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, 2016, 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 <lustre/lustre_idl.h>
  40 #include <lu_object.h>
  41 #include <dt_object.h>
  42 #include <lustre_fid.h>
  43 #include <lustre_lib.h>
  44 #include <lustre_net.h>
  45 #include <lustre/lustre_user.h>
  46 #include <md_object.h>
  47 #include <obd_class.h>
  48
  49 #include "lfsck_internal.h"
  50
  51 #define LFSCK_LAYOUT_MAGIC_V1           0xB173AE14
  52 #define LFSCK_LAYOUT_MAGIC_V2           0xB1734D76
  53 #define LFSCK_LAYOUT_MAGIC_V3           0xB17371B9
  54
  55 #define LFSCK_LAYOUT_MAGIC              LFSCK_LAYOUT_MAGIC_V3
  56
  57 struct lfsck_layout_seq {
  58         struct list_head         lls_list;
  59         __u64                    lls_seq;
  60         __u64                    lls_lastid;
  61         __u64                    lls_lastid_known;
  62         struct dt_object        *lls_lastid_obj;
  63         unsigned int             lls_dirty:1;
  64 };
  65
  66 struct lfsck_layout_slave_target {
  67         /* link into lfsck_layout_slave_data::llsd_master_list. */
  68         struct list_head        llst_list;
  69         /* The position for next record in the rbtree for iteration. */
  70         struct lu_fid           llst_fid;
  71         /* Dummy hash for iteration against the rbtree. */
  72         __u64                   llst_hash;
  73         __u64                   llst_gen;
  74         atomic_t                llst_ref;
  75         __u32                   llst_index;
  76         /* How many times we have failed to get the master status. */
  77         int                     llst_failures;
  78 };
  79
  80 struct lfsck_layout_slave_data {
  81         /* list for lfsck_layout_seq */
  82         struct list_head         llsd_seq_list;
  83
  84         /* list for the masters involve layout verification. */
  85         struct list_head         llsd_master_list;
  86         spinlock_t               llsd_lock;
  87         __u64                    llsd_touch_gen;
  88         struct dt_object        *llsd_rb_obj;
  89         struct rb_root           llsd_rb_root;
  90         rwlock_t                 llsd_rb_lock;
  91         unsigned int             llsd_rbtree_valid:1;
  92 };
  93
  94 struct lfsck_layout_slave_async_args {
  95         struct obd_export                *llsaa_exp;
  96         struct lfsck_component           *llsaa_com;
  97         struct lfsck_layout_slave_target *llsaa_llst;
  98 };
  99
 100 static inline bool lfsck_comp_extent_aligned(__u64 size)
 101 {
 102          return (size & (LOV_MIN_STRIPE_SIZE - 1)) == 0;
 103 }
 104
 105 static inline void
 106 lfsck_layout_llst_put(struct lfsck_layout_slave_target *llst)
 107 {
 108         if (atomic_dec_and_test(&llst->llst_ref)) {
 109                 LASSERT(list_empty(&llst->llst_list));
 110
 111                 OBD_FREE_PTR(llst);
 112         }
 113 }
 114
 115 static inline int
 116 lfsck_layout_llst_add(struct lfsck_layout_slave_data *llsd, __u32 index)
 117 {
 118         struct lfsck_layout_slave_target *llst;
 119         struct lfsck_layout_slave_target *tmp;
 120         int                               rc   = 0;
 121
 122         OBD_ALLOC_PTR(llst);
 123         if (llst == NULL)
 124                 return -ENOMEM;
 125
 126         INIT_LIST_HEAD(&llst->llst_list);
 127         llst->llst_gen = 0;
 128         llst->llst_index = index;
 129         atomic_set(&llst->llst_ref, 1);
 130
 131         spin_lock(&llsd->llsd_lock);
 132         list_for_each_entry(tmp, &llsd->llsd_master_list, llst_list) {
 133                 if (tmp->llst_index == index) {
 134                         rc = -EALREADY;
 135                         break;
 136                 }
 137         }
 138         if (rc == 0)
 139                 list_add_tail(&llst->llst_list, &llsd->llsd_master_list);
 140         spin_unlock(&llsd->llsd_lock);
 141
 142         if (rc != 0)
 143                 OBD_FREE_PTR(llst);
 144
 145         return rc;
 146 }
 147
 148 static inline void
 149 lfsck_layout_llst_del(struct lfsck_layout_slave_data *llsd,
 150                       struct lfsck_layout_slave_target *llst)
 151 {
 152         bool del = false;
 153
 154         spin_lock(&llsd->llsd_lock);
 155         if (!list_empty(&llst->llst_list)) {
 156                 list_del_init(&llst->llst_list);
 157                 del = true;
 158         }
 159         spin_unlock(&llsd->llsd_lock);
 160
 161         if (del)
 162                 lfsck_layout_llst_put(llst);
 163 }
 164
 165 static inline struct lfsck_layout_slave_target *
 166 lfsck_layout_llst_find_and_del(struct lfsck_layout_slave_data *llsd,
 167                                __u32 index, bool unlink)
 168 {
 169         struct lfsck_layout_slave_target *llst;
 170
 171         spin_lock(&llsd->llsd_lock);
 172         list_for_each_entry(llst, &llsd->llsd_master_list, llst_list) {
 173                 if (llst->llst_index == index) {
 174                         if (unlink)
 175                                 list_del_init(&llst->llst_list);
 176                         else
 177                                 atomic_inc(&llst->llst_ref);
 178                         spin_unlock(&llsd->llsd_lock);
 179
 180                         return llst;
 181                 }
 182         }
 183         spin_unlock(&llsd->llsd_lock);
 184
 185         return NULL;
 186 }
 187
 188 static struct lfsck_layout_req *
 189 lfsck_layout_assistant_req_init(struct lfsck_assistant_object *lso,
 190                                 struct dt_object *child, __u32 comp_id,
 191                                 __u32 ost_idx, __u32 lov_idx)
 192 {
 193         struct lfsck_layout_req *llr;
 194
 195         OBD_ALLOC_PTR(llr);
 196         if (llr == NULL)
 197                 return ERR_PTR(-ENOMEM);
 198
 199         INIT_LIST_HEAD(&llr->llr_lar.lar_list);
 200         llr->llr_lar.lar_parent = lfsck_assistant_object_get(lso);
 201         llr->llr_child = child;
 202         llr->llr_comp_id = comp_id;
 203         llr->llr_ost_idx = ost_idx;
 204         llr->llr_lov_idx = lov_idx;
 205
 206         return llr;
 207 }
 208
 209 static void lfsck_layout_assistant_req_fini(const struct lu_env *env,
 210                                             struct lfsck_assistant_req *lar)
 211 {
 212         struct lfsck_layout_req *llr =
 213                         container_of0(lar, struct lfsck_layout_req, llr_lar);
 214
 215         lfsck_object_put(env, llr->llr_child);
 216         lfsck_assistant_object_put(env, lar->lar_parent);
 217         OBD_FREE_PTR(llr);
 218 }
 219
 220 static int
 221 lfsck_layout_assistant_sync_failures_interpret(const struct lu_env *env,
 222                                                struct ptlrpc_request *req,
 223                                                void *args, int rc)
 224 {
 225         if (rc == 0) {
 226                 struct lfsck_async_interpret_args *laia = args;
 227                 struct lfsck_tgt_desc             *ltd  = laia->laia_ltd;
 228
 229                 ltd->ltd_synced_failures = 1;
 230                 atomic_dec(laia->laia_count);
 231         }
 232
 233         return 0;
 234 }
 235
 236 /**
 237  * Notify remote LFSCK instances about former failures.
 238  *
 239  * The local LFSCK instance has recorded which OSTs have ever failed to respond
 240  * some LFSCK verification requests (maybe because of network issues or the OST
 241  * itself trouble). During the respond gap, the OST may missed some OST-objects
 242  * verification, then the OST cannot know whether related OST-objects have been
 243  * referenced by related MDT-objects or not, then in the second-stage scanning,
 244  * these OST-objects will be regarded as orphan, if the OST-object contains bad
 245  * parent FID for back reference, then it will misguide the LFSCK to make wrong
 246  * fixing for the fake orphan.
 247  *
 248  * To avoid above trouble, when layout LFSCK finishes the first-stage scanning,
 249  * it will scan the bitmap for the ever failed OSTs, and notify them that they
 250  * have ever missed some OST-object verification and should skip the handling
 251  * for orphan OST-objects on all MDTs that are in the layout LFSCK.
 252  *
 253  * \param[in] env       pointer to the thread context
 254  * \param[in] com       pointer to the lfsck component
 255  * \param[in] lr        pointer to the lfsck request
 256  */
 257 static void lfsck_layout_assistant_sync_failures(const struct lu_env *env,
 258                                                  struct lfsck_component *com,
 259                                                  struct lfsck_request *lr)
 260 {
 261         struct lfsck_async_interpret_args *laia  =
 262                                 &lfsck_env_info(env)->lti_laia2;
 263         struct lfsck_assistant_data       *lad   = com->lc_data;
 264         struct lfsck_layout               *lo    = com->lc_file_ram;
 265         struct lfsck_instance             *lfsck = com->lc_lfsck;
 266         struct lfsck_tgt_descs            *ltds  = &lfsck->li_ost_descs;
 267         struct lfsck_tgt_desc             *ltd;
 268         struct ptlrpc_request_set         *set;
 269         atomic_t                           count;
 270         __u32                              idx;
 271         int                                rc    = 0;
 272         ENTRY;
 273
 274         if (!lad->lad_incomplete)
 275                 RETURN_EXIT;
 276
 277         /* If the MDT has ever failed to verfiy some OST-objects,
 278          * then sync failures with them firstly. */
 279         lr->lr_flags2 = lo->ll_flags | LF_INCOMPLETE;
 280
 281         atomic_set(&count, 0);
 282         memset(laia, 0, sizeof(*laia));
 283         laia->laia_count = &count;
 284         set = ptlrpc_prep_set();
 285         if (set == NULL)
 286                 GOTO(out, rc = -ENOMEM);
 287
 288         down_read(&ltds->ltd_rw_sem);
 289         cfs_foreach_bit(lad->lad_bitmap, idx) {
 290                 ltd = lfsck_ltd2tgt(ltds, idx);
 291                 LASSERT(ltd != NULL);
 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(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 {
 333         __u32 magic;
 334         __u32 pattern;
 335
 336         magic = le32_to_cpu(lmm->lmm_magic);
 337         /* If magic crashed, keep it there. Sometime later, during OST-object
 338          * orphan handling, if some OST-object(s) back-point to it, it can be
 339          * verified and repaired. */
 340         if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3) {
 341                 int rc;
 342
 343                 if ((magic & LOV_MAGIC_MASK) == LOV_MAGIC_MAGIC)
 344                         rc = -EOPNOTSUPP;
 345                 else
 346                         rc = -EINVAL;
 347
 348                 CDEBUG(D_LFSCK, "%s LOV EA magic %u for the file "DFID"\n",
 349                        rc == -EINVAL ? "Unknown" : "Unsupported",
 350                        magic, PFID(lfsck_dto2fid(obj)));
 351
 352                 return rc;
 353         }
 354
 355         pattern = le32_to_cpu(lmm->lmm_pattern);
 356         /* XXX: currently, we only support LOV_PATTERN_RAID0. */
 357         if (lov_pattern(pattern) != LOV_PATTERN_RAID0) {
 358                 CDEBUG(D_LFSCK, "Unsupported LOV EA pattern %u for the file "
 359                        DFID"\n", pattern, PFID(lfsck_dto2fid(obj)));
 360
 361                 return -EOPNOTSUPP;
 362         }
 363
 364         return 0;
 365 }
 366
 367 static int lfsck_layout_verify_header(struct dt_object *obj,
 368                                       struct lov_mds_md_v1 *lmm)
 369 {
 370         int rc = 0;
 371
 372         if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1) {
 373                 struct lov_comp_md_v1 *lcm = (struct lov_comp_md_v1 *)lmm;
 374                 int i;
 375                 __u16 count = le16_to_cpu(lcm->lcm_entry_count);
 376
 377                 if (unlikely(count == 0)) {
 378                         CDEBUG(D_LFSCK, "the PFL file "DFID" contains invalid "
 379                                "components count 0\n",
 380                                PFID(lfsck_dto2fid(obj)));
 381
 382                         return -EINVAL;
 383                 }
 384
 385                 for (i = 0; i < count; i++) {
 386                         struct lov_comp_md_entry_v1 *lcme =
 387                                                 &lcm->lcm_entries[i];
 388                         __u64 start = le64_to_cpu(lcme->lcme_extent.e_start);
 389                         __u64 end = le64_to_cpu(lcme->lcme_extent.e_end);
 390                         __u32 comp_id = le32_to_cpu(lcme->lcme_id);
 391
 392                         if (unlikely(comp_id == LCME_ID_INVAL ||
 393                                      comp_id > LCME_ID_MAX)) {
 394                                 CDEBUG(D_LFSCK, "found invalid FPL ID %u "
 395                                        "for the file "DFID" at idx %d\n",
 396                                        comp_id, PFID(lfsck_dto2fid(obj)), i);
 397
 398                                 return -EINVAL;
 399                         }
 400
 401                         if (unlikely(start >= end ||
 402                                      !lfsck_comp_extent_aligned(start) ||
 403                                      (!lfsck_comp_extent_aligned(end) &&
 404                                       end != LUSTRE_EOF))) {
 405                                 CDEBUG(D_LFSCK, "found invalid FPL extent "
 406                                        "range [%llu - %llu) for the file "
 407                                        DFID" at idx %d\n",
 408                                        start, end, PFID(lfsck_dto2fid(obj)), i);
 409
 410                                 return -EINVAL;
 411                         }
 412
 413                         rc = lfsck_layout_verify_header_v1v3(obj,
 414                                 (struct lov_mds_md_v1 *)((char *)lmm +
 415                                 le32_to_cpu(lcme->lcme_offset)));
 416                         if (rc)
 417                                 return rc;
 418                 }
 419         } else {
 420                 rc = lfsck_layout_verify_header_v1v3(obj, lmm);
 421         }
 422
 423         return rc;
 424 }
 425
 426 static int lfsck_layout_get_lovea(const struct lu_env *env,
 427                                   struct dt_object *obj, struct lu_buf *buf)
 428 {
 429         int rc;
 430         int rc1;
 431
 432 again:
 433         rc = dt_xattr_get(env, obj, buf, XATTR_NAME_LOV);
 434         if (rc == -ERANGE) {
 435                 rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_LOV);
 436                 if (rc <= 0)
 437                         return rc;
 438
 439                 lu_buf_realloc(buf, rc);
 440                 if (buf->lb_buf == NULL)
 441                         return -ENOMEM;
 442
 443                 goto again;
 444         }
 445
 446         if (rc == -ENODATA)
 447                 rc = 0;
 448
 449         if (rc <= 0)
 450                 return rc;
 451
 452         if (unlikely(buf->lb_buf == NULL)) {
 453                 lu_buf_alloc(buf, rc);
 454                 if (buf->lb_buf == NULL)
 455                         return -ENOMEM;
 456
 457                 goto again;
 458         }
 459
 460         rc1 = lfsck_layout_verify_header(obj, buf->lb_buf);
 461
 462         return rc1 ? rc1 : rc;
 463 }
 464
 465 #define LFSCK_RBTREE_BITMAP_SIZE        PAGE_SIZE
 466 #define LFSCK_RBTREE_BITMAP_WIDTH       (LFSCK_RBTREE_BITMAP_SIZE << 3)
 467 #define LFSCK_RBTREE_BITMAP_MASK        (LFSCK_RBTREE_BITMAP_WIDTH - 1)
 468
 469 struct lfsck_rbtree_node {
 470         struct rb_node   lrn_node;
 471         __u64            lrn_seq;
 472         __u32            lrn_first_oid;
 473         atomic_t         lrn_known_count;
 474         atomic_t         lrn_accessed_count;
 475         void            *lrn_known_bitmap;
 476         void            *lrn_accessed_bitmap;
 477 };
 478
 479 static inline int lfsck_rbtree_cmp(struct lfsck_rbtree_node *lrn,
 480                                    __u64 seq, __u32 oid)
 481 {
 482         if (seq < lrn->lrn_seq)
 483                 return -1;
 484
 485         if (seq > lrn->lrn_seq)
 486                 return 1;
 487
 488         if (oid < lrn->lrn_first_oid)
 489                 return -1;
 490
 491         if (oid - lrn->lrn_first_oid >= LFSCK_RBTREE_BITMAP_WIDTH)
 492                 return 1;
 493
 494         return 0;
 495 }
 496
 497 /* The caller should hold llsd->llsd_rb_lock. */
 498 static struct lfsck_rbtree_node *
 499 lfsck_rbtree_search(struct lfsck_layout_slave_data *llsd,
 500                     const struct lu_fid *fid, bool *exact)
 501 {
 502         struct rb_node           *node  = llsd->llsd_rb_root.rb_node;
 503         struct rb_node           *prev  = NULL;
 504         struct lfsck_rbtree_node *lrn   = NULL;
 505         int                       rc    = 0;
 506
 507         if (exact != NULL)
 508                 *exact = true;
 509
 510         while (node != NULL) {
 511                 prev = node;
 512                 lrn = rb_entry(node, struct lfsck_rbtree_node, lrn_node);
 513                 rc = lfsck_rbtree_cmp(lrn, fid_seq(fid), fid_oid(fid));
 514                 if (rc < 0)
 515                         node = node->rb_left;
 516                 else if (rc > 0)
 517                         node = node->rb_right;
 518                 else
 519                         return lrn;
 520         }
 521
 522         if (exact == NULL)
 523                 return NULL;
 524
 525         /* If there is no exactly matched one, then to the next valid one. */
 526         *exact = false;
 527
 528         /* The rbtree is empty. */
 529         if (rc == 0)
 530                 return NULL;
 531
 532         if (rc < 0)
 533                 return lrn;
 534
 535         node = rb_next(prev);
 536
 537         /* The end of the rbtree. */
 538         if (node == NULL)
 539                 return NULL;
 540
 541         lrn = rb_entry(node, struct lfsck_rbtree_node, lrn_node);
 542
 543         return lrn;
 544 }
 545
 546 static struct lfsck_rbtree_node *lfsck_rbtree_new(const struct lu_env *env,
 547                                                   const struct lu_fid *fid)
 548 {
 549         struct lfsck_rbtree_node *lrn;
 550
 551         OBD_ALLOC_PTR(lrn);
 552         if (lrn == NULL)
 553                 return ERR_PTR(-ENOMEM);
 554
 555         OBD_ALLOC(lrn->lrn_known_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 556         if (lrn->lrn_known_bitmap == NULL) {
 557                 OBD_FREE_PTR(lrn);
 558
 559                 return ERR_PTR(-ENOMEM);
 560         }
 561
 562         OBD_ALLOC(lrn->lrn_accessed_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 563         if (lrn->lrn_accessed_bitmap == NULL) {
 564                 OBD_FREE(lrn->lrn_known_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 565                 OBD_FREE_PTR(lrn);
 566
 567                 return ERR_PTR(-ENOMEM);
 568         }
 569
 570         RB_CLEAR_NODE(&lrn->lrn_node);
 571         lrn->lrn_seq = fid_seq(fid);
 572         lrn->lrn_first_oid = fid_oid(fid) & ~LFSCK_RBTREE_BITMAP_MASK;
 573         atomic_set(&lrn->lrn_known_count, 0);
 574         atomic_set(&lrn->lrn_accessed_count, 0);
 575
 576         return lrn;
 577 }
 578
 579 static void lfsck_rbtree_free(struct lfsck_rbtree_node *lrn)
 580 {
 581         OBD_FREE(lrn->lrn_accessed_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 582         OBD_FREE(lrn->lrn_known_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 583         OBD_FREE_PTR(lrn);
 584 }
 585
 586 /* The caller should hold lock. */
 587 static struct lfsck_rbtree_node *
 588 lfsck_rbtree_insert(struct lfsck_layout_slave_data *llsd,
 589                     struct lfsck_rbtree_node *lrn)
 590 {
 591         struct rb_node           **pos    = &llsd->llsd_rb_root.rb_node;
 592         struct rb_node            *parent = NULL;
 593         struct lfsck_rbtree_node  *tmp;
 594         int                        rc;
 595
 596         while (*pos != NULL) {
 597                 parent = *pos;
 598                 tmp = rb_entry(parent, struct lfsck_rbtree_node, lrn_node);
 599                 rc = lfsck_rbtree_cmp(tmp, lrn->lrn_seq, lrn->lrn_first_oid);
 600                 if (rc < 0)
 601                         pos = &(*pos)->rb_left;
 602                 else if (rc > 0)
 603                         pos = &(*pos)->rb_right;
 604                 else
 605                         return tmp;
 606         }
 607
 608         rb_link_node(&lrn->lrn_node, parent, pos);
 609         rb_insert_color(&lrn->lrn_node, &llsd->llsd_rb_root);
 610
 611         return lrn;
 612 }
 613
 614 extern const struct dt_index_operations lfsck_orphan_index_ops;
 615
 616 static int lfsck_rbtree_setup(const struct lu_env *env,
 617                               struct lfsck_component *com)
 618 {
 619         struct lu_fid                   *fid    = &lfsck_env_info(env)->lti_fid;
 620         struct lfsck_instance           *lfsck  = com->lc_lfsck;
 621         struct dt_device                *dev    = lfsck->li_bottom;
 622         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
 623         struct dt_object                *obj;
 624
 625         fid->f_seq = FID_SEQ_LAYOUT_RBTREE;
 626         fid->f_oid = lfsck_dev_idx(lfsck);
 627         fid->f_ver = 0;
 628         obj = dt_locate(env, dev, fid);
 629         if (IS_ERR(obj))
 630                 RETURN(PTR_ERR(obj));
 631
 632         /* Generate an in-RAM object to stand for the layout rbtree.
 633          * Scanning the layout rbtree will be via the iteration over
 634          * the object. In the future, the rbtree may be written onto
 635          * disk with the object.
 636          *
 637          * Mark the object to be as exist. */
 638         obj->do_lu.lo_header->loh_attr |= LOHA_EXISTS;
 639         obj->do_index_ops = &lfsck_orphan_index_ops;
 640         llsd->llsd_rb_obj = obj;
 641         llsd->llsd_rbtree_valid = 1;
 642         dev->dd_record_fid_accessed = 1;
 643
 644         CDEBUG(D_LFSCK, "%s: layout LFSCK init OST-objects accessing bitmap\n",
 645                lfsck_lfsck2name(lfsck));
 646
 647         return 0;
 648 }
 649
 650 static void lfsck_rbtree_cleanup(const struct lu_env *env,
 651                                  struct lfsck_component *com)
 652 {
 653         struct lfsck_instance           *lfsck = com->lc_lfsck;
 654         struct lfsck_layout_slave_data  *llsd  = com->lc_data;
 655         struct rb_node                  *node  = rb_first(&llsd->llsd_rb_root);
 656         struct rb_node                  *next;
 657         struct lfsck_rbtree_node        *lrn;
 658
 659         lfsck->li_bottom->dd_record_fid_accessed = 0;
 660         /* Invalid the rbtree, then no others will use it. */
 661         write_lock(&llsd->llsd_rb_lock);
 662         llsd->llsd_rbtree_valid = 0;
 663         write_unlock(&llsd->llsd_rb_lock);
 664
 665         while (node != NULL) {
 666                 next = rb_next(node);
 667                 lrn = rb_entry(node, struct lfsck_rbtree_node, lrn_node);
 668                 rb_erase(node, &llsd->llsd_rb_root);
 669                 lfsck_rbtree_free(lrn);
 670                 node = next;
 671         }
 672
 673         if (llsd->llsd_rb_obj != NULL) {
 674                 lfsck_object_put(env, llsd->llsd_rb_obj);
 675                 llsd->llsd_rb_obj = NULL;
 676         }
 677
 678         CDEBUG(D_LFSCK, "%s: layout LFSCK fini OST-objects accessing bitmap\n",
 679                lfsck_lfsck2name(lfsck));
 680 }
 681
 682 static void lfsck_rbtree_update_bitmap(const struct lu_env *env,
 683                                        struct lfsck_component *com,
 684                                        const struct lu_fid *fid,
 685                                        bool accessed)
 686 {
 687         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
 688         struct lfsck_rbtree_node        *lrn;
 689         bool                             insert = false;
 690         int                              idx;
 691         int                              rc     = 0;
 692         ENTRY;
 693
 694         if (unlikely(!fid_is_sane(fid) || fid_is_last_id(fid)))
 695                 RETURN_EXIT;
 696
 697         if (!fid_is_idif(fid) && !fid_is_norm(fid))
 698                 RETURN_EXIT;
 699
 700         read_lock(&llsd->llsd_rb_lock);
 701         if (!llsd->llsd_rbtree_valid)
 702                 GOTO(unlock, rc = 0);
 703
 704         lrn = lfsck_rbtree_search(llsd, fid, NULL);
 705         if (lrn == NULL) {
 706                 struct lfsck_rbtree_node *tmp;
 707
 708                 LASSERT(!insert);
 709
 710                 read_unlock(&llsd->llsd_rb_lock);
 711                 tmp = lfsck_rbtree_new(env, fid);
 712                 if (IS_ERR(tmp))
 713                         GOTO(out, rc = PTR_ERR(tmp));
 714
 715                 insert = true;
 716                 write_lock(&llsd->llsd_rb_lock);
 717                 if (!llsd->llsd_rbtree_valid) {
 718                         lfsck_rbtree_free(tmp);
 719                         GOTO(unlock, rc = 0);
 720                 }
 721
 722                 lrn = lfsck_rbtree_insert(llsd, tmp);
 723                 if (lrn != tmp)
 724                         lfsck_rbtree_free(tmp);
 725         }
 726
 727         idx = fid_oid(fid) & LFSCK_RBTREE_BITMAP_MASK;
 728         /* Any accessed object must be a known object. */
 729         if (!test_and_set_bit(idx, lrn->lrn_known_bitmap))
 730                 atomic_inc(&lrn->lrn_known_count);
 731         if (accessed && !test_and_set_bit(idx, lrn->lrn_accessed_bitmap))
 732                 atomic_inc(&lrn->lrn_accessed_count);
 733
 734         GOTO(unlock, rc = 0);
 735
 736 unlock:
 737         if (insert)
 738                 write_unlock(&llsd->llsd_rb_lock);
 739         else
 740                 read_unlock(&llsd->llsd_rb_lock);
 741 out:
 742         if (rc != 0 && accessed) {
 743                 struct lfsck_layout *lo = com->lc_file_ram;
 744
 745                 CDEBUG(D_LFSCK, "%s: fail to update OST-objects accessing "
 746                        "bitmap, and will cause incorrect LFSCK OST-object "
 747                        "handling, so disable it to cancel orphan handling "
 748                        "for related device. rc = %d\n",
 749                        lfsck_lfsck2name(com->lc_lfsck), rc);
 750
 751                 lo->ll_flags |= LF_INCOMPLETE;
 752                 lfsck_rbtree_cleanup(env, com);
 753         }
 754 }
 755
 756 static inline void lldk_le_to_cpu(struct lfsck_layout_dangling_key *des,
 757                                   const struct lfsck_layout_dangling_key *src)
 758 {
 759         fid_le_to_cpu(&des->lldk_fid, &src->lldk_fid);
 760         des->lldk_comp_id = le32_to_cpu(src->lldk_comp_id);
 761         des->lldk_ea_off = le32_to_cpu(src->lldk_ea_off);
 762 }
 763
 764 static inline void lldk_cpu_to_le(struct lfsck_layout_dangling_key *des,
 765                                   const struct lfsck_layout_dangling_key *src)
 766 {
 767         fid_cpu_to_le(&des->lldk_fid, &src->lldk_fid);
 768         des->lldk_comp_id = cpu_to_le32(src->lldk_comp_id);
 769         des->lldk_ea_off = cpu_to_le32(src->lldk_ea_off);
 770 }
 771
 772 static inline void lldk_be_to_cpu(struct lfsck_layout_dangling_key *des,
 773                                   const struct lfsck_layout_dangling_key *src)
 774 {
 775         fid_be_to_cpu(&des->lldk_fid, &src->lldk_fid);
 776         des->lldk_comp_id = be32_to_cpu(src->lldk_comp_id);
 777         des->lldk_ea_off = be32_to_cpu(src->lldk_ea_off);
 778 }
 779
 780 static inline void lldk_cpu_to_be(struct lfsck_layout_dangling_key *des,
 781                                   const struct lfsck_layout_dangling_key *src)
 782 {
 783         fid_cpu_to_be(&des->lldk_fid, &src->lldk_fid);
 784         des->lldk_comp_id = cpu_to_be32(src->lldk_comp_id);
 785         des->lldk_ea_off = cpu_to_be32(src->lldk_ea_off);
 786 }
 787
 788 static void lfsck_layout_le_to_cpu(struct lfsck_layout *des,
 789                                    const struct lfsck_layout *src)
 790 {
 791         int i;
 792
 793         des->ll_magic = le32_to_cpu(src->ll_magic);
 794         des->ll_status = le32_to_cpu(src->ll_status);
 795         des->ll_flags = le32_to_cpu(src->ll_flags);
 796         des->ll_success_count = le32_to_cpu(src->ll_success_count);
 797         des->ll_run_time_phase1 = le32_to_cpu(src->ll_run_time_phase1);
 798         des->ll_run_time_phase2 = le32_to_cpu(src->ll_run_time_phase2);
 799         des->ll_time_last_complete = le64_to_cpu(src->ll_time_last_complete);
 800         des->ll_time_latest_start = le64_to_cpu(src->ll_time_latest_start);
 801         des->ll_time_last_checkpoint =
 802                                 le64_to_cpu(src->ll_time_last_checkpoint);
 803         des->ll_pos_latest_start = le64_to_cpu(src->ll_pos_latest_start);
 804         des->ll_pos_last_checkpoint = le64_to_cpu(src->ll_pos_last_checkpoint);
 805         des->ll_pos_first_inconsistent =
 806                         le64_to_cpu(src->ll_pos_first_inconsistent);
 807         des->ll_objs_checked_phase1 = le64_to_cpu(src->ll_objs_checked_phase1);
 808         des->ll_objs_failed_phase1 = le64_to_cpu(src->ll_objs_failed_phase1);
 809         des->ll_objs_checked_phase2 = le64_to_cpu(src->ll_objs_checked_phase2);
 810         des->ll_objs_failed_phase2 = le64_to_cpu(src->ll_objs_failed_phase2);
 811         for (i = 0; i < LLIT_MAX; i++)
 812                 des->ll_objs_repaired[i] =
 813                                 le64_to_cpu(src->ll_objs_repaired[i]);
 814         des->ll_objs_skipped = le64_to_cpu(src->ll_objs_skipped);
 815         des->ll_bitmap_size = le32_to_cpu(src->ll_bitmap_size);
 816         lldk_le_to_cpu(&des->ll_lldk_latest_scanned_phase2,
 817                        &src->ll_lldk_latest_scanned_phase2);
 818 }
 819
 820 static void lfsck_layout_cpu_to_le(struct lfsck_layout *des,
 821                                    const struct lfsck_layout *src)
 822 {
 823         int i;
 824
 825         des->ll_magic = cpu_to_le32(src->ll_magic);
 826         des->ll_status = cpu_to_le32(src->ll_status);
 827         des->ll_flags = cpu_to_le32(src->ll_flags);
 828         des->ll_success_count = cpu_to_le32(src->ll_success_count);
 829         des->ll_run_time_phase1 = cpu_to_le32(src->ll_run_time_phase1);
 830         des->ll_run_time_phase2 = cpu_to_le32(src->ll_run_time_phase2);
 831         des->ll_time_last_complete = cpu_to_le64(src->ll_time_last_complete);
 832         des->ll_time_latest_start = cpu_to_le64(src->ll_time_latest_start);
 833         des->ll_time_last_checkpoint =
 834                                 cpu_to_le64(src->ll_time_last_checkpoint);
 835         des->ll_pos_latest_start = cpu_to_le64(src->ll_pos_latest_start);
 836         des->ll_pos_last_checkpoint = cpu_to_le64(src->ll_pos_last_checkpoint);
 837         des->ll_pos_first_inconsistent =
 838                         cpu_to_le64(src->ll_pos_first_inconsistent);
 839         des->ll_objs_checked_phase1 = cpu_to_le64(src->ll_objs_checked_phase1);
 840         des->ll_objs_failed_phase1 = cpu_to_le64(src->ll_objs_failed_phase1);
 841         des->ll_objs_checked_phase2 = cpu_to_le64(src->ll_objs_checked_phase2);
 842         des->ll_objs_failed_phase2 = cpu_to_le64(src->ll_objs_failed_phase2);
 843         for (i = 0; i < LLIT_MAX; i++)
 844                 des->ll_objs_repaired[i] =
 845                                 cpu_to_le64(src->ll_objs_repaired[i]);
 846         des->ll_objs_skipped = cpu_to_le64(src->ll_objs_skipped);
 847         des->ll_bitmap_size = cpu_to_le32(src->ll_bitmap_size);
 848         lldk_cpu_to_le(&des->ll_lldk_latest_scanned_phase2,
 849                        &src->ll_lldk_latest_scanned_phase2);
 850 }
 851
 852 /**
 853  * Load the OST bitmap from the lfsck_layout trace file.
 854  *
 855  * \param[in] env       pointer to the thread context
 856  * \param[in] com       pointer to the lfsck component
 857  *
 858  * \retval              0 for success
 859  * \retval              negative error number on failure or data corruption
 860  */
 861 static int lfsck_layout_load_bitmap(const struct lu_env *env,
 862                                     struct lfsck_component *com)
 863 {
 864         struct dt_object                *obj    = com->lc_obj;
 865         struct lfsck_assistant_data     *lad    = com->lc_data;
 866         struct lfsck_layout             *lo     = com->lc_file_ram;
 867         struct cfs_bitmap                       *bitmap = lad->lad_bitmap;
 868         loff_t                           pos    = com->lc_file_size;
 869         ssize_t                          size;
 870         __u32                            nbits;
 871         int                              rc;
 872         ENTRY;
 873
 874         if (com->lc_lfsck->li_ost_descs.ltd_tgts_bitmap->size >
 875             lo->ll_bitmap_size)
 876                 nbits = com->lc_lfsck->li_ost_descs.ltd_tgts_bitmap->size;
 877         else
 878                 nbits = lo->ll_bitmap_size;
 879
 880         if (unlikely(nbits < BITS_PER_LONG))
 881                 nbits = BITS_PER_LONG;
 882
 883         if (nbits > bitmap->size) {
 884                 __u32 new_bits = bitmap->size;
 885                 struct cfs_bitmap *new_bitmap;
 886
 887                 while (new_bits < nbits)
 888                         new_bits <<= 1;
 889
 890                 new_bitmap = CFS_ALLOCATE_BITMAP(new_bits);
 891                 if (new_bitmap == NULL)
 892                         RETURN(-ENOMEM);
 893
 894                 lad->lad_bitmap = new_bitmap;
 895                 CFS_FREE_BITMAP(bitmap);
 896                 bitmap = new_bitmap;
 897         }
 898
 899         if (lo->ll_bitmap_size == 0) {
 900                 lad->lad_incomplete = 0;
 901                 CFS_RESET_BITMAP(bitmap);
 902
 903                 RETURN(0);
 904         }
 905
 906         size = (lo->ll_bitmap_size + 7) >> 3;
 907         rc = dt_read(env, obj, lfsck_buf_get(env, bitmap->data, size), &pos);
 908         if (rc != size)
 909                 RETURN(rc >= 0 ? -EINVAL : rc);
 910
 911         if (cfs_bitmap_check_empty(bitmap))
 912                 lad->lad_incomplete = 0;
 913         else
 914                 lad->lad_incomplete = 1;
 915
 916         RETURN(0);
 917 }
 918
 919 /**
 920  * Load the layout LFSCK trace file from disk.
 921  *
 922  * The layout LFSCK trace file records the layout LFSCK status information
 923  * and other statistics, such as how many objects have been scanned, and how
 924  * many objects have been repaired, and etc. It also contains the bitmap for
 925  * failed OSTs during the layout LFSCK. All these information will be loaded
 926  * from disk to RAM when the layout LFSCK component setup.
 927  *
 928  * \param[in] env       pointer to the thread context
 929  * \param[in] com       pointer to the lfsck component
 930  *
 931  * \retval              positive number for file data corruption, the caller
 932  *                      should reset the layout LFSCK trace file
 933  * \retval              0 for success
 934  * \retval              negative error number on failure
 935  */
 936 static int lfsck_layout_load(const struct lu_env *env,
 937                              struct lfsck_component *com)
 938 {
 939         struct lfsck_layout             *lo     = com->lc_file_ram;
 940         ssize_t                          size   = com->lc_file_size;
 941         loff_t                           pos    = 0;
 942         int                              rc;
 943
 944         rc = dt_read(env, com->lc_obj,
 945                      lfsck_buf_get(env, com->lc_file_disk, size), &pos);
 946         if (rc == 0) {
 947                 return -ENOENT;
 948         } else if (rc < 0) {
 949                 CDEBUG(D_LFSCK, "%s: failed to load lfsck_layout: rc = %d\n",
 950                        lfsck_lfsck2name(com->lc_lfsck), rc);
 951                 return rc;
 952         } else if (rc != size) {
 953                 CDEBUG(D_LFSCK, "%s: lfsck_layout size %u != %u; reset it\n",
 954                        lfsck_lfsck2name(com->lc_lfsck), rc, (unsigned int)size);
 955                 return 1;
 956         }
 957
 958         lfsck_layout_le_to_cpu(lo, com->lc_file_disk);
 959         if (lo->ll_magic != LFSCK_LAYOUT_MAGIC) {
 960                 CDEBUG(D_LFSCK, "%s: invalid lfsck_layout magic %#x != %#x, "
 961                        "to be reset\n", lfsck_lfsck2name(com->lc_lfsck),
 962                        lo->ll_magic, LFSCK_LAYOUT_MAGIC);
 963                 return 1;
 964         }
 965
 966         return 0;
 967 }
 968
 969 /**
 970  * Store the layout LFSCK trace file on disk.
