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, 2017, 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 <lu_object.h>
  34 #include <dt_object.h>
  35 #include <md_object.h>
  36 #include <lustre_fid.h>
  37 #include <lustre_lib.h>
  38 #include <lustre_net.h>
  39
  40 #include "lfsck_internal.h"
  41
  42 #define LFSCK_NAMESPACE_MAGIC_V1        0xA0629D03
  43 #define LFSCK_NAMESPACE_MAGIC_V2        0xA0621A0B
  44 #define LFSCK_NAMESPACE_MAGIC_V3        0xA06249FF
  45
  46 /* For Lustre-2.x (x <= 6), the namespace LFSCK used LFSCK_NAMESPACE_MAGIC_V1
  47  * as the trace file magic. When downgrade to such old release, the old LFSCK
  48  * will not recognize the new LFSCK_NAMESPACE_MAGIC_V2 in the new trace file,
  49  * then it will reset the whole LFSCK, and will not cause start failure. The
  50  * similar case will happen when upgrade from such old release. */
  51 #define LFSCK_NAMESPACE_MAGIC           LFSCK_NAMESPACE_MAGIC_V3
  52
  53 enum lfsck_nameentry_check {
  54         LFSCK_NAMEENTRY_DEAD            = 1, /* The object has been unlinked. */
  55         LFSCK_NAMEENTRY_REMOVED         = 2, /* The entry has been removed. */
  56         LFSCK_NAMEENTRY_RECREATED       = 3, /* The entry has been recreated. */
  57 };
  58
  59 static struct lfsck_namespace_req *
  60 lfsck_namespace_assistant_req_init(struct lfsck_instance *lfsck,
  61                                    struct lfsck_assistant_object *lso,
  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_parent = lfsck_assistant_object_get(lso);
  74         lnr->lnr_lmv = lfsck_lmv_get(lfsck->li_lmv);
  75         lnr->lnr_fid = ent->lde_fid;
  76         lnr->lnr_dir_cookie = ent->lde_hash;
  77         lnr->lnr_attr = ent->lde_attrs;
  78         lnr->lnr_size = size;
  79         lnr->lnr_type = type;
  80         lnr->lnr_namelen = ent->lde_namelen;
  81         memcpy(lnr->lnr_name, ent->lde_name, ent->lde_namelen);
  82
  83         return lnr;
  84 }
  85
  86 static void lfsck_namespace_assistant_req_fini(const struct lu_env *env,
  87                                                struct lfsck_assistant_req *lar)
  88 {
  89         struct lfsck_namespace_req *lnr =
  90                 container_of(lar, struct lfsck_namespace_req, lnr_lar);
  91
  92         if (lnr->lnr_lmv != NULL)
  93                 lfsck_lmv_put(env, lnr->lnr_lmv);
  94
  95         lfsck_assistant_object_put(env, lar->lar_parent);
  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 = le64_to_cpu(src->ln_run_time_phase1);
 107         dst->ln_run_time_phase2 = le64_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         dst->ln_time_latest_reset = le64_to_cpu(src->ln_time_latest_reset);
 167         dst->ln_linkea_overflow_cleared =
 168                                 le64_to_cpu(src->ln_linkea_overflow_cleared);
 169         dst->ln_agent_entries_repaired =
 170                                 le64_to_cpu(src->ln_agent_entries_repaired);
 171 }
 172
 173 static void lfsck_namespace_cpu_to_le(struct lfsck_namespace *dst,
 174                                       struct lfsck_namespace *src)
 175 {
 176         dst->ln_magic = cpu_to_le32(src->ln_magic);
 177         dst->ln_status = cpu_to_le32(src->ln_status);
 178         dst->ln_flags = cpu_to_le32(src->ln_flags);
 179         dst->ln_success_count = cpu_to_le32(src->ln_success_count);
 180         dst->ln_run_time_phase1 = cpu_to_le64(src->ln_run_time_phase1);
 181         dst->ln_run_time_phase2 = cpu_to_le64(src->ln_run_time_phase2);
 182         dst->ln_time_last_complete = cpu_to_le64(src->ln_time_last_complete);
 183         dst->ln_time_latest_start = cpu_to_le64(src->ln_time_latest_start);
 184         dst->ln_time_last_checkpoint =
 185                                 cpu_to_le64(src->ln_time_last_checkpoint);
 186         lfsck_position_cpu_to_le(&dst->ln_pos_latest_start,
 187                                  &src->ln_pos_latest_start);
 188         lfsck_position_cpu_to_le(&dst->ln_pos_last_checkpoint,
 189                                  &src->ln_pos_last_checkpoint);
 190         lfsck_position_cpu_to_le(&dst->ln_pos_first_inconsistent,
 191                                  &src->ln_pos_first_inconsistent);
 192         dst->ln_items_checked = cpu_to_le64(src->ln_items_checked);
 193         dst->ln_items_repaired = cpu_to_le64(src->ln_items_repaired);
 194         dst->ln_items_failed = cpu_to_le64(src->ln_items_failed);
 195         dst->ln_dirs_checked = cpu_to_le64(src->ln_dirs_checked);
 196         dst->ln_objs_checked_phase2 = cpu_to_le64(src->ln_objs_checked_phase2);
 197         dst->ln_objs_repaired_phase2 =
 198                                 cpu_to_le64(src->ln_objs_repaired_phase2);
 199         dst->ln_objs_failed_phase2 = cpu_to_le64(src->ln_objs_failed_phase2);
 200         dst->ln_objs_nlink_repaired = cpu_to_le64(src->ln_objs_nlink_repaired);
 201         fid_cpu_to_le(&dst->ln_fid_latest_scanned_phase2,
 202                       &src->ln_fid_latest_scanned_phase2);
 203         dst->ln_dirent_repaired = cpu_to_le64(src->ln_dirent_repaired);
 204         dst->ln_linkea_repaired = cpu_to_le64(src->ln_linkea_repaired);
 205         dst->ln_mul_linked_checked = cpu_to_le64(src->ln_mul_linked_checked);
 206         dst->ln_mul_linked_repaired = cpu_to_le64(src->ln_mul_linked_repaired);
 207         dst->ln_unknown_inconsistency =
 208                                 cpu_to_le64(src->ln_unknown_inconsistency);
 209         dst->ln_unmatched_pairs_repaired =
 210                                 cpu_to_le64(src->ln_unmatched_pairs_repaired);
 211         dst->ln_dangling_repaired = cpu_to_le64(src->ln_dangling_repaired);
 212         dst->ln_mul_ref_repaired = cpu_to_le64(src->ln_mul_ref_repaired);
 213         dst->ln_bad_type_repaired = cpu_to_le64(src->ln_bad_type_repaired);
 214         dst->ln_lost_dirent_repaired =
 215                                 cpu_to_le64(src->ln_lost_dirent_repaired);
 216         dst->ln_striped_dirs_scanned =
 217                                 cpu_to_le64(src->ln_striped_dirs_scanned);
 218         dst->ln_striped_dirs_repaired =
 219                                 cpu_to_le64(src->ln_striped_dirs_repaired);
 220         dst->ln_striped_dirs_failed =
 221                                 cpu_to_le64(src->ln_striped_dirs_failed);
 222         dst->ln_striped_dirs_disabled =
 223                                 cpu_to_le64(src->ln_striped_dirs_disabled);
 224         dst->ln_striped_dirs_skipped =
 225                                 cpu_to_le64(src->ln_striped_dirs_skipped);
 226         dst->ln_striped_shards_scanned =
 227                                 cpu_to_le64(src->ln_striped_shards_scanned);
 228         dst->ln_striped_shards_repaired =
 229                                 cpu_to_le64(src->ln_striped_shards_repaired);
 230         dst->ln_striped_shards_failed =
 231                                 cpu_to_le64(src->ln_striped_shards_failed);
 232         dst->ln_striped_shards_skipped =
 233                                 cpu_to_le64(src->ln_striped_shards_skipped);
 234         dst->ln_name_hash_repaired = cpu_to_le64(src->ln_name_hash_repaired);
 235         dst->ln_local_lpf_scanned = cpu_to_le64(src->ln_local_lpf_scanned);
 236         dst->ln_local_lpf_moved = cpu_to_le64(src->ln_local_lpf_moved);
 237         dst->ln_local_lpf_skipped = cpu_to_le64(src->ln_local_lpf_skipped);
 238         dst->ln_local_lpf_failed = cpu_to_le64(src->ln_local_lpf_failed);
 239         dst->ln_bitmap_size = cpu_to_le32(src->ln_bitmap_size);
 240         dst->ln_time_latest_reset = cpu_to_le64(src->ln_time_latest_reset);
 241         dst->ln_linkea_overflow_cleared =
 242                                 cpu_to_le64(src->ln_linkea_overflow_cleared);
 243         dst->ln_agent_entries_repaired =
 244                                 cpu_to_le64(src->ln_agent_entries_repaired);
 245 }
 246
 247 static void lfsck_namespace_record_failure(const struct lu_env *env,
 248                                            struct lfsck_instance *lfsck,
 249                                            struct lfsck_namespace *ns)
 250 {
 251         struct lfsck_position pos;
 252
 253         ns->ln_items_failed++;
 254         lfsck_pos_fill(env, lfsck, &pos, false);
 255         if (lfsck_pos_is_zero(&ns->ln_pos_first_inconsistent) ||
 256             lfsck_pos_is_eq(&pos, &ns->ln_pos_first_inconsistent) < 0) {
 257                 ns->ln_pos_first_inconsistent = pos;
 258
 259                 CDEBUG(D_LFSCK, "%s: namespace LFSCK hit first non-repaired "
 260                        "inconsistency at the pos [%llu, "DFID", %#llx]\n",
 261                        lfsck_lfsck2name(lfsck),
 262                        ns->ln_pos_first_inconsistent.lp_oit_cookie,
 263                        PFID(&ns->ln_pos_first_inconsistent.lp_dir_parent),
 264                        ns->ln_pos_first_inconsistent.lp_dir_cookie);
 265         }
 266 }
 267
 268 /**
 269  * Load the MDT bitmap from the lfsck_namespace trace file.
 270  *
 271  * \param[in] env       pointer to the thread context
 272  * \param[in] com       pointer to the lfsck component
 273  *
 274  * \retval              0 for success
 275  * \retval              negative error number on failure or data corruption
 276  */
 277 static int lfsck_namespace_load_bitmap(const struct lu_env *env,
 278                                        struct lfsck_component *com)
 279 {
 280         struct dt_object                *obj    = com->lc_obj;
 281         struct lfsck_assistant_data     *lad    = com->lc_data;
 282         struct lfsck_namespace          *ns     = com->lc_file_ram;
 283         struct cfs_bitmap                       *bitmap = lad->lad_bitmap;
 284         ssize_t                          size;
 285         __u32                            nbits;
 286         int                              rc;
 287         ENTRY;
 288
 289         if (com->lc_lfsck->li_mdt_descs.ltd_tgts_bitmap->size >
 290             ns->ln_bitmap_size)
 291                 nbits = com->lc_lfsck->li_mdt_descs.ltd_tgts_bitmap->size;
 292         else
 293                 nbits = ns->ln_bitmap_size;
 294
 295         if (unlikely(nbits < BITS_PER_LONG))
 296                 nbits = BITS_PER_LONG;
 297
 298         if (nbits > bitmap->size) {
 299                 __u32 new_bits = bitmap->size;
 300                 struct cfs_bitmap *new_bitmap;
 301
 302                 while (new_bits < nbits)
 303                         new_bits <<= 1;
 304
 305                 new_bitmap = CFS_ALLOCATE_BITMAP(new_bits);
 306                 if (new_bitmap == NULL)
 307                         RETURN(-ENOMEM);
 308
 309                 lad->lad_bitmap = new_bitmap;
 310                 CFS_FREE_BITMAP(bitmap);
 311                 bitmap = new_bitmap;
 312         }
 313
 314         if (ns->ln_bitmap_size == 0) {
 315                 clear_bit(LAD_INCOMPLETE, &lad->lad_flags);
 316                 CFS_RESET_BITMAP(bitmap);
 317
 318                 RETURN(0);
 319         }
 320
 321         size = (ns->ln_bitmap_size + 7) >> 3;
 322         rc = dt_xattr_get(env, obj,
 323                           lfsck_buf_get(env, bitmap->data, size),
 324                           XATTR_NAME_LFSCK_BITMAP);
 325         if (rc != size)
 326                 RETURN(rc >= 0 ? -EINVAL : rc);
 327
 328         if (cfs_bitmap_check_empty(bitmap))
 329                 clear_bit(LAD_INCOMPLETE, &lad->lad_flags);
 330         else
 331                 set_bit(LAD_INCOMPLETE, &lad->lad_flags);
 332
 333         RETURN(0);
 334 }
 335
 336 /**
 337  * Load namespace LFSCK statistics information from the trace file.
