (ldata->ld_leh->leh_reccount == 1) {
2832                 rc = lfsck_namespace_dsd_single(env, com, child, pfid, ldata,
2833                                                 lh, type, NULL);
2834
2835                 if (rc == 0 && fid_is_zero(pfid) && linkea_count > 0)
2836                         rc = lfsck_namespace_rebuild_linkea(env, com, child,
2837                                                             ldata);
2838
2839                 RETURN(rc);
2840         }
2841
2842         /* All linkEA entries are invalid and removed, then handle the @child
2843          * as an orphan.*/
2844         if (ldata->ld_leh->leh_reccount == 0) {
2845                 rc = lfsck_namespace_dsd_orphan(env, com, child, pfid, lh,
2846                                                 type);
2847
2848                 RETURN(rc);
2849         }
2850
2851         /* If the dangling name entry for the orphan directory object has
2852          * been remvoed, then just check whether the directory object is
2853          * still under the .lustre/lost+found/MDTxxxx/ or not. */
2854         if (lpf) {
2855                 lpf = false;
2856                 goto again;
2857         }
2858
2859         /* There is no linkEA entry that matches the ".." name entry. Find
2860          * the first linkEA entry that both parent and name entry exist to
2861          * rebuild a new ".." name entry. */
2862         if (once) {
2863                 once = false;
2864                 goto again;
2865         }
2866
2867         RETURN(rc);
2868 }
2869
2870 /**
2871  * Repair the object's nlink attribute.
2872  *
2873  * If all the known name entries have been verified, then the object's hard
2874  * link attribute should match the object's linkEA entries count unless the
2875  * object's has too much hard link to be recorded in the linkEA. Such cases
2876  * should have been marked in the LFSCK trace file. Otherwise, trust the
2877  * linkEA to update the object's nlink attribute.
2878  *
2879  * \param[in] env       pointer to the thread context
2880  * \param[in] com       pointer to the lfsck component
2881  * \param[in] obj       pointer to the dt_object to be handled
2882  * \param[in,out] la    pointer to buffer to object's attribute before
2883  *                      and after the repairing
2884  *
2885  * \retval              positive number for repaired cases
2886  * \retval              0 if nothing to be repaired
2887  * \retval              negative error number on failure
2888  */
2889 static int lfsck_namespace_repair_nlink(const struct lu_env *env,
2890                                         struct lfsck_component *com,
2891                                         struct dt_object *obj,
2892                                         struct lu_attr *la)
2893 {
2894         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2895         struct lu_fid                   *tfid   = &info->lti_fid3;
2896         struct lfsck_namespace          *ns     = com->lc_file_ram;
2897         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2898         struct dt_device                *dev    = lfsck_obj2dev(obj);
2899         const struct lu_fid             *cfid   = lfsck_dto2fid(obj);
2900         struct thandle                  *th     = NULL;
2901         struct linkea_data               ldata  = { NULL };
2902         struct lustre_handle             lh     = { 0 };
2903         __u32                            old    = la->la_nlink;
2904         int                              idx;
2905         int                              rc     = 0;
2906         __u8                             flags;
2907         ENTRY;
2908
2909         LASSERT(!dt_object_remote(obj));
2910         LASSERT(S_ISREG(lfsck_object_type(obj)));
2911
2912         rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
2913                               MDS_INODELOCK_UPDATE, LCK_PW);
2914         if (rc != 0)
2915                 GOTO(log, rc);
2916
2917         th = dt_trans_create(env, dev);
2918         if (IS_ERR(th))
2919                 GOTO(log, rc = PTR_ERR(th));
2920
2921         la->la_valid = LA_NLINK;
2922         rc = dt_declare_attr_set(env, obj, la, th);
2923         if (rc != 0)
2924                 GOTO(stop, rc);
2925
2926         rc = dt_trans_start_local(env, dev, th);
2927         if (rc != 0)
2928                 GOTO(stop, rc);
2929
2930         dt_write_lock(env, obj, 0);
2931         /* If the LFSCK is marked as LF_INCOMPLETE, then means some MDT has
2932          * ever tried to verify some remote MDT-object that resides on this
2933          * MDT, but this MDT failed to respond such request. So means there
2934          * may be some remote name entry on other MDT that references this