 971  *
 972  * The layout LFSCK trace file records the layout LFSCK status information
 973  * and other statistics, such as how many objects have been scanned, and how
 974  * many objects have been repaired, and etc. It also contains the bitmap for
 975  * failed OSTs during the layout LFSCK. All these information will be synced
 976  * from RAM to disk periodically.
 977  *
 978  * \param[in] env       pointer to the thread context
 979  * \param[in] com       pointer to the lfsck component
 980  *
 981  * \retval              0 for success
 982  * \retval              negative error number on failure
 983  */
 984 static int lfsck_layout_store(const struct lu_env *env,
 985                               struct lfsck_component *com)
 986 {
 987         struct dt_object        *obj    = com->lc_obj;
 988         struct lfsck_instance   *lfsck  = com->lc_lfsck;
 989         struct lfsck_layout     *lo_ram = com->lc_file_ram;
 990         struct lfsck_layout     *lo     = com->lc_file_disk;
 991         struct thandle          *th;
 992         struct dt_device        *dev    = lfsck_obj2dev(obj);
 993         struct cfs_bitmap       *bitmap = NULL;
 994         loff_t                   pos;
 995         ssize_t                  size   = com->lc_file_size;
 996         __u32                    nbits  = 0;
 997         int                      rc;
 998         ENTRY;
 999
1000         if (lfsck->li_master) {
1001                 struct lfsck_assistant_data *lad = com->lc_data;
1002
1003                 bitmap = lad->lad_bitmap;
1004                 nbits = bitmap->size;
1005
1006                 LASSERT(nbits > 0);
1007                 LASSERTF((nbits & 7) == 0, "Invalid nbits %u\n", nbits);
1008         }
1009
1010         lo_ram->ll_bitmap_size = nbits;
1011         lfsck_layout_cpu_to_le(lo, lo_ram);
1012         th = dt_trans_create(env, dev);
1013         if (IS_ERR(th))
1014                 GOTO(log, rc = PTR_ERR(th));
1015
1016         rc = dt_declare_record_write(env, obj, lfsck_buf_get(env, lo, size),
1017                                      (loff_t)0, th);
1018         if (rc != 0)
1019                 GOTO(out, rc);
1020
1021         if (bitmap != NULL) {
1022                 rc = dt_declare_record_write(env, obj,
1023                                 lfsck_buf_get(env, bitmap->data, nbits >> 3),
1024                                 (loff_t)size, th);
1025                 if (rc != 0)
1026                         GOTO(out, rc);
1027         }
1028
1029         rc = dt_trans_start_local(env, dev, th);
1030         if (rc != 0)
1031                 GOTO(out, rc);
1032
1033         pos = 0;
1034         rc = dt_record_write(env, obj, lfsck_buf_get(env, lo, size), &pos, th);
1035         if (rc != 0)
1036                 GOTO(out, rc);
1037
1038         if (bitmap != NULL) {
1039                 pos = size;
1040                 rc = dt_record_write(env, obj,
1041                                 lfsck_buf_get(env, bitmap->data, nbits >> 3),
1042                                 &pos, th);
1043         }
1044
1045         GOTO(out, rc);
1046
1047 out:
1048         dt_trans_stop(env, dev, th);
1049
1050 log:
1051         if (rc != 0)
1052                 CDEBUG(D_LFSCK, "%s: fail to store lfsck_layout: rc = %d\n",
1053                        lfsck_lfsck2name(lfsck), rc);
1054
1055         return rc;
1056 }
1057
1058 static int lfsck_layout_init(const struct lu_env *env,
1059                              struct lfsck_component *com)
1060 {
1061         struct lfsck_layout *lo = com->lc_file_ram;
1062         int rc;
1063
1064         memset(lo, 0, com->lc_file_size);
1065         lo->ll_magic = LFSCK_LAYOUT_MAGIC;
1066         lo->ll_status = LS_INIT;
1067         down_write(&com->lc_sem);
1068         rc = lfsck_layout_store(env, com);
1069         if (rc == 0 && com->lc_lfsck->li_master)
1070                 rc = lfsck_load_sub_trace_files(env, com,
1071                         &dt_lfsck_layout_dangling_features, LFSCK_LAYOUT, true);
1072         up_write(&com->lc_sem);
1073
1074         return rc;
1075 }
1076
1077 static int fid_is_for_ostobj(const struct lu_env *env,
1078                              struct lfsck_instance *lfsck,
1079                              struct dt_object *obj, const struct lu_fid *fid)
1080 {
1081         struct seq_server_site  *ss     = lfsck_dev_site(lfsck);
1082         struct lu_seq_range     *range  = &lfsck_env_info(env)->lti_range;
1083         struct lustre_ost_attrs *loa;
1084         int                      rc;
1085
1086         fld_range_set_any(range);
1087         rc = fld_server_lookup(env, ss->ss_server_fld, fid_seq(fid), range);
1088         if (rc == 0) {
1089                 if (fld_range_is_ost(range))
1090                         return 1;
1091
1092                 return 0;
1093         }
1094
1095         loa = &lfsck_env_info(env)->lti_loa;
1096         rc = dt_xattr_get(env, obj, lfsck_buf_get(env, loa, sizeof(*loa)),
1097                           XATTR_NAME_LMA);
1098         if (rc >= sizeof(struct lustre_mdt_attrs)) {
1099                 lustre_lma_swab(&loa->loa_lma);
1100
1101                 return loa->loa_lma.lma_compat & LMAC_FID_ON_OST ? 1 : 0;
1102         }
1103
1104         rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_FID);
1105
1106         return rc > 0;
1107 }
1108
1109 static struct lfsck_layout_seq *
1110 lfsck_layout_seq_lookup(struct lfsck_layout_slave_data *llsd, __u64 seq)
1111 {
1112         struct lfsck_layout_seq *lls;
1113
1114         list_for_each_entry(lls, &llsd->llsd_seq_list, lls_list) {
1115                 if (lls->lls_seq == seq)
1116                         return lls;
1117
1118                 if (lls->lls_seq > seq)
1119                         return NULL;
1120         }
1121
1122         return NULL;
1123 }
1124
1125 static void
1126 lfsck_layout_seq_insert(struct lfsck_layout_slave_data *llsd,
1127                         struct lfsck_layout_seq *lls)
1128 {
1129         struct lfsck_layout_seq *tmp;
1130         struct list_head        *pos = &llsd->llsd_seq_list;
1131
1132         list_for_each_entry(tmp, &llsd->llsd_seq_list, lls_list) {
1133                 if (lls->lls_seq < tmp->lls_seq) {
1134                         pos = &tmp->lls_list;
1135                         break;
1136                 }
1137         }
1138         list_add_tail(&lls->lls_list, pos);
1139 }
1140
1141 static int
1142 lfsck_layout_lastid_create(const struct lu_env *env,
1143                            struct lfsck_instance *lfsck,
1144                            struct dt_object *obj)
1145 {
1146         struct lfsck_thread_info *info   = lfsck_env_info(env);
1147         struct lu_attr           *la     = &info->lti_la;
1148         struct dt_object_format  *dof    = &info->lti_dof;
1149         struct lfsck_bookmark    *bk     = &lfsck->li_bookmark_ram;
1150         struct dt_device         *dt     = lfsck_obj2dev(obj);
1151         struct thandle           *th;
1152         __u64                     lastid = 0;
1153         loff_t                    pos    = 0;
1154         int                       rc;
1155         ENTRY;
1156
1157         if (bk->lb_param & LPF_DRYRUN)
1158                 return 0;
1159
1160         memset(la, 0, sizeof(*la));
1161         la->la_mode = S_IFREG |  S_IRUGO | S_IWUSR;
1162         la->la_valid = LA_MODE | LA_UID | LA_GID;
1163         memset(dof, 0, sizeof(*dof));
1164         dof->dof_type = dt_mode_to_dft(S_IFREG);
1165
1166         th = dt_trans_create(env, dt);
1167         if (IS_ERR(th))
1168                 GOTO(log, rc = PTR_ERR(th));
1169
1170         rc = dt_declare_create(env, obj, la, NULL, dof, th);
1171         if (rc != 0)
1172                 GOTO(stop, rc);
1173
1174         rc = dt_declare_record_write(env, obj,
1175                                      lfsck_buf_get(env, &lastid,
1176                                                    sizeof(lastid)),
1177                                      pos, th);
1178         if (rc != 0)
1179                 GOTO(stop, rc);
1180
1181         rc = dt_trans_start_local(env, dt, th);
1182         if (rc != 0)
1183                 GOTO(stop, rc);
1184
1185         dt_write_lock(env, obj, 0);
1186         if (likely(dt_object_exists(obj) == 0)) {
1187                 rc = dt_create(env, obj, la, NULL, dof, th);
1188                 if (rc == 0)
1189                         rc = dt_record_write(env, obj,
1190                                 lfsck_buf_get(env, &lastid, sizeof(lastid)),
1191                                 &pos, th);
1192         }
1193         dt_write_unlock(env, obj);
1194
1195         GOTO(stop, rc);
1196
1197 stop:
1198         dt_trans_stop(env, dt, th);
1199
1200 log:
1201         CDEBUG(D_LFSCK, "%s: layout LFSCK will create LAST_ID for <seq> "
1202                "%#llx: rc = %d\n",
1203                lfsck_lfsck2name(lfsck), fid_seq(lfsck_dto2fid(obj)), rc);
1204
1205         return rc;
1206 }
1207
1208 static int
1209 lfsck_layout_lastid_reload(const struct lu_env *env,
1210                            struct lfsck_component *com,
1211                            struct lfsck_layout_seq *lls)
1212 {
1213         __u64   lastid;
1214         loff_t  pos     = 0;
1215         int     rc;
1216
1217         dt_read_lock(env, lls->lls_lastid_obj, 0);
1218         rc = dt_record_read(env, lls->lls_lastid_obj,
1219                             lfsck_buf_get(env, &lastid, sizeof(lastid)), &pos);
1220         dt_read_unlock(env, lls->lls_lastid_obj);
1221         if (unlikely(rc != 0))
1222                 return rc;
1223
1224         lastid = le64_to_cpu(lastid);
1225         if (lastid < lls->lls_lastid_known) {
1226                 struct lfsck_instance   *lfsck  = com->lc_lfsck;
1227                 struct lfsck_layout     *lo     = com->lc_file_ram;
1228
1229                 lls->lls_lastid = lls->lls_lastid_known;
1230                 lls->lls_dirty = 1;
1231                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
1232                         LASSERT(lfsck->li_out_notify != NULL);
1233
1234                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1235                                              LE_LASTID_REBUILDING);
1236                         lo->ll_flags |= LF_CRASHED_LASTID;
1237
1238                         CDEBUG(D_LFSCK, "%s: layout LFSCK finds crashed "
1239                                "LAST_ID file (1) for the sequence %#llx"
1240                                ", old value %llu, known value %llu\n",
1241                                lfsck_lfsck2name(lfsck), lls->lls_seq,
1242                                lastid, lls->lls_lastid);
1243                 }
1244         } else if (lastid >= lls->lls_lastid) {
1245                 lls->lls_lastid = lastid;
1246                 lls->lls_dirty = 0;
1247         }
1248
1249         return 0;
1250 }
1251
1252 static int
1253 lfsck_layout_lastid_store(const struct lu_env *env,
1254                           struct lfsck_component *com)
1255 {
1256         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1257         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
1258         struct dt_device                *dt     = lfsck->li_bottom;
1259         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
1260         struct lfsck_layout_seq         *lls;
1261         struct thandle                  *th;
1262         __u64                            lastid;
1263         int                              rc     = 0;
1264         int                              rc1    = 0;
1265
1266         list_for_each_entry(lls, &llsd->llsd_seq_list, lls_list) {
1267                 loff_t pos = 0;
1268
1269                 if (!lls->lls_dirty)
1270                         continue;
1271
1272                 CDEBUG(D_LFSCK, "%s: layout LFSCK will sync the LAST_ID for "
1273                        "<seq> %#llx as <oid> %llu\n",
1274                        lfsck_lfsck2name(lfsck), lls->lls_seq, lls->lls_lastid);
1275
1276                 if (bk->lb_param & LPF_DRYRUN) {
1277                         lls->lls_dirty = 0;
1278                         continue;
1279                 }
1280
1281                 th = dt_trans_create(env, dt);
1282                 if (IS_ERR(th)) {
1283                         rc1 = PTR_ERR(th);
1284                         CDEBUG(D_LFSCK, "%s: layout LFSCK failed to store "
1285                                "the LAST_ID for <seq> %#llx(1): rc = %d\n",
1286                                lfsck_lfsck2name(com->lc_lfsck),
1287                                lls->lls_seq, rc1);
1288                         continue;
1289                 }
1290
1291                 lastid = cpu_to_le64(lls->lls_lastid);
1292                 rc = dt_declare_record_write(env, lls->lls_lastid_obj,
1293                                              lfsck_buf_get(env, &lastid,
1294                                                            sizeof(lastid)),
1295                                              pos, th);
1296                 if (rc != 0)
1297                         goto stop;
1298
1299                 rc = dt_trans_start_local(env, dt, th);
1300                 if (rc != 0)
1301                         goto stop;
1302
1303                 dt_write_lock(env, lls->lls_lastid_obj, 0);
1304                 rc = dt_record_write(env, lls->lls_lastid_obj,
1305                                      lfsck_buf_get(env, &lastid,
1306                                      sizeof(lastid)), &pos, th);
1307                 dt_write_unlock(env, lls->lls_lastid_obj);
1308                 if (rc == 0)
1309                         lls->lls_dirty = 0;
1310
1311 stop:
1312                 dt_trans_stop(env, dt, th);
1313                 if (rc != 0) {
1314                         rc1 = rc;
1315                         CDEBUG(D_LFSCK, "%s: layout LFSCK failed to store "
1316                                "the LAST_ID for <seq> %#llx(2): rc = %d\n",
1317                                lfsck_lfsck2name(com->lc_lfsck),
1318                                lls->lls_seq, rc1);
1319                 }
1320         }
1321
1322         return rc1;
1323 }
1324
1325 static int
1326 lfsck_layout_lastid_load(const struct lu_env *env,
1327                          struct lfsck_component *com,
1328                          struct lfsck_layout_seq *lls)
1329 {
1330         struct lfsck_instance   *lfsck  = com->lc_lfsck;
1331         struct lfsck_layout     *lo     = com->lc_file_ram;
1332         struct lu_fid           *fid    = &lfsck_env_info(env)->lti_fid;
1333         struct dt_object        *obj;
1334         loff_t                   pos    = 0;
1335         int                      rc;
1336         ENTRY;
1337
1338         lu_last_id_fid(fid, lls->lls_seq, lfsck_dev_idx(lfsck));
1339         obj = dt_locate(env, lfsck->li_bottom, fid);
1340         if (IS_ERR(obj))
1341                 RETURN(PTR_ERR(obj));
1342
1343         /* LAST_ID crashed, to be rebuilt */
1344         if (dt_object_exists(obj) == 0) {
1345                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
1346                         LASSERT(lfsck->li_out_notify != NULL);
1347
1348                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1349                                              LE_LASTID_REBUILDING);
1350                         lo->ll_flags |= LF_CRASHED_LASTID;
1351
1352                         CDEBUG(D_LFSCK, "%s: layout LFSCK cannot find the "
1353                                "LAST_ID file for sequence %#llx\n",
1354                                lfsck_lfsck2name(lfsck), lls->lls_seq);
1355
1356                         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DELAY4) &&
1357                             cfs_fail_val > 0) {
1358                                 struct l_wait_info lwi = LWI_TIMEOUT(
1359                                                 cfs_time_seconds(cfs_fail_val),
1360                                                 NULL, NULL);
1361
1362                                 /* Some others may changed the cfs_fail_val
1363                                  * as zero after above check, re-check it for
1364                                  * sure to avoid falling into wait for ever. */
1365                                 if (likely(lwi.lwi_timeout > 0)) {
1366                                         struct ptlrpc_thread *thread =
1367                                                 &lfsck->li_thread;
1368
1369                                         up_write(&com->lc_sem);
1370                                         l_wait_event(thread->t_ctl_waitq,
1371                                                      !thread_is_running(thread),
1372                                                      &lwi);
1373                                         down_write(&com->lc_sem);
1374                                 }
1375                         }
1376                 }
1377
1378                 rc = lfsck_layout_lastid_create(env, lfsck, obj);
1379         } else {
1380                 dt_read_lock(env, obj, 0);
1381                 rc = dt_read(env, obj,
1382                         lfsck_buf_get(env, &lls->lls_lastid, sizeof(__u64)),
1383                         &pos);
1384                 dt_read_unlock(env, obj);
1385                 if (rc != 0 && rc != sizeof(__u64))
1386                         GOTO(out, rc = (rc > 0 ? -EFAULT : rc));
1387
1388                 if (rc == 0 && !(lo->ll_flags & LF_CRASHED_LASTID)) {
1389                         LASSERT(lfsck->li_out_notify != NULL);
1390
1391                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1392                                              LE_LASTID_REBUILDING);
1393                         lo->ll_flags |= LF_CRASHED_LASTID;
1394
1395                         CDEBUG(D_LFSCK, "%s: layout LFSCK finds invalid "
1396                                "LAST_ID file for the sequence %#llx"
1397                                ": rc = %d\n",
1398                                lfsck_lfsck2name(lfsck), lls->lls_seq, rc);
1399                 }
1400
1401                 lls->lls_lastid = le64_to_cpu(lls->lls_lastid);
1402                 rc = 0;
1403         }
1404
1405         GOTO(out, rc);
1406
1407 out:
1408         if (rc != 0)
1409                 lfsck_object_put(env, obj);
1410         else
1411                 lls->lls_lastid_obj = obj;
1412
1413         return rc;
1414 }
1415
1416 static void lfsck_layout_record_failure(const struct lu_env *env,
1417                                         struct lfsck_instance *lfsck,
1418                                         struct lfsck_layout *lo)
1419 {
1420         __u64 cookie;
1421
1422         lo->ll_objs_failed_phase1++;
1423         cookie = lfsck->li_obj_oit->do_index_ops->dio_it.store(env,
1424                                                         lfsck->li_di_oit);
1425         if (lo->ll_pos_first_inconsistent == 0 ||
1426             lo->ll_pos_first_inconsistent < cookie) {
1427                 lo->ll_pos_first_inconsistent = cookie;
1428
1429                 CDEBUG(D_LFSCK, "%s: layout LFSCK hit first non-repaired "
1430                        "inconsistency at the pos [%llu]\n",
1431                        lfsck_lfsck2name(lfsck),
1432                        lo->ll_pos_first_inconsistent);
1433         }
1434 }
1435
1436 static int lfsck_layout_double_scan_result(const struct lu_env *env,
1437                                            struct lfsck_component *com,
1438                                            int rc)
1439 {
1440         struct lfsck_instance   *lfsck = com->lc_lfsck;
1441         struct lfsck_layout     *lo    = com->lc_file_ram;
1442
1443         down_write(&com->lc_sem);
1444         lo->ll_run_time_phase2 += cfs_duration_sec(cfs_time_current() +
1445                                   HALF_SEC - com->lc_time_last_checkpoint);
1446         lo->ll_time_last_checkpoint = cfs_time_current_sec();
1447         lo->ll_objs_checked_phase2 += com->lc_new_checked;
1448
1449         if (rc > 0) {
1450                 if (lo->ll_flags & LF_INCOMPLETE) {
1451                         lo->ll_status = LS_PARTIAL;
1452                 } else {
1453                         if (lfsck->li_master) {
1454                                 struct lfsck_assistant_data *lad = com->lc_data;
1455
1456                                 if (lad->lad_incomplete)
1457                                         lo->ll_status = LS_PARTIAL;
1458                                 else
1459                                         lo->ll_status = LS_COMPLETED;
1460                         } else {
1461                                 lo->ll_status = LS_COMPLETED;
1462                         }
1463                 }
1464                 lo->ll_flags &= ~LF_SCANNED_ONCE;
1465                 if (!(lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN))
1466                         lo->ll_flags &= ~LF_INCONSISTENT;
1467                 lo->ll_time_last_complete = lo->ll_time_last_checkpoint;
1468                 lo->ll_success_count++;
1469         } else if (rc == 0) {
1470                 if (lfsck->li_status != 0)
1471                         lo->ll_status = lfsck->li_status;
1472                 else
1473                         lo->ll_status = LS_STOPPED;
1474         } else {
1475                 lo->ll_status = LS_FAILED;
1476         }
1477
1478         rc = lfsck_layout_store(env, com);
1479         up_write(&com->lc_sem);
1480
1481         return rc;
1482 }
1483
1484 static int lfsck_layout_trans_stop(const struct lu_env *env,
1485                                    struct dt_device *dev,
1486                                    struct thandle *handle, int result)
1487 {
1488         int rc;
1489
1490         /* XXX: If there is something worng or it needs to repair nothing,
1491          *      then notify the lower to stop the modification. Currently,
1492          *      we use th_result for such purpose, that may be replaced by
1493          *      some rollback mechanism in the future. */
1494         handle->th_result = result;
1495         rc = dt_trans_stop(env, dev, handle);
1496         if (result != 0)
1497                 return result > 0 ? 0 : result;
1498
1499         return rc == 0 ? 1 : rc;
1500 }
1501
1502 static int lfsck_layout_ins_dangling_rec(const struct lu_env *env,
1503                                          struct lfsck_component *com,
1504                                          const struct lu_fid *pfid,
1505                                          const struct lu_fid *cfid,
1506                                          __u32 comp_id, __u32 ea_off,
1507                                          __u32 ost_idx)
1508 {
1509         struct lfsck_layout_dangling_key *key = &lfsck_env_info(env)->lti_lldk;
1510         struct lu_fid *rec = &lfsck_env_info(env)->lti_fid3;
1511         struct dt_device *dev;
1512         struct dt_object *obj;
1513         struct thandle *th = NULL;
1514         int idx;
1515         int rc = 0;
1516         ENTRY;
1517
1518         idx = lfsck_sub_trace_file_fid2idx(pfid);
1519         obj = com->lc_sub_trace_objs[idx].lsto_obj;
1520         dev = lfsck_obj2dev(obj);
1521
1522         fid_cpu_to_be(&key->lldk_fid, pfid);
1523         key->lldk_comp_id = cpu_to_be32(comp_id);
1524         key->lldk_ea_off = cpu_to_be32(ea_off);
1525
1526         fid_cpu_to_be(rec, cfid);
1527         rec->f_ver = cpu_to_be32(ost_idx);
1528
1529         mutex_lock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1530
1531         th = dt_trans_create(env, dev);
1532         if (IS_ERR(th))
1533                 GOTO(unlock, rc = PTR_ERR(th));
1534
1535         rc = dt_declare_insert(env, obj,
1536                                (const struct dt_rec *)rec,
1537                                (const struct dt_key *)key, th);
1538         if (rc)
1539                 GOTO(unlock, rc);
1540
1541         rc = dt_trans_start_local(env, dev, th);
1542         if (rc)
1543                 GOTO(unlock, rc);
1544
1545         rc = dt_insert(env, obj, (const struct dt_rec *)rec,
1546                        (const struct dt_key *)key, th, 1);
1547
1548         GOTO(unlock, rc);
1549
1550 unlock:
1551         if (th && !IS_ERR(th))
1552                 dt_trans_stop(env, dev, th);
1553
1554         mutex_unlock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1555
1556         CDEBUG(D_LFSCK, "%s: insert the paris "DFID" => "DFID", comp_id = %u, "
1557                "ea_off = %u, ost_idx = %u, into the trace file for further "
1558                "dangling check: rc = %d\n", lfsck_lfsck2name(com->lc_lfsck),
1559                PFID(pfid), PFID(cfid), comp_id, ea_off, ost_idx, rc);
1560
1561         return rc;
1562 }
1563
1564 static int lfsck_layout_del_dangling_rec(const struct lu_env *env,
1565                                          struct lfsck_component *com,
1566                                          const struct lu_fid *fid,
1567                                          __u32 comp_id, __u32 ea_off)
1568 {
1569         struct lfsck_layout_dangling_key *key = &lfsck_env_info(env)->lti_lldk;
1570         struct dt_device *dev;
1571         struct dt_object *obj;
1572         struct thandle *th = NULL;
1573         int idx;
1574         int rc = 0;
1575         ENTRY;
1576
1577         idx = lfsck_sub_trace_file_fid2idx(fid);
1578         obj = com->lc_sub_trace_objs[idx].lsto_obj;
1579         dev = lfsck_obj2dev(obj);
1580
1581         fid_cpu_to_be(&key->lldk_fid, fid);
1582         key->lldk_comp_id = cpu_to_be32(comp_id);
1583         key->lldk_ea_off = cpu_to_be32(ea_off);
1584
1585         mutex_lock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1586
1587         th = dt_trans_create(env, dev);
1588         if (IS_ERR(th))
1589                 GOTO(unlock, rc = PTR_ERR(th));
1590
1591         rc = dt_declare_delete(env, obj, (const struct dt_key *)key, th);
1592         if (rc)
1593                 GOTO(unlock, rc);
1594
1595         rc = dt_trans_start_local(env, dev, th);
1596         if (rc)
1597                 GOTO(unlock, rc);
1598
1599         rc = dt_delete(env, obj, (const struct dt_key *)key, th);
1600
1601         GOTO(unlock, rc);
1602
1603 unlock:
1604         if (th && !IS_ERR(th))
1605                 dt_trans_stop(env, dev, th);
1606
1607         mutex_unlock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1608
1609         CDEBUG(D_LFSCK, "%s: delete the dangling record for "DFID
1610                ", comp_id = %u, ea_off = %u from the trace file: rc = %d\n",
1611                lfsck_lfsck2name(com->lc_lfsck), PFID(fid), comp_id, ea_off, rc);
1612
1613         return rc;
1614 }
1615
1616 /**
1617  * Get the system default stripe size.
1618  *
1619  * \param[in] env       pointer to the thread context
1620  * \param[in] lfsck     pointer to the lfsck instance
1621  * \param[out] size     pointer to the default stripe size
1622  *
1623  * \retval              0 for success
1624  * \retval              negative error number on failure
1625  */
1626 static int lfsck_layout_get_def_stripesize(const struct lu_env *env,
1627                                            struct lfsck_instance *lfsck,
1628                                            __u32 *size)
1629 {
1630         struct lov_user_md      *lum = &lfsck_env_info(env)->lti_lum;
1631         struct dt_object        *root;
1632         int                      rc;
1633
1634         root = dt_locate(env, lfsck->li_next, &lfsck->li_local_root_fid);
1635         if (IS_ERR(root))
1636                 return PTR_ERR(root);
1637
1638         /* Get the default stripe size via xattr_get on the backend root. */
1639         rc = dt_xattr_get(env, root, lfsck_buf_get(env, lum, sizeof(*lum)),
1640                           XATTR_NAME_LOV);
1641         if (rc > 0) {
1642                 /* The lum->lmm_stripe_size is LE mode. The *size also
1643                  * should be LE mode. So it is unnecessary to convert. */
1644                 *size = lum->lmm_stripe_size;
1645                 rc = 0;
1646         } else if (unlikely(rc == 0)) {
1647                 rc = -EINVAL;
1648         }
1649
1650         lfsck_object_put(env, root);
1651
1652         return rc;
1653 }
1654
1655 /**
1656  * \retval       +1: repaired
1657  * \retval        0: did nothing
1658  * \retval      -ve: on error
1659  */
1660 static int lfsck_layout_refill_lovea(const struct lu_env *env,
1661                                      struct lfsck_instance *lfsck,
1662                                      struct thandle *handle,
1663                                      struct dt_object *parent,
1664                                      const struct lu_fid *cfid,
1665                                      struct lu_buf *buf,
1666                                      struct lov_mds_md_v1 *lmm,
1667                                      struct lov_ost_data_v1 *slot,
1668                                      int fl, __u32 ost_idx, int size)
1669 {
1670         struct ost_id           *oi     = &lfsck_env_info(env)->lti_oi;
1671         struct lu_buf            ea_buf;
1672         int                      rc;
1673         __u32                    magic;
1674         __u32                    pattern;
1675         __u16                    count;
1676         ENTRY;
1677
1678         magic = le32_to_cpu(lmm->lmm_magic);
1679         pattern = le32_to_cpu(lmm->lmm_pattern);
1680         count = le16_to_cpu(lmm->lmm_stripe_count);
1681
1682         fid_to_ostid(cfid, oi);
1683         ostid_cpu_to_le(oi, &slot->l_ost_oi);
1684         slot->l_ost_gen = cpu_to_le32(0);
1685         slot->l_ost_idx = cpu_to_le32(ost_idx);
1686
1687         if (pattern & LOV_PATTERN_F_HOLE) {
1688                 struct lov_ost_data_v1 *objs;
1689                 int                     i;
1690
1691                 if (magic == LOV_MAGIC_V1)
1692                         objs = &lmm->lmm_objects[0];
1693                 else
1694                         objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
1695                 for (i = 0; i < count; i++, objs++) {
1696                         if (lovea_slot_is_dummy(objs))
1697                                 break;
1698                 }
1699
1700                 /* If the @slot is the last dummy slot to be refilled,
1701                  * then drop LOV_PATTERN_F_HOLE from lmm::lmm_pattern. */
1702                 if (i == count) {
1703                         lmm->lmm_pattern =
1704                                 cpu_to_le32(pattern & ~LOV_PATTERN_F_HOLE);
1705
1706                         CDEBUG(D_LFSCK, "%s: remove layout HOLE for "DFID
1707                                ": parent "DFID"\n", lfsck_lfsck2name(lfsck),
1708                                PFID(cfid), PFID(lfsck_dto2fid(parent)));
1709                 }
1710         }
1711
1712         lfsck_buf_init(&ea_buf, buf->lb_buf, size);
1713         rc = dt_xattr_set(env, parent, &ea_buf, XATTR_NAME_LOV, fl, handle);
1714         if (rc == 0)
1715                 rc = 1;
1716
1717         RETURN(rc);
1718 }
1719
1720 static struct lov_ost_data_v1 *
1721 __lfsck_layout_new_v1_lovea(struct lov_mds_md_v1 *lmm,
1722                             const struct lu_fid *pfid,
1723                             __u32 stripe_size, __u32 ea_off,
1724                             __u32 pattern, __u16 count)
1725 {
1726         lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
1727         lmm->lmm_pattern = cpu_to_le32(pattern);
1728         fid_to_lmm_oi(pfid, &lmm->lmm_oi);
1729         lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
1730         lmm->lmm_stripe_size = cpu_to_le32(stripe_size);
1731         lmm->lmm_stripe_count = cpu_to_le16(count);
1732         lmm->lmm_layout_gen = cpu_to_le16(1);
1733         memset(&lmm->lmm_objects[0], 0,
1734                sizeof(struct lov_ost_data_v1) * count);
1735
1736         return &lmm->lmm_objects[ea_off];
1737 }
1738
1739 static int lfsck_layout_new_v1_lovea(const struct lu_env *env,
1740                                      struct lfsck_instance *lfsck,
1741                                      struct ost_layout *ol,
1742                                      struct dt_object *parent,
1743                                      struct lu_buf *buf, __u32 ea_off,
1744                                      struct lov_mds_md_v1 **lmm,
1745                                      struct lov_ost_data_v1 **objs)
1746 {
1747         int size;
1748         __u32 stripe_size = ol->ol_stripe_size;
1749         __u32 pattern = LOV_PATTERN_RAID0;
1750         __u16 count;
1751
1752         if (ol->ol_stripe_count != 0)
1753                 count = ol->ol_stripe_count;
1754         else
1755                 count = ea_off + 1;
1756
1757         size = lov_mds_md_size(count, LOV_MAGIC_V1);
1758         LASSERTF(buf->lb_len >= size,
1759                  "buffer len %d is less than real size %d\n",
1760                  (int)buf->lb_len, size);
1761
1762         if (stripe_size == 0) {
1763                 int rc;
1764
1765                 rc = lfsck_layout_get_def_stripesize(env, lfsck, &stripe_size);
1766                 if (rc)
1767                         return rc;
1768         }
1769
1770         *lmm = buf->lb_buf;
1771         if (ol->ol_stripe_count > 1 ||
1772             (ol->ol_stripe_count == 0 && ea_off != 0)) {
1773                 pattern |= LOV_PATTERN_F_HOLE;
1774                 memset(&(*lmm)->lmm_objects[0], 0,
1775                        count * sizeof(struct lov_ost_data_v1));
1776         }
1777
1778         *objs = __lfsck_layout_new_v1_lovea(*lmm, lfsck_dto2fid(parent),
1779                                 stripe_size, ea_off, pattern, count);
1780
1781         return size;
1782 }
1783
1784 static int lfsck_layout_new_comp_lovea(const struct lu_env *env,
1785                                       struct ost_layout *ol,
1786                                       struct dt_object *parent,
1787                                       struct lu_buf *buf, __u32 ea_off,
1788                                       struct lov_mds_md_v1 **lmm,
1789                                       struct lov_ost_data_v1 **objs)
1790 {
1791         struct lov_comp_md_v1 *lcm;
1792         struct lov_comp_md_entry_v1 *lcme;
1793         __u32 pattern = LOV_PATTERN_RAID0;
1794         __u32 offset = sizeof(*lcm) + sizeof(*lcme);
1795         int lcme_size = lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
1796         int size = offset + lcme_size;
1797
1798         LASSERTF(buf->lb_len >= size,
1799                  "buffer len %d is less than real size %d\n",
1800                  (int)buf->lb_len, size);
1801
1802         lcm = buf->lb_buf;
1803         lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1804         lcm->lcm_size = cpu_to_le32(size);
1805         lcm->lcm_layout_gen = cpu_to_le32(1);
1806         lcm->lcm_flags = 0;
1807         lcm->lcm_entry_count = cpu_to_le16(1);
1808
1809         lcme = &lcm->lcm_entries[0];
1810         lcme->lcme_id = cpu_to_le32(ol->ol_comp_id);
1811         lcme->lcme_flags = cpu_to_le32(LCME_FL_INIT);
1812         lcme->lcme_extent.e_start = cpu_to_le64(ol->ol_comp_start);
1813         lcme->lcme_extent.e_end = cpu_to_le64(ol->ol_comp_end);
1814         lcme->lcme_offset = cpu_to_le32(offset);
1815         lcme->lcme_size = cpu_to_le32(lcme_size);
1816         if (ol->ol_stripe_count > 1)
1817                 pattern |= LOV_PATTERN_F_HOLE;
1818
1819         *lmm = buf->lb_buf + offset;
1820         *objs = __lfsck_layout_new_v1_lovea(*lmm, lfsck_dto2fid(parent),
1821                                             ol->ol_stripe_size, ea_off,
1822                                             pattern, ol->ol_stripe_count);
1823
1824         return size;
1825 }
1826
1827 static int lfsck_layout_add_comp_comp(const struct lu_env *env,
1828                                      struct lfsck_instance *lfsck,
1829                                      struct thandle *handle,
1830                                      struct ost_layout *ol,
1831                                      struct dt_object *parent,
1832                                      const struct lu_fid *cfid,
1833                                      struct lu_buf *buf, __u32 ost_idx,
1834                                      __u32 ea_off, int pos)
1835 {
1836         struct lov_comp_md_v1 *lcm = buf->lb_buf;
1837         struct lov_comp_md_entry_v1 *lcme;
1838         struct lov_mds_md_v1 *lmm;
1839         struct lov_ost_data_v1 *objs;
1840         int added = sizeof(*lcme) +
1841                     lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
1842         int size = le32_to_cpu(lcm->lcm_size) + added;
1843         int rc;
1844         int i;
1845         __u32 offset;
1846         __u32 pattern = LOV_PATTERN_RAID0;
1847         __u16 count = le16_to_cpu(lcm->lcm_entry_count);
1848         ENTRY;
1849
1850         lu_buf_check_and_grow(buf, size);
1851         /* set the lcm again because lu_buf_check_and_grow() may
1852          * have reallocated the buf. */
1853         lcm = buf->lb_buf;
1854         lcm->lcm_size = cpu_to_le32(size);
1855         le32_add_cpu(&lcm->lcm_layout_gen, 1);
1856         lcm->lcm_entry_count = cpu_to_le16(count + 1);
1857
1858         /* 1. Move the component bodies from [pos, count-1] to [pos+1, count]
1859          *    with distance of 'added'. */
1860         if (pos < count) {
1861                 size = 0;
1862                 for (i = pos; i < count; i++) {
1863                         lcme = &lcm->lcm_entries[i];
1864                         size += le32_to_cpu(lcme->lcme_size);
1865                 }
1866
1867                 offset = le32_to_cpu(lcm->lcm_entries[pos].lcme_offset);
1868                 memmove(buf->lb_buf + offset + added,
1869                         buf->lb_buf + offset, size);
1870         }
1871
1872         size = 0;
1873         /* 2. Move the component header [0, pos-1] to [0, pos-1] with distance
1874          *    of 'sizeof(struct lov_comp_md_entry_v1)' */
1875         if (pos > 0) {
1876                 for (i = 0; i < pos; i++) {
1877                         lcme = &lcm->lcm_entries[i];
1878                         size += le32_to_cpu(lcme->lcme_size);
1879                 }
1880
1881                 offset = le32_to_cpu(lcm->lcm_entries[0].lcme_offset);
1882                 memmove(buf->lb_buf + offset + sizeof(*lcme),
1883                         buf->lb_buf + offset, size);
1884         }
1885
1886         /* 3. Recalculate the enter offset for the component [pos, count-1] */
1887         for (i = count - 1; i >= pos; i--) {
1888                 lcm->lcm_entries[i + 1] = lcm->lcm_entries[i];
1889                 lcm->lcm_entries[i + 1].lcme_offset =
1890                         cpu_to_le32(le32_to_cpu(lcm->lcm_entries[i + 1].
