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         rc = ostid_set_id(oi, 0);
3120         if (rc)
3121                 GOTO(log, rc);
3122
3123         rc = ostid_to_fid(fid, oi, ltd->ltd_index);
3124         if (rc != 0)
3125                 GOTO(log, rc);
3126
3127         obj = lfsck_object_find_by_dev(env, ltd->ltd_tgt, fid);
3128         if (unlikely(IS_ERR(obj)))
3129                 GOTO(log, rc = PTR_ERR(obj));
3130
3131         rc = obj->do_ops->do_index_try(env, obj,
3132                                        &dt_lfsck_layout_orphan_features);
3133         if (rc != 0)
3134                 GOTO(put, rc);
3135
3136         iops = &obj->do_index_ops->dio_it;
3137         di = iops->init(env, obj, 0);
3138         if (IS_ERR(di))
3139                 GOTO(put, rc = PTR_ERR(di));
3140
3141         rc = iops->load(env, di, 0);
3142         if (rc == -ESRCH) {
3143                 /* -ESRCH means that the orphan OST-objects rbtree has been
3144                  * cleanup because of the OSS server restart or other errors. */
3145                 lfsck_lad_set_bitmap(env, com, ltd->ltd_index);
3146                 GOTO(fini, rc);
3147         }
3148
3149         if (rc == 0)
3150                 rc = iops->next(env, di);
3151         else if (rc > 0)
3152                 rc = 0;
3153
3154         if (rc < 0)
3155                 GOTO(fini, rc);
3156
3157         if (rc > 0)
3158                 GOTO(fini, rc = 0);
3159
3160         do {
3161                 struct dt_key           *key;
3162                 struct lu_orphan_rec_v2 *rec = &info->lti_rec;
3163
3164                 if (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val) &&
3165                     unlikely(!thread_is_running(&lfsck->li_thread)))
3166                         break;
3167
3168                 key = iops->key(env, di);
3169                 com->lc_fid_latest_scanned_phase2 = *(struct lu_fid *)key;
3170                 /* Remote target OST may be runnning old LFSCK */
3171                 memset(rec, 0, sizeof(*rec));
3172                 rc = iops->rec(env, di, (struct dt_rec *)rec, 0);
3173                 if (rc == 0)
3174                         rc = lfsck_layout_scan_orphan_one(env, com, ltd, rec,
3175                                         &com->lc_fid_latest_scanned_phase2);
3176                 if (rc != 0 && bk->lb_param & LPF_FAILOUT)
3177                         GOTO(fini, rc);
3178
3179                 lfsck_control_speed_by_self(com);
3180                 do {
3181                         rc = iops->next(env, di);
3182                 } while (rc < 0 && !(bk->lb_param & LPF_FAILOUT));
3183         } while (rc == 0);
3184
3185         GOTO(fini, rc);
3186
3187 fini:
3188         iops->put(env, di);
3189         iops->fini(env, di);
3190 put:
3191         lfsck_object_put(env, obj);
3192
3193 log:
3194         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant finished the orphan "
3195                "scanning for OST%04x: rc = %d\n",
3196                lfsck_lfsck2name(lfsck), ltd->ltd_index, rc);
3197
3198         return rc > 0 ? 0 : rc;
3199 }
3200
3201 static int lfsck_lmm2layout(struct lov_mds_md_v1 *lmm, struct ost_layout *ol,
3202                             __u32 comp_id)
3203 {
3204         __u32 magic = le32_to_cpu(lmm->lmm_magic);
3205         int rc = 0;
3206         ENTRY;
3207
3208         if (magic == LOV_MAGIC_V1 || magic == LOV_MAGIC_V3) {
3209                 ol->ol_stripe_size = lmm->lmm_stripe_size;
3210                 ol->ol_stripe_count = lmm->lmm_stripe_count;
3211                 ol->ol_comp_start = 0;
3212                 ol->ol_comp_end = 0;
3213                 ol->ol_comp_id = 0;
3214         } else if (magic == LOV_MAGIC_COMP_V1) {
3215                 struct lov_comp_md_v1 *lcm = (struct lov_comp_md_v1 *)lmm;
3216                 struct lov_comp_md_entry_v1 *lcme = NULL;
3217                 __u16 count = le16_to_cpu(lcm->lcm_entry_count);
3218                 int i;
3219
3220                 for (i = 0; i < count; i++) {
3221                         lcme = &lcm->lcm_entries[i];
3222                         if (le32_to_cpu(lcme->lcme_id) == comp_id) {
3223                                 LASSERT(le32_to_cpu(lcme->lcme_flags) &
3224                                         LCME_FL_INIT);
3225
3226                                 break;
3227                         }
3228                 }
3229
3230                 /* The comp has been removed, do nothing. */
3231                 if (i == count)
3232                         GOTO(out, rc = 1);
3233
3234                 lmm = (void *)lmm + le32_to_cpu(lcme->lcme_offset);
3235                 ol->ol_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
3236                 ol->ol_stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
3237                 ol->ol_comp_start = le64_to_cpu(lcme->lcme_extent.e_start);
3238                 ol->ol_comp_end = le64_to_cpu(lcme->lcme_extent.e_end);
3239                 ol->ol_comp_id = le32_to_cpu(lcme->lcme_id);
3240         } else {
3241                 GOTO(out, rc = -EINVAL);
3242         }
3243
3244         EXIT;
3245
3246 out:
3247         return rc;
3248 }
3249
3250 /**
3251  * Repair the MDT-object with dangling LOV EA reference.
3252  *
3253  * we need to repair the inconsistency according to the users' requirement:
3254  *
3255  * 1) Keep the inconsistency there and report the inconsistency case,
3256  *    then give the chance to the application to find related issues,
3257  *    and the users can make the decision about how to handle it with
3258  *    more human knownledge. (by default)
3259  *
3260  * 2) Re-create the missing OST-object with the FID/owner information.
3261  *
3262  * \param[in] env       pointer to the thread context
3263  * \param[in] com       the layout LFSCK component
3264  * \param[in] parent    the MDT-object with dangling LOV EA reference
3265  * \param[in] child     the OST-object to be created
3266  * \param[in] comp_id   the component ID of the OST-object in the LOV EA
3267  * \param[in] ea_off    the offset of the OST-object in the LOV EA
3268  * \param[in] ost_idx   the index of OST on which the OST-object resides
3269  *
3270  * \retval              +1 for repair successfully
3271  * \retval              0 for did nothing
3272  * \retval              negative error number on failure
3273  */
3274 static int __lfsck_layout_repair_dangling(const struct lu_env *env,
3275                                           struct lfsck_component *com,
3276                                           struct dt_object *parent,
3277                                           struct dt_object *child,
3278                                           __u32 comp_id, __u32 ea_off,
3279                                           __u32 ost_idx, bool log)
3280 {
3281         struct lfsck_thread_info *info = lfsck_env_info(env);
3282         struct filter_fid *ff = &info->lti_ff;
3283         struct ost_layout *ol = &ff->ff_layout;
3284         struct dt_object_format *dof = &info->lti_dof;
3285         struct lu_attr *la = &info->lti_la;
3286         struct lfsck_instance *lfsck = com->lc_lfsck;
3287         struct dt_device *dev = lfsck_obj2dev(child);
3288         const struct lu_fid *pfid = lfsck_dto2fid(parent);
3289         const struct lu_fid *cfid = lfsck_dto2fid(child);
3290         struct lu_buf *tbuf = &info->lti_big_buf;
3291         struct thandle *handle;
3292         struct lu_buf *buf;
3293         struct lustre_handle lh = { 0 };
3294         int rc;
3295         ENTRY;
3296
3297         if (!(lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ))
3298                 GOTO(log, rc = 1);
3299
3300         rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
3301                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
3302                               LCK_EX);
3303         if (rc != 0)
3304                 GOTO(log, rc);
3305
3306         rc = dt_attr_get(env, parent, la);
3307         if (rc != 0)
3308                 GOTO(unlock1, rc);
3309
3310         la->la_mode = S_IFREG | 0666;
3311         la->la_atime = la->la_mtime = la->la_ctime = 0;
3312         la->la_valid = LA_TYPE | LA_MODE | LA_UID | LA_GID |
3313                        LA_ATIME | LA_MTIME | LA_CTIME;
3314         memset(dof, 0, sizeof(*dof));
3315         ff->ff_parent.f_seq = cpu_to_le64(pfid->f_seq);
3316         ff->ff_parent.f_oid = cpu_to_le32(pfid->f_oid);
3317         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
3318          * MDT-object's FID::f_ver, instead it is the OST-object index in its
3319          * parent MDT-object's layout EA. */
3320         ff->ff_parent.f_stripe_idx = cpu_to_le32(ea_off);
3321
3322         rc = lfsck_layout_get_lovea(env, parent, tbuf);
3323         if (rc < 0)
3324                 GOTO(unlock1, rc);
3325
3326         rc = lfsck_lmm2layout(tbuf->lb_buf, ol, comp_id);
3327         if (rc)
3328                 GOTO(unlock1, rc);
3329
3330         buf = lfsck_buf_get(env, ff, sizeof(struct filter_fid));
3331         handle = dt_trans_create(env, dev);
3332         if (IS_ERR(handle))
3333                 GOTO(unlock1, rc = PTR_ERR(handle));
3334
3335         rc = dt_declare_create(env, child, la, NULL, dof, handle);
3336         if (rc != 0)
3337                 GOTO(stop, rc);
3338
3339         rc = dt_declare_xattr_set(env, child, buf, XATTR_NAME_FID,
3340                                   LU_XATTR_CREATE, handle);
3341         if (rc != 0)
3342                 GOTO(stop, rc);
3343
3344         rc = dt_trans_start_local(env, dev, handle);
3345         if (rc != 0)
3346                 GOTO(stop, rc);
3347
3348         dt_read_lock(env, parent, 0);
3349         if (unlikely(lfsck_is_dead_obj(parent)))
3350                 GOTO(unlock2, rc = 0);
3351
3352         if (lfsck->li_bookmark_ram.lb_param & LPF_DELAY_CREATE_OSTOBJ) {
3353                 struct ost_id *oi = &info->lti_oi;
3354                 struct lu_fid *tfid = &info->lti_fid2;
3355                 struct lu_buf *lovea = &info->lti_big_buf;
3356                 struct lov_mds_md_v1 *lmm;
3357                 struct lov_ost_data_v1 *objs;
3358                 __u32 magic;
3359                 int count;
3360                 int idx2;
3361
3362                 rc = lfsck_layout_get_lovea(env, parent, lovea);
3363                 if (rc <= 0)
3364                         GOTO(unlock2, rc);
3365
3366                 lmm = lovea->lb_buf;
3367                 magic = le32_to_cpu(lmm->lmm_magic);
3368                 if (magic == LOV_MAGIC_COMP_V1) {
3369                         struct lov_comp_md_v1 *lcm = buf->lb_buf;
3370                         struct lov_comp_md_entry_v1 *lcme;
3371                         __u16 count = le16_to_cpu(lcm->lcm_entry_count);
3372                         int i;
3373
3374                         for (i = 0; i < count; i++) {
3375                                 lcme = &lcm->lcm_entries[i];
3376                                 if (le32_to_cpu(lcme->lcme_id) == comp_id) {
3377                                         LASSERT(le32_to_cpu(lcme->lcme_flags) &
3378                                                 LCME_FL_INIT);
3379
3380                                         lmm = lovea->lb_buf +
3381                                                 le32_to_cpu(lcme->lcme_offset);
3382                                         magic = le32_to_cpu(lmm->lmm_magic);
3383                                         goto check;
3384                                 }
3385                         }
3386
3387                         /* Someone removed the component, do nothing. */
3388                         GOTO(unlock2, rc = 0);
3389                 }
3390
3391 check:
3392                 count = le16_to_cpu(lmm->lmm_stripe_count);
3393                 /* Someone changed the LOV EA, do nothing. */
3394                 if (count <= ea_off)
3395                         GOTO(unlock2, rc = 0);
3396
3397                 if (magic == LOV_MAGIC_V1) {
3398                         objs = &lmm->lmm_objects[ea_off];
3399                 } else {
3400                         LASSERT(magic == LOV_MAGIC_V3);
3401
3402                         objs = &((struct lov_mds_md_v3 *)lmm)->\
3403                                                         lmm_objects[ea_off];
3404                 }
3405
3406                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
3407                 idx2 = le32_to_cpu(objs->l_ost_idx);
3408                 rc = ostid_to_fid(tfid, oi, idx2);
3409                 /* Someone changed the LOV EA, do nothing. */
3410                 if (rc != 0 || !lu_fid_eq(tfid, cfid))
3411                         GOTO(unlock2, rc);
3412         }
3413
3414         rc = dt_create(env, child, la, NULL, dof, handle);
3415         if (rc != 0)
3416                 GOTO(unlock2, rc);
3417
3418         rc = dt_xattr_set(env, child, buf, XATTR_NAME_FID, LU_XATTR_CREATE,
3419                           handle);
3420
3421         GOTO(unlock2, rc);
3422
3423 unlock2:
3424         dt_read_unlock(env, parent);
3425
3426 stop:
3427         rc = lfsck_layout_trans_stop(env, dev, handle, rc);
3428
3429 unlock1:
3430         lfsck_ibits_unlock(&lh, LCK_EX);
3431
3432 log:
3433         if (rc && log)
3434                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant found "
3435                        "dangling reference for: parent "DFID", child "
3436                        DFID", comp_id %u, ea_off %u, ost_idx %u, %s: "
3437                        "rc = %d\n",
3438                        lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
3439                        comp_id, ea_off, ost_idx,
3440                        (lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ) ?
3441                                 "Create the lost OST-object as required" :
3442                                 "Keep the MDT-object there by default", rc);
3443
3444         return rc;
3445 }
3446
3447 /**
3448  * Repair the MDT-object with dangling LOV EA reference.
3449  *
3450  * Prepare parameters and call __lfsck_layout_repair_dangling()
3451  * to repair the dangling LOV EA reference.
3452  *
3453  * \param[in] env       pointer to the thread context
3454  * \param[in] com       the layout LFSCK component
3455  * \param[in] pfid      the MDT-object's FID
3456  * \param[in] cfid      the FID for the OST-object to be created
3457  * \param[in] comp_id   the component ID of the OST-object in the LOV EA
3458  * \param[in] ea_off    the offset of the OST-object in the LOV EA
3459  * \param[in] ost_idx   the index of OST on which the OST-object resides
3460  *
3461  * \retval              +1 for repair successfully
3462  * \retval              0 for did nothing
3463  * \retval              negative error number on failure
3464  */
3465 static int lfsck_layout_repair_dangling(const struct lu_env *env,
3466                                         struct lfsck_component *com,
3467                                         const struct lu_fid *pfid,
3468                                         const struct lu_fid *cfid,
3469                                         __u32 comp_id, __u32 ea_off,
3470                                         __u32 ost_idx)
3471 {
3472         struct lfsck_instance *lfsck = com->lc_lfsck;
3473         struct dt_object *parent = NULL;
3474         struct dt_object *child = NULL;
3475         struct lfsck_tgt_desc *ltd;
3476         int rc;
3477         ENTRY;
3478
3479         parent = lfsck_object_find_bottom(env, lfsck, pfid);
3480         if (IS_ERR(parent))
3481                 GOTO(log, rc = PTR_ERR(parent));
3482
3483         /* The MDT-object has been removed. */
3484         if (dt_object_exists(parent) == 0)
3485                 GOTO(log, rc = 0);
3486
3487         ltd = lfsck_ltd2tgt(&lfsck->li_ost_descs, ost_idx);
3488         if (unlikely(ltd == NULL))
3489                 GOTO(log, rc = -ENODEV);
3490
3491         child = lfsck_object_find_by_dev(env, ltd->ltd_tgt, cfid);
3492         if (IS_ERR(child))
3493                 GOTO(log, rc = PTR_ERR(child));
3494
3495         /* The OST-object has been created. */
3496         if (unlikely(dt_object_exists(child) != 0))
3497                 GOTO(log, rc = 0);
3498
3499         rc = __lfsck_layout_repair_dangling(env, com, parent, child,
3500                                             comp_id, ea_off, ost_idx, false);
3501
3502         GOTO(log, rc);
3503
3504 log:
3505         if (child != NULL && !IS_ERR(child))
3506                 lfsck_object_put(env, child);
3507
3508         if (parent != NULL && !IS_ERR(parent))
3509                 lfsck_object_put(env, parent);
3510
3511         if (rc)
3512                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant found "
3513                        "dangling reference for: parent "DFID", child "
3514                        DFID", comp_id %u, ea_off %u, ost_idx %u, %s: rc = %d\n",
3515                        lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
3516                        comp_id, ea_off, ost_idx,
3517                        (lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ) ?
3518                                 "Create the lost OST-object as required" :
3519                                 "Keep the MDT-object there by default", rc);
3520
3521         return rc;
3522 }
3523
3524 /* If the OST-object does not recognize the MDT-object as its parent, and
3525  * there is no other MDT-object claims as its parent, then just trust the
3526  * given MDT-object as its parent. So update the OST-object filter_fid. */
3527 static int lfsck_layout_repair_unmatched_pair(const struct lu_env *env,
3528                                               struct lfsck_component *com,
3529                                               struct dt_object *parent,
3530                                               struct lfsck_layout_req *llr,
3531                                               struct lu_attr *la)
3532 {
3533         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3534         struct filter_fid               *ff     = &info->lti_ff;
3535         struct ost_layout               *ol     = &ff->ff_layout;
3536         struct dt_object                *child  = llr->llr_child;
3537         struct dt_device                *dev    = lfsck_obj2dev(child);
3538         const struct lu_fid             *tfid   = lu_object_fid(&parent->do_lu);
3539         struct lu_buf                   *tbuf   = &info->lti_big_buf;
3540         struct thandle                  *handle;
3541         struct lu_buf                   *buf;
3542         struct lustre_handle             lh     = { 0 };
3543         int                              rc;
3544         ENTRY;
3545
3546         rc = lfsck_ibits_lock(env, com->lc_lfsck, parent, &lh,
3547                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
3548                               LCK_EX);
3549         if (rc != 0)
3550                 GOTO(log, rc);
3551
3552         ff->ff_parent.f_seq = cpu_to_le64(tfid->f_seq);
3553         ff->ff_parent.f_oid = cpu_to_le32(tfid->f_oid);
3554         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
3555          * MDT-object's FID::f_ver, instead it is the OST-object index in its
3556          * parent MDT-object's layout EA. */
3557         ff->ff_parent.f_stripe_idx = cpu_to_le32(llr->llr_lov_idx);
3558
3559         rc = lfsck_layout_get_lovea(env, parent, tbuf);
3560         if (rc < 0)
3561                 GOTO(unlock1, rc);
3562
3563         rc = lfsck_lmm2layout(tbuf->lb_buf, ol, llr->llr_comp_id);
3564         if (rc)
3565                 GOTO(unlock1, rc);
3566
3567         buf = lfsck_buf_get(env, ff, sizeof(*ff));
3568
3569         handle = dt_trans_create(env, dev);
3570         if (IS_ERR(handle))
3571                 GOTO(unlock1, rc = PTR_ERR(handle));
3572
3573         rc = dt_declare_xattr_set(env, child, buf, XATTR_NAME_FID, 0, handle);
3574         if (rc != 0)
3575                 GOTO(stop, rc);
3576
3577         rc = dt_attr_get(env, parent, la);
3578         if (rc != 0)
3579                 GOTO(stop, rc);
3580
3581         la->la_valid = LA_UID | LA_GID;
3582         rc = dt_declare_attr_set(env, child, la, handle);
3583         if (rc != 0)
3584                 GOTO(stop, rc);
3585
3586         rc = dt_trans_start_local(env, dev, handle);
3587         if (rc != 0)
3588                 GOTO(stop, rc);
3589
3590         dt_write_lock(env, parent, 0);
3591         if (unlikely(lfsck_is_dead_obj(parent)))
3592                 GOTO(unlock2, rc = 1);
3593
3594         rc = dt_xattr_set(env, child, buf, XATTR_NAME_FID, 0, handle);
3595         if (rc != 0)
3596                 GOTO(unlock2, rc);
3597
3598         /* Get the latest parent's owner. */
3599         rc = dt_attr_get(env, parent, la);
3600         if (rc != 0)
3601                 GOTO(unlock2, rc);
3602
3603         la->la_valid = LA_UID | LA_GID;
3604         rc = dt_attr_set(env, child, la, handle);
3605
3606         GOTO(unlock2, rc);
3607
3608 unlock2:
3609         dt_write_unlock(env, parent);
3610
3611 stop:
3612         rc = lfsck_layout_trans_stop(env, dev, handle, rc);
3613
3614 unlock1:
3615         lfsck_ibits_unlock(&lh, LCK_EX);
3616
3617 log:
3618         if (rc)
3619                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired "
3620                        "unmatched MDT-OST pair for: parent "DFID
3621                        ", child "DFID", comp_id %u, OST-index %u, "
3622                        "stripe-index %u, owner %u/%u: rc = %d\n",
3623                        lfsck_lfsck2name(com->lc_lfsck),
3624                        PFID(lfsck_dto2fid(parent)),
3625                        PFID(lfsck_dto2fid(child)),
3626                        llr->llr_comp_id, llr->llr_ost_idx, llr->llr_lov_idx,
3627                        la->la_uid, la->la_gid, rc);
3628
3629         return rc;
3630 }
3631
3632 /* If there are more than one MDT-objects claim as the OST-object's parent,
3633  * and the OST-object only recognizes one of them, then we need to generate
3634  * new OST-object(s) with new fid(s) for the non-recognized MDT-object(s). */
3635 static int lfsck_layout_repair_multiple_references(const struct lu_env *env,
3636                                                    struct lfsck_component *com,
3637                                                    struct dt_object *parent,
3638                                                    struct lfsck_layout_req *llr,
3639                                                    struct lu_attr *la)
3640 {
3641         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3642         struct dt_allocation_hint       *hint   = &info->lti_hint;
3643         struct dt_object_format         *dof    = &info->lti_dof;
3644         struct ost_id                   *oi     = &info->lti_oi;
3645         struct lu_buf                   *buf    = &info->lti_big_buf;
3646         struct lfsck_instance           *lfsck  = com->lc_lfsck;
3647         struct dt_device                *dev;
3648         struct lu_device                *d      =
3649                                 &lfsck_obj2dev(llr->llr_child)->dd_lu_dev;
3650         struct lu_object                *o;
3651         struct lu_object                *n;
3652         struct dt_object                *child  = NULL;
3653         struct thandle                  *handle = NULL;
3654         struct lov_mds_md_v1            *lmm;
3655         struct lov_ost_data_v1          *objs;
3656         const struct lu_fid             *pfid   = lfsck_dto2fid(parent);
3657         struct lu_fid                    tfid;
3658         struct lustre_handle             lh     = { 0 };
3659         __u32                            magic;
3660         __u32                            index;
3661         int                              rc;
3662         ENTRY;
3663
3664         /* We use two separated transactions to repair the inconsistency.
