lustre/lfsck/lfsck_layout.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License version 2 for more details.  A copy is
  14  * included in the COPYING file that accompanied this code.
  15
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2014, 2017, Intel Corporation.
  24  */
  25 /*
  26  * lustre/lfsck/lfsck_layout.c
  27  *
  28  * Author: Fan, Yong <fan.yong@intel.com>
  29  */
  30
  31 #ifndef EXPORT_SYMTAB
  32 # define EXPORT_SYMTAB
  33 #endif
  34 #define DEBUG_SUBSYSTEM S_LFSCK
  35
  36 #include <linux/bitops.h>
  37 #include <linux/rbtree.h>
  38
  39 #include <lu_object.h>
  40 #include <dt_object.h>
  41 #include <lustre_fid.h>
  42 #include <lustre_lib.h>
  43 #include <lustre_net.h>
  44 #include <md_object.h>
  45 #include <obd_class.h>
  46
  47 #include "lfsck_internal.h"
  48
  49 #define LFSCK_LAYOUT_MAGIC_V1           0xB173AE14
  50 #define LFSCK_LAYOUT_MAGIC_V2           0xB1734D76
  51 #define LFSCK_LAYOUT_MAGIC_V3           0xB17371B9
  52 #define LFSCK_LAYOUT_MAGIC_V4           0xB1732FED
  53
  54 #define LFSCK_LAYOUT_MAGIC              LFSCK_LAYOUT_MAGIC_V4
  55
  56 struct lfsck_layout_seq {
  57         struct list_head         lls_list;
  58         __u64                    lls_seq;
  59         __u64                    lls_lastid;
  60         __u64                    lls_lastid_known;
  61         struct dt_object        *lls_lastid_obj;
  62         unsigned int             lls_dirty:1;
  63 };
  64
  65 struct lfsck_layout_slave_target {
  66         /* link into lfsck_layout_slave_data::llsd_master_list. */
  67         struct list_head        llst_list;
  68         /* The position for next record in the rbtree for iteration. */
  69         struct lu_fid           llst_fid;
  70         /* Dummy hash for iteration against the rbtree. */
  71         __u64                   llst_hash;
  72         __u64                   llst_gen;
  73         atomic_t                llst_ref;
  74         __u32                   llst_index;
  75         /* How many times we have failed to get the master status. */
  76         int                     llst_failures;
  77 };
  78
  79 struct lfsck_layout_slave_data {
  80         /* list for lfsck_layout_seq */
  81         struct list_head         llsd_seq_list;
  82
  83         /* list for the masters involve layout verification. */
  84         struct list_head         llsd_master_list;
  85         spinlock_t               llsd_lock;
  86         __u64                    llsd_touch_gen;
  87         struct dt_object        *llsd_rb_obj;
  88         struct rb_root           llsd_rb_root;
  89         struct rw_semaphore      llsd_rb_rwsem;
  90         unsigned int             llsd_rbtree_valid:1;
  91 };
  92
  93 struct lfsck_layout_slave_async_args {
  94         struct obd_export                *llsaa_exp;
  95         struct lfsck_component           *llsaa_com;
  96         struct lfsck_layout_slave_target *llsaa_llst;
  97 };
  98
  99 static inline bool lfsck_comp_extent_aligned(__u64 size)
 100 {
 101          return (size & (LOV_MIN_STRIPE_SIZE - 1)) == 0;
 102 }
 103
 104 static inline void
 105 lfsck_layout_llst_put(struct lfsck_layout_slave_target *llst)
 106 {
 107         if (atomic_dec_and_test(&llst->llst_ref)) {
 108                 LASSERT(list_empty(&llst->llst_list));
 109
 110                 OBD_FREE_PTR(llst);
 111         }
 112 }
 113
 114 static inline int
 115 lfsck_layout_llst_add(struct lfsck_layout_slave_data *llsd, __u32 index)
 116 {
 117         struct lfsck_layout_slave_target *llst;
 118         struct lfsck_layout_slave_target *tmp;
 119         int                               rc   = 0;
 120
 121         OBD_ALLOC_PTR(llst);
 122         if (llst == NULL)
 123                 return -ENOMEM;
 124
 125         INIT_LIST_HEAD(&llst->llst_list);
 126         llst->llst_gen = 0;
 127         llst->llst_index = index;
 128         atomic_set(&llst->llst_ref, 1);
 129
 130         spin_lock(&llsd->llsd_lock);
 131         list_for_each_entry(tmp, &llsd->llsd_master_list, llst_list) {
 132                 if (tmp->llst_index == index) {
 133                         rc = -EALREADY;
 134                         break;
 135                 }
 136         }
 137         if (rc == 0)
 138                 list_add_tail(&llst->llst_list, &llsd->llsd_master_list);
 139         spin_unlock(&llsd->llsd_lock);
 140
 141         if (rc != 0)
 142                 OBD_FREE_PTR(llst);
 143
 144         return rc;
 145 }
 146
 147 static inline void
 148 lfsck_layout_llst_del(struct lfsck_layout_slave_data *llsd,
 149                       struct lfsck_layout_slave_target *llst)
 150 {
 151         bool del = false;
 152
 153         spin_lock(&llsd->llsd_lock);
 154         if (!list_empty(&llst->llst_list)) {
 155                 list_del_init(&llst->llst_list);
 156                 del = true;
 157         }
 158         spin_unlock(&llsd->llsd_lock);
 159
 160         if (del)
 161                 lfsck_layout_llst_put(llst);
 162 }
 163
 164 static inline struct lfsck_layout_slave_target *
 165 lfsck_layout_llst_find_and_del(struct lfsck_layout_slave_data *llsd,
 166                                __u32 index, bool unlink)
 167 {
 168         struct lfsck_layout_slave_target *llst;
 169
 170         spin_lock(&llsd->llsd_lock);
 171         list_for_each_entry(llst, &llsd->llsd_master_list, llst_list) {
 172                 if (llst->llst_index == index) {
 173                         if (unlink)
 174                                 list_del_init(&llst->llst_list);
 175                         else
 176                                 atomic_inc(&llst->llst_ref);
 177                         spin_unlock(&llsd->llsd_lock);
 178
 179                         return llst;
 180                 }
 181         }
 182         spin_unlock(&llsd->llsd_lock);
 183
 184         return NULL;
 185 }
 186
 187 static struct lfsck_layout_req *
 188 lfsck_layout_assistant_req_init(struct lfsck_assistant_object *lso,
 189                                 struct dt_object *child, __u32 comp_id,
 190                                 __u32 ost_idx, __u32 lov_idx)
 191 {
 192         struct lfsck_layout_req *llr;
 193
 194         OBD_ALLOC_PTR(llr);
 195         if (llr == NULL)
 196                 return ERR_PTR(-ENOMEM);
 197
 198         INIT_LIST_HEAD(&llr->llr_lar.lar_list);
 199         llr->llr_lar.lar_parent = lfsck_assistant_object_get(lso);
 200         llr->llr_child = child;
 201         llr->llr_comp_id = comp_id;
 202         llr->llr_ost_idx = ost_idx;
 203         llr->llr_lov_idx = lov_idx;
 204
 205         return llr;
 206 }
 207
 208 static void lfsck_layout_assistant_req_fini(const struct lu_env *env,
 209                                             struct lfsck_assistant_req *lar)
 210 {
 211         struct lfsck_layout_req *llr =
 212                         container_of0(lar, struct lfsck_layout_req, llr_lar);
 213
 214         lfsck_object_put(env, llr->llr_child);
 215         lfsck_assistant_object_put(env, lar->lar_parent);
 216         OBD_FREE_PTR(llr);
 217 }
 218
 219 static int
 220 lfsck_layout_assistant_sync_failures_interpret(const struct lu_env *env,
 221                                                struct ptlrpc_request *req,
 222                                                void *args, int rc)
 223 {
 224         if (rc == 0) {
 225                 struct lfsck_async_interpret_args *laia = args;
 226                 struct lfsck_tgt_desc             *ltd  = laia->laia_ltd;
 227
 228                 ltd->ltd_synced_failures = 1;
 229                 atomic_dec(laia->laia_count);
 230         }
 231
 232         return 0;
 233 }
 234
 235 /**
 236  * Notify remote LFSCK instances about former failures.
 237  *
 238  * The local LFSCK instance has recorded which OSTs have ever failed to respond
 239  * some LFSCK verification requests (maybe because of network issues or the OST
 240  * itself trouble). During the respond gap, the OST may missed some OST-objects
 241  * verification, then the OST cannot know whether related OST-objects have been
 242  * referenced by related MDT-objects or not, then in the second-stage scanning,
 243  * these OST-objects will be regarded as orphan, if the OST-object contains bad
 244  * parent FID for back reference, then it will misguide the LFSCK to make wrong
 245  * fixing for the fake orphan.
 246  *
 247  * To avoid above trouble, when layout LFSCK finishes the first-stage scanning,
 248  * it will scan the bitmap for the ever failed OSTs, and notify them that they
 249  * have ever missed some OST-object verification and should skip the handling
 250  * for orphan OST-objects on all MDTs that are in the layout LFSCK.
 251  *
 252  * \param[in] env       pointer to the thread context
 253  * \param[in] com       pointer to the lfsck component
 254  * \param[in] lr        pointer to the lfsck request
 255  */
 256 static void lfsck_layout_assistant_sync_failures(const struct lu_env *env,
 257                                                  struct lfsck_component *com,
 258                                                  struct lfsck_request *lr)
 259 {
 260         struct lfsck_async_interpret_args *laia  =
 261                                 &lfsck_env_info(env)->lti_laia2;
 262         struct lfsck_assistant_data       *lad   = com->lc_data;
 263         struct lfsck_layout               *lo    = com->lc_file_ram;
 264         struct lfsck_instance             *lfsck = com->lc_lfsck;
 265         struct lfsck_tgt_descs            *ltds  = &lfsck->li_ost_descs;
 266         struct lfsck_tgt_desc             *ltd;
 267         struct ptlrpc_request_set         *set;
 268         atomic_t                           count;
 269         __u32                              idx;
 270         int                                rc    = 0;
 271         ENTRY;
 272
 273         if (!lad->lad_incomplete)
 274                 RETURN_EXIT;
 275
 276         /* If the MDT has ever failed to verfiy some OST-objects,
 277          * then sync failures with them firstly. */
 278         lr->lr_flags2 = lo->ll_flags | LF_INCOMPLETE;
 279
 280         atomic_set(&count, 0);
 281         memset(laia, 0, sizeof(*laia));
 282         laia->laia_count = &count;
 283         set = ptlrpc_prep_set();
 284         if (set == NULL)
 285                 GOTO(out, rc = -ENOMEM);
 286
 287         down_read(&ltds->ltd_rw_sem);
 288         cfs_foreach_bit(lad->lad_bitmap, idx) {
 289                 ltd = lfsck_ltd2tgt(ltds, idx);
 290                 if (unlikely(!ltd))
 291                         continue;
 292
 293                 laia->laia_ltd = ltd;
 294                 rc = lfsck_async_request(env, ltd->ltd_exp, lr, set,
 295                                 lfsck_layout_assistant_sync_failures_interpret,
 296                                 laia, LFSCK_NOTIFY);
 297                 if (rc != 0) {
 298                         CDEBUG(D_LFSCK, "%s: LFSCK assistant fail to "
 299                                "notify target %x for %s phase1 done: "
 300                                "rc = %d\n", lfsck_lfsck2name(com->lc_lfsck),
 301                                ltd->ltd_index, lad->lad_name, rc);
 302
 303                         break;
 304                 }
 305
 306                 atomic_inc(&count);
 307         }
 308         up_read(&ltds->ltd_rw_sem);
 309
 310         if (rc == 0 && atomic_read(&count) > 0)
 311                 rc = ptlrpc_set_wait(env, set);
 312
 313         ptlrpc_set_destroy(set);
 314
 315         if (rc == 0 && atomic_read(&count) > 0)
 316                 rc = -EINVAL;
 317
 318         GOTO(out, rc);
 319
 320 out:
 321         if (rc != 0)
 322                 /* If failed to sync failures with the OSTs, then have to
 323                  * mark the whole LFSCK as LF_INCOMPLETE to skip the whole
 324                  * subsequent orphan OST-object handling. */
 325                 lo->ll_flags |= LF_INCOMPLETE;
 326
 327         lr->lr_flags2 = lo->ll_flags;
 328 }
 329
 330 static int lfsck_layout_verify_header_v1v3(struct dt_object *obj,
 331                                            struct lov_mds_md_v1 *lmm,
 332                                            __u64 start, __u32 comp_id)
 333 {
 334         __u32 magic;
 335         __u32 pattern;
 336
 337         magic = le32_to_cpu(lmm->lmm_magic);
 338         /* If magic crashed, keep it there. Sometime later, during OST-object
 339          * orphan handling, if some OST-object(s) back-point to it, it can be
 340          * verified and repaired. */
 341         if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3) {
 342                 int rc;
 343
 344                 if ((magic & LOV_MAGIC_MASK) == LOV_MAGIC_MAGIC)
 345                         rc = -EOPNOTSUPP;
 346                 else
 347                         rc = -EINVAL;
 348
 349                 CDEBUG(D_LFSCK, "%s LOV EA magic %u for the file "DFID"\n",
 350                        rc == -EINVAL ? "Unknown" : "Unsupported",
 351                        magic, PFID(lfsck_dto2fid(obj)));
 352
 353                 return rc;
 354         }
 355
 356         pattern = le32_to_cpu(lmm->lmm_pattern);
 357
 358 #if 0
 359         /* XXX: DoM file verification will be supportted via LU-11081. */
 360         if (lov_pattern(pattern) == LOV_PATTERN_MDT) {
 361                 if (start != 0) {
 362                         CDEBUG(D_LFSCK, "The DoM entry for "DFID" is not "
 363                                "the first component in the mirror %x/%llu\n",
 364                                PFID(lfsck_dto2fid(obj)), comp_id, start);
 365
 366                         return -EINVAL;
 367                 }
 368         }
 369 #endif
 370
 371         if (lov_pattern(pattern) != LOV_PATTERN_RAID0) {
 372                 CDEBUG(D_LFSCK, "Unsupported LOV EA pattern %u for the file "
 373                        DFID" in the component %x\n",
 374                        pattern, PFID(lfsck_dto2fid(obj)), comp_id);
 375
 376                 return -EOPNOTSUPP;
 377         }
 378
 379         return 0;
 380 }
 381
 382 static int lfsck_layout_verify_header(struct dt_object *obj,
 383                                       struct lov_mds_md_v1 *lmm)
 384 {
 385         int rc = 0;
 386
 387         if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
 388                 struct lov_comp_md_v1 *lcm = (struct lov_comp_md_v1 *)lmm;
 389                 int i;
 390                 __u16 count = le16_to_cpu(lcm->lcm_entry_count);
 391
 392                 if (unlikely(count == 0)) {
 393                         CDEBUG(D_LFSCK, "the PFL file "DFID" contains invalid "
 394                                "components count 0\n",
 395                                PFID(lfsck_dto2fid(obj)));
 396
 397                         return -EINVAL;
 398                 }
 399
 400                 for (i = 0; i < count && !rc; i++) {
 401                         struct lov_comp_md_entry_v1 *lcme =
 402                                                 &lcm->lcm_entries[i];
 403                         __u64 start = le64_to_cpu(lcme->lcme_extent.e_start);
 404                         __u64 end = le64_to_cpu(lcme->lcme_extent.e_end);
 405                         __u32 comp_id = le32_to_cpu(lcme->lcme_id);
 406
 407                         if (unlikely(comp_id == LCME_ID_INVAL ||
 408                                      comp_id > LCME_ID_MAX)) {
 409                                 CDEBUG(D_LFSCK, "found invalid FPL ID %u "
 410                                        "for the file "DFID" at idx %d\n",
 411                                        comp_id, PFID(lfsck_dto2fid(obj)), i);
 412
 413                                 return -EINVAL;
 414                         }
 415
 416                         if (unlikely(start >= end ||
 417                                      !lfsck_comp_extent_aligned(start) ||
 418                                      (!lfsck_comp_extent_aligned(end) &&
 419                                       end != LUSTRE_EOF))) {
 420                                 CDEBUG(D_LFSCK, "found invalid FPL extent "
 421                                        "range [%llu - %llu) for the file "
 422                                        DFID" at idx %d\n",
 423                                        start, end, PFID(lfsck_dto2fid(obj)), i);
 424
 425                                 return -EINVAL;
 426                         }
 427
 428                         rc = lfsck_layout_verify_header_v1v3(obj,
 429                                         (struct lov_mds_md_v1 *)((char *)lmm +
 430                                         le32_to_cpu(lcme->lcme_offset)), start,
 431                                         comp_id);
 432                 }
 433         } else {
 434                 rc = lfsck_layout_verify_header_v1v3(obj, lmm, 1, 0);
 435         }
 436
 437         return rc;
 438 }
 439
 440 static int lfsck_layout_get_lovea(const struct lu_env *env,
 441                                   struct dt_object *obj, struct lu_buf *buf)
 442 {
 443         int rc;
 444         int rc1;
 445
 446 again:
 447         rc = dt_xattr_get(env, obj, buf, XATTR_NAME_LOV);
 448         if (rc == -ERANGE) {
 449                 rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_LOV);
 450                 if (rc <= 0)