1891                                                 lcme_offset) + added);
1892         }
1893
1894         /* 4. Recalculate the enter offset for the component [0, pos) */
1895         for (i = 0; i < pos; i++) {
1896                 lcm->lcm_entries[i].lcme_offset =
1897                         cpu_to_le32(le32_to_cpu(lcm->lcm_entries[i].
1898                                                 lcme_offset) + sizeof(*lcme));
1899         }
1900
1901         offset = sizeof(*lcm) + sizeof(*lcme) * (count + 1) + size;
1902         /* 4. Insert the new component header (entry) at the slot 'pos'. */
1903         lcme = &lcm->lcm_entries[pos];
1904         lcme->lcme_id = cpu_to_le32(ol->ol_comp_id);
1905         lcme->lcme_flags = cpu_to_le32(LCME_FL_INIT);
1906         lcme->lcme_extent.e_start = cpu_to_le64(ol->ol_comp_start);
1907         lcme->lcme_extent.e_end = cpu_to_le64(ol->ol_comp_end);
1908         lcme->lcme_offset = cpu_to_le32(offset);
1909         lcme->lcme_size = cpu_to_le32(lov_mds_md_size(ol->ol_stripe_count,
1910                                                       LOV_MAGIC_V1));
1911
1912         if (ol->ol_stripe_count > 1)
1913                 pattern |= LOV_PATTERN_F_HOLE;
1914
1915         lmm = buf->lb_buf + offset;
1916         /* 5. Insert teh new component body at the 'offset'. */
1917         objs = __lfsck_layout_new_v1_lovea(lmm, lfsck_dto2fid(parent),
1918                                            ol->ol_stripe_size, ea_off,
1919                                            pattern, ol->ol_stripe_count);
1920
1921         rc = lfsck_layout_refill_lovea(env, lfsck, handle, parent, cfid, buf,
1922                                        lmm, objs, LU_XATTR_REPLACE, ost_idx,
1923                                        le32_to_cpu(lcm->lcm_size));
1924
1925         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant add new COMP for "
1926                DFID": parent "DFID", OST-index %u, stripe-index %u, "
1927                "stripe_size %u, stripe_count %u, comp_id %u, comp_start %llu, "
1928                "comp_end %llu, %s LOV EA hole: rc = %d\n",
1929                lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
1930                ost_idx, ea_off, ol->ol_stripe_size, ol->ol_stripe_count,
1931                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
1932                le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE ?
1933                "with" : "without", rc);
1934
1935         RETURN(rc);
1936 }
1937
1938 static int lfsck_layout_extend_v1v3_lovea(const struct lu_env *env,
1939                                           struct lfsck_instance *lfsck,
1940                                           struct thandle *handle,
1941                                           struct ost_layout *ol,
1942                                           struct dt_object *parent,
1943                                           const struct lu_fid *cfid,
1944                                           struct lu_buf *buf, __u32 ost_idx,
1945                                           __u32 ea_off)
1946 {
1947         struct lov_mds_md_v1 *lmm = buf->lb_buf;
1948         struct lov_ost_data_v1 *objs;
1949         __u16 count = le16_to_cpu(lmm->lmm_stripe_count);
1950         __u32 magic = le32_to_cpu(lmm->lmm_magic);
1951         int size;
1952         int gap;
1953         int rc;
1954         ENTRY;
1955
1956         /* The original LOVEA maybe re-generated via old filter_fid, at
1957          * that time, we do not know the stripe count and stripe size. */
1958         if (ol->ol_stripe_count > count)
1959                 count = ol->ol_stripe_count;
1960         if (ol->ol_stripe_size != 0 &&
1961             ol->ol_stripe_size != le32_to_cpu(lmm->lmm_stripe_size))
1962                 lmm->lmm_stripe_size = cpu_to_le32(ol->ol_stripe_size);
1963
1964         if (magic == LOV_MAGIC_V1)
1965                 objs = &lmm->lmm_objects[count];
1966         else
1967                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[count];
1968
1969         gap = ea_off - count;
1970         if (gap >= 0)
1971                 count = ea_off + 1;
1972
1973         size = lov_mds_md_size(count, magic);
1974         LASSERTF(buf->lb_len >= size,
1975                  "buffer len %d is less than real size %d\n",
1976                  (int)buf->lb_len, size);
1977
1978         if (gap > 0) {
1979                 memset(objs, 0, gap * sizeof(*objs));
1980                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_HOLE);
1981         }
1982
1983         lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
1984         lmm->lmm_stripe_count = cpu_to_le16(count);
1985         objs += gap;
1986
1987         rc = lfsck_layout_refill_lovea(env, lfsck, handle, parent, cfid, buf,
1988                                 lmm, objs, LU_XATTR_REPLACE, ost_idx, size);
1989
1990         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant extend layout EA for "
1991                DFID": parent "DFID", OST-index %u, stripe-index %u, "
1992                "stripe_size %u, stripe_count %u, comp_id %u, comp_start %llu, "
1993                "comp_end %llu, %s LOV EA hole: rc = %d\n",
1994                lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
1995                ost_idx, ea_off, ol->ol_stripe_size, ol->ol_stripe_count,
1996                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
1997                le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE ?
1998                "with" : "without", rc);
1999
2000         RETURN(rc);
2001 }
2002
2003 /**
2004  * \retval       +1: repaired
2005  * \retval        0: did nothing
2006  * \retval      -ve: on error
2007  */
2008 static int lfsck_layout_update_lovea(const struct lu_env *env,
2009                                      struct lfsck_instance *lfsck,
2010                                      struct thandle *handle,
2011                                      struct ost_layout *ol,
2012                                      struct dt_object *parent,
2013                                      const struct lu_fid *cfid,
2014                                      struct lu_buf *buf, int fl,
2015                                      __u32 ost_idx, __u32 ea_off)
2016 {
2017         struct lov_mds_md_v1 *lmm = NULL;
2018         struct lov_ost_data_v1 *objs = NULL;
2019         int rc = 0;
2020         ENTRY;
2021
2022         if (ol->ol_comp_id != 0)
2023                 rc = lfsck_layout_new_comp_lovea(env, ol, parent, buf, ea_off,
2024                                                 &lmm, &objs);
2025         else
2026                 rc = lfsck_layout_new_v1_lovea(env, lfsck, ol, parent, buf,
2027                                                ea_off, &lmm, &objs);
2028
2029         if (rc > 0)
2030                 rc = lfsck_layout_refill_lovea(env, lfsck, handle, parent, cfid,
2031                                                buf, lmm, objs, fl, ost_idx, rc);
2032
2033         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant created layout EA for "
2034                DFID": parent "DFID", OST-index %u, stripe-index %u, "
2035                "stripe_size %u, stripe_count %u, comp_id %u, comp_start %llu, "
2036                "comp_end %llu, fl %d, %s LOV EA hole: rc = %d\n",
2037                lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
2038                ost_idx, ea_off, ol->ol_stripe_size, ol->ol_stripe_count,
2039                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end, fl,
2040                le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE ?
2041                "with" : "without", rc);
2042
2043         RETURN(rc);
2044 }
2045
2046 static int __lfsck_layout_update_pfid(const struct lu_env *env,
2047                                       struct dt_object *child,
2048                                       const struct lu_fid *pfid,
2049                                       const struct ost_layout *ol, __u32 offset)
2050 {
2051         struct dt_device        *dev    = lfsck_obj2dev(child);
2052         struct filter_fid       *ff     = &lfsck_env_info(env)->lti_ff;
2053         struct thandle          *handle;
2054         struct lu_buf            buf    = { NULL };
2055         int                      rc;
2056
2057         ff->ff_parent.f_seq = cpu_to_le64(pfid->f_seq);
2058         ff->ff_parent.f_oid = cpu_to_le32(pfid->f_oid);
2059         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
2060          * MDT-object's FID::f_ver, instead it is the OST-object index in its
2061          * parent MDT-object's layout EA. */
2062         ff->ff_parent.f_stripe_idx = cpu_to_le32(offset);
2063         ost_layout_cpu_to_le(&ff->ff_layout, ol);
2064         lfsck_buf_init(&buf, ff, sizeof(*ff));
2065
2066         handle = dt_trans_create(env, dev);
2067         if (IS_ERR(handle))
2068                 RETURN(PTR_ERR(handle));
2069
2070         rc = dt_declare_xattr_set(env, child, &buf, XATTR_NAME_FID, 0, handle);
2071         if (rc != 0)
2072                 GOTO(stop, rc);
2073
2074         rc = dt_trans_start_local(env, dev, handle);
2075         if (rc != 0)
2076                 GOTO(stop, rc);
2077
2078         rc = dt_xattr_set(env, child, &buf, XATTR_NAME_FID, 0, handle);
2079
2080         GOTO(stop, rc);
2081
2082 stop:
2083         dt_trans_stop(env, dev, handle);
2084
2085         return rc;
2086 }
2087
2088 /**
2089  * \retval       +1: repaired
2090  * \retval        0: did nothing
2091  * \retval      -ve: on error
2092  */
2093 static int lfsck_layout_update_pfid(const struct lu_env *env,
2094                                     struct lfsck_component *com,
2095                                     struct dt_object *parent,
2096                                     struct lu_fid *cfid,
2097                                     struct dt_device *cdev,
2098                                     struct ost_layout *ol, __u32 ea_off)
2099 {
2100         struct dt_object        *child;
2101         int                      rc     = 0;
2102         ENTRY;
2103
2104         child = lfsck_object_find_by_dev(env, cdev, cfid);
2105         if (IS_ERR(child))
2106                 RETURN(PTR_ERR(child));
2107
2108         rc = __lfsck_layout_update_pfid(env, child,
2109                                         lu_object_fid(&parent->do_lu),
2110                                         ol, ea_off);
2111         lfsck_object_put(env, child);
2112
2113         RETURN(rc == 0 ? 1 : rc);
2114 }
2115
2116 static int lfsck_lovea_size(struct ost_layout *ol, __u32 ea_off)
2117 {
2118         if (ol->ol_comp_id != 0)
2119                 return sizeof(struct lov_comp_md_v1) +
2120                        sizeof(struct lov_comp_md_entry_v1) +
2121                        lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
2122
2123         if (ol->ol_stripe_count != 0)
2124                 return lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
2125
2126         return lov_mds_md_size(ea_off + 1, LOV_MAGIC_V1);
2127 }
2128
2129 /**
2130  * This function will create the MDT-object with the given (partial) LOV EA.
2131  *
2132  * Under some data corruption cases, the MDT-object of the file may be lost,
2133  * but its OST-objects, or some of them are there. The layout LFSCK needs to
2134  * re-create the MDT-object with the orphan OST-object(s) information.
2135  *
2136  * On the other hand, the LFSCK may has created some OST-object for repairing
2137  * dangling LOV EA reference, but as the LFSCK processing, it may find that
2138  * the old OST-object is there and should replace the former new created OST
2139  * object. Unfortunately, some others have modified such newly created object.
2140  * To keep the data (both new and old), the LFSCK will create MDT-object with
2141  * new FID to reference the original OST-object.
2142  *
2143  * \param[in] env       pointer to the thread context
2144  * \param[in] com       pointer to the lfsck component
2145  * \param[in] ltd       pointer to target device descriptor
2146  * \param[in] rec       pointer to the record for the orphan OST-object
2147  * \param[in] cfid      pointer to FID for the orphan OST-object
2148  * \param[in] infix     additional information, such as the FID for original
2149  *                      MDT-object and the stripe offset in the LOV EA
2150  * \param[in] type      the type for describing why the orphan MDT-object is
2151  *                      created. The rules are as following:
2152  *
2153  *  type "C":           Multiple OST-objects claim the same MDT-object and the
2154  *                      same slot in the layout EA. Then the LFSCK will create
2155  *                      new MDT-object(s) to hold the conflict OST-object(s).
2156  *
2157  *  type "N":           The orphan OST-object does not know which one was the
2158  *                      real parent MDT-object, so the LFSCK uses new FID for
2159  *                      its parent MDT-object.
2160  *
2161  *  type "R":           The orphan OST-object knows its parent MDT-object FID,
2162  *                      but does not know the position (the file name) in the
2163  *                      layout.
2164  *
2165  *  type "D":           The MDT-object is a directory, it may knows its parent
2166  *                      but because there is no valid linkEA, the LFSCK cannot
2167  *                      know where to put it back to the namespace.
2168  *  type "O":           The MDT-object has no linkEA, and there is no name
2169  *                      entry that references the MDT-object.
2170  *
2171  *  type "P":           The orphan object to be created was a parent directory
2172  *                      of some MDT-object which linkEA shows that the @orphan
2173  *                      object is missing.
2174  *
2175  * The orphan name will be like:
2176  * ${FID}-${infix}-${type}-${conflict_version}
2177  *
2178  * \param[in] ea_off    the stripe offset in the LOV EA
2179  *
2180  * \retval              positive on repaired something
2181  * \retval              0 if needs to repair nothing
2182  * \retval              negative error number on failure
2183  */
2184 static int lfsck_layout_recreate_parent(const struct lu_env *env,
2185                                         struct lfsck_component *com,
2186                                         struct lfsck_tgt_desc *ltd,
2187                                         struct lu_orphan_rec_v2 *rec,
2188                                         struct lu_fid *cfid,
2189                                         const char *infix,
2190                                         const char *type,
2191                                         __u32 ea_off)
2192 {
2193         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2194         struct dt_insert_rec            *dtrec  = &info->lti_dt_rec;
2195         char                            *name   = info->lti_key;
2196         struct lu_attr                  *la     = &info->lti_la2;
2197         struct dt_object_format         *dof    = &info->lti_dof;
2198         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2199         struct ost_layout               *ol     = &rec->lor_layout;
2200         struct lu_fid                   *pfid   = &rec->lor_rec.lor_fid;
2201         struct lu_fid                   *tfid   = &info->lti_fid3;
2202         struct dt_device                *dev    = lfsck->li_bottom;
2203         struct dt_object                *lpf    = lfsck->li_lpf_obj;
2204         struct dt_object                *pobj   = NULL;
2205         struct dt_object                *cobj   = NULL;
2206         struct thandle                  *th     = NULL;
2207         struct lu_buf                   *ea_buf = &info->lti_big_buf;
2208         struct lu_buf                    lov_buf;
2209         struct lfsck_lock_handle        *llh    = &info->lti_llh;
2210         struct linkea_data               ldata  = { NULL };
2211         struct lu_buf                    linkea_buf;
2212         const struct lu_name            *pname;
2213         int                              size   = 0;
2214         int                              idx    = 0;
2215         int                              rc     = 0;
2216         ENTRY;
2217
2218         if (unlikely(lpf == NULL))
2219                 GOTO(log, rc = -ENXIO);
2220
2221         /* We use two separated transactions to repair the inconsistency.
2222          *
2223          * 1) create the MDT-object locally.
2224          * 2) update the OST-object's PFID EA if necessary.
2225          *
2226          * If 1) succeed, but 2) failed, then the OST-object's PFID EA will be
2227          * updated when the layout LFSCK run next time.
2228          *
2229          * If 1) failed, but 2) succeed, then such MDT-object will be re-created
2230          * when the layout LFSCK run next time. */
2231
2232         if (fid_is_zero(pfid)) {
2233                 rc = lfsck_fid_alloc(env, lfsck, pfid, false);
2234                 if (rc != 0)
2235                         GOTO(log, rc);
2236
2237                 cobj = lfsck_object_find_by_dev(env, ltd->ltd_tgt, cfid);
2238                 if (IS_ERR(cobj))
2239                         GOTO(log, rc = PTR_ERR(cobj));
2240         }
2241
2242         pobj = lfsck_object_find_by_dev(env, dev, pfid);
2243         if (IS_ERR(pobj))
2244                 GOTO(log, rc = PTR_ERR(pobj));
2245
2246         LASSERT(infix != NULL);
2247         LASSERT(type != NULL);
2248
2249         memset(la, 0, sizeof(*la));
2250         la->la_uid = rec->lor_rec.lor_uid;
2251         la->la_gid = rec->lor_rec.lor_gid;
2252         la->la_mode = S_IFREG | S_IRUSR;
2253         la->la_valid = LA_MODE | LA_UID | LA_GID;
2254
2255         memset(dof, 0, sizeof(*dof));
2256         dof->dof_type = dt_mode_to_dft(S_IFREG);
2257         /* Because the dof->dof_reg.striped = 0, the LOD will not create
2258          * the stripe(s). The LFSCK will specify the LOV EA via
2259          * lfsck_layout_update_lovea(). */
2260
2261         size = lfsck_lovea_size(ol, ea_off);
2262         if (ea_buf->lb_len < size) {
2263                 lu_buf_realloc(ea_buf, size);
2264                 if (ea_buf->lb_buf == NULL)
2265                         GOTO(log, rc = -ENOMEM);
2266         }
2267
2268 again:
2269         do {
2270                 snprintf(name, NAME_MAX, DFID"%s-%s-%d", PFID(pfid), infix,
2271                          type, idx++);
2272                 rc = dt_lookup(env, lfsck->li_lpf_obj, (struct dt_rec *)tfid,
2273                                (const struct dt_key *)name);
2274                 if (rc != 0 && rc != -ENOENT)
2275                         GOTO(log, rc);
2276         } while (rc == 0);
2277
2278         rc = lfsck_lock(env, lfsck, lfsck->li_lpf_obj, name, llh,
2279                         MDS_INODELOCK_UPDATE, LCK_PW);
2280         if (rc != 0)
2281                 GOTO(log, rc);
2282
2283         /* Re-check whether the name conflict with othrs after taken
2284          * the ldlm lock. */
2285         rc = dt_lookup(env, lfsck->li_lpf_obj, (struct dt_rec *)tfid,
2286                        (const struct dt_key *)name);
2287         if (unlikely(rc == 0)) {
2288                 lfsck_unlock(llh);
2289                 goto again;
2290         }
2291
2292         if (rc != -ENOENT)
2293                 GOTO(unlock, rc);
2294
2295         pname = lfsck_name_get_const(env, name, strlen(name));
2296         rc = linkea_links_new(&ldata, &lfsck_env_info(env)->lti_linkea_buf,
2297                               pname, lfsck_dto2fid(lfsck->li_lpf_obj));
2298         if (rc != 0)
2299                 GOTO(unlock, rc);
2300
2301         /* The 1st transaction. */
2302         th = dt_trans_create(env, dev);
2303         if (IS_ERR(th))
2304                 GOTO(unlock, rc = PTR_ERR(th));
2305
2306         rc = dt_declare_create(env, pobj, la, NULL, dof, th);
2307         if (rc != 0)
2308                 GOTO(stop, rc);
2309
2310         lfsck_buf_init(&lov_buf, ea_buf->lb_buf, size);
2311         rc = dt_declare_xattr_set(env, pobj, &lov_buf, XATTR_NAME_LOV,
2312                                   LU_XATTR_CREATE, th);
2313         if (rc != 0)
2314                 GOTO(stop, rc);
2315
2316         dtrec->rec_fid = pfid;
2317         dtrec->rec_type = S_IFREG;
2318         rc = dt_declare_insert(env, lpf,
2319                                (const struct dt_rec *)dtrec,
2320                                (const struct dt_key *)name, th);
2321         if (rc != 0)
2322                 GOTO(stop, rc);
2323
2324         lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
2325                        ldata.ld_leh->leh_len);
2326         rc = dt_declare_xattr_set(env, pobj, &linkea_buf,
2327                                   XATTR_NAME_LINK, 0, th);
2328         if (rc != 0)
2329                 GOTO(stop, rc);
2330
2331         rc = dt_trans_start_local(env, dev, th);
2332         if (rc != 0)
2333                 GOTO(stop, rc);
2334
2335         dt_write_lock(env, pobj, 0);
2336         rc = dt_create(env, pobj, la, NULL, dof, th);
2337         if (rc == 0)
2338                 rc = lfsck_layout_update_lovea(env, lfsck, th, ol, pobj, cfid,
2339                         &lov_buf, LU_XATTR_CREATE, ltd->ltd_index, ea_off);
2340         dt_write_unlock(env, pobj);
2341         if (rc < 0)
2342                 GOTO(stop, rc);
2343
2344         rc = dt_insert(env, lpf, (const struct dt_rec *)dtrec,
2345                        (const struct dt_key *)name, th, 1);
2346         if (rc != 0)
2347                 GOTO(stop, rc);
2348
2349         rc = dt_xattr_set(env, pobj, &linkea_buf, XATTR_NAME_LINK, 0, th);
2350         if (rc == 0 && cobj != NULL) {
2351                 dt_trans_stop(env, dev, th);
2352                 th = NULL;
2353
2354                 /* The 2nd transaction. */
2355                 rc = __lfsck_layout_update_pfid(env, cobj, pfid, ol, ea_off);
2356         }
2357
2358         GOTO(stop, rc);
2359
2360 stop:
2361         if (th != NULL)
2362                 dt_trans_stop(env, dev, th);
2363
2364 unlock:
2365         lfsck_unlock(llh);
2366
2367 log:
2368         if (cobj != NULL && !IS_ERR(cobj))
2369                 lfsck_object_put(env, cobj);
2370         if (pobj != NULL && !IS_ERR(pobj))
2371                 lfsck_object_put(env, pobj);
2372
2373         if (rc < 0)
2374                 CDEBUG(D_LFSCK, "%s layout LFSCK assistant failed to "
2375                        "recreate the lost MDT-object: parent "DFID
2376                        ", child "DFID", OST-index %u, stripe-index %u, "
2377                        "infix %s, type %s: rc = %d\n",
2378                        lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
2379                        ltd->ltd_index, ea_off, infix, type, rc);
2380
2381         return rc >= 0 ? 1 : rc;
2382 }
2383
2384 static int lfsck_layout_master_conditional_destroy(const struct lu_env *env,
2385                                                    struct lfsck_component *com,
2386                                                    const struct lu_fid *fid,
2387                                                    __u32 index)
2388 {
2389         struct lfsck_thread_info *info  = lfsck_env_info(env);
2390         struct lfsck_request     *lr    = &info->lti_lr;
2391         struct lfsck_instance    *lfsck = com->lc_lfsck;
2392         struct lfsck_tgt_desc    *ltd;
2393         struct ptlrpc_request    *req;
2394         struct lfsck_request     *tmp;
2395         struct obd_export        *exp;
2396         int                       rc    = 0;
2397         ENTRY;
2398
2399         ltd = lfsck_tgt_get(&lfsck->li_ost_descs, index);
2400         if (unlikely(ltd == NULL))
2401                 RETURN(-ENXIO);
2402
2403         exp = ltd->ltd_exp;
2404         if (!(exp_connect_flags(exp) & OBD_CONNECT_LFSCK))
2405                 GOTO(put, rc = -EOPNOTSUPP);
2406
2407         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
2408         if (req == NULL)
2409                 GOTO(put, rc = -ENOMEM);
2410
2411         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
2412         if (rc != 0) {
2413                 ptlrpc_request_free(req);
2414
2415                 GOTO(put, rc);
2416         }
2417
2418         memset(lr, 0, sizeof(*lr));
2419         lr->lr_event = LE_CONDITIONAL_DESTROY;
2420         lr->lr_active = LFSCK_TYPE_LAYOUT;
2421         lr->lr_fid = *fid;
2422
2423         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
2424         *tmp = *lr;
2425         ptlrpc_request_set_replen(req);
2426
2427         rc = ptlrpc_queue_wait(req);
2428         ptlrpc_req_finished(req);
2429
2430         GOTO(put, rc);
2431
2432 put:
2433         lfsck_tgt_put(ltd);
2434
2435         return rc;
2436 }
2437
2438 static int lfsck_layout_slave_conditional_destroy(const struct lu_env *env,
2439                                                   struct lfsck_component *com,
2440                                                   struct lfsck_request *lr)
2441 {
2442         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2443         struct lu_attr                  *la     = &info->lti_la;
2444         union ldlm_policy_data          *policy = &info->lti_policy;
2445         struct ldlm_res_id              *resid  = &info->lti_resid;
2446         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2447         struct dt_device                *dev    = lfsck->li_bottom;
2448         struct lu_fid                   *fid    = &lr->lr_fid;
2449         struct dt_object                *obj;
2450         struct thandle                  *th     = NULL;
2451         struct lustre_handle             lh     = { 0 };
2452         __u64                            flags  = 0;
2453         int                              rc     = 0;
2454         ENTRY;
2455
2456         obj = lfsck_object_find_by_dev(env, dev, fid);
2457         if (IS_ERR(obj))
2458                 RETURN(PTR_ERR(obj));
2459
2460         dt_read_lock(env, obj, 0);
2461         if (dt_object_exists(obj) == 0 ||
2462             lfsck_is_dead_obj(obj)) {
2463                 dt_read_unlock(env, obj);
2464
2465                 GOTO(put, rc = -ENOENT);
2466         }
2467
2468         /* Get obj's attr without lock firstly. */
2469         rc = dt_attr_get(env, obj, la);
2470         dt_read_unlock(env, obj);
2471         if (rc != 0)
2472                 GOTO(put, rc);
2473
2474         if (likely(la->la_ctime != 0 || la->la_mode & S_ISUID))
2475                 GOTO(put, rc = -ETXTBSY);
2476
2477         /* Acquire extent lock on [0, EOF] to sync with all possible written. */
2478         LASSERT(lfsck->li_namespace != NULL);
2479
2480         memset(policy, 0, sizeof(*policy));
2481         policy->l_extent.end = OBD_OBJECT_EOF;
2482         ost_fid_build_resid(fid, resid);
2483         rc = ldlm_cli_enqueue_local(lfsck->li_namespace, resid, LDLM_EXTENT,
2484                                     policy, LCK_EX, &flags, ldlm_blocking_ast,
2485                                     ldlm_completion_ast, NULL, NULL, 0,
2486                                     LVB_T_NONE, NULL, &lh);
2487         if (rc != ELDLM_OK)
2488                 GOTO(put, rc = -EIO);
2489
2490         dt_write_lock(env, obj, 0);
2491         /* Get obj's attr within lock again. */
2492         rc = dt_attr_get(env, obj, la);
2493         if (rc != 0)
2494                 GOTO(unlock, rc);
2495
2496         if (la->la_ctime != 0)
2497                 GOTO(unlock, rc = -ETXTBSY);
2498
2499         th = dt_trans_create(env, dev);
2500         if (IS_ERR(th))
2501                 GOTO(unlock, rc = PTR_ERR(th));
2502
2503         rc = dt_declare_ref_del(env, obj, th);
2504         if (rc != 0)
2505                 GOTO(stop, rc);
2506
2507         rc = dt_declare_destroy(env, obj, th);
2508         if (rc != 0)
2509                 GOTO(stop, rc);
2510
2511         rc = dt_trans_start_local(env, dev, th);
2512         if (rc != 0)
2513                 GOTO(stop, rc);
2514
2515         rc = dt_ref_del(env, obj, th);
2516         if (rc != 0)
2517                 GOTO(stop, rc);
2518
2519         rc = dt_destroy(env, obj, th);
2520         if (rc == 0)
2521                 CDEBUG(D_LFSCK, "%s: layout LFSCK destroyed the empty "
2522                        "OST-object "DFID" that was created for reparing "
2523                        "dangling referenced case. But the original missing "
2524                        "OST-object is found now.\n",
2525                        lfsck_lfsck2name(lfsck), PFID(fid));
2526
2527         GOTO(stop, rc);
2528
2529 stop:
2530         dt_trans_stop(env, dev, th);
2531
2532 unlock:
2533         dt_write_unlock(env, obj);
2534         ldlm_lock_decref(&lh, LCK_EX);
2535
2536 put:
2537         lfsck_object_put(env, obj);
2538
2539         return rc;
2540 }
2541
2542 /**
2543  * Some OST-object has occupied the specified layout EA slot.
2544  * Such OST-object may be generated by the LFSCK when repair
2545  * dangling referenced MDT-object, which can be indicated by
2546  * attr::la_ctime == 0 but without S_ISUID in la_mode. If it
2547  * is true and such OST-object has not been modified yet, we
2548  * will replace it with the orphan OST-object; otherwise the
2549  * LFSCK will create new MDT-object to reference the orphan.