3665          *
3666          * 1) create the child (OST-object).
3667          * 2) update the parent LOV EA according to the child's FID.
3668          *
3669          * If 1) succeed, but 2) failed or aborted, then such OST-object will be
3670          * handled as orphan when the layout LFSCK run next time.
3671          *
3672          * If 1) failed, but 2) succeed, then such OST-object will be re-created
3673          * as dangling referened case when the layout LFSCK run next time. */
3674
3675         /* The 1st transaction. */
3676         o = lu_object_anon(env, d, NULL);
3677         if (IS_ERR(o))
3678                 GOTO(log, rc = PTR_ERR(o));
3679
3680         n = lu_object_locate(o->lo_header, d->ld_type);
3681         if (unlikely(n == NULL)) {
3682                 lu_object_put_nocache(env, o);
3683
3684                 GOTO(log, rc = -EINVAL);
3685         }
3686
3687         child = container_of(n, struct dt_object, do_lu);
3688         memset(hint, 0, sizeof(*hint));
3689         rc = dt_attr_get(env, parent, la);
3690         if (rc != 0)
3691                 GOTO(log, rc);
3692
3693         la->la_valid = LA_UID | LA_GID;
3694         memset(dof, 0, sizeof(*dof));
3695
3696         dev = lfsck_obj2dev(child);
3697         handle = dt_trans_create(env, dev);
3698         if (IS_ERR(handle))
3699                 GOTO(log, rc = PTR_ERR(handle));
3700
3701         rc = dt_declare_create(env, child, la, hint, dof, handle);
3702         if (rc != 0)
3703                 GOTO(stop, rc);
3704
3705         rc = dt_trans_start_local(env, dev, handle);
3706         if (rc != 0)
3707                 GOTO(stop, rc);
3708
3709         rc = dt_create(env, child, la, hint, dof, handle);
3710         dt_trans_stop(env, dev, handle);
3711         handle = NULL;
3712         if (rc != 0)
3713                 GOTO(log, rc);
3714
3715         rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
3716                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
3717                               LCK_EX);
3718         if (rc != 0)
3719                 GOTO(log, rc);
3720
3721         /* The 2nd transaction. */
3722
3723         /* XXX: Generally, we should use bottom device (OSD) to update parent
3724          *      LOV EA. But because the LOD-object still references the wrong
3725          *      OSP-object that should be detached after the parent's LOV EA
3726          *      refreshed. Unfortunately, there is no suitable API for that.
3727          *      So we have to make the LOD to re-load the OSP-object(s) via
3728          *      replacing the LOV EA against the LOD-object.
3729          *
3730          *      Once the DNE2 patches have been landed, we can replace the
3731          *      LOD device with the OSD device. LU-6230. */
3732
3733         dev = lfsck->li_next;
3734         parent = lfsck_object_locate(dev, parent);
3735         if (IS_ERR(parent))
3736                 GOTO(log, rc = PTR_ERR(parent));
3737
3738         handle = dt_trans_create(env, dev);
3739         if (IS_ERR(handle))
3740                 GOTO(log, rc = PTR_ERR(handle));
3741
3742         rc = dt_declare_xattr_set(env, parent, buf, XATTR_NAME_LOV,
3743                                   LU_XATTR_REPLACE, handle);
3744         if (rc != 0)
3745                 GOTO(stop, rc);
3746
3747         rc = dt_trans_start_local(env, dev, handle);
3748         if (rc != 0)
3749                 GOTO(stop, rc);
3750
3751         dt_write_lock(env, parent, 0);
3752         if (unlikely(lfsck_is_dead_obj(parent)))
3753                 GOTO(unlock, rc = 0);
3754
3755         rc = lfsck_layout_get_lovea(env, parent, buf);
3756         if (unlikely(!rc || rc == -ENODATA))
3757                 GOTO(unlock, rc = 0);
3758
3759         lmm = buf->lb_buf;
3760         magic = le32_to_cpu(lmm->lmm_magic);
3761         if (magic == LOV_MAGIC_COMP_V1) {
3762                 struct lov_comp_md_v1 *lcm = buf->lb_buf;
3763                 struct lov_comp_md_entry_v1 *lcme;
3764                 __u16 count = le16_to_cpu(lcm->lcm_entry_count);
3765                 int i;
3766
3767                 LASSERT(llr->llr_comp_id != 0);
3768
3769                 for (i = 0; i < count; i++) {
3770                         lcme = &lcm->lcm_entries[i];
3771                         if (le32_to_cpu(lcme->lcme_id) == llr->llr_comp_id) {
3772                                 LASSERT(le32_to_cpu(lcme->lcme_flags) &
3773                                         LCME_FL_INIT);
3774
3775                                 le32_add_cpu(&lcm->lcm_layout_gen, 1);
3776                                 lmm = buf->lb_buf +
3777                                         le32_to_cpu(lcme->lcme_offset);
3778                                 magic = le32_to_cpu(lmm->lmm_magic);
3779                                 goto set;
3780                         }
3781                 }
3782
3783                 GOTO(unlock, rc = 0);
3784         }
3785
3786 set:
3787         if (magic == LOV_MAGIC_V1) {
3788                 objs = &lmm->lmm_objects[llr->llr_lov_idx];
3789         } else {
3790                 LASSERT(magic == LOV_MAGIC_V3);
3791                 objs =
3792                 &((struct lov_mds_md_v3 *)lmm)->lmm_objects[llr->llr_lov_idx];
3793         }
3794
3795         ostid_le_to_cpu(&objs->l_ost_oi, oi);
3796         index = le32_to_cpu(objs->l_ost_idx);
3797         rc = ostid_to_fid(&tfid, oi, index);
3798         /* Someone changed layout during the LFSCK, no need to repair then. */
3799         if (rc == 0 && !lu_fid_eq(&tfid, lu_object_fid(&llr->llr_child->do_lu)))
3800                 GOTO(unlock, rc = 0);
3801
3802         lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
3803         fid_to_ostid(lu_object_fid(&child->do_lu), oi);
3804         ostid_cpu_to_le(oi, &objs->l_ost_oi);
3805         objs->l_ost_gen = cpu_to_le32(0);
3806         objs->l_ost_idx = cpu_to_le32(llr->llr_ost_idx);
3807         rc = dt_xattr_set(env, parent, buf, XATTR_NAME_LOV,
3808                           LU_XATTR_REPLACE, handle);
3809
3810         GOTO(unlock, rc = (rc == 0 ? 1 : rc));
3811
3812 unlock:
3813         dt_write_unlock(env, parent);
3814
3815 stop:
3816         if (handle != NULL)
3817                 dt_trans_stop(env, dev, handle);
3818
3819 log:
3820         lfsck_ibits_unlock(&lh, LCK_EX);
3821         if (child != NULL)
3822                 lfsck_object_put(env, child);
3823
3824         if (rc)
3825                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired "
3826                        "multiple references for: parent "DFID", comp_id %u, "
3827                        "OST-index %u, stripe-index %u, owner %u/%u: rc = %d\n",
3828                        lfsck_lfsck2name(lfsck), PFID(pfid),
3829                        llr->llr_comp_id, llr->llr_ost_idx, llr->llr_lov_idx,
3830                        la->la_uid, la->la_gid, rc);
3831
3832         return rc;
3833 }
3834
3835 /* If the MDT-object and the OST-object have different owner information,
3836  * then trust the MDT-object, because the normal chown/chgrp handle order
3837  * is from MDT to OST, and it is possible that some chown/chgrp operation
3838  * is partly done. */
3839 static int lfsck_layout_repair_owner(const struct lu_env *env,
3840                                      struct lfsck_component *com,
3841                                      struct dt_object *parent,
3842                                      struct lfsck_layout_req *llr,
3843                                      struct lu_attr *pla,
3844                                      const struct lu_attr *cla)
3845 {
3846         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3847         struct lu_attr                  *tla    = &info->lti_la2;
3848         struct dt_object                *child  = llr->llr_child;
3849         struct dt_device                *dev    = lfsck_obj2dev(child);
3850         struct thandle                  *handle;
3851         int                              rc;
3852         ENTRY;
3853
3854         tla->la_uid = pla->la_uid;
3855         tla->la_gid = pla->la_gid;
3856         tla->la_valid = LA_UID | LA_GID;
3857         handle = dt_trans_create(env, dev);
3858         if (IS_ERR(handle))
3859                 GOTO(log, rc = PTR_ERR(handle));
3860
3861         rc = dt_declare_attr_set(env, child, tla, handle);
3862         if (rc != 0)
3863                 GOTO(stop, rc);
3864
3865         rc = dt_trans_start_local(env, dev, handle);
3866         if (rc != 0)
3867                 GOTO(stop, rc);
3868
3869         /* Use the dt_object lock to serialize with destroy and attr_set. */
3870         dt_read_lock(env, parent, 0);
3871         if (unlikely(lfsck_is_dead_obj(parent)))
3872                 GOTO(unlock, rc = 1);
3873
3874         /* Get the latest parent's owner. */
3875         rc = dt_attr_get(env, parent, pla);
3876         if (rc != 0)
3877                 GOTO(unlock, rc);
3878
3879         /* Some others chown/chgrp during the LFSCK, needs to do nothing. */
3880         if (unlikely(tla->la_uid != pla->la_uid ||
3881                      tla->la_gid != pla->la_gid))
3882                 rc = 1;
3883         else
3884                 rc = dt_attr_set(env, child, tla, handle);
3885
3886         GOTO(unlock, rc);
3887
3888 unlock:
3889         dt_read_unlock(env, parent);
3890
3891 stop:
3892         rc = lfsck_layout_trans_stop(env, dev, handle, rc);
3893
3894 log:
3895         if (rc != 0)
3896                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired "
3897                        "inconsistent file owner for: parent "DFID", child "DFID
3898                        ", OST-index %u, stripe-index %u, old owner %u/%u, "
3899                        "new owner %u/%u: rc = %d\n",
3900                        lfsck_lfsck2name(com->lc_lfsck),
3901                        PFID(lfsck_dto2fid(parent)), PFID(lfsck_dto2fid(child)),
3902                        llr->llr_ost_idx, llr->llr_lov_idx,
3903                        cla->la_uid, cla->la_gid, tla->la_uid, tla->la_gid, rc);
3904
3905         return rc;
3906 }
3907
3908 /* Check whether the OST-object correctly back points to the
3909  * MDT-object (@parent) via the XATTR_NAME_FID xattr (@pfid). */
3910 static int lfsck_layout_check_parent(const struct lu_env *env,
3911                                      struct lfsck_component *com,
3912                                      struct lfsck_assistant_object *lso,
3913                                      struct filter_fid *ff,
3914                                      const struct lu_fid *cfid,
3915                                      const struct lu_attr *cla,
3916                                      struct lfsck_layout_req *llr)
3917 {
3918         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3919         struct lu_buf                   *buf    = &info->lti_big_buf;
3920         struct lu_fid                   *pfid   = &info->lti_fid;
3921         struct dt_object                *tobj;
3922         struct lov_mds_md_v1            *lmm;
3923         struct lov_ost_data_v1          *objs;
3924         struct lustre_handle             lh     = { 0 };
3925         int                              rc;
3926         int                              i;
3927         __u32                            magic;
3928         __u32                            idx;
3929         __u16                            count;
3930         ENTRY;
3931
3932         *pfid = ff->ff_parent;
3933         idx = pfid->f_stripe_idx;
3934         pfid->f_ver = 0;
3935
3936         if (unlikely(!fid_is_sane(pfid)))
3937                 RETURN(LLIT_UNMATCHED_PAIR);
3938
3939         if (lu_fid_eq(pfid, &lso->lso_fid)) {
3940                 if (likely(llr->llr_lov_idx == idx))
3941                         RETURN(0);
3942
3943                 RETURN(LLIT_UNMATCHED_PAIR);
3944         }
3945
3946         tobj = lfsck_object_find_bottom(env, com->lc_lfsck, pfid);
3947         if (IS_ERR(tobj))
3948                 RETURN(PTR_ERR(tobj));
3949
3950         if (dt_object_exists(tobj) == 0 || lfsck_is_dead_obj(tobj) ||
3951             !S_ISREG(lfsck_object_type(tobj)))
3952                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
3953
3954         /* Load the tobj's layout EA, in spite of it is a local MDT-object or
3955          * remote one on another MDT. Then check whether the given OST-object
3956          * is in such layout. If yes, it is multiple referenced, otherwise it
3957          * is unmatched referenced case. */
3958         rc = lfsck_layout_get_lovea(env, tobj, buf);
3959         if (rc == 0 || rc == -ENOENT)
3960                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
3961
3962         if (rc < 0)
3963                 GOTO(out, rc);
3964
3965         lmm = buf->lb_buf;
3966         magic = le32_to_cpu(lmm->lmm_magic);
3967         if (magic == LOV_MAGIC_COMP_V1) {
3968                 struct lov_comp_md_v1 *lcm = buf->lb_buf;
3969                 struct lov_comp_md_entry_v1 *lcme;
3970
3971                 if (ff->ff_layout.ol_comp_id == 0)
3972                         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
3973
3974                 count = le16_to_cpu(lcm->lcm_entry_count);
3975                 for (i = 0; i < count; i++) {
3976                         lcme = &lcm->lcm_entries[i];
3977                         if (le32_to_cpu(lcme->lcme_id) ==
3978                             ff->ff_layout.ol_comp_id) {
3979                                 lmm = buf->lb_buf +
3980                                         le32_to_cpu(lcme->lcme_offset);
3981                                 magic = le32_to_cpu(lmm->lmm_magic);
3982                                 if (!(le32_to_cpu(lcme->lcme_flags) &
3983                                       LCME_FL_INIT))
3984                                         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
3985
3986                                 goto further;
3987                         }
3988                 }
3989
3990                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
3991         }
3992
3993 further:
3994         if (magic == LOV_MAGIC_V1) {
3995                 objs = &lmm->lmm_objects[0];
3996         } else {
3997                 LASSERT(magic == LOV_MAGIC_V3);
3998                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
3999         }
4000
4001         count = le16_to_cpu(lmm->lmm_stripe_count);
4002         for (i = 0; i < count; i++, objs++) {
4003                 struct lu_fid           *tfid   = &info->lti_fid2;
4004                 struct ost_id           *oi     = &info->lti_oi;
4005                 __u32                    idx2;
4006
4007                 if (lovea_slot_is_dummy(objs))
4008                         continue;
4009
4010                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
4011                 idx2 = le32_to_cpu(objs->l_ost_idx);
4012                 rc = ostid_to_fid(tfid, oi, idx2);
4013                 if (rc != 0) {
4014                         CDEBUG(D_LFSCK, "%s: the parent "DFID" contains "
4015                                "invalid layout EA at the slot %d, index %u\n",
4016                                lfsck_lfsck2name(com->lc_lfsck),
4017                                PFID(pfid), i, idx2);
4018
4019                         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4020                 }
4021
4022                 if (lu_fid_eq(cfid, tfid)) {
4023                         rc = lfsck_ibits_lock(env, com->lc_lfsck, tobj, &lh,
4024                                               MDS_INODELOCK_UPDATE |
4025                                               MDS_INODELOCK_LAYOUT |
4026                                               MDS_INODELOCK_XATTR,
4027                                               LCK_EX);
4028                         if (rc != 0)
4029                                 GOTO(out, rc);
4030
4031                         dt_read_lock(env, tobj, 0);
4032
4033                         /* For local MDT-object, re-check existence
4034                          * after taken the lock. */
4035                         if (!dt_object_remote(tobj)) {
4036                                 if (dt_object_exists(tobj) == 0 ||
4037                                     lfsck_is_dead_obj(tobj))
4038                                         rc = LLIT_UNMATCHED_PAIR;
4039                                 else
4040                                         rc = LLIT_MULTIPLE_REFERENCED;
4041
4042                                 GOTO(unlock, rc);
4043                         }
4044
4045                         /* For migration case, the new MDT-object and old
4046                          * MDT-object may reference the same OST-object at
4047                          * some migration internal time.
4048                          *
4049                          * For remote MDT-object, the local MDT may not know
4050                          * whether it has been removed or not.  Try checking
4051                          * for a non-existent xattr to check if this object
4052                          * has been been removed or not. */
4053                         rc = dt_xattr_get(env, tobj, &LU_BUF_NULL,
4054                                           XATTR_NAME_DUMMY);
4055                         if (unlikely(rc == -ENOENT || rc >= 0))
4056                                 rc = LLIT_UNMATCHED_PAIR;
4057                         else if (rc == -ENODATA)
4058                                 rc = LLIT_MULTIPLE_REFERENCED;
4059
4060                         GOTO(unlock, rc);
4061                 }
4062         }
4063
4064         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4065
4066 unlock:
4067         if (lustre_handle_is_used(&lh)) {
4068                 dt_read_unlock(env, tobj);
4069                 lfsck_ibits_unlock(&lh, LCK_EX);
4070         }
4071
4072 out:
4073         lfsck_object_put(env, tobj);
4074
4075         return rc;
4076 }
4077
4078 static int lfsck_layout_assistant_handler_p1(const struct lu_env *env,
4079                                              struct lfsck_component *com,
4080                                              struct lfsck_assistant_req *lar)
4081 {
4082         struct lfsck_layout_req              *llr    =
4083                         container_of0(lar, struct lfsck_layout_req, llr_lar);
4084         struct lfsck_assistant_object        *lso    = lar->lar_parent;
4085         struct lfsck_layout                  *lo     = com->lc_file_ram;
4086         struct lfsck_thread_info             *info   = lfsck_env_info(env);
4087         struct filter_fid                    *ff     = &info->lti_ff;
4088         struct lu_buf buf = { .lb_buf = ff,
4089                               .lb_len = sizeof(*ff) };
4090         struct dt_object                     *parent = NULL;
4091         struct dt_object                     *child  = llr->llr_child;
4092         struct lu_attr                       *pla    = &lso->lso_attr;
4093         struct lu_attr                       *cla    = &info->lti_la;
4094         struct lfsck_instance                *lfsck  = com->lc_lfsck;
4095         struct lfsck_bookmark                *bk     = &lfsck->li_bookmark_ram;
4096         enum lfsck_layout_inconsistency_type  type   = LLIT_NONE;
4097         int                                   rc;
4098         ENTRY;
4099
4100         if (lso->lso_dead)
4101                 RETURN(0);
4102
4103         CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_ASSISTANT_DIRECT, cfs_fail_val);
4104
4105         rc = dt_attr_get(env, child, cla);
4106         if (rc == -ENOENT) {
4107                 parent = lfsck_assistant_object_load(env, lfsck, lso);
4108                 if (IS_ERR(parent)) {
4109                         rc = PTR_ERR(parent);
4110
4111                         RETURN(rc == -ENOENT ? 0 : rc);
4112                 }
4113
4114                 type = LLIT_DANGLING;
4115                 goto repair;
4116         }
4117
4118         if (rc != 0)
4119                 GOTO(out, rc);
4120
4121         lfsck_buf_init(&buf, ff, sizeof(*ff));
4122         rc = dt_xattr_get(env, child, &buf, XATTR_NAME_FID);
4123         if (unlikely(rc > 0 && rc < sizeof(struct lu_fid))) {
4124                 type = LLIT_UNMATCHED_PAIR;
4125                 goto repair;
4126         }
4127
4128         if (rc < 0 && rc != -ENODATA)
4129                 GOTO(out, rc);
4130
4131         if (rc == 0 || rc == -ENODATA)
4132                 GOTO(check_owner, rc = 0);
4133
4134         filter_fid_le_to_cpu(ff, ff, sizeof(*ff));
4135         rc = lfsck_layout_check_parent(env, com, lso, ff,
4136                                        lu_object_fid(&child->do_lu), cla, llr);
4137         if (rc > 0) {
4138                 type = rc;
4139                 goto repair;
4140         }
4141
4142         if (rc < 0)
4143                 GOTO(out, rc);
4144
4145 check_owner:
4146         /* Someone may has changed the owner after the parent attr pre-loaded.