 451                         return !rc ? -ENODATA : rc;
 452
 453                 lu_buf_realloc(buf, rc);
 454                 if (buf->lb_buf == NULL)
 455                         return -ENOMEM;
 456
 457                 goto again;
 458         }
 459
 460         if (rc <= 0)
 461                 return !rc ? -ENODATA : rc;
 462
 463         if (unlikely(buf->lb_buf == NULL)) {
 464                 lu_buf_alloc(buf, rc);
 465                 if (buf->lb_buf == NULL)
 466                         return -ENOMEM;
 467
 468                 goto again;
 469         }
 470
 471         rc1 = lfsck_layout_verify_header(obj, buf->lb_buf);
 472
 473         return rc1 ? rc1 : rc;
 474 }
 475
 476 #define LFSCK_RBTREE_BITMAP_SIZE        PAGE_SIZE
 477 #define LFSCK_RBTREE_BITMAP_WIDTH       (LFSCK_RBTREE_BITMAP_SIZE << 3)
 478 #define LFSCK_RBTREE_BITMAP_MASK        (LFSCK_RBTREE_BITMAP_WIDTH - 1)
 479
 480 struct lfsck_rbtree_node {
 481         struct rb_node   lrn_node;
 482         __u64            lrn_seq;
 483         __u32            lrn_first_oid;
 484         atomic_t         lrn_known_count;
 485         atomic_t         lrn_accessed_count;
 486         void            *lrn_known_bitmap;
 487         void            *lrn_accessed_bitmap;
 488 };
 489
 490 static inline int lfsck_rbtree_cmp(struct lfsck_rbtree_node *lrn,
 491                                    __u64 seq, __u32 oid)
 492 {
 493         if (seq < lrn->lrn_seq)
 494                 return -1;
 495
 496         if (seq > lrn->lrn_seq)
 497                 return 1;
 498
 499         if (oid < lrn->lrn_first_oid)
 500                 return -1;
 501
 502         if (oid - lrn->lrn_first_oid >= LFSCK_RBTREE_BITMAP_WIDTH)
 503                 return 1;
 504
 505         return 0;
 506 }
 507
 508 /* The caller should hold llsd->llsd_rb_lock. */
 509 static struct lfsck_rbtree_node *
 510 lfsck_rbtree_search(struct lfsck_layout_slave_data *llsd,
 511                     const struct lu_fid *fid, bool *exact)
 512 {
 513         struct rb_node           *node  = llsd->llsd_rb_root.rb_node;
 514         struct rb_node           *prev  = NULL;
 515         struct lfsck_rbtree_node *lrn   = NULL;
 516         int                       rc    = 0;
 517
 518         if (exact != NULL)
 519                 *exact = true;
 520
 521         while (node != NULL) {
 522                 prev = node;
 523                 lrn = rb_entry(node, struct lfsck_rbtree_node, lrn_node);
 524                 rc = lfsck_rbtree_cmp(lrn, fid_seq(fid), fid_oid(fid));
 525                 if (rc < 0)
 526                         node = node->rb_left;
 527                 else if (rc > 0)
 528                         node = node->rb_right;
 529                 else
 530                         return lrn;
 531         }
 532
 533         if (exact == NULL)
 534                 return NULL;
 535
 536         /* If there is no exactly matched one, then to the next valid one. */
 537         *exact = false;
 538
 539         /* The rbtree is empty. */
 540         if (rc == 0)
 541                 return NULL;
 542
 543         if (rc < 0)
 544                 return lrn;
 545
 546         node = rb_next(prev);
 547
 548         /* The end of the rbtree. */
 549         if (node == NULL)
 550                 return NULL;
 551
 552         lrn = rb_entry(node, struct lfsck_rbtree_node, lrn_node);
 553
 554         return lrn;
 555 }
 556
 557 static struct lfsck_rbtree_node *lfsck_rbtree_new(const struct lu_env *env,
 558                                                   const struct lu_fid *fid)
 559 {
 560         struct lfsck_rbtree_node *lrn;
 561
 562         OBD_ALLOC_PTR(lrn);
 563         if (lrn == NULL)
 564                 return ERR_PTR(-ENOMEM);
 565
 566         OBD_ALLOC(lrn->lrn_known_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 567         if (lrn->lrn_known_bitmap == NULL) {
 568                 OBD_FREE_PTR(lrn);
 569
 570                 return ERR_PTR(-ENOMEM);
 571         }
 572
 573         OBD_ALLOC(lrn->lrn_accessed_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 574         if (lrn->lrn_accessed_bitmap == NULL) {
 575                 OBD_FREE(lrn->lrn_known_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 576                 OBD_FREE_PTR(lrn);
 577
 578                 return ERR_PTR(-ENOMEM);
 579         }
 580
 581         RB_CLEAR_NODE(&lrn->lrn_node);
 582         lrn->lrn_seq = fid_seq(fid);
 583         lrn->lrn_first_oid = fid_oid(fid) & ~LFSCK_RBTREE_BITMAP_MASK;
 584         atomic_set(&lrn->lrn_known_count, 0);
 585         atomic_set(&lrn->lrn_accessed_count, 0);
 586
 587         return lrn;
 588 }
 589
 590 static void lfsck_rbtree_free(struct lfsck_rbtree_node *lrn)
 591 {
 592         OBD_FREE(lrn->lrn_accessed_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 593         OBD_FREE(lrn->lrn_known_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 594         OBD_FREE_PTR(lrn);
 595 }
 596
 597 /* The caller should hold lock. */
 598 static struct lfsck_rbtree_node *
 599 lfsck_rbtree_insert(struct lfsck_layout_slave_data *llsd,
 600                     struct lfsck_rbtree_node *lrn)
 601 {
 602         struct rb_node           **pos    = &llsd->llsd_rb_root.rb_node;
 603         struct rb_node            *parent = NULL;
 604         struct lfsck_rbtree_node  *tmp;
 605         int                        rc;
 606
 607         while (*pos != NULL) {
 608                 parent = *pos;
 609                 tmp = rb_entry(parent, struct lfsck_rbtree_node, lrn_node);
 610                 rc = lfsck_rbtree_cmp(tmp, lrn->lrn_seq, lrn->lrn_first_oid);
 611                 if (rc < 0)
 612                         pos = &(*pos)->rb_left;
 613                 else if (rc > 0)
 614                         pos = &(*pos)->rb_right;
 615                 else
 616                         return tmp;
 617         }
 618
 619         rb_link_node(&lrn->lrn_node, parent, pos);
 620         rb_insert_color(&lrn->lrn_node, &llsd->llsd_rb_root);
 621
 622         return lrn;
 623 }
 624
 625 extern const struct dt_index_operations lfsck_orphan_index_ops;
 626
 627 static int lfsck_rbtree_setup(const struct lu_env *env,
 628                               struct lfsck_component *com)
 629 {
 630         struct lu_fid                   *fid    = &lfsck_env_info(env)->lti_fid;
 631         struct lfsck_instance           *lfsck  = com->lc_lfsck;
 632         struct dt_device                *dev    = lfsck->li_bottom;
 633         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
 634         struct dt_object                *obj;
 635
 636         fid->f_seq = FID_SEQ_LAYOUT_RBTREE;
 637         fid->f_oid = lfsck_dev_idx(lfsck);
 638         fid->f_ver = 0;
 639         obj = dt_locate(env, dev, fid);
 640         if (IS_ERR(obj))
 641                 RETURN(PTR_ERR(obj));
 642
 643         /* Generate an in-RAM object to stand for the layout rbtree.
 644          * Scanning the layout rbtree will be via the iteration over
 645          * the object. In the future, the rbtree may be written onto
 646          * disk with the object.
 647          *
 648          * Mark the object to be as exist. */
 649         obj->do_lu.lo_header->loh_attr |= LOHA_EXISTS;
 650         obj->do_index_ops = &lfsck_orphan_index_ops;
 651         llsd->llsd_rb_obj = obj;
 652         llsd->llsd_rbtree_valid = 1;
 653         dev->dd_record_fid_accessed = 1;
 654
 655         CDEBUG(D_LFSCK, "%s: layout LFSCK init OST-objects accessing bitmap\n",
 656                lfsck_lfsck2name(lfsck));
 657
 658         return 0;
 659 }
 660
 661 static void lfsck_rbtree_cleanup(const struct lu_env *env,
 662                                  struct lfsck_component *com)
 663 {
 664         struct lfsck_instance           *lfsck = com->lc_lfsck;
 665         struct lfsck_layout_slave_data  *llsd  = com->lc_data;
 666         struct rb_node                  *node  = rb_first(&llsd->llsd_rb_root);
 667         struct rb_node                  *next;
 668         struct lfsck_rbtree_node        *lrn;
 669
 670         lfsck->li_bottom->dd_record_fid_accessed = 0;
 671         /* Invalid the rbtree, then no others will use it. */
 672         down_write(&llsd->llsd_rb_rwsem);
 673         llsd->llsd_rbtree_valid = 0;
 674         up_write(&llsd->llsd_rb_rwsem);
 675
 676         while (node != NULL) {
 677                 next = rb_next(node);
 678                 lrn = rb_entry(node, struct lfsck_rbtree_node, lrn_node);
 679                 rb_erase(node, &llsd->llsd_rb_root);
 680                 lfsck_rbtree_free(lrn);
 681                 node = next;
 682         }
 683
 684         if (llsd->llsd_rb_obj != NULL) {
 685                 lfsck_object_put(env, llsd->llsd_rb_obj);
 686                 llsd->llsd_rb_obj = NULL;
 687         }
 688
 689         CDEBUG(D_LFSCK, "%s: layout LFSCK fini OST-objects accessing bitmap\n",
 690                lfsck_lfsck2name(lfsck));
 691 }
 692
 693 static void lfsck_rbtree_update_bitmap(const struct lu_env *env,
 694                                        struct lfsck_component *com,
 695                                        const struct lu_fid *fid,
 696                                        bool accessed)
 697 {
 698         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
 699         struct lfsck_rbtree_node        *lrn;
 700         bool                             insert = false;
 701         int                              idx;
 702         int                              rc     = 0;
 703         ENTRY;
 704
 705         if (unlikely(!fid_is_sane(fid) || fid_is_last_id(fid)))
 706                 RETURN_EXIT;
 707
 708         if (!fid_is_idif(fid) && !fid_is_norm(fid))
 709                 RETURN_EXIT;
 710
 711         down_read(&llsd->llsd_rb_rwsem);
 712         if (!llsd->llsd_rbtree_valid)
 713                 GOTO(unlock, rc = 0);
 714
 715         lrn = lfsck_rbtree_search(llsd, fid, NULL);
 716         if (lrn == NULL) {
 717                 struct lfsck_rbtree_node *tmp;
 718
 719                 LASSERT(!insert);
 720
 721                 up_read(&llsd->llsd_rb_rwsem);
 722                 tmp = lfsck_rbtree_new(env, fid);
 723                 if (IS_ERR(tmp))
 724                         GOTO(out, rc = PTR_ERR(tmp));
 725
 726                 insert = true;
 727                 down_write(&llsd->llsd_rb_rwsem);
 728                 if (!llsd->llsd_rbtree_valid) {
 729                         lfsck_rbtree_free(tmp);
 730                         GOTO(unlock, rc = 0);
 731                 }
 732
 733                 lrn = lfsck_rbtree_insert(llsd, tmp);
 734                 if (lrn != tmp)
 735                         lfsck_rbtree_free(tmp);
 736         }
 737
 738         idx = fid_oid(fid) & LFSCK_RBTREE_BITMAP_MASK;
 739         /* Any accessed object must be a known object. */
 740         if (!test_and_set_bit(idx, lrn->lrn_known_bitmap))
 741                 atomic_inc(&lrn->lrn_known_count);
 742         if (accessed && !test_and_set_bit(idx, lrn->lrn_accessed_bitmap))
 743                 atomic_inc(&lrn->lrn_accessed_count);
 744
 745         GOTO(unlock, rc = 0);
 746
 747 unlock:
 748         if (insert)
 749                 up_write(&llsd->llsd_rb_rwsem);
 750         else
 751                 up_read(&llsd->llsd_rb_rwsem);
 752 out:
 753         if (rc != 0 && accessed) {
 754                 struct lfsck_layout *lo = com->lc_file_ram;
 755
 756                 CDEBUG(D_LFSCK, "%s: fail to update OST-objects accessing "
 757                        "bitmap, and will cause incorrect LFSCK OST-object "
 758                        "handling, so disable it to cancel orphan handling "
 759                        "for related device. rc = %d\n",
 760                        lfsck_lfsck2name(com->lc_lfsck), rc);
 761
 762                 lo->ll_flags |= LF_INCOMPLETE;
 763                 lfsck_rbtree_cleanup(env, com);
 764         }
 765 }
 766
 767 static inline void lldk_le_to_cpu(struct lfsck_layout_dangling_key *des,
 768                                   const struct lfsck_layout_dangling_key *src)
 769 {
 770         fid_le_to_cpu(&des->lldk_fid, &src->lldk_fid);
 771         des->lldk_comp_id = le32_to_cpu(src->lldk_comp_id);
 772         des->lldk_ea_off = le32_to_cpu(src->lldk_ea_off);
 773 }
 774
 775 static inline void lldk_cpu_to_le(struct lfsck_layout_dangling_key *des,
 776                                   const struct lfsck_layout_dangling_key *src)
 777 {
 778         fid_cpu_to_le(&des->lldk_fid, &src->lldk_fid);
 779         des->lldk_comp_id = cpu_to_le32(src->lldk_comp_id);
 780         des->lldk_ea_off = cpu_to_le32(src->lldk_ea_off);
 781 }
 782
 783 static inline void lldk_be_to_cpu(struct lfsck_layout_dangling_key *des,
 784                                   const struct lfsck_layout_dangling_key *src)
 785 {
 786         fid_be_to_cpu(&des->lldk_fid, &src->lldk_fid);
 787         des->lldk_comp_id = be32_to_cpu(src->lldk_comp_id);
 788         des->lldk_ea_off = be32_to_cpu(src->lldk_ea_off);
 789 }
 790
 791 static inline void lldk_cpu_to_be(struct lfsck_layout_dangling_key *des,
 792                                   const struct lfsck_layout_dangling_key *src)
 793 {
 794         fid_cpu_to_be(&des->lldk_fid, &src->lldk_fid);
 795         des->lldk_comp_id = cpu_to_be32(src->lldk_comp_id);
 796         des->lldk_ea_off = cpu_to_be32(src->lldk_ea_off);
 797 }
 798
 799 static void lfsck_layout_le_to_cpu(struct lfsck_layout *des,
 800                                    const struct lfsck_layout *src)
 801 {
 802         int i;
 803
 804         des->ll_magic = le32_to_cpu(src->ll_magic);
 805         des->ll_status = le32_to_cpu(src->ll_status);
 806         des->ll_flags = le32_to_cpu(src->ll_flags);
 807         des->ll_success_count = le32_to_cpu(src->ll_success_count);
 808         des->ll_run_time_phase1 = le64_to_cpu(src->ll_run_time_phase1);
 809         des->ll_run_time_phase2 = le64_to_cpu(src->ll_run_time_phase2);
 810         des->ll_time_last_complete = le64_to_cpu(src->ll_time_last_complete);
 811         des->ll_time_latest_start = le64_to_cpu(src->ll_time_latest_start);
 812         des->ll_time_last_checkpoint =
 813                                 le64_to_cpu(src->ll_time_last_checkpoint);
 814         des->ll_pos_latest_start = le64_to_cpu(src->ll_pos_latest_start);
 815         des->ll_pos_last_checkpoint = le64_to_cpu(src->ll_pos_last_checkpoint);
 816         des->ll_pos_first_inconsistent =
 817                         le64_to_cpu(src->ll_pos_first_inconsistent);
 818         des->ll_objs_checked_phase1 = le64_to_cpu(src->ll_objs_checked_phase1);
 819         des->ll_objs_failed_phase1 = le64_to_cpu(src->ll_objs_failed_phase1);
 820         des->ll_objs_checked_phase2 = le64_to_cpu(src->ll_objs_checked_phase2);
 821         des->ll_objs_failed_phase2 = le64_to_cpu(src->ll_objs_failed_phase2);
 822         for (i = 0; i < LLIT_MAX; i++)
 823                 des->ll_objs_repaired[i] =
 824                                 le64_to_cpu(src->ll_objs_repaired[i]);
 825         des->ll_objs_skipped = le64_to_cpu(src->ll_objs_skipped);
 826         des->ll_bitmap_size = le32_to_cpu(src->ll_bitmap_size);
 827         lldk_le_to_cpu(&des->ll_lldk_latest_scanned_phase2,
 828                        &src->ll_lldk_latest_scanned_phase2);
 829 }
 830
 831 static void lfsck_layout_cpu_to_le(struct lfsck_layout *des,
 832                                    const struct lfsck_layout *src)
 833 {
 834         int i;
 835
 836         des->ll_magic = cpu_to_le32(src->ll_magic);
 837         des->ll_status = cpu_to_le32(src->ll_status);
 838         des->ll_flags = cpu_to_le32(src->ll_flags);
 839         des->ll_success_count = cpu_to_le32(src->ll_success_count);
 840         des->ll_run_time_phase1 = cpu_to_le64(src->ll_run_time_phase1);
 841         des->ll_run_time_phase2 = cpu_to_le64(src->ll_run_time_phase2);
 842         des->ll_time_last_complete = cpu_to_le64(src->ll_time_last_complete);
 843         des->ll_time_latest_start = cpu_to_le64(src->ll_time_latest_start);
 844         des->ll_time_last_checkpoint =
 845                                 cpu_to_le64(src->ll_time_last_checkpoint);
 846         des->ll_pos_latest_start = cpu_to_le64(src->ll_pos_latest_start);
 847         des->ll_pos_last_checkpoint = cpu_to_le64(src->ll_pos_last_checkpoint);
 848         des->ll_pos_first_inconsistent =
 849                         cpu_to_le64(src->ll_pos_first_inconsistent);
 850         des->ll_objs_checked_phase1 = cpu_to_le64(src->ll_objs_checked_phase1);
 851         des->ll_objs_failed_phase1 = cpu_to_le64(src->ll_objs_failed_phase1);
 852         des->ll_objs_checked_phase2 = cpu_to_le64(src->ll_objs_checked_phase2);
 853         des->ll_objs_failed_phase2 = cpu_to_le64(src->ll_objs_failed_phase2);
 854         for (i = 0; i < LLIT_MAX; i++)
 855                 des->ll_objs_repaired[i] =
 856                                 cpu_to_le64(src->ll_objs_repaired[i]);
 857         des->ll_objs_skipped = cpu_to_le64(src->ll_objs_skipped);
 858         des->ll_bitmap_size = cpu_to_le32(src->ll_bitmap_size);
 859         lldk_cpu_to_le(&des->ll_lldk_latest_scanned_phase2,
 860                        &src->ll_lldk_latest_scanned_phase2);
 861 }
 862
 863 /**
 864  * Load the OST bitmap from the lfsck_layout trace file.