 338  *
 339  * \param[in] env       pointer to the thread context
 340  * \param[in] com       pointer to the lfsck component
 341  *
 342  * \retval              0 for success
 343  * \retval              negative error number on failure
 344  */
 345 static int lfsck_namespace_load(const struct lu_env *env,
 346                                 struct lfsck_component *com)
 347 {
 348         int len = com->lc_file_size;
 349         int rc;
 350
 351         rc = dt_xattr_get(env, com->lc_obj,
 352                           lfsck_buf_get(env, com->lc_file_disk, len),
 353                           XATTR_NAME_LFSCK_NAMESPACE);
 354         if (rc == len) {
 355                 struct lfsck_namespace *ns = com->lc_file_ram;
 356
 357                 lfsck_namespace_le_to_cpu(ns,
 358                                 (struct lfsck_namespace *)com->lc_file_disk);
 359                 if (ns->ln_magic != LFSCK_NAMESPACE_MAGIC) {
 360                         CDEBUG(D_LFSCK, "%s: invalid lfsck_namespace magic "
 361                                "%#x != %#x\n", lfsck_lfsck2name(com->lc_lfsck),
 362                                ns->ln_magic, LFSCK_NAMESPACE_MAGIC);
 363                         rc = -ESTALE;
 364                 } else {
 365                         rc = 0;
 366                 }
 367         } else if (rc != -ENODATA) {
 368                 CDEBUG(D_LFSCK, "%s: fail to load lfsck_namespace, "
 369                        "expected = %d: rc = %d\n",
 370                        lfsck_lfsck2name(com->lc_lfsck), len, rc);
 371                 if (rc >= 0)
 372                         rc = -ESTALE;
 373         }
 374
 375         return rc;
 376 }
 377
 378 static int lfsck_namespace_store(const struct lu_env *env,
 379                                  struct lfsck_component *com)
 380 {
 381         struct dt_object                *obj    = com->lc_obj;
 382         struct lfsck_instance           *lfsck  = com->lc_lfsck;
 383         struct lfsck_namespace          *ns     = com->lc_file_ram;
 384         struct lfsck_assistant_data     *lad    = com->lc_data;
 385         struct dt_device                *dev    = lfsck_obj2dev(obj);
 386         struct cfs_bitmap               *bitmap = NULL;
 387         struct thandle                  *handle;
 388         __u32                            nbits  = 0;
 389         int                              len    = com->lc_file_size;
 390         int                              rc;
 391         ENTRY;
 392
 393         if (lad != NULL) {
 394                 bitmap = lad->lad_bitmap;
 395                 nbits = bitmap->size;
 396
 397                 LASSERT(nbits > 0);
 398                 LASSERTF((nbits & 7) == 0, "Invalid nbits %u\n", nbits);
 399         }
 400
 401         ns->ln_bitmap_size = nbits;
 402         lfsck_namespace_cpu_to_le((struct lfsck_namespace *)com->lc_file_disk,
 403                                   ns);
 404         handle = dt_trans_create(env, dev);
 405         if (IS_ERR(handle))
 406                 GOTO(log, rc = PTR_ERR(handle));
 407
 408         rc = dt_declare_xattr_set(env, obj,
 409                                   lfsck_buf_get(env, com->lc_file_disk, len),
 410                                   XATTR_NAME_LFSCK_NAMESPACE, 0, handle);
 411         if (rc != 0)
 412                 GOTO(out, rc);
 413
 414         if (bitmap != NULL) {
 415                 rc = dt_declare_xattr_set(env, obj,
 416                                 lfsck_buf_get(env, bitmap->data, nbits >> 3),
 417                                 XATTR_NAME_LFSCK_BITMAP, 0, handle);
 418                 if (rc != 0)
 419                         GOTO(out, rc);
 420         }
 421
 422         rc = dt_trans_start_local(env, dev, handle);
 423         if (rc != 0)
 424                 GOTO(out, rc);
 425
 426         rc = dt_xattr_set(env, obj,
 427                           lfsck_buf_get(env, com->lc_file_disk, len),
 428                           XATTR_NAME_LFSCK_NAMESPACE, 0, handle);
 429         if (rc == 0 && bitmap != NULL)
 430                 rc = dt_xattr_set(env, obj,
 431                                   lfsck_buf_get(env, bitmap->data, nbits >> 3),
 432                                   XATTR_NAME_LFSCK_BITMAP, 0, handle);
 433
 434         GOTO(out, rc);
 435
 436 out:
 437         dt_trans_stop(env, dev, handle);
 438
 439 log:
 440         if (rc != 0)
 441                 CDEBUG(D_LFSCK, "%s: fail to store lfsck_namespace: rc = %d\n",
 442                        lfsck_lfsck2name(lfsck), rc);
 443         return rc;
 444 }
 445
 446 static int lfsck_namespace_init(const struct lu_env *env,
 447                                 struct lfsck_component *com)
 448 {
 449         struct lfsck_namespace *ns = com->lc_file_ram;
 450         int rc;
 451
 452         memset(ns, 0, sizeof(*ns));
 453         ns->ln_magic = LFSCK_NAMESPACE_MAGIC;
 454         ns->ln_status = LS_INIT;
 455         ns->ln_time_latest_reset = ktime_get_real_seconds();
 456         down_write(&com->lc_sem);
 457         rc = lfsck_namespace_store(env, com);
 458         if (rc == 0)
 459                 rc = lfsck_load_sub_trace_files(env, com,
 460                         &dt_lfsck_namespace_features, LFSCK_NAMESPACE, true);
 461         up_write(&com->lc_sem);
 462
 463         return rc;
 464 }
 465
 466 /**
 467  * Update the namespace LFSCK trace file for the given @fid
 468  *
 469  * \param[in] env       pointer to the thread context
 470  * \param[in] com       pointer to the lfsck component
 471  * \param[in] fid       the fid which flags to be updated in the lfsck
 472  *                      trace file
 473  * \param[in] add       true if add new flags, otherwise remove flags
 474  *
 475  * \retval              0 for success or nothing to be done
 476  * \retval              negative error number on failure
 477  */
 478 int lfsck_namespace_trace_update(const struct lu_env *env,
 479                                  struct lfsck_component *com,
 480                                  const struct lu_fid *fid,
 481                                  const __u8 flags, bool add)
 482 {
 483         struct lfsck_instance   *lfsck  = com->lc_lfsck;
 484         struct dt_object        *obj;
 485         struct lu_fid           *key    = &lfsck_env_info(env)->lti_fid3;
 486         struct dt_device        *dev;
 487         struct thandle          *th     = NULL;
 488         int                      idx;
 489         int                      rc     = 0;
 490         __u8                     old    = 0;
 491         __u8                     new    = 0;
 492         ENTRY;
 493
 494         LASSERT(flags != 0);
 495
 496         if (unlikely(!fid_is_sane(fid)))
 497                 RETURN(0);
 498
 499         idx = lfsck_sub_trace_file_fid2idx(fid);
 500         mutex_lock(&com->lc_sub_trace_objs[idx].lsto_mutex);
 501         obj = com->lc_sub_trace_objs[idx].lsto_obj;
 502         if (unlikely(obj == NULL)) {
 503                 mutex_unlock(&com->lc_sub_trace_objs[idx].lsto_mutex);
 504                 RETURN(0);
 505         }
 506
 507         lfsck_object_get(obj);
 508         dev = lfsck_obj2dev(obj);
 509         fid_cpu_to_be(key, fid);
 510         rc = dt_lookup(env, obj, (struct dt_rec *)&old,
 511                        (const struct dt_key *)key);
 512         if (rc == -ENOENT) {
 513                 if (!add)
 514                         GOTO(unlock, rc = 0);
 515
 516                 old = 0;
 517                 new = flags;
 518         } else if (rc == 0) {
 519                 if (add) {
 520                         if ((old & flags) == flags)
 521                                 GOTO(unlock, rc = 0);
 522
 523                         new = old | flags;
 524                 } else {
 525                         if ((old & flags) == 0)
 526                                 GOTO(unlock, rc = 0);
 527
 528                         new = old & ~flags;
 529                 }
 530         } else {
 531                 GOTO(log, rc);
 532         }
 533
 534         th = dt_trans_create(env, dev);
 535         if (IS_ERR(th))
 536                 GOTO(log, rc = PTR_ERR(th));
 537
 538         if (old != 0) {
 539                 rc = dt_declare_delete(env, obj,
 540                                        (const struct dt_key *)key, th);
 541                 if (rc != 0)
 542                         GOTO(log, rc);
 543         }
 544
 545         if (new != 0) {
 546                 rc = dt_declare_insert(env, obj,
 547                                        (const struct dt_rec *)&new,
 548                                        (const struct dt_key *)key, th);
 549                 if (rc != 0)
 550                         GOTO(log, rc);
 551         }
 552
 553         rc = dt_trans_start_local(env, dev, th);
 554         if (rc != 0)
 555                 GOTO(log, rc);
 556
 557         if (old != 0) {
 558                 rc = dt_delete(env, obj, (const struct dt_key *)key, th);
 559                 if (rc != 0)
 560                         GOTO(log, rc);
 561         }
 562
 563         if (new != 0) {
 564                 rc = dt_insert(env, obj, (const struct dt_rec *)&new,
 565                                (const struct dt_key *)key, th);
 566                 if (rc != 0)
 567                         GOTO(log, rc);
 568         }
 569
 570         GOTO(log, rc);
 571
 572 log:
 573         if (th != NULL && !IS_ERR(th))
 574                 dt_trans_stop(env, dev, th);
 575
 576         CDEBUG(D_LFSCK, "%s: namespace LFSCK %s flags for "DFID" in the "
 577                "trace file, flags %x, old %x, new %x: rc = %d\n",
 578                lfsck_lfsck2name(lfsck), add ? "add" : "del", PFID(fid),
 579                (__u32)flags, (__u32)old, (__u32)new, rc);
 580
 581 unlock:
 582         mutex_unlock(&com->lc_sub_trace_objs[idx].lsto_mutex);
 583         lfsck_object_put(env, obj);
 584
 585         return rc;
 586 }
 587
 588 int lfsck_namespace_check_exist(const struct lu_env *env,
 589                                 struct dt_object *dir,
 590                                 struct dt_object *obj, const char *name)
 591 {
 592         struct lu_fid    *fid = &lfsck_env_info(env)->lti_fid;
 593         int               rc;
 594         ENTRY;
 595
 596         if (unlikely(lfsck_is_dead_obj(obj)))
 597                 RETURN(LFSCK_NAMEENTRY_DEAD);
 598
 599         rc = dt_lookup(env, dir, (struct dt_rec *)fid,
 600                        (const struct dt_key *)name);
 601         if (rc == -ENOENT)
 602                 RETURN(LFSCK_NAMEENTRY_REMOVED);
 603
 604         if (rc < 0)
 605                 RETURN(rc);
 606
 607         if (!lu_fid_eq(fid, lfsck_dto2fid(obj)))
 608                 RETURN(LFSCK_NAMEENTRY_RECREATED);
 609
 610         RETURN(0);
 611 }
 612
 613 static int lfsck_declare_namespace_exec_dir(const struct lu_env *env,
 614                                             struct dt_object *obj,
 615                                             struct thandle *handle)
 616 {
 617         int rc;
 618
 619         /* For remote updating LINKEA, there may be further LFSCK action
 620          * on remote MDT after the updating, so update the LINKEA ASAP. */
 621         if (dt_object_remote(obj))
 622                 handle->th_sync = 1;
 623
 624         /* For destroying all invalid linkEA entries. */
 625         rc = dt_declare_xattr_del(env, obj, XATTR_NAME_LINK, handle);
 626         if (rc == 0)
 627                 /* For insert new linkEA entry. */
 628                 rc = dt_declare_xattr_set(env, obj,
 629                         lfsck_buf_get_const(env, NULL, MAX_LINKEA_SIZE),
 630                         XATTR_NAME_LINK, 0, handle);
 631         return rc;
 632 }
 633
 634 int __lfsck_links_read(const struct lu_env *env, struct dt_object *obj,
 635                        struct linkea_data *ldata, bool with_rec)
 636 {
 637         int rc;
 638
 639         if (ldata->ld_buf->lb_buf == NULL)
 640                 return -ENOMEM;
 641
 642         if (!dt_object_exists(obj))
 643                 return -ENOENT;
 644
 645         rc = dt_xattr_get(env, obj, ldata->ld_buf, XATTR_NAME_LINK);
 646         if (rc == -ERANGE) {
 647                 /* Buf was too small, figure out what we need. */
 648                 rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_LINK);
 649                 if (unlikely(rc == 0))
 650                         return -ENODATA;
 651
 652                 if (rc < 0)
 653                         return rc;
 654
 655                 lu_buf_realloc(ldata->ld_buf, rc);
 656                 if (ldata->ld_buf->lb_buf == NULL)
 657                         return -ENOMEM;
 658
 659                 rc = dt_xattr_get(env, obj, ldata->ld_buf, XATTR_NAME_LINK);
 660         }
 661
 662         if (unlikely(rc == 0))
 663                 return -ENODATA;
 664
 665         if (rc > 0) {
 666                 if (with_rec)
 667                         rc = linkea_init_with_rec(ldata);
 668                 else
 669                         rc = linkea_init(ldata);
 670         }
 671
 672         return rc;
 673 }
 674
 675 /**
 676  * Remove linkEA for the given object.
 677  *
 678  * The caller should take the ldlm lock before the calling.
 679  *
 680  * \param[in] env       pointer to the thread context
 681  * \param[in] com       pointer to the lfsck component
 682  * \param[in] obj       pointer to the dt_object to be handled
 683  *
 684  * \retval              0 for repaired cases
 685  * \retval              negative error number on failure
 686  */
 687 static int lfsck_namespace_links_remove(const struct lu_env *env,
 688                                         struct lfsck_component *com,
 689                                         struct dt_object *obj)
 690 {
 691         struct lfsck_instance           *lfsck  = com->lc_lfsck;
 692         struct dt_device                *dev    = lfsck_obj2dev(obj);
 693         struct thandle                  *th     = NULL;
 694         int                              rc     = 0;
 695         ENTRY;
 696
 697         LASSERT(dt_object_remote(obj) == 0);
 698
 699         th = dt_trans_create(env, dev);
 700         if (IS_ERR(th))
 701                 GOTO(log, rc = PTR_ERR(th));
 702
 703         rc = dt_declare_xattr_del(env, obj, XATTR_NAME_LINK, th);
 704         if (rc != 0)
 705                 GOTO(stop, rc);
 706
 707         rc = dt_trans_start_local(env, dev, th);
 708         if (rc != 0)
 709                 GOTO(stop, rc);
 710
 711         dt_write_lock(env, obj, 0);
 712         if (unlikely(lfsck_is_dead_obj(obj)))
 713                 GOTO(unlock, rc = -ENOENT);
 714
 715         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
 716                 GOTO(unlock, rc = 0);
 717
 718         rc = dt_xattr_del(env, obj, XATTR_NAME_LINK, th);
 719
 720         GOTO(unlock, rc);
 721
 722 unlock:
 723         dt_write_unlock(env, obj);
 724
 725 stop:
 726         dt_trans_stop(env, dev, th);
 727
 728 log:
 729         CDEBUG(D_LFSCK, "%s: namespace LFSCK remove invalid linkEA "
 730                "for the object "DFID": rc = %d\n",
 731                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(obj)), rc);
 732
 733         if (rc == 0) {
 734                 struct lfsck_namespace *ns = com->lc_file_ram;
 735
 736                 ns->ln_flags |= LF_INCONSISTENT;
 737         }
 738
 739         return rc;
 740 }
 741
 742 static int lfsck_links_write(const struct lu_env *env, struct dt_object *obj,
 743                              struct linkea_data *ldata, struct thandle *handle)
 744 {
 745         struct lu_buf buf;
 746         int rc;
 747
 748         lfsck_buf_init(&buf, ldata->ld_buf->lb_buf, ldata->ld_leh->leh_len);
 749
 750 again:
 751         rc = dt_xattr_set(env, obj, &buf, XATTR_NAME_LINK, 0, handle);
 752         if (unlikely(rc == -ENOSPC)) {
 753                 rc = linkea_overflow_shrink(ldata);
 754                 if (likely(rc > 0)) {
 755                         buf.lb_len = rc;
 756                         goto again;
 757                 }
 758         }
 759
 760         return rc;
 761 }
 762
 763 static inline bool linkea_reclen_is_valid(const struct linkea_data *ldata)
 764 {
 765         if (ldata->ld_reclen <= 0)
 766                 return false;
 767
 768         if ((char *)ldata->ld_lee + ldata->ld_reclen >
 769             (char *)ldata->ld_leh + ldata->ld_leh->leh_len)
 770                 return false;
 771
 772         return true;
 773 }
 774
 775 static inline bool linkea_entry_is_valid(const struct linkea_data *ldata,
 776                                          const struct lu_name *cname,
 777                                          const struct lu_fid *pfid)
 778 {
 779         if (!linkea_reclen_is_valid(ldata))
 780                 return false;
 781
 782         if (cname->ln_namelen <= 0 || cname->ln_namelen > NAME_MAX)
 783                 return false;
 784
 785         if (!fid_is_sane(pfid))
 786                 return false;
 787
 788         return true;
 789 }
 790
 791 static int lfsck_namespace_unpack_linkea_entry(struct linkea_data *ldata,
 792                                                struct lu_name *cname,
 793                                                struct lu_fid *pfid,
 794                                                char *buf, const int buflen)
 795 {
 796         linkea_entry_unpack(ldata->ld_lee, &ldata->ld_reclen, cname, pfid);
 797         if (unlikely(!linkea_entry_is_valid(ldata, cname, pfid)))
 798                 return -EINVAL;
 799
 800         /* To guarantee the 'name' is terminated with '0'. */
 801         memcpy(buf, cname->ln_name, cname->ln_namelen);
 802         buf[cname->ln_namelen] = 0;
 803         cname->ln_name = buf;
 804
 805         return 0;
 806 }
 807
 808 static void lfsck_linkea_del_buf(struct linkea_data *ldata,
 809                                  const struct lu_name *lname)
 810 {
 811         LASSERT(ldata->ld_leh != NULL && ldata->ld_lee != NULL);
 812
 813         /* If current record is corrupted, all the subsequent
 814          * records will be dropped. */
 815         if (unlikely(!linkea_reclen_is_valid(ldata))) {
 816                 void *ptr = ldata->ld_lee;
 817
 818                 ldata->ld_leh->leh_len = sizeof(struct link_ea_header);
 819                 ldata->ld_leh->leh_reccount = 0;
 820                 linkea_first_entry(ldata);
 821                 while (ldata->ld_lee != NULL &&
 822                        (char *)ldata->ld_lee < (char *)ptr) {
 823                         int reclen = (ldata->ld_lee->lee_reclen[0] << 8) |
 824                                      ldata->ld_lee->lee_reclen[1];
 825
 826                         ldata->ld_leh->leh_len += reclen;
 827                         ldata->ld_leh->leh_reccount++;
 828                         ldata->ld_lee = (struct link_ea_entry *)
 829                                         ((char *)ldata->ld_lee + reclen);
 830                 }
 831
 832                 ldata->ld_lee = NULL;
 833         } else {
 834                 linkea_del_buf(ldata, lname);
 835         }
 836 }
 837
 838 static int lfsck_namespace_filter_linkea_entry(struct linkea_data *ldata,
 839                                                struct lu_name *cname,
 840                                                struct lu_fid *pfid,
 841                                                bool remove)
 842 {
 843         struct link_ea_entry    *oldlee;
 844         int                      oldlen;
 845         int                      repeated = 0;
 846
 847         oldlee = ldata->ld_lee;
 848         oldlen = ldata->ld_reclen;
 849         linkea_next_entry(ldata);
 850         while (ldata->ld_lee != NULL) {
 851                 ldata->ld_reclen = (ldata->ld_lee->lee_reclen[0] << 8) |
 852                                    ldata->ld_lee->lee_reclen[1];
 853                 if (unlikely(!linkea_reclen_is_valid(ldata))) {
 854                         lfsck_linkea_del_buf(ldata, NULL);
 855                         LASSERT(ldata->ld_lee == NULL);
 856                 } else if (unlikely(ldata->ld_reclen == oldlen &&
 857                              memcmp(ldata->ld_lee, oldlee, oldlen) == 0)) {
 858                         repeated++;
 859                         if (!remove)
 860                                 break;
 861
 862                         lfsck_linkea_del_buf(ldata, cname);
 863                 } else {
 864                         linkea_next_entry(ldata);
 865                 }
 866         }
 867         ldata->ld_lee = oldlee;
 868         ldata->ld_reclen = oldlen;
 869
 870         return repeated;
 871 }
 872
 873 /**
 874  * Insert orphan into .lustre/lost+found/MDTxxxx/ locally.