4147          * It can be handled later inside the lfsck_layout_repair_owner(). */
4148         if (unlikely(cla->la_uid != pla->la_uid ||
4149                      cla->la_gid != pla->la_gid)) {
4150                 type = LLIT_INCONSISTENT_OWNER;
4151                 goto repair;
4152         }
4153
4154 repair:
4155         if (type == LLIT_NONE)
4156                 GOTO(out, rc = 0);
4157
4158         if (bk->lb_param & LPF_DRYRUN)
4159                 GOTO(out, rc = 1);
4160
4161         if (parent == NULL) {
4162                 parent = lfsck_assistant_object_load(env, lfsck, lso);
4163                 if (IS_ERR(parent)) {
4164                         rc = PTR_ERR(parent);
4165
4166                         if (rc == -ENOENT)
4167                                 RETURN(0);
4168
4169                         GOTO(out, rc);
4170                 }
4171         }
4172
4173         switch (type) {
4174         case LLIT_DANGLING:
4175                 if (bk->lb_param & LPF_DELAY_CREATE_OSTOBJ)
4176                         rc = lfsck_layout_ins_dangling_rec(env, com,
4177                                 lfsck_dto2fid(parent), lfsck_dto2fid(child),
4178                                 llr->llr_comp_id, llr->llr_lov_idx,
4179                                 llr->llr_ost_idx);
4180                 else
4181                         rc = __lfsck_layout_repair_dangling(env, com, parent,
4182                                                             llr->llr_child,
4183                                                             llr->llr_comp_id,
4184                                                             llr->llr_lov_idx,
4185                                                             llr->llr_ost_idx,
4186                                                             true);
4187                 break;
4188         case LLIT_UNMATCHED_PAIR:
4189                 rc = lfsck_layout_repair_unmatched_pair(env, com, parent,
4190                                                         llr, pla);
4191                 break;
4192         case LLIT_MULTIPLE_REFERENCED:
4193                 rc = lfsck_layout_repair_multiple_references(env, com, parent,
4194                                                              llr, pla);
4195                 break;
4196         case LLIT_INCONSISTENT_OWNER:
4197                 rc = lfsck_layout_repair_owner(env, com, parent, llr, pla, cla);
4198                 break;
4199         default:
4200                 rc = 0;
4201                 break;
4202         }
4203
4204         GOTO(out, rc);
4205
4206 out:
4207         down_write(&com->lc_sem);
4208         if (rc < 0) {
4209                 struct lfsck_assistant_data *lad = com->lc_data;
4210
4211                 if (unlikely(lad->lad_exit)) {
4212                         rc = 0;
4213                 } else if (rc == -ENOTCONN || rc == -ESHUTDOWN ||
4214                            rc == -ETIMEDOUT || rc == -EHOSTDOWN ||
4215                            rc == -EHOSTUNREACH) {
4216                         /* If cannot touch the target server,
4217                          * mark the LFSCK as INCOMPLETE. */
4218                         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant fail to "
4219                                "talk with OST %x: rc = %d\n",
4220                                lfsck_lfsck2name(lfsck), llr->llr_ost_idx, rc);
4221                         lfsck_lad_set_bitmap(env, com, llr->llr_ost_idx);
4222                         lo->ll_objs_skipped++;
4223                         rc = 0;
4224                 } else {
4225                         lfsck_layout_record_failure(env, lfsck, lo);
4226                 }
4227         } else if (rc > 0 && (type != LLIT_DANGLING ||
4228                               !(bk->lb_param & LPF_DELAY_CREATE_OSTOBJ))) {
4229                 LASSERTF(type > LLIT_NONE && type <= LLIT_MAX,
4230                          "unknown type = %d\n", type);
4231
4232                 lo->ll_objs_repaired[type - 1]++;
4233                 if (bk->lb_param & LPF_DRYRUN &&
4234                     unlikely(lo->ll_pos_first_inconsistent == 0))
4235                         lo->ll_pos_first_inconsistent =
4236                         lfsck->li_obj_oit->do_index_ops->dio_it.store(env,
4237                                                         lfsck->li_di_oit);
4238         }
4239         up_write(&com->lc_sem);
4240
4241         if (parent != NULL && !IS_ERR(parent))
4242                 lfsck_object_put(env, parent);
4243
4244         return rc;
4245 }
4246
4247 static int
4248 lfsck_layout_double_scan_one_trace_file(const struct lu_env *env,
4249                                         struct lfsck_component *com,
4250                                         struct dt_object *obj, bool first)
4251 {
4252         struct lfsck_instance *lfsck = com->lc_lfsck;
4253         struct ptlrpc_thread *thread = &lfsck->li_thread;
4254         struct lfsck_bookmark *bk = &lfsck->li_bookmark_ram;
4255         struct lfsck_layout *lo = com->lc_file_ram;
4256         const struct dt_it_ops *iops = &obj->do_index_ops->dio_it;
4257         struct dt_it *di;
4258         struct dt_key *key;
4259         struct lfsck_layout_dangling_key *parent =
4260                                         &lfsck_env_info(env)->lti_lldk;
4261         struct lu_fid *cfid = &lfsck_env_info(env)->lti_fid3;
4262         __u32 ost_idx;
4263         int rc;
4264         ENTRY;
4265
4266         di = iops->init(env, obj, 0);
4267         if (IS_ERR(di))
4268                 RETURN(PTR_ERR(di));
4269
4270         if (first)
4271                 lldk_cpu_to_be(parent, &lo->ll_lldk_latest_scanned_phase2);
4272         else
4273                 memset(parent, 0, sizeof(*parent));
4274         rc = iops->get(env, di, (const struct dt_key *)parent);
4275         if (rc < 0)
4276                 GOTO(fini, rc);
4277
4278         if (first) {
4279                 /* The start one either has been processed or does not exist,
4280                  * skip it. */
4281                 rc = iops->next(env, di);
4282                 if (rc != 0)
4283                         GOTO(put, rc);
4284         }
4285
4286         do {
4287                 if (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val) &&
4288                     unlikely(!thread_is_running(thread)))
4289                         GOTO(put, rc = 0);
4290
4291                 key = iops->key(env, di);
4292                 if (IS_ERR(key)) {
4293                         rc = PTR_ERR(key);
4294                         if (rc == -ENOENT)
4295                                 GOTO(put, rc = 1);
4296
4297                         goto checkpoint;
4298                 }
4299
4300                 lldk_be_to_cpu(parent,
4301                                 (const struct lfsck_layout_dangling_key *)key);
4302                 if (!fid_is_sane(&parent->lldk_fid)) {
4303                         rc = 0;
4304                         goto checkpoint;
4305                 }
4306
4307                 rc = iops->rec(env, di, (struct dt_rec *)cfid, 0);
4308                 if (rc == 0) {
4309                         fid_be_to_cpu(cfid, cfid);
4310                         ost_idx = cfid->f_ver;
4311                         cfid->f_ver = 0;
4312                         if (!fid_is_sane(cfid)) {
4313                                 rc = 0;
4314                                 goto checkpoint;
4315                         }
4316
4317                         rc = lfsck_layout_repair_dangling(env, com,
4318                                         &parent->lldk_fid, cfid,
4319                                         parent->lldk_comp_id,
4320                                         parent->lldk_ea_off, ost_idx);
4321                 }
4322
4323 checkpoint:
4324                 down_write(&com->lc_sem);
4325                 com->lc_new_checked++;
4326                 com->lc_new_scanned++;
4327                 if (rc >= 0)
4328                         lo->ll_lldk_latest_scanned_phase2 = *parent;
4329
4330                 if (rc > 0)
4331                         lo->ll_objs_repaired[LLIT_DANGLING - 1]++;
4332                 else if (rc < 0)
4333                         lo->ll_objs_failed_phase2++;
4334                 up_write(&com->lc_sem);
4335
4336                 if (rc < 0 && bk->lb_param & LPF_FAILOUT)
4337                         GOTO(put, rc);
4338
4339                 if (unlikely(cfs_time_beforeq(com->lc_time_next_checkpoint,
4340                                               cfs_time_current())) &&
4341                     com->lc_new_checked != 0) {
4342                         down_write(&com->lc_sem);
4343                         lo->ll_run_time_phase2 +=
4344                                 cfs_duration_sec(cfs_time_current() +
4345                                 HALF_SEC - com->lc_time_last_checkpoint);
4346                         lo->ll_time_last_checkpoint = cfs_time_current_sec();
4347                         lo->ll_objs_checked_phase2 += com->lc_new_checked;
4348                         com->lc_new_checked = 0;
4349                         lfsck_layout_store(env, com);
4350                         up_write(&com->lc_sem);
4351
4352                         com->lc_time_last_checkpoint = cfs_time_current();
4353                         com->lc_time_next_checkpoint =
4354                                 com->lc_time_last_checkpoint +
4355                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
4356                 }
4357
4358                 lfsck_control_speed_by_self(com);
4359                 if (unlikely(!thread_is_running(thread)))
4360                         GOTO(put, rc = 0);
4361
4362                 rc = iops->next(env, di);
4363         } while (rc == 0);
4364
4365         GOTO(put, rc);
4366
4367 put:
4368         iops->put(env, di);
4369
4370 fini:
4371         iops->fini(env, di);
4372
4373         return rc;
4374 }
4375
4376 static int lfsck_layout_assistant_handler_p2(const struct lu_env *env,
4377                                              struct lfsck_component *com)
4378 {
4379         struct lfsck_assistant_data     *lad    = com->lc_data;
4380         struct lfsck_instance           *lfsck  = com->lc_lfsck;
4381         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
4382         struct lfsck_tgt_descs          *ltds   = &lfsck->li_ost_descs;
4383         struct lfsck_tgt_desc           *ltd;
4384         int                              rc     = 0;
4385         ENTRY;
4386
4387         CDEBUG(D_LFSCK, "%s: layout LFSCK phase2 scan start\n",
4388                lfsck_lfsck2name(lfsck));
4389
4390         spin_lock(&ltds->ltd_lock);
4391         while (!list_empty(&lad->lad_ost_phase2_list)) {
4392                 ltd = list_entry(lad->lad_ost_phase2_list.next,
4393                                  struct lfsck_tgt_desc,
4394                                  ltd_layout_phase_list);
4395                 list_del_init(&ltd->ltd_layout_phase_list);
4396                 if (bk->lb_param & LPF_OST_ORPHAN) {
4397                         spin_unlock(&ltds->ltd_lock);
4398                         rc = lfsck_layout_scan_orphan(env, com, ltd);
4399                         if (rc != 0 && bk->lb_param & LPF_FAILOUT)
4400                                 RETURN(rc);
4401
4402                         if (unlikely(lad->lad_exit ||
4403                                      !thread_is_running(&lfsck->li_thread)))
4404                                 RETURN(0);
4405                         spin_lock(&ltds->ltd_lock);
4406                 }
4407         }
4408
4409         if (list_empty(&lad->lad_ost_phase1_list))
4410                 rc = 1;
4411         else
4412                 rc = 0;
4413         spin_unlock(&ltds->ltd_lock);
4414
4415         if (rc == 1 && bk->lb_param & LPF_OST_ORPHAN) {
4416                 struct lfsck_layout *lo = com->lc_file_ram;
4417                 int i;
4418
4419                 com->lc_new_checked = 0;
4420                 com->lc_new_scanned = 0;
4421                 com->lc_time_last_checkpoint = cfs_time_current();
4422                 com->lc_time_next_checkpoint = com->lc_time_last_checkpoint +
4423                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
4424
4425                 i = lfsck_sub_trace_file_fid2idx(
4426                                 &lo->ll_lldk_latest_scanned_phase2.lldk_fid);
4427                 rc = lfsck_layout_double_scan_one_trace_file(env, com,
4428                                 com->lc_sub_trace_objs[i].lsto_obj, true);
4429                 while (rc > 0 && ++i < LFSCK_STF_COUNT)
4430                         rc = lfsck_layout_double_scan_one_trace_file(env, com,
4431                                 com->lc_sub_trace_objs[i].lsto_obj, false);
4432
4433                 CDEBUG(D_LFSCK, "%s: layout LFSCK phase2 scan dangling stop "
4434                        "at the No. %d trace file: rc = %d\n",
4435                        lfsck_lfsck2name(lfsck), i, rc);
4436         }
4437
4438         CDEBUG(D_LFSCK, "%s: layout LFSCK phase2 scan stop: rc = %d\n",
4439                lfsck_lfsck2name(lfsck), rc);
4440
4441         RETURN(rc);
4442 }
4443
4444 static int
4445 lfsck_layout_slave_async_interpret(const struct lu_env *env,
4446                                    struct ptlrpc_request *req,
4447                                    void *args, int rc)
4448 {
4449         struct lfsck_layout_slave_async_args *llsaa = args;
4450         struct obd_export                    *exp   = llsaa->llsaa_exp;
4451         struct lfsck_component               *com   = llsaa->llsaa_com;
4452         struct lfsck_layout_slave_target     *llst  = llsaa->llsaa_llst;
4453         struct lfsck_layout_slave_data       *llsd  = com->lc_data;
4454         struct lfsck_reply                   *lr    = NULL;
4455         bool                                  done  = false;
4456
4457         if (rc != 0) {
4458                 /* It is probably caused by network trouble, or target crash,
4459                  * it will try several times (depends on the obd_timeout, and
4460                  * will not less than 3 times). But to make the LFSCK can go
4461                  * ahead, we should not try for ever. After some try but still
4462                  * hit failure, it will assume that the target exit the LFSCK
4463                  * prcoessing and stop try. */
4464                 if (rc == -ENOTCONN || rc == -ESHUTDOWN) {
4465                         int max_try = max_t(int, obd_timeout / 30, 3);
4466
4467                         if (++(llst->llst_failures) > max_try)
4468                                 done = true;
4469                 } else {
4470                         done = true;
4471                 }
4472         } else {
4473                 llst->llst_failures = 0;
4474                 lr = req_capsule_server_get(&req->rq_pill, &RMF_LFSCK_REPLY);
4475                 if (lr->lr_status != LS_SCANNING_PHASE1 &&
4476                     lr->lr_status != LS_SCANNING_PHASE2)
4477                         done = true;
4478         }
4479
4480         if (done) {
4481                 CDEBUG(D_LFSCK, "%s: layout LFSCK slave gets the MDT %x "
4482                        "status %d, failures_try %d\n", lfsck_lfsck2name(com->lc_lfsck),
4483                        llst->llst_index, lr != NULL ? lr->lr_status : rc,
4484                        llst->llst_failures);
4485
4486                 lfsck_layout_llst_del(llsd, llst);
4487         }
4488
4489         lfsck_layout_llst_put(llst);
4490         lfsck_component_put(env, com);
4491         class_export_put(exp);
4492
4493         return 0;
4494 }
4495
4496 static int lfsck_layout_async_query(const struct lu_env *env,
4497                                     struct lfsck_component *com,
4498                                     struct obd_export *exp,
4499                                     struct lfsck_layout_slave_target *llst,
4500                                     struct lfsck_request *lr,
4501                                     struct ptlrpc_request_set *set)
4502 {
4503         struct lfsck_layout_slave_async_args *llsaa;
4504         struct ptlrpc_request                *req;
4505         struct lfsck_request                 *tmp;
4506         int                                   rc;
4507         ENTRY;
4508
4509         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_QUERY);
4510         if (req == NULL)
4511                 RETURN(-ENOMEM);
4512
4513         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_QUERY);
4514         if (rc != 0) {
4515                 ptlrpc_request_free(req);
4516                 RETURN(rc);
4517         }
4518
4519         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
4520         *tmp = *lr;
4521         ptlrpc_request_set_replen(req);
4522
4523         llsaa = ptlrpc_req_async_args(req);
4524         llsaa->llsaa_exp = exp;
4525         llsaa->llsaa_com = lfsck_component_get(com);
4526         llsaa->llsaa_llst = llst;
4527         req->rq_interpret_reply = lfsck_layout_slave_async_interpret;
4528         req->rq_allow_intr = 1;
4529         ptlrpc_set_add_req(set, req);
4530
4531         RETURN(0);
4532 }
4533
4534 static int lfsck_layout_async_notify(const struct lu_env *env,
4535                                      struct obd_export *exp,
4536                                      struct lfsck_request *lr,
4537                                      struct ptlrpc_request_set *set)
4538 {
4539         struct ptlrpc_request   *req;
4540         struct lfsck_request    *tmp;
4541         int                      rc;
4542         ENTRY;
4543
4544         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
4545         if (req == NULL)
4546                 RETURN(-ENOMEM);
4547
4548         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
4549         if (rc != 0) {
4550                 ptlrpc_request_free(req);
4551                 RETURN(rc);
4552         }
4553
4554         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
4555         *tmp = *lr;
4556         ptlrpc_request_set_replen(req);
4557         req->rq_allow_intr = 1;
4558         ptlrpc_set_add_req(set, req);
4559
4560         RETURN(0);
4561 }
4562
4563 static int
4564 lfsck_layout_slave_query_master(const struct lu_env *env,
4565                                 struct lfsck_component *com)
4566 {
4567         struct lfsck_request             *lr    = &lfsck_env_info(env)->lti_lr;
4568         struct lfsck_instance            *lfsck = com->lc_lfsck;
4569         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
4570         struct lfsck_layout_slave_target *llst;
4571         struct obd_export                *exp;
4572         struct ptlrpc_request_set        *set;
4573         int                               rc    = 0;
4574         int                               rc1   = 0;
4575         ENTRY;
4576
4577         set = ptlrpc_prep_set();
4578         if (set == NULL)
4579                 GOTO(log, rc = -ENOMEM);
4580
4581         memset(lr, 0, sizeof(*lr));
4582         lr->lr_event = LE_QUERY;
4583         lr->lr_active = LFSCK_TYPE_LAYOUT;
4584
4585         llsd->llsd_touch_gen++;
4586         spin_lock(&llsd->llsd_lock);
4587         while (!list_empty(&llsd->llsd_master_list)) {
4588                 llst = list_entry(llsd->llsd_master_list.next,
4589                                   struct lfsck_layout_slave_target,
4590                                   llst_list);
4591                 if (llst->llst_gen == llsd->llsd_touch_gen)
4592                         break;
4593
4594                 llst->llst_gen = llsd->llsd_touch_gen;
4595                 list_move_tail(&llst->llst_list,
4596                                &llsd->llsd_master_list);
4597                 atomic_inc(&llst->llst_ref);
4598                 spin_unlock(&llsd->llsd_lock);
4599
4600                 exp = lustre_find_lwp_by_index(lfsck->li_obd->obd_name,
4601                                                llst->llst_index);
4602                 if (exp == NULL) {
4603                         lfsck_layout_llst_del(llsd, llst);
4604                         lfsck_layout_llst_put(llst);
4605                         spin_lock(&llsd->llsd_lock);
4606                         continue;
4607                 }
4608
4609                 rc = lfsck_layout_async_query(env, com, exp, llst, lr, set);
4610                 if (rc != 0) {
4611                         CDEBUG(D_LFSCK, "%s: layout LFSCK slave fail to "
4612                                "query %s for layout: rc = %d\n",
4613                                lfsck_lfsck2name(lfsck),
4614                                exp->exp_obd->obd_name, rc);
4615
4616                         rc1 = rc;
4617                         lfsck_layout_llst_put(llst);
4618                         class_export_put(exp);
4619                 }
4620                 spin_lock(&llsd->llsd_lock);
4621         }
4622         spin_unlock(&llsd->llsd_lock);
4623
4624         rc = ptlrpc_set_wait(set);
4625         ptlrpc_set_destroy(set);
4626
4627         GOTO(log, rc = (rc1 != 0 ? rc1 : rc));
4628
4629 log:
4630         CDEBUG(D_LFSCK, "%s: layout LFSCK slave queries master: rc = %d\n",
4631                lfsck_lfsck2name(com->lc_lfsck), rc);
4632
4633         return rc;
4634 }
4635
4636 static void
4637 lfsck_layout_slave_notify_master(const struct lu_env *env,
4638                                  struct lfsck_component *com,
4639                                  enum lfsck_events event, int result)
4640 {
4641         struct lfsck_layout              *lo    = com->lc_file_ram;
4642         struct lfsck_instance            *lfsck = com->lc_lfsck;
4643         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
4644         struct lfsck_request             *lr    = &lfsck_env_info(env)->lti_lr;
4645         struct lfsck_layout_slave_target *llst;
4646         struct obd_export                *exp;
4647         struct ptlrpc_request_set        *set;
4648         int                               rc;
4649         ENTRY;
4650
4651         CDEBUG(D_LFSCK, "%s: layout LFSCK slave notifies master\n",
4652                lfsck_lfsck2name(com->lc_lfsck));
4653
4654         set = ptlrpc_prep_set();
4655         if (set == NULL)
4656                 RETURN_EXIT;
4657
4658         memset(lr, 0, sizeof(*lr));
4659         lr->lr_event = event;
4660         lr->lr_flags = LEF_FROM_OST;
4661         lr->lr_status = result;
4662         lr->lr_index = lfsck_dev_idx(lfsck);
4663         lr->lr_active = LFSCK_TYPE_LAYOUT;
4664         lr->lr_flags2 = lo->ll_flags;
4665         llsd->llsd_touch_gen++;
4666         spin_lock(&llsd->llsd_lock);
4667         while (!list_empty(&llsd->llsd_master_list)) {
4668                 llst = list_entry(llsd->llsd_master_list.next,
4669                                   struct lfsck_layout_slave_target,
4670                                   llst_list);
4671                 if (llst->llst_gen == llsd->llsd_touch_gen)
4672                         break;
4673
4674                 llst->llst_gen = llsd->llsd_touch_gen;
4675                 list_move_tail(&llst->llst_list,
4676                                &llsd->llsd_master_list);
4677                 atomic_inc(&llst->llst_ref);
4678                 spin_unlock(&llsd->llsd_lock);
4679
4680                 exp = lustre_find_lwp_by_index(lfsck->li_obd->obd_name,
4681                                                llst->llst_index);
4682                 if (exp == NULL) {
4683                         lfsck_layout_llst_del(llsd, llst);
4684                         lfsck_layout_llst_put(llst);
4685                         spin_lock(&llsd->llsd_lock);
4686                         continue;
4687                 }
4688
4689                 rc = lfsck_layout_async_notify(env, exp, lr, set);
4690                 if (rc != 0)
4691                         CDEBUG(D_LFSCK, "%s: layout LFSCK slave fail to "
4692                                "notify %s for layout: rc = %d\n",
4693                                lfsck_lfsck2name(lfsck),
4694                                exp->exp_obd->obd_name, rc);
4695
4696                 lfsck_layout_llst_put(llst);
4697                 class_export_put(exp);
4698                 spin_lock(&llsd->llsd_lock);
4699         }
4700         spin_unlock(&llsd->llsd_lock);
4701
4702         ptlrpc_set_wait(set);
4703         ptlrpc_set_destroy(set);
4704
4705         RETURN_EXIT;
4706 }
4707
4708 /*
4709  * \ret -ENODATA: unrecognized stripe
4710  * \ret = 0     : recognized stripe
4711  * \ret < 0     : other failures
4712  */
4713 static int lfsck_layout_master_check_pairs(const struct lu_env *env,
4714                                            struct lfsck_component *com,
4715                                            struct lu_fid *cfid,
4716                                            struct lu_fid *pfid, __u32 comp_id)
4717 {
4718         struct lfsck_thread_info        *info   = lfsck_env_info(env);
4719         struct lu_buf                   *buf    = &info->lti_big_buf;
4720         struct ost_id                   *oi     = &info->lti_oi;
4721         struct dt_object                *obj;
4722         struct lov_mds_md_v1            *lmm;
4723         struct lov_ost_data_v1          *objs;
4724         __u32                            idx    = pfid->f_stripe_idx;
4725         __u32                            magic;
4726         int                              rc     = 0;
4727         int                              i;
4728         __u16                            count;
4729         ENTRY;
4730
4731         pfid->f_ver = 0;
4732         obj = lfsck_object_find_bottom(env, com->lc_lfsck, pfid);
4733         if (IS_ERR(obj))
4734                 RETURN(PTR_ERR(obj));
4735
4736         dt_read_lock(env, obj, 0);
4737         if (unlikely(dt_object_exists(obj) == 0 ||
4738                      lfsck_is_dead_obj(obj)))
4739                 GOTO(unlock, rc = -ENOENT);
4740
4741         if (!S_ISREG(lfsck_object_type(obj)))
4742                 GOTO(unlock, rc = -ENODATA);
4743
4744         rc = lfsck_layout_get_lovea(env, obj, buf);
4745         if (rc < 0)
4746                 GOTO(unlock, rc);
4747
4748         if (rc == 0)
4749                 GOTO(unlock, rc = -ENODATA);
4750
4751         lmm = buf->lb_buf;
4752         magic = le32_to_cpu(lmm->lmm_magic);
4753         if (magic == LOV_MAGIC_COMP_V1) {
4754                 struct lov_comp_md_v1 *lcm = buf->lb_buf;
4755                 struct lov_comp_md_entry_v1 *lcme;
4756
4757                 if (comp_id == 0)
4758                         GOTO(unlock, rc = -ENODATA);
4759
4760                 count = le16_to_cpu(lcm->lcm_entry_count);
4761                 for (i = 0; i < count; i++) {
4762                         lcme = &lcm->lcm_entries[i];
4763                         if (le32_to_cpu(lcme->lcme_id) == comp_id) {
4764                                 lmm = buf->lb_buf +
4765                                         le32_to_cpu(lcme->lcme_offset);
4766                                 magic = le32_to_cpu(lmm->lmm_magic);
4767                                 if (!(le32_to_cpu(lcme->lcme_flags) &
4768                                       LCME_FL_INIT))
4769                                         GOTO(unlock, rc = -ENODATA);
4770
4771                                 goto further;
4772                         }
4773                 }
4774
4775                 GOTO(unlock, rc = -ENODATA);
4776         }
4777
4778 further:
4779         if (magic == LOV_MAGIC_V1) {
4780                 objs = &lmm->lmm_objects[0];
4781         } else {
4782                 LASSERT(magic == LOV_MAGIC_V3);
4783                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
4784         }
4785
4786         fid_to_ostid(cfid, oi);
4787         count = le16_to_cpu(lmm->lmm_stripe_count);
4788         for (i = 0; i < count; i++, objs++) {
4789                 struct ost_id oi2;
4790
4791                 ostid_le_to_cpu(&objs->l_ost_oi, &oi2);
4792                 if (memcmp(oi, &oi2, sizeof(*oi)) == 0)
4793                         GOTO(unlock, rc = (i != idx ? -ENODATA : 0));
4794         }
4795
4796         GOTO(unlock, rc = -ENODATA);
4797
4798 unlock:
4799         dt_read_unlock(env, obj);
4800         lfsck_object_put(env, obj);
4801
4802         return rc;
4803 }
4804
4805 /*
4806  * The LFSCK-on-OST will ask the LFSCK-on-MDT to check whether the given
4807  * MDT-object/OST-object pairs match or not to aviod transfer MDT-object
4808  * layout EA from MDT to OST. On one hand, the OST no need to understand
4809  * the layout EA structure; on the other hand, it may cause trouble when
4810  * transfer large layout EA from MDT to OST via normal OUT RPC.
