lustre/lfsck/lfsck_layout.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License version 2 for more details.  A copy is
  14  * included in the COPYING file that accompanied this code.
  15
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2014, 2017, Intel Corporation.
  24  */
  25 /*
  26  * lustre/lfsck/lfsck_layout.c
  27  *
  28  * Author: Fan, Yong <fan.yong@intel.com>
  29  */
  30
  31 #ifndef EXPORT_SYMTAB
  32 # define EXPORT_SYMTAB
  33 #endif
  34 #define DEBUG_SUBSYSTEM S_LFSCK
  35
  36 #include <linux/bitops.h>
  37 #include <linux/rbtree.h>
  38
  39 #include <lu_object.h>
  40 #include <dt_object.h>
  41 #include <lustre_fid.h>
  42 #include <lustre_lib.h>
  43 #include <lustre_net.h>
  44 #include <md_object.h>
  45 #include <obd_class.h>
  46
  47 #include "lfsck_internal.h"
  48
  49 #define LFSCK_LAYOUT_MAGIC_V1           0xB173AE14
  50 #define LFSCK_LAYOUT_MAGIC_V2           0xB1734D76
  51 #define LFSCK_LAYOUT_MAGIC_V3           0xB17371B9
  52 #define LFSCK_LAYOUT_MAGIC_V4           0xB1732FED
  53
  54 #define LFSCK_LAYOUT_MAGIC              LFSCK_LAYOUT_MAGIC_V4
  55
  56 struct lfsck_layout_seq {
  57         struct list_head         lls_list;
  58         __u64                    lls_seq;
  59         __u64                    lls_lastid;
  60         __u64                    lls_lastid_known;
  61         struct dt_object        *lls_lastid_obj;
  62         unsigned int             lls_dirty:1;
  63 };
  64
  65 struct lfsck_layout_slave_target {
  66         /* link into lfsck_layout_slave_data::llsd_master_list. */
  67         struct list_head        llst_list;
  68         /* The position for next record in the rbtree for iteration. */
  69         struct lu_fid           llst_fid;
  70         /* Dummy hash for iteration against the rbtree. */
  71         __u64                   llst_hash;
  72         __u64                   llst_gen;
  73         atomic_t                llst_ref;
  74         __u32                   llst_index;
  75         /* How many times we have failed to get the master status. */
  76         int                     llst_failures;
  77 };
  78
  79 struct lfsck_layout_slave_data {
  80         /* list for lfsck_layout_seq */
  81         struct list_head         llsd_seq_list;
  82
  83         /* list for the masters involve layout verification. */
  84         struct list_head         llsd_master_list;
  85         spinlock_t               llsd_lock;
  86         __u64                    llsd_touch_gen;
  87         struct dt_object        *llsd_rb_obj;
  88         struct rb_root           llsd_rb_root;
  89         struct rw_semaphore      llsd_rb_rwsem;
  90         unsigned int             llsd_rbtree_valid:1;
  91 };
  92
  93 struct lfsck_layout_slave_async_args {
  94         struct obd_export                *llsaa_exp;
  95         struct lfsck_component           *llsaa_com;
  96         struct lfsck_layout_slave_target *llsaa_llst;
  97 };
  98
  99 static inline bool lfsck_comp_extent_aligned(__u64 border, __u32 size)
 100 {
 101         return (border & (size - 1)) == 0;
 102 }
 103
 104 static inline void
 105 lfsck_layout_llst_put(struct lfsck_layout_slave_target *llst)
 106 {
 107         if (atomic_dec_and_test(&llst->llst_ref)) {
 108                 LASSERT(list_empty(&llst->llst_list));
 109
 110                 OBD_FREE_PTR(llst);
 111         }
 112 }
 113
 114 static inline int
 115 lfsck_layout_llst_add(struct lfsck_layout_slave_data *llsd, __u32 index)
 116 {
 117         struct lfsck_layout_slave_target *llst;
 118         struct lfsck_layout_slave_target *tmp;
 119         int                               rc   = 0;
 120
 121         OBD_ALLOC_PTR(llst);
 122         if (llst == NULL)
 123                 return -ENOMEM;
 124
 125         INIT_LIST_HEAD(&llst->llst_list);
 126         llst->llst_gen = 0;
 127         llst->llst_index = index;
 128         atomic_set(&llst->llst_ref, 1);
 129
 130         spin_lock(&llsd->llsd_lock);
 131         list_for_each_entry(tmp, &llsd->llsd_master_list, llst_list) {
 132                 if (tmp->llst_index == index) {
 133                         rc = -EALREADY;
 134                         break;
 135                 }
 136         }
 137         if (rc == 0)
 138                 list_add_tail(&llst->llst_list, &llsd->llsd_master_list);
 139         spin_unlock(&llsd->llsd_lock);
 140
 141         if (rc != 0)
 142                 OBD_FREE_PTR(llst);
 143
 144         return rc;
 145 }
 146
 147 static inline void
 148 lfsck_layout_llst_del(struct lfsck_layout_slave_data *llsd,
 149                       struct lfsck_layout_slave_target *llst)
 150 {
 151         bool del = false;
 152
 153         spin_lock(&llsd->llsd_lock);
 154         if (!list_empty(&llst->llst_list)) {
 155                 list_del_init(&llst->llst_list);
 156                 del = true;
 157         }
 158         spin_unlock(&llsd->llsd_lock);
 159
 160         if (del)
 161                 lfsck_layout_llst_put(llst);
 162 }
 163
 164 static inline struct lfsck_layout_slave_target *
 165 lfsck_layout_llst_find_and_del(struct lfsck_layout_slave_data *llsd,
 166                                __u32 index, bool unlink)
 167 {
 168         struct lfsck_layout_slave_target *llst;
 169
 170         spin_lock(&llsd->llsd_lock);
 171         list_for_each_entry(llst, &llsd->llsd_master_list, llst_list) {
 172                 if (llst->llst_index == index) {
 173                         if (unlink)
 174                                 list_del_init(&llst->llst_list);
 175                         else
 176                                 atomic_inc(&llst->llst_ref);
 177                         spin_unlock(&llsd->llsd_lock);
 178
 179                         return llst;
 180                 }
 181         }
 182         spin_unlock(&llsd->llsd_lock);
 183
 184         return NULL;
 185 }
 186
 187 static struct lfsck_layout_req *
 188 lfsck_layout_assistant_req_init(struct lfsck_assistant_object *lso,
 189                                 struct dt_object *child, __u32 comp_id,
 190                                 __u32 ost_idx, __u32 lov_idx)
 191 {
 192         struct lfsck_layout_req *llr;
 193
 194         OBD_ALLOC_PTR(llr);
 195         if (llr == NULL)
 196                 return ERR_PTR(-ENOMEM);
 197
 198         INIT_LIST_HEAD(&llr->llr_lar.lar_list);
 199         llr->llr_lar.lar_parent = lfsck_assistant_object_get(lso);
 200         llr->llr_child = child;
 201         llr->llr_comp_id = comp_id;
 202         llr->llr_ost_idx = ost_idx;
 203         llr->llr_lov_idx = lov_idx;
 204
 205         return llr;
 206 }
 207
 208 static void lfsck_layout_assistant_req_fini(const struct lu_env *env,
 209                                             struct lfsck_assistant_req *lar)
 210 {
 211         struct lfsck_layout_req *llr =
 212                 container_of(lar, struct lfsck_layout_req, llr_lar);
 213
 214         lfsck_object_put(env, llr->llr_child);
 215         lfsck_assistant_object_put(env, lar->lar_parent);
 216         OBD_FREE_PTR(llr);
 217 }
 218
 219 static int
 220 lfsck_layout_assistant_sync_failures_interpret(const struct lu_env *env,
 221                                                struct ptlrpc_request *req,
 222                                                void *args, int rc)
 223 {
 224         if (rc == 0) {
 225                 struct lfsck_async_interpret_args *laia = args;
 226                 struct lfsck_tgt_desc             *ltd  = laia->laia_ltd;
 227
 228                 ltd->ltd_synced_failures = 1;
 229                 atomic_dec(laia->laia_count);
 230         }
 231
 232         return 0;
 233 }
 234
 235 /**
 236  * Notify remote LFSCK instances about former failures.