 875  *
 876  * Add the specified orphan MDT-object to the .lustre/lost+found/MDTxxxx/
 877  * with the given type to generate the name, the detailed rules for name
 878  * have been described as following.
 879  *
 880  * The function also generates the linkEA corresponding to the name entry
 881  * under the .lustre/lost+found/MDTxxxx/ for the orphan MDT-object.
 882  *
 883  * \param[in] env       pointer to the thread context
 884  * \param[in] com       pointer to the lfsck component
 885  * \param[in] orphan    pointer to the orphan MDT-object
 886  * \param[in] infix     additional information for the orphan name, such as
 887  *                      the FID for original
 888  * \param[in] type      the type for describing why the orphan MDT-object is
 889  *                      created. The rules are as following:
 890  *
 891  *  type "D":           The MDT-object is a directory, it may knows its parent
 892  *                      but because there is no valid linkEA, the LFSCK cannot
 893  *                      know where to put it back to the namespace.
 894  *  type "O":           The MDT-object has no linkEA, and there is no name
 895  *                      entry that references the MDT-object.
 896  *
 897  *  type "S":           The orphan MDT-object is a shard of a striped directory
 898  *
 899  * \see lfsck_layout_recreate_parent() for more types.
 900  *
 901  * The orphan name will be like:
 902  * ${FID}-${infix}-${type}-${conflict_version}
 903  *
 904  * \param[out] count    if some others inserted some linkEA entries by race,
 905  *                      then return the linkEA entries count.
 906  *
 907  * \retval              positive number for repaired cases
 908  * \retval              0 if needs to repair nothing
 909  * \retval              negative error number on failure
 910  */
 911 static int lfsck_namespace_insert_orphan(const struct lu_env *env,
 912                                          struct lfsck_component *com,
 913                                          struct dt_object *orphan,
 914                                          const char *infix, const char *type,
 915                                          int *count)
 916 {
 917         struct lfsck_thread_info        *info   = lfsck_env_info(env);
 918         struct lu_name                  *cname  = &info->lti_name;
 919         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
 920         struct lu_attr                  *la     = &info->lti_la2;
 921         const struct lu_fid             *cfid   = lfsck_dto2fid(orphan);
 922         const struct lu_fid             *pfid;
 923         struct lu_fid                    tfid;
 924         struct lfsck_instance           *lfsck  = com->lc_lfsck;
 925         struct dt_device                *dev    = lfsck_obj2dev(orphan);
 926         struct dt_object                *parent;
 927         struct thandle                  *th     = NULL;
 928         struct lfsck_lock_handle        *pllh   = &info->lti_llh;
 929         struct lustre_handle             clh    = { 0 };
 930         struct linkea_data               ldata2 = { NULL };
 931         struct lu_buf                    linkea_buf;
 932         int                              namelen;
 933         int                              idx    = 0;
 934         int                              rc     = 0;
 935         bool                             exist  = false;
 936         ENTRY;
 937
 938         cname->ln_name = NULL;
 939         if (unlikely(lfsck->li_lpf_obj == NULL))
 940                 GOTO(log, rc = -ENXIO);
 941
 942         parent = lfsck->li_lpf_obj;
 943         pfid = lfsck_dto2fid(parent);
 944
 945 again:
 946         do {
 947                 namelen = snprintf(info->lti_key, NAME_MAX, DFID"%s-%s-%d",
 948                                    PFID(cfid), infix, type, idx++);
 949                 rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
 950                                (const struct dt_key *)info->lti_key);
 951                 if (rc != 0 && rc != -ENOENT)
 952                         GOTO(log, rc);
 953
 954                 if (unlikely(rc == 0 && lu_fid_eq(cfid, &tfid)))
 955                         exist = true;
 956         } while (rc == 0 && !exist);
 957
 958         rc = lfsck_lock(env, lfsck, parent, info->lti_key, pllh,
 959                         MDS_INODELOCK_UPDATE, LCK_PW);
 960         if (rc != 0)
 961                 GOTO(log, rc);
 962
 963         /* Re-check whether the name conflict with othrs after taken
 964          * the ldlm lock. */
 965         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
 966                        (const struct dt_key *)info->lti_key);
 967         if (rc == 0) {
 968                 if (!lu_fid_eq(cfid, &tfid)) {
 969                         exist = false;
 970                         lfsck_unlock(pllh);
 971                         goto again;
 972                 }
 973
 974                 exist = true;
 975         } else if (rc != -ENOENT) {
 976                 GOTO(log, rc);
 977         } else {
 978                 exist = false;
 979         }
 980
 981         cname->ln_name = info->lti_key;
 982         cname->ln_namelen = namelen;
 983         rc = linkea_links_new(&ldata2, &info->lti_linkea_buf2,
 984                               cname, pfid);
 985         if (rc != 0)
 986                 GOTO(log, rc);
 987
 988         rc = lfsck_ibits_lock(env, lfsck, orphan, &clh,
 989                               MDS_INODELOCK_UPDATE | MDS_INODELOCK_LOOKUP |
 990                               MDS_INODELOCK_XATTR, LCK_EX);
 991         if (rc != 0)
 992                 GOTO(log, rc);
 993
 994         lfsck_buf_init(&linkea_buf, ldata2.ld_buf->lb_buf,
 995                        ldata2.ld_leh->leh_len);
 996         th = dt_trans_create(env, dev);
 997         if (IS_ERR(th))
 998                 GOTO(log, rc = PTR_ERR(th));
 999
1000         if (S_ISDIR(lfsck_object_type(orphan))) {
1001                 rc = dt_declare_delete(env, orphan,
1002                                        (const struct dt_key *)dotdot, th);
1003                 if (rc != 0)
1004                         GOTO(stop, rc);
1005
1006                 rec->rec_type = S_IFDIR;
1007                 rec->rec_fid = pfid;
1008                 rc = dt_declare_insert(env, orphan, (const struct dt_rec *)rec,
1009                                        (const struct dt_key *)dotdot, th);
1010                 if (rc != 0)
1011                         GOTO(stop, rc);
1012         }
1013
1014         rc = dt_declare_xattr_set(env, orphan, &linkea_buf,
1015                                   XATTR_NAME_LINK, 0, th);
1016         if (rc != 0)
1017                 GOTO(stop, rc);
1018
1019         if (!exist) {
1020                 rec->rec_type = lfsck_object_type(orphan) & S_IFMT;
1021                 rec->rec_fid = cfid;
1022                 rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
1023                                        (const struct dt_key *)cname->ln_name,
1024                                        th);
1025                 if (rc != 0)
1026                         GOTO(stop, rc);
1027
1028                 if (S_ISDIR(rec->rec_type)) {
1029                         rc = dt_declare_ref_add(env, parent, th);
1030                         if (rc != 0)
1031                                 GOTO(stop, rc);
1032                 }
1033         }
1034
1035         memset(la, 0, sizeof(*la));
1036         la->la_ctime = ktime_get_real_seconds();
1037         la->la_valid = LA_CTIME;
1038         rc = dt_declare_attr_set(env, orphan, la, th);
1039         if (rc != 0)
1040                 GOTO(stop, rc);
1041
1042         rc = dt_trans_start_local(env, dev, th);
1043         if (rc != 0)
1044                 GOTO(stop, rc);
1045
1046         dt_write_lock(env, orphan, 0);
1047         rc = lfsck_links_read2_with_rec(env, orphan, &ldata2);
1048         if (likely(rc == -ENODATA || rc == -EINVAL)) {
1049                 if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1050                         GOTO(unlock, rc = 1);
1051
1052                 if (S_ISDIR(lfsck_object_type(orphan))) {
1053                         rc = dt_delete(env, orphan,
1054                                        (const struct dt_key *)dotdot, th);
1055                         if (rc != 0)
1056                                 GOTO(unlock, rc);
1057
1058                         rec->rec_type = S_IFDIR;
1059                         rec->rec_fid = pfid;
1060                         rc = dt_insert(env, orphan, (const struct dt_rec *)rec,
1061                                        (const struct dt_key *)dotdot, th);
1062                         if (rc != 0)
1063                                 GOTO(unlock, rc);
1064                 }
1065
1066                 rc = dt_xattr_set(env, orphan, &linkea_buf, XATTR_NAME_LINK, 0,
1067                                   th);
1068         } else {
1069                 if (rc == 0 && count != NULL)
1070                         *count = ldata2.ld_leh->leh_reccount;
1071
1072                 GOTO(unlock, rc);
1073         }
1074         dt_write_unlock(env, orphan);
1075
1076         if (rc == 0 && !exist) {
1077                 rec->rec_type = lfsck_object_type(orphan) & S_IFMT;
1078                 rec->rec_fid = cfid;
1079                 rc = dt_insert(env, parent, (const struct dt_rec *)rec,
1080                                (const struct dt_key *)cname->ln_name, th);
1081                 if (rc == 0 && S_ISDIR(rec->rec_type)) {
1082                         dt_write_lock(env, parent, 0);
1083                         rc = dt_ref_add(env, parent, th);
1084                         dt_write_unlock(env, parent);
1085                 }
1086         }
1087
1088         if (rc == 0)
1089                 rc = dt_attr_set(env, orphan, la, th);
1090
1091         GOTO(stop, rc = (rc == 0 ? 1 : rc));
1092
1093 unlock:
1094         dt_write_unlock(env, orphan);
1095
1096 stop:
1097         dt_trans_stop(env, dev, th);
1098
1099 log:
1100         lfsck_ibits_unlock(&clh, LCK_EX);
1101         lfsck_unlock(pllh);
1102         CDEBUG(D_LFSCK, "%s: namespace LFSCK insert orphan for the "
1103                "object "DFID", name = %s: rc = %d\n",
1104                lfsck_lfsck2name(lfsck), PFID(cfid),
1105                cname->ln_name != NULL ? cname->ln_name : "<NULL>", rc);
1106
1107         if (rc != 0) {
1108                 struct lfsck_namespace *ns = com->lc_file_ram;
1109
1110                 ns->ln_flags |= LF_INCONSISTENT;
1111         }
1112
1113         return rc;
1114 }
1115
1116 static int lfsck_lmv_set(const struct lu_env *env,
1117                          struct lfsck_instance *lfsck,
1118                          struct dt_object *obj,
1119                          struct lmv_mds_md_v1 *lmv)
1120 {
1121         struct dt_device *dev = lfsck->li_next;
1122         struct thandle *th = NULL;
1123         struct lu_buf buf = { lmv, sizeof(*lmv) };
1124         int rc;
1125
1126         ENTRY;
1127
1128         th = dt_trans_create(env, dev);
1129         if (IS_ERR(th))
1130                 RETURN(PTR_ERR(th));
1131
1132         rc = dt_declare_xattr_set(env, obj, &buf, XATTR_NAME_LMV, 0, th);
1133         if (rc)
1134                 GOTO(stop, rc);
1135
1136         rc = dt_trans_start_local(env, dev, th);
1137         if (rc != 0)
1138                 GOTO(stop, rc);
1139
1140         rc = dt_xattr_set(env, obj, &buf, XATTR_NAME_LMV, 0, th);
1141         if (rc)
1142                 GOTO(stop, rc);
1143
1144         EXIT;
1145 stop:
1146         dt_trans_stop(env, dev, th);
1147
1148         return rc;
1149 }
1150
1151 static int lfsck_lmv_delete(const struct lu_env *env,
1152                             struct lfsck_instance *lfsck,
1153                             struct dt_object *obj)
1154 {
1155         struct dt_device *dev = lfsck->li_next;
1156         struct thandle *th = NULL;
1157         int rc;
1158
1159         ENTRY;
1160
1161         th = dt_trans_create(env, dev);
1162         if (IS_ERR(th))
1163                 RETURN(PTR_ERR(th));
1164
1165         rc = dt_declare_xattr_del(env, obj, XATTR_NAME_LMV, th);
1166         if (rc)
1167                 GOTO(stop, rc);
1168
1169         rc = dt_trans_start_local(env, dev, th);
1170         if (rc != 0)
1171                 GOTO(stop, rc);
1172
1173         rc = dt_xattr_del(env, obj, XATTR_NAME_LMV, th);
1174         if (rc)
1175                 GOTO(stop, rc);
1176
1177         EXIT;
1178 stop:
1179         dt_trans_stop(env, dev, th);
1180
1181         return rc;
1182 }
1183
1184 static inline int lfsck_object_is_shard(const struct lu_env *env,
1185                                         struct lfsck_instance *lfsck,
1186                                         struct dt_object *obj,
1187                                         const struct lu_name *lname)
1188 {
1189         struct lfsck_thread_info *info = lfsck_env_info(env);
1190         struct lmv_mds_md_v1 *lmv = &info->lti_lmv;
1191         int rc;
1192
1193         rc = lfsck_shard_name_to_index(env, lname->ln_name, lname->ln_namelen,
1194                                        lfsck_object_type(obj),
1195                                        lfsck_dto2fid(obj));
1196         if (rc < 0)
1197                 return 0;
1198
1199         rc = lfsck_read_stripe_lmv(env, lfsck, obj, lmv);
1200         if (rc == -ENODATA)
1201                 return 0;
1202
1203         if (!rc && lmv->lmv_magic == LMV_MAGIC_STRIPE)
1204                 return 1;
1205
1206         return rc;
1207 }
1208
1209 /**
1210  * Add the specified name entry back to namespace.