 865  *
 866  * \param[in] env       pointer to the thread context
 867  * \param[in] com       pointer to the lfsck component
 868  *
 869  * \retval              0 for success
 870  * \retval              negative error number on failure or data corruption
 871  */
 872 static int lfsck_layout_load_bitmap(const struct lu_env *env,
 873                                     struct lfsck_component *com)
 874 {
 875         struct dt_object                *obj    = com->lc_obj;
 876         struct lfsck_assistant_data     *lad    = com->lc_data;
 877         struct lfsck_layout             *lo     = com->lc_file_ram;
 878         struct cfs_bitmap                       *bitmap = lad->lad_bitmap;
 879         loff_t                           pos    = com->lc_file_size;
 880         ssize_t                          size;
 881         __u32                            nbits;
 882         int                              rc;
 883         ENTRY;
 884
 885         if (com->lc_lfsck->li_ost_descs.ltd_tgts_bitmap->size >
 886             lo->ll_bitmap_size)
 887                 nbits = com->lc_lfsck->li_ost_descs.ltd_tgts_bitmap->size;
 888         else
 889                 nbits = lo->ll_bitmap_size;
 890
 891         if (unlikely(nbits < BITS_PER_LONG))
 892                 nbits = BITS_PER_LONG;
 893
 894         if (nbits > bitmap->size) {
 895                 __u32 new_bits = bitmap->size;
 896                 struct cfs_bitmap *new_bitmap;
 897
 898                 while (new_bits < nbits)
 899                         new_bits <<= 1;
 900
 901                 new_bitmap = CFS_ALLOCATE_BITMAP(new_bits);
 902                 if (new_bitmap == NULL)
 903                         RETURN(-ENOMEM);
 904
 905                 lad->lad_bitmap = new_bitmap;
 906                 CFS_FREE_BITMAP(bitmap);
 907                 bitmap = new_bitmap;
 908         }
 909
 910         if (lo->ll_bitmap_size == 0) {
 911                 lad->lad_incomplete = 0;
 912                 CFS_RESET_BITMAP(bitmap);
 913
 914                 RETURN(0);
 915         }
 916
 917         size = (lo->ll_bitmap_size + 7) >> 3;
 918         rc = dt_read(env, obj, lfsck_buf_get(env, bitmap->data, size), &pos);
 919         if (rc != size)
 920                 RETURN(rc >= 0 ? -EINVAL : rc);
 921
 922         if (cfs_bitmap_check_empty(bitmap))
 923                 lad->lad_incomplete = 0;
 924         else
 925                 lad->lad_incomplete = 1;
 926
 927         RETURN(0);
 928 }
 929
 930 /**
 931  * Load the layout LFSCK trace file from disk.
 932  *
 933  * The layout LFSCK trace file records the layout LFSCK status information
 934  * and other statistics, such as how many objects have been scanned, and how
 935  * many objects have been repaired, and etc. It also contains the bitmap for
 936  * failed OSTs during the layout LFSCK. All these information will be loaded
 937  * from disk to RAM when the layout LFSCK component setup.
 938  *
 939  * \param[in] env       pointer to the thread context
 940  * \param[in] com       pointer to the lfsck component
 941  *
 942  * \retval              positive number for file data corruption, the caller
 943  *                      should reset the layout LFSCK trace file
 944  * \retval              0 for success
 945  * \retval              negative error number on failure
 946  */
 947 static int lfsck_layout_load(const struct lu_env *env,
 948                              struct lfsck_component *com)
 949 {
 950         struct lfsck_layout             *lo     = com->lc_file_ram;
 951         ssize_t                          size   = com->lc_file_size;
 952         loff_t                           pos    = 0;
 953         int                              rc;
 954
 955         rc = dt_read(env, com->lc_obj,
 956                      lfsck_buf_get(env, com->lc_file_disk, size), &pos);
 957         if (rc == 0) {
 958                 return -ENOENT;
 959         } else if (rc < 0) {
 960                 CDEBUG(D_LFSCK, "%s: failed to load lfsck_layout: rc = %d\n",
 961                        lfsck_lfsck2name(com->lc_lfsck), rc);
 962                 return rc;
 963         } else if (rc != size) {
 964                 CDEBUG(D_LFSCK, "%s: lfsck_layout size %u != %u; reset it\n",
 965                        lfsck_lfsck2name(com->lc_lfsck), rc, (unsigned int)size);
 966                 return 1;
 967         }
 968
 969         lfsck_layout_le_to_cpu(lo, com->lc_file_disk);
 970         if (lo->ll_magic != LFSCK_LAYOUT_MAGIC) {
 971                 CDEBUG(D_LFSCK, "%s: invalid lfsck_layout magic %#x != %#x, "
 972                        "to be reset\n", lfsck_lfsck2name(com->lc_lfsck),
 973                        lo->ll_magic, LFSCK_LAYOUT_MAGIC);
 974                 return 1;
 975         }
 976
 977         return 0;
 978 }
 979
 980 /**
 981  * Store the layout LFSCK trace file on disk.
 982  *
 983  * The layout LFSCK trace file records the layout LFSCK status information
 984  * and other statistics, such as how many objects have been scanned, and how
 985  * many objects have been repaired, and etc. It also contains the bitmap for
 986  * failed OSTs during the layout LFSCK. All these information will be synced
 987  * from RAM to disk periodically.
 988  *
 989  * \param[in] env       pointer to the thread context
 990  * \param[in] com       pointer to the lfsck component
 991  *
 992  * \retval              0 for success
 993  * \retval              negative error number on failure
 994  */
 995 static int lfsck_layout_store(const struct lu_env *env,
 996                               struct lfsck_component *com)
 997 {
 998         struct dt_object        *obj    = com->lc_obj;
 999         struct lfsck_instance   *lfsck  = com->lc_lfsck;
1000         struct lfsck_layout     *lo_ram = com->lc_file_ram;
1001         struct lfsck_layout     *lo     = com->lc_file_disk;
1002         struct thandle          *th;
1003         struct dt_device        *dev    = lfsck_obj2dev(obj);
1004         struct cfs_bitmap       *bitmap = NULL;
1005         loff_t                   pos;
1006         ssize_t                  size   = com->lc_file_size;
1007         __u32                    nbits  = 0;
1008         int                      rc;
1009         ENTRY;
1010
1011         if (lfsck->li_master) {
1012                 struct lfsck_assistant_data *lad = com->lc_data;
1013
1014                 bitmap = lad->lad_bitmap;
1015                 nbits = bitmap->size;
1016
1017                 LASSERT(nbits > 0);
1018                 LASSERTF((nbits & 7) == 0, "Invalid nbits %u\n", nbits);
1019         }
1020
1021         lo_ram->ll_bitmap_size = nbits;
1022         lfsck_layout_cpu_to_le(lo, lo_ram);
1023         th = dt_trans_create(env, dev);
1024         if (IS_ERR(th))
1025                 GOTO(log, rc = PTR_ERR(th));
1026
1027         rc = dt_declare_record_write(env, obj, lfsck_buf_get(env, lo, size),
1028                                      (loff_t)0, th);
1029         if (rc != 0)
1030                 GOTO(out, rc);
1031
1032         if (bitmap != NULL) {
1033                 rc = dt_declare_record_write(env, obj,
1034                                 lfsck_buf_get(env, bitmap->data, nbits >> 3),
1035                                 (loff_t)size, th);
1036                 if (rc != 0)
1037                         GOTO(out, rc);
1038         }
1039
1040         rc = dt_trans_start_local(env, dev, th);
1041         if (rc != 0)
1042                 GOTO(out, rc);
1043
1044         pos = 0;
1045         rc = dt_record_write(env, obj, lfsck_buf_get(env, lo, size), &pos, th);
1046         if (rc != 0)
1047                 GOTO(out, rc);
1048
1049         if (bitmap != NULL) {
1050                 pos = size;
1051                 rc = dt_record_write(env, obj,
1052                                 lfsck_buf_get(env, bitmap->data, nbits >> 3),
1053                                 &pos, th);
1054         }
1055
1056         GOTO(out, rc);
1057
1058 out:
1059         dt_trans_stop(env, dev, th);
1060
1061 log:
1062         if (rc != 0)
1063                 CDEBUG(D_LFSCK, "%s: fail to store lfsck_layout: rc = %d\n",
1064                        lfsck_lfsck2name(lfsck), rc);
1065
1066         return rc;
1067 }
1068
1069 static int lfsck_layout_init(const struct lu_env *env,
1070                              struct lfsck_component *com)
1071 {
1072         struct lfsck_layout *lo = com->lc_file_ram;
1073         int rc;
1074
1075         memset(lo, 0, com->lc_file_size);
1076         lo->ll_magic = LFSCK_LAYOUT_MAGIC;
1077         lo->ll_status = LS_INIT;
1078         down_write(&com->lc_sem);
1079         rc = lfsck_layout_store(env, com);
1080         if (rc == 0 && com->lc_lfsck->li_master)
1081                 rc = lfsck_load_sub_trace_files(env, com,
1082                         &dt_lfsck_layout_dangling_features, LFSCK_LAYOUT, true);
1083         up_write(&com->lc_sem);
1084
1085         return rc;
1086 }
1087
1088 static int fid_is_for_ostobj(const struct lu_env *env,
1089                              struct lfsck_instance *lfsck,
1090                              struct dt_object *obj, const struct lu_fid *fid)
1091 {
1092         struct seq_server_site  *ss     = lfsck_dev_site(lfsck);
1093         struct lu_seq_range     *range  = &lfsck_env_info(env)->lti_range;
1094         struct lustre_ost_attrs *loa;
1095         int                      rc;
1096
1097         fld_range_set_any(range);
1098         rc = fld_server_lookup(env, ss->ss_server_fld, fid_seq(fid), range);
1099         if (rc == 0) {
1100                 if (fld_range_is_ost(range))
1101                         return 1;
1102
1103                 return 0;
1104         }
1105
1106         loa = &lfsck_env_info(env)->lti_loa;
1107         rc = dt_xattr_get(env, obj, lfsck_buf_get(env, loa, sizeof(*loa)),
1108                           XATTR_NAME_LMA);
1109         if (rc >= sizeof(struct lustre_mdt_attrs)) {
1110                 lustre_lma_swab(&loa->loa_lma);
1111
1112                 return loa->loa_lma.lma_compat & LMAC_FID_ON_OST ? 1 : 0;
1113         }
1114
1115         rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_FID);
1116
1117         return rc > 0;
1118 }
1119
1120 static struct lfsck_layout_seq *
1121 lfsck_layout_seq_lookup(struct lfsck_layout_slave_data *llsd, __u64 seq)
1122 {
1123         struct lfsck_layout_seq *lls;
1124
1125         list_for_each_entry(lls, &llsd->llsd_seq_list, lls_list) {
1126                 if (lls->lls_seq == seq)
1127                         return lls;
1128
1129                 if (lls->lls_seq > seq)
1130                         return NULL;
1131         }
1132
1133         return NULL;
1134 }
1135
1136 static void
1137 lfsck_layout_seq_insert(struct lfsck_layout_slave_data *llsd,
1138                         struct lfsck_layout_seq *lls)
1139 {
1140         struct lfsck_layout_seq *tmp;
1141         struct list_head        *pos = &llsd->llsd_seq_list;
1142
1143         list_for_each_entry(tmp, &llsd->llsd_seq_list, lls_list) {
1144                 if (lls->lls_seq < tmp->lls_seq) {
1145                         pos = &tmp->lls_list;
1146                         break;
1147                 }
1148         }
1149         list_add_tail(&lls->lls_list, pos);
1150 }
1151
1152 static int
1153 lfsck_layout_lastid_create(const struct lu_env *env,
1154                            struct lfsck_instance *lfsck,
1155                            struct dt_object *obj)
1156 {
1157         struct lfsck_thread_info *info   = lfsck_env_info(env);
1158         struct lu_attr           *la     = &info->lti_la;
1159         struct dt_object_format  *dof    = &info->lti_dof;
1160         struct lfsck_bookmark    *bk     = &lfsck->li_bookmark_ram;
1161         struct dt_device         *dt     = lfsck_obj2dev(obj);
1162         struct thandle           *th;
1163         __u64                     lastid = 0;
1164         loff_t                    pos    = 0;
1165         int                       rc;
1166         ENTRY;
1167
1168         if (bk->lb_param & LPF_DRYRUN)
1169                 return 0;
1170
1171         memset(la, 0, sizeof(*la));
1172         la->la_mode = S_IFREG |  S_IRUGO | S_IWUSR;
1173         la->la_valid = LA_MODE | LA_UID | LA_GID;
1174         memset(dof, 0, sizeof(*dof));
1175         dof->dof_type = dt_mode_to_dft(S_IFREG);
1176
1177         th = dt_trans_create(env, dt);
1178         if (IS_ERR(th))
1179                 GOTO(log, rc = PTR_ERR(th));
1180
1181         rc = dt_declare_create(env, obj, la, NULL, dof, th);
1182         if (rc != 0)
1183                 GOTO(stop, rc);
1184
1185         rc = dt_declare_record_write(env, obj,
1186                                      lfsck_buf_get(env, &lastid,
1187                                                    sizeof(lastid)),
1188                                      pos, th);
1189         if (rc != 0)
1190                 GOTO(stop, rc);
1191
1192         rc = dt_trans_start_local(env, dt, th);
1193         if (rc != 0)
1194                 GOTO(stop, rc);
1195
1196         dt_write_lock(env, obj, 0);
1197         if (likely(dt_object_exists(obj) == 0)) {
1198                 rc = dt_create(env, obj, la, NULL, dof, th);
1199                 if (rc == 0)
1200                         rc = dt_record_write(env, obj,
1201                                 lfsck_buf_get(env, &lastid, sizeof(lastid)),
1202                                 &pos, th);
1203         }
1204         dt_write_unlock(env, obj);
1205
1206         GOTO(stop, rc);
1207
1208 stop:
1209         dt_trans_stop(env, dt, th);
1210
1211 log:
1212         CDEBUG(D_LFSCK, "%s: layout LFSCK will create LAST_ID for <seq> "
1213                "%#llx: rc = %d\n",
1214                lfsck_lfsck2name(lfsck), fid_seq(lfsck_dto2fid(obj)), rc);
1215
1216         return rc;
1217 }
1218
1219 static int
1220 lfsck_layout_lastid_reload(const struct lu_env *env,
1221                            struct lfsck_component *com,
1222                            struct lfsck_layout_seq *lls)
1223 {
1224         __u64   lastid;
1225         loff_t  pos     = 0;
1226         int     rc;
1227
1228         dt_read_lock(env, lls->lls_lastid_obj, 0);
1229         rc = dt_record_read(env, lls->lls_lastid_obj,
1230                             lfsck_buf_get(env, &lastid, sizeof(lastid)), &pos);
1231         dt_read_unlock(env, lls->lls_lastid_obj);
1232         if (unlikely(rc != 0))
1233                 return rc;
1234
1235         lastid = le64_to_cpu(lastid);
1236         if (lastid < lls->lls_lastid_known) {
1237                 struct lfsck_instance   *lfsck  = com->lc_lfsck;
1238                 struct lfsck_layout     *lo     = com->lc_file_ram;
1239
1240                 lls->lls_lastid = lls->lls_lastid_known;
1241                 lls->lls_dirty = 1;
1242                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
1243                         LASSERT(lfsck->li_out_notify != NULL);
1244
1245                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1246                                              LE_LASTID_REBUILDING);
1247                         lo->ll_flags |= LF_CRASHED_LASTID;
1248
1249                         CDEBUG(D_LFSCK, "%s: layout LFSCK finds crashed "
1250                                "LAST_ID file (1) for the sequence %#llx"
1251                                ", old value %llu, known value %llu\n",
1252                                lfsck_lfsck2name(lfsck), lls->lls_seq,
1253                                lastid, lls->lls_lastid);
1254                 }
1255         } else if (lastid >= lls->lls_lastid) {
1256                 lls->lls_lastid = lastid;
1257                 lls->lls_dirty = 0;
1258         }
1259
1260         return 0;
1261 }
1262
1263 static int
1264 lfsck_layout_lastid_store(const struct lu_env *env,
1265                           struct lfsck_component *com)
1266 {