 237  *
 238  * The local LFSCK instance has recorded which OSTs have ever failed to respond
 239  * some LFSCK verification requests (maybe because of network issues or the OST
 240  * itself trouble). During the respond gap, the OST may missed some OST-objects
 241  * verification, then the OST cannot know whether related OST-objects have been
 242  * referenced by related MDT-objects or not, then in the second-stage scanning,
 243  * these OST-objects will be regarded as orphan, if the OST-object contains bad
 244  * parent FID for back reference, then it will misguide the LFSCK to make wrong
 245  * fixing for the fake orphan.
 246  *
 247  * To avoid above trouble, when layout LFSCK finishes the first-stage scanning,
 248  * it will scan the bitmap for the ever failed OSTs, and notify them that they
 249  * have ever missed some OST-object verification and should skip the handling
 250  * for orphan OST-objects on all MDTs that are in the layout LFSCK.
 251  *
 252  * \param[in] env       pointer to the thread context
 253  * \param[in] com       pointer to the lfsck component
 254  * \param[in] lr        pointer to the lfsck request
 255  */
 256 static void lfsck_layout_assistant_sync_failures(const struct lu_env *env,
 257                                                  struct lfsck_component *com,
 258                                                  struct lfsck_request *lr)
 259 {
 260         struct lfsck_async_interpret_args *laia  =
 261                                 &lfsck_env_info(env)->lti_laia2;
 262         struct lfsck_assistant_data       *lad   = com->lc_data;
 263         struct lfsck_layout               *lo    = com->lc_file_ram;
 264         struct lfsck_instance             *lfsck = com->lc_lfsck;
 265         struct lfsck_tgt_descs            *ltds  = &lfsck->li_ost_descs;
 266         struct lfsck_tgt_desc             *ltd;
 267         struct ptlrpc_request_set         *set;
 268         atomic_t                           count;
 269         __u32                              idx;
 270         int                                rc    = 0;
 271         ENTRY;
 272
 273         if (!test_bit(LAD_INCOMPLETE, &lad->lad_flags))
 274                 RETURN_EXIT;
 275
 276         /* If the MDT has ever failed to verfiy some OST-objects,
 277          * then sync failures with them firstly. */
 278         lr->lr_flags2 = lo->ll_flags | LF_INCOMPLETE;
 279
 280         atomic_set(&count, 0);
 281         memset(laia, 0, sizeof(*laia));
 282         laia->laia_count = &count;
 283         set = ptlrpc_prep_set();
 284         if (set == NULL)
 285                 GOTO(out, rc = -ENOMEM);
 286
 287         down_read(&ltds->ltd_rw_sem);
 288         cfs_foreach_bit(lad->lad_bitmap, idx) {
 289                 ltd = lfsck_ltd2tgt(ltds, idx);
 290                 if (unlikely(!ltd))
 291                         continue;
 292
 293                 laia->laia_ltd = ltd;
 294                 rc = lfsck_async_request(env, ltd->ltd_exp, lr, set,
 295                                 lfsck_layout_assistant_sync_failures_interpret,
 296                                 laia, LFSCK_NOTIFY);
 297                 if (rc != 0) {
 298                         CDEBUG(D_LFSCK, "%s: LFSCK assistant fail to "
 299                                "notify target %x for %s phase1 done: "
 300                                "rc = %d\n", lfsck_lfsck2name(com->lc_lfsck),
 301                                ltd->ltd_index, lad->lad_name, rc);
 302
 303                         break;
 304                 }
 305
 306                 atomic_inc(&count);
 307         }
 308         up_read(&ltds->ltd_rw_sem);
 309
 310         if (rc == 0 && atomic_read(&count) > 0)
 311                 rc = ptlrpc_set_wait(env, set);
 312
 313         ptlrpc_set_destroy(set);
 314
 315         if (rc == 0 && atomic_read(&count) > 0)
 316                 rc = -EINVAL;
 317
 318         GOTO(out, rc);
 319
 320 out:
 321         if (rc != 0)
 322                 /* If failed to sync failures with the OSTs, then have to
 323                  * mark the whole LFSCK as LF_INCOMPLETE to skip the whole
 324                  * subsequent orphan OST-object handling. */
 325                 lo->ll_flags |= LF_INCOMPLETE;
 326
 327         lr->lr_flags2 = lo->ll_flags;
 328 }
 329
 330 static int lfsck_layout_verify_header_v1v3(struct dt_object *obj,
 331                                            struct lov_mds_md_v1 *lmm,
 332                                            __u64 start, __u64 end,
 333                                            __u32 comp_id,
 334                                            bool ext, bool *dom)
 335 {
 336         __u32 magic;
 337         __u32 pattern;
 338         __u32 size;
 339
 340         magic = le32_to_cpu(lmm->lmm_magic);
 341         /* If magic crashed, keep it there. Sometime later, during OST-object
 342          * orphan handling, if some OST-object(s) back-point to it, it can be
 343          * verified and repaired. */
 344         if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3) {
 345                 int rc;
 346
 347                 if ((magic & LOV_MAGIC_MASK) == LOV_MAGIC_MAGIC)
 348                         rc = -EOPNOTSUPP;
 349                 else
 350                         rc = -EINVAL;
 351
 352                 CDEBUG(D_LFSCK, "%s LOV EA magic 0x%X for the file "DFID"\n",
 353                        rc == -EINVAL ? "Unknown" : "Unsupported",
 354                        magic, PFID(lfsck_dto2fid(obj)));
 355
 356                 return rc;
 357         }
 358
 359         pattern = le32_to_cpu(lmm->lmm_pattern);
 360         *dom = !!(lov_pattern(pattern) == LOV_PATTERN_MDT);
 361
 362         /* XXX: DoM file verification will be supportted via LU-11081. */
 363         if (lov_pattern(pattern) == LOV_PATTERN_MDT) {
 364 #if 0
 365                 if (start != 0) {
 366                         CDEBUG(D_LFSCK, "The DoM entry for "DFID" is not "