2550  *
2551  * \retval       +1: repaired
2552  * \retval        0: did nothing
2553  * \retval      -ve: on error
2554  */
2555 static int lfsck_layout_conflict_create(const struct lu_env *env,
2556                                         struct lfsck_component *com,
2557                                         struct lfsck_tgt_desc *ltd,
2558                                         struct lu_orphan_rec_v2 *rec,
2559                                         struct dt_object *parent,
2560                                         struct lu_fid *cfid,
2561                                         struct lu_buf *ea_buf,
2562                                         struct lov_mds_md_v1 *lmm,
2563                                         struct lov_ost_data_v1 *slot,
2564                                         __u32 ea_off, int lovea_size)
2565 {
2566         struct lfsck_thread_info *info          = lfsck_env_info(env);
2567         struct lu_fid            *cfid2         = &info->lti_fid2;
2568         struct ost_id            *oi            = &info->lti_oi;
2569         struct dt_device         *dev           = lfsck_obj2dev(parent);
2570         struct thandle           *th            = NULL;
2571         struct lustre_handle      lh            = { 0 };
2572         __u32                     ost_idx2      = le32_to_cpu(slot->l_ost_idx);
2573         int                       rc            = 0;
2574         ENTRY;
2575
2576         while (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val)) {
2577                 if (unlikely(!thread_is_running(&com->lc_lfsck->li_thread)))
2578                         RETURN(0);
2579         }
2580
2581         ostid_le_to_cpu(&slot->l_ost_oi, oi);
2582         rc = ostid_to_fid(cfid2, oi, ost_idx2);
2583         if (rc != 0)
2584                 GOTO(out, rc);
2585
2586         rc = lfsck_ibits_lock(env, com->lc_lfsck, parent, &lh,
2587                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
2588                               LCK_EX);
2589         if (rc != 0)
2590                 GOTO(out, rc);
2591
2592         rc = lfsck_layout_master_conditional_destroy(env, com, cfid2, ost_idx2);
2593
2594         /* If the conflict OST-obejct is not created for fixing dangling
2595          * referenced MDT-object in former LFSCK check/repair, or it has
2596          * been modified by others, then we cannot destroy it. Re-create
2597          * a new MDT-object for the orphan OST-object. */
2598         if (rc == -ETXTBSY) {
2599                 /* No need the layout lock on the original parent. */
2600                 lfsck_ibits_unlock(&lh, LCK_EX);
2601
2602                 fid_zero(&rec->lor_rec.lor_fid);
2603                 snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf),
2604                          "-"DFID"-%x", PFID(lu_object_fid(&parent->do_lu)),
2605                          ea_off);
2606                 rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
2607                                                 info->lti_tmpbuf, "C", ea_off);
2608
2609                 RETURN(rc);
2610         }
2611
2612         if (rc != 0 && rc != -ENOENT)
2613                 GOTO(unlock, rc);
2614
2615         th = dt_trans_create(env, dev);
2616         if (IS_ERR(th))
2617                 GOTO(unlock, rc = PTR_ERR(th));
2618
2619         rc = dt_declare_xattr_set(env, parent, ea_buf, XATTR_NAME_LOV,
2620                                   LU_XATTR_REPLACE, th);
2621         if (rc != 0)
2622                 GOTO(stop, rc);
2623
2624         rc = dt_trans_start_local(env, dev, th);
2625         if (rc != 0)
2626                 GOTO(stop, rc);
2627
2628         dt_write_lock(env, parent, 0);
2629         lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
2630         rc = lfsck_layout_refill_lovea(env, com->lc_lfsck, th, parent, cfid,
2631                                        ea_buf, lmm, slot, LU_XATTR_REPLACE,
2632                                        ltd->ltd_index, lovea_size);
2633         dt_write_unlock(env, parent);
2634
2635         GOTO(stop, rc);
2636
2637 stop:
2638         dt_trans_stop(env, dev, th);
2639
2640 unlock:
2641         lfsck_ibits_unlock(&lh, LCK_EX);
2642
2643 out:
2644         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant replaced the conflict "
2645                "OST-object "DFID" on the OST %x with the orphan "DFID" on "
2646                "the OST %x: parent "DFID", stripe-index %u: rc = %d\n",
2647                lfsck_lfsck2name(com->lc_lfsck), PFID(cfid2), ost_idx2,
2648                PFID(cfid), ltd->ltd_index, PFID(lfsck_dto2fid(parent)),
2649                ea_off, rc);
2650
2651         return rc >= 0 ? 1 : rc;
2652 }
2653
2654 /**
2655  * \retval       +1: repaired
2656  * \retval        0: did nothing
2657  * \retval      -ve: on error
2658  */
2659 static int lfsck_layout_recreate_lovea(const struct lu_env *env,
2660                                        struct lfsck_component *com,
2661                                        struct lfsck_tgt_desc *ltd,
2662                                        struct lu_orphan_rec_v2 *rec,
2663                                        struct dt_object *parent,
2664                                        struct lu_fid *cfid,
2665                                        __u32 ost_idx, __u32 ea_off)
2666 {
2667         struct lfsck_thread_info *info          = lfsck_env_info(env);
2668         struct lu_buf            *buf           = &info->lti_big_buf;
2669         struct lu_fid            *fid           = &info->lti_fid2;
2670         struct ost_id            *oi            = &info->lti_oi;
2671         struct lfsck_instance    *lfsck         = com->lc_lfsck;
2672         struct dt_device         *dt            = lfsck_obj2dev(parent);
2673         struct lfsck_bookmark    *bk            = &lfsck->li_bookmark_ram;
2674         struct ost_layout        *ol            = &rec->lor_layout;
2675         struct lov_comp_md_v1    *lcm           = NULL;
2676         struct lov_comp_md_entry_v1 *lcme       = NULL;
2677         struct thandle           *handle        = NULL;
2678         size_t                    lovea_size;
2679         struct lov_mds_md_v1     *lmm;
2680         struct lov_ost_data_v1   *objs;
2681         struct lustre_handle      lh            = { 0 };
2682         __u32                     magic;
2683         __u32 flags = 0;
2684         int                       fl            = 0;
2685         int                       rc            = 0;
2686         int                       rc1;
2687         int                       i;
2688         __u16                     count;
2689         bool                      locked        = false;
2690         ENTRY;
2691
2692         rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
2693                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
2694                               LCK_EX);
2695         if (rc != 0) {
2696                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant failed to recreate "
2697                        "LOV EA for "DFID": parent "DFID", OST-index %u, "
2698                        "stripe-index %u, comp_id %u, comp_start %llu, "
2699                        "comp_end %llu: rc = %d\n",
2700                        lfsck_lfsck2name(lfsck), PFID(cfid),
2701                        PFID(lfsck_dto2fid(parent)), ost_idx, ea_off,
2702                        ol->ol_comp_id, ol->ol_comp_start,
2703                        ol->ol_comp_end, rc);
2704
2705                 RETURN(rc);
2706         }
2707
2708 again:
2709         if (locked) {
2710                 dt_write_unlock(env, parent);
2711                 locked = false;
2712         }
2713
2714         if (handle != NULL) {
2715                 dt_trans_stop(env, dt, handle);
2716                 handle = NULL;
2717         }
2718
2719         if (rc < 0)
2720                 GOTO(unlock_layout, rc);
2721
2722         lovea_size = rc;
2723         if (buf->lb_len < lovea_size) {
2724                 lu_buf_realloc(buf, lovea_size);
2725                 if (buf->lb_buf == NULL)
2726                         GOTO(unlock_layout, rc = -ENOMEM);
2727         }
2728
2729         if (!(bk->lb_param & LPF_DRYRUN)) {
2730                 handle = dt_trans_create(env, dt);
2731                 if (IS_ERR(handle))
2732                         GOTO(unlock_layout, rc = PTR_ERR(handle));
2733
2734                 rc = dt_declare_xattr_set(env, parent, buf, XATTR_NAME_LOV,
2735                                           fl, handle);
2736                 if (rc != 0)
2737                         GOTO(stop, rc);
2738
2739                 rc = dt_trans_start_local(env, dt, handle);
2740                 if (rc != 0)
2741                         GOTO(stop, rc);
2742         }
2743
2744         dt_write_lock(env, parent, 0);
2745         locked = true;
2746         rc = dt_xattr_get(env, parent, buf, XATTR_NAME_LOV);
2747         if (rc == -ERANGE) {
2748                 rc = dt_xattr_get(env, parent, &LU_BUF_NULL, XATTR_NAME_LOV);
2749                 LASSERT(rc != 0);
2750                 goto again;
2751         } else if (rc == -ENODATA || rc == 0) {
2752                 lovea_size = lfsck_lovea_size(ol, ea_off);
2753                 /* If the declared is not big enough, re-try. */
2754                 if (buf->lb_len < lovea_size) {
2755                         rc = lovea_size;
2756                         goto again;
2757                 }
2758                 fl = LU_XATTR_CREATE;
2759         } else if (rc < 0) {
2760                 GOTO(unlock_parent, rc);
2761         } else if (unlikely(buf->lb_len == 0)) {
2762                 goto again;
2763         } else {
2764                 fl = LU_XATTR_REPLACE;
2765                 lovea_size = rc;
2766         }
2767
2768         if (fl == LU_XATTR_CREATE) {
2769                 if (bk->lb_param & LPF_DRYRUN)
2770                         GOTO(unlock_parent, rc = 1);
2771
2772                 LASSERT(buf->lb_len >= lovea_size);
2773
2774                 rc = lfsck_layout_update_lovea(env, lfsck, handle, ol, parent,
2775                                                cfid, buf, fl, ost_idx, ea_off);
2776
2777                 GOTO(unlock_parent, rc);
2778         }
2779
2780         lmm = buf->lb_buf;
2781         rc1 = lfsck_layout_verify_header(parent, lmm);
2782
2783         /* If the LOV EA crashed, the rebuild it. */
2784         if (rc1 == -EINVAL) {
2785                 if (bk->lb_param & LPF_DRYRUN)
2786                         GOTO(unlock_parent, rc = 1);
2787
2788                 LASSERT(buf->lb_len >= lovea_size);
2789
2790                 rc = lfsck_layout_update_lovea(env, lfsck, handle, ol, parent,
2791                                                cfid, buf, fl, ost_idx, ea_off);
2792
2793                 GOTO(unlock_parent, rc);
2794         }
2795
2796         /* For other unknown magic/pattern, keep the current LOV EA. */
2797         if (rc1 != 0)
2798                 GOTO(unlock_parent, rc = rc1);
2799
2800         magic = le32_to_cpu(lmm->lmm_magic);
2801         if (magic == LOV_MAGIC_COMP_V1) {
2802                 __u64 start;
2803                 __u64 end;
2804
2805                 lcm = buf->lb_buf;
2806                 count = le16_to_cpu(lcm->lcm_entry_count);
2807                 for (i = 0; i < count; i++) {
2808                         lcme = &lcm->lcm_entries[i];
2809                         start = le64_to_cpu(lcme->lcme_extent.e_start);
2810                         end = le64_to_cpu(lcme->lcme_extent.e_end);
2811
2812                         if (end <= ol->ol_comp_start)
2813                                 continue;
2814
2815                         if (start >= ol->ol_comp_end)
2816                                 break;
2817
2818                         lmm = buf->lb_buf + le32_to_cpu(lcme->lcme_offset);
2819                         magic = le32_to_cpu(lmm->lmm_magic);
2820                         flags = le32_to_cpu(lcme->lcme_flags);
2821                         goto further;
2822                 }
2823
2824                 rc = lfsck_layout_add_comp_comp(env, lfsck, handle, ol, parent,
2825                                                cfid, buf, ost_idx, ea_off, i);
2826
2827                 GOTO(unlock_parent, rc);
2828         }
2829
2830 further:
2831         count = le16_to_cpu(lmm->lmm_stripe_count);
2832         if (count == 0)
2833                 GOTO(unlock_parent, rc = -EINVAL);
2834         LASSERT(count > 0);
2835
2836         /* Exceed the current end of MDT-object layout EA. Then extend it. */
2837         if (count <= ea_off) {
2838                 if (bk->lb_param & LPF_DRYRUN)
2839                         GOTO(unlock_parent, rc = 1);
2840
2841                 lovea_size = lov_mds_md_size(ea_off + 1, magic);
2842                 /* If the declared is not big enough, re-try. */
2843                 if (buf->lb_len < lovea_size) {
2844                         rc = lovea_size;
2845                         goto again;
2846                 }
2847
2848                 if (lcme && !(flags & LCME_FL_INIT))
2849                         lcme->lcme_flags = cpu_to_le32(flags | LCME_FL_INIT);
2850
2851                 rc = lfsck_layout_extend_v1v3_lovea(env, lfsck, handle, ol,
2852                                         parent, cfid, buf, ost_idx, ea_off);
2853
2854                 GOTO(unlock_parent, rc);
2855         }
2856
2857         LASSERTF(rc > 0, "invalid rc = %d\n", rc);
2858
2859         if (magic == LOV_MAGIC_V1) {
2860                 objs = &lmm->lmm_objects[0];
2861         } else {
2862                 LASSERT(magic == LOV_MAGIC_V3);
2863                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
2864         }
2865
2866         for (i = 0; i < count; i++, objs++) {
2867                 /* The MDT-object was created via lfsck_layout_recover_create()
2868                  * by others before, and we fill the dummy layout EA. */
2869                 if ((lcme && !(flags & LCME_FL_INIT)) ||
2870                      lovea_slot_is_dummy(objs)) {
2871                         if (i != ea_off)
2872                                 continue;
2873
2874                         if (bk->lb_param & LPF_DRYRUN)
2875                                 GOTO(unlock_parent, rc = 1);
2876
2877                         lmm->lmm_layout_gen =
2878                             cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
2879                         if (lcme) {
2880                                 LASSERT(lcm);
2881
2882                                 if (le32_to_cpu(lmm->lmm_stripe_size) !=
2883                                         ol->ol_stripe_size ||
2884                                     le16_to_cpu(lmm->lmm_stripe_count) !=
2885                                         ol->ol_stripe_count ||
2886                                     le64_to_cpu(lcme->lcme_extent.e_start) !=
2887                                         ol->ol_comp_start ||
2888                                     le64_to_cpu(lcme->lcme_extent.e_end) !=
2889                                         ol->ol_comp_end) {
2890                                         CDEBUG(D_LFSCK, "%s: found invalid "
2891                                         "component for "DFID ": parent "DFID
2892                                         ", stripe-index %u, stripe_size %u, "
2893                                         "stripe_count %u, comp_id %u, "
2894                                         "comp_start %llu, comp_end %llu, "
2895                                         "cur_stripe_size %u, "
2896                                         "cur_stripe_count %u, "
2897                                         "cur_comp_start %llu, "
2898                                         "cur_comp_end %llu\n",
2899                                         lfsck_lfsck2name(lfsck), PFID(cfid),
2900                                         PFID(lfsck_dto2fid(parent)), ea_off,
2901                                         ol->ol_stripe_size,
2902                                         ol->ol_stripe_count, ol->ol_comp_id,
2903                                         ol->ol_comp_start, ol->ol_comp_end,
2904                                         le32_to_cpu(lmm->lmm_stripe_size),
2905                                         le16_to_cpu(lmm->lmm_stripe_count),
2906                                         le64_to_cpu(lcme->lcme_extent.e_start),
2907                                         le64_to_cpu(lcme->lcme_extent.e_end));
2908
2909                                         GOTO(unlock_parent, rc = -EINVAL);
2910                                 }
2911
2912                                 le32_add_cpu(&lcm->lcm_layout_gen, 1);
2913                                 lovea_size = le32_to_cpu(lcm->lcm_size);
2914                                 if (!(flags & LCME_FL_INIT))
2915                                         lcme->lcme_flags = cpu_to_le32(flags |
2916                                                                 LCME_FL_INIT);
2917                         }
2918
2919                         LASSERTF(buf->lb_len >= lovea_size,
2920                                  "buffer len %d is less than real size %d\n",
2921                                  (int)buf->lb_len, (int)lovea_size);
2922
2923                         rc = lfsck_layout_refill_lovea(env, lfsck, handle,
2924                                                 parent, cfid, buf, lmm, objs,
2925                                                 fl, ost_idx, lovea_size);
2926
2927                         CDEBUG(D_LFSCK, "%s layout LFSCK assistant fill "
2928                                "dummy layout slot for "DFID": parent "DFID
2929                                ", OST-index %u, stripe-index %u: rc = %d\n",
2930                                lfsck_lfsck2name(lfsck), PFID(cfid),
2931                                PFID(lfsck_dto2fid(parent)), ost_idx, i, rc);
2932
2933                         GOTO(unlock_parent, rc);
2934                 }
2935
2936                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
2937                 rc = ostid_to_fid(fid, oi, le32_to_cpu(objs->l_ost_idx));
2938                 if (rc != 0) {
2939                         CDEBUG(D_LFSCK, "%s: the parent "DFID" contains "
2940                                "invalid layout EA at the slot %d, index %u\n",
2941                                lfsck_lfsck2name(lfsck),
2942                                PFID(lfsck_dto2fid(parent)), i,
2943                                le32_to_cpu(objs->l_ost_idx));
2944
2945                         GOTO(unlock_parent, rc);
2946                 }
2947
2948                 /* It should be rare case, the slot is there, but the LFSCK
2949                  * does not handle it during the first-phase cycle scanning. */
2950                 if (unlikely(lu_fid_eq(fid, cfid))) {
2951                         if (i == ea_off) {
2952                                 GOTO(unlock_parent, rc = 0);
2953                         } else {
2954                                 /* Rare case that the OST-object index
2955                                  * does not match the parent MDT-object
2956                                  * layout EA. We trust the later one. */
2957                                 if (bk->lb_param & LPF_DRYRUN)
2958                                         GOTO(unlock_parent, rc = 1);
2959
2960                                 dt_write_unlock(env, parent);
2961                                 if (handle != NULL)
2962                                         dt_trans_stop(env, dt, handle);
2963                                 lfsck_ibits_unlock(&lh, LCK_EX);
2964                                 rc = lfsck_layout_update_pfid(env, com, parent,
2965                                                         cfid, ltd->ltd_tgt,
2966                                                         ol, i);
2967
2968                                 CDEBUG(D_LFSCK, "%s layout LFSCK assistant "
2969                                        "updated OST-object's pfid for "DFID
2970                                        ": parent "DFID", OST-index %u, "
2971                                        "stripe-index %u: rc = %d\n",
2972                                        lfsck_lfsck2name(lfsck), PFID(cfid),
2973                                        PFID(lfsck_dto2fid(parent)),
2974                                        ltd->ltd_index, i, rc);
2975
2976                                 RETURN(rc);
2977                         }
2978                 }
2979         }
2980
2981         /* The MDT-object exists, but related layout EA slot is occupied
2982          * by others. */
2983         if (bk->lb_param & LPF_DRYRUN)
2984                 GOTO(unlock_parent, rc = 1);
2985
2986         dt_write_unlock(env, parent);
2987         if (handle != NULL)
2988                 dt_trans_stop(env, dt, handle);
2989         lfsck_ibits_unlock(&lh, LCK_EX);
2990         if (magic == LOV_MAGIC_V1)
2991                 objs = &lmm->lmm_objects[ea_off];
2992         else
2993                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[ea_off];
2994         rc = lfsck_layout_conflict_create(env, com, ltd, rec, parent, cfid,
2995                                           buf, lmm, objs, ea_off, lovea_size);
2996
2997         RETURN(rc);
2998
2999 unlock_parent:
3000         if (locked)
3001                 dt_write_unlock(env, parent);
3002
3003 stop:
3004         if (handle != NULL)
3005                 dt_trans_stop(env, dt, handle);
3006
3007 unlock_layout:
3008         lfsck_ibits_unlock(&lh, LCK_EX);
3009
3010         return rc;
3011 }
3012
3013 static int lfsck_layout_scan_orphan_one(const struct lu_env *env,
3014                                         struct lfsck_component *com,
3015                                         struct lfsck_tgt_desc *ltd,
3016                                         struct lu_orphan_rec_v2 *rec,
3017                                         struct lu_fid *cfid)
3018 {
3019         struct lfsck_layout     *lo     = com->lc_file_ram;
3020         struct lu_fid           *pfid   = &rec->lor_rec.lor_fid;
3021         struct dt_object        *parent = NULL;
3022         __u32                    ea_off = pfid->f_stripe_idx;
3023         int                      rc     = 0;
3024         ENTRY;
3025
3026         if (!fid_is_sane(cfid))
3027                 GOTO(out, rc = -EINVAL);
3028
3029         pfid->f_ver = 0;
3030         if (fid_is_zero(pfid)) {
3031                 rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
3032                                                   "", "N", ea_off);
3033                 GOTO(out, rc);
3034         }
3035
3036         if (!fid_is_sane(pfid))
3037                 GOTO(out, rc = -EINVAL);
3038
3039         parent = lfsck_object_find_by_dev(env, com->lc_lfsck->li_bottom, pfid);
3040         if (IS_ERR(parent))
3041                 GOTO(out, rc = PTR_ERR(parent));
3042
3043         if (unlikely(dt_object_remote(parent) != 0))
3044                 GOTO(put, rc = -EXDEV);
3045
3046         if (dt_object_exists(parent) == 0) {
3047                 lfsck_object_put(env, parent);
3048                 rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
3049                                                   "", "R", ea_off);
3050                 GOTO(out, rc);
3051         }
3052
3053         if (!S_ISREG(lu_object_attr(&parent->do_lu)))
3054                 GOTO(put, rc = -EISDIR);
3055
3056         /* The orphan OST-object claims to be the parent's stripe, then
3057          * related dangling record in the trace file is meaningless. */
3058         rc = lfsck_layout_del_dangling_rec(env, com, pfid,
3059                                            rec->lor_layout.ol_comp_id, ea_off);
3060         if (rc && rc != -ENOENT)
3061                 GOTO(put, rc);
3062
3063         rc = lfsck_layout_recreate_lovea(env, com, ltd, rec, parent, cfid,
3064                                          ltd->ltd_index, ea_off);
3065
3066         GOTO(put, rc);
3067
3068 put:
3069         if (rc <= 0)
3070                 lfsck_object_put(env, parent);
3071         else
3072                 /* The layout EA is changed, need to be reloaded next time. */
3073                 dt_object_put_nocache(env, parent);
3074
3075 out:
3076         down_write(&com->lc_sem);
3077         com->lc_new_scanned++;
3078         com->lc_new_checked++;
3079         if (rc > 0) {
3080                 lo->ll_objs_repaired[LLIT_ORPHAN - 1]++;
3081                 rc = 0;
3082         } else if (rc < 0) {
3083                 lo->ll_objs_failed_phase2++;
3084         }
3085         up_write(&com->lc_sem);
3086
3087         return rc;
3088 }
3089
3090 static int lfsck_layout_scan_orphan(const struct lu_env *env,
3091                                     struct lfsck_component *com,
3092                                     struct lfsck_tgt_desc *ltd)
3093 {
3094         struct lfsck_assistant_data     *lad    = com->lc_data;
3095         struct lfsck_instance           *lfsck  = com->lc_lfsck;
3096         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
3097         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3098         struct ost_id                   *oi     = &info->lti_oi;
3099         struct lu_fid                   *fid    = &info->lti_fid;
3100         struct dt_object                *obj;
3101         const struct dt_it_ops          *iops;
3102         struct dt_it                    *di;
3103         int                              rc     = 0;
3104         ENTRY;
3105
3106         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant starts the orphan "
3107                "scanning for OST%04x\n",
3108                lfsck_lfsck2name(lfsck), ltd->ltd_index);
3109
3110         if (cfs_bitmap_check(lad->lad_bitmap, ltd->ltd_index)) {
3111                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant skip the orphan "
3112                        "scanning for OST%04x\n",
3113                        lfsck_lfsck2name(lfsck), ltd->ltd_index);
3114
3115                 RETURN(0);
3116         }
3117
3118         ostid_set_seq(oi, FID_SEQ_IDIF);
3119         ostid_set_id(oi, 0);
3120         rc = ostid_to_fid(fid, oi, ltd->ltd_index);
3121         if (rc != 0)
3122                 GOTO(log, rc);
3123
3124         obj = lfsck_object_find_by_dev(env, ltd->ltd_tgt, fid);
3125         if (unlikely(IS_ERR(obj)))
3126                 GOTO(log, rc = PTR_ERR(obj));
3127
3128         rc = obj->do_ops->do_index_try(env, obj,
3129                                        &dt_lfsck_layout_orphan_features);
3130         if (rc != 0)
3131                 GOTO(put, rc);
3132
3133         iops = &obj->do_index_ops->dio_it;
3134         di = iops->init(env, obj, 0);
3135         if (IS_ERR(di))
3136                 GOTO(put, rc = PTR_ERR(di));
3137
3138         rc = iops->load(env, di, 0);
3139         if (rc == -ESRCH) {
3140                 /* -ESRCH means that the orphan OST-objects rbtree has been
3141                  * cleanup because of the OSS server restart or other errors. */
3142                 lfsck_lad_set_bitmap(env, com, ltd->ltd_index);
3143                 GOTO(fini, rc);
3144         }
3145
3146         if (rc == 0)
3147                 rc = iops->next(env, di);
3148         else if (rc > 0)
3149                 rc = 0;
3150
3151         if (rc < 0)
3152                 GOTO(fini, rc);
3153
3154         if (rc > 0)
3155                 GOTO(fini, rc = 0);
3156
3157         do {
3158                 struct dt_key           *key;
3159                 struct lu_orphan_rec_v2 *rec = &info->lti_rec;
3160
3161                 if (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val) &&
3162                     unlikely(!thread_is_running(&lfsck->li_thread)))
3163                         break;
3164
3165                 key = iops->key(env, di);
3166                 com->lc_fid_latest_scanned_phase2 = *(struct lu_fid *)key;
3167                 /* Remote target OST may be runnning old LFSCK */
3168                 memset(rec, 0, sizeof(*rec));
3169                 rc = iops->rec(env, di, (struct dt_rec *)rec, 0);
3170                 if (rc == 0)
3171                         rc = lfsck_layout_scan_orphan_one(env, com, ltd, rec,
3172                                         &com->lc_fid_latest_scanned_phase2);
3173                 if (rc != 0 && bk->lb_param & LPF_FAILOUT)
3174                         GOTO(fini, rc);
3175
3176                 lfsck_control_speed_by_self(com);
3177                 do {
3178                         rc = iops->next(env, di);
3179                 } while (rc < 0 && !(bk->lb_param & LPF_FAILOUT));
3180         } while (rc == 0);
3181
3182         GOTO(fini, rc);
3183
3184 fini:
3185         iops->put(env, di);
3186         iops->fini(env, di);
3187 put:
3188         lfsck_object_put(env, obj);
3189
3190 log:
3191         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant finished the orphan "
3192                "scanning for OST%04x: rc = %d\n",
3193                lfsck_lfsck2name(lfsck), ltd->ltd_index, rc);
3194
3195         return rc > 0 ? 0 : rc;
3196 }
3197
3198 static int lfsck_lmm2layout(struct lov_mds_md_v1 *lmm, struct ost_layout *ol,
3199                             __u32 comp_id)
3200 {
3201         __u32 magic = le32_to_cpu(lmm->lmm_magic);
3202         int rc = 0;
3203         ENTRY;
3204
3205         if (magic == LOV_MAGIC_V1 || magic == LOV_MAGIC_V3) {
3206                 ol->ol_stripe_size = lmm->lmm_stripe_size;
3207                 ol->ol_stripe_count = lmm->lmm_stripe_count;
3208                 ol->ol_comp_start = 0;
3209                 ol->ol_comp_end = 0;
3210                 ol->ol_comp_id = 0;
3211         } else if (magic == LOV_MAGIC_COMP_V1) {
3212                 struct lov_comp_md_v1 *lcm = (struct lov_comp_md_v1 *)lmm;
3213                 struct lov_comp_md_entry_v1 *lcme = NULL;
3214                 __u16 count = le16_to_cpu(lcm->lcm_entry_count);
3215                 int i;
3216
3217                 for (i = 0; i < count; i++) {
3218                         lcme = &lcm->lcm_entries[i];
3219                         if (le32_to_cpu(lcme->lcme_id) == comp_id) {
3220                                 LASSERT(le32_to_cpu(lcme->lcme_flags) &
3221                                         LCME_FL_INIT);
3222
3223                                 break;
3224                         }
3225                 }
3226
3227                 /* The comp has been removed, do nothing. */
3228                 if (i == count)
3229                         GOTO(out, rc = 1);
3230
3231                 lmm = (void *)lmm + le32_to_cpu(lcme->lcme_offset);
3232                 ol->ol_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
3233                 ol->ol_stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
3234                 ol->ol_comp_start = le64_to_cpu(lcme->lcme_extent.e_start);
3235                 ol->ol_comp_end = le64_to_cpu(lcme->lcme_extent.e_end);
3236                 ol->ol_comp_id = le32_to_cpu(lcme->lcme_id);
3237         } else {
3238                 GOTO(out, rc = -EINVAL);
3239         }
3240
3241         EXIT;
3242
3243 out:
3244         return rc;
3245 }
3246
3247 /**
3248  * Repair the MDT-object with dangling LOV EA reference.
3249  *
3250  * we need to repair the inconsistency according to the users' requirement:
3251  *
3252  * 1) Keep the inconsistency there and report the inconsistency case,
3253  *    then give the chance to the application to find related issues,
3254  *    and the users can make the decision about how to handle it with
3255  *    more human knownledge. (by default)
3256  *
3257  * 2) Re-create the missing OST-object with the FID/owner information.
3258  *
3259  * \param[in] env       pointer to the thread context
3260  * \param[in] com       the layout LFSCK component
3261  * \param[in] parent    the MDT-object with dangling LOV EA reference
3262  * \param[in] child     the OST-object to be created
3263  * \param[in] comp_id   the component ID of the OST-object in the LOV EA
3264  * \param[in] ea_off    the offset of the OST-object in the LOV EA
3265  * \param[in] ost_idx   the index of OST on which the OST-object resides
3266  *
3267  * \retval              +1 for repair successfully
3268  * \retval              0 for did nothing
3269  * \retval              negative error number on failure
3270  */
3271 static int __lfsck_layout_repair_dangling(const struct lu_env *env,
3272                                           struct lfsck_component *com,
3273                                           struct dt_object *parent,
3274                                           struct dt_object *child,
3275                                           __u32 comp_id, __u32 ea_off,
3276                                           __u32 ost_idx, bool log)
3277 {
3278         struct lfsck_thread_info *info = lfsck_env_info(env);
3279         struct filter_fid *ff = &info->lti_ff;
3280         struct ost_layout *ol = &ff->ff_layout;
3281         struct dt_object_format *dof = &info->lti_dof;
3282         struct lu_attr *la = &info->lti_la;
3283         struct lfsck_instance *lfsck = com->lc_lfsck;
3284         struct dt_device *dev = lfsck_obj2dev(child);
3285         const struct lu_fid *pfid = lfsck_dto2fid(parent);
3286         const struct lu_fid *cfid = lfsck_dto2fid(child);
3287         struct lu_buf *tbuf = &info->lti_big_buf;
3288         struct thandle *handle;
3289         struct lu_buf *buf;
3290         struct lustre_handle lh = { 0 };
3291         int rc;
3292         ENTRY;
3293
3294         if (!(lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ))
3295                 GOTO(log, rc = 1);
3296
3297         rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
3298                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
3299                               LCK_EX);
3300         if (rc != 0)
3301                 GOTO(log, rc);
3302
3303         rc = dt_attr_get(env, parent, la);
3304         if (rc != 0)
3305                 GOTO(unlock1, rc);
3306
3307         la->la_mode = S_IFREG | 0666;
3308         la->la_atime = la->la_mtime = la->la_ctime = 0;
3309         la->la_valid = LA_TYPE | LA_MODE | LA_UID | LA_GID |
3310                        LA_ATIME | LA_MTIME | LA_CTIME;
3311         memset(dof, 0, sizeof(*dof));
3312         ff->ff_parent.f_seq = cpu_to_le64(pfid->f_seq);
3313         ff->ff_parent.f_oid = cpu_to_le32(pfid->f_oid);
3314         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
3315          * MDT-object's FID::f_ver, instead it is the OST-object index in its
3316          * parent MDT-object's layout EA. */
3317         ff->ff_parent.f_stripe_idx = cpu_to_le32(ea_off);
3318
3319         rc = lfsck_layout_get_lovea(env, parent, tbuf);
3320         if (rc < 0)
3321                 GOTO(unlock1, rc);
3322
3323         rc = lfsck_lmm2layout(tbuf->lb_buf, ol, comp_id);
3324         if (rc)
3325                 GOTO(unlock1, rc);
3326
3327         buf = lfsck_buf_get(env, ff, sizeof(struct filter_fid));
3328         handle = dt_trans_create(env, dev);
3329         if (IS_ERR(handle))
3330                 GOTO(unlock1, rc = PTR_ERR(handle));
3331
3332         rc = dt_declare_create(env, child, la, NULL, dof, handle);
3333         if (rc != 0)
3334                 GOTO(stop, rc);
3335
3336         rc = dt_declare_xattr_set(env, child, buf, XATTR_NAME_FID,
3337                                   LU_XATTR_CREATE, handle);
3338         if (rc != 0)
3339                 GOTO(stop, rc);
3340
3341         rc = dt_trans_start_local(env, dev, handle);
3342         if (rc != 0)
3343                 GOTO(stop, rc);
3344
3345         dt_read_lock(env, parent, 0);
3346         if (unlikely(lfsck_is_dead_obj(parent)))
3347                 GOTO(unlock2, rc = 0);
3348
3349         if (lfsck->li_bookmark_ram.lb_param & LPF_DELAY_CREATE_OSTOBJ) {
3350                 struct ost_id *oi = &info->lti_oi;
3351                 struct lu_fid *tfid = &info->lti_fid2;
3352                 struct lu_buf *lovea = &info->lti_big_buf;
3353                 struct lov_mds_md_v1 *lmm;
3354                 struct lov_ost_data_v1 *objs;
3355                 __u32 magic;
3356                 int count;
3357                 int idx2;
3358
3359                 rc = lfsck_layout_get_lovea(env, parent, lovea);
3360                 if (rc <= 0)
3361                         GOTO(unlock2, rc);
3362
3363                 lmm = lovea->lb_buf;
3364                 magic = le32_to_cpu(lmm->lmm_magic);
3365                 if (magic == LOV_MAGIC_COMP_V1) {
3366                         struct lov_comp_md_v1 *lcm = buf->lb_buf;
3367                         struct lov_comp_md_entry_v1 *lcme;
3368                         __u16 count = le16_to_cpu(lcm->lcm_entry_count);
3369                         int i;
3370
3371                         for (i = 0; i < count; i++) {
3372                                 lcme = &lcm->lcm_entries[i];
3373                                 if (le32_to_cpu(lcme->lcme_id) == comp_id) {
3374                                         LASSERT(le32_to_cpu(lcme->lcme_flags) &
3375                                                 LCME_FL_INIT);
3376
3377                                         lmm = lovea->lb_buf +
3378                                                 le32_to_cpu(lcme->lcme_offset);
3379                                         magic = le32_to_cpu(lmm->lmm_magic);
3380                                         goto check;
3381                                 }
3382                         }
3383
3384                         /* Someone removed the component, do nothing. */
3385                         GOTO(unlock2, rc = 0);
3386                 }
3387
3388 check:
3389                 count = le16_to_cpu(lmm->lmm_stripe_count);
3390                 /* Someone changed the LOV EA, do nothing. */
3391                 if (count <= ea_off)
3392                         GOTO(unlock2, rc = 0);
3393
3394                 if (magic == LOV_MAGIC_V1) {
3395                         objs = &lmm->lmm_objects[ea_off];
3396                 } else {
3397                         LASSERT(magic == LOV_MAGIC_V3);
3398
3399                         objs = &((struct lov_mds_md_v3 *)lmm)->\
3400                                                         lmm_objects[ea_off];
3401                 }
3402
3403                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
3404                 idx2 = le32_to_cpu(objs->l_ost_idx);
3405                 rc = ostid_to_fid(tfid, oi, idx2);
3406                 /* Someone changed the LOV EA, do nothing. */
3407                 if (rc != 0 || !lu_fid_eq(tfid, cfid))
3408                         GOTO(unlock2, rc);
3409         }
3410
3411         rc = dt_create(env, child, la, NULL, dof, handle);
3412         if (rc != 0)
3413                 GOTO(unlock2, rc);
3414
3415         rc = dt_xattr_set(env, child, buf, XATTR_NAME_FID, LU_XATTR_CREATE,
3416                           handle);
3417
3418         GOTO(unlock2, rc);
3419
3420 unlock2:
3421         dt_read_unlock(env, parent);
3422
3423 stop:
3424         rc = lfsck_layout_trans_stop(env, dev, handle, rc);
3425
3426 unlock1:
3427         lfsck_ibits_unlock(&lh, LCK_EX);
3428
3429 log:
3430         if (rc && log)
3431                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant found "
3432                        "dangling reference for: parent "DFID", child "
3433                        DFID", comp_id %u, ea_off %u, ost_idx %u, %s: "
3434                        "rc = %d\n",
3435                        lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
3436                        comp_id, ea_off, ost_idx,
3437                        (lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ) ?
3438                                 "Create the lost OST-object as required" :
3439                                 "Keep the MDT-object there by default", rc);
3440
3441         return rc;
3442 }
3443
3444 /**
3445  * Repair the MDT-object with dangling LOV EA reference.
3446  *
3447  * Prepare parameters and call __lfsck_layout_repair_dangling()
3448  * to repair the dangling LOV EA reference.
3449  *
3450  * \param[in] env       pointer to the thread context
3451  * \param[in] com       the layout LFSCK component
3452  * \param[in] pfid      the MDT-object's FID
3453  * \param[in] cfid      the FID for the OST-object to be created
3454  * \param[in] comp_id   the component ID of the OST-object in the LOV EA
3455  * \param[in] ea_off    the offset of the OST-object in the LOV EA
3456  * \param[in] ost_idx   the index of OST on which the OST-object resides
3457  *
3458  * \retval              +1 for repair successfully
3459  * \retval              0 for did nothing
3460  * \retval              negative error number on failure
3461  */
3462 static int lfsck_layout_repair_dangling(const struct lu_env *env,
3463                                         struct lfsck_component *com,
3464                                         const struct lu_fid *pfid,
3465                                         const struct lu_fid *cfid,
3466                                         __u32 comp_id, __u32 ea_off,
3467                                         __u32 ost_idx)
3468 {
3469         struct lfsck_instance *lfsck = com->lc_lfsck;
3470         struct dt_object *parent = NULL;
3471         struct dt_object *child = NULL;
3472         struct lfsck_tgt_desc *ltd;
3473         int rc;
3474         ENTRY;
3475
3476         parent = lfsck_object_find_bottom(env, lfsck, pfid);
3477         if (IS_ERR(parent))
3478                 GOTO(log, rc = PTR_ERR(parent));
3479
3480         /* The MDT-object has been removed. */
3481         if (dt_object_exists(parent) == 0)
3482                 GOTO(log, rc = 0);
3483
3484         ltd = lfsck_ltd2tgt(&lfsck->li_ost_descs, ost_idx);
3485         if (unlikely(ltd == NULL))
3486                 GOTO(log, rc = -ENODEV);
3487
3488         child = lfsck_object_find_by_dev(env, ltd->ltd_tgt, cfid);
3489         if (IS_ERR(child))
3490                 GOTO(log, rc = PTR_ERR(child));
3491
3492         /* The OST-object has been created. */
3493         if (unlikely(dt_object_exists(child) != 0))
3494                 GOTO(log, rc = 0);
3495
3496         rc = __lfsck_layout_repair_dangling(env, com, parent, child,
3497                                             comp_id, ea_off, ost_idx, false);
3498
3499         GOTO(log, rc);
3500
3501 log:
3502         if (child != NULL && !IS_ERR(child))
3503                 lfsck_object_put(env, child);
3504
3505         if (parent != NULL && !IS_ERR(parent))
3506                 lfsck_object_put(env, parent);
3507
3508         if (rc)
3509                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant found "
3510                        "dangling reference for: parent "DFID", child "
3511                        DFID", comp_id %u, ea_off %u, ost_idx %u, %s: rc = %d\n",
3512                        lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
3513                        comp_id, ea_off, ost_idx,
3514                        (lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ) ?
3515                                 "Create the lost OST-object as required" :
3516                                 "Keep the MDT-object there by default", rc);
3517
3518         return rc;
3519 }
3520
3521 /* If the OST-object does not recognize the MDT-object as its parent, and
3522  * there is no other MDT-object claims as its parent, then just trust the
3523  * given MDT-object as its parent. So update the OST-object filter_fid. */
3524 static int lfsck_layout_repair_unmatched_pair(const struct lu_env *env,
3525                                               struct lfsck_component *com,
3526                                               struct dt_object *parent,
3527                                               struct lfsck_layout_req *llr,
3528                                               struct lu_attr *la)
3529 {
3530         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3531         struct filter_fid               *ff     = &info->lti_ff;
3532         struct ost_layout               *ol     = &ff->ff_layout;
3533         struct dt_object                *child  = llr->llr_child;
3534         struct dt_device                *dev    = lfsck_obj2dev(child);
3535         const struct lu_fid             *tfid   = lu_object_fid(&parent->do_lu);
3536         struct lu_buf                   *tbuf   = &info->lti_big_buf;
3537         struct thandle                  *handle;
3538         struct lu_buf                   *buf;
3539         struct lustre_handle             lh     = { 0 };
3540         int                              rc;
3541         ENTRY;
3542
3543         rc = lfsck_ibits_lock(env, com->lc_lfsck, parent, &lh,
3544                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
3545                               LCK_EX);
3546         if (rc != 0)
3547                 GOTO(log, rc);
3548
3549         ff->ff_parent.f_seq = cpu_to_le64(tfid->f_seq);
3550         ff->ff_parent.f_oid = cpu_to_le32(tfid->f_oid);
3551         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
3552          * MDT-object's FID::f_ver, instead it is the OST-object index in its
3553          * parent MDT-object's layout EA. */
3554         ff->ff_parent.f_stripe_idx = cpu_to_le32(llr->llr_lov_idx);
3555
3556         rc = lfsck_layout_get_lovea(env, parent, tbuf);
3557         if (rc < 0)
3558                 GOTO(unlock1, rc);
3559
3560         rc = lfsck_lmm2layout(tbuf->lb_buf, ol, llr->llr_comp_id);
3561         if (rc)
3562                 GOTO(unlock1, rc);
3563
3564         buf = lfsck_buf_get(env, ff, sizeof(*ff));
3565
3566         handle = dt_trans_create(env, dev);
3567         if (IS_ERR(handle))
3568                 GOTO(unlock1, rc = PTR_ERR(handle));
3569
3570         rc = dt_declare_xattr_set(env, child, buf, XATTR_NAME_FID, 0, handle);
3571         if (rc != 0)
3572                 GOTO(stop, rc);
3573
3574         rc = dt_attr_get(env, parent, la);
3575         if (rc != 0)
3576                 GOTO(stop, rc);
3577
3578         la->la_valid = LA_UID | LA_GID;
3579         rc = dt_declare_attr_set(env, child, la, handle);
3580         if (rc != 0)
3581                 GOTO(stop, rc);
3582
3583         rc = dt_trans_start_local(env, dev, handle);
3584         if (rc != 0)
3585                 GOTO(stop, rc);
3586
3587         dt_write_lock(env, parent, 0);
3588         if (unlikely(lfsck_is_dead_obj(parent)))
3589                 GOTO(unlock2, rc = 1);
3590
3591         rc = dt_xattr_set(env, child, buf, XATTR_NAME_FID, 0, handle);
3592         if (rc != 0)
3593                 GOTO(unlock2, rc);
3594
3595         /* Get the latest parent's owner. */
3596         rc = dt_attr_get(env, parent, la);
3597         if (rc != 0)
3598                 GOTO(unlock2, rc);
3599
3600         la->la_valid = LA_UID | LA_GID;
3601         rc = dt_attr_set(env, child, la, handle);
3602
3603         GOTO(unlock2, rc);
3604
3605 unlock2:
3606         dt_write_unlock(env, parent);
3607
3608 stop:
3609         rc = lfsck_layout_trans_stop(env, dev, handle, rc);
3610
3611 unlock1:
3612         lfsck_ibits_unlock(&lh, LCK_EX);
3613
3614 log:
3615         if (rc)
3616                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired "
3617                        "unmatched MDT-OST pair for: parent "DFID
3618                        ", child "DFID", comp_id %u, OST-index %u, "
3619                        "stripe-index %u, owner %u/%u: rc = %d\n",
3620                        lfsck_lfsck2name(com->lc_lfsck),
3621                        PFID(lfsck_dto2fid(parent)),
3622                        PFID(lfsck_dto2fid(child)),
3623                        llr->llr_comp_id, llr->llr_ost_idx, llr->llr_lov_idx,
3624                        la->la_uid, la->la_gid, rc);
3625
3626         return rc;
3627 }
3628
3629 /* If there are more than one MDT-objects claim as the OST-object's parent,
3630  * and the OST-object only recognizes one of them, then we need to generate
3631  * new OST-object(s) with new fid(s) for the non-recognized MDT-object(s). */
3632 static int lfsck_layout_repair_multiple_references(const struct lu_env *env,
3633                                                    struct lfsck_component *com,
3634                                                    struct dt_object *parent,
3635                                                    struct lfsck_layout_req *llr,
3636                                                    struct lu_attr *la)
3637 {
3638         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3639         struct dt_allocation_hint       *hint   = &info->lti_hint;
3640         struct dt_object_format         *dof    = &info->lti_dof;
3641         struct ost_id                   *oi     = &info->lti_oi;
3642         struct lu_buf                   *buf    = &info->lti_big_buf;
3643         struct lfsck_instance           *lfsck  = com->lc_lfsck;
3644         struct dt_device                *dev;
3645         struct lu_device                *d      =
3646                                 &lfsck_obj2dev(llr->llr_child)->dd_lu_dev;
3647         struct lu_object                *o;
3648         struct lu_object                *n;
3649         struct dt_object                *child  = NULL;
3650         struct thandle                  *handle = NULL;
3651         struct lov_mds_md_v1            *lmm;
3652         struct lov_ost_data_v1          *objs;
3653         const struct lu_fid             *pfid   = lfsck_dto2fid(parent);
3654         struct lu_fid                    tfid;
3655         struct lustre_handle             lh     = { 0 };
3656         __u32                            magic;
3657         __u32                            index;
3658         int                              rc;
3659         ENTRY;
3660
3661         /* We use two separated transactions to repair the inconsistency.
3662          *
3663          * 1) create the child (OST-object).
3664          * 2) update the parent LOV EA according to the child's FID.
3665          *
3666          * If 1) succeed, but 2) failed or aborted, then such OST-object will be
3667          * handled as orphan when the layout LFSCK run next time.