4811  *
4812  * \ret > 0: unrecognized stripe
4813  * \ret = 0: recognized stripe
4814  * \ret < 0: other failures
4815  */
4816 static int lfsck_layout_slave_check_pairs(const struct lu_env *env,
4817                                           struct lfsck_component *com,
4818                                           struct lu_fid *cfid,
4819                                           struct lu_fid *pfid, __u32 comp_id)
4820 {
4821         struct lfsck_instance    *lfsck  = com->lc_lfsck;
4822         struct obd_device        *obd    = lfsck->li_obd;
4823         struct seq_server_site   *ss     = lfsck_dev_site(lfsck);
4824         struct obd_export        *exp    = NULL;
4825         struct ptlrpc_request    *req    = NULL;
4826         struct lfsck_request     *lr;
4827         struct lu_seq_range      *range  = &lfsck_env_info(env)->lti_range;
4828         int                       rc     = 0;
4829         ENTRY;
4830
4831         if (unlikely(fid_is_idif(pfid)))
4832                 RETURN(1);
4833
4834         fld_range_set_any(range);
4835         rc = fld_server_lookup(env, ss->ss_server_fld, fid_seq(pfid), range);
4836         if (rc != 0)
4837                 RETURN(rc == -ENOENT ? 1 : rc);
4838
4839         if (unlikely(!fld_range_is_mdt(range)))
4840                 RETURN(1);
4841
4842         exp = lustre_find_lwp_by_index(obd->obd_name, range->lsr_index);
4843         if (unlikely(exp == NULL))
4844                 RETURN(1);
4845
4846         if (!(exp_connect_flags(exp) & OBD_CONNECT_LFSCK))
4847                 GOTO(out, rc = -EOPNOTSUPP);
4848
4849         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
4850         if (req == NULL)
4851                 GOTO(out, rc = -ENOMEM);
4852
4853         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
4854         if (rc != 0) {
4855                 ptlrpc_request_free(req);
4856
4857                 GOTO(out, rc);
4858         }
4859
4860         lr = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
4861         memset(lr, 0, sizeof(*lr));
4862         lr->lr_event = LE_PAIRS_VERIFY;
4863         lr->lr_active = LFSCK_TYPE_LAYOUT;
4864         lr->lr_fid = *cfid; /* OST-object itself FID. */
4865         lr->lr_fid2 = *pfid; /* The claimed parent FID. */
4866         lr->lr_comp_id = comp_id;
4867
4868         ptlrpc_request_set_replen(req);
4869         rc = ptlrpc_queue_wait(req);
4870         ptlrpc_req_finished(req);
4871
4872         if (rc == -ENOENT || rc == -ENODATA)
4873                 rc = 1;
4874
4875         GOTO(out, rc);
4876
4877 out:
4878         if (exp != NULL)
4879                 class_export_put(exp);
4880
4881         return rc;
4882 }
4883
4884 static int lfsck_layout_slave_repair_pfid(const struct lu_env *env,
4885                                           struct lfsck_component *com,
4886                                           struct lfsck_req_local *lrl)
4887 {
4888         struct dt_object        *obj;
4889         int                      rc     = 0;
4890         ENTRY;
4891
4892         obj = lfsck_object_find_bottom(env, com->lc_lfsck, &lrl->lrl_fid);
4893         if (IS_ERR(obj))
4894                 GOTO(log, rc = PTR_ERR(obj));
4895
4896         dt_write_lock(env, obj, 0);
4897         if (unlikely(dt_object_exists(obj) == 0 ||
4898                      lfsck_is_dead_obj(obj)))
4899                 GOTO(unlock, rc = 0);
4900
4901         rc = __lfsck_layout_update_pfid(env, obj, &lrl->lrl_ff_client.ff_parent,
4902                                         &lrl->lrl_ff_client.ff_layout,
4903                                         lrl->lrl_ff_client.ff_parent.f_ver);
4904
4905         GOTO(unlock, rc);
4906
4907 unlock:
4908         dt_write_unlock(env, obj);
4909         lfsck_object_put(env, obj);
4910
4911 log:
4912         CDEBUG(D_LFSCK, "%s: layout LFSCK slave repaired pfid for "DFID
4913                ", parent "DFID": rc = %d\n", lfsck_lfsck2name(com->lc_lfsck),
4914                PFID(&lrl->lrl_fid), PFID(&lrl->lrl_ff_client.ff_parent), rc);
4915
4916         return rc;
4917 }
4918
4919 /* layout APIs */
4920
4921 static void lfsck_layout_slave_quit(const struct lu_env *env,
4922                                     struct lfsck_component *com);
4923
4924 static int lfsck_layout_reset(const struct lu_env *env,
4925                               struct lfsck_component *com, bool init)
4926 {
4927         struct lfsck_layout     *lo    = com->lc_file_ram;
4928         int                      rc;
4929
4930         down_write(&com->lc_sem);
4931         if (init) {
4932                 memset(lo, 0, com->lc_file_size);
4933         } else {
4934                 __u32 count = lo->ll_success_count;
4935                 __u64 last_time = lo->ll_time_last_complete;
4936
4937                 memset(lo, 0, com->lc_file_size);
4938                 lo->ll_success_count = count;
4939                 lo->ll_time_last_complete = last_time;
4940         }
4941
4942         lo->ll_magic = LFSCK_LAYOUT_MAGIC;
4943         lo->ll_status = LS_INIT;
4944
4945         if (com->lc_lfsck->li_master) {
4946                 struct lfsck_assistant_data *lad = com->lc_data;
4947
4948                 lad->lad_incomplete = 0;
4949                 CFS_RESET_BITMAP(lad->lad_bitmap);
4950         }
4951
4952         rc = lfsck_layout_store(env, com);
4953         if (rc == 0 && com->lc_lfsck->li_master)
4954                 rc = lfsck_load_sub_trace_files(env, com,
4955                         &dt_lfsck_layout_dangling_features, LFSCK_LAYOUT, true);
4956         up_write(&com->lc_sem);
4957
4958         CDEBUG(D_LFSCK, "%s: layout LFSCK reset: rc = %d\n",
4959                lfsck_lfsck2name(com->lc_lfsck), rc);
4960
4961         return rc;
4962 }
4963
4964 static void lfsck_layout_fail(const struct lu_env *env,
4965                               struct lfsck_component *com, bool new_checked)
4966 {
4967         struct lfsck_layout *lo = com->lc_file_ram;
4968
4969         down_write(&com->lc_sem);
4970         if (new_checked)
4971                 com->lc_new_checked++;
4972         lfsck_layout_record_failure(env, com->lc_lfsck, lo);
4973         up_write(&com->lc_sem);
4974 }
4975
4976 static int lfsck_layout_master_checkpoint(const struct lu_env *env,
4977                                           struct lfsck_component *com, bool init)
4978 {
4979         struct lfsck_instance   *lfsck   = com->lc_lfsck;
4980         struct lfsck_layout     *lo      = com->lc_file_ram;
4981         int                      rc;
4982
4983         if (!init) {
4984                 rc = lfsck_checkpoint_generic(env, com);
4985                 if (rc != 0)
4986                         return rc > 0 ? 0 : rc;
4987         }
4988
4989         down_write(&com->lc_sem);
4990         if (init) {
4991                 lo->ll_pos_latest_start =
4992                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
4993         } else {
4994                 lo->ll_pos_last_checkpoint =
4995                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
4996                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
4997                                 HALF_SEC - lfsck->li_time_last_checkpoint);
4998                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
4999                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5000                 com->lc_new_checked = 0;
5001         }
5002
5003         rc = lfsck_layout_store(env, com);
5004         up_write(&com->lc_sem);
5005
5006         CDEBUG(D_LFSCK, "%s: layout LFSCK master checkpoint at the pos ["
5007                "%llu], status = %d: rc = %d\n", lfsck_lfsck2name(lfsck),
5008                lfsck->li_pos_current.lp_oit_cookie, lo->ll_status, rc);
5009
5010         return rc;
5011 }
5012
5013 static int lfsck_layout_slave_checkpoint(const struct lu_env *env,
5014                                          struct lfsck_component *com, bool init)
5015 {
5016         struct lfsck_instance   *lfsck = com->lc_lfsck;
5017         struct lfsck_layout     *lo    = com->lc_file_ram;
5018         int                      rc;
5019
5020         if (com->lc_new_checked == 0 && !init)
5021                 return 0;
5022
5023         down_write(&com->lc_sem);
5024         if (init) {
5025                 lo->ll_pos_latest_start =
5026                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5027         } else {
5028                 lo->ll_pos_last_checkpoint =
5029                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5030                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
5031                                 HALF_SEC - lfsck->li_time_last_checkpoint);
5032                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
5033                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5034                 com->lc_new_checked = 0;
5035         }
5036
5037         rc = lfsck_layout_store(env, com);
5038         up_write(&com->lc_sem);
5039
5040         CDEBUG(D_LFSCK, "%s: layout LFSCK slave checkpoint at the pos ["
5041                "%llu], status = %d: rc = %d\n", lfsck_lfsck2name(lfsck),
5042                lfsck->li_pos_current.lp_oit_cookie, lo->ll_status, rc);
5043
5044         return rc;
5045 }
5046
5047 static int lfsck_layout_prep(const struct lu_env *env,
5048                              struct lfsck_component *com,
5049                              struct lfsck_start *start)
5050 {
5051         struct lfsck_instance   *lfsck  = com->lc_lfsck;
5052         struct lfsck_layout     *lo     = com->lc_file_ram;
5053         struct lfsck_position   *pos    = &com->lc_pos_start;
5054
5055         fid_zero(&pos->lp_dir_parent);
5056         pos->lp_dir_cookie = 0;
5057         if (lo->ll_status == LS_COMPLETED ||
5058             lo->ll_status == LS_PARTIAL ||
5059             /* To handle orphan, must scan from the beginning. */
5060             (start != NULL && start->ls_flags & LPF_OST_ORPHAN)) {
5061                 int rc;
5062
5063                 rc = lfsck_layout_reset(env, com, false);
5064                 if (rc == 0)
5065                         rc = lfsck_set_param(env, lfsck, start, true);
5066
5067                 if (rc != 0) {
5068                         CDEBUG(D_LFSCK, "%s: layout LFSCK prep failed: "
5069                                "rc = %d\n", lfsck_lfsck2name(lfsck), rc);
5070
5071                         return rc;
5072                 }
5073         }
5074
5075         down_write(&com->lc_sem);
5076         lo->ll_time_latest_start = cfs_time_current_sec();
5077         spin_lock(&lfsck->li_lock);
5078         if (lo->ll_flags & LF_SCANNED_ONCE) {
5079                 if (!lfsck->li_drop_dryrun ||
5080                     lo->ll_pos_first_inconsistent == 0) {
5081                         lo->ll_status = LS_SCANNING_PHASE2;
5082                         list_move_tail(&com->lc_link,
5083                                        &lfsck->li_list_double_scan);
5084                         pos->lp_oit_cookie = 0;
5085                 } else {
5086                         int i;
5087
5088                         lo->ll_status = LS_SCANNING_PHASE1;
5089                         lo->ll_run_time_phase1 = 0;
5090                         lo->ll_run_time_phase2 = 0;
5091                         lo->ll_objs_checked_phase1 = 0;
5092                         lo->ll_objs_checked_phase2 = 0;
5093                         lo->ll_objs_failed_phase1 = 0;
5094                         lo->ll_objs_failed_phase2 = 0;
5095                         for (i = 0; i < LLIT_MAX; i++)
5096                                 lo->ll_objs_repaired[i] = 0;
5097
5098                         pos->lp_oit_cookie = lo->ll_pos_first_inconsistent;
5099                         fid_zero(&com->lc_fid_latest_scanned_phase2);
5100                 }
5101         } else {
5102                 lo->ll_status = LS_SCANNING_PHASE1;
5103                 if (!lfsck->li_drop_dryrun ||
5104                     lo->ll_pos_first_inconsistent == 0)
5105                         pos->lp_oit_cookie = lo->ll_pos_last_checkpoint + 1;
5106                 else
5107                         pos->lp_oit_cookie = lo->ll_pos_first_inconsistent;
5108         }
5109         spin_unlock(&lfsck->li_lock);
5110         up_write(&com->lc_sem);
5111
5112         return 0;
5113 }
5114
5115 static int lfsck_layout_slave_prep(const struct lu_env *env,
5116                                    struct lfsck_component *com,
5117                                    struct lfsck_start_param *lsp)
5118 {
5119         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
5120         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5121         struct lfsck_layout             *lo     = com->lc_file_ram;
5122         struct lfsck_start              *start  = lsp->lsp_start;
5123         int                              rc;
5124
5125         rc = lfsck_layout_prep(env, com, start);
5126         if (rc != 0)
5127                 return rc;
5128
5129         if (lo->ll_flags & LF_CRASHED_LASTID &&
5130             list_empty(&llsd->llsd_master_list)) {
5131                 LASSERT(lfsck->li_out_notify != NULL);
5132
5133                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
5134                                      LE_LASTID_REBUILDING);
5135         }
5136
5137         if (!lsp->lsp_index_valid)
5138                 return 0;
5139
5140         rc = lfsck_layout_llst_add(llsd, lsp->lsp_index);
5141         if (rc == 0 && start != NULL && start->ls_flags & LPF_OST_ORPHAN) {
5142                 LASSERT(!llsd->llsd_rbtree_valid);
5143
5144                 write_lock(&llsd->llsd_rb_lock);
5145                 rc = lfsck_rbtree_setup(env, com);
5146                 write_unlock(&llsd->llsd_rb_lock);
5147         }
5148
5149         CDEBUG(D_LFSCK, "%s: layout LFSCK slave prep done, start pos ["
5150                "%llu]\n", lfsck_lfsck2name(lfsck),
5151                com->lc_pos_start.lp_oit_cookie);
5152
5153         return rc;
5154 }
5155
5156 static int lfsck_layout_master_prep(const struct lu_env *env,
5157                                     struct lfsck_component *com,
5158                                     struct lfsck_start_param *lsp)
5159 {
5160         int rc;
5161         ENTRY;
5162
5163         rc = lfsck_layout_load_bitmap(env, com);
5164         if (rc != 0) {
5165                 rc = lfsck_layout_reset(env, com, false);
5166                 if (rc == 0)
5167                         rc = lfsck_set_param(env, com->lc_lfsck,
5168                                              lsp->lsp_start, true);
5169
5170                 if (rc != 0)
5171                         GOTO(log, rc);
5172         }
5173
5174         rc = lfsck_layout_prep(env, com, lsp->lsp_start);
5175         if (rc != 0)
5176                 RETURN(rc);
5177
5178         rc = lfsck_start_assistant(env, com, lsp);
5179
5180         GOTO(log, rc);
5181
5182 log:
5183         CDEBUG(D_LFSCK, "%s: layout LFSCK master prep done, start pos ["
5184                "%llu]\n", lfsck_lfsck2name(com->lc_lfsck),
5185                com->lc_pos_start.lp_oit_cookie);
5186
5187         return 0;
5188 }
5189
5190 /* Pre-fetch the attribute for each stripe in the given layout EA. */
5191 static int lfsck_layout_scan_stripes(const struct lu_env *env,
5192                                      struct lfsck_component *com,
5193                                      struct dt_object *parent,
5194                                      struct lov_mds_md_v1 *lmm, __u32 comp_id)
5195 {
5196         struct lfsck_thread_info        *info    = lfsck_env_info(env);
5197         struct lfsck_instance           *lfsck   = com->lc_lfsck;
5198         struct lfsck_bookmark           *bk      = &lfsck->li_bookmark_ram;
5199         struct lfsck_layout             *lo      = com->lc_file_ram;
5200         struct lfsck_assistant_data     *lad     = com->lc_data;
5201         struct lfsck_assistant_object   *lso     = NULL;
5202         struct lov_ost_data_v1          *objs;
5203         struct lfsck_tgt_descs          *ltds    = &lfsck->li_ost_descs;
5204         struct ptlrpc_thread            *mthread = &lfsck->li_thread;
5205         struct ptlrpc_thread            *athread = &lad->lad_thread;
5206         struct l_wait_info               lwi     = { 0 };
5207         struct lu_buf                    buf;
5208         int                              rc      = 0;
5209         int                              i;
5210         __u32                            magic;
5211         __u16                            count;
5212         ENTRY;
5213
5214         lfsck_buf_init(&buf, &info->lti_ff, sizeof(struct filter_fid));
5215         magic = le32_to_cpu(lmm->lmm_magic);
5216         if (magic == LOV_MAGIC_V1) {
5217                 objs = &lmm->lmm_objects[0];
5218         } else {
5219                 LASSERT(magic == LOV_MAGIC_V3);
5220                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
5221         }
5222
5223         count = le16_to_cpu(lmm->lmm_stripe_count);
5224         for (i = 0; i < count; i++, objs++) {
5225                 struct lu_fid           *fid    = &info->lti_fid;
5226                 struct ost_id           *oi     = &info->lti_oi;
5227                 struct lfsck_layout_req *llr;
5228                 struct lfsck_tgt_desc   *tgt    = NULL;
5229                 struct dt_object        *cobj   = NULL;
5230                 __u32                    index;
5231                 bool                     wakeup = false;
5232
5233                 if (unlikely(lovea_slot_is_dummy(objs)))
5234                         continue;
5235
5236                 l_wait_event(mthread->t_ctl_waitq,
5237                              lad->lad_prefetched < bk->lb_async_windows ||
5238                              !thread_is_running(mthread) ||
5239                              thread_is_stopped(athread),
5240                              &lwi);
5241
5242                 if (unlikely(!thread_is_running(mthread)) ||
5243                              thread_is_stopped(athread))
5244                         GOTO(out, rc = 0);
5245
5246                 if (unlikely(lfsck_is_dead_obj(parent)))
5247                         GOTO(out, rc = 0);
5248
5249                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
5250                 index = le32_to_cpu(objs->l_ost_idx);
5251                 rc = ostid_to_fid(fid, oi, index);
5252                 if (rc != 0) {
5253                         CDEBUG(D_LFSCK, "%s: get invalid layout EA for "DFID
5254                                ": "DOSTID", idx %u, comp_id %u\n",
5255                                lfsck_lfsck2name(lfsck),
5256                                PFID(lfsck_dto2fid(parent)), POSTID(oi),
5257                                index, comp_id);
5258                         goto next;
5259                 }
5260
5261                 tgt = lfsck_tgt_get(ltds, index);
5262                 if (unlikely(tgt == NULL)) {
5263                         CDEBUG(D_LFSCK, "%s: cannot talk with OST %x which "
5264                                "did not join the layout LFSCK, comp_id %u\n",
5265                                lfsck_lfsck2name(lfsck), index, comp_id);
5266                         lfsck_lad_set_bitmap(env, com, index);
5267                         goto next;
5268                 }
5269
5270                 /* There is potential deadlock race condition between object
5271                  * destroy and layout LFSCK. Consider the following scenario:
5272                  *
5273                  * 1) The LFSCK thread obtained the parent object firstly, at
5274                  *    that time, the parent object has not been destroyed yet.