 367                                "the first component in the mirror %x/%llu\n",
 368                                PFID(lfsck_dto2fid(obj)), comp_id, start);
 369
 370                         return -EINVAL;
 371                 }
 372 #endif
 373         } else if (!lov_pattern_supported_normal_comp(lov_pattern(pattern))) {
 374                 CDEBUG(D_LFSCK, "Unsupported LOV EA pattern %u for the file "
 375                        DFID" in the component %x\n",
 376                        pattern, PFID(lfsck_dto2fid(obj)), comp_id);
 377
 378                 return -EOPNOTSUPP;
 379         }
 380
 381         size = le32_to_cpu(lmm->lmm_stripe_size);
 382         if (!ext && end != LUSTRE_EOF && start != end &&
 383             !lfsck_comp_extent_aligned(end, size)){
 384                 CDEBUG(D_LFSCK, "not aligned border in PFL extent range "
 385                        "[%llu - %llu) stripesize %u for the file "DFID
 386                        " at idx %d\n", start, end, size,
 387                        PFID(lfsck_dto2fid(obj)), comp_id);
 388
 389                 return -EINVAL;
 390         }
 391
 392         return 0;
 393 }
 394
 395 static int lfsck_layout_verify_header_foreign(struct dt_object *obj,
 396                                               struct lov_foreign_md *lfm,
 397                                               size_t len)
 398 {
 399         /* magic has been verified already */
 400         __u32 value_len = le32_to_cpu(lfm->lfm_length);
 401         /* type and flags are not checked for instance */
 402
 403         CDEBUG(D_INFO, "foreign LOV EA, magic %x, len %u, type %x, flags %x, for file "DFID"\n",
 404                le32_to_cpu(lfm->lfm_magic), value_len,
 405                le32_to_cpu(lfm->lfm_type), le32_to_cpu(lfm->lfm_flags),
 406                PFID(lfsck_dto2fid(obj)));
 407
 408         if (len != value_len + offsetof(typeof(*lfm), lfm_value))
 409                 CDEBUG(D_LFSCK, "foreign LOV EA internal size %u does not match EA full size %zu for file "DFID"\n",
 410                        value_len, len, PFID(lfsck_dto2fid(obj)));
 411
 412         /* nothing to repair */
 413         return -ENODATA;
 414 }
 415
 416 static int lfsck_layout_verify_header(struct dt_object *obj,
 417                                       struct lov_mds_md_v1 *lmm, size_t len)
 418 {
 419         bool p_dom = false;
 420         int rc = 0;
 421
 422         if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_COMP_V1 ||
 423             le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_SEL) {
 424                 struct lov_comp_md_v1 *lcm = (struct lov_comp_md_v1 *)lmm;
 425                 bool p_zero = false;
 426                 int i;
 427                 __u16 count = le16_to_cpu(lcm->lcm_entry_count);
 428
 429                 if (unlikely(count == 0)) {
 430                         CDEBUG(D_LFSCK, "the PFL file "DFID" contains invalid "
 431                                "components count 0\n",
 432                                PFID(lfsck_dto2fid(obj)));
 433
 434                         return -EINVAL;
 435                 }
 436
 437                 for (i = 0; i < count && !rc; i++) {
 438                         struct lov_comp_md_entry_v1 *lcme =
 439                                                 &lcm->lcm_entries[i];
 440                         __u64 start = le64_to_cpu(lcme->lcme_extent.e_start);
 441                         __u64 end = le64_to_cpu(lcme->lcme_extent.e_end);
 442                         __u32 comp_id = le32_to_cpu(lcme->lcme_id);
 443                         bool ext, inited, zero;
 444                         __u32 flags;
 445
 446                         if (unlikely(comp_id == LCME_ID_INVAL ||
 447                                      comp_id > LCME_ID_MAX)) {
 448                                 CDEBUG(D_LFSCK, "found invalid PFL ID %u "
 449                                        "for the file "DFID" at idx %d\n",
 450                                        comp_id, PFID(lfsck_dto2fid(obj)), i);
 451
 452                                 return -EINVAL;
 453                         }
 454
 455                         flags = le32_to_cpu(lcme->lcme_flags);
 456                         ext = flags & LCME_FL_EXTENSION;
 457                         inited = flags & LCME_FL_INIT;
 458                         zero = !!(start == end);
 459
 460                         if ((i == 0) && zero) {
 461                                 CDEBUG(D_LFSCK, "invalid PFL comp %d: [%llu "
 462                                        "- %llu) for "DFID"\n", i, start, end,
 463                                        PFID(lfsck_dto2fid(obj)));
 464                                 return -EINVAL;
 465                         }
 466
 467                         if ((zero && (inited || (i + 1 == count))) ||
 468                             (start > end)) {
 469                                 CDEBUG(D_LFSCK, "invalid PFL comp %d/%d: "
 470                                        "[%llu, %llu) for "DFID", %sinited\n",
 471                                        i, count, start, end,
 472                                        PFID(lfsck_dto2fid(obj)),
 473                                        inited ? "" : "NOT ");
 474                                 return -EINVAL;
 475                         }
 476
 477                         if (!ext && p_zero) {
 478                                 CDEBUG(D_LFSCK, "invalid PFL comp %d: [%llu, "
 479                                        "%llu) for "DFID": NOT extension "
 480                                        "after 0-length component\n", i,
 481                                        start, end, PFID(lfsck_dto2fid(obj)));
 482                                 return -EINVAL;
 483                         }
 484
 485                         if (ext && (inited || p_dom || zero)) {
 486                                 CDEBUG(D_LFSCK, "invalid PFL comp %d: [%llu, "
 487                                        "%llu) for "DFID": %s\n", i,