1211  *
1212  * If there is a linkEA entry that back references a name entry under
1213  * some parent directory, but such parent directory does not have the
1214  * claimed name entry. On the other hand, the linkEA entries count is
1215  * not larger than the MDT-object's hard link count. Under such case,
1216  * it is quite possible that the name entry is lost. Then the LFSCK
1217  * should add the name entry back to the namespace.
1218  *
1219  * If \a child is shard, which means \a parent is a striped directory,
1220  * if \a parent has LMV, we need to delete it before insertion because
1221  * now parent's striping is broken and can't be parsed correctly.
1222  *
1223  * \param[in] env       pointer to the thread context
1224  * \param[in] com       pointer to the lfsck component
1225  * \param[in] parent    pointer to the directory under which the name entry
1226  *                      will be inserted into
1227  * \param[in] child     pointer to the object referenced by the name entry
1228  *                      that to be inserted into the parent
1229  * \param[in] lname     the name for the child in the parent directory
1230  *
1231  * \retval              positive number for repaired cases
1232  * \retval              0 if nothing to be repaired
1233  * \retval              negative error number on failure
1234  */
1235 static int lfsck_namespace_insert_normal(const struct lu_env *env,
1236                                          struct lfsck_component *com,
1237                                          struct dt_object *parent,
1238                                          struct dt_object *child,
1239                                          const struct lu_name *lname)
1240 {
1241         struct lfsck_thread_info *info = lfsck_env_info(env);
1242         struct lu_attr *la = &info->lti_la;
1243         struct dt_insert_rec *rec = &info->lti_dt_rec;
1244         struct lfsck_instance *lfsck = com->lc_lfsck;
1245         /* The child and its name may be on different MDTs. */
1246         const struct lu_fid *pfid = lfsck_dto2fid(parent);
1247         const struct lu_fid *cfid = lfsck_dto2fid(child);
1248         struct dt_device *dev = lfsck->li_next;
1249         struct thandle *th = NULL;
1250         struct lfsck_lock_handle *llh = &info->lti_llh;
1251         struct lmv_mds_md_v1 *lmv = &info->lti_lmv;
1252         struct lu_buf buf = { lmv, sizeof(*lmv) };
1253         /* whether parent's LMV is deleted before insertion */
1254         bool parent_lmv_deleted = false;
1255         /* whether parent's LMV is missing */
1256         bool parent_lmv_lost = false;
1257         int rc = 0;
1258
1259         ENTRY;
1260
1261         /* @parent/@child may be based on lfsck->li_bottom,
1262          * but here we need the object based on the lfsck->li_next. */
1263
1264         parent = lfsck_object_locate(dev, parent);
1265         if (IS_ERR(parent))
1266                 GOTO(log, rc = PTR_ERR(parent));
1267
1268         child = lfsck_object_locate(dev, child);
1269         if (IS_ERR(child))
1270                 GOTO(log, rc = PTR_ERR(child));
1271
1272         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1273                 GOTO(log, rc = 1);
1274
1275         rc = lfsck_lock(env, lfsck, parent, lname->ln_name, llh,
1276                         MDS_INODELOCK_LOOKUP | MDS_INODELOCK_UPDATE |
1277                         MDS_INODELOCK_XATTR, LCK_EX);
1278         if (rc)
1279                 GOTO(log, rc);
1280
1281         rc = lfsck_object_is_shard(env, lfsck, child, lname);
1282         if (rc < 0)
1283                 GOTO(unlock, rc);
1284
1285         if (rc == 1) {
1286                 rc = lfsck_read_stripe_lmv(env, lfsck, parent, lmv);
1287                 if (!rc) {
1288                         /*
1289                          * To add a shard, we need to convert parent to a
1290                          * plain directory by deleting its LMV, and after
1291                          * insertion set it back.
1292                          */
1293                         rc = lfsck_lmv_delete(env, lfsck, parent);
1294                         if (rc)
1295                                 GOTO(unlock, rc);
1296                         parent_lmv_deleted = true;
1297                         lmv->lmv_layout_version++;
1298                         lfsck_lmv_header_cpu_to_le(lmv, lmv);
1299                 } else if (rc == -ENODATA) {
1300                         struct lu_seq_range *range = &info->lti_range;
1301                         struct seq_server_site *ss = lfsck_dev_site(lfsck);
1302
1303                         rc = lfsck_read_stripe_lmv(env, lfsck, child, lmv);
1304                         if (rc)
1305                                 GOTO(unlock, rc);
1306
1307                         fld_range_set_mdt(range);
1308                         rc = fld_server_lookup(env, ss->ss_server_fld,
1309                                        fid_seq(lfsck_dto2fid(parent)), range);
1310                         if (rc)
1311                                 GOTO(unlock, rc);
1312
1313                         parent_lmv_lost = true;
1314                         lmv->lmv_magic = LMV_MAGIC;
1315                         lmv->lmv_master_mdt_index = range->lsr_index;
1316                         lmv->lmv_layout_version++;
1317                         lfsck_lmv_header_cpu_to_le(lmv, lmv);
1318                 } else {
1319                         GOTO(unlock, rc);
1320                 }
1321         }
1322
1323         if (unlikely(!dt_try_as_dir(env, parent)))
1324                 GOTO(unlock, rc = -ENOTDIR);
1325
1326         th = dt_trans_create(env, dev);
1327         if (IS_ERR(th))
1328                 GOTO(unlock, rc = PTR_ERR(th));
1329
1330         rec->rec_type = lfsck_object_type(child) & S_IFMT;
1331         rec->rec_fid = cfid;
1332         rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
1333                                (const struct dt_key *)lname->ln_name, th);
1334         if (rc != 0)
1335                 GOTO(stop, rc);
1336
1337         if (S_ISDIR(rec->rec_type)) {
1338                 rc = dt_declare_ref_add(env, parent, th);
1339                 if (rc != 0)
1340                         GOTO(stop, rc);
1341         }
1342
1343         if (parent_lmv_lost) {
1344                 rc = dt_declare_xattr_set(env, parent, &buf, XATTR_NAME_LMV,
1345                                           0, th);
1346                 if (rc)
1347                         GOTO(stop, rc);
1348         }
1349
1350         la->la_ctime = ktime_get_real_seconds();
1351         la->la_valid = LA_CTIME;
1352         rc = dt_declare_attr_set(env, parent, la, th);
1353         if (rc != 0)
1354                 GOTO(stop, rc);
1355
1356         rc = dt_declare_attr_set(env, child, la, th);
1357         if (rc != 0)
1358                 GOTO(stop, rc);
1359
1360         rc = dt_trans_start_local(env, dev, th);
1361         if (rc != 0)
1362                 GOTO(stop, rc);
1363
1364         rc = dt_insert(env, parent, (const struct dt_rec *)rec,
1365                        (const struct dt_key *)lname->ln_name, th);
1366         if (rc != 0)
1367                 GOTO(stop, rc);
1368
1369         if (S_ISDIR(rec->rec_type)) {
1370                 dt_write_lock(env, parent, 0);
1371                 rc = dt_ref_add(env, parent, th);
1372                 dt_write_unlock(env, parent);
1373                 if (rc != 0)
1374                         GOTO(stop, rc);
1375         }
1376
1377         if (parent_lmv_lost) {
1378                 rc = dt_xattr_set(env, parent, &buf, XATTR_NAME_LMV, 0, th);
1379                 if (rc)
1380                         GOTO(stop, rc);
1381         }
1382
1383         rc = dt_attr_set(env, parent, la, th);
1384         if (rc != 0)
1385                 GOTO(stop, rc);
1386
1387         rc = dt_attr_set(env, child, la, th);
1388
1389         GOTO(stop, rc = (rc == 0 ? 1 : rc));
1390
1391 stop:
1392         dt_trans_stop(env, dev, th);
1393
1394 unlock:
1395         if (parent_lmv_deleted)
1396                 lfsck_lmv_set(env, lfsck, parent, lmv);
1397
1398         lfsck_unlock(llh);
1399
1400 log:
1401         CDEBUG(D_LFSCK, "%s: namespace LFSCK insert object "DFID" with "
1402                "the name %s and type %o to the parent "DFID": rc = %d\n",
1403                lfsck_lfsck2name(lfsck), PFID(cfid), lname->ln_name,
1404                lfsck_object_type(child) & S_IFMT, PFID(pfid), rc);
1405
1406         if (rc != 0) {
1407                 struct lfsck_namespace *ns = com->lc_file_ram;
1408
1409                 ns->ln_flags |= LF_INCONSISTENT;
1410                 if (rc > 0)
1411                         ns->ln_lost_dirent_repaired++;
1412         }
1413
1414         return rc;
1415 }
1416
1417 /**
1418  * Create the specified orphan directory.
1419  *
1420  * For the case that the parent MDT-object stored in some MDT-object's
1421  * linkEA entry is lost, the LFSCK will re-create the parent object as
1422  * an orphan and insert it into .lustre/lost+found/MDTxxxx/ directory
1423  * with the name ${FID}-P-${conflict_version}.
1424  *
1425  * \param[in] env       pointer to the thread context
1426  * \param[in] com       pointer to the lfsck component
1427  * \param[in] orphan    pointer to the orphan MDT-object to be created
1428  * \param[in] lmv       pointer to master LMV EA that will be set to the orphan
1429  *
1430  * \retval              positive number for repaired cases
1431  * \retval              negative error number on failure
1432  */
1433 static int lfsck_namespace_create_orphan_dir(const struct lu_env *env,
1434                                              struct lfsck_component *com,
1435                                              struct dt_object *orphan,
1436                                              struct lmv_mds_md_v1 *lmv)
1437 {
1438         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1439         struct lu_attr                  *la     = &info->lti_la;
1440         struct dt_allocation_hint       *hint   = &info->lti_hint;
1441         struct dt_object_format         *dof    = &info->lti_dof;
1442         struct lu_name                  *cname  = &info->lti_name2;
1443         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
1444         struct lmv_mds_md_v1            *lmv2   = &info->lti_lmv2;
1445         const struct lu_fid             *cfid   = lfsck_dto2fid(orphan);
1446         struct lu_fid                    tfid;
1447         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1448         struct lfsck_namespace          *ns     = com->lc_file_ram;
1449         struct dt_device                *dev    = lfsck_obj2dev(orphan);
1450         struct dt_object                *parent = NULL;
1451         struct thandle                  *th     = NULL;
1452         struct lfsck_lock_handle        *llh    = &info->lti_llh;
1453         struct linkea_data               ldata  = { NULL };
1454         struct lu_buf                    linkea_buf;
1455         struct lu_buf                    lmv_buf;
1456         char                             name[32];
1457         int                              namelen;
1458         int                              idx    = 0;
1459         int                              rc     = 0;
1460         int                              rc1    = 0;
1461         ENTRY;
1462
1463         LASSERT(!dt_object_exists(orphan));
1464
1465         cname->ln_name = NULL;
1466         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
1467                 GOTO(log, rc = 1);
1468
1469         if (dt_object_remote(orphan)) {
1470                 LASSERT(lfsck->li_lpf_root_obj != NULL);
1471
1472                 idx = lfsck_find_mdt_idx_by_fid(env, lfsck, cfid);
1473                 if (idx < 0)
1474                         GOTO(log, rc = idx);
1475
1476                 snprintf(name, 8, "MDT%04x", idx);
1477                 rc = dt_lookup(env, lfsck->li_lpf_root_obj,
1478                                (struct dt_rec *)&tfid,
1479                                (const struct dt_key *)name);
1480                 if (rc != 0)
1481                         GOTO(log, rc = (rc == -ENOENT ? -ENXIO : rc));
1482
1483                 parent = lfsck_object_find_bottom(env, lfsck, &tfid);
1484                 if (IS_ERR(parent))
1485                         GOTO(log, rc = PTR_ERR(parent));
1486
1487                 if (unlikely(!dt_try_as_dir(env, parent)))
1488                         GOTO(log, rc = -ENOTDIR);
1489         } else {
1490                 if (unlikely(lfsck->li_lpf_obj == NULL))
1491                         GOTO(log, rc = -ENXIO);
1492
1493                 parent = lfsck->li_lpf_obj;
1494         }
1495
1496         dev = lfsck_find_dev_by_fid(env, lfsck, cfid);
1497         if (IS_ERR(dev))
1498                 GOTO(log, rc = PTR_ERR(dev));
1499
1500         idx = 0;
1501
1502 again:
1503         do {
1504                 namelen = snprintf(name, 31, DFID"-P-%d",
1505                                    PFID(cfid), idx++);
1506                 rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
1507                                (const struct dt_key *)name);
1508                 if (rc != 0 && rc != -ENOENT)
1509                         GOTO(log, rc);
1510         } while (rc == 0);
1511
1512         rc = lfsck_lock(env, lfsck, parent, name, llh,
1513                         MDS_INODELOCK_UPDATE, LCK_PW);
1514         if (rc != 0)
1515                 GOTO(log, rc);
1516
1517         /* Re-check whether the name conflict with othrs after taken
1518          * the ldlm lock. */
1519         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
1520                        (const struct dt_key *)name);
1521         if (unlikely(rc == 0)) {
1522                 lfsck_unlock(llh);
1523                 goto again;
1524         }
1525
1526         if (rc != -ENOENT)
1527                 GOTO(unlock1, rc);
1528
1529         cname->ln_name = name;
1530         cname->ln_namelen = namelen;
1531
1532         memset(la, 0, sizeof(*la));
1533         la->la_mode = S_IFDIR | 0700;
1534         la->la_valid = LA_TYPE | LA_MODE | LA_UID | LA_GID |
1535                        LA_ATIME | LA_MTIME | LA_CTIME;
1536
1537         orphan->do_ops->do_ah_init(env, hint, parent, orphan,
1538                                    la->la_mode & S_IFMT);
1539
1540         memset(dof, 0, sizeof(*dof));
1541         dof->dof_type = dt_mode_to_dft(S_IFDIR);
1542
1543         rc = linkea_links_new(&ldata, &info->lti_linkea_buf2,
1544                               cname, lfsck_dto2fid(parent));
1545         if (rc != 0)
1546                 GOTO(unlock1, rc);
1547
1548         th = dt_trans_create(env, dev);
1549         if (IS_ERR(th))
1550                 GOTO(unlock1, rc = PTR_ERR(th));
1551
1552         /* Sync the remote transaction to guarantee that the subsequent
1553          * lock against the @orphan can find the @orphan in time. */
1554         if (dt_object_remote(orphan))
1555                 th->th_sync = 1;
1556
1557         rc = dt_declare_create(env, orphan, la, hint, dof, th);