1267         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1268         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
1269         struct dt_device                *dt     = lfsck->li_bottom;
1270         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
1271         struct lfsck_layout_seq         *lls;
1272         struct thandle                  *th;
1273         __u64                            lastid;
1274         int                              rc     = 0;
1275         int                              rc1    = 0;
1276
1277         list_for_each_entry(lls, &llsd->llsd_seq_list, lls_list) {
1278                 loff_t pos = 0;
1279
1280                 if (!lls->lls_dirty)
1281                         continue;
1282
1283                 CDEBUG(D_LFSCK, "%s: layout LFSCK will sync the LAST_ID for "
1284                        "<seq> %#llx as <oid> %llu\n",
1285                        lfsck_lfsck2name(lfsck), lls->lls_seq, lls->lls_lastid);
1286
1287                 if (bk->lb_param & LPF_DRYRUN) {
1288                         lls->lls_dirty = 0;
1289                         continue;
1290                 }
1291
1292                 th = dt_trans_create(env, dt);
1293                 if (IS_ERR(th)) {
1294                         rc1 = PTR_ERR(th);
1295                         CDEBUG(D_LFSCK, "%s: layout LFSCK failed to store "
1296                                "the LAST_ID for <seq> %#llx(1): rc = %d\n",
1297                                lfsck_lfsck2name(com->lc_lfsck),
1298                                lls->lls_seq, rc1);
1299                         continue;
1300                 }
1301
1302                 lastid = cpu_to_le64(lls->lls_lastid);
1303                 rc = dt_declare_record_write(env, lls->lls_lastid_obj,
1304                                              lfsck_buf_get(env, &lastid,
1305                                                            sizeof(lastid)),
1306                                              pos, th);
1307                 if (rc != 0)
1308                         goto stop;
1309
1310                 rc = dt_trans_start_local(env, dt, th);
1311                 if (rc != 0)
1312                         goto stop;
1313
1314                 dt_write_lock(env, lls->lls_lastid_obj, 0);
1315                 rc = dt_record_write(env, lls->lls_lastid_obj,
1316                                      lfsck_buf_get(env, &lastid,
1317                                      sizeof(lastid)), &pos, th);
1318                 dt_write_unlock(env, lls->lls_lastid_obj);
1319                 if (rc == 0)
1320                         lls->lls_dirty = 0;
1321
1322 stop:
1323                 dt_trans_stop(env, dt, th);
1324                 if (rc != 0) {
1325                         rc1 = rc;
1326                         CDEBUG(D_LFSCK, "%s: layout LFSCK failed to store "
1327                                "the LAST_ID for <seq> %#llx(2): rc = %d\n",
1328                                lfsck_lfsck2name(com->lc_lfsck),
1329                                lls->lls_seq, rc1);
1330                 }
1331         }
1332
1333         return rc1;
1334 }
1335
1336 static int
1337 lfsck_layout_lastid_load(const struct lu_env *env,
1338                          struct lfsck_component *com,
1339                          struct lfsck_layout_seq *lls)
1340 {
1341         struct lfsck_instance   *lfsck  = com->lc_lfsck;
1342         struct lfsck_layout     *lo     = com->lc_file_ram;
1343         struct lu_fid           *fid    = &lfsck_env_info(env)->lti_fid;
1344         struct dt_object        *obj;
1345         loff_t                   pos    = 0;
1346         int                      rc;
1347         ENTRY;
1348
1349         lu_last_id_fid(fid, lls->lls_seq, lfsck_dev_idx(lfsck));
1350         obj = dt_locate(env, lfsck->li_bottom, fid);
1351         if (IS_ERR(obj))
1352                 RETURN(PTR_ERR(obj));
1353
1354         /* LAST_ID crashed, to be rebuilt */
1355         if (dt_object_exists(obj) == 0) {
1356                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
1357                         LASSERT(lfsck->li_out_notify != NULL);
1358
1359                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1360                                              LE_LASTID_REBUILDING);
1361                         lo->ll_flags |= LF_CRASHED_LASTID;
1362
1363                         CDEBUG(D_LFSCK, "%s: layout LFSCK cannot find the "
1364                                "LAST_ID file for sequence %#llx\n",
1365                                lfsck_lfsck2name(lfsck), lls->lls_seq);
1366
1367                         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DELAY4) &&
1368                             cfs_fail_val > 0) {
1369                                 struct l_wait_info lwi = LWI_TIMEOUT(
1370                                                 cfs_time_seconds(cfs_fail_val),
1371                                                 NULL, NULL);
1372
1373                                 /* Some others may changed the cfs_fail_val
1374                                  * as zero after above check, re-check it for
1375                                  * sure to avoid falling into wait for ever. */
1376                                 if (likely(lwi.lwi_timeout > 0)) {
1377                                         struct ptlrpc_thread *thread =
1378                                                 &lfsck->li_thread;
1379
1380                                         up_write(&com->lc_sem);
1381                                         l_wait_event(thread->t_ctl_waitq,
1382                                                      !thread_is_running(thread),
1383                                                      &lwi);
1384                                         down_write(&com->lc_sem);
1385                                 }
1386                         }
1387                 }
1388
1389                 rc = lfsck_layout_lastid_create(env, lfsck, obj);
1390         } else {
1391                 dt_read_lock(env, obj, 0);
1392                 rc = dt_read(env, obj,
1393                         lfsck_buf_get(env, &lls->lls_lastid, sizeof(__u64)),
1394                         &pos);
1395                 dt_read_unlock(env, obj);
1396                 if (rc != 0 && rc != sizeof(__u64))
1397                         GOTO(out, rc = (rc > 0 ? -EFAULT : rc));
1398
1399                 if (rc == 0 && !(lo->ll_flags & LF_CRASHED_LASTID)) {
1400                         LASSERT(lfsck->li_out_notify != NULL);
1401
1402                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1403                                              LE_LASTID_REBUILDING);
1404                         lo->ll_flags |= LF_CRASHED_LASTID;
1405
1406                         CDEBUG(D_LFSCK, "%s: layout LFSCK finds invalid "
1407                                "LAST_ID file for the sequence %#llx"
1408                                ": rc = %d\n",
1409                                lfsck_lfsck2name(lfsck), lls->lls_seq, rc);
1410                 }
1411
1412                 lls->lls_lastid = le64_to_cpu(lls->lls_lastid);
1413                 rc = 0;
1414         }
1415
1416         GOTO(out, rc);
1417
1418 out:
1419         if (rc != 0)
1420                 lfsck_object_put(env, obj);
1421         else
1422                 lls->lls_lastid_obj = obj;
1423
1424         return rc;
1425 }
1426
1427 static void lfsck_layout_record_failure(const struct lu_env *env,
1428                                         struct lfsck_instance *lfsck,
1429                                         struct lfsck_layout *lo)
1430 {
1431         __u64 cookie;
1432
1433         lo->ll_objs_failed_phase1++;
1434         cookie = lfsck->li_obj_oit->do_index_ops->dio_it.store(env,
1435                                                         lfsck->li_di_oit);
1436         if (lo->ll_pos_first_inconsistent == 0 ||
1437             lo->ll_pos_first_inconsistent < cookie) {
1438                 lo->ll_pos_first_inconsistent = cookie;
1439
1440                 CDEBUG(D_LFSCK, "%s: layout LFSCK hit first non-repaired "
1441                        "inconsistency at the pos [%llu]\n",
1442                        lfsck_lfsck2name(lfsck),
1443                        lo->ll_pos_first_inconsistent);
1444         }
1445 }
1446
1447 static int lfsck_layout_double_scan_result(const struct lu_env *env,
1448                                            struct lfsck_component *com,
1449                                            int rc)
1450 {
1451         struct lfsck_instance   *lfsck = com->lc_lfsck;
1452         struct lfsck_layout     *lo    = com->lc_file_ram;
1453
1454         CDEBUG(D_LFSCK, "%s: layout LFSCK double scan: rc = %d\n",
1455                lfsck_lfsck2name(lfsck), rc);
1456
1457         down_write(&com->lc_sem);
1458         lo->ll_run_time_phase2 += ktime_get_seconds() -
1459                                   com->lc_time_last_checkpoint;
1460         lo->ll_time_last_checkpoint = ktime_get_real_seconds();
1461         lo->ll_objs_checked_phase2 += com->lc_new_checked;
1462
1463         if (rc > 0) {
1464                 if (lo->ll_flags & LF_INCOMPLETE) {
1465                         lo->ll_status = LS_PARTIAL;
1466                 } else {
1467                         if (lfsck->li_master) {
1468                                 struct lfsck_assistant_data *lad = com->lc_data;
1469
1470                                 if (lad->lad_incomplete)
1471                                         lo->ll_status = LS_PARTIAL;
1472                                 else
1473                                         lo->ll_status = LS_COMPLETED;
1474                         } else {
1475                                 lo->ll_status = LS_COMPLETED;
1476                         }
1477                 }
1478                 lo->ll_flags &= ~LF_SCANNED_ONCE;
1479                 if (!(lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN))
1480                         lo->ll_flags &= ~LF_INCONSISTENT;
1481                 lo->ll_time_last_complete = lo->ll_time_last_checkpoint;
1482                 lo->ll_success_count++;
1483         } else if (rc == 0) {
1484                 if (lfsck->li_status != 0)
1485                         lo->ll_status = lfsck->li_status;
1486                 else
1487                         lo->ll_status = LS_STOPPED;
1488         } else {
1489                 lo->ll_status = LS_FAILED;
1490         }
1491
1492         rc = lfsck_layout_store(env, com);
1493         up_write(&com->lc_sem);
1494
1495         CDEBUG(D_LFSCK, "%s: layout LFSCK double scan result %u: rc = %d\n",
1496                lfsck_lfsck2name(lfsck), lo->ll_status, rc);
1497
1498         return rc;
1499 }
1500
1501 static int lfsck_layout_trans_stop(const struct lu_env *env,
1502                                    struct dt_device *dev,
1503                                    struct thandle *handle, int result)
1504 {
1505         int rc;
1506
1507         /* XXX: If there is something worng or it needs to repair nothing,
1508          *      then notify the lower to stop the modification. Currently,
1509          *      we use th_result for such purpose, that may be replaced by
1510          *      some rollback mechanism in the future. */
1511         handle->th_result = result;
1512         rc = dt_trans_stop(env, dev, handle);
1513         if (result != 0)
1514                 return result > 0 ? 0 : result;
1515
1516         return rc == 0 ? 1 : rc;
1517 }
1518
1519 static int lfsck_layout_ins_dangling_rec(const struct lu_env *env,
1520                                          struct lfsck_component *com,
1521                                          const struct lu_fid *pfid,
1522                                          const struct lu_fid *cfid,
1523                                          __u32 comp_id, __u32 ea_off,
1524                                          __u32 ost_idx)
1525 {
1526         struct lfsck_layout_dangling_key *key = &lfsck_env_info(env)->lti_lldk;
1527         struct lu_fid *rec = &lfsck_env_info(env)->lti_fid3;
1528         struct dt_device *dev;
1529         struct dt_object *obj;
1530         struct thandle *th = NULL;
1531         int idx;
1532         int rc = 0;
1533         ENTRY;
1534
1535         idx = lfsck_sub_trace_file_fid2idx(pfid);
1536         obj = com->lc_sub_trace_objs[idx].lsto_obj;
1537         dev = lfsck_obj2dev(obj);
1538
1539         fid_cpu_to_be(&key->lldk_fid, pfid);
1540         key->lldk_comp_id = cpu_to_be32(comp_id);
1541         key->lldk_ea_off = cpu_to_be32(ea_off);
1542
1543         fid_cpu_to_be(rec, cfid);
1544         rec->f_ver = cpu_to_be32(ost_idx);
1545
1546         mutex_lock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1547
1548         th = dt_trans_create(env, dev);
1549         if (IS_ERR(th))
1550                 GOTO(unlock, rc = PTR_ERR(th));
1551
1552         rc = dt_declare_insert(env, obj,
1553                                (const struct dt_rec *)rec,
1554                                (const struct dt_key *)key, th);
1555         if (rc)
1556                 GOTO(unlock, rc);
1557
1558         rc = dt_trans_start_local(env, dev, th);
1559         if (rc)
1560                 GOTO(unlock, rc);
1561
1562         rc = dt_insert(env, obj, (const struct dt_rec *)rec,
1563                        (const struct dt_key *)key, th);
1564
1565         GOTO(unlock, rc);
1566
1567 unlock:
1568         if (th && !IS_ERR(th))
1569                 dt_trans_stop(env, dev, th);
1570
1571         mutex_unlock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1572
1573         CDEBUG(D_LFSCK, "%s: insert the paris "DFID" => "DFID", comp_id = %u, "
1574                "ea_off = %u, ost_idx = %u, into the trace file for further "
1575                "dangling check: rc = %d\n", lfsck_lfsck2name(com->lc_lfsck),
1576                PFID(pfid), PFID(cfid), comp_id, ea_off, ost_idx, rc);
1577
1578         return rc;
1579 }
1580
1581 static int lfsck_layout_del_dangling_rec(const struct lu_env *env,
1582                                          struct lfsck_component *com,
1583                                          const struct lu_fid *fid,
1584                                          __u32 comp_id, __u32 ea_off)
1585 {
1586         struct lfsck_layout_dangling_key *key = &lfsck_env_info(env)->lti_lldk;
1587         struct dt_device *dev;
1588         struct dt_object *obj;
1589         struct thandle *th = NULL;
1590         int idx;
1591         int rc = 0;
1592         ENTRY;
1593
1594         idx = lfsck_sub_trace_file_fid2idx(fid);
1595         obj = com->lc_sub_trace_objs[idx].lsto_obj;
1596         dev = lfsck_obj2dev(obj);
1597
1598         fid_cpu_to_be(&key->lldk_fid, fid);
1599         key->lldk_comp_id = cpu_to_be32(comp_id);
1600         key->lldk_ea_off = cpu_to_be32(ea_off);
1601
1602         mutex_lock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1603
1604         th = dt_trans_create(env, dev);
1605         if (IS_ERR(th))
1606                 GOTO(unlock, rc = PTR_ERR(th));
1607
1608         rc = dt_declare_delete(env, obj, (const struct dt_key *)key, th);
1609         if (rc)
1610                 GOTO(unlock, rc);
1611
1612         rc = dt_trans_start_local(env, dev, th);
1613         if (rc)
1614                 GOTO(unlock, rc);
1615
1616         rc = dt_delete(env, obj, (const struct dt_key *)key, th);
1617
1618         GOTO(unlock, rc);
1619
1620 unlock:
1621         if (th && !IS_ERR(th))
1622                 dt_trans_stop(env, dev, th);
1623
1624         mutex_unlock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1625
1626         CDEBUG(D_LFSCK, "%s: delete the dangling record for "DFID
1627                ", comp_id = %u, ea_off = %u from the trace file: rc = %d\n",
1628                lfsck_lfsck2name(com->lc_lfsck), PFID(fid), comp_id, ea_off, rc);
1629
1630         return rc;
1631 }
1632
1633 /**
1634  * Get the system default stripe size.