 488                                        start, end, PFID(lfsck_dto2fid(obj)),
 489                                        inited ? "inited extension" :
 490                                        p_dom ? "extension follows DOM" :
 491                                        zero ? "zero length extension" : "");
 492                                 return -EINVAL;
 493                         }
 494
 495                         rc = lfsck_layout_verify_header_v1v3(obj,
 496                                         (struct lov_mds_md_v1 *)((char *)lmm +
 497                                         le32_to_cpu(lcme->lcme_offset)), start,
 498                                         end, comp_id, ext, &p_dom);
 499
 500                         p_zero = zero;
 501                 }
 502         } else if (le32_to_cpu(lmm->lmm_magic) == LOV_MAGIC_FOREIGN) {
 503                 rc = lfsck_layout_verify_header_foreign(obj,
 504                                                 (struct lov_foreign_md *)lmm,
 505                                                 len);
 506         } else {
 507                 rc = lfsck_layout_verify_header_v1v3(obj, lmm, 0, LUSTRE_EOF,
 508                                                      0, false, &p_dom);
 509         }
 510
 511         return rc;
 512 }
 513
 514 static int lfsck_layout_get_lovea(const struct lu_env *env,
 515                                   struct dt_object *obj, struct lu_buf *buf)
 516 {
 517         int rc;
 518         int rc1;
 519
 520 again:
 521         rc = dt_xattr_get(env, obj, buf, XATTR_NAME_LOV);
 522         if (rc == -ERANGE) {
 523                 rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_LOV);
 524                 if (rc <= 0)
 525                         return !rc ? -ENODATA : rc;
 526
 527                 lu_buf_realloc(buf, rc);
 528                 if (buf->lb_buf == NULL)
 529                         return -ENOMEM;
 530
 531                 goto again;
 532         }
 533
 534         if (rc <= 0)
 535                 return !rc ? -ENODATA : rc;
 536
 537         if (unlikely(buf->lb_buf == NULL)) {
 538                 lu_buf_alloc(buf, rc);
 539                 if (buf->lb_buf == NULL)
 540                         return -ENOMEM;
 541
 542                 goto again;
 543         }
 544
 545         rc1 = lfsck_layout_verify_header(obj, buf->lb_buf, rc);
 546
 547         return rc1 ? rc1 : rc;
 548 }
 549
 550 #define LFSCK_RBTREE_BITMAP_SIZE        PAGE_SIZE
 551 #define LFSCK_RBTREE_BITMAP_WIDTH       (LFSCK_RBTREE_BITMAP_SIZE << 3)
 552 #define LFSCK_RBTREE_BITMAP_MASK        (LFSCK_RBTREE_BITMAP_WIDTH - 1)
 553
 554 struct lfsck_rbtree_node {
 555         struct rb_node   lrn_node;
 556         __u64            lrn_seq;
 557         __u32            lrn_first_oid;
 558         atomic_t         lrn_known_count;
 559         atomic_t         lrn_accessed_count;
 560         void            *lrn_known_bitmap;
 561         void            *lrn_accessed_bitmap;
 562 };
 563
 564 static inline int lfsck_rbtree_cmp(struct lfsck_rbtree_node *lrn,
 565                                    __u64 seq, __u32 oid)
 566 {
 567         if (seq < lrn->lrn_seq)
 568                 return -1;
 569
 570         if (seq > lrn->lrn_seq)
 571                 return 1;
 572
 573         if (oid < lrn->lrn_first_oid)
 574                 return -1;
 575
 576         if (oid - lrn->lrn_first_oid >= LFSCK_RBTREE_BITMAP_WIDTH)
 577                 return 1;
 578
 579         return 0;
 580 }
 581
 582 /* The caller should hold llsd->llsd_rb_lock. */
 583 static struct lfsck_rbtree_node *
 584 lfsck_rbtree_search(struct lfsck_layout_slave_data *llsd,
 585                     const struct lu_fid *fid, bool *exact)
 586 {
 587         struct rb_node           *node  = llsd->llsd_rb_root.rb_node;
 588         struct rb_node           *prev  = NULL;
 589         struct lfsck_rbtree_node *lrn   = NULL;
 590         int                       rc    = 0;
 591
 592         if (exact != NULL)
 593                 *exact = true;
 594
 595         while (node != NULL) {
 596                 prev = node;
 597                 lrn = rb_entry(node, struct lfsck_rbtree_node, lrn_node);
 598                 rc = lfsck_rbtree_cmp(lrn, fid_seq(fid), fid_oid(fid));
 599                 if (rc < 0)
 600                         node = node->rb_left;
 601                 else if (rc > 0)
 602                         node = node->rb_right;
 603                 else
 604                         return lrn;
 605         }
 606
 607         if (exact == NULL)
 608                 return NULL;
 609
 610         /* If there is no exactly matched one, then to the next valid one. */
 611         *exact = false;
 612
 613         /* The rbtree is empty. */
 614         if (rc == 0)
 615                 return NULL;
 616
 617         if (rc < 0)
 618                 return lrn;
 619
 620         node = rb_next(prev);
 621
 622         /* The end of the rbtree. */
 623         if (node == NULL)
 624                 return NULL;
 625
 626         lrn = rb_entry(node, struct lfsck_rbtree_node, lrn_node);
 627
 628         return lrn;
 629 }
 630
 631 static struct lfsck_rbtree_node *lfsck_rbtree_new(const struct lu_env *env,
 632                                                   const struct lu_fid *fid)
 633 {
 634         struct lfsck_rbtree_node *lrn;
 635
 636         OBD_ALLOC_PTR(lrn);
 637         if (lrn == NULL)
 638                 return ERR_PTR(-ENOMEM);
 639
 640         OBD_ALLOC(lrn->lrn_known_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 641         if (lrn->lrn_known_bitmap == NULL) {
 642                 OBD_FREE_PTR(lrn);
 643
 644                 return ERR_PTR(-ENOMEM);
 645         }
 646
 647         OBD_ALLOC(lrn->lrn_accessed_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 648         if (lrn->lrn_accessed_bitmap == NULL) {
 649                 OBD_FREE(lrn->lrn_known_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 650                 OBD_FREE_PTR(lrn);