1558         if (rc != 0)
1559                 GOTO(stop, rc);
1560
1561         if (unlikely(!dt_try_as_dir(env, orphan)))
1562                 GOTO(stop, rc = -ENOTDIR);
1563
1564         rc = dt_declare_ref_add(env, orphan, th);
1565         if (rc != 0)
1566                 GOTO(stop, rc);
1567
1568         rec->rec_type = S_IFDIR;
1569         rec->rec_fid = cfid;
1570         rc = dt_declare_insert(env, orphan, (const struct dt_rec *)rec,
1571                                (const struct dt_key *)dot, th);
1572         if (rc != 0)
1573                 GOTO(stop, rc);
1574
1575         rec->rec_fid = lfsck_dto2fid(parent);
1576         rc = dt_declare_insert(env, orphan, (const struct dt_rec *)rec,
1577                                (const struct dt_key *)dotdot, th);
1578         if (rc != 0)
1579                 GOTO(stop, rc);
1580
1581         if (lmv != NULL) {
1582                 lmv->lmv_magic = LMV_MAGIC;
1583                 lmv->lmv_master_mdt_index = lfsck_dev_idx(lfsck);
1584                 lfsck_lmv_header_cpu_to_le(lmv2, lmv);
1585                 lfsck_buf_init(&lmv_buf, lmv2, sizeof(*lmv2));
1586                 rc = dt_declare_xattr_set(env, orphan, &lmv_buf, XATTR_NAME_LMV,
1587                                           0, th);
1588                 if (rc != 0)
1589                         GOTO(stop, rc);
1590         }
1591
1592         lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
1593                        ldata.ld_leh->leh_len);
1594         rc = dt_declare_xattr_set(env, orphan, &linkea_buf,
1595                                   XATTR_NAME_LINK, 0, th);
1596         if (rc != 0)
1597                 GOTO(stop, rc);
1598
1599         rec->rec_fid = cfid;
1600         rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
1601                                (const struct dt_key *)name, th);
1602         if (rc == 0)
1603                 rc = dt_declare_ref_add(env, parent, th);
1604
1605         if (rc != 0)
1606                 GOTO(stop, rc);
1607
1608         rc = dt_trans_start_local(env, dev, th);
1609         if (rc != 0)
1610                 GOTO(stop, rc);
1611
1612         dt_write_lock(env, orphan, 0);
1613         rc = dt_create(env, orphan, la, hint, dof, th);
1614         if (rc != 0)
1615                 GOTO(unlock2, rc);
1616
1617         rc = dt_ref_add(env, orphan, th);
1618         if (rc != 0)
1619                 GOTO(unlock2, rc);
1620
1621         rec->rec_fid = cfid;
1622         rc = dt_insert(env, orphan, (const struct dt_rec *)rec,
1623                        (const struct dt_key *)dot, th);
1624         if (rc != 0)
1625                 GOTO(unlock2, rc);
1626
1627         rec->rec_fid = lfsck_dto2fid(parent);
1628         rc = dt_insert(env, orphan, (const struct dt_rec *)rec,
1629                        (const struct dt_key *)dotdot, th);
1630         if (rc != 0)
1631                 GOTO(unlock2, rc);
1632
1633         if (lmv != NULL) {
1634                 rc = dt_xattr_set(env, orphan, &lmv_buf, XATTR_NAME_LMV, 0, th);
1635                 if (rc != 0)
1636                         GOTO(unlock2, rc);
1637         }
1638
1639         rc = dt_xattr_set(env, orphan, &linkea_buf,
1640                           XATTR_NAME_LINK, 0, th);
1641         dt_write_unlock(env, orphan);
1642         if (rc != 0)
1643                 GOTO(stop, rc);
1644
1645         rec->rec_fid = cfid;
1646         rc = dt_insert(env, parent, (const struct dt_rec *)rec,
1647                        (const struct dt_key *)name, th);
1648         if (rc == 0) {
1649                 dt_write_lock(env, parent, 0);
1650                 rc = dt_ref_add(env, parent, th);
1651                 dt_write_unlock(env, parent);
1652         }
1653
1654         GOTO(stop, rc = (rc == 0 ? 1 : rc));
1655
1656 unlock2:
1657         dt_write_unlock(env, orphan);
1658
1659 stop:
1660         rc1 = dt_trans_stop(env, dev, th);
1661         if (rc1 != 0 && rc > 0)
1662                 rc = rc1;
1663
1664 unlock1:
1665         lfsck_unlock(llh);
1666
1667 log:
1668         CDEBUG(D_LFSCK, "%s: namespace LFSCK create orphan dir for "
1669                "the object "DFID", name = %s: rc = %d\n",
1670                lfsck_lfsck2name(lfsck), PFID(cfid),
1671                cname->ln_name != NULL ? cname->ln_name : "<NULL>", rc);
1672
1673         if (parent != NULL && !IS_ERR(parent) && parent != lfsck->li_lpf_obj)
1674                 lfsck_object_put(env, parent);
1675
1676         if (rc != 0)
1677                 ns->ln_flags |= LF_INCONSISTENT;
1678
1679         return rc;
1680 }
1681
1682 /**
1683  * Remove the specified entry from the linkEA.
1684  *
1685  * Locate the linkEA entry with the given @cname and @pfid, then
1686  * remove this entry or the other entries those are repeated with
1687  * this entry.
1688  *
1689  * \param[in] env       pointer to the thread context
1690  * \param[in] com       pointer to the lfsck component
1691  * \param[in] obj       pointer to the dt_object to be handled
1692  * \param[in,out]ldata  pointer to the buffer that holds the linkEA
1693  * \param[in] cname     the name for the child in the parent directory
1694  * \param[in] pfid      the parent directory's FID for the linkEA
1695  * \param[in] next      if true, then remove the first found linkEA
1696  *                      entry, and move the ldata->ld_lee to next entry
1697  *
1698  * \retval              positive number for repaired cases
1699  * \retval              0 if nothing to be repaired
1700  * \retval              negative error number on failure
1701  */
1702 static int lfsck_namespace_shrink_linkea(const struct lu_env *env,
1703                                          struct lfsck_component *com,
1704                                          struct dt_object *obj,
1705                                          struct linkea_data *ldata,
1706                                          struct lu_name *cname,
1707                                          struct lu_fid *pfid,
1708                                          bool next)
1709 {
1710         struct lfsck_instance           *lfsck     = com->lc_lfsck;
1711         struct dt_device                *dev       = lfsck_obj2dev(obj);
1712         struct lfsck_bookmark           *bk        = &lfsck->li_bookmark_ram;
1713         struct thandle                  *th        = NULL;
1714         struct lustre_handle             lh        = { 0 };
1715         struct linkea_data               ldata_new = { NULL };
1716         struct lu_buf                    linkea_buf;
1717         int                              buflen    = 0;
1718         int                              rc        = 0;
1719         ENTRY;
1720
1721         rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
1722                               MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
1723                               LCK_EX);
1724         if (rc != 0)
1725                 GOTO(log, rc);
1726
1727         if (next)
1728                 lfsck_linkea_del_buf(ldata, cname);
1729         else
1730                 lfsck_namespace_filter_linkea_entry(ldata, cname, pfid,
1731                                                     true);
1732         if (ldata->ld_leh->leh_reccount > 0 ||
1733             unlikely(ldata->ld_leh->leh_overflow_time)) {
1734                 lfsck_buf_init(&linkea_buf, ldata->ld_buf->lb_buf,
1735                                ldata->ld_leh->leh_len);
1736                 buflen = linkea_buf.lb_len;
1737         }
1738
1739 again:
1740         th = dt_trans_create(env, dev);
1741         if (IS_ERR(th))
1742                 GOTO(unlock1, rc = PTR_ERR(th));
1743
1744         if (buflen != 0)
1745                 rc = dt_declare_xattr_set(env, obj, &linkea_buf,
1746                                           XATTR_NAME_LINK, 0, th);
1747         else
1748                 rc = dt_declare_xattr_del(env, obj, XATTR_NAME_LINK, th);
1749         if (rc != 0)
1750                 GOTO(stop, rc);
1751
1752         rc = dt_trans_start_local(env, dev, th);
1753         if (rc != 0)
1754                 GOTO(stop, rc);
1755
1756         dt_write_lock(env, obj, 0);
1757         if (unlikely(lfsck_is_dead_obj(obj)))
1758                 GOTO(unlock2, rc = -ENOENT);
1759
1760         rc = lfsck_links_read2_with_rec(env, obj, &ldata_new);
1761         if (rc)
1762                 GOTO(unlock2, rc = (rc == -ENODATA ? 0 : rc));
1763
1764         /* The specified linkEA entry has been removed by race. */
1765         rc = linkea_links_find(&ldata_new, cname, pfid);
1766         if (rc != 0)
1767                 GOTO(unlock2, rc = 0);
1768
1769         if (bk->lb_param & LPF_DRYRUN)
1770                 GOTO(unlock2, rc = 1);
1771
1772         if (next)
1773                 lfsck_linkea_del_buf(&ldata_new, cname);
1774         else
1775                 lfsck_namespace_filter_linkea_entry(&ldata_new, cname, pfid,
1776                                                     true);
1777
1778         /*
1779          * linkea may change because it doesn't take lock in the first read, if
1780          * it becomes larger, restart from beginning.
1781          */
1782         if ((ldata_new.ld_leh->leh_reccount > 0 ||
1783              unlikely(ldata_new.ld_leh->leh_overflow_time)) &&
1784             buflen < ldata_new.ld_leh->leh_len) {
1785                 dt_write_unlock(env, obj);
1786                 dt_trans_stop(env, dev, th);
1787                 lfsck_buf_init(&linkea_buf, ldata_new.ld_buf->lb_buf,
1788                                ldata_new.ld_leh->leh_len);
1789                 buflen = linkea_buf.lb_len;
1790                 goto again;
1791         }
1792
1793         if (buflen)
1794                 rc = lfsck_links_write(env, obj, &ldata_new, th);
1795         else
1796                 rc = dt_xattr_del(env, obj, XATTR_NAME_LINK, th);
1797
1798         GOTO(unlock2, rc = (rc == 0 ? 1 : rc));
1799
1800 unlock2:
1801         dt_write_unlock(env, obj);
1802
1803 stop:
1804         dt_trans_stop(env, dev, th);
1805
1806 unlock1:
1807         lfsck_ibits_unlock(&lh, LCK_EX);
1808
1809 log:
1810         CDEBUG(D_LFSCK, "%s: namespace LFSCK remove %s linkEA entry "
1811                "for the object: "DFID", parent "DFID", name %.*s\n",
1812                lfsck_lfsck2name(lfsck), next ? "invalid" : "redundant",
1813                PFID(lfsck_dto2fid(obj)), PFID(pfid), cname->ln_namelen,
1814                cname->ln_name);
1815
1816         if (rc != 0) {
1817                 struct lfsck_namespace *ns = com->lc_file_ram;
1818
1819                 ns->ln_flags |= LF_INCONSISTENT;
1820         }
1821
1822         return rc;
1823 }
1824
1825 /**
1826  * Conditionally remove the specified entry from the linkEA.
1827  *
1828  * Take the parent lock firstly, then check whether the specified
1829  * name entry exists or not: if yes, do nothing; otherwise, call
1830  * lfsck_namespace_shrink_linkea() to remove the linkea entry.
1831  *
1832  * \param[in] env       pointer to the thread context
1833  * \param[in] com       pointer to the lfsck component
1834  * \param[in] parent    pointer to the parent directory
1835  * \param[in] child     pointer to the child object that holds the linkEA
1836  * \param[in,out]ldata  pointer to the buffer that holds the linkEA
1837  * \param[in] cname     the name for the child in the parent directory
1838  * \param[in] pfid      the parent directory's FID for the linkEA
1839  *
1840  * \retval              positive number for repaired cases
1841  * \retval              0 if nothing to be repaired
1842  * \retval              negative error number on failure
1843  */
1844 static int lfsck_namespace_shrink_linkea_cond(const struct lu_env *env,
1845                                               struct lfsck_component *com,
1846                                               struct dt_object *parent,
1847                                               struct dt_object *child,
1848                                               struct linkea_data *ldata,
1849                                               struct lu_name *cname,
1850                                               struct lu_fid *pfid)
1851 {
1852         struct lfsck_thread_info *info  = lfsck_env_info(env);
1853         struct lu_fid            *cfid  = &info->lti_fid3;
1854         struct lfsck_lock_handle *llh   = &info->lti_llh;
1855         int                       rc;
1856         ENTRY;
1857
1858         rc = lfsck_lock(env, com->lc_lfsck, parent, cname->ln_name, llh,
1859                         MDS_INODELOCK_UPDATE, LCK_PR);
1860         if (rc != 0)
1861                 RETURN(rc);
1862
1863         dt_read_lock(env, parent, 0);
1864         if (unlikely(lfsck_is_dead_obj(parent))) {
1865                 dt_read_unlock(env, parent);
1866                 lfsck_unlock(llh);
1867                 rc = lfsck_namespace_shrink_linkea(env, com, child, ldata,
1868                                                    cname, pfid, true);
1869
1870                 RETURN(rc);
1871         }
1872
1873         rc = dt_lookup(env, parent, (struct dt_rec *)cfid,
1874                        (const struct dt_key *)cname->ln_name);
1875         dt_read_unlock(env, parent);
1876
1877         /* It is safe to release the ldlm lock, because when the logic come
1878          * here, we have got all the needed information above whether the
1879          * linkEA entry is valid or not. It is not important that others
1880          * may add new linkEA entry after the ldlm lock released. If other
1881          * has removed the specified linkEA entry by race, then it is OK,
1882          * because the subsequent lfsck_namespace_shrink_linkea() can handle
1883          * such case. */
1884         lfsck_unlock(llh);
1885         if (rc == -ENOENT) {
1886                 rc = lfsck_namespace_shrink_linkea(env, com, child, ldata,
1887                                                    cname, pfid, true);
1888
1889                 RETURN(rc);
1890         }
1891
1892         if (rc != 0)
1893                 RETURN(rc);
1894
1895         /* The LFSCK just found some internal status of cross-MDTs
1896          * create operation. That is normal. */
1897         if (lu_fid_eq(cfid, lfsck_dto2fid(child))) {
1898                 linkea_next_entry(ldata);
1899
1900                 RETURN(0);
1901         }
1902
1903         rc = lfsck_namespace_shrink_linkea(env, com, child, ldata, cname,
1904                                            pfid, true);
1905
1906         RETURN(rc);
1907 }
1908
1909 /**
1910  * Conditionally replace name entry in the parent.
1911  *
1912  * As required, the LFSCK may re-create the lost MDT-object for dangling
1913  * name entry, but such repairing may be wrong because of bad FID in the
1914  * name entry. As the LFSCK processing, the real MDT-object may be found,
1915  * then the LFSCK should check whether the former re-created MDT-object
1916  * has been modified or not, if not, then destroy it and update the name
1917  * entry in the parent to reference the real MDT-object.