3668          *
3669          * If 1) failed, but 2) succeed, then such OST-object will be re-created
3670          * as dangling referened case when the layout LFSCK run next time. */
3671
3672         /* The 1st transaction. */
3673         o = lu_object_anon(env, d, NULL);
3674         if (IS_ERR(o))
3675                 GOTO(log, rc = PTR_ERR(o));
3676
3677         n = lu_object_locate(o->lo_header, d->ld_type);
3678         if (unlikely(n == NULL)) {
3679                 lu_object_put_nocache(env, o);
3680
3681                 GOTO(log, rc = -EINVAL);
3682         }
3683
3684         child = container_of(n, struct dt_object, do_lu);
3685         memset(hint, 0, sizeof(*hint));
3686         rc = dt_attr_get(env, parent, la);
3687         if (rc != 0)
3688                 GOTO(log, rc);
3689
3690         la->la_valid = LA_UID | LA_GID;
3691         memset(dof, 0, sizeof(*dof));
3692
3693         dev = lfsck_obj2dev(child);
3694         handle = dt_trans_create(env, dev);
3695         if (IS_ERR(handle))
3696                 GOTO(log, rc = PTR_ERR(handle));
3697
3698         rc = dt_declare_create(env, child, la, hint, dof, handle);
3699         if (rc != 0)
3700                 GOTO(stop, rc);
3701
3702         rc = dt_trans_start_local(env, dev, handle);
3703         if (rc != 0)
3704                 GOTO(stop, rc);
3705
3706         rc = dt_create(env, child, la, hint, dof, handle);
3707         dt_trans_stop(env, dev, handle);
3708         handle = NULL;
3709         if (rc != 0)
3710                 GOTO(log, rc);
3711
3712         rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
3713                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
3714                               LCK_EX);
3715         if (rc != 0)
3716                 GOTO(log, rc);
3717
3718         /* The 2nd transaction. */
3719
3720         /* XXX: Generally, we should use bottom device (OSD) to update parent
3721          *      LOV EA. But because the LOD-object still references the wrong
3722          *      OSP-object that should be detached after the parent's LOV EA
3723          *      refreshed. Unfortunately, there is no suitable API for that.
3724          *      So we have to make the LOD to re-load the OSP-object(s) via
3725          *      replacing the LOV EA against the LOD-object.
3726          *
3727          *      Once the DNE2 patches have been landed, we can replace the
3728          *      LOD device with the OSD device. LU-6230. */
3729
3730         dev = lfsck->li_next;
3731         parent = lfsck_object_locate(dev, parent);
3732         if (IS_ERR(parent))
3733                 GOTO(log, rc = PTR_ERR(parent));
3734
3735         handle = dt_trans_create(env, dev);
3736         if (IS_ERR(handle))
3737                 GOTO(log, rc = PTR_ERR(handle));
3738
3739         rc = dt_declare_xattr_set(env, parent, buf, XATTR_NAME_LOV,
3740                                   LU_XATTR_REPLACE, handle);
3741         if (rc != 0)
3742                 GOTO(stop, rc);
3743
3744         rc = dt_trans_start_local(env, dev, handle);
3745         if (rc != 0)
3746                 GOTO(stop, rc);
3747
3748         dt_write_lock(env, parent, 0);
3749         if (unlikely(lfsck_is_dead_obj(parent)))
3750                 GOTO(unlock, rc = 0);
3751
3752         rc = lfsck_layout_get_lovea(env, parent, buf);
3753         if (unlikely(!rc || rc == -ENODATA))
3754                 GOTO(unlock, rc = 0);
3755
3756         lmm = buf->lb_buf;
3757         magic = le32_to_cpu(lmm->lmm_magic);
3758         if (magic == LOV_MAGIC_COMP_V1) {
3759                 struct lov_comp_md_v1 *lcm = buf->lb_buf;
3760                 struct lov_comp_md_entry_v1 *lcme;
3761                 __u16 count = le16_to_cpu(lcm->lcm_entry_count);
3762                 int i;
3763
3764                 LASSERT(llr->llr_comp_id != 0);
3765
3766                 for (i = 0; i < count; i++) {
3767                         lcme = &lcm->lcm_entries[i];
3768                         if (le32_to_cpu(lcme->lcme_id) == llr->llr_comp_id) {
3769                                 LASSERT(le32_to_cpu(lcme->lcme_flags) &
3770                                         LCME_FL_INIT);
3771
3772                                 le32_add_cpu(&lcm->lcm_layout_gen, 1);
3773                                 lmm = buf->lb_buf +
3774                                         le32_to_cpu(lcme->lcme_offset);
3775                                 magic = le32_to_cpu(lmm->lmm_magic);
3776                                 goto set;
3777                         }
3778                 }
3779
3780                 GOTO(unlock, rc = 0);
3781         }
3782
3783 set:
3784         if (magic == LOV_MAGIC_V1) {
3785                 objs = &lmm->lmm_objects[llr->llr_lov_idx];
3786         } else {
3787                 LASSERT(magic == LOV_MAGIC_V3);
3788                 objs =
3789                 &((struct lov_mds_md_v3 *)lmm)->lmm_objects[llr->llr_lov_idx];
3790         }
3791
3792         ostid_le_to_cpu(&objs->l_ost_oi, oi);
3793         index = le32_to_cpu(objs->l_ost_idx);
3794         rc = ostid_to_fid(&tfid, oi, index);
3795         /* Someone changed layout during the LFSCK, no need to repair then. */
3796         if (rc == 0 && !lu_fid_eq(&tfid, lu_object_fid(&llr->llr_child->do_lu)))
3797                 GOTO(unlock, rc = 0);
3798
3799         lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
3800         fid_to_ostid(lu_object_fid(&child->do_lu), oi);
3801         ostid_cpu_to_le(oi, &objs->l_ost_oi);
3802         objs->l_ost_gen = cpu_to_le32(0);
3803         objs->l_ost_idx = cpu_to_le32(llr->llr_ost_idx);
3804         rc = dt_xattr_set(env, parent, buf, XATTR_NAME_LOV,
3805                           LU_XATTR_REPLACE, handle);
3806
3807         GOTO(unlock, rc = (rc == 0 ? 1 : rc));
3808
3809 unlock:
3810         dt_write_unlock(env, parent);
3811
3812 stop:
3813         if (handle != NULL)
3814                 dt_trans_stop(env, dev, handle);
3815
3816 log:
3817         lfsck_ibits_unlock(&lh, LCK_EX);
3818         if (child != NULL)
3819                 lfsck_object_put(env, child);
3820
3821         if (rc)
3822                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired "
3823                        "multiple references for: parent "DFID", comp_id %u, "
3824                        "OST-index %u, stripe-index %u, owner %u/%u: rc = %d\n",
3825                        lfsck_lfsck2name(lfsck), PFID(pfid),
3826                        llr->llr_comp_id, llr->llr_ost_idx, llr->llr_lov_idx,
3827                        la->la_uid, la->la_gid, rc);
3828
3829         return rc;
3830 }
3831
3832 /* If the MDT-object and the OST-object have different owner information,
3833  * then trust the MDT-object, because the normal chown/chgrp handle order
3834  * is from MDT to OST, and it is possible that some chown/chgrp operation
3835  * is partly done. */
3836 static int lfsck_layout_repair_owner(const struct lu_env *env,
3837                                      struct lfsck_component *com,
3838                                      struct dt_object *parent,
3839                                      struct lfsck_layout_req *llr,
3840                                      struct lu_attr *pla,
3841                                      const struct lu_attr *cla)
3842 {
3843         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3844         struct lu_attr                  *tla    = &info->lti_la2;
3845         struct dt_object                *child  = llr->llr_child;
3846         struct dt_device                *dev    = lfsck_obj2dev(child);
3847         struct thandle                  *handle;
3848         int                              rc;
3849         ENTRY;
3850
3851         tla->la_uid = pla->la_uid;
3852         tla->la_gid = pla->la_gid;
3853         tla->la_valid = LA_UID | LA_GID;
3854         handle = dt_trans_create(env, dev);
3855         if (IS_ERR(handle))
3856                 GOTO(log, rc = PTR_ERR(handle));
3857
3858         rc = dt_declare_attr_set(env, child, tla, handle);
3859         if (rc != 0)
3860                 GOTO(stop, rc);
3861
3862         rc = dt_trans_start_local(env, dev, handle);
3863         if (rc != 0)
3864                 GOTO(stop, rc);
3865
3866         /* Use the dt_object lock to serialize with destroy and attr_set. */
3867         dt_read_lock(env, parent, 0);
3868         if (unlikely(lfsck_is_dead_obj(parent)))
3869                 GOTO(unlock, rc = 1);
3870
3871         /* Get the latest parent's owner. */
3872         rc = dt_attr_get(env, parent, pla);
3873         if (rc != 0)
3874                 GOTO(unlock, rc);
3875
3876         /* Some others chown/chgrp during the LFSCK, needs to do nothing. */
3877         if (unlikely(tla->la_uid != pla->la_uid ||
3878                      tla->la_gid != pla->la_gid))
3879                 rc = 1;
3880         else
3881                 rc = dt_attr_set(env, child, tla, handle);
3882
3883         GOTO(unlock, rc);
3884
3885 unlock:
3886         dt_read_unlock(env, parent);
3887
3888 stop:
3889         rc = lfsck_layout_trans_stop(env, dev, handle, rc);
3890
3891 log:
3892         if (rc != 0)
3893                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired "
3894                        "inconsistent file owner for: parent "DFID", child "DFID
3895                        ", OST-index %u, stripe-index %u, old owner %u/%u, "
3896                        "new owner %u/%u: rc = %d\n",
3897                        lfsck_lfsck2name(com->lc_lfsck),
3898                        PFID(lfsck_dto2fid(parent)), PFID(lfsck_dto2fid(child)),
3899                        llr->llr_ost_idx, llr->llr_lov_idx,
3900                        cla->la_uid, cla->la_gid, tla->la_uid, tla->la_gid, rc);
3901
3902         return rc;
3903 }
3904
3905 /* Check whether the OST-object correctly back points to the
3906  * MDT-object (@parent) via the XATTR_NAME_FID xattr (@pfid). */
3907 static int lfsck_layout_check_parent(const struct lu_env *env,
3908                                      struct lfsck_component *com,
3909                                      struct lfsck_assistant_object *lso,
3910                                      struct filter_fid *ff,
3911                                      const struct lu_fid *cfid,
3912                                      const struct lu_attr *cla,
3913                                      struct lfsck_layout_req *llr)
3914 {
3915         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3916         struct lu_buf                   *buf    = &info->lti_big_buf;
3917         struct lu_fid                   *pfid   = &info->lti_fid;
3918         struct dt_object                *tobj;
3919         struct lov_mds_md_v1            *lmm;
3920         struct lov_ost_data_v1          *objs;
3921         struct lustre_handle             lh     = { 0 };
3922         int                              rc;
3923         int                              i;
3924         __u32                            magic;
3925         __u32                            idx;
3926         __u16                            count;
3927         ENTRY;
3928
3929         *pfid = ff->ff_parent;
3930         idx = pfid->f_stripe_idx;
3931         pfid->f_ver = 0;
3932
3933         if (unlikely(!fid_is_sane(pfid)))
3934                 RETURN(LLIT_UNMATCHED_PAIR);
3935
3936         if (lu_fid_eq(pfid, &lso->lso_fid)) {
3937                 if (likely(llr->llr_lov_idx == idx))
3938                         RETURN(0);
3939
3940                 RETURN(LLIT_UNMATCHED_PAIR);
3941         }
3942
3943         tobj = lfsck_object_find_bottom(env, com->lc_lfsck, pfid);
3944         if (IS_ERR(tobj))
3945                 RETURN(PTR_ERR(tobj));
3946
3947         if (dt_object_exists(tobj) == 0 || lfsck_is_dead_obj(tobj) ||
3948             !S_ISREG(lfsck_object_type(tobj)))
3949                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
3950
3951         /* Load the tobj's layout EA, in spite of it is a local MDT-object or
3952          * remote one on another MDT. Then check whether the given OST-object
3953          * is in such layout. If yes, it is multiple referenced, otherwise it
3954          * is unmatched referenced case. */
3955         rc = lfsck_layout_get_lovea(env, tobj, buf);
3956         if (rc == 0 || rc == -ENOENT)
3957                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
3958
3959         if (rc < 0)
3960                 GOTO(out, rc);
3961
3962         lmm = buf->lb_buf;
3963         magic = le32_to_cpu(lmm->lmm_magic);
3964         if (magic == LOV_MAGIC_COMP_V1) {
3965                 struct lov_comp_md_v1 *lcm = buf->lb_buf;
3966                 struct lov_comp_md_entry_v1 *lcme;
3967
3968                 if (ff->ff_layout.ol_comp_id == 0)
3969                         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
3970
3971                 count = le16_to_cpu(lcm->lcm_entry_count);
3972                 for (i = 0; i < count; i++) {
3973                         lcme = &lcm->lcm_entries[i];
3974                         if (le32_to_cpu(lcme->lcme_id) ==
3975                             ff->ff_layout.ol_comp_id) {
3976                                 lmm = buf->lb_buf +
3977                                         le32_to_cpu(lcme->lcme_offset);
3978                                 magic = le32_to_cpu(lmm->lmm_magic);
3979                                 if (!(le32_to_cpu(lcme->lcme_flags) &
3980                                       LCME_FL_INIT))
3981                                         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
3982
3983                                 goto further;
3984                         }
3985                 }
3986
3987                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
3988         }
3989
3990 further:
3991         if (magic == LOV_MAGIC_V1) {
3992                 objs = &lmm->lmm_objects[0];
3993         } else {
3994                 LASSERT(magic == LOV_MAGIC_V3);
3995                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
3996         }
3997
3998         count = le16_to_cpu(lmm->lmm_stripe_count);
3999         for (i = 0; i < count; i++, objs++) {
4000                 struct lu_fid           *tfid   = &info->lti_fid2;
4001                 struct ost_id           *oi     = &info->lti_oi;
4002                 __u32                    idx2;
4003
4004                 if (lovea_slot_is_dummy(objs))
4005                         continue;
4006
4007                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
4008                 idx2 = le32_to_cpu(objs->l_ost_idx);
4009                 rc = ostid_to_fid(tfid, oi, idx2);
4010                 if (rc != 0) {
4011                         CDEBUG(D_LFSCK, "%s: the parent "DFID" contains "
4012                                "invalid layout EA at the slot %d, index %u\n",
4013                                lfsck_lfsck2name(com->lc_lfsck),
4014                                PFID(pfid), i, idx2);
4015
4016                         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4017                 }
4018
4019                 if (lu_fid_eq(cfid, tfid)) {
4020                         rc = lfsck_ibits_lock(env, com->lc_lfsck, tobj, &lh,
4021                                               MDS_INODELOCK_UPDATE |
4022                                               MDS_INODELOCK_LAYOUT |
4023                                               MDS_INODELOCK_XATTR,
4024                                               LCK_EX);
4025                         if (rc != 0)
4026                                 GOTO(out, rc);
4027
4028                         dt_read_lock(env, tobj, 0);
4029
4030                         /* For local MDT-object, re-check existence
4031                          * after taken the lock. */
4032                         if (!dt_object_remote(tobj)) {
4033                                 if (dt_object_exists(tobj) == 0 ||
4034                                     lfsck_is_dead_obj(tobj))
4035                                         rc = LLIT_UNMATCHED_PAIR;
4036                                 else
4037                                         rc = LLIT_MULTIPLE_REFERENCED;
4038
4039                                 GOTO(unlock, rc);
4040                         }
4041
4042                         /* For migration case, the new MDT-object and old
4043                          * MDT-object may reference the same OST-object at
4044                          * some migration internal time.
4045                          *
4046                          * For remote MDT-object, the local MDT may not know
4047                          * whether it has been removed or not.  Try checking
4048                          * for a non-existent xattr to check if this object
4049                          * has been been removed or not. */
4050                         rc = dt_xattr_get(env, tobj, &LU_BUF_NULL,
4051                                           XATTR_NAME_DUMMY);
4052                         if (unlikely(rc == -ENOENT || rc >= 0))
4053                                 rc = LLIT_UNMATCHED_PAIR;
4054                         else if (rc == -ENODATA)
4055                                 rc = LLIT_MULTIPLE_REFERENCED;
4056
4057                         GOTO(unlock, rc);
4058                 }
4059         }
4060
4061         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4062
4063 unlock:
4064         if (lustre_handle_is_used(&lh)) {
4065                 dt_read_unlock(env, tobj);
4066                 lfsck_ibits_unlock(&lh, LCK_EX);
4067         }
4068
4069 out:
4070         lfsck_object_put(env, tobj);
4071
4072         return rc;
4073 }
4074
4075 static int lfsck_layout_assistant_handler_p1(const struct lu_env *env,
4076                                              struct lfsck_component *com,
4077                                              struct lfsck_assistant_req *lar)
4078 {
4079         struct lfsck_layout_req              *llr    =
4080                         container_of0(lar, struct lfsck_layout_req, llr_lar);
4081         struct lfsck_assistant_object        *lso    = lar->lar_parent;
4082         struct lfsck_layout                  *lo     = com->lc_file_ram;
4083         struct lfsck_thread_info             *info   = lfsck_env_info(env);
4084         struct filter_fid                    *ff     = &info->lti_ff;
4085         struct lu_buf buf = { .lb_buf = ff,
4086                               .lb_len = sizeof(*ff) };
4087         struct dt_object                     *parent = NULL;
4088         struct dt_object                     *child  = llr->llr_child;
4089         struct lu_attr                       *pla    = &lso->lso_attr;
4090         struct lu_attr                       *cla    = &info->lti_la;
4091         struct lfsck_instance                *lfsck  = com->lc_lfsck;
4092         struct lfsck_bookmark                *bk     = &lfsck->li_bookmark_ram;
4093         enum lfsck_layout_inconsistency_type  type   = LLIT_NONE;
4094         int                                   rc;
4095         ENTRY;
4096
4097         if (lso->lso_dead)
4098                 RETURN(0);
4099
4100         CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_ASSISTANT_DIRECT, cfs_fail_val);
4101
4102         rc = dt_attr_get(env, child, cla);
4103         if (rc == -ENOENT) {
4104                 parent = lfsck_assistant_object_load(env, lfsck, lso);
4105                 if (IS_ERR(parent)) {
4106                         rc = PTR_ERR(parent);
4107
4108                         RETURN(rc == -ENOENT ? 0 : rc);
4109                 }
4110
4111                 type = LLIT_DANGLING;
4112                 goto repair;
4113         }
4114
4115         if (rc != 0)
4116                 GOTO(out, rc);
4117
4118         lfsck_buf_init(&buf, ff, sizeof(*ff));
4119         rc = dt_xattr_get(env, child, &buf, XATTR_NAME_FID);
4120         if (unlikely(rc > 0 && rc < sizeof(struct lu_fid))) {
4121                 type = LLIT_UNMATCHED_PAIR;
4122                 goto repair;
4123         }
4124
4125         if (rc < 0 && rc != -ENODATA)
4126                 GOTO(out, rc);
4127
4128         if (rc == 0 || rc == -ENODATA)
4129                 GOTO(check_owner, rc = 0);
4130
4131         filter_fid_le_to_cpu(ff, ff, sizeof(*ff));
4132         rc = lfsck_layout_check_parent(env, com, lso, ff,
4133                                        lu_object_fid(&child->do_lu), cla, llr);
4134         if (rc > 0) {
4135                 type = rc;
4136                 goto repair;
4137         }
4138
4139         if (rc < 0)
4140                 GOTO(out, rc);
4141
4142 check_owner:
4143         /* Someone may has changed the owner after the parent attr pre-loaded.
4144          * It can be handled later inside the lfsck_layout_repair_owner(). */
4145         if (unlikely(cla->la_uid != pla->la_uid ||
4146                      cla->la_gid != pla->la_gid)) {
4147                 type = LLIT_INCONSISTENT_OWNER;
4148                 goto repair;
4149         }
4150
4151 repair:
4152         if (type == LLIT_NONE)
4153                 GOTO(out, rc = 0);
4154
4155         if (bk->lb_param & LPF_DRYRUN)
4156                 GOTO(out, rc = 1);
4157
4158         if (parent == NULL) {
4159                 parent = lfsck_assistant_object_load(env, lfsck, lso);
4160                 if (IS_ERR(parent)) {
4161                         rc = PTR_ERR(parent);
4162
4163                         if (rc == -ENOENT)
4164                                 RETURN(0);
4165
4166                         GOTO(out, rc);
4167                 }
4168         }
4169
4170         switch (type) {
4171         case LLIT_DANGLING:
4172                 if (bk->lb_param & LPF_DELAY_CREATE_OSTOBJ)
4173                         rc = lfsck_layout_ins_dangling_rec(env, com,
4174                                 lfsck_dto2fid(parent), lfsck_dto2fid(child),
4175                                 llr->llr_comp_id, llr->llr_lov_idx,
4176                                 llr->llr_ost_idx);
4177                 else
4178                         rc = __lfsck_layout_repair_dangling(env, com, parent,
4179                                                             llr->llr_child,
4180                                                             llr->llr_comp_id,
4181                                                             llr->llr_lov_idx,
4182                                                             llr->llr_ost_idx,
4183                                                             true);
4184                 break;
4185         case LLIT_UNMATCHED_PAIR:
4186                 rc = lfsck_layout_repair_unmatched_pair(env, com, parent,
4187                                                         llr, pla);
4188                 break;
4189         case LLIT_MULTIPLE_REFERENCED:
4190                 rc = lfsck_layout_repair_multiple_references(env, com, parent,
4191                                                              llr, pla);
4192                 break;
4193         case LLIT_INCONSISTENT_OWNER:
4194                 rc = lfsck_layout_repair_owner(env, com, parent, llr, pla, cla);
4195                 break;
4196         default:
4197                 rc = 0;
4198                 break;
4199         }
4200
4201         GOTO(out, rc);
4202
4203 out:
4204         down_write(&com->lc_sem);
4205         if (rc < 0) {
4206                 struct lfsck_assistant_data *lad = com->lc_data;
4207
4208                 if (unlikely(lad->lad_exit)) {
4209                         rc = 0;
4210                 } else if (rc == -ENOTCONN || rc == -ESHUTDOWN ||
4211                            rc == -ETIMEDOUT || rc == -EHOSTDOWN ||
4212                            rc == -EHOSTUNREACH) {
4213                         /* If cannot touch the target server,
4214                          * mark the LFSCK as INCOMPLETE. */
4215                         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant fail to "
4216                                "talk with OST %x: rc = %d\n",
4217                                lfsck_lfsck2name(lfsck), llr->llr_ost_idx, rc);
4218                         lfsck_lad_set_bitmap(env, com, llr->llr_ost_idx);
4219                         lo->ll_objs_skipped++;
4220                         rc = 0;
4221                 } else {
4222                         lfsck_layout_record_failure(env, lfsck, lo);
4223                 }
4224         } else if (rc > 0 && (type != LLIT_DANGLING ||
4225                               !(bk->lb_param & LPF_DELAY_CREATE_OSTOBJ))) {
4226                 LASSERTF(type > LLIT_NONE && type <= LLIT_MAX,
4227                          "unknown type = %d\n", type);
4228
4229                 lo->ll_objs_repaired[type - 1]++;
4230                 if (bk->lb_param & LPF_DRYRUN &&
4231                     unlikely(lo->ll_pos_first_inconsistent == 0))
4232                         lo->ll_pos_first_inconsistent =
4233                         lfsck->li_obj_oit->do_index_ops->dio_it.store(env,
4234                                                         lfsck->li_di_oit);
4235         }
4236         up_write(&com->lc_sem);
4237
4238         if (parent != NULL && !IS_ERR(parent))
4239                 lfsck_object_put(env, parent);
4240
4241         return rc;
4242 }
4243
4244 static int
4245 lfsck_layout_double_scan_one_trace_file(const struct lu_env *env,
4246                                         struct lfsck_component *com,
4247                                         struct dt_object *obj, bool first)
4248 {
4249         struct lfsck_instance *lfsck = com->lc_lfsck;
4250         struct ptlrpc_thread *thread = &lfsck->li_thread;
4251         struct lfsck_bookmark *bk = &lfsck->li_bookmark_ram;
4252         struct lfsck_layout *lo = com->lc_file_ram;
4253         const struct dt_it_ops *iops = &obj->do_index_ops->dio_it;
4254         struct dt_it *di;
4255         struct dt_key *key;
4256         struct lfsck_layout_dangling_key *parent =
4257                                         &lfsck_env_info(env)->lti_lldk;
4258         struct lu_fid *cfid = &lfsck_env_info(env)->lti_fid3;
4259         __u32 ost_idx;
4260         int rc;
4261         ENTRY;
4262
4263         di = iops->init(env, obj, 0);
4264         if (IS_ERR(di))
4265                 RETURN(PTR_ERR(di));
4266
4267         if (first)
4268                 lldk_cpu_to_be(parent, &lo->ll_lldk_latest_scanned_phase2);
4269         else
4270                 memset(parent, 0, sizeof(*parent));
4271         rc = iops->get(env, di, (const struct dt_key *)parent);
4272         if (rc < 0)
4273                 GOTO(fini, rc);
4274
4275         if (first) {
4276                 /* The start one either has been processed or does not exist,
4277                  * skip it. */
4278                 rc = iops->next(env, di);
4279                 if (rc != 0)
4280                         GOTO(put, rc);
4281         }
4282
4283         do {
4284                 if (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val) &&
4285                     unlikely(!thread_is_running(thread)))
4286                         GOTO(put, rc = 0);
4287
4288                 key = iops->key(env, di);
4289                 if (IS_ERR(key)) {
4290                         rc = PTR_ERR(key);
4291                         if (rc == -ENOENT)
4292                                 GOTO(put, rc = 1);
4293
4294                         goto checkpoint;
4295                 }
4296
4297                 lldk_be_to_cpu(parent,
4298                                 (const struct lfsck_layout_dangling_key *)key);
4299                 if (!fid_is_sane(&parent->lldk_fid)) {
4300                         rc = 0;
4301                         goto checkpoint;
4302                 }
4303
4304                 rc = iops->rec(env, di, (struct dt_rec *)cfid, 0);
4305                 if (rc == 0) {
4306                         fid_be_to_cpu(cfid, cfid);
4307                         ost_idx = cfid->f_ver;
4308                         cfid->f_ver = 0;
4309                         if (!fid_is_sane(cfid)) {
4310                                 rc = 0;
4311                                 goto checkpoint;
4312                         }
4313
4314                         rc = lfsck_layout_repair_dangling(env, com,
4315                                         &parent->lldk_fid, cfid,
4316                                         parent->lldk_comp_id,
4317                                         parent->lldk_ea_off, ost_idx);
4318                 }
4319
4320 checkpoint:
4321                 down_write(&com->lc_sem);
4322                 com->lc_new_checked++;
4323                 com->lc_new_scanned++;
4324                 if (rc >= 0)
4325                         lo->ll_lldk_latest_scanned_phase2 = *parent;
4326
4327                 if (rc > 0)
4328                         lo->ll_objs_repaired[LLIT_DANGLING - 1]++;
4329                 else if (rc < 0)
4330                         lo->ll_objs_failed_phase2++;
4331                 up_write(&com->lc_sem);
4332
4333                 if (rc < 0 && bk->lb_param & LPF_FAILOUT)
4334                         GOTO(put, rc);
4335
4336                 if (unlikely(cfs_time_beforeq(com->lc_time_next_checkpoint,
4337                                               cfs_time_current())) &&
4338                     com->lc_new_checked != 0) {
4339                         down_write(&com->lc_sem);
4340                         lo->ll_run_time_phase2 +=
4341                                 cfs_duration_sec(cfs_time_current() +
4342                                 HALF_SEC - com->lc_time_last_checkpoint);
4343                         lo->ll_time_last_checkpoint = cfs_time_current_sec();
4344                         lo->ll_objs_checked_phase2 += com->lc_new_checked;
4345                         com->lc_new_checked = 0;
4346                         lfsck_layout_store(env, com);
4347                         up_write(&com->lc_sem);
4348
4349                         com->lc_time_last_checkpoint = cfs_time_current();
4350                         com->lc_time_next_checkpoint =
4351                                 com->lc_time_last_checkpoint +
4352                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
4353                 }
4354
4355                 lfsck_control_speed_by_self(com);
4356                 if (unlikely(!thread_is_running(thread)))
4357                         GOTO(put, rc = 0);
4358
4359                 rc = iops->next(env, di);
4360         } while (rc == 0);
4361
4362         GOTO(put, rc);
4363
4364 put:
4365         iops->put(env, di);
4366
4367 fini:
4368         iops->fini(env, di);
4369
4370         return rc;
4371 }
4372
4373 static int lfsck_layout_assistant_handler_p2(const struct lu_env *env,
4374                                              struct lfsck_component *com)
4375 {
4376         struct lfsck_assistant_data     *lad    = com->lc_data;
4377         struct lfsck_instance           *lfsck  = com->lc_lfsck;
4378         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
4379         struct lfsck_tgt_descs          *ltds   = &lfsck->li_ost_descs;
4380         struct lfsck_tgt_desc           *ltd;
4381         int                              rc     = 0;
4382         ENTRY;
4383
4384         CDEBUG(D_LFSCK, "%s: layout LFSCK phase2 scan start\n",
4385                lfsck_lfsck2name(lfsck));
4386
4387         spin_lock(&ltds->ltd_lock);
4388         while (!list_empty(&lad->lad_ost_phase2_list)) {
4389                 ltd = list_entry(lad->lad_ost_phase2_list.next,
4390                                  struct lfsck_tgt_desc,
4391                                  ltd_layout_phase_list);
4392                 list_del_init(&ltd->ltd_layout_phase_list);
4393                 if (bk->lb_param & LPF_OST_ORPHAN) {
4394                         spin_unlock(&ltds->ltd_lock);
4395                         rc = lfsck_layout_scan_orphan(env, com, ltd);
4396                         if (rc != 0 && bk->lb_param & LPF_FAILOUT)
4397                                 RETURN(rc);
4398
4399                         if (unlikely(lad->lad_exit ||
4400                                      !thread_is_running(&lfsck->li_thread)))
4401                                 RETURN(0);
4402                         spin_lock(&ltds->ltd_lock);
4403                 }
4404         }
4405
4406         if (list_empty(&lad->lad_ost_phase1_list))
4407                 rc = 1;
4408         else
4409                 rc = 0;
4410         spin_unlock(&ltds->ltd_lock);
4411
4412         if (rc == 1 && bk->lb_param & LPF_OST_ORPHAN) {
4413                 struct lfsck_layout *lo = com->lc_file_ram;
4414                 int i;
4415
4416                 com->lc_new_checked = 0;
4417                 com->lc_new_scanned = 0;
4418                 com->lc_time_last_checkpoint = cfs_time_current();
4419                 com->lc_time_next_checkpoint = com->lc_time_last_checkpoint +
4420                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
4421
4422                 i = lfsck_sub_trace_file_fid2idx(
4423                                 &lo->ll_lldk_latest_scanned_phase2.lldk_fid);
4424                 rc = lfsck_layout_double_scan_one_trace_file(env, com,
4425                                 com->lc_sub_trace_objs[i].lsto_obj, true);
4426                 while (rc > 0 && ++i < LFSCK_STF_COUNT)
4427                         rc = lfsck_layout_double_scan_one_trace_file(env, com,
4428                                 com->lc_sub_trace_objs[i].lsto_obj, false);
4429
4430                 CDEBUG(D_LFSCK, "%s: layout LFSCK phase2 scan dangling stop "
4431                        "at the No. %d trace file: rc = %d\n",
4432                        lfsck_lfsck2name(lfsck), i, rc);
4433         }
4434
4435         CDEBUG(D_LFSCK, "%s: layout LFSCK phase2 scan stop: rc = %d\n",
4436                lfsck_lfsck2name(lfsck), rc);
4437
4438         RETURN(rc);
4439 }
4440
4441 static int
4442 lfsck_layout_slave_async_interpret(const struct lu_env *env,
4443                                    struct ptlrpc_request *req,
4444                                    void *args, int rc)
4445 {
4446         struct lfsck_layout_slave_async_args *llsaa = args;
4447         struct obd_export                    *exp   = llsaa->llsaa_exp;
4448         struct lfsck_component               *com   = llsaa->llsaa_com;
4449         struct lfsck_layout_slave_target     *llst  = llsaa->llsaa_llst;
4450         struct lfsck_layout_slave_data       *llsd  = com->lc_data;
4451         struct lfsck_reply                   *lr    = NULL;
4452         bool                                  done  = false;
4453
4454         if (rc != 0) {
4455                 /* It is probably caused by network trouble, or target crash,
4456                  * it will try several times (depends on the obd_timeout, and
4457                  * will not less than 3 times). But to make the LFSCK can go
4458                  * ahead, we should not try for ever. After some try but still
4459                  * hit failure, it will assume that the target exit the LFSCK
4460                  * prcoessing and stop try. */
4461                 if (rc == -ENOTCONN || rc == -ESHUTDOWN) {
4462                         int max_try = max_t(int, obd_timeout / 30, 3);
4463
4464                         if (++(llst->llst_failures) > max_try)
4465                                 done = true;
4466                 } else {
4467                         done = true;
4468                 }
4469         } else {
4470                 llst->llst_failures = 0;
4471                 lr = req_capsule_server_get(&req->rq_pill, &RMF_LFSCK_REPLY);
4472                 if (lr->lr_status != LS_SCANNING_PHASE1 &&
4473                     lr->lr_status != LS_SCANNING_PHASE2)
4474                         done = true;
4475         }
4476
4477         if (done) {
4478                 CDEBUG(D_LFSCK, "%s: layout LFSCK slave gets the MDT %x "
4479                        "status %d, failures_try %d\n", lfsck_lfsck2name(com->lc_lfsck),
4480                        llst->llst_index, lr != NULL ? lr->lr_status : rc,
4481                        llst->llst_failures);
4482
4483                 lfsck_layout_llst_del(llsd, llst);
4484         }
4485
4486         lfsck_layout_llst_put(llst);
4487         lfsck_component_put(env, com);
4488         class_export_put(exp);
4489
4490         return 0;
4491 }
4492
4493 static int lfsck_layout_async_query(const struct lu_env *env,
4494                                     struct lfsck_component *com,
4495                                     struct obd_export *exp,
4496                                     struct lfsck_layout_slave_target *llst,
4497                                     struct lfsck_request *lr,
4498                                     struct ptlrpc_request_set *set)
4499 {
4500         struct lfsck_layout_slave_async_args *llsaa;
4501         struct ptlrpc_request                *req;
4502         struct lfsck_request                 *tmp;
4503         int                                   rc;
4504         ENTRY;
4505
4506         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_QUERY);
4507         if (req == NULL)
4508                 RETURN(-ENOMEM);
4509
4510         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_QUERY);
4511         if (rc != 0) {
4512                 ptlrpc_request_free(req);
4513                 RETURN(rc);
4514         }
4515
4516         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
4517         *tmp = *lr;
4518         ptlrpc_request_set_replen(req);
4519
4520         llsaa = ptlrpc_req_async_args(req);
4521         llsaa->llsaa_exp = exp;
4522         llsaa->llsaa_com = lfsck_component_get(com);
4523         llsaa->llsaa_llst = llst;
4524         req->rq_interpret_reply = lfsck_layout_slave_async_interpret;
4525         req->rq_allow_intr = 1;
4526         ptlrpc_set_add_req(set, req);
4527
4528         RETURN(0);
4529 }
4530
4531 static int lfsck_layout_async_notify(const struct lu_env *env,
4532                                      struct obd_export *exp,
4533                                      struct lfsck_request *lr,
4534                                      struct ptlrpc_request_set *set)
4535 {
4536         struct ptlrpc_request   *req;
4537         struct lfsck_request    *tmp;
4538         int                      rc;
4539         ENTRY;
4540
4541         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
4542         if (req == NULL)
4543                 RETURN(-ENOMEM);
4544
4545         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
4546         if (rc != 0) {
4547                 ptlrpc_request_free(req);
4548                 RETURN(rc);
4549         }
4550
4551         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
4552         *tmp = *lr;
4553         ptlrpc_request_set_replen(req);
4554         req->rq_allow_intr = 1;
4555         ptlrpc_set_add_req(set, req);
4556
4557         RETURN(0);
4558 }
4559
4560 static int
4561 lfsck_layout_slave_query_master(const struct lu_env *env,
4562                                 struct lfsck_component *com)
4563 {
4564         struct lfsck_request             *lr    = &lfsck_env_info(env)->lti_lr;
4565         struct lfsck_instance            *lfsck = com->lc_lfsck;
4566         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
4567         struct lfsck_layout_slave_target *llst;
4568         struct obd_export                *exp;
4569         struct ptlrpc_request_set        *set;
4570         int                               rc    = 0;
4571         int                               rc1   = 0;
4572         ENTRY;
4573
4574         set = ptlrpc_prep_set();
4575         if (set == NULL)
4576                 GOTO(log, rc = -ENOMEM);
4577
4578         memset(lr, 0, sizeof(*lr));
4579         lr->lr_event = LE_QUERY;
4580         lr->lr_active = LFSCK_TYPE_LAYOUT;
4581
4582         llsd->llsd_touch_gen++;
4583         spin_lock(&llsd->llsd_lock);
4584         while (!list_empty(&llsd->llsd_master_list)) {
4585                 llst = list_entry(llsd->llsd_master_list.next,
4586                                   struct lfsck_layout_slave_target,
4587                                   llst_list);
4588                 if (llst->llst_gen == llsd->llsd_touch_gen)
4589                         break;
4590
4591                 llst->llst_gen = llsd->llsd_touch_gen;
4592                 list_move_tail(&llst->llst_list,
4593                                &llsd->llsd_master_list);
4594                 atomic_inc(&llst->llst_ref);
4595                 spin_unlock(&llsd->llsd_lock);
4596
4597                 exp = lustre_find_lwp_by_index(lfsck->li_obd->obd_name,
4598                                                llst->llst_index);
4599                 if (exp == NULL) {
4600                         lfsck_layout_llst_del(llsd, llst);
4601                         lfsck_layout_llst_put(llst);
4602                         spin_lock(&llsd->llsd_lock);
4603                         continue;
4604                 }
4605
4606                 rc = lfsck_layout_async_query(env, com, exp, llst, lr, set);
4607                 if (rc != 0) {
4608                         CDEBUG(D_LFSCK, "%s: layout LFSCK slave fail to "
4609                                "query %s for layout: rc = %d\n",
4610                                lfsck_lfsck2name(lfsck),
4611                                exp->exp_obd->obd_name, rc);
4612
4613                         rc1 = rc;
4614                         lfsck_layout_llst_put(llst);
4615                         class_export_put(exp);
4616                 }
4617                 spin_lock(&llsd->llsd_lock);
4618         }
4619         spin_unlock(&llsd->llsd_lock);
4620
4621         rc = ptlrpc_set_wait(set);
4622         ptlrpc_set_destroy(set);
4623
4624         GOTO(log, rc = (rc1 != 0 ? rc1 : rc));
4625
4626 log:
4627         CDEBUG(D_LFSCK, "%s: layout LFSCK slave queries master: rc = %d\n",
4628                lfsck_lfsck2name(com->lc_lfsck), rc);
4629
4630         return rc;
4631 }
4632
4633 static void
4634 lfsck_layout_slave_notify_master(const struct lu_env *env,
4635                                  struct lfsck_component *com,
4636                                  enum lfsck_events event, int result)
4637 {
4638         struct lfsck_layout              *lo    = com->lc_file_ram;
4639         struct lfsck_instance            *lfsck = com->lc_lfsck;
4640         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
4641         struct lfsck_request             *lr    = &lfsck_env_info(env)->lti_lr;
4642         struct lfsck_layout_slave_target *llst;
4643         struct obd_export                *exp;
4644         struct ptlrpc_request_set        *set;
4645         int                               rc;
4646         ENTRY;
4647
4648         CDEBUG(D_LFSCK, "%s: layout LFSCK slave notifies master\n",
4649                lfsck_lfsck2name(com->lc_lfsck));
4650
4651         set = ptlrpc_prep_set();
4652         if (set == NULL)
4653                 RETURN_EXIT;
4654
4655         memset(lr, 0, sizeof(*lr));
4656         lr->lr_event = event;
4657         lr->lr_flags = LEF_FROM_OST;
4658         lr->lr_status = result;
4659         lr->lr_index = lfsck_dev_idx(lfsck);
4660         lr->lr_active = LFSCK_TYPE_LAYOUT;
4661         lr->lr_flags2 = lo->ll_flags;
4662         llsd->llsd_touch_gen++;
4663         spin_lock(&llsd->llsd_lock);
4664         while (!list_empty(&llsd->llsd_master_list)) {
4665                 llst = list_entry(llsd->llsd_master_list.next,
4666                                   struct lfsck_layout_slave_target,
4667                                   llst_list);
4668                 if (llst->llst_gen == llsd->llsd_touch_gen)
4669                         break;
4670
4671                 llst->llst_gen = llsd->llsd_touch_gen;
4672                 list_move_tail(&llst->llst_list,
4673                                &llsd->llsd_master_list);
4674                 atomic_inc(&llst->llst_ref);
4675                 spin_unlock(&llsd->llsd_lock);
4676
4677                 exp = lustre_find_lwp_by_index(lfsck->li_obd->obd_name,
4678                                                llst->llst_index);
4679                 if (exp == NULL) {
4680                         lfsck_layout_llst_del(llsd, llst);
4681                         lfsck_layout_llst_put(llst);
4682                         spin_lock(&llsd->llsd_lock);
4683                         continue;
4684                 }
4685
4686                 rc = lfsck_layout_async_notify(env, exp, lr, set);
4687                 if (rc != 0)
4688                         CDEBUG(D_LFSCK, "%s: layout LFSCK slave fail to "
4689                                "notify %s for layout: rc = %d\n",
4690                                lfsck_lfsck2name(lfsck),
4691                                exp->exp_obd->obd_name, rc);
4692
4693                 lfsck_layout_llst_put(llst);
4694                 class_export_put(exp);
4695                 spin_lock(&llsd->llsd_lock);
4696         }
4697         spin_unlock(&llsd->llsd_lock);
4698
4699         ptlrpc_set_wait(set);
4700         ptlrpc_set_destroy(set);
4701
4702         RETURN_EXIT;
4703 }
4704
4705 /*
4706  * \ret -ENODATA: unrecognized stripe
4707  * \ret = 0     : recognized stripe
4708  * \ret < 0     : other failures
4709  */
4710 static int lfsck_layout_master_check_pairs(const struct lu_env *env,
4711                                            struct lfsck_component *com,
4712                                            struct lu_fid *cfid,
4713                                            struct lu_fid *pfid, __u32 comp_id)
4714 {
4715         struct lfsck_thread_info        *info   = lfsck_env_info(env);
4716         struct lu_buf                   *buf    = &info->lti_big_buf;
4717         struct ost_id                   *oi     = &info->lti_oi;
4718         struct dt_object                *obj;
4719         struct lov_mds_md_v1            *lmm;
4720         struct lov_ost_data_v1          *objs;
4721         __u32                            idx    = pfid->f_stripe_idx;
4722         __u32                            magic;
4723         int                              rc     = 0;
4724         int                              i;
4725         __u16                            count;
4726         ENTRY;
4727
4728         pfid->f_ver = 0;
4729         obj = lfsck_object_find_bottom(env, com->lc_lfsck, pfid);
4730         if (IS_ERR(obj))
4731                 RETURN(PTR_ERR(obj));
4732
4733         dt_read_lock(env, obj, 0);
4734         if (unlikely(dt_object_exists(obj) == 0 ||
4735                      lfsck_is_dead_obj(obj)))
4736                 GOTO(unlock, rc = -ENOENT);
4737
4738         if (!S_ISREG(lfsck_object_type(obj)))
4739                 GOTO(unlock, rc = -ENODATA);
4740
4741         rc = lfsck_layout_get_lovea(env, obj, buf);
4742         if (rc < 0)
4743                 GOTO(unlock, rc);
4744
4745         if (rc == 0)
4746                 GOTO(unlock, rc = -ENODATA);
4747
4748         lmm = buf->lb_buf;
4749         magic = le32_to_cpu(lmm->lmm_magic);
4750         if (magic == LOV_MAGIC_COMP_V1) {
4751                 struct lov_comp_md_v1 *lcm = buf->lb_buf;
4752                 struct lov_comp_md_entry_v1 *lcme;
4753
4754                 if (comp_id == 0)
4755                         GOTO(unlock, rc = -ENODATA);
4756
4757                 count = le16_to_cpu(lcm->lcm_entry_count);
4758                 for (i = 0; i < count; i++) {
4759                         lcme = &lcm->lcm_entries[i];
4760                         if (le32_to_cpu(lcme->lcme_id) == comp_id) {
4761                                 lmm = buf->lb_buf +
4762                                         le32_to_cpu(lcme->lcme_offset);
4763                                 magic = le32_to_cpu(lmm->lmm_magic);
4764                                 if (!(le32_to_cpu(lcme->lcme_flags) &
4765                                       LCME_FL_INIT))
4766                                         GOTO(unlock, rc = -ENODATA);
4767
4768                                 goto further;
4769                         }
4770                 }
4771
4772                 GOTO(unlock, rc = -ENODATA);
4773         }
4774
4775 further:
4776         if (magic == LOV_MAGIC_V1) {
4777                 objs = &lmm->lmm_objects[0];
4778         } else {
4779                 LASSERT(magic == LOV_MAGIC_V3);
4780                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
4781         }
4782
4783         fid_to_ostid(cfid, oi);
4784         count = le16_to_cpu(lmm->lmm_stripe_count);
4785         for (i = 0; i < count; i++, objs++) {
4786                 struct ost_id oi2;
4787
4788                 ostid_le_to_cpu(&objs->l_ost_oi, &oi2);
4789                 if (memcmp(oi, &oi2, sizeof(*oi)) == 0)
4790                         GOTO(unlock, rc = (i != idx ? -ENODATA : 0));
4791         }
4792
4793         GOTO(unlock, rc = -ENODATA);
4794
4795 unlock:
4796         dt_read_unlock(env, obj);
4797         lfsck_object_put(env, obj);
4798
4799         return rc;
4800 }
4801
4802 /*
4803  * The LFSCK-on-OST will ask the LFSCK-on-MDT to check whether the given
4804  * MDT-object/OST-object pairs match or not to aviod transfer MDT-object
4805  * layout EA from MDT to OST. On one hand, the OST no need to understand
4806  * the layout EA structure; on the other hand, it may cause trouble when
4807  * transfer large layout EA from MDT to OST via normal OUT RPC.