 651
 652                 return ERR_PTR(-ENOMEM);
 653         }
 654
 655         RB_CLEAR_NODE(&lrn->lrn_node);
 656         lrn->lrn_seq = fid_seq(fid);
 657         lrn->lrn_first_oid = fid_oid(fid) & ~LFSCK_RBTREE_BITMAP_MASK;
 658         atomic_set(&lrn->lrn_known_count, 0);
 659         atomic_set(&lrn->lrn_accessed_count, 0);
 660
 661         return lrn;
 662 }
 663
 664 static void lfsck_rbtree_free(struct lfsck_rbtree_node *lrn)
 665 {
 666         OBD_FREE(lrn->lrn_accessed_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 667         OBD_FREE(lrn->lrn_known_bitmap, LFSCK_RBTREE_BITMAP_SIZE);
 668         OBD_FREE_PTR(lrn);
 669 }
 670
 671 /* The caller should hold lock. */
 672 static struct lfsck_rbtree_node *
 673 lfsck_rbtree_insert(struct lfsck_layout_slave_data *llsd,
 674                     struct lfsck_rbtree_node *lrn)
 675 {
 676         struct rb_node           **pos    = &llsd->llsd_rb_root.rb_node;
 677         struct rb_node            *parent = NULL;
 678         struct lfsck_rbtree_node  *tmp;
 679         int                        rc;
 680
 681         while (*pos != NULL) {
 682                 parent = *pos;
 683                 tmp = rb_entry(parent, struct lfsck_rbtree_node, lrn_node);
 684                 rc = lfsck_rbtree_cmp(tmp, lrn->lrn_seq, lrn->lrn_first_oid);
 685                 if (rc < 0)
 686                         pos = &(*pos)->rb_left;
 687                 else if (rc > 0)
 688                         pos = &(*pos)->rb_right;
 689                 else
 690                         return tmp;
 691         }
 692
 693         rb_link_node(&lrn->lrn_node, parent, pos);
 694         rb_insert_color(&lrn->lrn_node, &llsd->llsd_rb_root);
 695
 696         return lrn;
 697 }
 698
 699 static const struct dt_index_operations lfsck_orphan_index_ops;
 700
 701 static int lfsck_rbtree_setup(const struct lu_env *env,
 702                               struct lfsck_component *com)
 703 {
 704         struct lu_fid                   *fid    = &lfsck_env_info(env)->lti_fid;
 705         struct lfsck_instance           *lfsck  = com->lc_lfsck;
 706         struct dt_device                *dev    = lfsck->li_bottom;
 707         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
 708         struct dt_object                *obj;
 709
 710         fid->f_seq = FID_SEQ_LAYOUT_RBTREE;
 711         fid->f_oid = lfsck_dev_idx(lfsck);
 712         fid->f_ver = 0;
 713         obj = dt_locate(env, dev, fid);
 714         if (IS_ERR(obj))
 715                 RETURN(PTR_ERR(obj));
 716
 717         /* Generate an in-RAM object to stand for the layout rbtree.
 718          * Scanning the layout rbtree will be via the iteration over
 719          * the object. In the future, the rbtree may be written onto
 720          * disk with the object.
 721          *
 722          * Mark the object to be as exist. */
 723         obj->do_lu.lo_header->loh_attr |= LOHA_EXISTS;
 724         obj->do_index_ops = &lfsck_orphan_index_ops;
 725         llsd->llsd_rb_obj = obj;
 726         llsd->llsd_rbtree_valid = 1;
 727         dev->dd_record_fid_accessed = 1;
 728
 729         CDEBUG(D_LFSCK, "%s: layout LFSCK init OST-objects accessing bitmap\n",
 730                lfsck_lfsck2name(lfsck));
 731
 732         return 0;
 733 }
 734
 735 static void lfsck_rbtree_cleanup(const struct lu_env *env,
 736                                  struct lfsck_component *com)
 737 {
 738         struct lfsck_instance           *lfsck = com->lc_lfsck;
 739         struct lfsck_layout_slave_data  *llsd  = com->lc_data;
 740         struct rb_node                  *node  = rb_first(&llsd->llsd_rb_root);
 741         struct rb_node                  *next;
 742         struct lfsck_rbtree_node        *lrn;
 743
 744         lfsck->li_bottom->dd_record_fid_accessed = 0;
 745         /* Invalid the rbtree, then no others will use it. */
 746         down_write(&llsd->llsd_rb_rwsem);
 747         llsd->llsd_rbtree_valid = 0;
 748         up_write(&llsd->llsd_rb_rwsem);
 749
 750         while (node != NULL) {
 751                 next = rb_next(node);
 752                 lrn = rb_entry(node, struct lfsck_rbtree_node, lrn_node);
 753                 rb_erase(node, &llsd->llsd_rb_root);
 754                 lfsck_rbtree_free(lrn);
 755                 node = next;
 756         }
 757
 758         if (llsd->llsd_rb_obj != NULL) {
 759                 lfsck_object_put(env, llsd->llsd_rb_obj);
 760                 llsd->llsd_rb_obj = NULL;
 761         }
 762
 763         CDEBUG(D_LFSCK, "%s: layout LFSCK fini OST-objects accessing bitmap\n",
 764                lfsck_lfsck2name(lfsck));
 765 }
 766
 767 static void lfsck_rbtree_update_bitmap(const struct lu_env *env,
 768                                        struct lfsck_component *com,
 769                                        const struct lu_fid *fid,
 770                                        bool accessed)
 771 {
 772         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
 773         struct lfsck_rbtree_node        *lrn;
 774         bool                             insert = false;
 775         int                              idx;
 776         int                              rc     = 0;
 777         ENTRY;
 778
 779         if (unlikely(!fid_is_sane(fid) || fid_is_last_id(fid)))
 780                 RETURN_EXIT;
 781
 782         if (!fid_is_idif(fid) && !fid_is_norm(fid))
 783                 RETURN_EXIT;
 784
 785         down_read(&llsd->llsd_rb_rwsem);
 786         if (!llsd->llsd_rbtree_valid)
 787                 GOTO(unlock, rc = 0);
 788
 789         lrn = lfsck_rbtree_search(llsd, fid, NULL);
 790         if (lrn == NULL) {
 791                 struct lfsck_rbtree_node *tmp;
 792
 793                 LASSERT(!insert);
 794
 795                 up_read(&llsd->llsd_rb_rwsem);
 796                 tmp = lfsck_rbtree_new(env, fid);
 797                 if (IS_ERR(tmp))
 798                         GOTO(out, rc = PTR_ERR(tmp));