1918  *
1919  * \param[in] env       pointer to the thread context
1920  * \param[in] com       pointer to the lfsck component
1921  * \param[in] parent    pointer to the parent directory
1922  * \param[in] child     pointer to the MDT-object that may be the real
1923  *                      MDT-object corresponding to the name entry in parent
1924  * \param[in] cfid      the current FID in the name entry
1925  * \param[in] cname     contains the name of the child in the parent directory
1926  *
1927  * \retval              positive number for repaired cases
1928  * \retval              0 if nothing to be repaired
1929  * \retval              negative error number on failure
1930  */
1931 static int lfsck_namespace_replace_cond(const struct lu_env *env,
1932                                         struct lfsck_component *com,
1933                                         struct dt_object *parent,
1934                                         struct dt_object *child,
1935                                         const struct lu_fid *cfid,
1936                                         const struct lu_name *cname)
1937 {
1938         struct lfsck_thread_info        *info   = lfsck_env_info(env);
1939         struct lu_attr                  *la     = &info->lti_la;
1940         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
1941         struct lu_fid                    tfid;
1942         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1943         /* The child and its name may be on different MDTs. */
1944         struct dt_device                *dev    = lfsck->li_next;
1945         const char                      *name   = cname->ln_name;
1946         const struct lu_fid             *pfid   = lfsck_dto2fid(parent);
1947         struct dt_object                *cobj   = NULL;
1948         struct lfsck_lock_handle        *pllh   = &info->lti_llh;
1949         struct lustre_handle             clh    = { 0 };
1950         struct linkea_data               ldata  = { NULL };
1951         struct thandle                  *th     = NULL;
1952         bool                             exist  = true;
1953         int                              rc     = 0;
1954         ENTRY;
1955
1956         /* @parent/@child may be based on lfsck->li_bottom,
1957          * but here we need the object based on the lfsck->li_next. */
1958
1959         parent = lfsck_object_locate(dev, parent);
1960         if (IS_ERR(parent))
1961                 GOTO(log, rc = PTR_ERR(parent));
1962
1963         if (unlikely(!dt_try_as_dir(env, parent)))
1964                 GOTO(log, rc = -ENOTDIR);
1965
1966         rc = lfsck_lock(env, lfsck, parent, name, pllh,
1967                         MDS_INODELOCK_UPDATE, LCK_PW);
1968         if (rc != 0)
1969                 GOTO(log, rc);
1970
1971         if (!fid_is_sane(cfid)) {
1972                 exist = false;
1973                 goto replace;
1974         }
1975
1976         cobj = lfsck_object_find_by_dev(env, dev, cfid);
1977         if (IS_ERR(cobj)) {
1978                 rc = PTR_ERR(cobj);
1979                 if (rc == -ENOENT) {
1980                         exist = false;
1981                         goto replace;
1982                 }
1983
1984                 GOTO(log, rc);
1985         }
1986
1987         if (!dt_object_exists(cobj)) {
1988                 exist = false;
1989                 goto replace;
1990         }
1991
1992         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
1993                        (const struct dt_key *)name);
1994         if (rc == -ENOENT) {
1995                 exist = false;
1996                 goto replace;
1997         }
1998
1999         if (rc != 0)
2000                 GOTO(log, rc);
2001
2002         /* Someone changed the name entry, cannot replace it. */
2003         if (!lu_fid_eq(cfid, &tfid))
2004                 GOTO(log, rc = 0);
2005
2006         /* lock the object to be destroyed. */
2007         rc = lfsck_ibits_lock(env, lfsck, cobj, &clh,
2008                               MDS_INODELOCK_UPDATE |
2009                               MDS_INODELOCK_UPDATE | MDS_INODELOCK_XATTR,
2010                               LCK_EX);
2011         if (rc != 0)
2012                 GOTO(log, rc);
2013
2014         if (unlikely(lfsck_is_dead_obj(cobj))) {
2015                 exist = false;
2016                 goto replace;
2017         }
2018
2019         rc = dt_attr_get(env, cobj, la);
2020         if (rc != 0)
2021                 GOTO(log, rc);
2022
2023         /* The object has been modified by other(s), or it is not created by
2024          * LFSCK, the two cases are indistinguishable. So cannot replace it. */
2025         if (la->la_ctime != 0)
2026                 GOTO(log, rc);
2027
2028         if (S_ISREG(la->la_mode)) {
2029                 rc = dt_xattr_get(env, cobj, &LU_BUF_NULL, XATTR_NAME_LOV);
2030                 /* If someone has created related OST-object(s),
2031                  * then keep it. */
2032                 if ((rc > 0) || (rc < 0 && rc != -ENODATA))
2033                         GOTO(log, rc = (rc > 0 ? 0 : rc));
2034         }
2035
2036 replace:
2037         dt_read_lock(env, child, 0);
2038         rc = lfsck_links_read2_with_rec(env, child, &ldata);
2039         dt_read_unlock(env, child);
2040
2041         /* Someone changed the child, no need to replace. */
2042         if (rc == -ENODATA)
2043                 GOTO(log, rc = 0);
2044
2045         if (rc != 0)
2046                 GOTO(log, rc);
2047
2048         rc = linkea_links_find(&ldata, cname, pfid);
2049         /* Someone moved the child, no need to replace. */
2050         if (rc != 0)
2051                 GOTO(log, rc = 0);
2052
2053         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
2054                 GOTO(log, rc = 1);
2055
2056         th = dt_trans_create(env, dev);
2057         if (IS_ERR(th))
2058                 GOTO(log, rc = PTR_ERR(th));
2059
2060         if (exist) {
2061                 rc = dt_declare_destroy(env, cobj, th);
2062                 if (rc != 0)
2063                         GOTO(stop, rc);
2064         }
2065
2066         rc = dt_declare_delete(env, parent, (const struct dt_key *)name, th);
2067         if (rc != 0)
2068                 GOTO(stop, rc);
2069
2070         rec->rec_type = S_IFDIR;
2071         rec->rec_fid = lfsck_dto2fid(child);
2072         rc = dt_declare_insert(env, parent, (const struct dt_rec *)rec,
2073                                (const struct dt_key *)name, th);
2074         if (rc != 0)
2075                 GOTO(stop, rc);
2076
2077         rc = dt_trans_start_local(env, dev, th);
2078         if (rc != 0)
2079                 GOTO(stop, rc);
2080
2081         if (exist) {
2082                 rc = dt_destroy(env, cobj, th);
2083                 if (rc != 0)
2084                         GOTO(stop, rc);
2085         }
2086
2087         /* The old name entry maybe not exist. */
2088         rc = dt_delete(env, parent, (const struct dt_key *)name, th);
2089         if (rc != 0 && rc != -ENOENT)
2090                 GOTO(stop, rc);
2091
2092         rc = dt_insert(env, parent, (const struct dt_rec *)rec,
2093                        (const struct dt_key *)name, th);
2094
2095         GOTO(stop, rc = (rc == 0 ? 1 : rc));
2096
2097 stop:
2098         dt_trans_stop(env, dev, th);
2099
2100 log:
2101         lfsck_ibits_unlock(&clh, LCK_EX);
2102         lfsck_unlock(pllh);
2103
2104         if (cobj != NULL && !IS_ERR(cobj))
2105                 lfsck_object_put(env, cobj);
2106
2107         CDEBUG(D_LFSCK, "%s: namespace LFSCK conditionally destroy the "
2108                "object "DFID" because of conflict with the object "DFID
2109                " under the parent "DFID" with name %s: rc = %d\n",
2110                lfsck_lfsck2name(lfsck), PFID(cfid),
2111                PFID(lfsck_dto2fid(child)), PFID(pfid), name, rc);
2112
2113         return rc;
2114 }
2115
2116 /**
2117  * Overwrite the linkEA for the object with the given ldata.
2118  *
2119  * The caller should take the ldlm lock before the calling.
2120  *
2121  * \param[in] env       pointer to the thread context
2122  * \param[in] com       pointer to the lfsck component
2123  * \param[in] obj       pointer to the dt_object to be handled
2124  * \param[in] ldata     pointer to the new linkEA data
2125  *
2126  * \retval              positive number for repaired cases
2127  * \retval              0 if nothing to be repaired
2128  * \retval              negative error number on failure
2129  */
2130 int lfsck_namespace_rebuild_linkea(const struct lu_env *env,
2131                                    struct lfsck_component *com,
2132                                    struct dt_object *obj,
2133                                    struct linkea_data *ldata)
2134 {
2135         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2136         struct dt_device                *dev    = lfsck_obj2dev(obj);
2137         struct thandle                  *th     = NULL;
2138         struct lu_buf                    linkea_buf;
2139         int                              rc     = 0;
2140         ENTRY;
2141
2142         th = dt_trans_create(env, dev);
2143         if (IS_ERR(th))
2144                 GOTO(log, rc = PTR_ERR(th));
2145
2146         lfsck_buf_init(&linkea_buf, ldata->ld_buf->lb_buf,
2147                        ldata->ld_leh->leh_len);
2148         rc = dt_declare_xattr_set(env, obj, &linkea_buf,
2149                                   XATTR_NAME_LINK, 0, th);
2150         if (rc != 0)
2151                 GOTO(stop, rc);
2152
2153         rc = dt_trans_start_local(env, dev, th);
2154         if (rc != 0)
2155                 GOTO(stop, rc);
2156
2157         dt_write_lock(env, obj, 0);
2158         if (unlikely(lfsck_is_dead_obj(obj)))
2159                 GOTO(unlock, rc = 0);
2160
2161         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
2162                 GOTO(unlock, rc = 1);
2163
2164         rc = dt_xattr_set(env, obj, &linkea_buf,
2165                           XATTR_NAME_LINK, 0, th);
2166
2167         GOTO(unlock, rc = (rc == 0 ? 1 : rc));
2168
2169 unlock:
2170         dt_write_unlock(env, obj);
2171
2172 stop:
2173         dt_trans_stop(env, dev, th);
2174
2175 log:
2176         CDEBUG(D_LFSCK, "%s: namespace LFSCK rebuild linkEA for the "
2177                "object "DFID": rc = %d\n",
2178                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(obj)), rc);
2179
2180         if (rc != 0) {
2181                 struct lfsck_namespace *ns = com->lc_file_ram;
2182
2183                 ns->ln_flags |= LF_INCONSISTENT;
2184         }
2185
2186         return rc;
2187 }
2188
2189 /**
2190  * Repair invalid name entry.
2191  *
2192  * If the name entry contains invalid information, such as bad file type
2193  * or (and) corrupted object FID, then either remove the name entry or
2194  * udpate the name entry with the given (right) information.
2195  *
2196  * \param[in] env       pointer to the thread context
2197  * \param[in] com       pointer to the lfsck component
2198  * \param[in] parent    pointer to the parent directory
2199  * \param[in] child     pointer to the object referenced by the name entry
2200  * \param[in] name      the old name of the child under the parent directory
2201  * \param[in] name2     the new name of the child under the parent directory
2202  * \param[in] type      the type claimed by the name entry
2203  * \param[in] update    update the name entry if true; otherwise, remove it
2204  * \param[in] dec       decrease the parent nlink count if true
2205  *
2206  * \retval              positive number for repaired successfully
2207  * \retval              0 if nothing to be repaired
2208  * \retval              negative error number on failure
2209  */
2210 int lfsck_namespace_repair_dirent(const struct lu_env *env,
2211                                   struct lfsck_component *com,
2212                                   struct dt_object *parent,
2213                                   struct dt_object *child,
2214                                   const char *name, const char *name2,
2215                                   __u16 type, bool update, bool dec)
2216 {
2217         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2218         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
2219         const struct lu_fid             *pfid   = lfsck_dto2fid(parent);
2220         struct lu_fid                   cfid    = {0};
2221         struct lu_fid                    tfid;
2222         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2223         struct dt_device                *dev    = lfsck->li_next;
2224         struct thandle                  *th     = NULL;
2225         struct lfsck_lock_handle        *llh    = &info->lti_llh;
2226         struct lustre_handle             lh     = { 0 };
2227         int                              rc     = 0;
2228         ENTRY;
2229
2230         if (child)
2231                 cfid = *lfsck_dto2fid(child);
2232         parent = lfsck_object_locate(dev, parent);
2233         if (IS_ERR(parent))
2234                 GOTO(log, rc = PTR_ERR(parent));
2235
2236         if (unlikely(!dt_try_as_dir(env, parent)))
2237                 GOTO(log, rc = -ENOTDIR);
2238
2239         if (!update || strcmp(name, name2) == 0)
2240                 rc = lfsck_lock(env, lfsck, parent, name, llh,
2241                                 MDS_INODELOCK_UPDATE, LCK_PW);
2242         else
2243                 rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
2244                                       MDS_INODELOCK_UPDATE, LCK_PW);
2245         if (rc != 0)
2246                 GOTO(log, rc);
2247
2248         th = dt_trans_create(env, dev);
2249         if (IS_ERR(th))
2250                 GOTO(unlock1, rc = PTR_ERR(th));
2251
2252         rc = dt_declare_delete(env, parent, (const struct dt_key *)name, th);
2253         if (rc != 0)
2254                 GOTO(stop, rc);
2255
2256         if (update) {
2257                 rec->rec_type = lfsck_object_type(child) & S_IFMT;
2258                 LASSERT(!fid_is_zero(&cfid));
2259                 rec->rec_fid = &cfid;
2260                 rc = dt_declare_insert(env, parent,
2261                                        (const struct dt_rec *)rec,
2262                                        (const struct dt_key *)name2, th);
2263                 if (rc != 0)
2264                         GOTO(stop, rc);
2265         }
2266
2267         if (dec && S_ISDIR(type)) {
2268                 rc = dt_declare_ref_del(env, parent, th);
2269                 if (rc != 0)
2270                         GOTO(stop, rc);
2271         }
2272
2273         rc = dt_trans_start_local(env, dev, th);
2274         if (rc != 0)
2275                 GOTO(stop, rc);
2276
2277
2278         dt_write_lock(env, parent, 0);
2279         rc = dt_lookup(env, dt_object_child(parent), (struct dt_rec *)&tfid,
2280                        (const struct dt_key *)name);
2281         /* Someone has removed the bad name entry by race. */
2282         if (rc == -ENOENT)
2283                 GOTO(unlock2, rc = 0);
2284
2285         if (rc != 0)
2286                 GOTO(unlock2, rc);
2287
2288         /* Someone has removed the bad name entry and reused it for other
2289          * object by race. */
2290         if (!lu_fid_eq(&tfid, &cfid))
2291                 GOTO(unlock2, rc = 0);
2292
2293         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
2294                 GOTO(unlock2, rc = 1);
2295
2296         rc = dt_delete(env, parent, (const struct dt_key *)name, th);
2297         if (rc != 0)
2298                 GOTO(unlock2, rc);
2299
2300         if (update) {
2301                 rc = dt_insert(env, parent,
2302                                (const struct dt_rec *)rec,
2303                                (const struct dt_key *)name2, th);
2304                 if (rc != 0)
2305                         GOTO(unlock2, rc);
2306         }
2307
2308         if (dec && S_ISDIR(type)) {
2309                 rc = dt_ref_del(env, parent, th);
2310                 if (rc != 0)
2311                         GOTO(unlock2, rc);
2312         }
2313
2314         GOTO(unlock2, rc = (rc == 0 ? 1 : rc));
2315
2316 unlock2:
2317         dt_write_unlock(env, parent);
2318
2319 stop:
2320         dt_trans_stop(env, dev, th);
2321
2322         /* We are not sure whether the child will become orphan or not.