1635  *
1636  * \param[in] env       pointer to the thread context
1637  * \param[in] lfsck     pointer to the lfsck instance
1638  * \param[out] size     pointer to the default stripe size
1639  *
1640  * \retval              0 for success
1641  * \retval              negative error number on failure
1642  */
1643 static int lfsck_layout_get_def_stripesize(const struct lu_env *env,
1644                                            struct lfsck_instance *lfsck,
1645                                            __u32 *size)
1646 {
1647         struct lov_user_md      *lum = &lfsck_env_info(env)->lti_lum;
1648         struct dt_object        *root;
1649         int                      rc;
1650
1651         root = dt_locate(env, lfsck->li_next, &lfsck->li_local_root_fid);
1652         if (IS_ERR(root))
1653                 return PTR_ERR(root);
1654
1655         /* Get the default stripe size via xattr_get on the backend root. */
1656         rc = dt_xattr_get(env, root, lfsck_buf_get(env, lum, sizeof(*lum)),
1657                           XATTR_NAME_LOV);
1658         if (rc > 0) {
1659                 /* The lum->lmm_stripe_size is LE mode. The *size also
1660                  * should be LE mode. So it is unnecessary to convert. */
1661                 *size = lum->lmm_stripe_size;
1662                 rc = 0;
1663         } else if (unlikely(rc == 0)) {
1664                 rc = -EINVAL;
1665         }
1666
1667         lfsck_object_put(env, root);
1668
1669         return rc;
1670 }
1671
1672 /**
1673  * \retval       +1: repaired
1674  * \retval        0: did nothing
1675  * \retval      -ve: on error
1676  */
1677 static int lfsck_layout_refill_lovea(const struct lu_env *env,
1678                                      struct lfsck_instance *lfsck,
1679                                      struct thandle *handle,
1680                                      struct dt_object *parent,
1681                                      const struct lu_fid *cfid,
1682                                      struct lu_buf *buf,
1683                                      struct lov_mds_md_v1 *lmm,
1684                                      struct lov_ost_data_v1 *slot,
1685                                      int fl, __u32 ost_idx, int size)
1686 {
1687         struct ost_id           *oi     = &lfsck_env_info(env)->lti_oi;
1688         struct lu_buf            ea_buf;
1689         int                      rc;
1690         __u32                    magic;
1691         __u32                    pattern;
1692         __u16                    count;
1693         ENTRY;
1694
1695         magic = le32_to_cpu(lmm->lmm_magic);
1696         pattern = le32_to_cpu(lmm->lmm_pattern);
1697         count = le16_to_cpu(lmm->lmm_stripe_count);
1698
1699         fid_to_ostid(cfid, oi);
1700         ostid_cpu_to_le(oi, &slot->l_ost_oi);
1701         slot->l_ost_gen = cpu_to_le32(0);
1702         slot->l_ost_idx = cpu_to_le32(ost_idx);
1703
1704         if (pattern & LOV_PATTERN_F_HOLE) {
1705                 struct lov_ost_data_v1 *objs;
1706                 int                     i;
1707
1708                 if (magic == LOV_MAGIC_V1)
1709                         objs = &lmm->lmm_objects[0];
1710                 else
1711                         objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
1712                 for (i = 0; i < count; i++, objs++) {
1713                         if (lovea_slot_is_dummy(objs))
1714                                 break;
1715                 }
1716
1717                 /* If the @slot is the last dummy slot to be refilled,
1718                  * then drop LOV_PATTERN_F_HOLE from lmm::lmm_pattern. */
1719                 if (i == count) {
1720                         lmm->lmm_pattern =
1721                                 cpu_to_le32(pattern & ~LOV_PATTERN_F_HOLE);
1722
1723                         CDEBUG(D_LFSCK, "%s: remove layout HOLE for "DFID
1724                                ": parent "DFID"\n", lfsck_lfsck2name(lfsck),
1725                                PFID(cfid), PFID(lfsck_dto2fid(parent)));
1726                 }
1727         }
1728
1729         lfsck_buf_init(&ea_buf, buf->lb_buf, size);
1730         rc = dt_xattr_set(env, parent, &ea_buf, XATTR_NAME_LOV, fl, handle);
1731         if (rc == 0)
1732                 rc = 1;
1733
1734         RETURN(rc);
1735 }
1736
1737 static struct lov_ost_data_v1 *
1738 __lfsck_layout_new_v1_lovea(struct lov_mds_md_v1 *lmm,
1739                             const struct lu_fid *pfid,
1740                             __u32 stripe_size, __u32 ea_off,
1741                             __u32 pattern, __u16 count)
1742 {
1743         lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
1744         lmm->lmm_pattern = cpu_to_le32(pattern);
1745         fid_to_lmm_oi(pfid, &lmm->lmm_oi);
1746         lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
1747         lmm->lmm_stripe_size = cpu_to_le32(stripe_size);
1748         lmm->lmm_stripe_count = cpu_to_le16(count);
1749         lmm->lmm_layout_gen = cpu_to_le16(1);
1750         memset(&lmm->lmm_objects[0], 0,
1751                sizeof(struct lov_ost_data_v1) * count);
1752
1753         return &lmm->lmm_objects[ea_off];
1754 }
1755
1756 static int lfsck_layout_new_v1_lovea(const struct lu_env *env,
1757                                      struct lfsck_instance *lfsck,
1758                                      struct ost_layout *ol,
1759                                      struct dt_object *parent,
1760                                      struct lu_buf *buf, __u32 ea_off,
1761                                      struct lov_mds_md_v1 **lmm,
1762                                      struct lov_ost_data_v1 **objs)
1763 {
1764         int size;
1765         __u32 stripe_size = ol->ol_stripe_size;
1766         __u32 pattern = LOV_PATTERN_RAID0;
1767         __u16 count;
1768
1769         if (ol->ol_stripe_count != 0)
1770                 count = ol->ol_stripe_count;
1771         else
1772                 count = ea_off + 1;
1773
1774         size = lov_mds_md_size(count, LOV_MAGIC_V1);
1775         LASSERTF(buf->lb_len >= size,
1776                  "buffer len %d is less than real size %d\n",
1777                  (int)buf->lb_len, size);
1778
1779         if (stripe_size == 0) {
1780                 int rc;
1781
1782                 rc = lfsck_layout_get_def_stripesize(env, lfsck, &stripe_size);
1783                 if (rc)
1784                         return rc;
1785         }
1786
1787         *lmm = buf->lb_buf;
1788         if (ol->ol_stripe_count > 1 ||
1789             (ol->ol_stripe_count == 0 && ea_off != 0)) {
1790                 pattern |= LOV_PATTERN_F_HOLE;
1791                 memset(&(*lmm)->lmm_objects[0], 0,
1792                        count * sizeof(struct lov_ost_data_v1));
1793         }
1794
1795         *objs = __lfsck_layout_new_v1_lovea(*lmm, lfsck_dto2fid(parent),
1796                                 stripe_size, ea_off, pattern, count);
1797
1798         return size;
1799 }
1800
1801 static int lfsck_layout_new_comp_lovea(const struct lu_env *env,
1802                                        struct lu_orphan_rec_v3 *rec,
1803                                        struct dt_object *parent,
1804                                        struct lu_buf *buf, __u32 ea_off,
1805                                        struct lov_mds_md_v1 **lmm,
1806                                        struct lov_ost_data_v1 **objs)
1807 {
1808         struct ost_layout *ol = &rec->lor_layout;
1809         struct lov_comp_md_v1 *lcm;
1810         struct lov_comp_md_entry_v1 *lcme;
1811         __u32 pattern = LOV_PATTERN_RAID0;
1812         __u32 offset = sizeof(*lcm) + sizeof(*lcme);
1813         int lcme_size = lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
1814         int size = offset + lcme_size;
1815
1816         LASSERTF(buf->lb_len >= size,
1817                  "buffer len %d is less than real size %d\n",
1818                  (int)buf->lb_len, size);
1819
1820         lcm = buf->lb_buf;
1821         lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1822         lcm->lcm_size = cpu_to_le32(size);
1823         if (rec->lor_range) {
1824                 lcm->lcm_layout_gen = cpu_to_le32(rec->lor_layout_version +
1825                                                   rec->lor_range);
1826                 lcm->lcm_flags = cpu_to_le16(LCM_FL_WRITE_PENDING);
1827         } else if (rec->lor_layout_version) {
1828                 lcm->lcm_layout_gen = cpu_to_le32(rec->lor_layout_version +
1829                                                   rec->lor_range);
1830                 lcm->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1831         } else {
1832                 lcm->lcm_layout_gen = cpu_to_le32(1);
1833                 lcm->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1834         }
1835         lcm->lcm_entry_count = cpu_to_le16(1);
1836         /* Currently, we do not know how many mirrors will be, set it as zero
1837          * at the beginning. It will be updated when more mirrors are found. */
1838         lcm->lcm_mirror_count = 0;
1839
1840         lcme = &lcm->lcm_entries[0];
1841         lcme->lcme_id = cpu_to_le32(ol->ol_comp_id);
1842         lcme->lcme_flags = cpu_to_le32(LCME_FL_INIT);
1843         lcme->lcme_extent.e_start = cpu_to_le64(ol->ol_comp_start);
1844         lcme->lcme_extent.e_end = cpu_to_le64(ol->ol_comp_end);
1845         lcme->lcme_offset = cpu_to_le32(offset);
1846         lcme->lcme_size = cpu_to_le32(lcme_size);
1847         lcme->lcme_layout_gen = lcm->lcm_layout_gen;
1848         if (ol->ol_stripe_count > 1)
1849                 pattern |= LOV_PATTERN_F_HOLE;
1850
1851         *lmm = buf->lb_buf + offset;
1852         *objs = __lfsck_layout_new_v1_lovea(*lmm, lfsck_dto2fid(parent),
1853                                             ol->ol_stripe_size, ea_off,
1854                                             pattern, ol->ol_stripe_count);
1855
1856         return size;
1857 }
1858
1859 static void lfsck_layout_update_lcm(struct lov_comp_md_v1 *lcm,
1860                                     struct lov_comp_md_entry_v1 *lcme,
1861                                     __u32 version, __u32 range)
1862 {
1863         struct lov_comp_md_entry_v1 *tmp;
1864         __u64 start = le64_to_cpu(lcme->lcme_extent.e_start);
1865         __u64 end = le64_to_cpu(lcme->lcme_extent.e_end);
1866         __u32 gen = version + range;
1867         __u32 tmp_gen;
1868         int i;
1869         __u16 count = le16_to_cpu(lcm->lcm_entry_count);
1870         __u16 flags = le16_to_cpu(lcm->lcm_flags);
1871
1872         if (!gen)
1873                 gen = 1;
1874         lcme->lcme_layout_gen = cpu_to_le32(gen);
1875         if (le32_to_cpu(lcm->lcm_layout_gen) < gen)
1876                 lcm->lcm_layout_gen = cpu_to_le32(gen);
1877
1878         if (range)
1879                 lcm->lcm_flags = cpu_to_le16(LCM_FL_WRITE_PENDING);
1880         else if (flags == LCM_FL_NONE && le16_to_cpu(lcm->lcm_mirror_count) > 0)
1881                 lcm->lcm_flags = cpu_to_le16(LCM_FL_RDONLY);
1882
1883         for (i = 0; i < count; i++) {
1884                 tmp = &lcm->lcm_entries[i];
1885                 if (le64_to_cpu(tmp->lcme_extent.e_end) <= start)
1886                         continue;
1887
1888                 if (le64_to_cpu(tmp->lcme_extent.e_start) >= end)
1889                         continue;
1890
1891                 if (le32_to_cpu(tmp->lcme_flags) & LCME_FL_STALE)
1892                         continue;
1893
1894                 tmp_gen = le32_to_cpu(tmp->lcme_layout_gen);
1895                 /* "lcme_layout_gen == 0" but without LCME_FL_STALE flag,
1896                  * then it should be the latest version of all mirrors. */
1897                 if (tmp_gen == 0 || tmp_gen > gen) {
1898                         lcme->lcme_flags = cpu_to_le32(
1899                                 le32_to_cpu(lcme->lcme_flags) | LCME_FL_STALE);
1900                         break;
1901                 }
1902
1903                 if (tmp_gen < gen)
1904                         tmp->lcme_flags = cpu_to_le32(
1905                                 le32_to_cpu(tmp->lcme_flags) | LCME_FL_STALE);
1906         }
1907 }
1908
1909 static int lfsck_layout_add_comp(const struct lu_env *env,
1910                                  struct lfsck_instance *lfsck,
1911                                  struct thandle *handle,
1912                                  struct lu_orphan_rec_v3 *rec,
1913                                  struct dt_object *parent,
1914                                  const struct lu_fid *cfid,
1915                                  struct lu_buf *buf, __u32 ost_idx,
1916                                  __u32 ea_off, int pos, bool new_mirror)
1917 {
1918         struct ost_layout *ol = &rec->lor_layout;
1919         struct lov_comp_md_v1 *lcm = buf->lb_buf;
1920         struct lov_comp_md_entry_v1 *lcme;
1921         struct lov_mds_md_v1 *lmm;
1922         struct lov_ost_data_v1 *objs;
1923         int added = sizeof(*lcme) +
1924                     lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
1925         int size = le32_to_cpu(lcm->lcm_size) + added;
1926         int rc;
1927         int i;
1928         __u32 offset;
1929         __u32 pattern = LOV_PATTERN_RAID0;
1930         __u16 count = le16_to_cpu(lcm->lcm_entry_count);
1931         ENTRY;
1932
1933         lu_buf_check_and_grow(buf, size);
1934         /* set the lcm again because lu_buf_check_and_grow() may
1935          * have reallocated the buf. */
1936         lcm = buf->lb_buf;
1937         lcm->lcm_size = cpu_to_le32(size);
1938         lcm->lcm_entry_count = cpu_to_le16(count + 1);
1939         if (new_mirror)
1940                 le16_add_cpu(&lcm->lcm_mirror_count, 1);
1941
1942         /* 1. Move the component bodies from [pos, count-1] to [pos+1, count]
1943          *    with distance of 'added'. */
1944         if (pos < count) {
1945                 size = 0;
1946                 for (i = pos; i < count; i++) {
1947                         lcme = &lcm->lcm_entries[i];
1948                         size += le32_to_cpu(lcme->lcme_size);
1949                 }
1950
1951                 offset = le32_to_cpu(lcm->lcm_entries[pos].lcme_offset);
1952                 memmove(buf->lb_buf + offset + added,
1953                         buf->lb_buf + offset, size);
1954         }
1955
1956         size = 0;
1957         /* 2. Move the component header [0, pos-1] to [0, pos-1] with distance
1958          *    of 'sizeof(struct lov_comp_md_entry_v1)' */
1959         if (pos > 0) {
1960                 for (i = 0; i < pos; i++) {
1961                         lcme = &lcm->lcm_entries[i];
1962                         size += le32_to_cpu(lcme->lcme_size);
1963                 }
1964
1965                 offset = le32_to_cpu(lcm->lcm_entries[0].lcme_offset);
1966                 memmove(buf->lb_buf + offset + sizeof(*lcme),
1967                         buf->lb_buf + offset, size);
1968         }
1969
1970         /* 3. Recalculate the enter offset for the component [pos, count-1] */
1971         for (i = count - 1; i >= pos; i--) {
1972                 lcm->lcm_entries[i + 1] = lcm->lcm_entries[i];
1973                 lcm->lcm_entries[i + 1].lcme_offset =
1974                         cpu_to_le32(le32_to_cpu(lcm->lcm_entries[i + 1].