2935          * object with another name, so we cannot know whether this linkEA
2936          * is valid or not. So keep it there and maybe resolved when next
2937          * LFSCK run. */
2938         if (ns->ln_flags & LF_INCOMPLETE)
2939                 GOTO(unlock, rc = 0);
2940
2941         fid_cpu_to_be(tfid, cfid);
2942         idx = lfsck_sub_trace_file_fid2idx(cfid);
2943         rc = dt_lookup(env, com->lc_sub_trace_objs[idx].lsto_obj,
2944                        (struct dt_rec *)&flags, (const struct dt_key *)tfid);
2945         if (rc != 0)
2946                 GOTO(unlock, rc);
2947
2948         if (flags & LNTF_SKIP_NLINK)
2949                 GOTO(unlock, rc = 0);
2950
2951         rc = dt_attr_get(env, obj, la);
2952         if (rc != 0)
2953                 GOTO(unlock, rc = (rc == -ENOENT ? 0 : rc));
2954
2955         rc = lfsck_links_read2(env, obj, &ldata);
2956         if (rc != 0)
2957                 GOTO(unlock, rc = (rc == -ENODATA ? 0 : rc));
2958
2959         if (la->la_nlink == ldata.ld_leh->leh_reccount ||
2960             unlikely(la->la_nlink == 0))
2961                 GOTO(unlock, rc = 0);
2962
2963         la->la_nlink = ldata.ld_leh->leh_reccount;
2964         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
2965                 GOTO(unlock, rc = 1);
2966
2967         rc = dt_attr_set(env, obj, la, th);
2968
2969         GOTO(unlock, rc = (rc == 0 ? 1 : rc));
2970
2971 unlock:
2972         dt_write_unlock(env, obj);
2973
2974 stop:
2975         dt_trans_stop(env, dev, th);
2976
2977 log:
2978         lfsck_ibits_unlock(&lh, LCK_PW);
2979         CDEBUG(D_LFSCK, "%s: namespace LFSCK repaired the object "DFID"'s "
2980                "nlink count from %u to %u: rc = %d\n",
2981                lfsck_lfsck2name(lfsck), PFID(cfid), old, la->la_nlink, rc);
2982
2983         if (rc != 0)
2984                 ns->ln_flags |= LF_INCONSISTENT;
2985
2986         return rc;
2987 }
2988
2989 /**
2990  * Double scan the directory object for namespace LFSCK.
2991  *
2992  * This function will verify the <parent, child> pairs in the namespace tree:
2993  * the parent references the child via some name entry that should be in the
2994  * child's linkEA entry, the child should back references the parent via its
2995  * ".." name entry.
2996  *
2997  * The LFSCK will scan every linkEA entry in turn until find out the first
2998  * matched pairs. If found, then all other linkEA entries will be dropped.
2999  * If all the linkEA entries cannot match the ".." name entry, then there
3000  * are serveral possible cases:
3001  *
3002  * 1) If there is only one linkEA entry, then trust it as long as the PFID
3003  *    in the linkEA entry is valid.
3004  *
3005  * 2) If there are multiple linkEA entries, then try to find the linkEA
3006  *    that matches the ".." name entry. If found, then all other entries
3007  *    are invalid; otherwise, it is quite possible that the ".." name entry
3008  *    is corrupted. Under such case, the LFSCK will rebuild the ".." name
3009  *    entry according to the first valid linkEA entry (both the parent and
3010  *    the name entry should exist).
3011  *
3012  * 3) If the directory object has no (valid) linkEA entry, then the
3013  *    directory object will be handled as pure orphan and inserted
3014  *    in the .lustre/lost+found/MDTxxxx/ with the name:
3015  *    ${self_FID}-${PFID}-D-${conflict_version}
3016  *
3017  * \param[in] env       pointer to the thread context
3018  * \param[in] com       pointer to the lfsck component
3019  * \param[in] child     pointer to the directory object to be handled
3020  * \param[in] flags     to indicate the specical checking on the @child
3021  *
3022  * \retval              positive number for repaired cases
3023  * \retval              0 if nothing to be repaired
3024  * \retval              negative error number on failure
3025  */
3026 static int lfsck_namespace_double_scan_dir(const struct lu_env *env,
3027                                            struct lfsck_component *com,
3028                                            struct dt_object *child, __u8 flags)
3029 {
3030         struct lfsck_thread_info *info          = lfsck_env_info(env);
3031         const struct lu_fid      *cfid          = lfsck_dto2fid(child);
3032         struct lu_fid            *pfid          = &info->lti_fid2;
3033         struct lfsck_namespace   *ns            = com->lc_file_ram;
3034         struct lfsck_instance    *lfsck         = com->lc_lfsck;
3035         struct lustre_handle      lh            = { 0 };
3036         struct linkea_data        ldata         = { NULL };
3037         bool                      unknown       = false;