5275                  *
5276                  * 2) One RPC service thread destroyed the parent and all its
5277                  *    children objects. Because the LFSCK is referencing the
5278                  *    parent object, then the parent object will be marked as
5279                  *    dying in RAM. On the other hand, the parent object is
5280                  *    referencing all its children objects, then all children
5281                  *    objects will be marked as dying in RAM also.
5282                  *
5283                  * 3) The LFSCK thread tries to find some child object with
5284                  *    the parent object referenced. Then it will find that the
5285                  *    child object is dying. According to the object visibility
5286                  *    rules: the object with dying flag cannot be returned to
5287                  *    others. So the LFSCK thread has to wait until the dying
5288                  *    object has been purged from RAM, then it can allocate a
5289                  *    new object (with the same FID) in RAM. Unfortunately, the
5290                  *    LFSCK thread itself is referencing the parent object, and
5291                  *    cause the parent object cannot be purged, then cause the
5292                  *    child object cannot be purged also. So the LFSCK thread
5293                  *    will fall into deadlock.
5294                  *
5295                  * We introduce non-blocked version lu_object_find() to allow
5296                  * the LFSCK thread to return failure immediately (instead of
5297                  * wait) when it finds dying (child) object, then the LFSCK
5298                  * thread can check whether the parent object is dying or not.
5299                  * So avoid above deadlock. LU-5395 */
5300                 cobj = lfsck_object_find_by_dev_nowait(env, tgt->ltd_tgt, fid);
5301                 if (IS_ERR(cobj)) {
5302                         if (lfsck_is_dead_obj(parent)) {
5303                                 lfsck_tgt_put(tgt);
5304
5305                                 GOTO(out, rc = 0);
5306                         }
5307
5308                         rc = PTR_ERR(cobj);
5309                         goto next;
5310                 }
5311
5312                 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_ASSISTANT_DIRECT)) {
5313                         rc = dt_declare_attr_get(env, cobj);
5314                         if (rc != 0)
5315                                 goto next;
5316
5317                         rc = dt_declare_xattr_get(env, cobj, &buf,
5318                                                   XATTR_NAME_FID);
5319                         if (rc != 0)
5320                                 goto next;
5321                 }
5322
5323                 if (lso == NULL) {
5324                         struct lu_attr *attr = &info->lti_la;
5325
5326                         rc = dt_attr_get(env, parent, attr);
5327                         if (rc != 0)
5328                                 goto next;
5329
5330                         lso = lfsck_assistant_object_init(env,
5331                                 lfsck_dto2fid(parent), attr,
5332                                 lfsck->li_pos_current.lp_oit_cookie, false);
5333                         if (IS_ERR(lso)) {
5334                                 rc = PTR_ERR(lso);
5335                                 lso = NULL;
5336
5337                                 goto next;
5338                         }
5339                 }
5340
5341                 llr = lfsck_layout_assistant_req_init(lso, cobj, comp_id,
5342                                                       index, i);
5343                 if (IS_ERR(llr)) {
5344                         rc = PTR_ERR(llr);
5345                         goto next;
5346                 }
5347
5348                 cobj = NULL;
5349                 spin_lock(&lad->lad_lock);
5350                 if (lad->lad_assistant_status < 0) {
5351                         spin_unlock(&lad->lad_lock);
5352                         lfsck_layout_assistant_req_fini(env, &llr->llr_lar);
5353                         lfsck_tgt_put(tgt);
5354                         RETURN(lad->lad_assistant_status);
5355                 }
5356
5357                 list_add_tail(&llr->llr_lar.lar_list, &lad->lad_req_list);
5358                 if (lad->lad_prefetched == 0)
5359                         wakeup = true;
5360
5361                 lad->lad_prefetched++;
5362                 spin_unlock(&lad->lad_lock);
5363                 if (wakeup)
5364                         wake_up_all(&athread->t_ctl_waitq);
5365
5366 next:
5367                 down_write(&com->lc_sem);
5368                 com->lc_new_checked++;
5369                 if (rc < 0)
5370                         lfsck_layout_record_failure(env, lfsck, lo);
5371                 up_write(&com->lc_sem);
5372
5373                 if (cobj != NULL && !IS_ERR(cobj))
5374                         lfsck_object_put(env, cobj);
5375
5376                 if (likely(tgt != NULL))
5377                         lfsck_tgt_put(tgt);
5378
5379                 if (rc < 0 && bk->lb_param & LPF_FAILOUT)
5380                         GOTO(out, rc);
5381         }
5382
5383         GOTO(out, rc = 0);
5384
5385 out:
5386         if (lso != NULL)
5387                 lfsck_assistant_object_put(env, lso);
5388
5389         return rc;
5390 }
5391
5392 /* For the given object, read its layout EA locally. For each stripe, pre-fetch
5393  * the OST-object's attribute and generate an structure lfsck_layout_req on the
5394  * list ::lad_req_list.
5395  *
5396  * For each request on above list, the lfsck_layout_assistant thread compares
5397  * the OST side attribute with local attribute, if inconsistent, then repair it.
5398  *
5399  * All above processing is async mode with pipeline. */
5400 static int lfsck_layout_master_exec_oit(const struct lu_env *env,
5401                                         struct lfsck_component *com,
5402                                         struct dt_object *obj)
5403 {
5404         struct lfsck_thread_info        *info   = lfsck_env_info(env);
5405         struct ost_id                   *oi     = &info->lti_oi;
5406         struct lfsck_layout             *lo     = com->lc_file_ram;
5407         struct lfsck_assistant_data     *lad    = com->lc_data;
5408         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5409         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
5410         struct thandle                  *handle = NULL;
5411         struct lu_buf                   *buf    = &info->lti_big_buf;
5412         struct lov_mds_md_v1            *lmm    = NULL;
5413         struct dt_device                *dev    = lfsck_obj2dev(obj);
5414         struct lustre_handle             lh     = { 0 };
5415         struct lu_buf                    ea_buf = { NULL };
5416         struct lov_comp_md_v1           *lcm    = NULL;
5417         struct lov_comp_md_entry_v1     *lcme   = NULL;
5418         int                              rc     = 0;
5419         int                              size   = 0;
5420         __u32                            magic  = 0;
5421         __u16                            count  = 0;
5422         bool                             locked = false;
5423         bool                             stripe = false;
5424         bool                             bad_oi = false;
5425         ENTRY;
5426
5427         if (!S_ISREG(lfsck_object_type(obj)))
5428                 GOTO(out, rc = 0);
5429
5430         if (lad->lad_assistant_status < 0)
5431                 GOTO(out, rc = -ESRCH);
5432
5433         fid_to_lmm_oi(lfsck_dto2fid(obj), oi);
5434         lmm_oi_cpu_to_le(oi, oi);
5435         dt_read_lock(env, obj, 0);
5436         locked = true;
5437
5438 again:
5439         bad_oi = false;
5440         if (dt_object_exists(obj) == 0 ||
5441             lfsck_is_dead_obj(obj))
5442                 GOTO(out, rc = 0);
5443
5444         rc = lfsck_layout_get_lovea(env, obj, buf);
5445         if (rc <= 0)
5446                 /* Skip bad lov EA during the 1st cycle scanning, and
5447                  * try to recover it via orphan in the 2nd scanning. */
5448                 GOTO(out, rc = (rc == -EINVAL ? 0 : rc));
5449
5450         size = rc;
5451         lmm = buf->lb_buf;
5452         magic = le32_to_cpu(lmm->lmm_magic);
5453         if (magic == LOV_MAGIC_COMP_V1) {
5454                 int i;
5455
5456                 lcm = buf->lb_buf;
5457                 count = le16_to_cpu(lcm->lcm_entry_count);
5458                 for (i = 0; i < count; i++) {
5459                         lcme = &lcm->lcm_entries[i];
5460                         lmm = buf->lb_buf + le32_to_cpu(lcme->lcme_offset);
5461                         if (memcmp(oi, &lmm->lmm_oi, sizeof(*oi)) != 0)
5462                                 goto fix;
5463                 }
5464
5465                 GOTO(out, stripe = true);
5466         } else if (memcmp(oi, &lmm->lmm_oi, sizeof(*oi)) == 0) {
5467                 GOTO(out, stripe = true);
5468         }
5469
5470 fix:
5471         /* Inconsistent lmm_oi, should be repaired. */
5472         bad_oi = true;
5473
5474         if (bk->lb_param & LPF_DRYRUN) {
5475                 lo->ll_objs_repaired[LLIT_OTHERS - 1]++;
5476
5477                 GOTO(out, stripe = true);
5478         }
5479
5480         if (!lustre_handle_is_used(&lh)) {
5481                 dt_read_unlock(env, obj);
5482                 locked = false;
5483                 rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
5484                                       MDS_INODELOCK_LAYOUT |
5485                                       MDS_INODELOCK_XATTR, LCK_EX);
5486                 if (rc != 0)
5487                         GOTO(out, rc);
5488
5489                 handle = dt_trans_create(env, dev);
5490                 if (IS_ERR(handle))
5491                         GOTO(out, rc = PTR_ERR(handle));
5492
5493                 lfsck_buf_init(&ea_buf, lmm, size);
5494                 rc = dt_declare_xattr_set(env, obj, &ea_buf, XATTR_NAME_LOV,
5495                                           LU_XATTR_REPLACE, handle);
5496                 if (rc != 0)
5497                         GOTO(out, rc);
5498
5499                 rc = dt_trans_start_local(env, dev, handle);
5500                 if (rc != 0)
5501                         GOTO(out, rc);
5502
5503                 dt_write_lock(env, obj, 0);
5504                 locked = true;
5505
5506                 goto again;
5507         }
5508
5509         if (magic == LOV_MAGIC_COMP_V1) {
5510                 int i;
5511
5512                 for (i = 0; i < count; i++) {
5513                         lcme = &lcm->lcm_entries[i];
5514                         lmm = buf->lb_buf + le32_to_cpu(lcme->lcme_offset);
5515                         lmm->lmm_oi = *oi;
5516                 }
5517         } else {
5518                 lmm->lmm_oi = *oi;
5519         }
5520
5521         rc = dt_xattr_set(env, obj, &ea_buf, XATTR_NAME_LOV,
5522                           LU_XATTR_REPLACE, handle);
5523         if (rc != 0)
5524                 GOTO(out, rc);
5525
5526         lo->ll_objs_repaired[LLIT_OTHERS - 1]++;
5527
5528         GOTO(out, stripe = true);
5529
5530 out:
5531         if (locked) {
5532                 if (lustre_handle_is_used(&lh))
5533                         dt_write_unlock(env, obj);
5534                 else
5535                         dt_read_unlock(env, obj);
5536         }
5537
5538         if (handle != NULL && !IS_ERR(handle))
5539                 dt_trans_stop(env, dev, handle);
5540
5541         lfsck_ibits_unlock(&lh, LCK_EX);
5542
5543         if (bad_oi)
5544                 CDEBUG(D_LFSCK, "%s: layout LFSCK master %s bad lmm_oi for "
5545                        DFID": rc = %d\n", lfsck_lfsck2name(lfsck),
5546                        bk->lb_param & LPF_DRYRUN ? "found" : "repaired",
5547                        PFID(lfsck_dto2fid(obj)), rc);
5548
5549         if (stripe) {
5550                 if (magic == LOV_MAGIC_COMP_V1) {
5551                         int i;
5552
5553                         for (i = 0; i < count; i++) {
5554                                 lcme = &lcm->lcm_entries[i];
5555                                 if (!(le32_to_cpu(lcme->lcme_flags) &
5556                                       LCME_FL_INIT))
5557                                         continue;
5558
5559                                 rc = lfsck_layout_scan_stripes(env, com, obj,
5560                                         (struct lov_mds_md_v1 *)(buf->lb_buf +
5561                                         le32_to_cpu(lcme->lcme_offset)),
5562                                         le32_to_cpu(lcme->lcme_id));
5563                         }
5564                 } else {
5565                         rc = lfsck_layout_scan_stripes(env, com, obj, lmm, 0);
5566                 }
5567         } else {
5568                 down_write(&com->lc_sem);
5569                 com->lc_new_checked++;
5570                 if (rc < 0)
5571                         lfsck_layout_record_failure(env, lfsck, lo);
5572                 up_write(&com->lc_sem);
5573         }
5574
5575         return rc;
5576 }
5577
5578 static int lfsck_layout_slave_exec_oit(const struct lu_env *env,
5579                                        struct lfsck_component *com,
5580                                        struct dt_object *obj)
5581 {
5582         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5583         struct lfsck_layout             *lo     = com->lc_file_ram;
5584         const struct lu_fid             *fid    = lfsck_dto2fid(obj);
5585         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
5586         struct lfsck_layout_seq         *lls;
5587         __u64                            seq;
5588         __u64                            oid;
5589         int                              rc;
5590         ENTRY;
5591
5592         LASSERT(llsd != NULL);
5593
5594         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DELAY5) &&
5595             cfs_fail_val == lfsck_dev_idx(lfsck)) {
5596                 struct l_wait_info       lwi = LWI_TIMEOUT(cfs_time_seconds(1),
5597                                                            NULL, NULL);
5598                 struct ptlrpc_thread    *thread = &lfsck->li_thread;
5599
5600                 l_wait_event(thread->t_ctl_waitq,
5601                              !thread_is_running(thread),
5602                              &lwi);
5603         }
5604
5605         lfsck_rbtree_update_bitmap(env, com, fid, false);
5606
5607         down_write(&com->lc_sem);
5608         if (fid_is_idif(fid))
5609                 seq = 0;
5610         else if (!fid_is_norm(fid) ||
5611                  !fid_is_for_ostobj(env, lfsck, obj, fid))
5612                 GOTO(unlock, rc = 0);
5613         else
5614                 seq = fid_seq(fid);
5615         com->lc_new_checked++;
5616
5617         lls = lfsck_layout_seq_lookup(llsd, seq);
5618         if (lls == NULL) {
5619                 OBD_ALLOC_PTR(lls);
5620                 if (unlikely(lls == NULL))
5621                         GOTO(unlock, rc = -ENOMEM);
5622
5623                 INIT_LIST_HEAD(&lls->lls_list);
5624                 lls->lls_seq = seq;
5625                 rc = lfsck_layout_lastid_load(env, com, lls);
5626                 if (rc != 0) {
5627                         CDEBUG(D_LFSCK, "%s: layout LFSCK failed to "
5628                               "load LAST_ID for %#llx: rc = %d\n",
5629                               lfsck_lfsck2name(com->lc_lfsck), seq, rc);
5630                         lo->ll_objs_failed_phase1++;
5631                         OBD_FREE_PTR(lls);
5632                         GOTO(unlock, rc);
5633                 }
5634
5635                 lfsck_layout_seq_insert(llsd, lls);
5636         }
5637
5638         if (unlikely(fid_is_last_id(fid)))
5639                 GOTO(unlock, rc = 0);
5640
5641         if (fid_is_idif(fid))
5642                 oid = fid_idif_id(fid_seq(fid), fid_oid(fid), fid_ver(fid));
5643         else
5644                 oid = fid_oid(fid);
5645
5646         if (oid > lls->lls_lastid_known)
5647                 lls->lls_lastid_known = oid;
5648
5649         if (oid > lls->lls_lastid) {
5650                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
5651                         /* OFD may create new objects during LFSCK scanning. */
5652                         rc = lfsck_layout_lastid_reload(env, com, lls);
5653                         if (unlikely(rc != 0)) {
5654                                 CDEBUG(D_LFSCK, "%s: layout LFSCK failed to "
5655                                       "reload LAST_ID for %#llx: rc = %d\n",
5656                                       lfsck_lfsck2name(com->lc_lfsck),
5657                                       lls->lls_seq, rc);
5658
5659                                 GOTO(unlock, rc);
5660                         }
5661
5662                         if (oid <= lls->lls_lastid ||
5663                             lo->ll_flags & LF_CRASHED_LASTID)
5664                                 GOTO(unlock, rc = 0);
5665
5666                         LASSERT(lfsck->li_out_notify != NULL);
5667
5668                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
5669                                              LE_LASTID_REBUILDING);
5670                         lo->ll_flags |= LF_CRASHED_LASTID;
5671
5672                         CDEBUG(D_LFSCK, "%s: layout LFSCK finds crashed "
5673                                "LAST_ID file (2) for the sequence %#llx"
5674                                ", old value %llu, known value %llu\n",
5675                                lfsck_lfsck2name(lfsck), lls->lls_seq,
5676                                lls->lls_lastid, oid);
5677                 }
5678
5679                 lls->lls_lastid = oid;
5680                 lls->lls_dirty = 1;
5681         }
5682
5683         GOTO(unlock, rc = 0);
5684
5685 unlock:
5686         up_write(&com->lc_sem);
5687
5688         return rc;
5689 }
5690
5691 static int lfsck_layout_exec_dir(const struct lu_env *env,
5692                                  struct lfsck_component *com,
5693                                  struct lfsck_assistant_object *lso,
5694                                  struct lu_dirent *ent, __u16 type)
5695 {
5696         return 0;
5697 }
5698
5699 static int lfsck_layout_master_post(const struct lu_env *env,
5700                                     struct lfsck_component *com,
5701                                     int result, bool init)
5702 {
5703         struct lfsck_instance   *lfsck  = com->lc_lfsck;
5704         struct lfsck_layout     *lo     = com->lc_file_ram;
5705         int                      rc;
5706         ENTRY;
5707
5708         lfsck_post_generic(env, com, &result);
5709
5710         down_write(&com->lc_sem);
5711         spin_lock(&lfsck->li_lock);
5712         if (!init)
5713                 lo->ll_pos_last_checkpoint =
5714                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5715
5716         if (result > 0) {
5717                 if (lo->ll_flags & LF_INCOMPLETE)
5718                         lo->ll_status = LS_PARTIAL;
5719                 else
5720                         lo->ll_status = LS_SCANNING_PHASE2;
5721                 lo->ll_flags |= LF_SCANNED_ONCE;
5722                 lo->ll_flags &= ~LF_UPGRADE;
5723                 list_move_tail(&com->lc_link, &lfsck->li_list_double_scan);
5724         } else if (result == 0) {
5725                 if (lfsck->li_status != 0)
5726                         lo->ll_status = lfsck->li_status;
5727                 else
5728                         lo->ll_status = LS_STOPPED;
5729                 if (lo->ll_status != LS_PAUSED)
5730                         list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5731         } else {
5732                 lo->ll_status = LS_FAILED;
5733                 list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5734         }
5735         spin_unlock(&lfsck->li_lock);
5736
5737         if (!init) {
5738                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
5739                                 HALF_SEC - lfsck->li_time_last_checkpoint);
5740                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
5741                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5742                 com->lc_new_checked = 0;
5743         }
5744
5745         rc = lfsck_layout_store(env, com);
5746         up_write(&com->lc_sem);
5747
5748         CDEBUG(D_LFSCK, "%s: layout LFSCK master post done: rc = %d\n",
5749                lfsck_lfsck2name(lfsck), rc);
5750
5751         RETURN(rc);
5752 }
5753
5754 static int lfsck_layout_slave_post(const struct lu_env *env,
5755                                    struct lfsck_component *com,
5756                                    int result, bool init)
5757 {
5758         struct lfsck_instance   *lfsck = com->lc_lfsck;
5759         struct lfsck_layout     *lo    = com->lc_file_ram;
5760         int                      rc;
5761         bool                     done  = false;
5762
5763         down_write(&com->lc_sem);
5764         rc = lfsck_layout_lastid_store(env, com);
5765         if (rc != 0)
5766                 result = rc;
5767
5768         LASSERT(lfsck->li_out_notify != NULL);
5769
5770         spin_lock(&lfsck->li_lock);
5771         if (!init)
5772                 lo->ll_pos_last_checkpoint =
5773                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5774
5775         if (result > 0) {
5776                 lo->ll_status = LS_SCANNING_PHASE2;
5777                 lo->ll_flags |= LF_SCANNED_ONCE;
5778                 if (lo->ll_flags & LF_CRASHED_LASTID) {
5779                         done = true;
5780                         lo->ll_flags &= ~LF_CRASHED_LASTID;
5781
5782                         CDEBUG(D_LFSCK, "%s: layout LFSCK has rebuilt "
5783                                "crashed LAST_ID files successfully\n",
5784                                lfsck_lfsck2name(lfsck));
5785                 }
5786                 lo->ll_flags &= ~LF_UPGRADE;
5787                 list_move_tail(&com->lc_link, &lfsck->li_list_double_scan);
5788         } else if (result == 0) {
5789                 if (lfsck->li_status != 0)
5790                         lo->ll_status = lfsck->li_status;
5791                 else
5792                         lo->ll_status = LS_STOPPED;
5793                 if (lo->ll_status != LS_PAUSED)
5794                         list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5795         } else {
5796                 lo->ll_status = LS_FAILED;
5797                 list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5798         }
5799         spin_unlock(&lfsck->li_lock);
5800
5801         if (done)
5802                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
5803                                      LE_LASTID_REBUILT);
5804
5805         if (!init) {
5806                 lo->ll_run_time_phase1 += cfs_duration_sec(cfs_time_current() +
5807                                 HALF_SEC - lfsck->li_time_last_checkpoint);
5808                 lo->ll_time_last_checkpoint = cfs_time_current_sec();
5809                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5810                 com->lc_new_checked = 0;
5811         }
5812
5813         rc = lfsck_layout_store(env, com);
5814         up_write(&com->lc_sem);
5815
5816         lfsck_layout_slave_notify_master(env, com, LE_PHASE1_DONE, result);
5817
5818         CDEBUG(D_LFSCK, "%s: layout LFSCK slave post done: rc = %d\n",
5819                lfsck_lfsck2name(lfsck), rc);
5820
5821         return rc;
5822 }
5823
5824 static void lfsck_layout_dump(const struct lu_env *env,
5825                               struct lfsck_component *com, struct seq_file *m)
5826 {
5827         struct lfsck_instance   *lfsck = com->lc_lfsck;
5828         struct lfsck_bookmark   *bk    = &lfsck->li_bookmark_ram;
5829         struct lfsck_layout     *lo    = com->lc_file_ram;
5830
5831         down_read(&com->lc_sem);
5832         seq_printf(m, "name: lfsck_layout\n"
5833                    "magic: %#x\n"
5834                    "version: %d\n"
5835                    "status: %s\n",
5836                    lo->ll_magic,
5837                    bk->lb_version,
5838                    lfsck_status2name(lo->ll_status));
5839
5840         lfsck_bits_dump(m, lo->ll_flags, lfsck_flags_names, "flags");
5841
5842         lfsck_bits_dump(m, bk->lb_param, lfsck_param_names, "param");
5843
5844         lfsck_time_dump(m, lo->ll_time_last_complete, "last_completed");
5845
5846         lfsck_time_dump(m, lo->ll_time_latest_start, "latest_start");
5847
5848         lfsck_time_dump(m, lo->ll_time_last_checkpoint, "last_checkpoint");
5849
5850         seq_printf(m, "latest_start_position: %llu\n"
5851                    "last_checkpoint_position: %llu\n"
5852                    "first_failure_position: %llu\n",
5853                    lo->ll_pos_latest_start,
5854                    lo->ll_pos_last_checkpoint,
5855                    lo->ll_pos_first_inconsistent);