1975                                                 lcme_offset) + added);
1976         }
1977
1978         /* 4. Recalculate the enter offset for the component [0, pos) */
1979         for (i = 0; i < pos; i++) {
1980                 lcm->lcm_entries[i].lcme_offset =
1981                         cpu_to_le32(le32_to_cpu(lcm->lcm_entries[i].
1982                                                 lcme_offset) + sizeof(*lcme));
1983         }
1984
1985         offset = sizeof(*lcm) + sizeof(*lcme) * (count + 1) + size;
1986         /* 4. Insert the new component header (entry) at the slot 'pos'. */
1987         lcme = &lcm->lcm_entries[pos];
1988         lcme->lcme_id = cpu_to_le32(ol->ol_comp_id);
1989         lcme->lcme_flags = cpu_to_le32(LCME_FL_INIT);
1990         lcme->lcme_extent.e_start = cpu_to_le64(ol->ol_comp_start);
1991         lcme->lcme_extent.e_end = cpu_to_le64(ol->ol_comp_end);
1992         lcme->lcme_offset = cpu_to_le32(offset);
1993         lcme->lcme_size = cpu_to_le32(lov_mds_md_size(ol->ol_stripe_count,
1994                                                       LOV_MAGIC_V1));
1995
1996         if (ol->ol_stripe_count > 1)
1997                 pattern |= LOV_PATTERN_F_HOLE;
1998
1999         lmm = buf->lb_buf + offset;
2000         /* 5. Insert teh new component body at the 'offset'. */
2001         objs = __lfsck_layout_new_v1_lovea(lmm, lfsck_dto2fid(parent),
2002                                            ol->ol_stripe_size, ea_off,
2003                                            pattern, ol->ol_stripe_count);
2004
2005         /* 6. Update mirror related flags and version. */
2006         lfsck_layout_update_lcm(lcm, lcme, rec->lor_layout_version,
2007                                 rec->lor_range);
2008
2009         rc = lfsck_layout_refill_lovea(env, lfsck, handle, parent, cfid, buf,
2010                                        lmm, objs, LU_XATTR_REPLACE, ost_idx,
2011                                        le32_to_cpu(lcm->lcm_size));
2012
2013         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant add new COMP for "
2014                DFID": parent "DFID", OST-index %u, stripe-index %u, "
2015                "stripe_size %u, stripe_count %u, comp_id %u, comp_start %llu, "
2016                "comp_end %llu, layout version %u, range %u, "
2017                "%s LOV EA hole: rc = %d\n",
2018                lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
2019                ost_idx, ea_off, ol->ol_stripe_size, ol->ol_stripe_count,
2020                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
2021                rec->lor_layout_version, rec->lor_range,
2022                le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE ?
2023                "with" : "without", rc);
2024
2025         RETURN(rc);
2026 }
2027
2028 static int lfsck_layout_extend_v1v3_lovea(const struct lu_env *env,
2029                                           struct lfsck_instance *lfsck,
2030                                           struct thandle *handle,
2031                                           struct ost_layout *ol,
2032                                           struct dt_object *parent,
2033                                           const struct lu_fid *cfid,
2034                                           struct lu_buf *buf, __u32 ost_idx,
2035                                           __u32 ea_off)
2036 {
2037         struct lov_mds_md_v1 *lmm = buf->lb_buf;
2038         struct lov_ost_data_v1 *objs;
2039         __u16 count = le16_to_cpu(lmm->lmm_stripe_count);
2040         __u32 magic = le32_to_cpu(lmm->lmm_magic);
2041         int size;
2042         int gap;
2043         int rc;
2044         ENTRY;
2045
2046         /* The original LOVEA maybe re-generated via old filter_fid, at
2047          * that time, we do not know the stripe count and stripe size. */
2048         if (ol->ol_stripe_count > count)
2049                 count = ol->ol_stripe_count;
2050         if (ol->ol_stripe_size != 0 &&
2051             ol->ol_stripe_size != le32_to_cpu(lmm->lmm_stripe_size))
2052                 lmm->lmm_stripe_size = cpu_to_le32(ol->ol_stripe_size);
2053
2054         if (magic == LOV_MAGIC_V1)
2055                 objs = &lmm->lmm_objects[count];
2056         else
2057                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[count];
2058
2059         gap = ea_off - count;
2060         if (gap >= 0)
2061                 count = ea_off + 1;
2062
2063         size = lov_mds_md_size(count, magic);
2064         LASSERTF(buf->lb_len >= size,
2065                  "buffer len %d is less than real size %d\n",
2066                  (int)buf->lb_len, size);
2067
2068         if (gap > 0) {
2069                 memset(objs, 0, gap * sizeof(*objs));
2070                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_HOLE);
2071         }
2072
2073         lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
2074         lmm->lmm_stripe_count = cpu_to_le16(count);
2075         objs += gap;
2076
2077         rc = lfsck_layout_refill_lovea(env, lfsck, handle, parent, cfid, buf,
2078                                 lmm, objs, LU_XATTR_REPLACE, ost_idx, size);
2079
2080         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant extend layout EA for "
2081                DFID": parent "DFID", OST-index %u, stripe-index %u, "
2082                "stripe_size %u, stripe_count %u, comp_id %u, comp_start %llu, "
2083                "comp_end %llu, %s LOV EA hole: rc = %d\n",
2084                lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
2085                ost_idx, ea_off, ol->ol_stripe_size, ol->ol_stripe_count,
2086                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
2087                le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE ?
2088                "with" : "without", rc);
2089
2090         RETURN(rc);
2091 }
2092
2093 /**
2094  * \retval       +1: repaired
2095  * \retval        0: did nothing
2096  * \retval      -ve: on error
2097  */
2098 static int lfsck_layout_update_lovea(const struct lu_env *env,
2099                                      struct lfsck_instance *lfsck,
2100                                      struct thandle *handle,
2101                                      struct lu_orphan_rec_v3 *rec,
2102                                      struct dt_object *parent,
2103                                      const struct lu_fid *cfid,
2104                                      struct lu_buf *buf, int fl,
2105                                      __u32 ost_idx, __u32 ea_off)
2106 {
2107         struct ost_layout *ol = &rec->lor_layout;
2108         struct lov_mds_md_v1 *lmm = NULL;
2109         struct lov_ost_data_v1 *objs = NULL;
2110         int rc = 0;
2111         ENTRY;
2112
2113         if (ol->ol_comp_id != 0)
2114                 rc = lfsck_layout_new_comp_lovea(env, rec, parent, buf, ea_off,
2115                                                  &lmm, &objs);
2116         else
2117                 rc = lfsck_layout_new_v1_lovea(env, lfsck, &rec->lor_layout,
2118                                                parent, buf, ea_off, &lmm,
2119                                                &objs);
2120         if (rc > 0)
2121                 rc = lfsck_layout_refill_lovea(env, lfsck, handle, parent, cfid,
2122                                                buf, lmm, objs, fl, ost_idx, rc);
2123
2124         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant created layout EA for "
2125                DFID": parent "DFID", OST-index %u, stripe-index %u, "
2126                "stripe_size %u, stripe_count %u, comp_id %u, comp_start %llu, "
2127                "comp_end %llu, layout version %u, range %u, fl %d, "
2128                "%s LOV EA hole: rc = %d\n",
2129                lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
2130                ost_idx, ea_off, ol->ol_stripe_size, ol->ol_stripe_count,
2131                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
2132                rec->lor_layout_version, rec->lor_range, fl,
2133                le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE ?
2134                "with" : "without", rc);
2135
2136         RETURN(rc);
2137 }
2138
2139 static int __lfsck_layout_update_pfid(const struct lu_env *env,
2140                                       struct dt_object *child,
2141                                       const struct lu_fid *pfid,
2142                                       const struct ost_layout *ol, __u32 offset,
2143                                       __u32 version, __u32 range)
2144 {
2145         struct dt_device        *dev    = lfsck_obj2dev(child);
2146         struct filter_fid       *ff     = &lfsck_env_info(env)->lti_ff;
2147         struct thandle          *handle;
2148         struct lu_buf            buf    = { NULL };
2149         int                      rc;
2150
2151         ff->ff_parent.f_seq = cpu_to_le64(pfid->f_seq);
2152         ff->ff_parent.f_oid = cpu_to_le32(pfid->f_oid);
2153         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
2154          * MDT-object's FID::f_ver, instead it is the OST-object index in its
2155          * parent MDT-object's layout EA. */
2156         ff->ff_parent.f_stripe_idx = cpu_to_le32(offset);
2157         ost_layout_cpu_to_le(&ff->ff_layout, ol);
2158         ff->ff_layout_version = cpu_to_le32(version);
2159         ff->ff_range = cpu_to_le32(range);
2160         lfsck_buf_init(&buf, ff, sizeof(*ff));
2161
2162         handle = dt_trans_create(env, dev);
2163         if (IS_ERR(handle))
2164                 RETURN(PTR_ERR(handle));
2165
2166         rc = dt_declare_xattr_set(env, child, &buf, XATTR_NAME_FID, 0, handle);
2167         if (rc != 0)
2168                 GOTO(stop, rc);
2169
2170         rc = dt_trans_start_local(env, dev, handle);
2171         if (rc != 0)
2172                 GOTO(stop, rc);
2173
2174         rc = dt_xattr_set(env, child, &buf, XATTR_NAME_FID, 0, handle);
2175
2176         GOTO(stop, rc);
2177
2178 stop:
2179         dt_trans_stop(env, dev, handle);
2180
2181         return rc;
2182 }
2183
2184 /**
2185  * \retval       +1: repaired
2186  * \retval        0: did nothing
2187  * \retval      -ve: on error
2188  */
2189 static int lfsck_layout_update_pfid(const struct lu_env *env,
2190                                     struct lfsck_component *com,
2191                                     struct dt_object *parent,
2192                                     struct lu_fid *cfid,
2193                                     struct dt_device *cdev,
2194                                     struct lu_orphan_rec_v3 *rec, __u32 ea_off)
2195 {
2196         struct dt_object        *child;
2197         int                      rc     = 0;
2198         ENTRY;
2199
2200         child = lfsck_object_find_by_dev(env, cdev, cfid);
2201         if (IS_ERR(child))
2202                 RETURN(PTR_ERR(child));
2203
2204         rc = __lfsck_layout_update_pfid(env, child,
2205                                         lu_object_fid(&parent->do_lu),
2206                                         &rec->lor_layout, ea_off,
2207                                         rec->lor_layout_version,
2208                                         rec->lor_range);
2209         lfsck_object_put(env, child);
2210
2211         RETURN(rc == 0 ? 1 : rc);
2212 }
2213
2214 static int lfsck_lovea_size(struct ost_layout *ol, __u32 ea_off)
2215 {
2216         if (ol->ol_comp_id != 0)
2217                 return sizeof(struct lov_comp_md_v1) +
2218                        sizeof(struct lov_comp_md_entry_v1) +
2219                        lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
2220
2221         if (ol->ol_stripe_count != 0)
2222                 return lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
2223
2224         return lov_mds_md_size(ea_off + 1, LOV_MAGIC_V1);
2225 }
2226
2227 /**
2228  * This function will create the MDT-object with the given (partial) LOV EA.
2229  *
2230  * Under some data corruption cases, the MDT-object of the file may be lost,
2231  * but its OST-objects, or some of them are there. The layout LFSCK needs to
2232  * re-create the MDT-object with the orphan OST-object(s) information.
2233  *
2234  * On the other hand, the LFSCK may has created some OST-object for repairing
2235  * dangling LOV EA reference, but as the LFSCK processing, it may find that
2236  * the old OST-object is there and should replace the former new created OST
2237  * object. Unfortunately, some others have modified such newly created object.
2238  * To keep the data (both new and old), the LFSCK will create MDT-object with
2239  * new FID to reference the original OST-object.
2240  *
2241  * \param[in] env       pointer to the thread context
2242  * \param[in] com       pointer to the lfsck component
2243  * \param[in] ltd       pointer to target device descriptor
2244  * \param[in] rec       pointer to the record for the orphan OST-object
2245  * \param[in] cfid      pointer to FID for the orphan OST-object
2246  * \param[in] infix     additional information, such as the FID for original
2247  *                      MDT-object and the stripe offset in the LOV EA
2248  * \param[in] type      the type for describing why the orphan MDT-object is
2249  *                      created. The rules are as following:
2250  *
2251  *  type "C":           Multiple OST-objects claim the same MDT-object and the
2252  *                      same slot in the layout EA. Then the LFSCK will create
2253  *                      new MDT-object(s) to hold the conflict OST-object(s).
2254  *
2255  *  type "N":           The orphan OST-object does not know which one was the
2256  *                      real parent MDT-object, so the LFSCK uses new FID for
2257  *                      its parent MDT-object.
2258  *
2259  *  type "R":           The orphan OST-object knows its parent MDT-object FID,
2260  *                      but does not know the position (the file name) in the
2261  *                      layout.
2262  *
2263  *  type "D":           The MDT-object is a directory, it may knows its parent
2264  *                      but because there is no valid linkEA, the LFSCK cannot
2265  *                      know where to put it back to the namespace.
2266  *  type "O":           The MDT-object has no linkEA, and there is no name
2267  *                      entry that references the MDT-object.
2268  *
2269  *  type "P":           The orphan object to be created was a parent directory
2270  *                      of some MDT-object which linkEA shows that the @orphan
2271  *                      object is missing.
2272  *
2273  * The orphan name will be like:
2274  * ${FID}-${infix}-${type}-${conflict_version}
2275  *
2276  * \param[in] ea_off    the stripe offset in the LOV EA
2277  *
2278  * \retval              positive on repaired something
2279  * \retval              0 if needs to repair nothing
2280  * \retval              negative error number on failure
2281  */
2282 static int lfsck_layout_recreate_parent(const struct lu_env *env,
2283                                         struct lfsck_component *com,
2284                                         struct lfsck_tgt_desc *ltd,
2285                                         struct lu_orphan_rec_v3 *rec,
2286                                         struct lu_fid *cfid,
2287                                         const char *infix,
2288                                         const char *type,
2289                                         __u32 ea_off)
2290 {
2291         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2292         struct dt_insert_rec            *dtrec  = &info->lti_dt_rec;
2293         char                            *name   = info->lti_key;
2294         struct lu_attr                  *la     = &info->lti_la2;
2295         struct dt_object_format         *dof    = &info->lti_dof;
2296         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2297         struct lu_fid                   *pfid   = &rec->lor_rec.lor_fid;
2298         struct lu_fid                   *tfid   = &info->lti_fid3;
2299         struct dt_device                *dev    = lfsck->li_bottom;
2300         struct dt_object                *lpf    = lfsck->li_lpf_obj;
2301         struct dt_object                *pobj   = NULL;
2302         struct dt_object                *cobj   = NULL;
2303         struct thandle                  *th     = NULL;
2304         struct lu_buf                   *ea_buf = &info->lti_big_buf;
2305         struct lu_buf                    lov_buf;
2306         struct lfsck_lock_handle        *llh    = &info->lti_llh;
2307         struct linkea_data               ldata  = { NULL };
2308         struct lu_buf                    linkea_buf;
2309         const struct lu_name            *pname;
2310         int                              size   = 0;
2311         int                              idx    = 0;
2312         int                              rc     = 0;
2313         ENTRY;
2314
2315         if (unlikely(lpf == NULL))
2316                 GOTO(log, rc = -ENXIO);
2317
2318         /* We use two separated transactions to repair the inconsistency.