4808  *
4809  * \ret > 0: unrecognized stripe
4810  * \ret = 0: recognized stripe
4811  * \ret < 0: other failures
4812  */
4813 static int lfsck_layout_slave_check_pairs(const struct lu_env *env,
4814                                           struct lfsck_component *com,
4815                                           struct lu_fid *cfid,
4816                                           struct lu_fid *pfid, __u32 comp_id)
4817 {
4818         struct lfsck_instance    *lfsck  = com->lc_lfsck;
4819         struct obd_device        *obd    = lfsck->li_obd;
4820         struct seq_server_site   *ss     = lfsck_dev_site(lfsck);
4821         struct obd_export        *exp    = NULL;
4822         struct ptlrpc_request    *req    = NULL;
4823         struct lfsck_request     *lr;
4824         struct lu_seq_range      *range  = &lfsck_env_info(env)->lti_range;
4825         int                       rc     = 0;
4826         ENTRY;
4827
4828         if (unlikely(fid_is_idif(pfid)))
4829                 RETURN(1);
4830
4831         fld_range_set_any(range);
4832         rc = fld_server_lookup(env, ss->ss_server_fld, fid_seq(pfid), range);
4833         if (rc != 0)
4834                 RETURN(rc == -ENOENT ? 1 : rc);
4835
4836         if (unlikely(!fld_range_is_mdt(range)))
4837                 RETURN(1);
4838
4839         exp = lustre_find_lwp_by_index(obd->obd_name, range->lsr_index);
4840         if (unlikely(exp == NULL))
4841                 RETURN(1);
4842
4843         if (!(exp_connect_flags(exp) & OBD_CONNECT_LFSCK))
4844                 GOTO(out, rc = -EOPNOTSUPP);
4845
4846         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
4847         if (req == NULL)
4848                 GOTO(out, rc = -ENOMEM);
4849
4850         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
4851         if (rc != 0) {
4852                 ptlrpc_request_free(req);
4853
4854                 GOTO(out, rc);
4855         }
4856
4857         lr = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
4858         memset(lr, 0, sizeof(*lr));
4859         lr->lr_event = LE_PAIRS_VERIFY;
4860         lr->lr_active = LFSCK_TYPE_LAYOUT;
4861         lr->lr_fid = *cfid; /* OST-object itself FID. */
4862         lr->lr_fid2 = *pfid; /* The claimed parent FID. */
4863         lr->lr_comp_id = comp_id;
4864
4865         ptlrpc_request_set_replen(req);
4866         rc = ptlrpc_queue_wait(req);
4867         ptlrpc_req_finished(req);
4868
4869         if (rc == -ENOENT || rc == -ENODATA)
4870                 rc = 1;
4871
4872         GOTO(out, rc);
4873
4874 out:
4875         if (exp != NULL)
4876                 class_export_put(exp);
4877
4878         return rc;
4879 }
4880
4881 static int lfsck_layout_slave_repair_pfid(const struct lu_env *env,
4882                                           struct lfsck_component *com,
4883                                           struct lfsck_req_local *lrl)
4884 {
4885         struct dt_object        *obj;
4886         int                      rc     = 0;
4887         ENTRY;
4888
4889         obj = lfsck_object_find_bottom(env, com->lc_lfsck, &lrl->lrl_fid);
4890         if (IS_ERR(obj))
4891                 GOTO(log, rc = PTR_ERR(obj));
4892
4893         dt_write_lock(env, obj, 0);
4894         if (unlikely(dt_object_exists(obj) == 0 ||
4895                      lfsck_is_dead_obj(obj)))
4896                 GOTO(unlock, rc = 0);
4897
4898         rc = __lfsck_layout_update_pfid(env, obj, &lrl->lrl_ff_client.ff_parent,
4899                                         &lrl->lrl_ff_client.ff_layout,
4900                                         lrl->lrl_ff_client.ff_parent.f_ver);
4901
4902         GOTO(unlock, rc);
4903
4904 unlock:
4905         dt_write_unlock(env, obj);
4906         lfsck_object_put(env, obj);
4907
4908 log:
4909         CDEBUG(D_LFSCK, "%s: layout LFSCK slave repaired pfid for "DFID
4910                ", parent "DFID": rc = %d\n", lfsck_lfsck2name(com->lc_lfsck),
4911                PFID(&lrl->lrl_fid), PFID(&lrl->lrl_ff_client.ff_parent), rc);
4912
4913         return rc;
4914 }
4915
4916 /* layout APIs */
4917
4918 static void lfsck_layout_slave_quit(const struct lu_env *env,
4919                                     struct lfsck_component *com);
4920
4921 static int lfsck_layout_reset(const struct lu_env *env,
4922                               struct lfsck_component *com, bool init)
4923 {
4924         struct lfsck_layout     *lo    = com->lc_file_ram;
4925         int                      rc;
4926
4927         down_write(&com->lc_sem);
4928         if (init) {
4929                 memset(lo, 0, com->lc_file_size);
4930         } else {
4931                 __u32 count = lo->ll_success_count;
4932                 __u64 last_time = lo->ll_time_last_complete;
4933
4934                 memset(lo, 0, com->lc_file_size);
4935                 lo->ll_success_count = count;
4936                 lo->ll_time_last_complete = last_time;
4937         }
4938
4939         lo->ll_magic = LFSCK_LAYOUT_MAGIC;
4940         lo->ll_status = LS_INIT;
4941
4942         if (com->lc_lfsck->li_master) {
4943                 struct lfsck_assistant_data *lad = com->lc_data;
4944
4945                 lad->lad_incomplete = 0;
4946                 CFS_RESET_BITMAP(lad->lad_bitmap);
4947         }
4948
4949         rc = lfsck_layout_store(env, com);
4950         if (rc == 0 && com->lc_lfsck->li_master)
4951                 rc = lfsck_load_sub_trace_files(env, com,
4952                         &dt_lfsck_layout_dangling_features, LFSCK_LAYOUT, true);
4953         up_write(&com->lc_sem);
4954
4955         CDEBUG(D_LFSCK, "%s: layout LFSCK reset: rc = %d\n",
4956                lfsck_lfsck2name(com->lc_lfsck), rc);
4957
4958         return rc;
4959 }
4960
4961 static void lfsck_layout_fail(const struct lu_env *env,
4962                               struct lfsck_component *com, bool new_checked)
4963 {
4964         struct lfsck_layout *lo = com->lc_file_ram;
4965
4966         down_write(&com->lc_sem);
4967         if (new_checked)
4968                 com->lc_new_checked++;
4969         lfsck_layout_record_failure(env, com->lc_lfsck, lo);
4970         up_write(&com->lc_sem);
4971 }
4972
4973 static int lfsck_layout_master_checkpoint(const struct lu_env *env,
4974                                           struct lfsck_component *com, bool init)
4975 {
4976         struct lfsck_instance   *lfsck   = com->lc_lfsck;
4977         struct lfsck_layout     *lo      = com->lc_file_ram;
4978         int                      rc;
4979
4980         if (!init) {
4981                 rc = lfsck_checkpoint_generic(env, com);
4982                 if (rc != 0)
4983                         return rc > 0 ? 0 : rc;
4984         }
4985
4986         down_write(&com->lc_sem);
4987         if (init) {
4988                 lo->ll_pos_latest_start =
4989                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
4990         } else {
4991                 lo->ll_pos_last_checkpoint =
4992                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
4993                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
4994                                 HALF_SEC - lfsck->li_time_last_checkpoint);
4995                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
4996                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
4997                 com->lc_new_checked = 0;
4998         }
4999
5000         rc = lfsck_layout_store(env, com);
5001         up_write(&com->lc_sem);
5002
5003         CDEBUG(D_LFSCK, "%s: layout LFSCK master checkpoint at the pos ["
5004                "%llu], status = %d: rc = %d\n", lfsck_lfsck2name(lfsck),
5005                lfsck->li_pos_current.lp_oit_cookie, lo->ll_status, rc);
5006
5007         return rc;
5008 }
5009
5010 static int lfsck_layout_slave_checkpoint(const struct lu_env *env,
5011                                          struct lfsck_component *com, bool init)
5012 {
5013         struct lfsck_instance   *lfsck = com->lc_lfsck;
5014         struct lfsck_layout     *lo    = com->lc_file_ram;
5015         int                      rc;
5016
5017         if (com->lc_new_checked == 0 && !init)
5018                 return 0;
5019
5020         down_write(&com->lc_sem);
5021         if (init) {
5022                 lo->ll_pos_latest_start =
5023                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5024         } else {
5025                 lo->ll_pos_last_checkpoint =
5026                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5027                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
5028                                 HALF_SEC - lfsck->li_time_last_checkpoint);
5029                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
5030                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5031                 com->lc_new_checked = 0;
5032         }
5033
5034         rc = lfsck_layout_store(env, com);
5035         up_write(&com->lc_sem);
5036
5037         CDEBUG(D_LFSCK, "%s: layout LFSCK slave checkpoint at the pos ["
5038                "%llu], status = %d: rc = %d\n", lfsck_lfsck2name(lfsck),
5039                lfsck->li_pos_current.lp_oit_cookie, lo->ll_status, rc);
5040
5041         return rc;
5042 }
5043
5044 static int lfsck_layout_prep(const struct lu_env *env,
5045                              struct lfsck_component *com,
5046                              struct lfsck_start *start)
5047 {
5048         struct lfsck_instance   *lfsck  = com->lc_lfsck;
5049         struct lfsck_layout     *lo     = com->lc_file_ram;
5050         struct lfsck_position   *pos    = &com->lc_pos_start;
5051
5052         fid_zero(&pos->lp_dir_parent);
5053         pos->lp_dir_cookie = 0;
5054         if (lo->ll_status == LS_COMPLETED ||
5055             lo->ll_status == LS_PARTIAL ||
5056             /* To handle orphan, must scan from the beginning. */
5057             (start != NULL && start->ls_flags & LPF_OST_ORPHAN)) {
5058                 int rc;
5059
5060                 rc = lfsck_layout_reset(env, com, false);
5061                 if (rc == 0)
5062                         rc = lfsck_set_param(env, lfsck, start, true);
5063
5064                 if (rc != 0) {
5065                         CDEBUG(D_LFSCK, "%s: layout LFSCK prep failed: "
5066                                "rc = %d\n", lfsck_lfsck2name(lfsck), rc);
5067
5068                         return rc;
5069                 }
5070         }
5071
5072         down_write(&com->lc_sem);
5073         lo->ll_time_latest_start = cfs_time_current_sec();
5074         spin_lock(&lfsck->li_lock);
5075         if (lo->ll_flags & LF_SCANNED_ONCE) {
5076                 if (!lfsck->li_drop_dryrun ||
5077                     lo->ll_pos_first_inconsistent == 0) {
5078                         lo->ll_status = LS_SCANNING_PHASE2;
5079                         list_move_tail(&com->lc_link,
5080                                        &lfsck->li_list_double_scan);
5081                         pos->lp_oit_cookie = 0;
5082                 } else {
5083                         int i;
5084
5085                         lo->ll_status = LS_SCANNING_PHASE1;
5086                         lo->ll_run_time_phase1 = 0;
5087                         lo->ll_run_time_phase2 = 0;
5088                         lo->ll_objs_checked_phase1 = 0;
5089                         lo->ll_objs_checked_phase2 = 0;
5090                         lo->ll_objs_failed_phase1 = 0;
5091                         lo->ll_objs_failed_phase2 = 0;
5092                         for (i = 0; i < LLIT_MAX; i++)
5093                                 lo->ll_objs_repaired[i] = 0;
5094
5095                         pos->lp_oit_cookie = lo->ll_pos_first_inconsistent;
5096                         fid_zero(&com->lc_fid_latest_scanned_phase2);
5097                 }
5098         } else {
5099                 lo->ll_status = LS_SCANNING_PHASE1;
5100                 if (!lfsck->li_drop_dryrun ||
5101                     lo->ll_pos_first_inconsistent == 0)
5102                         pos->lp_oit_cookie = lo->ll_pos_last_checkpoint + 1;
5103                 else
5104                         pos->lp_oit_cookie = lo->ll_pos_first_inconsistent;
5105         }
5106         spin_unlock(&lfsck->li_lock);
5107         up_write(&com->lc_sem);
5108
5109         return 0;
5110 }
5111
5112 static int lfsck_layout_slave_prep(const struct lu_env *env,
5113                                    struct lfsck_component *com,
5114                                    struct lfsck_start_param *lsp)
5115 {
5116         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
5117         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5118         struct lfsck_layout             *lo     = com->lc_file_ram;
5119         struct lfsck_start              *start  = lsp->lsp_start;
5120         int                              rc;
5121
5122         rc = lfsck_layout_prep(env, com, start);
5123         if (rc != 0)
5124                 return rc;
5125
5126         if (lo->ll_flags & LF_CRASHED_LASTID &&
5127             list_empty(&llsd->llsd_master_list)) {
5128                 LASSERT(lfsck->li_out_notify != NULL);
5129
5130                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
5131                                      LE_LASTID_REBUILDING);
5132         }
5133
5134         if (!lsp->lsp_index_valid)
5135                 return 0;
5136
5137         rc = lfsck_layout_llst_add(llsd, lsp->lsp_index);
5138         if (rc == 0 && start != NULL && start->ls_flags & LPF_OST_ORPHAN) {
5139                 LASSERT(!llsd->llsd_rbtree_valid);
5140
5141                 write_lock(&llsd->llsd_rb_lock);
5142                 rc = lfsck_rbtree_setup(env, com);
5143                 write_unlock(&llsd->llsd_rb_lock);
5144         }
5145
5146         CDEBUG(D_LFSCK, "%s: layout LFSCK slave prep done, start pos ["
5147                "%llu]\n", lfsck_lfsck2name(lfsck),
5148                com->lc_pos_start.lp_oit_cookie);
5149
5150         return rc;
5151 }
5152
5153 static int lfsck_layout_master_prep(const struct lu_env *env,
5154                                     struct lfsck_component *com,
5155                                     struct lfsck_start_param *lsp)
5156 {
5157         int rc;
5158         ENTRY;
5159
5160         rc = lfsck_layout_load_bitmap(env, com);
5161         if (rc != 0) {
5162                 rc = lfsck_layout_reset(env, com, false);
5163                 if (rc == 0)
5164                         rc = lfsck_set_param(env, com->lc_lfsck,
5165                                              lsp->lsp_start, true);
5166
5167                 if (rc != 0)
5168                         GOTO(log, rc);
5169         }
5170
5171         rc = lfsck_layout_prep(env, com, lsp->lsp_start);
5172         if (rc != 0)
5173                 RETURN(rc);
5174
5175         rc = lfsck_start_assistant(env, com, lsp);
5176
5177         GOTO(log, rc);
5178
5179 log:
5180         CDEBUG(D_LFSCK, "%s: layout LFSCK master prep done, start pos ["
5181                "%llu]\n", lfsck_lfsck2name(com->lc_lfsck),
5182                com->lc_pos_start.lp_oit_cookie);
5183
5184         return 0;
5185 }
5186
5187 /* Pre-fetch the attribute for each stripe in the given layout EA. */
5188 static int lfsck_layout_scan_stripes(const struct lu_env *env,
5189                                      struct lfsck_component *com,
5190                                      struct dt_object *parent,
5191                                      struct lov_mds_md_v1 *lmm, __u32 comp_id)
5192 {
5193         struct lfsck_thread_info        *info    = lfsck_env_info(env);
5194         struct lfsck_instance           *lfsck   = com->lc_lfsck;
5195         struct lfsck_bookmark           *bk      = &lfsck->li_bookmark_ram;
5196         struct lfsck_layout             *lo      = com->lc_file_ram;
5197         struct lfsck_assistant_data     *lad     = com->lc_data;
5198         struct lfsck_assistant_object   *lso     = NULL;
5199         struct lov_ost_data_v1          *objs;
5200         struct lfsck_tgt_descs          *ltds    = &lfsck->li_ost_descs;
5201         struct ptlrpc_thread            *mthread = &lfsck->li_thread;
5202         struct ptlrpc_thread            *athread = &lad->lad_thread;
5203         struct l_wait_info               lwi     = { 0 };
5204         struct lu_buf                    buf;
5205         int                              rc      = 0;
5206         int                              i;
5207         __u32                            magic;
5208         __u16                            count;
5209         ENTRY;
5210
5211         lfsck_buf_init(&buf, &info->lti_ff, sizeof(struct filter_fid));
5212         magic = le32_to_cpu(lmm->lmm_magic);
5213         if (magic == LOV_MAGIC_V1) {
5214                 objs = &lmm->lmm_objects[0];
5215         } else {
5216                 LASSERT(magic == LOV_MAGIC_V3);
5217                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
5218         }
5219
5220         count = le16_to_cpu(lmm->lmm_stripe_count);
5221         for (i = 0; i < count; i++, objs++) {
5222                 struct lu_fid           *fid    = &info->lti_fid;
5223                 struct ost_id           *oi     = &info->lti_oi;
5224                 struct lfsck_layout_req *llr;
5225                 struct lfsck_tgt_desc   *tgt    = NULL;
5226                 struct dt_object        *cobj   = NULL;
5227                 __u32                    index;
5228                 bool                     wakeup = false;
5229
5230                 if (unlikely(lovea_slot_is_dummy(objs)))
5231                         continue;
5232
5233                 l_wait_event(mthread->t_ctl_waitq,
5234                              lad->lad_prefetched < bk->lb_async_windows ||
5235                              !thread_is_running(mthread) ||
5236                              thread_is_stopped(athread),
5237                              &lwi);
5238
5239                 if (unlikely(!thread_is_running(mthread)) ||
5240                              thread_is_stopped(athread))
5241                         GOTO(out, rc = 0);
5242
5243                 if (unlikely(lfsck_is_dead_obj(parent)))
5244                         GOTO(out, rc = 0);
5245
5246                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
5247                 index = le32_to_cpu(objs->l_ost_idx);
5248                 rc = ostid_to_fid(fid, oi, index);
5249                 if (rc != 0) {
5250                         CDEBUG(D_LFSCK, "%s: get invalid layout EA for "DFID
5251                                ": "DOSTID", idx %u, comp_id %u\n",
5252                                lfsck_lfsck2name(lfsck),
5253                                PFID(lfsck_dto2fid(parent)), POSTID(oi),
5254                                index, comp_id);
5255                         goto next;
5256                 }
5257
5258                 tgt = lfsck_tgt_get(ltds, index);
5259                 if (unlikely(tgt == NULL)) {
5260                         CDEBUG(D_LFSCK, "%s: cannot talk with OST %x which "
5261                                "did not join the layout LFSCK, comp_id %u\n",
5262                                lfsck_lfsck2name(lfsck), index, comp_id);
5263                         lfsck_lad_set_bitmap(env, com, index);
5264                         goto next;
5265                 }
5266
5267                 /* There is potential deadlock race condition between object
5268                  * destroy and layout LFSCK. Consider the following scenario:
5269                  *
5270                  * 1) The LFSCK thread obtained the parent object firstly, at
5271                  *    that time, the parent object has not been destroyed yet.
5272                  *
5273                  * 2) One RPC service thread destroyed the parent and all its
5274                  *    children objects. Because the LFSCK is referencing the
5275                  *    parent object, then the parent object will be marked as
5276                  *    dying in RAM. On the other hand, the parent object is
5277                  *    referencing all its children objects, then all children
5278                  *    objects will be marked as dying in RAM also.
5279                  *
5280                  * 3) The LFSCK thread tries to find some child object with
5281                  *    the parent object referenced. Then it will find that the
5282                  *    child object is dying. According to the object visibility
5283                  *    rules: the object with dying flag cannot be returned to
5284                  *    others. So the LFSCK thread has to wait until the dying
5285                  *    object has been purged from RAM, then it can allocate a
5286                  *    new object (with the same FID) in RAM. Unfortunately, the
5287                  *    LFSCK thread itself is referencing the parent object, and
5288                  *    cause the parent object cannot be purged, then cause the
5289                  *    child object cannot be purged also. So the LFSCK thread
5290                  *    will fall into deadlock.
5291                  *
5292                  * We introduce non-blocked version lu_object_find() to allow
5293                  * the LFSCK thread to return failure immediately (instead of
5294                  * wait) when it finds dying (child) object, then the LFSCK
5295                  * thread can check whether the parent object is dying or not.
5296                  * So avoid above deadlock. LU-5395 */
5297                 cobj = lfsck_object_find_by_dev_nowait(env, tgt->ltd_tgt, fid);
5298                 if (IS_ERR(cobj)) {
5299                         if (lfsck_is_dead_obj(parent)) {
5300                                 lfsck_tgt_put(tgt);
5301
5302                                 GOTO(out, rc = 0);
5303                         }
5304
5305                         rc = PTR_ERR(cobj);
5306                         goto next;
5307                 }
5308
5309                 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_ASSISTANT_DIRECT)) {
5310                         rc = dt_declare_attr_get(env, cobj);
5311                         if (rc != 0)
5312                                 goto next;
5313
5314                         rc = dt_declare_xattr_get(env, cobj, &buf,
5315                                                   XATTR_NAME_FID);
5316                         if (rc != 0)
5317                                 goto next;
5318                 }
5319
5320                 if (lso == NULL) {
5321                         struct lu_attr *attr = &info->lti_la;
5322
5323                         rc = dt_attr_get(env, parent, attr);
5324                         if (rc != 0)
5325                                 goto next;
5326
5327                         lso = lfsck_assistant_object_init(env,
5328                                 lfsck_dto2fid(parent), attr,
5329                                 lfsck->li_pos_current.lp_oit_cookie, false);
5330                         if (IS_ERR(lso)) {
5331                                 rc = PTR_ERR(lso);
5332                                 lso = NULL;
5333
5334                                 goto next;
5335                         }
5336                 }
5337
5338                 llr = lfsck_layout_assistant_req_init(lso, cobj, comp_id,
5339                                                       index, i);
5340                 if (IS_ERR(llr)) {
5341                         rc = PTR_ERR(llr);
5342                         goto next;
5343                 }
5344
5345                 cobj = NULL;
5346                 spin_lock(&lad->lad_lock);
5347                 if (lad->lad_assistant_status < 0) {
5348                         spin_unlock(&lad->lad_lock);
5349                         lfsck_layout_assistant_req_fini(env, &llr->llr_lar);
5350                         lfsck_tgt_put(tgt);
5351                         RETURN(lad->lad_assistant_status);
5352                 }
5353
5354                 list_add_tail(&llr->llr_lar.lar_list, &lad->lad_req_list);
5355                 if (lad->lad_prefetched == 0)
5356                         wakeup = true;
5357
5358                 lad->lad_prefetched++;
5359                 spin_unlock(&lad->lad_lock);
5360                 if (wakeup)
5361                         wake_up_all(&athread->t_ctl_waitq);
5362
5363 next:
5364                 down_write(&com->lc_sem);
5365                 com->lc_new_checked++;
5366                 if (rc < 0)
5367                         lfsck_layout_record_failure(env, lfsck, lo);
5368                 up_write(&com->lc_sem);
5369
5370                 if (cobj != NULL && !IS_ERR(cobj))
5371                         lfsck_object_put(env, cobj);
5372
5373                 if (likely(tgt != NULL))
5374                         lfsck_tgt_put(tgt);
5375
5376                 if (rc < 0 && bk->lb_param & LPF_FAILOUT)
5377                         GOTO(out, rc);
5378         }
5379
5380         GOTO(out, rc = 0);
5381
5382 out:
5383         if (lso != NULL)
5384                 lfsck_assistant_object_put(env, lso);
5385
5386         return rc;
5387 }
5388
5389 /* For the given object, read its layout EA locally. For each stripe, pre-fetch
5390  * the OST-object's attribute and generate an structure lfsck_layout_req on the
5391  * list ::lad_req_list.
5392  *
5393  * For each request on above list, the lfsck_layout_assistant thread compares
5394  * the OST side attribute with local attribute, if inconsistent, then repair it.