5856
5857         seq_printf(m, "success_count: %u\n"
5858                    "repaired_dangling: %llu\n"
5859                    "repaired_unmatched_pair: %llu\n"
5860                    "repaired_multiple_referenced: %llu\n"
5861                    "repaired_orphan: %llu\n"
5862                    "repaired_inconsistent_owner: %llu\n"
5863                    "repaired_others: %llu\n"
5864                    "skipped: %llu\n"
5865                    "failed_phase1: %llu\n"
5866                    "failed_phase2: %llu\n",
5867                    lo->ll_success_count,
5868                    lo->ll_objs_repaired[LLIT_DANGLING - 1],
5869                    lo->ll_objs_repaired[LLIT_UNMATCHED_PAIR - 1],
5870                    lo->ll_objs_repaired[LLIT_MULTIPLE_REFERENCED - 1],
5871                    lo->ll_objs_repaired[LLIT_ORPHAN - 1],
5872                    lo->ll_objs_repaired[LLIT_INCONSISTENT_OWNER - 1],
5873                    lo->ll_objs_repaired[LLIT_OTHERS - 1],
5874                    lo->ll_objs_skipped,
5875                    lo->ll_objs_failed_phase1,
5876                    lo->ll_objs_failed_phase2);
5877
5878         if (lo->ll_status == LS_SCANNING_PHASE1) {
5879                 __u64 pos;
5880                 cfs_duration_t duration = cfs_time_current() -
5881                                           lfsck->li_time_last_checkpoint;
5882                 __u64 checked = lo->ll_objs_checked_phase1 +
5883                                 com->lc_new_checked;
5884                 __u64 speed = checked;
5885                 __u64 new_checked = com->lc_new_checked *
5886                                     msecs_to_jiffies(MSEC_PER_SEC);
5887                 __u32 rtime = lo->ll_run_time_phase1 +
5888                               cfs_duration_sec(duration + HALF_SEC);
5889
5890                 if (duration != 0)
5891                         do_div(new_checked, duration);
5892                 if (rtime != 0)
5893                         do_div(speed, rtime);
5894                 seq_printf(m, "checked_phase1: %llu\n"
5895                            "checked_phase2: %llu\n"
5896                            "run_time_phase1: %u seconds\n"
5897                            "run_time_phase2: %u seconds\n"
5898                            "average_speed_phase1: %llu items/sec\n"
5899                            "average_speed_phase2: N/A\n"
5900                            "real-time_speed_phase1: %llu items/sec\n"
5901                            "real-time_speed_phase2: N/A\n",
5902                            checked,
5903                            lo->ll_objs_checked_phase2,
5904                            rtime,
5905                            lo->ll_run_time_phase2,
5906                            speed,
5907                            new_checked);
5908
5909                 if (likely(lfsck->li_di_oit)) {
5910                         const struct dt_it_ops *iops =
5911                                 &lfsck->li_obj_oit->do_index_ops->dio_it;
5912
5913                         /* The low layer otable-based iteration position may NOT
5914                          * exactly match the layout-based directory traversal
5915                          * cookie. Generally, it is not a serious issue. But the
5916                          * caller should NOT make assumption on that. */
5917                         pos = iops->store(env, lfsck->li_di_oit);
5918                         if (!lfsck->li_current_oit_processed)
5919                                 pos--;
5920                 } else {
5921                         pos = lo->ll_pos_last_checkpoint;
5922                 }
5923
5924                 seq_printf(m, "current_position: %llu\n", pos);
5925         } else if (lo->ll_status == LS_SCANNING_PHASE2) {
5926                 cfs_duration_t duration = cfs_time_current() -
5927                                           com->lc_time_last_checkpoint;
5928                 __u64 checked = lo->ll_objs_checked_phase2 +
5929                                 com->lc_new_checked;
5930                 __u64 speed1 = lo->ll_objs_checked_phase1;
5931                 __u64 speed2 = checked;
5932                 __u64 new_checked = com->lc_new_checked *
5933                                     msecs_to_jiffies(MSEC_PER_SEC);
5934                 __u32 rtime = lo->ll_run_time_phase2 +
5935                               cfs_duration_sec(duration + HALF_SEC);
5936
5937                 if (duration != 0)
5938                         do_div(new_checked, duration);
5939                 if (lo->ll_run_time_phase1 != 0)
5940                         do_div(speed1, lo->ll_run_time_phase1);
5941                 if (rtime != 0)
5942                         do_div(speed2, rtime);
5943                 seq_printf(m, "checked_phase1: %llu\n"
5944                            "checked_phase2: %llu\n"
5945                            "run_time_phase1: %u seconds\n"
5946                            "run_time_phase2: %u seconds\n"
5947                            "average_speed_phase1: %llu items/sec\n"
5948                            "average_speed_phase2: %llu items/sec\n"
5949                            "real-time_speed_phase1: N/A\n"
5950                            "real-time_speed_phase2: %llu items/sec\n"
5951                            "current_position: "DFID"\n",
5952                            lo->ll_objs_checked_phase1,
5953                            checked,
5954                            lo->ll_run_time_phase1,
5955                            rtime,
5956                            speed1,
5957                            speed2,
5958                            new_checked,
5959                            PFID(&com->lc_fid_latest_scanned_phase2));
5960         } else {
5961                 __u64 speed1 = lo->ll_objs_checked_phase1;
5962                 __u64 speed2 = lo->ll_objs_checked_phase2;
5963
5964                 if (lo->ll_run_time_phase1 != 0)
5965                         do_div(speed1, lo->ll_run_time_phase1);
5966                 if (lo->ll_run_time_phase2 != 0)
5967                         do_div(speed2, lo->ll_run_time_phase2);
5968                 seq_printf(m, "checked_phase1: %llu\n"
5969                            "checked_phase2: %llu\n"
5970                            "run_time_phase1: %u seconds\n"
5971                            "run_time_phase2: %u seconds\n"
5972                            "average_speed_phase1: %llu items/sec\n"
5973                            "average_speed_phase2: %llu objs/sec\n"
5974                            "real-time_speed_phase1: N/A\n"
5975                            "real-time_speed_phase2: N/A\n"
5976                            "current_position: N/A\n",
5977                            lo->ll_objs_checked_phase1,
5978                            lo->ll_objs_checked_phase2,
5979                            lo->ll_run_time_phase1,
5980                            lo->ll_run_time_phase2,
5981                            speed1,
5982                            speed2);
5983         }
5984
5985         up_read(&com->lc_sem);
5986 }
5987
5988 static int lfsck_layout_master_double_scan(const struct lu_env *env,
5989                                            struct lfsck_component *com)
5990 {
5991         struct lfsck_layout             *lo     = com->lc_file_ram;
5992         struct lfsck_assistant_data     *lad    = com->lc_data;
5993         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5994         struct lfsck_tgt_descs          *ltds;
5995         struct lfsck_tgt_desc           *ltd;
5996         struct lfsck_tgt_desc           *next;
5997         int                              rc;
5998
5999         rc = lfsck_double_scan_generic(env, com, lo->ll_status);
6000
6001         if (thread_is_stopped(&lad->lad_thread)) {
6002                 LASSERT(list_empty(&lad->lad_req_list));
6003                 LASSERT(list_empty(&lad->lad_ost_phase1_list));
6004                 LASSERT(list_empty(&lad->lad_mdt_phase1_list));
6005
6006                 ltds = &lfsck->li_ost_descs;
6007                 spin_lock(&ltds->ltd_lock);
6008                 list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase2_list,
6009                                          ltd_layout_phase_list) {
6010                         list_del_init(&ltd->ltd_layout_phase_list);
6011                 }
6012                 spin_unlock(&ltds->ltd_lock);
6013
6014                 ltds = &lfsck->li_mdt_descs;
6015                 spin_lock(&ltds->ltd_lock);
6016                 list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase2_list,
6017                                          ltd_layout_phase_list) {
6018                         list_del_init(&ltd->ltd_layout_phase_list);
6019                 }
6020                 spin_unlock(&ltds->ltd_lock);
6021         }
6022
6023         return rc;
6024 }
6025
6026 static int lfsck_layout_slave_double_scan(const struct lu_env *env,
6027                                           struct lfsck_component *com)
6028 {
6029         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6030         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
6031         struct lfsck_layout             *lo     = com->lc_file_ram;
6032         struct ptlrpc_thread            *thread = &lfsck->li_thread;
6033         int                              rc;
6034         ENTRY;
6035
6036         CDEBUG(D_LFSCK, "%s: layout LFSCK slave phase2 scan start\n",
6037                lfsck_lfsck2name(lfsck));
6038
6039         atomic_inc(&lfsck->li_double_scan_count);
6040
6041         if (lo->ll_flags & LF_INCOMPLETE)
6042                 GOTO(done, rc = 1);
6043
6044         com->lc_new_checked = 0;
6045         com->lc_new_scanned = 0;
6046         com->lc_time_last_checkpoint = cfs_time_current();
6047         com->lc_time_next_checkpoint = com->lc_time_last_checkpoint +
6048                                 cfs_time_seconds(LFSCK_CHECKPOINT_INTERVAL);
6049
6050         while (1) {
6051                 struct l_wait_info lwi = LWI_TIMEOUT(cfs_time_seconds(30),
6052                                                      NULL, NULL);
6053
6054                 rc = lfsck_layout_slave_query_master(env, com);
6055                 if (list_empty(&llsd->llsd_master_list)) {
6056                         if (unlikely(!thread_is_running(thread)))
6057                                 rc = 0;
6058                         else
6059                                 rc = 1;
6060
6061                         GOTO(done, rc);
6062                 }
6063
6064                 if (rc < 0)
6065                         GOTO(done, rc);
6066
6067                 rc = l_wait_event(thread->t_ctl_waitq,
6068                                   !thread_is_running(thread) ||
6069                                   lo->ll_flags & LF_INCOMPLETE ||
6070                                   list_empty(&llsd->llsd_master_list),
6071                                   &lwi);
6072                 if (unlikely(!thread_is_running(thread)))
6073                         GOTO(done, rc = 0);
6074
6075                 if (lo->ll_flags & LF_INCOMPLETE)
6076                         GOTO(done, rc = 1);
6077
6078                 if (rc == -ETIMEDOUT)
6079                         continue;
6080
6081                 GOTO(done, rc = (rc < 0 ? rc : 1));
6082         }
6083
6084 done:
6085         rc = lfsck_layout_double_scan_result(env, com, rc);
6086         lfsck_layout_slave_notify_master(env, com, LE_PHASE2_DONE,
6087                         (rc > 0 && lo->ll_flags & LF_INCOMPLETE) ? 0 : rc);
6088         lfsck_layout_slave_quit(env, com);
6089         if (atomic_dec_and_test(&lfsck->li_double_scan_count))
6090                 wake_up_all(&lfsck->li_thread.t_ctl_waitq);
6091
6092         CDEBUG(D_LFSCK, "%s: layout LFSCK slave phase2 scan finished, "
6093                "status %d: rc = %d\n",
6094                lfsck_lfsck2name(lfsck), lo->ll_status, rc);
6095
6096         return rc;
6097 }
6098
6099 static void lfsck_layout_master_data_release(const struct lu_env *env,
6100                                              struct lfsck_component *com)
6101 {
6102         struct lfsck_assistant_data     *lad    = com->lc_data;
6103         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6104         struct lfsck_tgt_descs          *ltds;
6105         struct lfsck_tgt_desc           *ltd;
6106         struct lfsck_tgt_desc           *next;
6107
6108         LASSERT(lad != NULL);
6109         LASSERT(thread_is_init(&lad->lad_thread) ||
6110                 thread_is_stopped(&lad->lad_thread));
6111         LASSERT(list_empty(&lad->lad_req_list));
6112
6113         com->lc_data = NULL;
6114
6115         ltds = &lfsck->li_ost_descs;
6116         spin_lock(&ltds->ltd_lock);
6117         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase1_list,
6118                                  ltd_layout_phase_list) {
6119                 list_del_init(&ltd->ltd_layout_phase_list);
6120         }
6121         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase2_list,
6122                                  ltd_layout_phase_list) {
6123                 list_del_init(&ltd->ltd_layout_phase_list);
6124         }
6125         list_for_each_entry_safe(ltd, next, &lad->lad_ost_list,
6126                                  ltd_layout_list) {
6127                 list_del_init(&ltd->ltd_layout_list);
6128         }
6129         spin_unlock(&ltds->ltd_lock);
6130
6131         ltds = &lfsck->li_mdt_descs;
6132         spin_lock(&ltds->ltd_lock);
6133         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase1_list,
6134                                  ltd_layout_phase_list) {
6135                 list_del_init(&ltd->ltd_layout_phase_list);
6136         }
6137         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase2_list,
6138                                  ltd_layout_phase_list) {
6139                 list_del_init(&ltd->ltd_layout_phase_list);
6140         }
6141         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_list,
6142                                  ltd_layout_list) {
6143                 list_del_init(&ltd->ltd_layout_list);
6144         }
6145         spin_unlock(&ltds->ltd_lock);
6146
6147         if (likely(lad->lad_bitmap != NULL))
6148                 CFS_FREE_BITMAP(lad->lad_bitmap);
6149
6150         OBD_FREE_PTR(lad);
6151 }
6152
6153 static void lfsck_layout_slave_data_release(const struct lu_env *env,
6154                                             struct lfsck_component *com)
6155 {
6156         struct lfsck_layout_slave_data *llsd = com->lc_data;
6157
6158         lfsck_layout_slave_quit(env, com);
6159         com->lc_data = NULL;
6160         OBD_FREE_PTR(llsd);
6161 }
6162
6163 static void lfsck_layout_master_quit(const struct lu_env *env,
6164                                      struct lfsck_component *com)
6165 {
6166         struct lfsck_assistant_data     *lad    = com->lc_data;
6167         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6168         struct lfsck_tgt_descs          *ltds;
6169         struct lfsck_tgt_desc           *ltd;
6170         struct lfsck_tgt_desc           *next;
6171
6172         LASSERT(lad != NULL);
6173
6174         lfsck_quit_generic(env, com);
6175
6176         LASSERT(thread_is_init(&lad->lad_thread) ||
6177                 thread_is_stopped(&lad->lad_thread));
6178         LASSERT(list_empty(&lad->lad_req_list));
6179
6180         ltds = &lfsck->li_ost_descs;
6181         spin_lock(&ltds->ltd_lock);
6182         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase1_list,
6183                                  ltd_layout_phase_list) {
6184                 list_del_init(&ltd->ltd_layout_phase_list);
6185         }
6186         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase2_list,
6187                                  ltd_layout_phase_list) {
6188                 list_del_init(&ltd->ltd_layout_phase_list);
6189         }
6190         spin_unlock(&ltds->ltd_lock);
6191
6192         ltds = &lfsck->li_mdt_descs;
6193         spin_lock(&ltds->ltd_lock);
6194         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase1_list,
6195                                  ltd_layout_phase_list) {
6196                 list_del_init(&ltd->ltd_layout_phase_list);
6197         }
6198         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase2_list,
6199                                  ltd_layout_phase_list) {
6200                 list_del_init(&ltd->ltd_layout_phase_list);
6201         }
6202         spin_unlock(&ltds->ltd_lock);
6203 }
6204
6205 static void lfsck_layout_slave_quit(const struct lu_env *env,
6206                                     struct lfsck_component *com)
6207 {
6208         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
6209         struct lfsck_layout_seq          *lls;
6210         struct lfsck_layout_seq          *next;
6211         struct lfsck_layout_slave_target *llst;
6212
6213         LASSERT(llsd != NULL);
6214
6215         down_write(&com->lc_sem);
6216         list_for_each_entry_safe(lls, next, &llsd->llsd_seq_list,
6217                                  lls_list) {
6218                 list_del_init(&lls->lls_list);
6219                 lfsck_object_put(env, lls->lls_lastid_obj);
6220                 OBD_FREE_PTR(lls);
6221         }
6222         up_write(&com->lc_sem);
6223
6224         spin_lock(&llsd->llsd_lock);
6225         while (!list_empty(&llsd->llsd_master_list)) {
6226                 llst = list_entry(llsd->llsd_master_list.next,
6227                                   struct lfsck_layout_slave_target, llst_list);
6228                 list_del_init(&llst->llst_list);
6229                 spin_unlock(&llsd->llsd_lock);
6230                 lfsck_layout_llst_put(llst);
6231                 spin_lock(&llsd->llsd_lock);
6232         }
6233         spin_unlock(&llsd->llsd_lock);
6234
6235         lfsck_rbtree_cleanup(env, com);
6236 }
6237
6238 static int lfsck_layout_master_in_notify(const struct lu_env *env,
6239                                          struct lfsck_component *com,
6240                                          struct lfsck_request *lr)
6241 {
6242         struct lfsck_instance           *lfsck = com->lc_lfsck;
6243         struct lfsck_layout             *lo    = com->lc_file_ram;
6244         struct lfsck_assistant_data     *lad   = com->lc_data;
6245         struct lfsck_tgt_descs          *ltds;
6246         struct lfsck_tgt_desc           *ltd;
6247         bool                             fail  = false;
6248         ENTRY;
6249
6250         if (lr->lr_event == LE_PAIRS_VERIFY) {
6251                 int rc;
6252
6253                 rc = lfsck_layout_master_check_pairs(env, com, &lr->lr_fid,
6254                                                      &lr->lr_fid2,
6255                                                      lr->lr_comp_id);
6256
6257                 RETURN(rc);
6258         }
6259
6260         CDEBUG(D_LFSCK, "%s: layout LFSCK master handles notify %u "
6261                "from %s %x, status %d, flags %x, flags2 %x\n",
6262                lfsck_lfsck2name(lfsck), lr->lr_event,
6263                (lr->lr_flags & LEF_FROM_OST) ? "OST" : "MDT",
6264                lr->lr_index, lr->lr_status, lr->lr_flags, lr->lr_flags2);
6265
6266         if (lr->lr_event != LE_PHASE1_DONE &&
6267             lr->lr_event != LE_PHASE2_DONE &&
6268             lr->lr_event != LE_PEER_EXIT)
6269                 RETURN(-EINVAL);
6270
6271         if (lr->lr_flags & LEF_FROM_OST)
6272                 ltds = &lfsck->li_ost_descs;
6273         else
6274                 ltds = &lfsck->li_mdt_descs;
6275         spin_lock(&ltds->ltd_lock);
6276         ltd = lfsck_ltd2tgt(ltds, lr->lr_index);
6277         if (ltd == NULL) {
6278                 spin_unlock(&ltds->ltd_lock);
6279
6280                 RETURN(-ENXIO);
6281         }
6282
6283         list_del_init(&ltd->ltd_layout_phase_list);
6284         switch (lr->lr_event) {
6285         case LE_PHASE1_DONE:
6286                 if (lr->lr_status <= 0 || lr->lr_flags2 & LF_INCOMPLETE) {
6287                         if (lr->lr_flags2 & LF_INCOMPLETE) {
6288                                 if (lr->lr_flags & LEF_FROM_OST)
6289                                         lfsck_lad_set_bitmap(env, com,
6290                                                              ltd->ltd_index);
6291                                 else
6292                                         lo->ll_flags |= LF_INCOMPLETE;
6293                         }
6294                         ltd->ltd_layout_done = 1;
6295                         list_del_init(&ltd->ltd_layout_list);
6296                         fail = true;
6297                         break;
6298                 }
6299
6300                 if (lr->lr_flags & LEF_FROM_OST) {
6301                         if (list_empty(&ltd->ltd_layout_list))
6302                                 list_add_tail(&ltd->ltd_layout_list,
6303                                               &lad->lad_ost_list);
6304                         list_add_tail(&ltd->ltd_layout_phase_list,
6305                                       &lad->lad_ost_phase2_list);
6306                 } else {
6307                         if (list_empty(&ltd->ltd_layout_list))
6308                                 list_add_tail(&ltd->ltd_layout_list,
6309                                               &lad->lad_mdt_list);
6310                         list_add_tail(&ltd->ltd_layout_phase_list,
6311                                       &lad->lad_mdt_phase2_list);
6312                 }
6313                 break;
6314         case LE_PHASE2_DONE:
6315                 ltd->ltd_layout_done = 1;
6316                 if (!list_empty(&ltd->ltd_layout_list)) {
6317                         list_del_init(&ltd->ltd_layout_list);
6318                         if (lr->lr_flags2 & LF_INCOMPLETE) {
6319                                 lfsck_lad_set_bitmap(env, com, ltd->ltd_index);
6320                                 fail = true;
6321                         }
6322                 }
6323
6324                 break;
6325         case LE_PEER_EXIT:
6326                 fail = true;
6327                 ltd->ltd_layout_done = 1;
6328                 list_del_init(&ltd->ltd_layout_list);
6329                 if (!(lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT) &&
6330                     !(lr->lr_flags & LEF_FROM_OST))
6331                                 lo->ll_flags |= LF_INCOMPLETE;
6332                 break;
6333         default:
6334                 break;
6335         }
6336         spin_unlock(&ltds->ltd_lock);
6337
6338         if (fail && lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT) {
6339                 struct lfsck_stop *stop = &lfsck_env_info(env)->lti_stop;
6340
6341                 memset(stop, 0, sizeof(*stop));
6342                 stop->ls_status = lr->lr_status;
6343                 stop->ls_flags = lr->lr_param & ~LPF_BROADCAST;
6344                 lfsck_stop(env, lfsck->li_bottom, stop);
6345         } else if (lfsck_phase2_next_ready(lad)) {
6346                 wake_up_all(&lad->lad_thread.t_ctl_waitq);
6347         }
6348
6349         RETURN(0);
6350 }
6351
6352 static int lfsck_layout_slave_in_notify_local(const struct lu_env *env,
6353                                               struct lfsck_component *com,
6354                                               struct lfsck_req_local *lrl,
6355                                               struct thandle *th)
6356 {
6357         ENTRY;
6358
6359         switch (lrl->lrl_event) {
6360         case LEL_FID_ACCESSED:
6361                 lfsck_rbtree_update_bitmap(env, com, &lrl->lrl_fid, true);
6362                 RETURN(0);
6363         case LEL_PAIRS_VERIFY_LOCAL: {
6364                 int rc;
6365
6366                 lrl->lrl_status = LPVS_INIT;
6367                 /* Firstly, if the MDT-object which is claimed via OST-object
6368                  * local stored PFID xattr recognizes the OST-object, then it
6369                  * must be that the client given PFID is wrong. */
6370                 rc = lfsck_layout_slave_check_pairs(env, com, &lrl->lrl_fid,
6371                                 &lrl->lrl_ff_local.ff_parent,
6372                                 lrl->lrl_ff_local.ff_layout.ol_comp_id);
6373                 if (rc <= 0)
6374                         RETURN(0);
6375
6376                 lrl->lrl_status = LPVS_INCONSISTENT;
6377                 /* The OST-object local stored PFID xattr is stale. We need to
6378                  * check whether the MDT-object that is claimed via the client
6379                  * given PFID information recognizes the OST-object or not. If
6380                  * matches, then need to update the OST-object's PFID xattr. */
6381                 rc = lfsck_layout_slave_check_pairs(env, com, &lrl->lrl_fid,
6382                                 &lrl->lrl_ff_client.ff_parent,
6383                                 lrl->lrl_ff_client.ff_layout.ol_comp_id);
6384                 /* For rc < 0 case:
6385                  * We are not sure whether the client given PFID information
6386                  * is correct or not, do nothing to avoid improper fixing.