2323          * Record it in the LFSCK trace file for further checking in
2324          * the second-stage scanning. */
2325         if (!update && !dec && child && rc == 0)
2326                 lfsck_namespace_trace_update(env, com, &cfid,
2327                                              LNTF_CHECK_LINKEA, true);
2328
2329 unlock1:
2330         /* It is harmless even if unlock the unused lock_handle */
2331         lfsck_ibits_unlock(&lh, LCK_PW);
2332         lfsck_unlock(llh);
2333
2334 log:
2335         CDEBUG(D_LFSCK, "%s: namespace LFSCK assistant found bad name "
2336                "entry for: parent "DFID", child "DFID", name %s, type "
2337                "in name entry %o, type claimed by child %o. repair it "
2338                "by %s with new name2 %s: rc = %d\n",
2339                lfsck_lfsck2name(lfsck), PFID(pfid), PFID(&cfid),
2340                name, type, update ? lfsck_object_type(child) : 0,
2341                update ? "updating" : "removing", name2, rc);
2342
2343         if (rc != 0) {
2344                 struct lfsck_namespace *ns = com->lc_file_ram;
2345
2346                 ns->ln_flags |= LF_INCONSISTENT;
2347         }
2348
2349         return rc;
2350 }
2351
2352 /**
2353  * Update the ".." name entry for the given object.
2354  *
2355  * The object's ".." is corrupted, this function will update the ".." name
2356  * entry with the given pfid, and the linkEA with the given ldata.
2357  *
2358  * The caller should take the ldlm lock before the calling.
2359  *
2360  * \param[in] env       pointer to the thread context
2361  * \param[in] com       pointer to the lfsck component
2362  * \param[in] obj       pointer to the dt_object to be handled
2363  * \param[in] pfid      the new fid for the object's ".." name entry
2364  * \param[in] cname     the name for the @obj in the parent directory
2365  *
2366  * \retval              positive number for repaired cases
2367  * \retval              0 if nothing to be repaired
2368  * \retval              negative error number on failure
2369  */
2370 static int lfsck_namespace_repair_unmatched_pairs(const struct lu_env *env,
2371                                                   struct lfsck_component *com,
2372                                                   struct dt_object *obj,
2373                                                   const struct lu_fid *pfid,
2374                                                   struct lu_name *cname)
2375 {
2376         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2377         struct dt_insert_rec            *rec    = &info->lti_dt_rec;
2378         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2379         struct dt_device                *dev    = lfsck_obj2dev(obj);
2380         struct thandle                  *th     = NULL;
2381         struct linkea_data               ldata  = { NULL };
2382         struct lu_buf                    linkea_buf;
2383         int                              rc     = 0;
2384         ENTRY;
2385
2386         LASSERT(!dt_object_remote(obj));
2387         LASSERT(S_ISDIR(lfsck_object_type(obj)));
2388
2389         rc = linkea_links_new(&ldata, &info->lti_big_buf, cname, pfid);
2390         if (rc != 0)
2391                 GOTO(log, rc);
2392
2393         lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
2394                        ldata.ld_leh->leh_len);
2395
2396         th = dt_trans_create(env, dev);
2397         if (IS_ERR(th))
2398                 GOTO(log, rc = PTR_ERR(th));
2399
2400         rc = dt_declare_delete(env, obj, (const struct dt_key *)dotdot, th);
2401         if (rc != 0)
2402                 GOTO(stop, rc);
2403
2404         rec->rec_type = S_IFDIR;
2405         rec->rec_fid = pfid;
2406         rc = dt_declare_insert(env, obj, (const struct dt_rec *)rec,
2407                                (const struct dt_key *)dotdot, th);
2408         if (rc != 0)
2409                 GOTO(stop, rc);
2410
2411         rc = dt_declare_xattr_set(env, obj, &linkea_buf,
2412                                   XATTR_NAME_LINK, 0, th);
2413         if (rc != 0)
2414                 GOTO(stop, rc);
2415
2416         rc = dt_trans_start_local(env, dev, th);
2417         if (rc != 0)
2418                 GOTO(stop, rc);
2419
2420         dt_write_lock(env, obj, 0);
2421         if (unlikely(lfsck_is_dead_obj(obj)))
2422                 GOTO(unlock, rc = 0);
2423
2424         if (lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN)
2425                 GOTO(unlock, rc = 1);
2426
2427         /* The old ".." name entry maybe not exist. */
2428         dt_delete(env, obj, (const struct dt_key *)dotdot, th);
2429
2430         rc = dt_insert(env, obj, (const struct dt_rec *)rec,
2431                        (const struct dt_key *)dotdot, th);
2432         if (rc != 0)
2433                 GOTO(unlock, rc);
2434
2435         rc = lfsck_links_write(env, obj, &ldata, th);
2436
2437         GOTO(unlock, rc = (rc == 0 ? 1 : rc));
2438
2439 unlock:
2440         dt_write_unlock(env, obj);
2441
2442 stop:
2443         dt_trans_stop(env, dev, th);
2444
2445 log:
2446         CDEBUG(D_LFSCK, "%s: namespace LFSCK rebuild dotdot name entry for "
2447                "the object "DFID", new parent "DFID": rc = %d\n",
2448                lfsck_lfsck2name(lfsck), PFID(lfsck_dto2fid(obj)),
2449                PFID(pfid), rc);
2450
2451         if (rc != 0) {
2452                 struct lfsck_namespace *ns = com->lc_file_ram;
2453
2454                 ns->ln_flags |= LF_INCONSISTENT;
2455         }
2456
2457         return rc;
2458 }
2459
2460 /**
2461  * Handle orphan @obj during Double Scan Directory.
2462  *
2463  * Remove the @obj's current (invalid) linkEA entries, and insert
2464  * it in the directory .lustre/lost+found/MDTxxxx/ with the name:
2465  * ${FID}-${PFID}-D-${conflict_version}
2466  *
2467  * The caller should take the ldlm lock before the calling.
2468  *
2469  * \param[in] env       pointer to the thread context
2470  * \param[in] com       pointer to the lfsck component
2471  * \param[in] obj       pointer to the orphan object to be handled
2472  * \param[in] pfid      the new fid for the object's ".." name entry
2473  * \param[in,out] lh    ldlm lock handler for the given @obj
2474  * \param[out] type     to tell the caller what the inconsistency is
2475  *
2476  * \retval              positive number for repaired cases
2477  * \retval              0 if nothing to be repaired
2478  * \retval              negative error number on failure
2479  */
2480 static int
2481 lfsck_namespace_dsd_orphan(const struct lu_env *env,
2482                            struct lfsck_component *com,
2483                            struct dt_object *obj,
2484                            const struct lu_fid *pfid,
2485                            struct lustre_handle *lh,
2486                            enum lfsck_namespace_inconsistency_type *type)
2487 {
2488         struct lfsck_thread_info *info = lfsck_env_info(env);
2489         struct lfsck_namespace   *ns   = com->lc_file_ram;
2490         int                       rc;
2491         ENTRY;
2492
2493         /* Remove the unrecognized linkEA. */
2494         rc = lfsck_namespace_links_remove(env, com, obj);
2495         lfsck_ibits_unlock(lh, LCK_EX);
2496         if (rc < 0 && rc != -ENODATA)
2497                 RETURN(rc);
2498
2499         *type = LNIT_MUL_REF;
2500
2501         /* If the LFSCK is marked as LF_INCOMPLETE, then means some MDT has
2502          * ever tried to verify some remote MDT-object that resides on this
2503          * MDT, but this MDT failed to respond such request. So means there
2504          * may be some remote name entry on other MDT that references this
2505          * object with another name, so we cannot know whether this linkEA
2506          * is valid or not. So keep it there and maybe resolved when next
2507          * LFSCK run. */
2508         if (ns->ln_flags & LF_INCOMPLETE)
2509                 RETURN(0);
2510
2511         /* The unique linkEA is invalid, even if the ".." name entry may be
2512          * valid, we still cannot know via which name entry this directory
2513          * will be referenced. Then handle it as pure orphan. */
2514         snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf),
2515                  "-"DFID, PFID(pfid));
2516         rc = lfsck_namespace_insert_orphan(env, com, obj,
2517                                            info->lti_tmpbuf, "D", NULL);
2518
2519         RETURN(rc);
2520 }
2521
2522 /**
2523  * Double Scan Directory object for single linkEA entry case.
2524  *
2525  * The given @child has unique linkEA entry. If the linkEA entry is valid,
2526  * then check whether the name is in the namespace or not, if not, add the
2527  * missing name entry back to namespace. If the linkEA entry is invalid,
2528  * then remove it and insert the @child in the .lustre/lost+found/MDTxxxx/
2529  * as an orphan.
2530  *
2531  * \param[in] env       pointer to the thread context
2532  * \param[in] com       pointer to the lfsck component
2533  * \param[in] child     pointer to the directory to be double scanned
2534  * \param[in] pfid      the FID corresponding to the ".." entry
2535  * \param[in] ldata     pointer to the linkEA data for the given @child
2536  * \param[in,out] lh    ldlm lock handler for the given @child
2537  * \param[out] type     to tell the caller what the inconsistency is
2538  * \param[in] retry     if found inconsistency, but the caller does not hold
2539  *                      ldlm lock on the @child, then set @retry as true
2540  * \param[in] unknown   set if does not know how to repair the inconsistency
2541  *
2542  * \retval              positive number for repaired cases
2543  * \retval              0 if nothing to be repaired
2544  * \retval              negative error number on failure
2545  */
2546 static int
2547 lfsck_namespace_dsd_single(const struct lu_env *env,
2548                            struct lfsck_component *com,
2549                            struct dt_object *child,
2550                            const struct lu_fid *pfid,
2551                            struct linkea_data *ldata,
2552                            struct lustre_handle *lh,
2553                            enum lfsck_namespace_inconsistency_type *type,
2554                            bool *retry, bool *unknown)
2555 {
2556         struct lfsck_thread_info *info          = lfsck_env_info(env);
2557         struct lu_name           *cname         = &info->lti_name;
2558         const struct lu_fid      *cfid          = lfsck_dto2fid(child);
2559         struct lu_fid             tfid;
2560         struct lfsck_namespace   *ns            = com->lc_file_ram;
2561         struct lfsck_instance    *lfsck         = com->lc_lfsck;
2562         struct dt_object         *parent        = NULL;
2563         struct lmv_mds_md_v1     *lmv;
2564         int                       rc            = 0;
2565         ENTRY;
2566
2567         rc = lfsck_namespace_unpack_linkea_entry(ldata, cname, &tfid,
2568                                                  info->lti_key,
2569                                                  sizeof(info->lti_key));
2570         /* The unique linkEA entry with bad parent will be handled as orphan. */
2571         if (rc != 0) {
2572                 if (!lustre_handle_is_used(lh) && retry != NULL)
2573                         *retry = true;
2574                 else
2575                         rc = lfsck_namespace_dsd_orphan(env, com, child,
2576                                                         pfid, lh, type);
2577
2578                 GOTO(out, rc);
2579         }
2580
2581         parent = lfsck_object_find_bottom(env, lfsck, &tfid);
2582         if (IS_ERR(parent))
2583                 GOTO(out, rc = PTR_ERR(parent));
2584
2585         /* We trust the unique linkEA entry in spite of whether it matches the
2586          * ".." name entry or not. Because even if the linkEA entry is wrong
2587          * and the ".." name entry is right, we still cannot know via which
2588          * name entry the child will be referenced, since all known entries
2589          * have been verified during the first-stage scanning. */
2590         if (!dt_object_exists(parent)) {
2591                 /* If the LFSCK is marked as LF_INCOMPLETE, then means some MDT
2592                  * has ever tried to verify some remote MDT-object that resides
2593                  * on this MDT, but this MDT failed to respond such request. So
2594                  * means there may be some remote name entry on other MDT that
2595                  * references this object with another name, so we cannot know
2596                  * whether this linkEA is valid or not. So keep it there and
2597                  * maybe resolved when next LFSCK run. */
2598                 if (ns->ln_flags & LF_INCOMPLETE)
2599                         GOTO(out, rc = 0);
2600
2601                 if (!lustre_handle_is_used(lh) && retry != NULL) {
2602                         *retry = true;
2603
2604                         GOTO(out, rc = 0);
2605                 }
2606
2607                 lfsck_ibits_unlock(lh, LCK_EX);
2608
2609 lost_parent:
2610                 lmv = &info->lti_lmv;
2611                 rc = lfsck_read_stripe_lmv(env, lfsck, child, lmv);
2612                 if (rc != 0 && rc != -ENODATA)
2613                         GOTO(out, rc);
2614
2615                 if (rc == -ENODATA || lmv->lmv_magic != LMV_MAGIC_STRIPE) {
2616                         lmv = NULL;
2617                 } else if (lfsck_shard_name_to_index(env,
2618                                         cname->ln_name, cname->ln_namelen,
2619                                         S_IFDIR, cfid) < 0) {
2620                         /* It is an invalid name entry, we
2621                          * cannot trust the parent also. */
2622                         rc = lfsck_namespace_shrink_linkea(env, com, child,
2623                                                 ldata, cname, &tfid, true);
2624                         if (rc < 0)
2625                                 GOTO(out, rc);
2626
2627                         snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf),
2628                                  "-"DFID, PFID(pfid));
2629                         rc = lfsck_namespace_insert_orphan(env, com, child,
2630                                                 info->lti_tmpbuf, "S", NULL);
2631
2632                         GOTO(out, rc);
2633                 }
2634
2635                 /* Create the lost parent as an orphan. */
2636                 rc = lfsck_namespace_create_orphan_dir(env, com, parent, lmv);
2637                 if (rc >= 0) {
2638                         /* Add the missing name entry to the parent. */
2639                         rc = lfsck_namespace_insert_normal(env, com, parent,
2640                                                            child, cname);
2641                         if (unlikely(rc == -EEXIST)) {
2642                                 /* Unfortunately, someone reused the name
2643                                  * under the parent by race. So we have
2644                                  * to remove the linkEA entry from
2645                                  * current child object. It means that the
2646                                  * LFSCK cannot recover the system
2647                                  * totally back to its original status,
2648                                  * but it is necessary to make the
2649                                  * current system to be consistent. */
2650                                 rc = lfsck_namespace_shrink_linkea(env,
2651                                                 com, child, ldata,
2652                                                 cname, &tfid, true);
2653                                 if (rc >= 0) {
2654                                         snprintf(info->lti_tmpbuf,
2655                                                  sizeof(info->lti_tmpbuf),
2656                                                  "-"DFID, PFID(pfid));
2657                                         rc = lfsck_namespace_insert_orphan(env,
2658                                                 com, child, info->lti_tmpbuf,
2659                                                 "D", NULL);
2660                                 }
2661                         }
2662                 }
2663
2664                 GOTO(out, rc);
2665         } /* !dt_object_exists(parent) */
2666
2667         /* The unique linkEA entry with bad parent will be handled as orphan. */
2668         if (unlikely(!dt_try_as_dir(env, parent))) {
2669                 if (!lustre_handle_is_used(lh) && retry != NULL)
2670                         *retry = true;
2671                 else
2672                         rc = lfsck_namespace_dsd_orphan(env, com, child,
2673                                                         pfid, lh, type);
2674
2675                 GOTO(out, rc);
2676         }
2677
2678         rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
2679                        (const struct dt_key *)cname->ln_name);
2680         if (rc == -ENOENT) {
2681                 /* If the LFSCK is marked as LF_INCOMPLETE, then means some MDT
2682                  * has ever tried to verify some remote MDT-object that resides
2683                  * on this MDT, but this MDT failed to respond such request. So
2684                  * means there may be some remote name entry on other MDT that
2685                  * references this object with another name, so we cannot know
2686                  * whether this linkEA is valid or not. So keep it there and
2687                  * maybe resolved when next LFSCK run. */
2688                 if (ns->ln_flags & LF_INCOMPLETE)
2689                         GOTO(out, rc = 0);
2690
2691                 if (!lustre_handle_is_used(lh) && retry != NULL) {
2692                         *retry = true;
2693
2694                         GOTO(out, rc = 0);
2695                 }
2696
2697                 lfsck_ibits_unlock(lh, LCK_EX);
2698                 rc = lfsck_namespace_check_name(env, lfsck, parent, child,
2699                                                 cname);
2700                 if (rc == -ENOENT)
2701                         goto lost_parent;
2702
2703                 if (rc < 0)
2704                         GOTO(out, rc);
2705
2706                 /* It is an invalid name entry, drop it. */
2707                 if (unlikely(rc > 0)) {
2708                         rc = lfsck_namespace_shrink_linkea(env, com, child,
2709                                                 ldata, cname, &tfid, true);
2710                         if (rc >= 0) {
2711                                 snprintf(info->lti_tmpbuf,
2712                                          sizeof(info->lti_tmpbuf),
2713                                          "-"DFID, PFID(pfid));
2714                                 rc = lfsck_namespace_insert_orphan(env, com,
2715                                         child, info->lti_tmpbuf, "D", NULL);
2716                         }
2717
2718                         GOTO(out, rc);
2719                 }
2720
2721                 /* Add the missing name entry back to the namespace. */
2722                 rc = lfsck_namespace_insert_normal(env, com, parent, child,
2723                                                    cname);
2724                 if (unlikely(rc == -ESTALE))
2725                         /* It may happen when the remote object has been
2726                          * removed, but the local MDT is not aware of that. */
2727                         goto lost_parent;
2728
2729                 if (unlikely(rc == -EEXIST)) {
2730                         /* Unfortunately, someone reused the name under the
2731                          * parent by race. So we have to remove the linkEA
2732                          * entry from current child object. It means that the
2733                          * LFSCK cannot recover the system totally back to
2734                          * its original status, but it is necessary to make
2735                          * the current system to be consistent.