5395  *
5396  * All above processing is async mode with pipeline. */
5397 static int lfsck_layout_master_exec_oit(const struct lu_env *env,
5398                                         struct lfsck_component *com,
5399                                         struct dt_object *obj)
5400 {
5401         struct lfsck_thread_info        *info   = lfsck_env_info(env);
5402         struct ost_id                   *oi     = &info->lti_oi;
5403         struct lfsck_layout             *lo     = com->lc_file_ram;
5404         struct lfsck_assistant_data     *lad    = com->lc_data;
5405         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5406         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
5407         struct thandle                  *handle = NULL;
5408         struct lu_buf                   *buf    = &info->lti_big_buf;
5409         struct lov_mds_md_v1            *lmm    = NULL;
5410         struct dt_device                *dev    = lfsck_obj2dev(obj);
5411         struct lustre_handle             lh     = { 0 };
5412         struct lu_buf                    ea_buf = { NULL };
5413         struct lov_comp_md_v1           *lcm    = NULL;
5414         struct lov_comp_md_entry_v1     *lcme   = NULL;
5415         int                              rc     = 0;
5416         int                              size   = 0;
5417         __u32                            magic  = 0;
5418         __u16                            count  = 0;
5419         bool                             locked = false;
5420         bool                             stripe = false;
5421         bool                             bad_oi = false;
5422         ENTRY;
5423
5424         if (!S_ISREG(lfsck_object_type(obj)))
5425                 GOTO(out, rc = 0);
5426
5427         if (lad->lad_assistant_status < 0)
5428                 GOTO(out, rc = -ESRCH);
5429
5430         fid_to_lmm_oi(lfsck_dto2fid(obj), oi);
5431         lmm_oi_cpu_to_le(oi, oi);
5432         dt_read_lock(env, obj, 0);
5433         locked = true;
5434
5435 again:
5436         bad_oi = false;
5437         if (dt_object_exists(obj) == 0 ||
5438             lfsck_is_dead_obj(obj))
5439                 GOTO(out, rc = 0);
5440
5441         rc = lfsck_layout_get_lovea(env, obj, buf);
5442         if (rc <= 0)
5443                 /* Skip bad lov EA during the 1st cycle scanning, and
5444                  * try to recover it via orphan in the 2nd scanning. */
5445                 GOTO(out, rc = (rc == -EINVAL ? 0 : rc));
5446
5447         size = rc;
5448         lmm = buf->lb_buf;
5449         magic = le32_to_cpu(lmm->lmm_magic);
5450         if (magic == LOV_MAGIC_COMP_V1) {
5451                 int i;
5452
5453                 lcm = buf->lb_buf;
5454                 count = le16_to_cpu(lcm->lcm_entry_count);
5455                 for (i = 0; i < count; i++) {
5456                         lcme = &lcm->lcm_entries[i];
5457                         lmm = buf->lb_buf + le32_to_cpu(lcme->lcme_offset);
5458                         if (memcmp(oi, &lmm->lmm_oi, sizeof(*oi)) != 0)
5459                                 goto fix;
5460                 }
5461
5462                 GOTO(out, stripe = true);
5463         } else if (memcmp(oi, &lmm->lmm_oi, sizeof(*oi)) == 0) {
5464                 GOTO(out, stripe = true);
5465         }
5466
5467 fix:
5468         /* Inconsistent lmm_oi, should be repaired. */
5469         bad_oi = true;
5470
5471         if (bk->lb_param & LPF_DRYRUN) {
5472                 lo->ll_objs_repaired[LLIT_OTHERS - 1]++;
5473
5474                 GOTO(out, stripe = true);
5475         }
5476
5477         if (!lustre_handle_is_used(&lh)) {
5478                 dt_read_unlock(env, obj);
5479                 locked = false;
5480                 rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
5481                                       MDS_INODELOCK_LAYOUT |
5482                                       MDS_INODELOCK_XATTR, LCK_EX);
5483                 if (rc != 0)
5484                         GOTO(out, rc);
5485
5486                 handle = dt_trans_create(env, dev);
5487                 if (IS_ERR(handle))
5488                         GOTO(out, rc = PTR_ERR(handle));
5489
5490                 lfsck_buf_init(&ea_buf, lmm, size);
5491                 rc = dt_declare_xattr_set(env, obj, &ea_buf, XATTR_NAME_LOV,
5492                                           LU_XATTR_REPLACE, handle);
5493                 if (rc != 0)
5494                         GOTO(out, rc);
5495
5496                 rc = dt_trans_start_local(env, dev, handle);
5497                 if (rc != 0)
5498                         GOTO(out, rc);
5499
5500                 dt_write_lock(env, obj, 0);
5501                 locked = true;
5502
5503                 goto again;
5504         }
5505
5506         if (magic == LOV_MAGIC_COMP_V1) {
5507                 int i;
5508
5509                 for (i = 0; i < count; i++) {
5510                         lcme = &lcm->lcm_entries[i];
5511                         lmm = buf->lb_buf + le32_to_cpu(lcme->lcme_offset);
5512                         lmm->lmm_oi = *oi;
5513                 }
5514         } else {
5515                 lmm->lmm_oi = *oi;
5516         }
5517
5518         rc = dt_xattr_set(env, obj, &ea_buf, XATTR_NAME_LOV,
5519                           LU_XATTR_REPLACE, handle);
5520         if (rc != 0)
5521                 GOTO(out, rc);
5522
5523         lo->ll_objs_repaired[LLIT_OTHERS - 1]++;
5524
5525         GOTO(out, stripe = true);
5526
5527 out:
5528         if (locked) {
5529                 if (lustre_handle_is_used(&lh))
5530                         dt_write_unlock(env, obj);
5531                 else
5532                         dt_read_unlock(env, obj);
5533         }
5534
5535         if (handle != NULL && !IS_ERR(handle))
5536                 dt_trans_stop(env, dev, handle);
5537
5538         lfsck_ibits_unlock(&lh, LCK_EX);
5539
5540         if (bad_oi)
5541                 CDEBUG(D_LFSCK, "%s: layout LFSCK master %s bad lmm_oi for "
5542                        DFID": rc = %d\n", lfsck_lfsck2name(lfsck),
5543                        bk->lb_param & LPF_DRYRUN ? "found" : "repaired",
5544                        PFID(lfsck_dto2fid(obj)), rc);
5545
5546         if (stripe) {
5547                 if (magic == LOV_MAGIC_COMP_V1) {
5548                         int i;
5549
5550                         for (i = 0; i < count; i++) {
5551                                 lcme = &lcm->lcm_entries[i];
5552                                 if (!(le32_to_cpu(lcme->lcme_flags) &
5553                                       LCME_FL_INIT))
5554                                         continue;
5555
5556                                 rc = lfsck_layout_scan_stripes(env, com, obj,
5557                                         (struct lov_mds_md_v1 *)(buf->lb_buf +
5558                                         le32_to_cpu(lcme->lcme_offset)),
5559                                         le32_to_cpu(lcme->lcme_id));
5560                         }
5561                 } else {
5562                         rc = lfsck_layout_scan_stripes(env, com, obj, lmm, 0);
5563                 }
5564         } else {
5565                 down_write(&com->lc_sem);
5566                 com->lc_new_checked++;
5567                 if (rc < 0)
5568                         lfsck_layout_record_failure(env, lfsck, lo);
5569                 up_write(&com->lc_sem);
5570         }
5571
5572         return rc;
5573 }
5574
5575 static int lfsck_layout_slave_exec_oit(const struct lu_env *env,
5576                                        struct lfsck_component *com,
5577                                        struct dt_object *obj)
5578 {
5579         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5580         struct lfsck_layout             *lo     = com->lc_file_ram;
5581         const struct lu_fid             *fid    = lfsck_dto2fid(obj);
5582         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
5583         struct lfsck_layout_seq         *lls;
5584         __u64                            seq;
5585         __u64                            oid;
5586         int                              rc;
5587         ENTRY;
5588
5589         LASSERT(llsd != NULL);
5590
5591         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DELAY5) &&
5592             cfs_fail_val == lfsck_dev_idx(lfsck)) {
5593                 struct l_wait_info       lwi = LWI_TIMEOUT(cfs_time_seconds(1),
5594                                                            NULL, NULL);
5595                 struct ptlrpc_thread    *thread = &lfsck->li_thread;
5596
5597                 l_wait_event(thread->t_ctl_waitq,
5598                              !thread_is_running(thread),
5599                              &lwi);
5600         }
5601
5602         lfsck_rbtree_update_bitmap(env, com, fid, false);
5603
5604         down_write(&com->lc_sem);
5605         if (fid_is_idif(fid))
5606                 seq = 0;
5607         else if (!fid_is_norm(fid) ||
5608                  !fid_is_for_ostobj(env, lfsck, obj, fid))
5609                 GOTO(unlock, rc = 0);
5610         else
5611                 seq = fid_seq(fid);
5612         com->lc_new_checked++;
5613
5614         lls = lfsck_layout_seq_lookup(llsd, seq);
5615         if (lls == NULL) {
5616                 OBD_ALLOC_PTR(lls);
5617                 if (unlikely(lls == NULL))
5618                         GOTO(unlock, rc = -ENOMEM);
5619
5620                 INIT_LIST_HEAD(&lls->lls_list);
5621                 lls->lls_seq = seq;
5622                 rc = lfsck_layout_lastid_load(env, com, lls);
5623                 if (rc != 0) {
5624                         CDEBUG(D_LFSCK, "%s: layout LFSCK failed to "
5625                               "load LAST_ID for %#llx: rc = %d\n",
5626                               lfsck_lfsck2name(com->lc_lfsck), seq, rc);
5627                         lo->ll_objs_failed_phase1++;
5628                         OBD_FREE_PTR(lls);
5629                         GOTO(unlock, rc);
5630                 }
5631
5632                 lfsck_layout_seq_insert(llsd, lls);
5633         }
5634
5635         if (unlikely(fid_is_last_id(fid)))
5636                 GOTO(unlock, rc = 0);
5637
5638         if (fid_is_idif(fid))
5639                 oid = fid_idif_id(fid_seq(fid), fid_oid(fid), fid_ver(fid));
5640         else
5641                 oid = fid_oid(fid);
5642
5643         if (oid > lls->lls_lastid_known)
5644                 lls->lls_lastid_known = oid;
5645
5646         if (oid > lls->lls_lastid) {
5647                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
5648                         /* OFD may create new objects during LFSCK scanning. */
5649                         rc = lfsck_layout_lastid_reload(env, com, lls);
5650                         if (unlikely(rc != 0)) {
5651                                 CDEBUG(D_LFSCK, "%s: layout LFSCK failed to "
5652                                       "reload LAST_ID for %#llx: rc = %d\n",
5653                                       lfsck_lfsck2name(com->lc_lfsck),
5654                                       lls->lls_seq, rc);
5655
5656                                 GOTO(unlock, rc);
5657                         }
5658
5659                         if (oid <= lls->lls_lastid ||
5660                             lo->ll_flags & LF_CRASHED_LASTID)
5661                                 GOTO(unlock, rc = 0);
5662
5663                         LASSERT(lfsck->li_out_notify != NULL);
5664
5665                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
5666                                              LE_LASTID_REBUILDING);
5667                         lo->ll_flags |= LF_CRASHED_LASTID;
5668
5669                         CDEBUG(D_LFSCK, "%s: layout LFSCK finds crashed "
5670                                "LAST_ID file (2) for the sequence %#llx"
5671                                ", old value %llu, known value %llu\n",
5672                                lfsck_lfsck2name(lfsck), lls->lls_seq,
5673                                lls->lls_lastid, oid);
5674                 }
5675
5676                 lls->lls_lastid = oid;
5677                 lls->lls_dirty = 1;
5678         }
5679
5680         GOTO(unlock, rc = 0);
5681
5682 unlock:
5683         up_write(&com->lc_sem);
5684
5685         return rc;
5686 }
5687
5688 static int lfsck_layout_exec_dir(const struct lu_env *env,
5689                                  struct lfsck_component *com,
5690                                  struct lfsck_assistant_object *lso,
5691                                  struct lu_dirent *ent, __u16 type)
5692 {
5693         return 0;
5694 }
5695
5696 static int lfsck_layout_master_post(const struct lu_env *env,
5697                                     struct lfsck_component *com,
5698                                     int result, bool init)
5699 {
5700         struct lfsck_instance   *lfsck  = com->lc_lfsck;
5701         struct lfsck_layout     *lo     = com->lc_file_ram;
5702         int                      rc;
5703         ENTRY;
5704
5705         lfsck_post_generic(env, com, &result);
5706
5707         down_write(&com->lc_sem);
5708         spin_lock(&lfsck->li_lock);
5709         if (!init)
5710                 lo->ll_pos_last_checkpoint =
5711                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5712
5713         if (result > 0) {
5714                 if (lo->ll_flags & LF_INCOMPLETE)
5715                         lo->ll_status = LS_PARTIAL;
5716                 else
5717                         lo->ll_status = LS_SCANNING_PHASE2;
5718                 lo->ll_flags |= LF_SCANNED_ONCE;
5719                 lo->ll_flags &= ~LF_UPGRADE;
5720                 list_move_tail(&com->lc_link, &lfsck->li_list_double_scan);
5721         } else if (result == 0) {
5722                 if (lfsck->li_status != 0)
5723                         lo->ll_status = lfsck->li_status;
5724                 else
5725                         lo->ll_status = LS_STOPPED;
5726                 if (lo->ll_status != LS_PAUSED)
5727                         list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5728         } else {
5729                 lo->ll_status = LS_FAILED;
5730                 list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5731         }
5732         spin_unlock(&lfsck->li_lock);
5733
5734         if (!init) {
5735                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
5736                                 HALF_SEC - lfsck->li_time_last_checkpoint);
5737                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
5738                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5739                 com->lc_new_checked = 0;
5740         }
5741
5742         rc = lfsck_layout_store(env, com);
5743         up_write(&com->lc_sem);
5744
5745         CDEBUG(D_LFSCK, "%s: layout LFSCK master post done: rc = %d\n",
5746                lfsck_lfsck2name(lfsck), rc);
5747
5748         RETURN(rc);
5749 }
5750
5751 static int lfsck_layout_slave_post(const struct lu_env *env,
5752                                    struct lfsck_component *com,
5753                                    int result, bool init)
5754 {
5755         struct lfsck_instance   *lfsck = com->lc_lfsck;
5756         struct lfsck_layout     *lo    = com->lc_file_ram;
5757         int                      rc;
5758         bool                     done  = false;
5759
5760         down_write(&com->lc_sem);
5761         rc = lfsck_layout_lastid_store(env, com);
5762         if (rc != 0)
5763                 result = rc;
5764
5765         LASSERT(lfsck->li_out_notify != NULL);
5766
5767         spin_lock(&lfsck->li_lock);
5768         if (!init)
5769                 lo->ll_pos_last_checkpoint =
5770                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5771
5772         if (result > 0) {
5773                 lo->ll_status = LS_SCANNING_PHASE2;
5774                 lo->ll_flags |= LF_SCANNED_ONCE;
5775                 if (lo->ll_flags & LF_CRASHED_LASTID) {
5776                         done = true;
5777                         lo->ll_flags &= ~LF_CRASHED_LASTID;
5778
5779                         CDEBUG(D_LFSCK, "%s: layout LFSCK has rebuilt "
5780                                "crashed LAST_ID files successfully\n",
5781                                lfsck_lfsck2name(lfsck));
5782                 }
5783                 lo->ll_flags &= ~LF_UPGRADE;
5784                 list_move_tail(&com->lc_link, &lfsck->li_list_double_scan);
5785         } else if (result == 0) {
5786                 if (lfsck->li_status != 0)
5787                         lo->ll_status = lfsck->li_status;
5788                 else
5789                         lo->ll_status = LS_STOPPED;
5790                 if (lo->ll_status != LS_PAUSED)
5791                         list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5792         } else {
5793                 lo->ll_status = LS_FAILED;
5794                 list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5795         }
5796         spin_unlock(&lfsck->li_lock);
5797
5798         if (done)
5799                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
5800                                      LE_LASTID_REBUILT);
5801
5802         if (!init) {
5803                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
5804                                 HALF_SEC - lfsck->li_time_last_checkpoint);
5805                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
5806                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5807                 com->lc_new_checked = 0;
5808         }
5809
5810         rc = lfsck_layout_store(env, com);
5811         up_write(&com->lc_sem);
5812
5813         lfsck_layout_slave_notify_master(env, com, LE_PHASE1_DONE, result);
5814
5815         CDEBUG(D_LFSCK, "%s: layout LFSCK slave post done: rc = %d\n",
5816                lfsck_lfsck2name(lfsck), rc);
5817
5818         return rc;
5819 }
5820
5821 static void lfsck_layout_dump(const struct lu_env *env,
5822                               struct lfsck_component *com, struct seq_file *m)
5823 {
5824         struct lfsck_instance   *lfsck = com->lc_lfsck;
5825         struct lfsck_bookmark   *bk    = &lfsck->li_bookmark_ram;
5826         struct lfsck_layout     *lo    = com->lc_file_ram;
5827
5828         down_read(&com->lc_sem);
5829         seq_printf(m, "name: lfsck_layout\n"
5830                    "magic: %#x\n"
5831                    "version: %d\n"
5832                    "status: %s\n",
5833                    lo->ll_magic,
5834                    bk->lb_version,
5835                    lfsck_status2name(lo->ll_status));
5836
5837         lfsck_bits_dump(m, lo->ll_flags, lfsck_flags_names, "flags");
5838
5839         lfsck_bits_dump(m, bk->lb_param, lfsck_param_names, "param");
5840
5841         lfsck_time_dump(m, lo->ll_time_last_complete, "last_completed");
5842
5843         lfsck_time_dump(m, lo->ll_time_latest_start, "latest_start");
5844
5845         lfsck_time_dump(m, lo->ll_time_last_checkpoint, "last_checkpoint");
5846
5847         seq_printf(m, "latest_start_position: %llu\n"
5848                    "last_checkpoint_position: %llu\n"
5849                    "first_failure_position: %llu\n",
5850                    lo->ll_pos_latest_start,
5851                    lo->ll_pos_last_checkpoint,
5852                    lo->ll_pos_first_inconsistent);
5853
5854         seq_printf(m, "success_count: %u\n"
5855                    "repaired_dangling: %llu\n"
5856                    "repaired_unmatched_pair: %llu\n"
5857                    "repaired_multiple_referenced: %llu\n"
5858                    "repaired_orphan: %llu\n"
5859                    "repaired_inconsistent_owner: %llu\n"
5860                    "repaired_others: %llu\n"
5861                    "skipped: %llu\n"
5862                    "failed_phase1: %llu\n"
5863                    "failed_phase2: %llu\n",
5864                    lo->ll_success_count,
5865                    lo->ll_objs_repaired[LLIT_DANGLING - 1],
5866                    lo->ll_objs_repaired[LLIT_UNMATCHED_PAIR - 1],
5867                    lo->ll_objs_repaired[LLIT_MULTIPLE_REFERENCED - 1],
5868                    lo->ll_objs_repaired[LLIT_ORPHAN - 1],
5869                    lo->ll_objs_repaired[LLIT_INCONSISTENT_OWNER - 1],
5870                    lo->ll_objs_repaired[LLIT_OTHERS - 1],
5871                    lo->ll_objs_skipped,
5872                    lo->ll_objs_failed_phase1,
5873                    lo->ll_objs_failed_phase2);
5874
5875         if (lo->ll_status == LS_SCANNING_PHASE1) {
5876                 __u64 pos;
5877                 cfs_duration_t duration = cfs_time_current() -
5878                                           lfsck->li_time_last_checkpoint;
5879                 __u64 checked = lo->ll_objs_checked_phase1 +
5880                                 com->lc_new_checked;
5881                 __u64 speed = checked;
5882                 __u64 new_checked = com->lc_new_checked *
5883                                     msecs_to_jiffies(MSEC_PER_SEC);
5884                 __u32 rtime = lo->ll_run_time_phase1 +
5885                               cfs_duration_sec(duration + HALF_SEC);
5886
5887                 if (duration != 0)
5888                         do_div(new_checked, duration);
5889                 if (rtime != 0)
5890                         do_div(speed, rtime);
5891                 seq_printf(m, "checked_phase1: %llu\n"
5892                            "checked_phase2: %llu\n"
5893                            "run_time_phase1: %u seconds\n"
5894                            "run_time_phase2: %u seconds\n"
5895                            "average_speed_phase1: %llu items/sec\n"
5896                            "average_speed_phase2: N/A\n"
5897                            "real-time_speed_phase1: %llu items/sec\n"
5898                            "real-time_speed_phase2: N/A\n",
5899                            checked,
5900                            lo->ll_objs_checked_phase2,
5901                            rtime,
5902                            lo->ll_run_time_phase2,
5903                            speed,
5904                            new_checked);
5905
5906                 if (likely(lfsck->li_di_oit)) {
5907                         const struct dt_it_ops *iops =
5908                                 &lfsck->li_obj_oit->do_index_ops->dio_it;
5909
5910                         /* The low layer otable-based iteration position may NOT
5911                          * exactly match the layout-based directory traversal
5912                          * cookie. Generally, it is not a serious issue. But the
5913                          * caller should NOT make assumption on that. */
5914                         pos = iops->store(env, lfsck->li_di_oit);
5915                         if (!lfsck->li_current_oit_processed)
5916                                 pos--;
5917                 } else {
5918                         pos = lo->ll_pos_last_checkpoint;
5919                 }
5920
5921                 seq_printf(m, "current_position: %llu\n", pos);
5922         } else if (lo->ll_status == LS_SCANNING_PHASE2) {
5923                 cfs_duration_t duration = cfs_time_current() -
5924                                           com->lc_time_last_checkpoint;
5925                 __u64 checked = lo->ll_objs_checked_phase2 +
5926                                 com->lc_new_checked;
5927                 __u64 speed1 = lo->ll_objs_checked_phase1;
5928                 __u64 speed2 = checked;
5929                 __u64 new_checked = com->lc_new_checked *
5930                                     msecs_to_jiffies(MSEC_PER_SEC);
5931                 __u32 rtime = lo->ll_run_time_phase2 +
5932                               cfs_duration_sec(duration + HALF_SEC);
5933
5934                 if (duration != 0)
5935                         do_div(new_checked, duration);
5936                 if (lo->ll_run_time_phase1 != 0)
5937                         do_div(speed1, lo->ll_run_time_phase1);
5938                 if (rtime != 0)
5939                         do_div(speed2, rtime);
5940                 seq_printf(m, "checked_phase1: %llu\n"
5941                            "checked_phase2: %llu\n"
5942                            "run_time_phase1: %u seconds\n"
5943                            "run_time_phase2: %u seconds\n"
5944                            "average_speed_phase1: %llu items/sec\n"
5945                            "average_speed_phase2: %llu items/sec\n"
5946                            "real-time_speed_phase1: N/A\n"
5947                            "real-time_speed_phase2: %llu items/sec\n"
5948                            "current_position: "DFID"\n",
5949                            lo->ll_objs_checked_phase1,
5950                            checked,
5951                            lo->ll_run_time_phase1,
5952                            rtime,
5953                            speed1,
5954                            speed2,
5955                            new_checked,
5956                            PFID(&com->lc_fid_latest_scanned_phase2));
5957         } else {
5958                 __u64 speed1 = lo->ll_objs_checked_phase1;
5959                 __u64 speed2 = lo->ll_objs_checked_phase2;
5960
5961                 if (lo->ll_run_time_phase1 != 0)
5962                         do_div(speed1, lo->ll_run_time_phase1);
5963                 if (lo->ll_run_time_phase2 != 0)
5964                         do_div(speed2, lo->ll_run_time_phase2);
5965                 seq_printf(m, "checked_phase1: %llu\n"
5966                            "checked_phase2: %llu\n"
5967                            "run_time_phase1: %u seconds\n"
5968                            "run_time_phase2: %u seconds\n"
5969                            "average_speed_phase1: %llu items/sec\n"
5970                            "average_speed_phase2: %llu objs/sec\n"
5971                            "real-time_speed_phase1: N/A\n"
5972                            "real-time_speed_phase2: N/A\n"
5973                            "current_position: N/A\n",
5974                            lo->ll_objs_checked_phase1,
5975                            lo->ll_objs_checked_phase2,
5976                            lo->ll_run_time_phase1,
5977                            lo->ll_run_time_phase2,
5978                            speed1,
5979                            speed2);
5980         }
5981
5982         up_read(&com->lc_sem);
5983 }
5984
5985 static int lfsck_layout_master_double_scan(const struct lu_env *env,
5986                                            struct lfsck_component *com)
5987 {
5988         struct lfsck_layout             *lo     = com->lc_file_ram;
5989         struct lfsck_assistant_data     *lad    = com->lc_data;
5990         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5991         struct lfsck_tgt_descs          *ltds;
5992         struct lfsck_tgt_desc           *ltd;
5993         struct lfsck_tgt_desc           *next;
5994         int                              rc;
5995
5996         rc = lfsck_double_scan_generic(env, com, lo->ll_status);
5997
5998         if (thread_is_stopped(&lad->lad_thread)) {
5999                 LASSERT(list_empty(&lad->lad_req_list));
6000                 LASSERT(list_empty(&lad->lad_ost_phase1_list));
6001                 LASSERT(list_empty(&lad->lad_mdt_phase1_list));
6002
6003                 ltds = &lfsck->li_ost_descs;
6004                 spin_lock(&ltds->ltd_lock);
6005                 list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase2_list,
6006                                          ltd_layout_phase_list) {
6007                         list_del_init(&ltd->ltd_layout_phase_list);
6008                 }
6009                 spin_unlock(&ltds->ltd_lock);
6010
6011                 ltds = &lfsck->li_mdt_descs;
6012                 spin_lock(&ltds->ltd_lock);
6013                 list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase2_list,
6014                                          ltd_layout_phase_list) {
6015                         list_del_init(&ltd->ltd_layout_phase_list);
6016                 }
6017                 spin_unlock(&ltds->ltd_lock);
6018         }
6019
6020         return rc;
6021 }
6022
6023 static int lfsck_layout_slave_double_scan(const struct lu_env *env,
6024                                           struct lfsck_component *com)
6025 {
6026         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6027         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
6028         struct lfsck_layout             *lo     = com->lc_file_ram;
6029         struct ptlrpc_thread            *thread = &lfsck->li_thread;
6030         int                              rc;
6031         ENTRY;
6032
6033         CDEBUG(D_LFSCK, "%s: layout LFSCK slave phase2 scan start\n",
6034                lfsck_lfsck2name(lfsck));
6035
6036         atomic_inc(&lfsck->li_double_scan_count);
6037
6038         if (lo->ll_flags & LF_INCOMPLETE)
6039                 GOTO(done, rc = 1);
6040
6041         com->lc_new_checked = 0;
6042         com->lc_new_scanned = 0;
6043         com->lc_time_last_checkpoint = cfs_time_current();
6044         com->lc_time_next_checkpoint = com->lc_time_last_checkpoint +
6045                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
6046
6047         while (1) {
6048                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(30),
6049                                                      NULL, NULL);
6050
6051                 rc = lfsck_layout_slave_query_master(env, com);
6052                 if (list_empty(&llsd->llsd_master_list)) {
6053                         if (unlikely(!thread_is_running(thread)))
6054                                 rc = 0;
6055                         else
6056                                 rc = 1;
6057
6058                         GOTO(done, rc);
6059                 }
6060
6061                 if (rc < 0)
6062                         GOTO(done, rc);
6063
6064                 rc = l_wait_event(thread->t_ctl_waitq,
6065                                   !thread_is_running(thread) ||
6066                                   lo->ll_flags & LF_INCOMPLETE ||
6067                                   list_empty(&llsd->llsd_master_list),
6068                                   &lwi);
6069                 if (unlikely(!thread_is_running(thread)))
6070                         GOTO(done, rc = 0);
6071
6072                 if (lo->ll_flags & LF_INCOMPLETE)
6073                         GOTO(done, rc = 1);
6074
6075                 if (rc == -ETIMEDOUT)
6076                         continue;
6077
6078                 GOTO(done, rc = (rc < 0 ? rc : 1));
6079         }
6080
6081 done:
6082         rc = lfsck_layout_double_scan_result(env, com, rc);
6083         lfsck_layout_slave_notify_master(env, com, LE_PHASE2_DONE,
6084                         (rc > 0 && lo->ll_flags & LF_INCOMPLETE) ? 0 : rc);
6085         lfsck_layout_slave_quit(env, com);
6086         if (atomic_dec_and_test(&lfsck->li_double_scan_count))
6087                 wake_up_all(&lfsck->li_thread.t_ctl_waitq);
6088
6089         CDEBUG(D_LFSCK, "%s: layout LFSCK slave phase2 scan finished, "
6090                "status %d: rc = %d\n",
6091                lfsck_lfsck2name(lfsck), lo->ll_status, rc);
6092
6093         return rc;
6094 }
6095
6096 static void lfsck_layout_master_data_release(const struct lu_env *env,
6097                                              struct lfsck_component *com)
6098 {
6099         struct lfsck_assistant_data     *lad    = com->lc_data;
6100         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6101         struct lfsck_tgt_descs          *ltds;
6102         struct lfsck_tgt_desc           *ltd;
6103         struct lfsck_tgt_desc           *next;
6104
6105         LASSERT(lad != NULL);
6106         LASSERT(thread_is_init(&lad->lad_thread) ||
6107                 thread_is_stopped(&lad->lad_thread));
6108         LASSERT(list_empty(&lad->lad_req_list));
6109
6110         com->lc_data = NULL;
6111
6112         ltds = &lfsck->li_ost_descs;
6113         spin_lock(&ltds->ltd_lock);
6114         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase1_list,
6115                                  ltd_layout_phase_list) {
6116                 list_del_init(&ltd->ltd_layout_phase_list);
6117         }
6118         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase2_list,
6119                                  ltd_layout_phase_list) {
6120                 list_del_init(&ltd->ltd_layout_phase_list);
6121         }
6122         list_for_each_entry_safe(ltd, next, &lad->lad_ost_list,
6123                                  ltd_layout_list) {
6124                 list_del_init(&ltd->ltd_layout_list);
6125         }
6126         spin_unlock(&ltds->ltd_lock);
6127
6128         ltds = &lfsck->li_mdt_descs;
6129         spin_lock(&ltds->ltd_lock);
6130         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase1_list,
6131                                  ltd_layout_phase_list) {
6132                 list_del_init(&ltd->ltd_layout_phase_list);
6133         }
6134         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase2_list,
6135                                  ltd_layout_phase_list) {
6136                 list_del_init(&ltd->ltd_layout_phase_list);
6137         }
6138         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_list,
6139                                  ltd_layout_list) {
6140                 list_del_init(&ltd->ltd_layout_list);
6141         }
6142         spin_unlock(&ltds->ltd_lock);
6143
6144         if (likely(lad->lad_bitmap != NULL))
6145                 CFS_FREE_BITMAP(lad->lad_bitmap);
6146
6147         OBD_FREE_PTR(lad);
6148 }
6149
6150 static void lfsck_layout_slave_data_release(const struct lu_env *env,
6151                                             struct lfsck_component *com)
6152 {
6153         struct lfsck_layout_slave_data *llsd = com->lc_data;
6154
6155         lfsck_layout_slave_quit(env, com);
6156         com->lc_data = NULL;
6157         OBD_FREE_PTR(llsd);
6158 }
6159
6160 static void lfsck_layout_master_quit(const struct lu_env *env,
6161                                      struct lfsck_component *com)
6162 {
6163         struct lfsck_assistant_data     *lad    = com->lc_data;
6164         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6165         struct lfsck_tgt_descs          *ltds;
6166         struct lfsck_tgt_desc           *ltd;
6167         struct lfsck_tgt_desc           *next;
6168
6169         LASSERT(lad != NULL);
6170
6171         lfsck_quit_generic(env, com);
6172
6173         LASSERT(thread_is_init(&lad->lad_thread) ||
6174                 thread_is_stopped(&lad->lad_thread));
6175         LASSERT(list_empty(&lad->lad_req_list));
6176
6177         ltds = &lfsck->li_ost_descs;
6178         spin_lock(&ltds->ltd_lock);
6179         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase1_list,
6180                                  ltd_layout_phase_list) {
6181                 list_del_init(&ltd->ltd_layout_phase_list);
6182         }
6183         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase2_list,
6184                                  ltd_layout_phase_list) {
6185                 list_del_init(&ltd->ltd_layout_phase_list);
6186         }
6187         spin_unlock(&ltds->ltd_lock);
6188
6189         ltds = &lfsck->li_mdt_descs;
6190         spin_lock(&ltds->ltd_lock);
6191         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase1_list,
6192                                  ltd_layout_phase_list) {
6193                 list_del_init(&ltd->ltd_layout_phase_list);
6194         }
6195         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase2_list,
6196                                  ltd_layout_phase_list) {
6197                 list_del_init(&ltd->ltd_layout_phase_list);
6198         }
6199         spin_unlock(&ltds->ltd_lock);
6200 }
6201
6202 static void lfsck_layout_slave_quit(const struct lu_env *env,
6203                                     struct lfsck_component *com)
6204 {
6205         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
6206         struct lfsck_layout_seq          *lls;
6207         struct lfsck_layout_seq          *next;
6208         struct lfsck_layout_slave_target *llst;
6209
6210         LASSERT(llsd != NULL);
6211
6212         down_write(&com->lc_sem);
6213         list_for_each_entry_safe(lls, next, &llsd->llsd_seq_list,
6214                                  lls_list) {
6215                 list_del_init(&lls->lls_list);
6216                 lfsck_object_put(env, lls->lls_lastid_obj);
6217                 OBD_FREE_PTR(lls);
6218         }
6219         up_write(&com->lc_sem);
6220
6221         spin_lock(&llsd->llsd_lock);
6222         while (!list_empty(&llsd->llsd_master_list)) {
6223                 llst = list_entry(llsd->llsd_master_list.next,
6224                                   struct lfsck_layout_slave_target, llst_list);
6225                 list_del_init(&llst->llst_list);
6226                 spin_unlock(&llsd->llsd_lock);
6227                 lfsck_layout_llst_put(llst);
6228                 spin_lock(&llsd->llsd_lock);
6229         }
6230         spin_unlock(&llsd->llsd_lock);
6231
6232         lfsck_rbtree_cleanup(env, com);
6233 }
6234
6235 static int lfsck_layout_master_in_notify(const struct lu_env *env,
6236                                          struct lfsck_component *com,
6237                                          struct lfsck_request *lr)
6238 {
6239         struct lfsck_instance           *lfsck = com->lc_lfsck;
6240         struct lfsck_layout             *lo    = com->lc_file_ram;
6241         struct lfsck_assistant_data     *lad   = com->lc_data;
6242         struct lfsck_tgt_descs          *ltds;
6243         struct lfsck_tgt_desc           *ltd;
6244         bool                             fail  = false;
6245         ENTRY;
6246
6247         if (lr->lr_event == LE_PAIRS_VERIFY) {
6248                 int rc;
6249
6250                 rc = lfsck_layout_master_check_pairs(env, com, &lr->lr_fid,
6251                                                      &lr->lr_fid2,
6252                                                      lr->lr_comp_id);
6253
6254                 RETURN(rc);
6255         }
6256
6257         CDEBUG(D_LFSCK, "%s: layout LFSCK master handles notify %u "
6258                "from %s %x, status %d, flags %x, flags2 %x\n",
6259                lfsck_lfsck2name(lfsck), lr->lr_event,
6260                (lr->lr_flags & LEF_FROM_OST) ? "OST" : "MDT",
6261                lr->lr_index, lr->lr_status, lr->lr_flags, lr->lr_flags2);
6262
6263         if (lr->lr_event != LE_PHASE1_DONE &&
6264             lr->lr_event != LE_PHASE2_DONE &&
6265             lr->lr_event != LE_PEER_EXIT)
6266                 RETURN(-EINVAL);
6267
6268         if (lr->lr_flags & LEF_FROM_OST)
6269                 ltds = &lfsck->li_ost_descs;
6270         else
6271                 ltds = &lfsck->li_mdt_descs;
6272         spin_lock(&ltds->ltd_lock);
6273         ltd = lfsck_ltd2tgt(ltds, lr->lr_index);
6274         if (ltd == NULL) {
6275                 spin_unlock(&ltds->ltd_lock);
6276
6277                 RETURN(-ENXIO);
6278         }
6279
6280         list_del_init(&ltd->ltd_layout_phase_list);
6281         switch (lr->lr_event) {
6282         case LE_PHASE1_DONE:
6283                 if (lr->lr_status <= 0 || lr->lr_flags2 & LF_INCOMPLETE) {
6284                         if (lr->lr_flags2 & LF_INCOMPLETE) {
6285                                 if (lr->lr_flags & LEF_FROM_OST)
6286                                         lfsck_lad_set_bitmap(env, com,
6287                                                              ltd->ltd_index);
6288                                 else
6289                                         lo->ll_flags |= LF_INCOMPLETE;
6290                         }
6291                         ltd->ltd_layout_done = 1;
6292                         list_del_init(&ltd->ltd_layout_list);
6293                         fail = true;
6294                         break;
6295                 }
6296
6297                 if (lr->lr_flags & LEF_FROM_OST) {
6298                         if (list_empty(&ltd->ltd_layout_list))
6299                                 list_add_tail(&ltd->ltd_layout_list,
6300                                               &lad->lad_ost_list);
6301                         list_add_tail(&ltd->ltd_layout_phase_list,
6302                                       &lad->lad_ost_phase2_list);
6303                 } else {
6304                         if (list_empty(&ltd->ltd_layout_list))
6305                                 list_add_tail(&ltd->ltd_layout_list,
6306                                               &lad->lad_mdt_list);
6307                         list_add_tail(&ltd->ltd_layout_phase_list,
6308                                       &lad->lad_mdt_phase2_list);
6309                 }
6310                 break;
6311         case LE_PHASE2_DONE:
6312                 ltd->ltd_layout_done = 1;
6313                 if (!list_empty(&ltd->ltd_layout_list)) {
6314                         list_del_init(&ltd->ltd_layout_list);
6315                         if (lr->lr_flags2 & LF_INCOMPLETE) {
6316                                 lfsck_lad_set_bitmap(env, com, ltd->ltd_index);
6317                                 fail = true;
6318                         }
6319                 }
6320
6321                 break;
6322         case LE_PEER_EXIT:
6323                 fail = true;
6324                 ltd->ltd_layout_done = 1;
6325                 list_del_init(&ltd->ltd_layout_list);
6326                 if (!(lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT) &&
6327                     !(lr->lr_flags & LEF_FROM_OST))
6328                                 lo->ll_flags |= LF_INCOMPLETE;
6329                 break;
6330         default:
6331                 break;
6332         }
6333         spin_unlock(&ltds->ltd_lock);
6334
6335         if (fail && lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT) {
6336                 struct lfsck_stop *stop = &lfsck_env_info(env)->lti_stop;
6337
6338                 memset(stop, 0, sizeof(*stop));
6339                 stop->ls_status = lr->lr_status;
6340                 stop->ls_flags = lr->lr_param & ~LPF_BROADCAST;
6341                 lfsck_stop(env, lfsck->li_bottom, stop);
6342         } else if (lfsck_phase2_next_ready(lad)) {
6343                 wake_up_all(&lad->lad_thread.t_ctl_waitq);
6344         }
6345
6346         RETURN(0);
6347 }
6348
6349 static int lfsck_layout_slave_in_notify_local(const struct lu_env *env,
6350                                               struct lfsck_component *com,
6351                                               struct lfsck_req_local *lrl,
6352                                               struct thandle *th)
6353 {
6354         ENTRY;
6355
6356         switch (lrl->lrl_event) {
6357         case LEL_FID_ACCESSED:
6358                 lfsck_rbtree_update_bitmap(env, com, &lrl->lrl_fid, true);
6359                 RETURN(0);
6360         case LEL_PAIRS_VERIFY_LOCAL: {
6361                 int rc;
6362
6363                 lrl->lrl_status = LPVS_INIT;
6364                 /* Firstly, if the MDT-object which is claimed via OST-object
6365                  * local stored PFID xattr recognizes the OST-object, then it
6366                  * must be that the client given PFID is wrong. */
6367                 rc = lfsck_layout_slave_check_pairs(env, com, &lrl->lrl_fid,
6368                                 &lrl->lrl_ff_local.ff_parent,
6369                                 lrl->lrl_ff_local.ff_layout.ol_comp_id);
6370                 if (rc <= 0)
6371                         RETURN(0);
6372
6373                 lrl->lrl_status = LPVS_INCONSISTENT;
6374                 /* The OST-object local stored PFID xattr is stale. We need to
6375                  * check whether the MDT-object that is claimed via the client
6376                  * given PFID information recognizes the OST-object or not. If
6377                  * matches, then need to update the OST-object's PFID xattr. */
6378                 rc = lfsck_layout_slave_check_pairs(env, com, &lrl->lrl_fid,
6379                                 &lrl->lrl_ff_client.ff_parent,
6380                                 lrl->lrl_ff_client.ff_layout.ol_comp_id);
6381                 /* For rc < 0 case:
6382                  * We are not sure whether the client given PFID information
6383                  * is correct or not, do nothing to avoid improper fixing.
6384                  *
6385                  * For rc > 0 case:
6386                  * The client given PFID information is also invalid, we can
6387                  * NOT fix the OST-object inconsistency.