6387                  *
6388                  * For rc > 0 case:
6389                  * The client given PFID information is also invalid, we can
6390                  * NOT fix the OST-object inconsistency.
6391                  */
6392                 if (!rc) {
6393                         lrl->lrl_status = LPVS_INCONSISTENT_TOFIX;
6394                         rc = lfsck_layout_slave_repair_pfid(env, com, lrl);
6395                 }
6396
6397                 RETURN(rc);
6398         }
6399         default:
6400                 break;
6401         }
6402
6403         RETURN(-EOPNOTSUPP);
6404 }
6405
6406 static int lfsck_layout_slave_in_notify(const struct lu_env *env,
6407                                         struct lfsck_component *com,
6408                                         struct lfsck_request *lr)
6409 {
6410         struct lfsck_instance *lfsck = com->lc_lfsck;
6411         struct lfsck_layout_slave_data *llsd = com->lc_data;
6412         struct lfsck_layout_slave_target *llst;
6413         int rc;
6414         ENTRY;
6415
6416         switch (lr->lr_event) {
6417         case LE_CONDITIONAL_DESTROY:
6418                 rc = lfsck_layout_slave_conditional_destroy(env, com, lr);
6419                 RETURN(rc);
6420         case LE_PHASE1_DONE: {
6421                 if (lr->lr_flags2 & LF_INCOMPLETE) {
6422                         struct lfsck_layout *lo = com->lc_file_ram;
6423
6424                         lo->ll_flags |= LF_INCOMPLETE;
6425                         llst = lfsck_layout_llst_find_and_del(llsd,
6426                                                               lr->lr_index,
6427                                                               true);
6428                         if (llst != NULL) {
6429                                 lfsck_layout_llst_put(llst);
6430                                 wake_up_all(&lfsck->li_thread.t_ctl_waitq);
6431                         }
6432                 }
6433
6434                 RETURN(0);
6435         }
6436         case LE_PHASE2_DONE:
6437         case LE_PEER_EXIT:
6438                 CDEBUG(D_LFSCK, "%s: layout LFSCK slave handle notify %u "
6439                        "from MDT %x, status %d\n", lfsck_lfsck2name(lfsck),
6440                        lr->lr_event, lr->lr_index, lr->lr_status);
6441                 break;
6442         default:
6443                 RETURN(-EINVAL);
6444         }
6445
6446         llst = lfsck_layout_llst_find_and_del(llsd, lr->lr_index, true);
6447         if (llst == NULL)
6448                 RETURN(0);
6449
6450         lfsck_layout_llst_put(llst);
6451         if (list_empty(&llsd->llsd_master_list))
6452                 wake_up_all(&lfsck->li_thread.t_ctl_waitq);
6453
6454         if (lr->lr_event == LE_PEER_EXIT &&
6455             (lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT ||
6456              (list_empty(&llsd->llsd_master_list) &&
6457               (lr->lr_status == LS_STOPPED ||
6458                lr->lr_status == LS_CO_STOPPED)))) {
6459                 struct lfsck_stop *stop = &lfsck_env_info(env)->lti_stop;
6460
6461                 memset(stop, 0, sizeof(*stop));
6462                 stop->ls_status = lr->lr_status;
6463                 stop->ls_flags = lr->lr_param & ~LPF_BROADCAST;
6464                 lfsck_stop(env, lfsck->li_bottom, stop);
6465         }
6466
6467         RETURN(0);
6468 }
6469
6470 static void lfsck_layout_repaired(struct lfsck_layout *lo, __u64 *count)
6471 {
6472         int i;
6473
6474         for (i = 0; i < LLIT_MAX; i++)
6475                 *count += lo->ll_objs_repaired[i];
6476 }
6477
6478 static int lfsck_layout_query_all(const struct lu_env *env,
6479                                   struct lfsck_component *com,
6480                                   __u32 *mdts_count, __u32 *osts_count,
6481                                   __u64 *repaired)
6482 {
6483         struct lfsck_layout *lo = com->lc_file_ram;
6484         struct lfsck_tgt_descs *ltds;
6485         struct lfsck_tgt_desc *ltd;
6486         int idx;
6487         int rc;
6488         ENTRY;
6489
6490         rc = lfsck_query_all(env, com);
6491         if (rc != 0)
6492                 RETURN(rc);
6493
6494         ltds = &com->lc_lfsck->li_mdt_descs;
6495         down_read(&ltds->ltd_rw_sem);
6496         cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
6497                 ltd = lfsck_ltd2tgt(ltds, idx);
6498                 LASSERT(ltd != NULL);
6499
6500                 mdts_count[ltd->ltd_layout_status]++;
6501                 *repaired += ltd->ltd_layout_repaired;
6502         }
6503         up_read(&ltds->ltd_rw_sem);
6504
6505         ltds = &com->lc_lfsck->li_ost_descs;
6506         down_read(&ltds->ltd_rw_sem);
6507         cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
6508                 ltd = lfsck_ltd2tgt(ltds, idx);
6509                 LASSERT(ltd != NULL);
6510
6511                 osts_count[ltd->ltd_layout_status]++;
6512                 *repaired += ltd->ltd_layout_repaired;
6513         }
6514         up_read(&ltds->ltd_rw_sem);
6515
6516         down_read(&com->lc_sem);
6517         mdts_count[lo->ll_status]++;
6518         lfsck_layout_repaired(lo, repaired);
6519         up_read(&com->lc_sem);
6520
6521         RETURN(0);
6522 }
6523
6524 static int lfsck_layout_query(const struct lu_env *env,
6525                               struct lfsck_component *com,
6526                               struct lfsck_request *req,
6527                               struct lfsck_reply *rep,
6528                               struct lfsck_query *que, int idx)
6529 {
6530         struct lfsck_layout *lo = com->lc_file_ram;
6531         int rc = 0;
6532
6533         if (que != NULL) {
6534                 LASSERT(com->lc_lfsck->li_master);
6535
6536                 rc = lfsck_layout_query_all(env, com,
6537                                             que->lu_mdts_count[idx],
6538                                             que->lu_osts_count[idx],
6539                                             &que->lu_repaired[idx]);
6540         } else {
6541                 down_read(&com->lc_sem);
6542                 rep->lr_status = lo->ll_status;
6543                 if (req->lr_flags & LEF_QUERY_ALL)
6544                         lfsck_layout_repaired(lo, &rep->lr_repaired);
6545                 up_read(&com->lc_sem);
6546         }
6547
6548         return rc;
6549 }
6550
6551 /* with lfsck::li_lock held */
6552 static int lfsck_layout_slave_join(const struct lu_env *env,
6553                                    struct lfsck_component *com,
6554                                    struct lfsck_start_param *lsp)
6555 {
6556         struct lfsck_instance            *lfsck = com->lc_lfsck;
6557         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
6558         struct lfsck_layout_slave_target *llst;
6559         struct lfsck_start               *start = lsp->lsp_start;
6560         int                               rc    = 0;
6561         ENTRY;
6562
6563         if (start == NULL || !(start->ls_flags & LPF_OST_ORPHAN))
6564                 RETURN(0);
6565
6566         if (!lsp->lsp_index_valid)
6567                 RETURN(-EINVAL);
6568
6569         /* If someone is running the LFSCK without orphan handling,
6570          * it will not maintain the object accessing rbtree. So we
6571          * cannot join it for orphan handling. */
6572         if (!llsd->llsd_rbtree_valid)
6573                 RETURN(-EBUSY);
6574
6575         spin_unlock(&lfsck->li_lock);
6576         rc = lfsck_layout_llst_add(llsd, lsp->lsp_index);
6577         spin_lock(&lfsck->li_lock);
6578         if (rc == 0 && !thread_is_running(&lfsck->li_thread)) {
6579                 spin_unlock(&lfsck->li_lock);
6580                 llst = lfsck_layout_llst_find_and_del(llsd, lsp->lsp_index,
6581                                                       true);
6582                 if (llst != NULL)
6583                         lfsck_layout_llst_put(llst);
6584                 spin_lock(&lfsck->li_lock);
6585                 rc = -EAGAIN;
6586         }
6587
6588         RETURN(rc);
6589 }
6590
6591 static struct lfsck_operations lfsck_layout_master_ops = {
6592         .lfsck_reset            = lfsck_layout_reset,
6593         .lfsck_fail             = lfsck_layout_fail,
6594         .lfsck_checkpoint       = lfsck_layout_master_checkpoint,
6595         .lfsck_prep             = lfsck_layout_master_prep,
6596         .lfsck_exec_oit         = lfsck_layout_master_exec_oit,
6597         .lfsck_exec_dir         = lfsck_layout_exec_dir,
6598         .lfsck_post             = lfsck_layout_master_post,
6599         .lfsck_dump             = lfsck_layout_dump,
6600         .lfsck_double_scan      = lfsck_layout_master_double_scan,
6601         .lfsck_data_release     = lfsck_layout_master_data_release,
6602         .lfsck_quit             = lfsck_layout_master_quit,
6603         .lfsck_in_notify        = lfsck_layout_master_in_notify,
6604         .lfsck_query            = lfsck_layout_query,
6605 };
6606
6607 static struct lfsck_operations lfsck_layout_slave_ops = {
6608         .lfsck_reset            = lfsck_layout_reset,
6609         .lfsck_fail             = lfsck_layout_fail,
6610         .lfsck_checkpoint       = lfsck_layout_slave_checkpoint,
6611         .lfsck_prep             = lfsck_layout_slave_prep,
6612         .lfsck_exec_oit         = lfsck_layout_slave_exec_oit,
6613         .lfsck_exec_dir         = lfsck_layout_exec_dir,
6614         .lfsck_post             = lfsck_layout_slave_post,
6615         .lfsck_dump             = lfsck_layout_dump,
6616         .lfsck_double_scan      = lfsck_layout_slave_double_scan,
6617         .lfsck_data_release     = lfsck_layout_slave_data_release,
6618         .lfsck_quit             = lfsck_layout_slave_quit,
6619         .lfsck_in_notify_local  = lfsck_layout_slave_in_notify_local,
6620         .lfsck_in_notify        = lfsck_layout_slave_in_notify,
6621         .lfsck_query            = lfsck_layout_query,
6622         .lfsck_join             = lfsck_layout_slave_join,
6623 };
6624
6625 static void lfsck_layout_assistant_fill_pos(const struct lu_env *env,
6626                                             struct lfsck_component *com,
6627                                             struct lfsck_position *pos)
6628 {
6629         struct lfsck_assistant_data     *lad = com->lc_data;
6630         struct lfsck_layout_req         *llr;
6631
6632         if (((struct lfsck_layout *)(com->lc_file_ram))->ll_status !=
6633             LS_SCANNING_PHASE1)
6634                 return;
6635
6636         if (list_empty(&lad->lad_req_list))
6637                 return;
6638
6639         llr = list_entry(lad->lad_req_list.next,
6640                          struct lfsck_layout_req,
6641                          llr_lar.lar_list);
6642         pos->lp_oit_cookie = llr->llr_lar.lar_parent->lso_oit_cookie - 1;
6643 }
6644
6645 struct lfsck_assistant_operations lfsck_layout_assistant_ops = {
6646         .la_handler_p1          = lfsck_layout_assistant_handler_p1,
6647         .la_handler_p2          = lfsck_layout_assistant_handler_p2,
6648         .la_fill_pos            = lfsck_layout_assistant_fill_pos,
6649         .la_double_scan_result  = lfsck_layout_double_scan_result,
6650         .la_req_fini            = lfsck_layout_assistant_req_fini,
6651         .la_sync_failures       = lfsck_layout_assistant_sync_failures,
6652 };
6653
6654 int lfsck_layout_setup(const struct lu_env *env, struct lfsck_instance *lfsck)
6655 {
6656         struct lfsck_component  *com;
6657         struct lfsck_layout     *lo;
6658         struct dt_object        *root = NULL;
6659         struct dt_object        *obj;
6660         int                      i;
6661         int                      rc;
6662         ENTRY;
6663
6664         OBD_ALLOC_PTR(com);
6665         if (com == NULL)
6666                 RETURN(-ENOMEM);
6667
6668         INIT_LIST_HEAD(&com->lc_link);
6669         INIT_LIST_HEAD(&com->lc_link_dir);
6670         init_rwsem(&com->lc_sem);
6671         atomic_set(&com->lc_ref, 1);
6672         com->lc_lfsck = lfsck;
6673         com->lc_type = LFSCK_TYPE_LAYOUT;
6674         if (lfsck->li_master) {
6675                 com->lc_ops = &lfsck_layout_master_ops;
6676                 com->lc_data = lfsck_assistant_data_init(
6677                                 &lfsck_layout_assistant_ops,
6678                                 LFSCK_LAYOUT);
6679                 if (com->lc_data == NULL)
6680                         GOTO(out, rc = -ENOMEM);
6681
6682                 for (i = 0; i < LFSCK_STF_COUNT; i++)
6683                         mutex_init(&com->lc_sub_trace_objs[i].lsto_mutex);
6684         } else {
6685                 struct lfsck_layout_slave_data *llsd;
6686
6687                 com->lc_ops = &lfsck_layout_slave_ops;
6688                 OBD_ALLOC_PTR(llsd);
6689                 if (llsd == NULL)
6690                         GOTO(out, rc = -ENOMEM);
6691
6692                 INIT_LIST_HEAD(&llsd->llsd_seq_list);
6693                 INIT_LIST_HEAD(&llsd->llsd_master_list);
6694                 spin_lock_init(&llsd->llsd_lock);
6695                 llsd->llsd_rb_root = RB_ROOT;
6696                 rwlock_init(&llsd->llsd_rb_lock);
6697                 com->lc_data = llsd;
6698         }
6699         com->lc_file_size = sizeof(*lo);
6700         OBD_ALLOC(com->lc_file_ram, com->lc_file_size);
6701         if (com->lc_file_ram == NULL)
6702                 GOTO(out, rc = -ENOMEM);
6703
6704         OBD_ALLOC(com->lc_file_disk, com->lc_file_size);
6705         if (com->lc_file_disk == NULL)
6706                 GOTO(out, rc = -ENOMEM);
6707
6708         root = dt_locate(env, lfsck->li_bottom, &lfsck->li_local_root_fid);
6709         if (IS_ERR(root))
6710                 GOTO(out, rc = PTR_ERR(root));
6711
6712         if (unlikely(!dt_try_as_dir(env, root)))
6713                 GOTO(out, rc = -ENOTDIR);
6714
6715         obj = local_file_find_or_create(env, lfsck->li_los, root,
6716                                         LFSCK_LAYOUT,
6717                                         S_IFREG | S_IRUGO | S_IWUSR);
6718         if (IS_ERR(obj))
6719                 GOTO(out, rc = PTR_ERR(obj));
6720
6721         com->lc_obj = obj;
6722         rc = lfsck_layout_load(env, com);
6723         if (rc > 0)
6724                 rc = lfsck_layout_reset(env, com, true);
6725         else if (rc == -ENOENT)
6726                 rc = lfsck_layout_init(env, com);
6727         else if (lfsck->li_master)
6728                 rc = lfsck_load_sub_trace_files(env, com,
6729                                 &dt_lfsck_layout_dangling_features,
6730                                 LFSCK_LAYOUT, false);
6731
6732         if (rc != 0)
6733                 GOTO(out, rc);
6734
6735         lo = com->lc_file_ram;
6736         switch (lo->ll_status) {
6737         case LS_INIT:
6738         case LS_COMPLETED:
6739         case LS_FAILED:
6740         case LS_STOPPED:
6741         case LS_PARTIAL:
6742                 spin_lock(&lfsck->li_lock);
6743                 list_add_tail(&com->lc_link, &lfsck->li_list_idle);
6744                 spin_unlock(&lfsck->li_lock);
6745                 break;
6746         default:
6747                 CERROR("%s: unknown lfsck_layout status %d\n",
6748                        lfsck_lfsck2name(lfsck), lo->ll_status);
6749                 /* fall through */
6750         case LS_SCANNING_PHASE1:
6751         case LS_SCANNING_PHASE2:
6752                 /* No need to store the status to disk right now.