2319          *
2320          * 1) create the MDT-object locally.
2321          * 2) update the OST-object's PFID EA if necessary.
2322          *
2323          * If 1) succeed, but 2) failed, then the OST-object's PFID EA will be
2324          * updated when the layout LFSCK run next time.
2325          *
2326          * If 1) failed, but 2) succeed, then such MDT-object will be re-created
2327          * when the layout LFSCK run next time. */
2328
2329         if (fid_is_zero(pfid)) {
2330                 rc = lfsck_fid_alloc(env, lfsck, pfid, false);
2331                 if (rc != 0)
2332                         GOTO(log, rc);
2333
2334                 cobj = lfsck_object_find_by_dev(env, ltd->ltd_tgt, cfid);
2335                 if (IS_ERR(cobj))
2336                         GOTO(log, rc = PTR_ERR(cobj));
2337         }
2338
2339         pobj = lfsck_object_find_by_dev(env, dev, pfid);
2340         if (IS_ERR(pobj))
2341                 GOTO(log, rc = PTR_ERR(pobj));
2342
2343         LASSERT(infix != NULL);
2344         LASSERT(type != NULL);
2345
2346         memset(la, 0, sizeof(*la));
2347         la->la_uid = rec->lor_rec.lor_uid;
2348         la->la_gid = rec->lor_rec.lor_gid;
2349         la->la_mode = S_IFREG | S_IRUSR;
2350         la->la_valid = LA_MODE | LA_UID | LA_GID;
2351
2352         memset(dof, 0, sizeof(*dof));
2353         dof->dof_type = dt_mode_to_dft(S_IFREG);
2354         /* Because the dof->dof_reg.striped = 0, the LOD will not create
2355          * the stripe(s). The LFSCK will specify the LOV EA via
2356          * lfsck_layout_update_lovea(). */
2357
2358         size = lfsck_lovea_size(&rec->lor_layout, ea_off);
2359         if (ea_buf->lb_len < size) {
2360                 lu_buf_realloc(ea_buf, size);
2361                 if (ea_buf->lb_buf == NULL)
2362                         GOTO(log, rc = -ENOMEM);
2363         }
2364
2365 again:
2366         do {
2367                 snprintf(name, NAME_MAX, DFID"%s-%s-%d", PFID(pfid), infix,
2368                          type, idx++);
2369                 rc = dt_lookup(env, lfsck->li_lpf_obj, (struct dt_rec *)tfid,
2370                                (const struct dt_key *)name);
2371                 if (rc != 0 && rc != -ENOENT)
2372                         GOTO(log, rc);
2373         } while (rc == 0);
2374
2375         rc = lfsck_lock(env, lfsck, lfsck->li_lpf_obj, name, llh,
2376                         MDS_INODELOCK_UPDATE, LCK_PW);
2377         if (rc != 0)
2378                 GOTO(log, rc);
2379
2380         /* Re-check whether the name conflict with othrs after taken
2381          * the ldlm lock. */
2382         rc = dt_lookup(env, lfsck->li_lpf_obj, (struct dt_rec *)tfid,
2383                        (const struct dt_key *)name);
2384         if (unlikely(rc == 0)) {
2385                 lfsck_unlock(llh);
2386                 goto again;
2387         }
2388
2389         if (rc != -ENOENT)
2390                 GOTO(unlock, rc);
2391
2392         pname = lfsck_name_get_const(env, name, strlen(name));
2393         rc = linkea_links_new(&ldata, &lfsck_env_info(env)->lti_linkea_buf,
2394                               pname, lfsck_dto2fid(lfsck->li_lpf_obj));
2395         if (rc != 0)
2396                 GOTO(unlock, rc);
2397
2398         /* The 1st transaction. */
2399         th = dt_trans_create(env, dev);
2400         if (IS_ERR(th))
2401                 GOTO(unlock, rc = PTR_ERR(th));
2402
2403         rc = dt_declare_create(env, pobj, la, NULL, dof, th);
2404         if (rc != 0)
2405                 GOTO(stop, rc);
2406
2407         lfsck_buf_init(&lov_buf, ea_buf->lb_buf, size);
2408         rc = dt_declare_xattr_set(env, pobj, &lov_buf, XATTR_NAME_LOV,
2409                                   LU_XATTR_CREATE, th);
2410         if (rc != 0)
2411                 GOTO(stop, rc);
2412
2413         dtrec->rec_fid = pfid;
2414         dtrec->rec_type = S_IFREG;
2415         rc = dt_declare_insert(env, lpf,
2416                                (const struct dt_rec *)dtrec,
2417                                (const struct dt_key *)name, th);
2418         if (rc != 0)
2419                 GOTO(stop, rc);
2420
2421         lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
2422                        ldata.ld_leh->leh_len);
2423         rc = dt_declare_xattr_set(env, pobj, &linkea_buf,
2424                                   XATTR_NAME_LINK, 0, th);
2425         if (rc != 0)
2426                 GOTO(stop, rc);
2427
2428         rc = dt_trans_start_local(env, dev, th);
2429         if (rc != 0)
2430                 GOTO(stop, rc);
2431
2432         dt_write_lock(env, pobj, 0);
2433         rc = dt_create(env, pobj, la, NULL, dof, th);
2434         if (rc == 0)
2435                 rc = lfsck_layout_update_lovea(env, lfsck, th, rec, pobj, cfid,
2436                         &lov_buf, LU_XATTR_CREATE, ltd->ltd_index, ea_off);
2437         dt_write_unlock(env, pobj);
2438         if (rc < 0)
2439                 GOTO(stop, rc);
2440
2441         rc = dt_insert(env, lpf, (const struct dt_rec *)dtrec,
2442                        (const struct dt_key *)name, th);
2443         if (rc != 0)
2444                 GOTO(stop, rc);
2445
2446         rc = dt_xattr_set(env, pobj, &linkea_buf, XATTR_NAME_LINK, 0, th);
2447         if (rc == 0 && cobj != NULL) {
2448                 dt_trans_stop(env, dev, th);
2449                 th = NULL;
2450
2451                 /* The 2nd transaction. */
2452                 rc = __lfsck_layout_update_pfid(env, cobj, pfid,
2453                                                 &rec->lor_layout, ea_off,
2454                                                 rec->lor_layout_version,
2455                                                 rec->lor_range);
2456         }
2457
2458         GOTO(stop, rc);
2459
2460 stop:
2461         if (th != NULL)
2462                 dt_trans_stop(env, dev, th);
2463
2464 unlock:
2465         lfsck_unlock(llh);
2466
2467 log:
2468         if (cobj != NULL && !IS_ERR(cobj))
2469                 lfsck_object_put(env, cobj);
2470         if (pobj != NULL && !IS_ERR(pobj))
2471                 lfsck_object_put(env, pobj);
2472
2473         if (rc < 0)
2474                 CDEBUG(D_LFSCK, "%s layout LFSCK assistant failed to "
2475                        "recreate the lost MDT-object: parent "DFID
2476                        ", child "DFID", OST-index %u, stripe-index %u, "
2477                        "infix %s, type %s: rc = %d\n",
2478                        lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
2479                        ltd->ltd_index, ea_off, infix, type, rc);
2480
2481         return rc >= 0 ? 1 : rc;
2482 }
2483
2484 static int lfsck_layout_master_conditional_destroy(const struct lu_env *env,
2485                                                    struct lfsck_component *com,
2486                                                    const struct lu_fid *fid,
2487                                                    __u32 index)
2488 {
2489         struct lfsck_thread_info *info  = lfsck_env_info(env);
2490         struct lfsck_request     *lr    = &info->lti_lr;
2491         struct lfsck_instance    *lfsck = com->lc_lfsck;
2492         struct lfsck_tgt_desc    *ltd;
2493         struct ptlrpc_request    *req;
2494         struct lfsck_request     *tmp;
2495         struct obd_export        *exp;
2496         int                       rc    = 0;
2497         ENTRY;
2498
2499         ltd = lfsck_tgt_get(&lfsck->li_ost_descs, index);
2500         if (unlikely(ltd == NULL))
2501                 RETURN(-ENXIO);
2502
2503         exp = ltd->ltd_exp;
2504         if (!(exp_connect_flags(exp) & OBD_CONNECT_LFSCK))
2505                 GOTO(put, rc = -EOPNOTSUPP);
2506
2507         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
2508         if (req == NULL)
2509                 GOTO(put, rc = -ENOMEM);
2510
2511         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
2512         if (rc != 0) {
2513                 ptlrpc_request_free(req);
2514
2515                 GOTO(put, rc);
2516         }
2517
2518         memset(lr, 0, sizeof(*lr));
2519         lr->lr_event = LE_CONDITIONAL_DESTROY;
2520         lr->lr_active = LFSCK_TYPE_LAYOUT;
2521         lr->lr_fid = *fid;
2522
2523         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
2524         *tmp = *lr;
2525         ptlrpc_request_set_replen(req);
2526
2527         rc = ptlrpc_queue_wait(req);
2528         ptlrpc_req_finished(req);
2529
2530         GOTO(put, rc);
2531
2532 put:
2533         lfsck_tgt_put(ltd);
2534
2535         return rc;
2536 }
2537
2538 static int lfsck_layout_slave_conditional_destroy(const struct lu_env *env,
2539                                                   struct lfsck_component *com,
2540                                                   struct lfsck_request *lr)
2541 {
2542         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2543         struct lu_attr                  *la     = &info->lti_la;
2544         union ldlm_policy_data          *policy = &info->lti_policy;
2545         struct ldlm_res_id              *resid  = &info->lti_resid;
2546         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2547         struct dt_device                *dev    = lfsck->li_bottom;
2548         struct lu_fid                   *fid    = &lr->lr_fid;
2549         struct dt_object                *obj;
2550         struct thandle                  *th     = NULL;
2551         struct lustre_handle             lh     = { 0 };
2552         __u64                            flags  = 0;
2553         int                              rc     = 0;
2554         ENTRY;
2555
2556         obj = lfsck_object_find_by_dev(env, dev, fid);
2557         if (IS_ERR(obj))
2558                 RETURN(PTR_ERR(obj));
2559
2560         dt_read_lock(env, obj, 0);
2561         if (dt_object_exists(obj) == 0 ||
2562             lfsck_is_dead_obj(obj)) {
2563                 dt_read_unlock(env, obj);
2564
2565                 GOTO(put, rc = -ENOENT);
2566         }
2567
2568         /* Get obj's attr without lock firstly. */
2569         rc = dt_attr_get(env, obj, la);
2570         dt_read_unlock(env, obj);
2571         if (rc != 0)
2572                 GOTO(put, rc);
2573
2574         if (likely(la->la_ctime != 0 || la->la_mode & S_ISUID))
2575                 GOTO(put, rc = -ETXTBSY);
2576
2577         /* Acquire extent lock on [0, EOF] to sync with all possible written. */
2578         LASSERT(lfsck->li_namespace != NULL);
2579
2580         memset(policy, 0, sizeof(*policy));
2581         policy->l_extent.end = OBD_OBJECT_EOF;
2582         ost_fid_build_resid(fid, resid);
2583         rc = ldlm_cli_enqueue_local(env, lfsck->li_namespace, resid,
2584                                     LDLM_EXTENT, policy, LCK_EX, &flags,
2585                                     ldlm_blocking_ast, ldlm_completion_ast,
2586                                     NULL, NULL, 0, LVB_T_NONE, NULL, &lh);
2587         if (rc != ELDLM_OK)
2588                 GOTO(put, rc = -EIO);
2589
2590         dt_write_lock(env, obj, 0);
2591         /* Get obj's attr within lock again. */
2592         rc = dt_attr_get(env, obj, la);
2593         if (rc != 0)
2594                 GOTO(unlock, rc);
2595
2596         if (la->la_ctime != 0)
2597                 GOTO(unlock, rc = -ETXTBSY);
2598
2599         th = dt_trans_create(env, dev);
2600         if (IS_ERR(th))
2601                 GOTO(unlock, rc = PTR_ERR(th));
2602
2603         rc = dt_declare_ref_del(env, obj, th);
2604         if (rc != 0)
2605                 GOTO(stop, rc);
2606
2607         rc = dt_declare_destroy(env, obj, th);
2608         if (rc != 0)
2609                 GOTO(stop, rc);
2610
2611         rc = dt_trans_start_local(env, dev, th);
2612         if (rc != 0)
2613                 GOTO(stop, rc);
2614
2615         rc = dt_ref_del(env, obj, th);
2616         if (rc != 0)
2617                 GOTO(stop, rc);
2618
2619         rc = dt_destroy(env, obj, th);
2620         if (rc == 0)
2621                 CDEBUG(D_LFSCK, "%s: layout LFSCK destroyed the empty "
2622                        "OST-object "DFID" that was created for reparing "
2623                        "dangling referenced case. But the original missing "
2624                        "OST-object is found now.\n",
2625                        lfsck_lfsck2name(lfsck), PFID(fid));
2626
2627         GOTO(stop, rc);
2628
2629 stop:
2630         dt_trans_stop(env, dev, th);
2631
2632 unlock:
2633         dt_write_unlock(env, obj);
2634         ldlm_lock_decref(&lh, LCK_EX);
2635
2636 put:
2637         lfsck_object_put(env, obj);
2638
2639         return rc;
2640 }
2641
2642 /**
2643  * Some OST-object has occupied the specified layout EA slot.
2644  * Such OST-object may be generated by the LFSCK when repair
2645  * dangling referenced MDT-object, which can be indicated by
2646  * attr::la_ctime == 0 but without S_ISUID in la_mode. If it
2647  * is true and such OST-object has not been modified yet, we
2648  * will replace it with the orphan OST-object; otherwise the
2649  * LFSCK will create new MDT-object to reference the orphan.