 799
 800                 insert = true;
 801                 down_write(&llsd->llsd_rb_rwsem);
 802                 if (!llsd->llsd_rbtree_valid) {
 803                         lfsck_rbtree_free(tmp);
 804                         GOTO(unlock, rc = 0);
 805                 }
 806
 807                 lrn = lfsck_rbtree_insert(llsd, tmp);
 808                 if (lrn != tmp)
 809                         lfsck_rbtree_free(tmp);
 810         }
 811
 812         idx = fid_oid(fid) & LFSCK_RBTREE_BITMAP_MASK;
 813         /* Any accessed object must be a known object. */
 814         if (!test_and_set_bit(idx, lrn->lrn_known_bitmap))
 815                 atomic_inc(&lrn->lrn_known_count);
 816         if (accessed && !test_and_set_bit(idx, lrn->lrn_accessed_bitmap))
 817                 atomic_inc(&lrn->lrn_accessed_count);
 818
 819         GOTO(unlock, rc = 0);
 820
 821 unlock:
 822         if (insert)
 823                 up_write(&llsd->llsd_rb_rwsem);
 824         else
 825                 up_read(&llsd->llsd_rb_rwsem);
 826 out:
 827         if (rc != 0 && accessed) {
 828                 struct lfsck_layout *lo = com->lc_file_ram;
 829
 830                 CDEBUG(D_LFSCK, "%s: fail to update OST-objects accessing "
 831                        "bitmap, and will cause incorrect LFSCK OST-object "
 832                        "handling, so disable it to cancel orphan handling "
 833                        "for related device. rc = %d\n",
 834                        lfsck_lfsck2name(com->lc_lfsck), rc);
 835
 836                 lo->ll_flags |= LF_INCOMPLETE;
 837                 lfsck_rbtree_cleanup(env, com);
 838         }
 839 }
 840
 841 static inline void lldk_le_to_cpu(struct lfsck_layout_dangling_key *des,
 842                                   const struct lfsck_layout_dangling_key *src)
 843 {
 844         fid_le_to_cpu(&des->lldk_fid, &src->lldk_fid);
 845         des->lldk_comp_id = le32_to_cpu(src->lldk_comp_id);
 846         des->lldk_ea_off = le32_to_cpu(src->lldk_ea_off);
 847 }
 848
 849 static inline void lldk_cpu_to_le(struct lfsck_layout_dangling_key *des,
 850                                   const struct lfsck_layout_dangling_key *src)
 851 {
 852         fid_cpu_to_le(&des->lldk_fid, &src->lldk_fid);
 853         des->lldk_comp_id = cpu_to_le32(src->lldk_comp_id);
 854         des->lldk_ea_off = cpu_to_le32(src->lldk_ea_off);
 855 }
 856
 857 static inline void lldk_be_to_cpu(struct lfsck_layout_dangling_key *des,
 858                                   const struct lfsck_layout_dangling_key *src)
 859 {
 860         fid_be_to_cpu(&des->lldk_fid, &src->lldk_fid);
 861         des->lldk_comp_id = be32_to_cpu(src->lldk_comp_id);
 862         des->lldk_ea_off = be32_to_cpu(src->lldk_ea_off);
 863 }
 864
 865 static inline void lldk_cpu_to_be(struct lfsck_layout_dangling_key *des,
 866                                   const struct lfsck_layout_dangling_key *src)
 867 {
 868         fid_cpu_to_be(&des->lldk_fid, &src->lldk_fid);
 869         des->lldk_comp_id = cpu_to_be32(src->lldk_comp_id);
 870         des->lldk_ea_off = cpu_to_be32(src->lldk_ea_off);
 871 }
 872
 873 static void lfsck_layout_le_to_cpu(struct lfsck_layout *des,
 874                                    const struct lfsck_layout *src)
 875 {
 876         int i;
 877
 878         des->ll_magic = le32_to_cpu(src->ll_magic);
 879         des->ll_status = le32_to_cpu(src->ll_status);
 880         des->ll_flags = le32_to_cpu(src->ll_flags);
 881         des->ll_success_count = le32_to_cpu(src->ll_success_count);
 882         des->ll_run_time_phase1 = le64_to_cpu(src->ll_run_time_phase1);
 883         des->ll_run_time_phase2 = le64_to_cpu(src->ll_run_time_phase2);
 884         des->ll_time_last_complete = le64_to_cpu(src->ll_time_last_complete);
 885         des->ll_time_latest_start = le64_to_cpu(src->ll_time_latest_start);
 886         des->ll_time_last_checkpoint =
 887                                 le64_to_cpu(src->ll_time_last_checkpoint);
 888         des->ll_pos_latest_start = le64_to_cpu(src->ll_pos_latest_start);
 889         des->ll_pos_last_checkpoint = le64_to_cpu(src->ll_pos_last_checkpoint);
 890         des->ll_pos_first_inconsistent =
 891                         le64_to_cpu(src->ll_pos_first_inconsistent);
 892         des->ll_objs_checked_phase1 = le64_to_cpu(src->ll_objs_checked_phase1);
 893         des->ll_objs_failed_phase1 = le64_to_cpu(src->ll_objs_failed_phase1);
 894         des->ll_objs_checked_phase2 = le64_to_cpu(src->ll_objs_checked_phase2);
 895         des->ll_objs_failed_phase2 = le64_to_cpu(src->ll_objs_failed_phase2);
 896         for (i = 0; i < LLIT_MAX; i++)
 897                 des->ll_objs_repaired[i] =
 898                                 le64_to_cpu(src->ll_objs_repaired[i]);
 899         des->ll_objs_skipped = le64_to_cpu(src->ll_objs_skipped);
 900         des->ll_bitmap_size = le32_to_cpu(src->ll_bitmap_size);
 901         lldk_le_to_cpu(&des->ll_lldk_latest_scanned_phase2,
 902                        &src->ll_lldk_latest_scanned_phase2);
 903 }
 904
 905 static void lfsck_layout_cpu_to_le(struct lfsck_layout *des,
 906                                    const struct lfsck_layout *src)
 907 {
 908         int i;
 909
 910         des->ll_magic = cpu_to_le32(src->ll_magic);
 911         des->ll_status = cpu_to_le32(src->ll_status);
 912         des->ll_flags = cpu_to_le32(src->ll_flags);
 913         des->ll_success_count = cpu_to_le32(src->ll_success_count);
 914         des->ll_run_time_phase1 = cpu_to_le64(src->ll_run_time_phase1);
 915         des->ll_run_time_phase2 = cpu_to_le64(src->ll_run_time_phase2);
 916         des->ll_time_last_complete = cpu_to_le64(src->ll_time_last_complete);
 917         des->ll_time_latest_start = cpu_to_le64(src->ll_time_latest_start);
 918         des->ll_time_last_checkpoint =
 919                                 cpu_to_le64(src->ll_time_last_checkpoint);
 920         des->ll_pos_latest_start = cpu_to_le64(src->ll_pos_latest_start);