6388                  */
6389                 if (!rc) {
6390                         lrl->lrl_status = LPVS_INCONSISTENT_TOFIX;
6391                         rc = lfsck_layout_slave_repair_pfid(env, com, lrl);
6392                 }
6393
6394                 RETURN(rc);
6395         }
6396         default:
6397                 break;
6398         }
6399
6400         RETURN(-EOPNOTSUPP);
6401 }
6402
6403 static int lfsck_layout_slave_in_notify(const struct lu_env *env,
6404                                         struct lfsck_component *com,
6405                                         struct lfsck_request *lr)
6406 {
6407         struct lfsck_instance *lfsck = com->lc_lfsck;
6408         struct lfsck_layout_slave_data *llsd = com->lc_data;
6409         struct lfsck_layout_slave_target *llst;
6410         int rc;
6411         ENTRY;
6412
6413         switch (lr->lr_event) {
6414         case LE_CONDITIONAL_DESTROY:
6415                 rc = lfsck_layout_slave_conditional_destroy(env, com, lr);
6416                 RETURN(rc);
6417         case LE_PHASE1_DONE: {
6418                 if (lr->lr_flags2 & LF_INCOMPLETE) {
6419                         struct lfsck_layout *lo = com->lc_file_ram;
6420
6421                         lo->ll_flags |= LF_INCOMPLETE;
6422                         llst = lfsck_layout_llst_find_and_del(llsd,
6423                                                               lr->lr_index,
6424                                                               true);
6425                         if (llst != NULL) {
6426                                 lfsck_layout_llst_put(llst);
6427                                 wake_up_all(&lfsck->li_thread.t_ctl_waitq);
6428                         }
6429                 }
6430
6431                 RETURN(0);
6432         }
6433         case LE_PHASE2_DONE:
6434         case LE_PEER_EXIT:
6435                 CDEBUG(D_LFSCK, "%s: layout LFSCK slave handle notify %u "
6436                        "from MDT %x, status %d\n", lfsck_lfsck2name(lfsck),
6437                        lr->lr_event, lr->lr_index, lr->lr_status);
6438                 break;
6439         default:
6440                 RETURN(-EINVAL);
6441         }
6442
6443         llst = lfsck_layout_llst_find_and_del(llsd, lr->lr_index, true);
6444         if (llst == NULL)
6445                 RETURN(0);
6446
6447         lfsck_layout_llst_put(llst);
6448         if (list_empty(&llsd->llsd_master_list))
6449                 wake_up_all(&lfsck->li_thread.t_ctl_waitq);
6450
6451         if (lr->lr_event == LE_PEER_EXIT &&
6452             (lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT ||
6453              (list_empty(&llsd->llsd_master_list) &&
6454               (lr->lr_status == LS_STOPPED ||
6455                lr->lr_status == LS_CO_STOPPED)))) {
6456                 struct lfsck_stop *stop = &lfsck_env_info(env)->lti_stop;
6457
6458                 memset(stop, 0, sizeof(*stop));
6459                 stop->ls_status = lr->lr_status;
6460                 stop->ls_flags = lr->lr_param & ~LPF_BROADCAST;
6461                 lfsck_stop(env, lfsck->li_bottom, stop);
6462         }
6463
6464         RETURN(0);
6465 }
6466
6467 static void lfsck_layout_repaired(struct lfsck_layout *lo, __u64 *count)
6468 {
6469         int i;
6470
6471         for (i = 0; i < LLIT_MAX; i++)
6472                 *count += lo->ll_objs_repaired[i];
6473 }
6474
6475 static int lfsck_layout_query_all(const struct lu_env *env,
6476                                   struct lfsck_component *com,
6477                                   __u32 *mdts_count, __u32 *osts_count,
6478                                   __u64 *repaired)
6479 {
6480         struct lfsck_layout *lo = com->lc_file_ram;
6481         struct lfsck_tgt_descs *ltds;
6482         struct lfsck_tgt_desc *ltd;
6483         int idx;
6484         int rc;
6485         ENTRY;
6486
6487         rc = lfsck_query_all(env, com);
6488         if (rc != 0)
6489                 RETURN(rc);
6490
6491         ltds = &com->lc_lfsck->li_mdt_descs;
6492         down_read(&ltds->ltd_rw_sem);
6493         cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
6494                 ltd = lfsck_ltd2tgt(ltds, idx);
6495                 LASSERT(ltd != NULL);
6496
6497                 mdts_count[ltd->ltd_layout_status]++;
6498                 *repaired += ltd->ltd_layout_repaired;
6499         }
6500         up_read(&ltds->ltd_rw_sem);
6501
6502         ltds = &com->lc_lfsck->li_ost_descs;
6503         down_read(&ltds->ltd_rw_sem);
6504         cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
6505                 ltd = lfsck_ltd2tgt(ltds, idx);
6506                 LASSERT(ltd != NULL);
6507
6508                 osts_count[ltd->ltd_layout_status]++;
6509                 *repaired += ltd->ltd_layout_repaired;
6510         }
6511         up_read(&ltds->ltd_rw_sem);
6512
6513         down_read(&com->lc_sem);
6514         mdts_count[lo->ll_status]++;
6515         lfsck_layout_repaired(lo, repaired);
6516         up_read(&com->lc_sem);
6517
6518         RETURN(0);
6519 }
6520
6521 static int lfsck_layout_query(const struct lu_env *env,
6522                               struct lfsck_component *com,
6523                               struct lfsck_request *req,
6524                               struct lfsck_reply *rep,
6525                               struct lfsck_query *que, int idx)
6526 {
6527         struct lfsck_layout *lo = com->lc_file_ram;
6528         int rc = 0;
6529
6530         if (que != NULL) {
6531                 LASSERT(com->lc_lfsck->li_master);
6532
6533                 rc = lfsck_layout_query_all(env, com,
6534                                             que->lu_mdts_count[idx],
6535                                             que->lu_osts_count[idx],
6536                                             &que->lu_repaired[idx]);
6537         } else {
6538                 down_read(&com->lc_sem);
6539                 rep->lr_status = lo->ll_status;
6540                 if (req->lr_flags & LEF_QUERY_ALL)
6541                         lfsck_layout_repaired(lo, &rep->lr_repaired);
6542                 up_read(&com->lc_sem);
6543         }
6544
6545         return rc;
6546 }
6547
6548 /* with lfsck::li_lock held */
6549 static int lfsck_layout_slave_join(const struct lu_env *env,
6550                                    struct lfsck_component *com,
6551                                    struct lfsck_start_param *lsp)
6552 {
6553         struct lfsck_instance            *lfsck = com->lc_lfsck;
6554         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
6555         struct lfsck_layout_slave_target *llst;
6556         struct lfsck_start               *start = lsp->lsp_start;
6557         int                               rc    = 0;
6558         ENTRY;
6559
6560         if (start == NULL || !(start->ls_flags & LPF_OST_ORPHAN))
6561                 RETURN(0);
6562
6563         if (!lsp->lsp_index_valid)
6564                 RETURN(-EINVAL);
6565
6566         /* If someone is running the LFSCK without orphan handling,
6567          * it will not maintain the object accessing rbtree. So we
6568          * cannot join it for orphan handling. */
6569         if (!llsd->llsd_rbtree_valid)
6570                 RETURN(-EBUSY);
6571
6572         spin_unlock(&lfsck->li_lock);
6573         rc = lfsck_layout_llst_add(llsd, lsp->lsp_index);
6574         spin_lock(&lfsck->li_lock);
6575         if (rc == 0 && !thread_is_running(&lfsck->li_thread)) {
6576                 spin_unlock(&lfsck->li_lock);
6577                 llst = lfsck_layout_llst_find_and_del(llsd, lsp->lsp_index,
6578                                                       true);
6579                 if (llst != NULL)
6580                         lfsck_layout_llst_put(llst);
6581                 spin_lock(&lfsck->li_lock);
6582                 rc = -EAGAIN;
6583         }
6584
6585         RETURN(rc);
6586 }
6587
6588 static struct lfsck_operations lfsck_layout_master_ops = {
6589         .lfsck_reset            = lfsck_layout_reset,
6590         .lfsck_fail             = lfsck_layout_fail,
6591         .lfsck_checkpoint       = lfsck_layout_master_checkpoint,
6592         .lfsck_prep             = lfsck_layout_master_prep,
6593         .lfsck_exec_oit         = lfsck_layout_master_exec_oit,
6594         .lfsck_exec_dir         = lfsck_layout_exec_dir,
6595         .lfsck_post             = lfsck_layout_master_post,
6596         .lfsck_dump             = lfsck_layout_dump,
6597         .lfsck_double_scan      = lfsck_layout_master_double_scan,
6598         .lfsck_data_release     = lfsck_layout_master_data_release,
6599         .lfsck_quit             = lfsck_layout_master_quit,
6600         .lfsck_in_notify        = lfsck_layout_master_in_notify,
6601         .lfsck_query            = lfsck_layout_query,
6602 };
6603
6604 static struct lfsck_operations lfsck_layout_slave_ops = {
6605         .lfsck_reset            = lfsck_layout_reset,
6606         .lfsck_fail             = lfsck_layout_fail,
6607         .lfsck_checkpoint       = lfsck_layout_slave_checkpoint,
6608         .lfsck_prep             = lfsck_layout_slave_prep,
6609         .lfsck_exec_oit         = lfsck_layout_slave_exec_oit,
6610         .lfsck_exec_dir         = lfsck_layout_exec_dir,
6611         .lfsck_post             = lfsck_layout_slave_post,
6612         .lfsck_dump             = lfsck_layout_dump,
6613         .lfsck_double_scan      = lfsck_layout_slave_double_scan,
6614         .lfsck_data_release     = lfsck_layout_slave_data_release,
6615         .lfsck_quit             = lfsck_layout_slave_quit,
6616         .lfsck_in_notify_local  = lfsck_layout_slave_in_notify_local,
6617         .lfsck_in_notify        = lfsck_layout_slave_in_notify,
6618         .lfsck_query            = lfsck_layout_query,
6619         .lfsck_join             = lfsck_layout_slave_join,
6620 };
6621
6622 static void lfsck_layout_assistant_fill_pos(const struct lu_env *env,
6623                                             struct lfsck_component *com,
6624                                             struct lfsck_position *pos)
6625 {
6626         struct lfsck_assistant_data     *lad = com->lc_data;
6627         struct lfsck_layout_req         *llr;
6628
6629         if (((struct lfsck_layout *)(com->lc_file_ram))->ll_status !=
6630             LS_SCANNING_PHASE1)
6631                 return;
6632
6633         if (list_empty(&lad->lad_req_list))
6634                 return;
6635
6636         llr = list_entry(lad->lad_req_list.next,
6637                          struct lfsck_layout_req,
6638                          llr_lar.lar_list);
6639         pos->lp_oit_cookie = llr->llr_lar.lar_parent->lso_oit_cookie - 1;
6640 }
6641
6642 struct lfsck_assistant_operations lfsck_layout_assistant_ops = {
6643         .la_handler_p1          = lfsck_layout_assistant_handler_p1,
6644         .la_handler_p2          = lfsck_layout_assistant_handler_p2,
6645         .la_fill_pos            = lfsck_layout_assistant_fill_pos,
6646         .la_double_scan_result  = lfsck_layout_double_scan_result,
6647         .la_req_fini            = lfsck_layout_assistant_req_fini,
6648         .la_sync_failures       = lfsck_layout_assistant_sync_failures,
6649 };
6650
6651 int lfsck_layout_setup(const struct lu_env *env, struct lfsck_instance *lfsck)
6652 {
6653         struct lfsck_component  *com;
6654         struct lfsck_layout     *lo;
6655         struct dt_object        *root = NULL;
6656         struct dt_object        *obj;
6657         int                      i;
6658         int                      rc;
6659         ENTRY;
6660
6661         OBD_ALLOC_PTR(com);
6662         if (com == NULL)
6663                 RETURN(-ENOMEM);
6664
6665         INIT_LIST_HEAD(&com->lc_link);
6666         INIT_LIST_HEAD(&com->lc_link_dir);
6667         init_rwsem(&com->lc_sem);
6668         atomic_set(&com->lc_ref, 1);
6669         com->lc_lfsck = lfsck;
6670         com->lc_type = LFSCK_TYPE_LAYOUT;
6671         if (lfsck->li_master) {
6672                 com->lc_ops = &lfsck_layout_master_ops;
6673                 com->lc_data = lfsck_assistant_data_init(
6674                                 &lfsck_layout_assistant_ops,
6675                                 LFSCK_LAYOUT);
6676                 if (com->lc_data == NULL)
6677                         GOTO(out, rc = -ENOMEM);
6678
6679                 for (i = 0; i < LFSCK_STF_COUNT; i++)
6680                         mutex_init(&com->lc_sub_trace_objs[i].lsto_mutex);
6681         } else {
6682                 struct lfsck_layout_slave_data *llsd;
6683
6684                 com->lc_ops = &lfsck_layout_slave_ops;
6685                 OBD_ALLOC_PTR(llsd);
6686                 if (llsd == NULL)
6687                         GOTO(out, rc = -ENOMEM);
6688
6689                 INIT_LIST_HEAD(&llsd->llsd_seq_list);
6690                 INIT_LIST_HEAD(&llsd->llsd_master_list);
6691                 spin_lock_init(&llsd->llsd_lock);
6692                 llsd->llsd_rb_root = RB_ROOT;
6693                 rwlock_init(&llsd->llsd_rb_lock);
6694                 com->lc_data = llsd;
6695         }
6696         com->lc_file_size = sizeof(*lo);
6697         OBD_ALLOC(com->lc_file_ram, com->lc_file_size);
6698         if (com->lc_file_ram == NULL)
6699                 GOTO(out, rc = -ENOMEM);
6700
6701         OBD_ALLOC(com->lc_file_disk, com->lc_file_size);
6702         if (com->lc_file_disk == NULL)
6703                 GOTO(out, rc = -ENOMEM);
6704
6705         root = dt_locate(env, lfsck->li_bottom, &lfsck->li_local_root_fid);
6706         if (IS_ERR(root))
6707                 GOTO(out, rc = PTR_ERR(root));
6708
6709         if (unlikely(!dt_try_as_dir(env, root)))
6710                 GOTO(out, rc = -ENOTDIR);
6711
6712         obj = local_file_find_or_create(env, lfsck->li_los, root,
6713                                         LFSCK_LAYOUT,
6714                                         S_IFREG | S_IRUGO | S_IWUSR);
6715         if (IS_ERR(obj))
6716                 GOTO(out, rc = PTR_ERR(obj));
6717
6718         com->lc_obj = obj;
6719         rc = lfsck_layout_load(env, com);
6720         if (rc > 0)
6721                 rc = lfsck_layout_reset(env, com, true);
6722         else if (rc == -ENOENT)
6723                 rc = lfsck_layout_init(env, com);
6724         else if (lfsck->li_master)
6725                 rc = lfsck_load_sub_trace_files(env, com,
6726                                 &dt_lfsck_layout_dangling_features,
6727                                 LFSCK_LAYOUT, false);
6728
6729         if (rc != 0)
6730                 GOTO(out, rc);
6731
6732         lo = com->lc_file_ram;
6733         switch (lo->ll_status) {
6734         case LS_INIT:
6735         case LS_COMPLETED:
6736         case LS_FAILED:
6737         case LS_STOPPED:
6738         case LS_PARTIAL:
6739                 spin_lock(&lfsck->li_lock);
6740                 list_add_tail(&com->lc_link, &lfsck->li_list_idle);
6741                 spin_unlock(&lfsck->li_lock);
6742                 break;
6743         default:
6744                 CERROR("%s: unknown lfsck_layout status %d\n",
6745                        lfsck_lfsck2name(lfsck), lo->ll_status);
6746                 /* fall through */
6747         case LS_SCANNING_PHASE1:
6748         case LS_SCANNING_PHASE2:
6749                 /* No need to store the status to disk right now.
6750                  * If the system crashed before the status stored,
6751                  * it will be loaded back when next time. */
6752                 lo->ll_status = LS_CRASHED;
6753                 if (!lfsck->li_master)
6754                         lo->ll_flags |= LF_INCOMPLETE;
6755                 /* fall through */
6756         case LS_PAUSED:
6757         case LS_CRASHED:
6758         case LS_CO_FAILED:
6759         case LS_CO_STOPPED:
6760         case LS_CO_PAUSED:
6761                 spin_lock(&lfsck->li_lock);
6762                 list_add_tail(&com->lc_link, &lfsck->li_list_scan);
6763                 spin_unlock(&lfsck->li_lock);
6764                 break;
6765         }
6766
6767         if (lo->ll_flags & LF_CRASHED_LASTID) {
6768                 LASSERT(lfsck->li_out_notify != NULL);
6769
6770                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
6771                                      LE_LASTID_REBUILDING);
6772         }
6773
6774         GOTO(out, rc = 0);
6775
6776 out:
6777         if (root != NULL && !IS_ERR(root))
6778                 lfsck_object_put(env, root);
6779
6780         if (rc != 0) {
6781                 lfsck_component_cleanup(env, com);
6782                 CERROR("%s: fail to init layout LFSCK component: rc = %d\n",
6783                        lfsck_lfsck2name(lfsck), rc);
6784         }
6785
6786         return rc;
6787 }
6788
6789 struct lfsck_orphan_it {
6790         struct lfsck_component           *loi_com;
6791         struct lfsck_rbtree_node         *loi_lrn;
6792         struct lfsck_layout_slave_target *loi_llst;
6793         struct lu_fid                     loi_key;
6794         struct lu_orphan_rec_v2           loi_rec;
6795         __u64                             loi_hash;
6796         unsigned int                      loi_over:1;
6797 };
6798
6799 static int lfsck_fid_match_idx(const struct lu_env *env,
6800                                struct lfsck_instance *lfsck,
6801                                const struct lu_fid *fid, int idx)
6802 {
6803         struct seq_server_site  *ss;
6804         struct lu_server_fld    *sf;
6805         struct lu_seq_range     *range = &lfsck_env_info(env)->lti_range;
6806         int                      rc;
6807
6808         /* All abnormal cases will be returned to MDT0. */
6809         if (!fid_is_norm(fid)) {
6810                 if (idx == 0)
6811                         return 1;
6812
6813                 return 0;
6814         }
6815
6816         ss = lfsck_dev_site(lfsck);
6817         if (unlikely(ss == NULL))
6818                 return -ENOTCONN;
6819
6820         sf = ss->ss_server_fld;
6821         LASSERT(sf != NULL);
6822
6823         fld_range_set_any(range);
6824         rc = fld_server_lookup(env, sf, fid_seq(fid), range);
6825         if (rc != 0)
6826                 return rc;
6827
6828         if (!fld_range_is_mdt(range))
6829                 return -EINVAL;
6830
6831         if (range->lsr_index == idx)
6832                 return 1;
6833
6834         return 0;
6835 }
6836
6837 static void lfsck_layout_destroy_orphan(const struct lu_env *env,
6838                                         struct dt_object *obj)
6839 {
6840         struct dt_device        *dev    = lfsck_obj2dev(obj);
6841         struct thandle          *handle;
6842         int                      rc;
6843         ENTRY;
6844
6845         handle = dt_trans_create(env, dev);
6846         if (IS_ERR(handle))
6847                 RETURN_EXIT;
6848
6849         rc = dt_declare_ref_del(env, obj, handle);
6850         if (rc != 0)
6851                 GOTO(stop, rc);
6852
6853         rc = dt_declare_destroy(env, obj, handle);
6854         if (rc != 0)
6855                 GOTO(stop, rc);
6856
6857         rc = dt_trans_start_local(env, dev, handle);
6858         if (rc != 0)
6859                 GOTO(stop, rc);
6860
6861         dt_write_lock(env, obj, 0);
6862         rc = dt_ref_del(env, obj, handle);
6863         if (rc == 0)
6864                 rc = dt_destroy(env, obj, handle);
6865         dt_write_unlock(env, obj);
6866
6867         GOTO(stop, rc);
6868
6869 stop:
6870         dt_trans_stop(env, dev, handle);
6871
6872         CDEBUG(D_LFSCK, "destroy orphan OST-object "DFID": rc = %d\n",
6873                PFID(lfsck_dto2fid(obj)), rc);
6874
6875         RETURN_EXIT;
6876 }
6877
6878 static int lfsck_orphan_index_lookup(const struct lu_env *env,
6879                                      struct dt_object *dt,
6880                                      struct dt_rec *rec,
6881                                      const struct dt_key *key)
6882 {
6883         return -EOPNOTSUPP;
6884 }
6885
6886 static int lfsck_orphan_index_declare_insert(const struct lu_env *env,
6887                                              struct dt_object *dt,
6888                                              const struct dt_rec *rec,
6889                                              const struct dt_key *key,
6890                                              struct thandle *handle)
6891 {
6892         return -EOPNOTSUPP;
6893 }
6894
6895 static int lfsck_orphan_index_insert(const struct lu_env *env,
6896                                      struct dt_object *dt,
6897                                      const struct dt_rec *rec,
6898                                      const struct dt_key *key,
6899                                      struct thandle *handle,
6900                                      int ignore_quota)
6901 {
6902         return -EOPNOTSUPP;
6903 }
6904
6905 static int lfsck_orphan_index_declare_delete(const struct lu_env *env,
6906                                              struct dt_object *dt,
6907                                              const struct dt_key *key,
6908                                              struct thandle *handle)
6909 {
6910         return -EOPNOTSUPP;
6911 }
6912
6913 static int lfsck_orphan_index_delete(const struct lu_env *env,
6914                                      struct dt_object *dt,
6915                                      const struct dt_key *key,
6916                                      struct thandle *handle)
6917 {
6918         return -EOPNOTSUPP;
6919 }
6920
6921 static struct dt_it *lfsck_orphan_it_init(const struct lu_env *env,
6922                                           struct dt_object *dt,
6923                                           __u32 attr)
6924 {
6925         struct dt_device                *dev    = lu2dt_dev(dt->do_lu.lo_dev);
6926         struct lfsck_instance           *lfsck;
6927         struct lfsck_component          *com    = NULL;
6928         struct lfsck_layout_slave_data  *llsd;
6929         struct lfsck_orphan_it          *it     = NULL;
6930         struct lfsck_layout             *lo;
6931         int                              rc     = 0;
6932         ENTRY;
6933
6934         lfsck = lfsck_instance_find(dev, true, false);
6935         if (unlikely(lfsck == NULL))
6936                 RETURN(ERR_PTR(-ENXIO));
6937
6938         com = lfsck_component_find(lfsck, LFSCK_TYPE_LAYOUT);
6939         if (unlikely(com == NULL))
6940                 GOTO(out, rc = -ENOENT);
6941
6942         lo = com->lc_file_ram;
6943         if (lo->ll_flags & LF_INCOMPLETE)
6944                 GOTO(out, rc = -ESRCH);
6945
6946         llsd = com->lc_data;
6947         if (!llsd->llsd_rbtree_valid)
6948                 GOTO(out, rc = -ESRCH);
6949
6950         OBD_ALLOC_PTR(it);
6951         if (it == NULL)
6952                 GOTO(out, rc = -ENOMEM);
6953
6954         it->loi_llst = lfsck_layout_llst_find_and_del(llsd, attr, false);
6955         if (it->loi_llst == NULL)
6956                 GOTO(out, rc = -ENXIO);
6957
6958         if (dev->dd_record_fid_accessed) {
6959                 /* The first iteration against the rbtree, scan the whole rbtree
6960                  * to remove the nodes which do NOT need to be handled. */
6961                 write_lock(&llsd->llsd_rb_lock);
6962                 if (dev->dd_record_fid_accessed) {
6963                         struct rb_node                  *node;
6964                         struct rb_node                  *next;
6965                         struct lfsck_rbtree_node        *lrn;
6966
6967                         /* No need to record the fid accessing anymore. */
6968                         dev->dd_record_fid_accessed = 0;
6969
6970                         node = rb_first(&llsd->llsd_rb_root);
6971                         while (node != NULL) {
6972                                 next = rb_next(node);
6973                                 lrn = rb_entry(node, struct lfsck_rbtree_node,
6974                                                lrn_node);
6975                                 if (atomic_read(&lrn->lrn_known_count) <=
6976                                     atomic_read(&lrn->lrn_accessed_count)) {
6977                                         rb_erase(node, &llsd->llsd_rb_root);
6978                                         lfsck_rbtree_free(lrn);
6979                                 }
6980                                 node = next;
6981                         }
6982                 }
6983                 write_unlock(&llsd->llsd_rb_lock);
6984         }
6985
6986         /* read lock the rbtree when init, and unlock when fini */
6987         read_lock(&llsd->llsd_rb_lock);
6988         it->loi_com = com;
6989         com = NULL;
6990
6991         GOTO(out, rc = 0);
6992
6993 out:
6994         if (com != NULL)
6995                 lfsck_component_put(env, com);
6996
6997         CDEBUG(D_LFSCK, "%s: init the orphan iteration: rc = %d\n",
6998                lfsck_lfsck2name(lfsck), rc);
6999
7000         lfsck_instance_put(env, lfsck);
7001         if (rc != 0) {
7002                 if (it != NULL)
7003                         OBD_FREE_PTR(it);
7004
7005                 it = (struct lfsck_orphan_it *)ERR_PTR(rc);
7006         }
7007
7008         return (struct dt_it *)it;
7009 }
7010
7011 static void lfsck_orphan_it_fini(const struct lu_env *env,
7012                                  struct dt_it *di)
7013 {
7014         struct lfsck_orphan_it           *it    = (struct lfsck_orphan_it *)di;
7015         struct lfsck_component           *com   = it->loi_com;
7016         struct lfsck_layout_slave_data   *llsd;
7017         struct lfsck_layout_slave_target *llst;
7018
7019         if (com != NULL) {
7020                 CDEBUG(D_LFSCK, "%s: fini the orphan iteration\n",
7021                        lfsck_lfsck2name(com->lc_lfsck));
7022
7023                 llsd = com->lc_data;
7024                 read_unlock(&llsd->llsd_rb_lock);
7025                 llst = it->loi_llst;
7026                 LASSERT(llst != NULL);
7027
7028                 /* Save the key and hash for iterate next. */
7029                 llst->llst_fid = it->loi_key;
7030                 llst->llst_hash = it->loi_hash;
7031                 lfsck_layout_llst_put(llst);
7032                 lfsck_component_put(env, com);
7033         }
7034         OBD_FREE_PTR(it);
7035 }
7036
7037 /**
7038  * \retval       +1: the iteration finished
7039  * \retval        0: on success, not finished
7040  * \retval      -ve: on error
7041  */
7042 static int lfsck_orphan_it_next(const struct lu_env *env,
7043                                 struct dt_it *di)
7044 {
7045         struct lfsck_thread_info        *info   = lfsck_env_info(env);
7046         struct filter_fid               *ff     = &info->lti_ff;
7047         struct lu_attr                  *la     = &info->lti_la;
7048         struct lfsck_orphan_it          *it     = (struct lfsck_orphan_it *)di;
7049         struct lu_fid                   *key    = &it->loi_key;
7050         struct lu_orphan_rec_v2         *rec    = &it->loi_rec;
7051         struct ost_layout               *ol     = &rec->lor_layout;
7052         struct lfsck_component          *com    = it->loi_com;
7053         struct lfsck_instance           *lfsck  = com->lc_lfsck;
7054         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
7055         struct dt_object                *obj;
7056         struct lfsck_rbtree_node        *lrn;
7057         int                              pos;
7058         int                              rc;
7059         __u32                            save;
7060         __u32                            idx    = it->loi_llst->llst_index;
7061         bool                             exact  = false;
7062         ENTRY;
7063
7064         if (it->loi_over)
7065                 RETURN(1);
7066
7067 again0:
7068         lrn = it->loi_lrn;
7069         if (lrn == NULL) {
7070                 lrn = lfsck_rbtree_search(llsd, key, &exact);
7071                 if (lrn == NULL) {
7072                         it->loi_over = 1;
7073                         RETURN(1);
7074                 }
7075
7076                 it->loi_lrn = lrn;
7077                 if (!exact) {
7078                         key->f_seq = lrn->lrn_seq;
7079                         key->f_oid = lrn->lrn_first_oid;
7080                         key->f_ver = 0;
7081                 }
7082         } else {
7083                 key->f_oid++;
7084                 if (unlikely(key->f_oid == 0)) {
7085                         key->f_seq++;
7086                         it->loi_lrn = NULL;
7087                         goto again0;
7088                 }
7089
7090                 if (key->f_oid >=
7091                     lrn->lrn_first_oid + LFSCK_RBTREE_BITMAP_WIDTH) {
7092                         it->loi_lrn = NULL;
7093                         goto again0;
7094                 }
7095         }
7096
7097         if (unlikely(atomic_read(&lrn->lrn_known_count) <=
7098                      atomic_read(&lrn->lrn_accessed_count))) {
7099                 struct rb_node *next = rb_next(&lrn->lrn_node);
7100
7101                 while (next != NULL) {
7102                         lrn = rb_entry(next, struct lfsck_rbtree_node,
7103                                        lrn_node);
7104                         if (atomic_read(&lrn->lrn_known_count) >
7105                             atomic_read(&lrn->lrn_accessed_count))
7106                                 break;
7107                         next = rb_next(next);
7108                 }
7109
7110                 if (next == NULL) {
7111                         it->loi_over = 1;
7112                         RETURN(1);
7113                 }
7114
7115                 it->loi_lrn = lrn;
7116                 key->f_seq = lrn->lrn_seq;
7117                 key->f_oid = lrn->lrn_first_oid;
7118                 key->f_ver = 0;
7119         }
7120
7121         pos = key->f_oid - lrn->lrn_first_oid;
7122
7123 again1:
7124         pos = find_next_bit(lrn->lrn_known_bitmap,
7125                             LFSCK_RBTREE_BITMAP_WIDTH, pos);
7126         if (pos >= LFSCK_RBTREE_BITMAP_WIDTH) {
7127                 key->f_oid = lrn->lrn_first_oid + pos;
7128                 if (unlikely(key->f_oid < lrn->lrn_first_oid)) {
7129                         key->f_seq++;
7130                         key->f_oid = 0;
7131                 }
7132                 it->loi_lrn = NULL;
7133                 goto again0;
7134         }
7135
7136         if (test_bit(pos, lrn->lrn_accessed_bitmap)) {
7137                 pos++;
7138                 goto again1;
7139         }
7140
7141         key->f_oid = lrn->lrn_first_oid + pos;
7142         obj = lfsck_object_find_bottom(env, lfsck, key);
7143         if (IS_ERR(obj)) {
7144                 rc = PTR_ERR(obj);
7145                 if (rc == -ENOENT) {
7146                         pos++;
7147                         goto again1;
7148                 }
7149                 RETURN(rc);
7150         }
7151
7152         dt_read_lock(env, obj, 0);
7153         if (dt_object_exists(obj) == 0 ||
7154             lfsck_is_dead_obj(obj)) {
7155                 dt_read_unlock(env, obj);
7156                 lfsck_object_put(env, obj);
7157                 pos++;
7158                 goto again1;
7159         }
7160
7161         rc = dt_attr_get(env, obj, la);
7162         if (rc != 0)
7163                 GOTO(out, rc);
7164
7165         rc = dt_xattr_get(env, obj, lfsck_buf_get(env, ff, sizeof(*ff)),
7166                           XATTR_NAME_FID);
7167         if (rc == -ENODATA) {
7168                 /* For the pre-created OST-object, update the bitmap to avoid
7169                  * others LFSCK (second phase) iteration to touch it again. */
7170                 if (la->la_ctime == 0) {
7171                         if (!test_and_set_bit(pos, lrn->lrn_accessed_bitmap))
7172                                 atomic_inc(&lrn->lrn_accessed_count);
7173
7174                         /* For the race between repairing dangling referenced
7175                          * MDT-object and unlink the file, it may left orphan
7176                          * OST-object there. Destroy it now! */
7177                         if (unlikely(!(la->la_mode & S_ISUID))) {
7178                                 dt_read_unlock(env, obj);
7179                                 lfsck_layout_destroy_orphan(env, obj);
7180                                 lfsck_object_put(env, obj);
7181                                 pos++;
7182                                 goto again1;
7183                         }
7184                 } else if (idx == 0) {
7185                         /* If the orphan OST-object has no parent information,
7186                          * regard it as referenced by the MDT-object on MDT0. */
7187                         fid_zero(&rec->lor_rec.lor_fid);
7188                         rec->lor_rec.lor_uid = la->la_uid;
7189                         rec->lor_rec.lor_gid = la->la_gid;
7190                         memset(ol, 0, sizeof(*ol));
7191
7192                         GOTO(out, rc = 0);
7193                 }
7194
7195                 dt_read_unlock(env, obj);
7196                 lfsck_object_put(env, obj);
7197                 pos++;
7198                 goto again1;
7199         }
7200
7201         if (rc < sizeof(struct lu_fid))
7202                 GOTO(out, rc = (rc < 0 ? rc : -EINVAL));
7203
7204         fid_le_to_cpu(&rec->lor_rec.lor_fid, &ff->ff_parent);
7205         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
7206          * MDT-object's FID::f_ver, instead it is the OST-object index in its
7207          * parent MDT-object's layout EA. */
7208         save = rec->lor_rec.lor_fid.f_stripe_idx;
7209         rec->lor_rec.lor_fid.f_ver = 0;
7210         rc = lfsck_fid_match_idx(env, lfsck, &rec->lor_rec.lor_fid, idx);
7211         /* If the orphan OST-object does not claim the MDT, then next.
7212          *
7213          * If we do not know whether it matches or not, then return it
7214          * to the MDT for further check. */
7215         if (rc == 0) {
7216                 dt_read_unlock(env, obj);
7217                 lfsck_object_put(env, obj);
7218                 pos++;
7219                 goto again1;
7220         }
7221
7222         rec->lor_rec.lor_fid.f_stripe_idx = save;
7223         rec->lor_rec.lor_uid = la->la_uid;
7224         rec->lor_rec.lor_gid = la->la_gid;
7225         ost_layout_le_to_cpu(ol, &ff->ff_layout);
7226
7227         CDEBUG(D_LFSCK, "%s: return orphan "DFID", PFID "DFID", owner %u:%u, "
7228                "stripe size %u, stripe count %u, COMP id %u, COMP start %llu, "
7229                "COMP end %llu\n", lfsck_lfsck2name(com->lc_lfsck), PFID(key),
7230                PFID(&rec->lor_rec.lor_fid), rec->lor_rec.lor_uid,
7231                rec->lor_rec.lor_gid, ol->ol_stripe_size, ol->ol_stripe_count,
7232                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end);
7233
7234         GOTO(out, rc = 0);
7235
7236 out:
7237         dt_read_unlock(env, obj);
7238         lfsck_object_put(env, obj);
7239         if (rc == 0)
7240                 it->loi_hash++;
7241
7242         return rc;
7243 }
7244
7245 /**
7246  * \retval       +1: locate to the exactly position
7247  * \retval        0: cannot locate to the exactly position,
7248  *                   call next() to move to a valid position.
7249  * \retval      -ve: on error
7250  */
7251 static int lfsck_orphan_it_get(const struct lu_env *env,
7252                                struct dt_it *di,
7253                                const struct dt_key *key)
7254 {
7255         struct lfsck_orphan_it  *it   = (struct lfsck_orphan_it *)di;
7256         int                      rc;
7257
7258         it->loi_key = *(struct lu_fid *)key;
7259         rc = lfsck_orphan_it_next(env, di);
7260         if (rc == 1)
7261                 return 0;
7262
7263         if (rc == 0)
7264                 return 1;
7265
7266         return rc;
7267 }
7268
7269 static void lfsck_orphan_it_put(const struct lu_env *env,
7270                                 struct dt_it *di)
7271 {
7272 }
7273
7274 static struct dt_key *lfsck_orphan_it_key(const struct lu_env *env,
7275                                           const struct dt_it *di)
7276 {
7277         struct lfsck_orphan_it *it = (struct lfsck_orphan_it *)di;
7278
7279         return (struct dt_key *)&it->loi_key;
7280 }
7281
7282 static int lfsck_orphan_it_key_size(const struct lu_env *env,
7283                                     const struct dt_it *di)
7284 {
7285         return sizeof(struct lu_fid);
7286 }
7287
7288 static int lfsck_orphan_it_rec(const struct lu_env *env,
7289                                const struct dt_it *di,
7290                                struct dt_rec *rec,
7291                                __u32 attr)
7292 {
7293         struct lfsck_orphan_it *it = (struct lfsck_orphan_it *)di;
7294
7295         *(struct lu_orphan_rec_v2 *)rec = it->loi_rec;
7296
7297         return 0;
7298 }
7299
7300 static __u64 lfsck_orphan_it_store(const struct lu_env *env,
7301                                    const struct dt_it *di)
7302 {
7303         struct lfsck_orphan_it  *it   = (struct lfsck_orphan_it *)di;
7304
7305         return it->loi_hash;
7306 }
7307
7308 /**
7309  * \retval       +1: locate to the exactly position
7310  * \retval        0: cannot locate to the exactly position,
7311  *                   call next() to move to a valid position.
7312  * \retval      -ve: on error
7313  */
7314 static int lfsck_orphan_it_load(const struct lu_env *env,
7315                                 const struct dt_it *di,
7316                                 __u64 hash)
7317 {
7318         struct lfsck_orphan_it           *it   = (struct lfsck_orphan_it *)di;
7319         struct lfsck_layout_slave_target *llst = it->loi_llst;
7320         int                               rc;
7321
7322         LASSERT(llst != NULL);
7323
7324         if (hash != llst->llst_hash) {
7325                 CDEBUG(D_LFSCK, "%s: the given hash %llu for orphan "
7326                        "iteration does not match the one when fini "
7327                        "%llu, to be reset.\n",
7328                        lfsck_lfsck2name(it->loi_com->lc_lfsck), hash,
7329                        llst->llst_hash);
7330                 fid_zero(&llst->llst_fid);
7331                 llst->llst_hash = 0;
7332         }
7333
7334         it->loi_key = llst->llst_fid;
7335         it->loi_hash = llst->llst_hash;
7336         rc = lfsck_orphan_it_next(env, (struct dt_it *)di);
7337         if (rc == 1)
7338                 return 0;
7339
7340         if (rc == 0)
7341                 return 1;
7342
7343         return rc;
7344 }
7345
7346 static int lfsck_orphan_it_key_rec(const struct lu_env *env,
7347                                    const struct dt_it *di,
7348                                    void *key_rec)
7349 {
7350         return 0;
7351 }
7352
7353 const struct dt_index_operations lfsck_orphan_index_ops = {
7354         .dio_lookup             = lfsck_orphan_index_lookup,
7355         .dio_declare_insert     = lfsck_orphan_index_declare_insert,
7356         .dio_insert             = lfsck_orphan_index_insert,
7357         .dio_declare_delete     = lfsck_orphan_index_declare_delete,
7358         .dio_delete             = lfsck_orphan_index_delete,
7359         .dio_it = {
7360                 .init           = lfsck_orphan_it_init,
7361                 .fini           = lfsck_orphan_it_fini,
7362                 .get            = lfsck_orphan_it_get,
7363                 .put            = lfsck_orphan_it_put,
7364                 .next           = lfsck_orphan_it_next,
7365                 .key            = lfsck_orphan_it_key,
7366                 .key_size       = lfsck_orphan_it_key_size,
7367                 .rec            = lfsck_orphan_it_rec,
7368                 .store          = lfsck_orphan_it_store,
7369                 .load           = lfsck_orphan_it_load,
7370                 .key_rec        = lfsck_orphan_it_key_rec,
7371         }
7372 };