2736                          *
2737                          * It also may be because of the LFSCK found some
2738                          * internal status of create operation. Under such
2739                          * case, nothing to be done. */
2740                         rc = lfsck_namespace_shrink_linkea_cond(env, com,
2741                                         parent, child, ldata, cname, &tfid);
2742                         if (rc >= 0) {
2743                                 snprintf(info->lti_tmpbuf,
2744                                          sizeof(info->lti_tmpbuf),
2745                                          "-"DFID, PFID(pfid));
2746                                 rc = lfsck_namespace_insert_orphan(env, com,
2747                                         child, info->lti_tmpbuf, "D", NULL);
2748                         }
2749                 }
2750
2751                 GOTO(out, rc);
2752         } /* rc == -ENOENT */
2753
2754         if (rc != 0)
2755                 GOTO(out, rc);
2756
2757         if (!lu_fid_eq(&tfid, cfid)) {
2758                 if (!lustre_handle_is_used(lh) && retry != NULL) {
2759                         *retry = true;
2760
2761                         GOTO(out, rc = 0);
2762                 }
2763
2764                 lfsck_ibits_unlock(lh, LCK_EX);
2765                 /* The name entry references another MDT-object that
2766                  * may be created by the LFSCK for repairing dangling
2767                  * name entry. Try to replace it. */
2768                 rc = lfsck_namespace_replace_cond(env, com, parent, child,
2769                                                   &tfid, cname);
2770                 if (rc == 0)
2771                         rc = lfsck_namespace_dsd_orphan(env, com, child,
2772                                                         pfid, lh, type);
2773
2774                 GOTO(out, rc);
2775         }
2776
2777         /* Zero FID may because the remote directroy object has invalid linkEA,
2778          * or lost linkEA. Under such case, the LFSCK on this MDT does not know
2779          * how to repair the inconsistency, but the namespace LFSCK on the MDT
2780          * where its name entry resides may has more information (name, FID) to
2781          * repair such inconsistency. So here, keep the inconsistency to avoid
2782          * some imporper repairing. */
2783         if (fid_is_zero(pfid)) {
2784                 if (unknown)
2785                         *unknown = true;
2786
2787                 GOTO(out, rc = 0);
2788         }
2789
2790         /* The ".." name entry is wrong, update it. */
2791         if (!lu_fid_eq(pfid, lfsck_dto2fid(parent))) {
2792                 if (!lustre_handle_is_used(lh) && retry != NULL) {
2793                         *retry = true;
2794
2795                         GOTO(out, rc = 0);
2796                 }
2797
2798                 *type = LNIT_UNMATCHED_PAIRS;
2799                 rc = lfsck_namespace_repair_unmatched_pairs(env, com, child,
2800                                                 lfsck_dto2fid(parent), cname);
2801         }
2802
2803         GOTO(out, rc);
2804
2805 out:
2806         if (parent != NULL && !IS_ERR(parent))
2807                 lfsck_object_put(env, parent);
2808
2809         return rc;
2810 }
2811
2812 /**
2813  * Double Scan Directory object for multiple linkEA entries case.
2814  *
2815  * The given @child has multiple linkEA entries. There is at most one linkEA
2816  * entry will be valid, all the others will be removed. Firstly, the function
2817  * will try to find out the linkEA entry for which the name entry exists under
2818  * the given parent (@pfid). If there is no linkEA entry that matches the given
2819  * ".." name entry, then tries to find out the first linkEA entry that both the
2820  * parent and the name entry exist to rebuild a new ".." name entry.
2821  *
2822  * \param[in] env       pointer to the thread context
2823  * \param[in] com       pointer to the lfsck component
2824  * \param[in] child     pointer to the directory to be double scanned
2825  * \param[in] pfid      the FID corresponding to the ".." entry
2826  * \param[in] ldata     pointer to the linkEA data for the given @child
2827  * \param[in,out] lh    ldlm lock handler for the given @child
2828  * \param[out] type     to tell the caller what the inconsistency is
2829  * \param[in] lpf       true if the ".." entry is under lost+found/MDTxxxx/
2830  * \param[in] unknown   set if does not know how to repair the inconsistency
2831  *
2832  * \retval              positive number for repaired cases
2833  * \retval              0 if nothing to be repaired
2834  * \retval              negative error number on failure
2835  */
2836 static int
2837 lfsck_namespace_dsd_multiple(const struct lu_env *env,
2838                              struct lfsck_component *com,
2839                              struct dt_object *child,
2840                              const struct lu_fid *pfid,
2841                              struct linkea_data *ldata,
2842                              struct lustre_handle *lh,
2843                              enum lfsck_namespace_inconsistency_type *type,
2844                              bool lpf, bool *unknown)
2845 {
2846         struct lfsck_thread_info *info          = lfsck_env_info(env);
2847         struct lu_name           *cname         = &info->lti_name;
2848         const struct lu_fid      *cfid          = lfsck_dto2fid(child);
2849         struct lu_fid            *pfid2         = &info->lti_fid3;
2850         struct lu_fid             tfid;
2851         struct lfsck_namespace   *ns            = com->lc_file_ram;
2852         struct lfsck_instance    *lfsck         = com->lc_lfsck;
2853         struct lfsck_bookmark    *bk            = &lfsck->li_bookmark_ram;
2854         struct dt_object         *parent        = NULL;
2855         struct linkea_data        ldata_new     = { NULL };
2856         int                       dirent_count  = 0;
2857         int                       rc            = 0;
2858         bool                      once          = true;
2859         ENTRY;
2860
2861 again:
2862         while (ldata->ld_lee != NULL) {
2863                 rc = lfsck_namespace_unpack_linkea_entry(ldata, cname, &tfid,
2864                                                          info->lti_key,
2865                                                          sizeof(info->lti_key));
2866                 /* Drop invalid linkEA entry. */
2867                 if (rc != 0) {
2868                         lfsck_linkea_del_buf(ldata, cname);
2869                         continue;
2870                 }
2871
2872                 /* Drop repeated linkEA entries. */
2873                 lfsck_namespace_filter_linkea_entry(ldata, cname, &tfid, true);
2874
2875                 /* If current dotdot is the .lustre/lost+found/MDTxxxx/,
2876                  * then it is possible that: the directry object has ever
2877                  * been lost, but its name entry was there. In the former
2878                  * LFSCK run, during the first-stage scanning, the LFSCK
2879                  * found the dangling name entry, but it did not recreate
2880                  * the lost object, and when moved to the second-stage
2881                  * scanning, some children objects of the lost directory
2882                  * object were found, then the LFSCK recreated such lost
2883                  * directory object as an orphan.
2884                  *
2885                  * When the LFSCK runs again, if the dangling name is still
2886                  * there, the LFSCK should move the orphan directory object
2887                  * back to the normal namespace. */
2888                 if (!lpf && !fid_is_zero(pfid) &&
2889                     !lu_fid_eq(pfid, &tfid) && once) {
2890                         linkea_next_entry(ldata);
2891                         continue;
2892                 }
2893
2894                 parent = lfsck_object_find_bottom(env, lfsck, &tfid);
2895                 if (IS_ERR(parent))
2896                         RETURN(PTR_ERR(parent));
2897
2898                 if (!dt_object_exists(parent)) {
2899                         lfsck_object_put(env, parent);
2900                         if (ldata->ld_leh->leh_reccount > 1) {
2901                                 /* If it is NOT the last linkEA entry, then
2902                                  * there is still other chance to make the
2903                                  * child to be visible via other parent, then
2904                                  * remove this linkEA entry. */
2905                                 lfsck_linkea_del_buf(ldata, cname);
2906                                 continue;
2907                         }
2908
2909                         break;
2910                 }
2911
2912                 /* The linkEA entry with bad parent will be removed. */
2913                 if (unlikely(!dt_try_as_dir(env, parent))) {
2914                         lfsck_object_put(env, parent);
2915                         lfsck_linkea_del_buf(ldata, cname);
2916                         continue;
2917                 }
2918
2919                 rc = dt_lookup(env, parent, (struct dt_rec *)&tfid,
2920                                (const struct dt_key *)cname->ln_name);
2921                 *pfid2 = *lfsck_dto2fid(parent);
2922                 if (rc == -ENOENT) {
2923                         lfsck_object_put(env, parent);
2924                         linkea_next_entry(ldata);
2925                         continue;
2926                 }
2927
2928                 if (rc != 0) {
2929                         lfsck_object_put(env, parent);
2930
2931                         RETURN(rc);
2932                 }
2933
2934                 if (lu_fid_eq(&tfid, cfid)) {
2935                         lfsck_object_put(env, parent);
2936                         /* If the parent (that is declared via linkEA entry)
2937                          * directory contains the specified child, but such
2938                          * parent does not match the dotdot name entry, then
2939                          * trust the linkEA. */
2940                         if (!fid_is_zero(pfid) && !lu_fid_eq(pfid, pfid2)) {
2941                                 *type = LNIT_UNMATCHED_PAIRS;
2942                                 rc = lfsck_namespace_repair_unmatched_pairs(env,
2943                                                 com, child, pfid2, cname);
2944
2945                                 RETURN(rc);
2946                         }
2947
2948 rebuild:
2949                         /* It is the most common case that we find the
2950                          * name entry corresponding to the linkEA entry
2951                          * that matches the ".." name entry. */
2952                         rc = linkea_links_new(&ldata_new, &info->lti_big_buf,
2953                                               cname, pfid2);
2954                         if (rc != 0)
2955                                 RETURN(rc);
2956
2957                         rc = lfsck_namespace_rebuild_linkea(env, com, child,
2958                                                             &ldata_new);
2959                         if (rc < 0)
2960                                 RETURN(rc);
2961
2962                         lfsck_linkea_del_buf(ldata, cname);
2963                         linkea_first_entry(ldata);
2964                         /* There may be some invalid dangling name entries under
2965                          * other parent directories, remove all of them. */
2966                         while (ldata->ld_lee != NULL) {
2967                                 rc = lfsck_namespace_unpack_linkea_entry(ldata,
2968                                                 cname, &tfid, info->lti_key,
2969                                                 sizeof(info->lti_key));
2970                                 if (rc != 0)
2971                                         goto next;
2972
2973                                 parent = lfsck_object_find_bottom(env, lfsck,
2974                                                                   &tfid);
2975                                 if (IS_ERR(parent)) {
2976                                         rc = PTR_ERR(parent);
2977                                         if (rc != -ENOENT &&
2978                                             bk->lb_param & LPF_FAILOUT)
2979                                                 RETURN(rc);
2980
2981                                         goto next;
2982                                 }
2983
2984                                 if (!dt_object_exists(parent)) {
2985                                         lfsck_object_put(env, parent);
2986                                         goto next;
2987                                 }
2988
2989                                 rc = lfsck_namespace_repair_dirent(env, com,
2990                                         parent, child, cname->ln_name,
2991                                         cname->ln_name, S_IFDIR, false, true);
2992                                 lfsck_object_put(env, parent);
2993                                 if (rc < 0) {
2994                                         if (bk->lb_param & LPF_FAILOUT)
2995                                                 RETURN(rc);
2996
2997                                         goto next;
2998                                 }
2999
3000                                 dirent_count += rc;
3001
3002 next:
3003                                 lfsck_linkea_del_buf(ldata, cname);
3004                         }
3005
3006                         ns->ln_dirent_repaired += dirent_count;
3007
3008                         RETURN(rc);
3009                 } /* lu_fid_eq(&tfid, lfsck_dto2fid(child)) */
3010
3011                 lfsck_ibits_unlock(lh, LCK_EX);
3012                 /* The name entry references another MDT-object that may be
3013                  * created by the LFSCK for repairing dangling name entry.
3014                  * Try to replace it. */
3015                 rc = lfsck_namespace_replace_cond(env, com, parent, child,
3016                                                   &tfid, cname);
3017                 lfsck_object_put(env, parent);
3018                 if (rc < 0)
3019                         RETURN(rc);
3020
3021                 if (rc > 0)
3022                         goto rebuild;
3023
3024                 lfsck_linkea_del_buf(ldata, cname);
3025         } /* while (ldata->ld_lee != NULL) */
3026
3027         /* If there is still linkEA overflow, return. */
3028         if (unlikely(ldata->ld_leh->leh_overflow_time))
3029                 RETURN(0);
3030
3031         linkea_first_entry(ldata);
3032         if (ldata->ld_leh->leh_reccount == 1) {
3033                 rc = lfsck_namespace_dsd_single(env, com, child, pfid, ldata,
3034                                                 lh, type, NULL, unknown);
3035
3036                 RETURN(rc);
3037         }
3038
3039         /* All linkEA entries are invalid and removed, then handle the @child
3040          * as an orphan.*/
3041         if (ldata->ld_leh->leh_reccount == 0) {
3042                 rc = lfsck_namespace_dsd_orphan(env, com, child, pfid, lh,