 921         des->ll_pos_last_checkpoint = cpu_to_le64(src->ll_pos_last_checkpoint);
 922         des->ll_pos_first_inconsistent =
 923                         cpu_to_le64(src->ll_pos_first_inconsistent);
 924         des->ll_objs_checked_phase1 = cpu_to_le64(src->ll_objs_checked_phase1);
 925         des->ll_objs_failed_phase1 = cpu_to_le64(src->ll_objs_failed_phase1);
 926         des->ll_objs_checked_phase2 = cpu_to_le64(src->ll_objs_checked_phase2);
 927         des->ll_objs_failed_phase2 = cpu_to_le64(src->ll_objs_failed_phase2);
 928         for (i = 0; i < LLIT_MAX; i++)
 929                 des->ll_objs_repaired[i] =
 930                                 cpu_to_le64(src->ll_objs_repaired[i]);
 931         des->ll_objs_skipped = cpu_to_le64(src->ll_objs_skipped);
 932         des->ll_bitmap_size = cpu_to_le32(src->ll_bitmap_size);
 933         lldk_cpu_to_le(&des->ll_lldk_latest_scanned_phase2,
 934                        &src->ll_lldk_latest_scanned_phase2);
 935 }
 936
 937 /**
 938  * Load the OST bitmap from the lfsck_layout trace file.
 939  *
 940  * \param[in] env       pointer to the thread context
 941  * \param[in] com       pointer to the lfsck component
 942  *
 943  * \retval              0 for success
 944  * \retval              negative error number on failure or data corruption
 945  */
 946 static int lfsck_layout_load_bitmap(const struct lu_env *env,
 947                                     struct lfsck_component *com)
 948 {
 949         struct dt_object                *obj    = com->lc_obj;
 950         struct lfsck_assistant_data     *lad    = com->lc_data;
 951         struct lfsck_layout             *lo     = com->lc_file_ram;
 952         struct cfs_bitmap                       *bitmap = lad->lad_bitmap;
 953         loff_t                           pos    = com->lc_file_size;
 954         ssize_t                          size;
 955         __u32                            nbits;
 956         int                              rc;
 957         ENTRY;
 958
 959         if (com->lc_lfsck->li_ost_descs.ltd_tgts_bitmap->size >
 960             lo->ll_bitmap_size)
 961                 nbits = com->lc_lfsck->li_ost_descs.ltd_tgts_bitmap->size;
 962         else
 963                 nbits = lo->ll_bitmap_size;
 964
 965         if (unlikely(nbits < BITS_PER_LONG))
 966                 nbits = BITS_PER_LONG;
 967
 968         if (nbits > bitmap->size) {
 969                 __u32 new_bits = bitmap->size;
 970                 struct cfs_bitmap *new_bitmap;
 971
 972                 while (new_bits < nbits)
 973                         new_bits <<= 1;
 974
 975                 new_bitmap = CFS_ALLOCATE_BITMAP(new_bits);
 976                 if (new_bitmap == NULL)
 977                         RETURN(-ENOMEM);
 978
 979                 lad->lad_bitmap = new_bitmap;
 980                 CFS_FREE_BITMAP(bitmap);
 981                 bitmap = new_bitmap;
 982         }
 983
 984         if (lo->ll_bitmap_size == 0) {
 985                 clear_bit(LAD_INCOMPLETE, &lad->lad_flags);
 986                 CFS_RESET_BITMAP(bitmap);
 987
 988                 RETURN(0);
 989         }
 990
 991         size = (lo->ll_bitmap_size + 7) >> 3;
 992         rc = dt_read(env, obj, lfsck_buf_get(env, bitmap->data, size), &pos);
 993         if (rc != size)
 994                 RETURN(rc >= 0 ? -EINVAL : rc);
 995
 996         if (cfs_bitmap_check_empty(bitmap))
 997                 clear_bit(LAD_INCOMPLETE, &lad->lad_flags);
 998         else
 999                 set_bit(LAD_INCOMPLETE, &lad->lad_flags);
1000
1001         RETURN(0);
1002 }
1003
1004 /**
1005  * Load the layout LFSCK trace file from disk.
1006  *
1007  * The layout LFSCK trace file records the layout LFSCK status information
1008  * and other statistics, such as how many objects have been scanned, and how
1009  * many objects have been repaired, and etc. It also contains the bitmap for
1010  * failed OSTs during the layout LFSCK. All these information will be loaded
1011  * from disk to RAM when the layout LFSCK component setup.
1012  *
1013  * \param[in] env       pointer to the thread context
1014  * \param[in] com       pointer to the lfsck component
1015  *
1016  * \retval              positive number for file data corruption, the caller
1017  *                      should reset the layout LFSCK trace file
1018  * \retval              0 for success
1019  * \retval              negative error number on failure
1020  */
1021 static int lfsck_layout_load(const struct lu_env *env,
1022                              struct lfsck_component *com)
1023 {
1024         struct lfsck_layout             *lo     = com->lc_file_ram;
1025         ssize_t                          size   = com->lc_file_size;
1026         loff_t                           pos    = 0;
1027         int                              rc;
1028
1029         rc = dt_read(env, com->lc_obj,
1030                      lfsck_buf_get(env, com->lc_file_disk, size), &pos);
1031         if (rc == 0) {
1032                 return -ENOENT;
1033         } else if (rc < 0) {
1034                 CDEBUG(D_LFSCK, "%s: failed to load lfsck_layout: rc = %d\n",
1035                        lfsck_lfsck2name(com->lc_lfsck), rc);
1036                 return rc;
1037         } else if (rc != size) {
1038                 CDEBUG(D_LFSCK, "%s: lfsck_layout size %u != %u; reset it\n",
1039                        lfsck_lfsck2name(com->lc_lfsck), rc, (unsigned int)size);
1040                 return 1;
1041         }
1042
1043         lfsck_layout_le_to_cpu(lo, com->lc_file_disk);
1044         if (lo->ll_magic != LFSCK_LAYOUT_MAGIC) {
1045                 CDEBUG(D_LFSCK, "%s: invalid lfsck_layout magic %#x != %#x, "
1046                        "to be reset\n", lfsck_lfsck2name(com->lc_lfsck),
1047                        lo->ll_magic, LFSCK_LAYOUT_MAGIC);
1048                 return 1;
1049         }
1050
1051         return 0;
1052 }
1053
1054 /**
1055  * Store the layout LFSCK trace file on disk.