6753                  * If the system crashed before the status stored,
6754                  * it will be loaded back when next time. */
6755                 lo->ll_status = LS_CRASHED;
6756                 if (!lfsck->li_master)
6757                         lo->ll_flags |= LF_INCOMPLETE;
6758                 /* fall through */
6759         case LS_PAUSED:
6760         case LS_CRASHED:
6761         case LS_CO_FAILED:
6762         case LS_CO_STOPPED:
6763         case LS_CO_PAUSED:
6764                 spin_lock(&lfsck->li_lock);
6765                 list_add_tail(&com->lc_link, &lfsck->li_list_scan);
6766                 spin_unlock(&lfsck->li_lock);
6767                 break;
6768         }
6769
6770         if (lo->ll_flags & LF_CRASHED_LASTID) {
6771                 LASSERT(lfsck->li_out_notify != NULL);
6772
6773                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
6774                                      LE_LASTID_REBUILDING);
6775         }
6776
6777         GOTO(out, rc = 0);
6778
6779 out:
6780         if (root != NULL && !IS_ERR(root))
6781                 lfsck_object_put(env, root);
6782
6783         if (rc != 0) {
6784                 lfsck_component_cleanup(env, com);
6785                 CERROR("%s: fail to init layout LFSCK component: rc = %d\n",
6786                        lfsck_lfsck2name(lfsck), rc);
6787         }
6788
6789         return rc;
6790 }
6791
6792 struct lfsck_orphan_it {
6793         struct lfsck_component           *loi_com;
6794         struct lfsck_rbtree_node         *loi_lrn;
6795         struct lfsck_layout_slave_target *loi_llst;
6796         struct lu_fid                     loi_key;
6797         struct lu_orphan_rec_v2           loi_rec;
6798         __u64                             loi_hash;
6799         unsigned int                      loi_over:1;
6800 };
6801
6802 static int lfsck_fid_match_idx(const struct lu_env *env,
6803                                struct lfsck_instance *lfsck,
6804                                const struct lu_fid *fid, int idx)
6805 {
6806         struct seq_server_site  *ss;
6807         struct lu_server_fld    *sf;
6808         struct lu_seq_range     *range = &lfsck_env_info(env)->lti_range;
6809         int                      rc;
6810
6811         /* All abnormal cases will be returned to MDT0. */
6812         if (!fid_is_norm(fid)) {
6813                 if (idx == 0)
6814                         return 1;
6815
6816                 return 0;
6817         }
6818
6819         ss = lfsck_dev_site(lfsck);
6820         if (unlikely(ss == NULL))
6821                 return -ENOTCONN;
6822
6823         sf = ss->ss_server_fld;
6824         LASSERT(sf != NULL);
6825
6826         fld_range_set_any(range);
6827         rc = fld_server_lookup(env, sf, fid_seq(fid), range);
6828         if (rc != 0)
6829                 return rc;
6830
6831         if (!fld_range_is_mdt(range))
6832                 return -EINVAL;
6833
6834         if (range->lsr_index == idx)
6835                 return 1;
6836
6837         return 0;
6838 }
6839
6840 static void lfsck_layout_destroy_orphan(const struct lu_env *env,
6841                                         struct dt_object *obj)
6842 {
6843         struct dt_device        *dev    = lfsck_obj2dev(obj);
6844         struct thandle          *handle;
6845         int                      rc;
6846         ENTRY;
6847
6848         handle = dt_trans_create(env, dev);
6849         if (IS_ERR(handle))
6850                 RETURN_EXIT;
6851
6852         rc = dt_declare_ref_del(env, obj, handle);
6853         if (rc != 0)
6854                 GOTO(stop, rc);
6855
6856         rc = dt_declare_destroy(env, obj, handle);
6857         if (rc != 0)
6858                 GOTO(stop, rc);
6859
6860         rc = dt_trans_start_local(env, dev, handle);
6861         if (rc != 0)
6862                 GOTO(stop, rc);
6863
6864         dt_write_lock(env, obj, 0);
6865         rc = dt_ref_del(env, obj, handle);
6866         if (rc == 0)
6867                 rc = dt_destroy(env, obj, handle);
6868         dt_write_unlock(env, obj);
6869
6870         GOTO(stop, rc);
6871
6872 stop:
6873         dt_trans_stop(env, dev, handle);
6874
6875         CDEBUG(D_LFSCK, "destroy orphan OST-object "DFID": rc = %d\n",
6876                PFID(lfsck_dto2fid(obj)), rc);
6877
6878         RETURN_EXIT;
6879 }
6880
6881 static int lfsck_orphan_index_lookup(const struct lu_env *env,
6882                                      struct dt_object *dt,
6883                                      struct dt_rec *rec,
6884                                      const struct dt_key *key)
6885 {
6886         return -EOPNOTSUPP;
6887 }
6888
6889 static int lfsck_orphan_index_declare_insert(const struct lu_env *env,
6890                                              struct dt_object *dt,
6891                                              const struct dt_rec *rec,
6892                                              const struct dt_key *key,
6893                                              struct thandle *handle)
6894 {
6895         return -EOPNOTSUPP;
6896 }
6897
6898 static int lfsck_orphan_index_insert(const struct lu_env *env,
6899                                      struct dt_object *dt,
6900                                      const struct dt_rec *rec,
6901                                      const struct dt_key *key,
6902                                      struct thandle *handle,
6903                                      int ignore_quota)
6904 {
6905         return -EOPNOTSUPP;
6906 }
6907
6908 static int lfsck_orphan_index_declare_delete(const struct lu_env *env,
6909                                              struct dt_object *dt,
6910                                              const struct dt_key *key,
6911                                              struct thandle *handle)
6912 {
6913         return -EOPNOTSUPP;
6914 }
6915
6916 static int lfsck_orphan_index_delete(const struct lu_env *env,
6917                                      struct dt_object *dt,
6918                                      const struct dt_key *key,
6919                                      struct thandle *handle)
6920 {
6921         return -EOPNOTSUPP;
6922 }
6923
6924 static struct dt_it *lfsck_orphan_it_init(const struct lu_env *env,
6925                                           struct dt_object *dt,
6926                                           __u32 attr)
6927 {
6928         struct dt_device                *dev    = lu2dt_dev(dt->do_lu.lo_dev);
6929         struct lfsck_instance           *lfsck;
6930         struct lfsck_component          *com    = NULL;
6931         struct lfsck_layout_slave_data  *llsd;
6932         struct lfsck_orphan_it          *it     = NULL;
6933         struct lfsck_layout             *lo;
6934         int                              rc     = 0;
6935         ENTRY;
6936
6937         lfsck = lfsck_instance_find(dev, true, false);
6938         if (unlikely(lfsck == NULL))
6939                 RETURN(ERR_PTR(-ENXIO));
6940
6941         com = lfsck_component_find(lfsck, LFSCK_TYPE_LAYOUT);
6942         if (unlikely(com == NULL))
6943                 GOTO(out, rc = -ENOENT);
6944
6945         lo = com->lc_file_ram;
6946         if (lo->ll_flags & LF_INCOMPLETE)
6947                 GOTO(out, rc = -ESRCH);
6948
6949         llsd = com->lc_data;
6950         if (!llsd->llsd_rbtree_valid)
6951                 GOTO(out, rc = -ESRCH);
6952
6953         OBD_ALLOC_PTR(it);
6954         if (it == NULL)
6955                 GOTO(out, rc = -ENOMEM);
6956
6957         it->loi_llst = lfsck_layout_llst_find_and_del(llsd, attr, false);
6958         if (it->loi_llst == NULL)
6959                 GOTO(out, rc = -ENXIO);
6960
6961         if (dev->dd_record_fid_accessed) {
6962                 /* The first iteration against the rbtree, scan the whole rbtree
6963                  * to remove the nodes which do NOT need to be handled. */
6964                 write_lock(&llsd->llsd_rb_lock);
6965                 if (dev->dd_record_fid_accessed) {
6966                         struct rb_node                  *node;
6967                         struct rb_node                  *next;
6968                         struct lfsck_rbtree_node        *lrn;
6969
6970                         /* No need to record the fid accessing anymore. */
6971                         dev->dd_record_fid_accessed = 0;
6972
6973                         node = rb_first(&llsd->llsd_rb_root);
6974                         while (node != NULL) {
6975                                 next = rb_next(node);
6976                                 lrn = rb_entry(node, struct lfsck_rbtree_node,
6977                                                lrn_node);
6978                                 if (atomic_read(&lrn->lrn_known_count) <=
6979                                     atomic_read(&lrn->lrn_accessed_count)) {
6980                                         rb_erase(node, &llsd->llsd_rb_root);
6981                                         lfsck_rbtree_free(lrn);
6982                                 }
6983                                 node = next;
6984                         }
6985                 }
6986                 write_unlock(&llsd->llsd_rb_lock);
6987         }
6988
6989         /* read lock the rbtree when init, and unlock when fini */
6990         read_lock(&llsd->llsd_rb_lock);
6991         it->loi_com = com;
6992         com = NULL;
6993
6994         GOTO(out, rc = 0);
6995
6996 out:
6997         if (com != NULL)
6998                 lfsck_component_put(env, com);
6999
7000         CDEBUG(D_LFSCK, "%s: init the orphan iteration: rc = %d\n",
7001                lfsck_lfsck2name(lfsck), rc);
7002
7003         lfsck_instance_put(env, lfsck);
7004         if (rc != 0) {
7005                 if (it != NULL)
7006                         OBD_FREE_PTR(it);
7007
7008                 it = (struct lfsck_orphan_it *)ERR_PTR(rc);
7009         }
7010
7011         return (struct dt_it *)it;
7012 }
7013
7014 static void lfsck_orphan_it_fini(const struct lu_env *env,
7015                                  struct dt_it *di)
7016 {
7017         struct lfsck_orphan_it           *it    = (struct lfsck_orphan_it *)di;
7018         struct lfsck_component           *com   = it->loi_com;
7019         struct lfsck_layout_slave_data   *llsd;
7020         struct lfsck_layout_slave_target *llst;
7021
7022         if (com != NULL) {
7023                 CDEBUG(D_LFSCK, "%s: fini the orphan iteration\n",
7024                        lfsck_lfsck2name(com->lc_lfsck));
7025
7026                 llsd = com->lc_data;
7027                 read_unlock(&llsd->llsd_rb_lock);
7028                 llst = it->loi_llst;
7029                 LASSERT(llst != NULL);
7030
7031                 /* Save the key and hash for iterate next. */
7032                 llst->llst_fid = it->loi_key;
7033                 llst->llst_hash = it->loi_hash;
7034                 lfsck_layout_llst_put(llst);
7035                 lfsck_component_put(env, com);
7036         }
7037         OBD_FREE_PTR(it);
7038 }
7039
7040 /**
7041  * \retval       +1: the iteration finished
7042  * \retval        0: on success, not finished
7043  * \retval      -ve: on error
7044  */
7045 static int lfsck_orphan_it_next(const struct lu_env *env,
7046                                 struct dt_it *di)
7047 {
7048         struct lfsck_thread_info        *info   = lfsck_env_info(env);
7049         struct filter_fid               *ff     = &info->lti_ff;
7050         struct lu_attr                  *la     = &info->lti_la;
7051         struct lfsck_orphan_it          *it     = (struct lfsck_orphan_it *)di;
7052         struct lu_fid                   *key    = &it->loi_key;
7053         struct lu_orphan_rec_v2         *rec    = &it->loi_rec;
7054         struct ost_layout               *ol     = &rec->lor_layout;
7055         struct lfsck_component          *com    = it->loi_com;
7056         struct lfsck_instance           *lfsck  = com->lc_lfsck;
7057         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
7058         struct dt_object                *obj;
7059         struct lfsck_rbtree_node        *lrn;
7060         int                              pos;
7061         int                              rc;
7062         __u32                            save;
7063         __u32                            idx    = it->loi_llst->llst_index;
7064         bool                             exact  = false;
7065         ENTRY;
7066
7067         if (it->loi_over)
7068                 RETURN(1);
7069
7070 again0:
7071         lrn = it->loi_lrn;
7072         if (lrn == NULL) {
7073                 lrn = lfsck_rbtree_search(llsd, key, &exact);
7074                 if (lrn == NULL) {
7075                         it->loi_over = 1;
7076                         RETURN(1);
7077                 }
7078
7079                 it->loi_lrn = lrn;
7080                 if (!exact) {
7081                         key->f_seq = lrn->lrn_seq;
7082                         key->f_oid = lrn->lrn_first_oid;
7083                         key->f_ver = 0;
7084                 }
7085         } else {
7086                 key->f_oid++;
7087                 if (unlikely(key->f_oid == 0)) {
7088                         key->f_seq++;
7089                         it->loi_lrn = NULL;
7090                         goto again0;
7091                 }
7092
7093                 if (key->f_oid >=
7094                     lrn->lrn_first_oid + LFSCK_RBTREE_BITMAP_WIDTH) {
7095                         it->loi_lrn = NULL;
7096                         goto again0;
7097                 }
7098         }
7099
7100         if (unlikely(atomic_read(&lrn->lrn_known_count) <=
7101                      atomic_read(&lrn->lrn_accessed_count))) {
7102                 struct rb_node *next = rb_next(&lrn->lrn_node);
7103
7104                 while (next != NULL) {
7105                         lrn = rb_entry(next, struct lfsck_rbtree_node,
7106                                        lrn_node);
7107                         if (atomic_read(&lrn->lrn_known_count) >
7108                             atomic_read(&lrn->lrn_accessed_count))
7109                                 break;
7110                         next = rb_next(next);
7111                 }
7112
7113                 if (next == NULL) {
7114                         it->loi_over = 1;
7115                         RETURN(1);
7116                 }
7117
7118                 it->loi_lrn = lrn;
7119                 key->f_seq = lrn->lrn_seq;
7120                 key->f_oid = lrn->lrn_first_oid;
7121                 key->f_ver = 0;
7122         }
7123
7124         pos = key->f_oid - lrn->lrn_first_oid;
7125
7126 again1:
7127         pos = find_next_bit(lrn->lrn_known_bitmap,
7128                             LFSCK_RBTREE_BITMAP_WIDTH, pos);
7129         if (pos >= LFSCK_RBTREE_BITMAP_WIDTH) {
7130                 key->f_oid = lrn->lrn_first_oid + pos;
7131                 if (unlikely(key->f_oid < lrn->lrn_first_oid)) {
7132                         key->f_seq++;
7133                         key->f_oid = 0;
7134                 }
7135                 it->loi_lrn = NULL;
7136                 goto again0;
7137         }
7138
7139         if (test_bit(pos, lrn->lrn_accessed_bitmap)) {
7140                 pos++;
7141                 goto again1;
7142         }
7143
7144         key->f_oid = lrn->lrn_first_oid + pos;
7145         obj = lfsck_object_find_bottom(env, lfsck, key);
7146         if (IS_ERR(obj)) {
7147                 rc = PTR_ERR(obj);
7148                 if (rc == -ENOENT) {
7149                         pos++;
7150                         goto again1;
7151                 }
7152                 RETURN(rc);
7153         }
7154
7155         dt_read_lock(env, obj, 0);
7156         if (dt_object_exists(obj) == 0 ||
7157             lfsck_is_dead_obj(obj)) {
7158                 dt_read_unlock(env, obj);
7159                 lfsck_object_put(env, obj);
7160                 pos++;
7161                 goto again1;
7162         }
7163
7164         rc = dt_attr_get(env, obj, la);
7165         if (rc != 0)
7166                 GOTO(out, rc);
7167
7168         rc = dt_xattr_get(env, obj, lfsck_buf_get(env, ff, sizeof(*ff)),
7169                           XATTR_NAME_FID);
7170         if (rc == -ENODATA) {
7171                 /* For the pre-created OST-object, update the bitmap to avoid
7172                  * others LFSCK (second phase) iteration to touch it again. */
7173                 if (la->la_ctime == 0) {
7174                         if (!test_and_set_bit(pos, lrn->lrn_accessed_bitmap))
7175                                 atomic_inc(&lrn->lrn_accessed_count);
7176
7177                         /* For the race between repairing dangling referenced
7178                          * MDT-object and unlink the file, it may left orphan
7179                          * OST-object there. Destroy it now! */
7180                         if (unlikely(!(la->la_mode & S_ISUID))) {
7181                                 dt_read_unlock(env, obj);
7182                                 lfsck_layout_destroy_orphan(env, obj);
7183                                 lfsck_object_put(env, obj);
7184                                 pos++;
7185                                 goto again1;
7186                         }
7187                 } else if (idx == 0) {
7188                         /* If the orphan OST-object has no parent information,
7189                          * regard it as referenced by the MDT-object on MDT0. */
7190                         fid_zero(&rec->lor_rec.lor_fid);
7191                         rec->lor_rec.lor_uid = la->la_uid;
7192                         rec->lor_rec.lor_gid = la->la_gid;
7193                         memset(ol, 0, sizeof(*ol));
7194
7195                         GOTO(out, rc = 0);
7196                 }
7197
7198                 dt_read_unlock(env, obj);
7199                 lfsck_object_put(env, obj);
7200                 pos++;
7201                 goto again1;
7202         }
7203
7204         if (rc < sizeof(struct lu_fid))
7205                 GOTO(out, rc = (rc < 0 ? rc : -EINVAL));
7206
7207         fid_le_to_cpu(&rec->lor_rec.lor_fid, &ff->ff_parent);
7208         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
7209          * MDT-object's FID::f_ver, instead it is the OST-object index in its
7210          * parent MDT-object's layout EA. */
7211         save = rec->lor_rec.lor_fid.f_stripe_idx;
7212         rec->lor_rec.lor_fid.f_ver = 0;
7213         rc = lfsck_fid_match_idx(env, lfsck, &rec->lor_rec.lor_fid, idx);
7214         /* If the orphan OST-object does not claim the MDT, then next.
7215          *
7216          * If we do not know whether it matches or not, then return it
7217          * to the MDT for further check. */
7218         if (rc == 0) {
7219                 dt_read_unlock(env, obj);
7220                 lfsck_object_put(env, obj);
7221                 pos++;
7222                 goto again1;
7223         }
7224
7225         rec->lor_rec.lor_fid.f_stripe_idx = save;
7226         rec->lor_rec.lor_uid = la->la_uid;
7227         rec->lor_rec.lor_gid = la->la_gid;
7228         ost_layout_le_to_cpu(ol, &ff->ff_layout);
7229
7230         CDEBUG(D_LFSCK, "%s: return orphan "DFID", PFID "DFID", owner %u:%u, "
7231                "stripe size %u, stripe count %u, COMP id %u, COMP start %llu, "
7232                "COMP end %llu\n", lfsck_lfsck2name(com->lc_lfsck), PFID(key),
7233                PFID(&rec->lor_rec.lor_fid), rec->lor_rec.lor_uid,
7234                rec->lor_rec.lor_gid, ol->ol_stripe_size, ol->ol_stripe_count,
7235                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end);
7236
7237         GOTO(out, rc = 0);
7238
7239 out:
7240         dt_read_unlock(env, obj);
7241         lfsck_object_put(env, obj);
7242         if (rc == 0)
7243                 it->loi_hash++;
7244
7245         return rc;
7246 }
7247
7248 /**
7249  * \retval       +1: locate to the exactly position
7250  * \retval        0: cannot locate to the exactly position,
7251  *                   call next() to move to a valid position.
7252  * \retval      -ve: on error
7253  */
7254 static int lfsck_orphan_it_get(const struct lu_env *env,
7255                                struct dt_it *di,
7256                                const struct dt_key *key)
7257 {
7258         struct lfsck_orphan_it  *it   = (struct lfsck_orphan_it *)di;
7259         int                      rc;
7260
7261         it->loi_key = *(struct lu_fid *)key;
7262         rc = lfsck_orphan_it_next(env, di);
7263         if (rc == 1)
7264                 return 0;
7265
7266         if (rc == 0)
7267                 return 1;
7268
7269         return rc;
7270 }
7271
7272 static void lfsck_orphan_it_put(const struct lu_env *env,
7273                                 struct dt_it *di)
7274 {
7275 }
7276
7277 static struct dt_key *lfsck_orphan_it_key(const struct lu_env *env,
7278                                           const struct dt_it *di)
7279 {
7280         struct lfsck_orphan_it *it = (struct lfsck_orphan_it *)di;
7281
7282         return (struct dt_key *)&it->loi_key;
7283 }
7284
7285 static int lfsck_orphan_it_key_size(const struct lu_env *env,
7286                                     const struct dt_it *di)
7287 {
7288         return sizeof(struct lu_fid);
7289 }
7290
7291 static int lfsck_orphan_it_rec(const struct lu_env *env,
7292                                const struct dt_it *di,
7293                                struct dt_rec *rec,
7294                                __u32 attr)
7295 {
7296         struct lfsck_orphan_it *it = (struct lfsck_orphan_it *)di;
7297
7298         *(struct lu_orphan_rec_v2 *)rec = it->loi_rec;
7299
7300         return 0;
7301 }
7302
7303 static __u64 lfsck_orphan_it_store(const struct lu_env *env,
7304                                    const struct dt_it *di)
7305 {
7306         struct lfsck_orphan_it  *it   = (struct lfsck_orphan_it *)di;
7307
7308         return it->loi_hash;
7309 }
7310
7311 /**
7312  * \retval       +1: locate to the exactly position
7313  * \retval        0: cannot locate to the exactly position,
7314  *                   call next() to move to a valid position.
7315  * \retval      -ve: on error
7316  */
7317 static int lfsck_orphan_it_load(const struct lu_env *env,
7318                                 const struct dt_it *di,
7319                                 __u64 hash)
7320 {
7321         struct lfsck_orphan_it           *it   = (struct lfsck_orphan_it *)di;
7322         struct lfsck_layout_slave_target *llst = it->loi_llst;
7323         int                               rc;
7324
7325         LASSERT(llst != NULL);
7326
7327         if (hash != llst->llst_hash) {
7328                 CDEBUG(D_LFSCK, "%s: the given hash %llu for orphan "
7329                        "iteration does not match the one when fini "
7330                        "%llu, to be reset.\n",
7331                        lfsck_lfsck2name(it->loi_com->lc_lfsck), hash,
7332                        llst->llst_hash);
7333                 fid_zero(&llst->llst_fid);
7334                 llst->llst_hash = 0;
7335         }
7336
7337         it->loi_key = llst->llst_fid;
7338         it->loi_hash = llst->llst_hash;
7339         rc = lfsck_orphan_it_next(env, (struct dt_it *)di);
7340         if (rc == 1)
7341                 return 0;
7342
7343         if (rc == 0)
7344                 return 1;
7345
7346         return rc;
7347 }
7348
7349 static int lfsck_orphan_it_key_rec(const struct lu_env *env,
7350                                    const struct dt_it *di,
7351                                    void *key_rec)
7352 {
7353         return 0;
7354 }
7355
7356 const struct dt_index_operations lfsck_orphan_index_ops = {
7357         .dio_lookup             = lfsck_orphan_index_lookup,
7358         .dio_declare_insert     = lfsck_orphan_index_declare_insert,
7359         .dio_insert             = lfsck_orphan_index_insert,
7360         .dio_declare_delete     = lfsck_orphan_index_declare_delete,
7361         .dio_delete             = lfsck_orphan_index_delete,
7362         .dio_it = {
7363                 .init           = lfsck_orphan_it_init,
7364                 .fini           = lfsck_orphan_it_fini,
7365                 .get            = lfsck_orphan_it_get,
7366                 .put            = lfsck_orphan_it_put,
7367                 .next           = lfsck_orphan_it_next,
7368                 .key            = lfsck_orphan_it_key,
7369                 .key_size       = lfsck_orphan_it_key_size,
7370                 .rec            = lfsck_orphan_it_rec,
7371                 .store          = lfsck_orphan_it_store,
7372                 .load           = lfsck_orphan_it_load,
7373                 .key_rec        = lfsck_orphan_it_key_rec,
7374         }
7375 };