2650  *
2651  * \retval       +1: repaired
2652  * \retval        0: did nothing
2653  * \retval      -ve: on error
2654  */
2655 static int lfsck_layout_conflict_create(const struct lu_env *env,
2656                                         struct lfsck_component *com,
2657                                         struct lfsck_tgt_desc *ltd,
2658                                         struct lu_orphan_rec_v3 *rec,
2659                                         struct dt_object *parent,
2660                                         struct lu_fid *cfid,
2661                                         struct lu_buf *ea_buf,
2662                                         struct lov_mds_md_v1 *lmm,
2663                                         struct lov_ost_data_v1 *slot,
2664                                         __u32 ea_off, int lovea_size)
2665 {
2666         struct lfsck_thread_info *info          = lfsck_env_info(env);
2667         struct lu_fid            *cfid2         = &info->lti_fid2;
2668         struct ost_id            *oi            = &info->lti_oi;
2669         struct dt_device         *dev           = lfsck_obj2dev(parent);
2670         struct thandle           *th            = NULL;
2671         struct lustre_handle      lh            = { 0 };
2672         __u32                     ost_idx2      = le32_to_cpu(slot->l_ost_idx);
2673         int                       rc            = 0;
2674         ENTRY;
2675
2676         while (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val)) {
2677                 if (unlikely(!thread_is_running(&com->lc_lfsck->li_thread)))
2678                         RETURN(0);
2679         }
2680
2681         ostid_le_to_cpu(&slot->l_ost_oi, oi);
2682         rc = ostid_to_fid(cfid2, oi, ost_idx2);
2683         if (rc != 0)
2684                 GOTO(out, rc);
2685
2686         rc = lfsck_ibits_lock(env, com->lc_lfsck, parent, &lh,
2687                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
2688                               LCK_EX);
2689         if (rc != 0)
2690                 GOTO(out, rc);
2691
2692         rc = lfsck_layout_master_conditional_destroy(env, com, cfid2, ost_idx2);
2693
2694         /* If the conflict OST-obejct is not created for fixing dangling
2695          * referenced MDT-object in former LFSCK check/repair, or it has
2696          * been modified by others, then we cannot destroy it. Re-create
2697          * a new MDT-object for the orphan OST-object. */
2698         if (rc == -ETXTBSY) {
2699                 /* No need the layout lock on the original parent. */
2700                 lfsck_ibits_unlock(&lh, LCK_EX);
2701
2702                 fid_zero(&rec->lor_rec.lor_fid);
2703                 snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf),
2704                          "-"DFID"-%x", PFID(lu_object_fid(&parent->do_lu)),
2705                          ea_off);
2706                 rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
2707                                                 info->lti_tmpbuf, "C", ea_off);
2708
2709                 RETURN(rc);
2710         }
2711
2712         if (rc != 0 && rc != -ENOENT)
2713                 GOTO(unlock, rc);
2714
2715         th = dt_trans_create(env, dev);
2716         if (IS_ERR(th))
2717                 GOTO(unlock, rc = PTR_ERR(th));
2718
2719         rc = dt_declare_xattr_set(env, parent, ea_buf, XATTR_NAME_LOV,
2720                                   LU_XATTR_REPLACE, th);
2721         if (rc != 0)
2722                 GOTO(stop, rc);
2723
2724         rc = dt_trans_start_local(env, dev, th);
2725         if (rc != 0)
2726                 GOTO(stop, rc);
2727
2728         dt_write_lock(env, parent, 0);
2729         lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
2730         rc = lfsck_layout_refill_lovea(env, com->lc_lfsck, th, parent, cfid,
2731                                        ea_buf, lmm, slot, LU_XATTR_REPLACE,
2732                                        ltd->ltd_index, lovea_size);
2733         dt_write_unlock(env, parent);
2734
2735         GOTO(stop, rc);
2736
2737 stop:
2738         dt_trans_stop(env, dev, th);
2739
2740 unlock:
2741         lfsck_ibits_unlock(&lh, LCK_EX);
2742
2743 out:
2744         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant replaced the conflict "
2745                "OST-object "DFID" on the OST %x with the orphan "DFID" on "
2746                "the OST %x: parent "DFID", stripe-index %u: rc = %d\n",
2747                lfsck_lfsck2name(com->lc_lfsck), PFID(cfid2), ost_idx2,
2748                PFID(cfid), ltd->ltd_index, PFID(lfsck_dto2fid(parent)),
2749                ea_off, rc);
2750
2751         return rc >= 0 ? 1 : rc;
2752 }
2753
2754 /**
2755  * \retval       +1: repaired
2756  * \retval        0: did nothing
2757  * \retval      -ve: on error
2758  */
2759 static int lfsck_layout_recreate_lovea(const struct lu_env *env,
2760                                        struct lfsck_component *com,
2761                                        struct lfsck_tgt_desc *ltd,
2762                                        struct lu_orphan_rec_v3 *rec,
2763                                        struct dt_object *parent,
2764                                        struct lu_fid *cfid,
2765                                        __u32 ost_idx, __u32 ea_off)
2766 {
2767         struct lfsck_thread_info *info          = lfsck_env_info(env);
2768         struct lu_buf            *buf           = &info->lti_big_buf;
2769         struct lu_fid            *fid           = &info->lti_fid2;
2770         struct ost_id            *oi            = &info->lti_oi;
2771         struct lfsck_instance    *lfsck         = com->lc_lfsck;
2772         struct dt_device         *dt            = lfsck_obj2dev(parent);
2773         struct lfsck_bookmark    *bk            = &lfsck->li_bookmark_ram;
2774         struct ost_layout        *ol            = &rec->lor_layout;
2775         struct lov_comp_md_v1    *lcm           = NULL;
2776         struct lov_comp_md_entry_v1 *lcme       = NULL;
2777         struct thandle           *handle        = NULL;
2778         size_t                    lovea_size;
2779         struct lov_mds_md_v1     *lmm;
2780         struct lov_ost_data_v1   *objs;
2781         struct lustre_handle      lh            = { 0 };
2782         __u32                     magic;
2783         __u32 flags = 0;
2784         int                       fl            = 0;
2785         int                       rc            = 0;
2786         int                       rc1;
2787         int                       i;
2788         int pos = 0;
2789         __u16 count;
2790         bool locked = false;
2791         bool new_mirror = true;
2792         ENTRY;
2793
2794         rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
2795                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
2796                               LCK_EX);
2797         if (rc != 0) {
2798                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant failed to recreate "
2799                        "LOV EA for "DFID": parent "DFID", OST-index %u, "
2800                        "stripe-index %u, comp_id %u, comp_start %llu, "
2801                        "comp_end %llu, layout version %u, range %u: rc = %d\n",
2802                        lfsck_lfsck2name(lfsck), PFID(cfid),
2803                        PFID(lfsck_dto2fid(parent)), ost_idx, ea_off,
2804                        ol->ol_comp_id, ol->ol_comp_start,
2805                        ol->ol_comp_end, rec->lor_layout_version,
2806                        rec->lor_range, rc);
2807
2808                 RETURN(rc);
2809         }
2810
2811 again:
2812         if (locked) {
2813                 dt_write_unlock(env, parent);
2814                 locked = false;
2815         }
2816
2817         if (handle != NULL) {
2818                 dt_trans_stop(env, dt, handle);
2819                 handle = NULL;
2820         }
2821
2822         if (rc < 0)
2823                 GOTO(unlock_layout, rc);
2824
2825         lovea_size = rc;
2826         if (buf->lb_len < lovea_size) {
2827                 lu_buf_realloc(buf, lovea_size);
2828                 if (buf->lb_buf == NULL)
2829                         GOTO(unlock_layout, rc = -ENOMEM);
2830         }
2831
2832         if (!(bk->lb_param & LPF_DRYRUN)) {
2833                 handle = dt_trans_create(env, dt);
2834                 if (IS_ERR(handle))
2835                         GOTO(unlock_layout, rc = PTR_ERR(handle));
2836
2837                 rc = dt_declare_xattr_set(env, parent, buf, XATTR_NAME_LOV,
2838                                           fl, handle);
2839                 if (rc != 0)
2840                         GOTO(stop, rc);
2841
2842                 rc = dt_trans_start_local(env, dt, handle);
2843                 if (rc != 0)
2844                         GOTO(stop, rc);
2845         }
2846
2847         dt_write_lock(env, parent, 0);
2848         locked = true;
2849         rc = dt_xattr_get(env, parent, buf, XATTR_NAME_LOV);
2850         if (rc == -ERANGE) {
2851                 rc = dt_xattr_get(env, parent, &LU_BUF_NULL, XATTR_NAME_LOV);
2852                 LASSERT(rc != 0);
2853                 goto again;
2854         } else if (rc == -ENODATA || rc == 0) {
2855                 lovea_size = lfsck_lovea_size(ol, ea_off);
2856                 /* If the declared is not big enough, re-try. */
2857                 if (buf->lb_len < lovea_size) {
2858                         rc = lovea_size;
2859                         goto again;
2860                 }
2861                 fl = LU_XATTR_CREATE;
2862         } else if (rc < 0) {
2863                 GOTO(unlock_parent, rc);
2864         } else if (unlikely(buf->lb_len == 0)) {
2865                 goto again;
2866         } else {
2867                 fl = LU_XATTR_REPLACE;
2868                 lovea_size = rc;
2869         }
2870
2871         if (fl == LU_XATTR_CREATE) {
2872                 if (bk->lb_param & LPF_DRYRUN)
2873                         GOTO(unlock_parent, rc = 1);
2874
2875                 LASSERT(buf->lb_len >= lovea_size);
2876
2877                 rc = lfsck_layout_update_lovea(env, lfsck, handle, rec, parent,
2878                                                cfid, buf, fl, ost_idx, ea_off);
2879
2880                 GOTO(unlock_parent, rc);
2881         }
2882
2883         lmm = buf->lb_buf;
2884         rc1 = lfsck_layout_verify_header(parent, lmm);
2885
2886         /* If the LOV EA crashed, the rebuild it. */
2887         if (rc1 == -EINVAL) {
2888                 if (bk->lb_param & LPF_DRYRUN)
2889                         GOTO(unlock_parent, rc = 1);
2890
2891                 LASSERT(buf->lb_len >= lovea_size);
2892
2893                 rc = lfsck_layout_update_lovea(env, lfsck, handle, rec, parent,
2894                                                cfid, buf, fl, ost_idx, ea_off);
2895
2896                 GOTO(unlock_parent, rc);
2897         }
2898
2899         /* For other unknown magic/pattern, keep the current LOV EA. */
2900         if (rc1 == -EOPNOTSUPP)
2901                 GOTO(unlock_parent, rc1 = 0);
2902
2903         if (rc1)
2904                 GOTO(unlock_parent, rc = rc1);
2905
2906         magic = le32_to_cpu(lmm->lmm_magic);
2907         if (magic == LOV_MAGIC_COMP_V1) {
2908                 __u64 start;
2909                 __u64 end;
2910                 __u16 mirror_id0 = mirror_id_of(ol->ol_comp_id);
2911                 __u16 mirror_id1;
2912
2913                 if (bk->lb_param & LPF_DRYRUN)
2914                         GOTO(unlock_parent, rc = 1);
2915
2916                 lcm = buf->lb_buf;
2917                 count = le16_to_cpu(lcm->lcm_entry_count);
2918                 for (i = 0; i < count; pos = ++i) {
2919                         lcme = &lcm->lcm_entries[i];
2920                         start = le64_to_cpu(lcme->lcme_extent.e_start);
2921                         end = le64_to_cpu(lcme->lcme_extent.e_end);
2922                         mirror_id1 = mirror_id_of(le32_to_cpu(lcme->lcme_id));
2923
2924                         if (mirror_id0 > mirror_id1)
2925                                 continue;
2926
2927                         if (mirror_id0 < mirror_id1)
2928                                 break;
2929
2930                         new_mirror = false;
2931                         if (end <= ol->ol_comp_start)
2932                                 continue;
2933
2934                         if (start >= ol->ol_comp_end)
2935                                 break;
2936
2937                         lmm = buf->lb_buf + le32_to_cpu(lcme->lcme_offset);
2938                         magic = le32_to_cpu(lmm->lmm_magic);
2939                         flags = le32_to_cpu(lcme->lcme_flags);
2940                         goto further;
2941                 }
2942
2943                 rc = lfsck_layout_add_comp(env, lfsck, handle, rec, parent,
2944                                 cfid, buf, ost_idx, ea_off, pos, new_mirror);
2945
2946                 GOTO(unlock_parent, rc);
2947         }
2948
2949 further:
2950         count = le16_to_cpu(lmm->lmm_stripe_count);
2951         if (count == 0)
2952                 GOTO(unlock_parent, rc = -EINVAL);
2953         LASSERT(count > 0);
2954
2955         /* Exceed the current end of MDT-object layout EA. Then extend it. */
2956         if (count <= ea_off) {
2957                 if (bk->lb_param & LPF_DRYRUN)
2958                         GOTO(unlock_parent, rc = 1);
2959
2960                 lovea_size = lov_mds_md_size(ea_off + 1, magic);
2961                 /* If the declared is not big enough, re-try. */
2962                 if (buf->lb_len < lovea_size) {
2963                         rc = lovea_size;
2964                         goto again;
2965                 }
2966
2967                 if (lcm) {
2968                         LASSERT(lcme);
2969
2970                         lcme->lcme_flags = cpu_to_le32(flags | LCME_FL_INIT);
2971                         lfsck_layout_update_lcm(lcm, lcme,
2972                                                 rec->lor_layout_version,
2973                                                 rec->lor_range);
2974                 }
2975
2976                 rc = lfsck_layout_extend_v1v3_lovea(env, lfsck, handle, ol,
2977                                         parent, cfid, buf, ost_idx, ea_off);
2978
2979                 GOTO(unlock_parent, rc);
2980         }
2981
2982         LASSERTF(rc > 0, "invalid rc = %d\n", rc);
2983
2984         if (magic == LOV_MAGIC_V1) {
2985                 objs = &lmm->lmm_objects[0];
2986         } else {
2987                 LASSERT(magic == LOV_MAGIC_V3);
2988                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
2989         }
2990
2991         for (i = 0; i < count; i++, objs++) {
2992                 /* The MDT-object was created via lfsck_layout_recover_create()
2993                  * by others before, and we fill the dummy layout EA. */
2994                 if ((lcme && !(flags & LCME_FL_INIT)) ||
2995                      lovea_slot_is_dummy(objs)) {
2996                         if (i != ea_off)
2997                                 continue;
2998
2999                         if (bk->lb_param & LPF_DRYRUN)
3000                                 GOTO(unlock_parent, rc = 1);
3001
3002                         lmm->lmm_layout_gen =
3003                             cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
3004                         if (lcme) {
3005                                 LASSERT(lcm);
3006
3007                                 if (le32_to_cpu(lmm->lmm_stripe_size) !=
3008                                         ol->ol_stripe_size ||
3009                                     le16_to_cpu(lmm->lmm_stripe_count) !=
3010                                         ol->ol_stripe_count ||
3011                                     le64_to_cpu(lcme->lcme_extent.e_start) !=
3012                                         ol->ol_comp_start ||
3013                                     le64_to_cpu(lcme->lcme_extent.e_end) !=
3014                                         ol->ol_comp_end) {
3015                                         CDEBUG(D_LFSCK, "%s: found invalid "
3016                                         "component for "DFID ": parent "DFID
3017                                         ", stripe-index %u, stripe_size %u, "
3018                                         "stripe_count %u, comp_id %u, "
3019                                         "comp_start %llu, comp_end %llu, "
3020                                         "cur_stripe_size %u, "
3021                                         "cur_stripe_count %u, "
3022                                         "cur_comp_start %llu, "
3023                                         "cur_comp_end %llu\n",
3024                                         lfsck_lfsck2name(lfsck), PFID(cfid),
3025                                         PFID(lfsck_dto2fid(parent)), ea_off,
3026                                         ol->ol_stripe_size,
3027                                         ol->ol_stripe_count, ol->ol_comp_id,
3028                                         ol->ol_comp_start, ol->ol_comp_end,
3029                                         le32_to_cpu(lmm->lmm_stripe_size),
3030                                         le16_to_cpu(lmm->lmm_stripe_count),
3031                                         le64_to_cpu(lcme->lcme_extent.e_start),
3032                                         le64_to_cpu(lcme->lcme_extent.e_end));
3033
3034                                         GOTO(unlock_parent, rc = -EINVAL);
3035                                 }
3036
3037                                 lovea_size = le32_to_cpu(lcm->lcm_size);
3038                                 lcme->lcme_flags = cpu_to_le32(flags |
3039                                                                LCME_FL_INIT);
3040                                 lfsck_layout_update_lcm(lcm, lcme,
3041                                                         rec->lor_layout_version,
3042                                                         rec->lor_range);
3043                         }
3044
3045                         LASSERTF(buf->lb_len >= lovea_size,
3046                                  "buffer len %d is less than real size %d\n",
3047                                  (int)buf->lb_len, (int)lovea_size);
3048
3049                         rc = lfsck_layout_refill_lovea(env, lfsck, handle,
3050                                                 parent, cfid, buf, lmm, objs,
3051                                                 fl, ost_idx, lovea_size);
3052
3053                         CDEBUG(D_LFSCK, "%s layout LFSCK assistant fill "
3054                                "dummy layout slot for "DFID": parent "DFID
3055                                ", OST-index %u, stripe-index %u: rc = %d\n",
3056                                lfsck_lfsck2name(lfsck), PFID(cfid),
3057                                PFID(lfsck_dto2fid(parent)), ost_idx, i, rc);
3058
3059                         GOTO(unlock_parent, rc);
3060                 }
3061
3062                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
3063                 rc = ostid_to_fid(fid, oi, le32_to_cpu(objs->l_ost_idx));
3064                 if (rc != 0) {
3065                         CDEBUG(D_LFSCK, "%s: the parent "DFID" contains "
3066                                "invalid layout EA at the slot %d, index %u\n",
3067                                lfsck_lfsck2name(lfsck),
3068                                PFID(lfsck_dto2fid(parent)), i,
3069                                le32_to_cpu(objs->l_ost_idx));
3070
3071                         GOTO(unlock_parent, rc);
3072                 }
3073
3074                 /* It should be rare case, the slot is there, but the LFSCK
3075                  * does not handle it during the first-phase cycle scanning. */
3076                 if (unlikely(lu_fid_eq(fid, cfid))) {
3077                         if (i == ea_off) {
3078                                 GOTO(unlock_parent, rc = 0);
3079                         } else {
3080                                 /* Rare case that the OST-object index
3081                                  * does not match the parent MDT-object
3082                                  * layout EA. We trust the later one. */
3083                                 if (bk->lb_param & LPF_DRYRUN)
3084                                         GOTO(unlock_parent, rc = 1);
3085
3086                                 dt_write_unlock(env, parent);
3087                                 if (handle != NULL)
3088                                         dt_trans_stop(env, dt, handle);
3089                    &n