1056  *
1057  * The layout LFSCK trace file records the layout LFSCK status information
1058  * and other statistics, such as how many objects have been scanned, and how
1059  * many objects have been repaired, and etc. It also contains the bitmap for
1060  * failed OSTs during the layout LFSCK. All these information will be synced
1061  * from RAM to disk periodically.
1062  *
1063  * \param[in] env       pointer to the thread context
1064  * \param[in] com       pointer to the lfsck component
1065  *
1066  * \retval              0 for success
1067  * \retval              negative error number on failure
1068  */
1069 static int lfsck_layout_store(const struct lu_env *env,
1070                               struct lfsck_component *com)
1071 {
1072         struct dt_object        *obj    = com->lc_obj;
1073         struct lfsck_instance   *lfsck  = com->lc_lfsck;
1074         struct lfsck_layout     *lo_ram = com->lc_file_ram;
1075         struct lfsck_layout     *lo     = com->lc_file_disk;
1076         struct thandle          *th;
1077         struct dt_device        *dev    = lfsck_obj2dev(obj);
1078         struct cfs_bitmap       *bitmap = NULL;
1079         loff_t                   pos;
1080         ssize_t                  size   = com->lc_file_size;
1081         __u32                    nbits  = 0;
1082         int                      rc;
1083         ENTRY;
1084
1085         if (lfsck->li_master) {
1086                 struct lfsck_assistant_data *lad = com->lc_data;
1087
1088                 bitmap = lad->lad_bitmap;
1089                 nbits = bitmap->size;
1090
1091                 LASSERT(nbits > 0);
1092                 LASSERTF((nbits & 7) == 0, "Invalid nbits %u\n", nbits);
1093         }
1094
1095         lo_ram->ll_bitmap_size = nbits;
1096         lfsck_layout_cpu_to_le(lo, lo_ram);
1097         th = dt_trans_create(env, dev);
1098         if (IS_ERR(th))
1099                 GOTO(log, rc = PTR_ERR(th));
1100
1101         rc = dt_declare_record_write(env, obj, lfsck_buf_get(env, lo, size),
1102                                      (loff_t)0, th);
1103         if (rc != 0)
1104                 GOTO(out, rc);
1105
1106         if (bitmap != NULL) {
1107                 rc = dt_declare_record_write(env, obj,
1108                                 lfsck_buf_get(env, bitmap->data, nbits >> 3),
1109                                 (loff_t)size, th);
1110                 if (rc != 0)
1111                         GOTO(out, rc);
1112         }
1113
1114         rc = dt_trans_start_local(env, dev, th);
1115         if (rc != 0)
1116                 GOTO(out, rc);
1117
1118         pos = 0;
1119         rc = dt_record_write(env, obj, lfsck_buf_get(env, lo, size), &pos, th);
1120         if (rc != 0)
1121                 GOTO(out, rc);
1122
1123         if (bitmap != NULL) {
1124                 pos = size;
1125                 rc = dt_record_write(env, obj,
1126                                 lfsck_buf_get(env, bitmap->data, nbits >> 3),
1127                                 &pos, th);
1128         }
1129
1130         GOTO(out, rc);
1131
1132 out:
1133         dt_trans_stop(env, dev, th);
1134
1135 log:
1136         if (rc != 0)
1137                 CDEBUG(D_LFSCK, "%s: fail to store lfsck_layout: rc = %d\n",
1138                        lfsck_lfsck2name(lfsck), rc);
1139
1140         return rc;
1141 }
1142
1143 static int lfsck_layout_init(const struct lu_env *env,
1144                              struct lfsck_component *com)
1145 {
1146         struct lfsck_layout *lo = com->lc_file_ram;
1147         int rc;
1148
1149         memset(lo, 0, com->lc_file_size);
1150         lo->ll_magic = LFSCK_LAYOUT_MAGIC;
1151         lo->ll_status = LS_INIT;
1152         down_write(&com->lc_sem);
1153         rc = lfsck_layout_store(env, com);
1154         if (rc == 0 && com->lc_lfsck->li_master)
1155                 rc = lfsck_load_sub_trace_files(env, com,
1156                         &dt_lfsck_layout_dangling_features, LFSCK_LAYOUT, true);
1157         up_write(&com->lc_sem);
1158
1159         return rc;
1160 }
1161
1162 static int fid_is_for_ostobj(const struct lu_env *env,
1163                              struct lfsck_instance *lfsck,
1164                              struct dt_object *obj, const struct lu_fid *fid)
1165 {
1166         struct seq_server_site  *ss     = lfsck_dev_site(lfsck);
1167         struct lu_seq_range     *range  = &lfsck_env_info(env)->lti_range;
1168         struct lustre_ost_attrs *loa;
1169         int                      rc;
1170
1171         fld_range_set_any(range);
1172         rc = fld_server_lookup(env, ss->ss_server_fld, fid_seq(fid), range);
1173         if (rc == 0) {
1174                 if (fld_range_is_ost(range))
1175                         return 1;
1176
1177                 return 0;
1178         }
1179
1180         loa = &lfsck_env_info(env)->lti_loa;
1181         rc = dt_xattr_get(env, obj, lfsck_buf_get(env, loa, sizeof(*loa)),
1182                           XATTR_NAME_LMA);
1183         if (rc >= (int)sizeof(struct lustre_mdt_attrs)) {
1184                 lustre_lma_swab(&loa->loa_lma);
1185
1186                 return loa->loa_lma.lma_compat & LMAC_FID_ON_OST ? 1 : 0;
1187         }
1188
1189         rc = dt_xattr_get(env, obj, &LU_BUF_NULL, XATTR_NAME_FID);
1190
1191         return rc > 0;
1192 }
1193
1194 static struct lfsck_layout_seq *
1195 lfsck_layout_seq_lookup(struct lfsck_layout_slave_data *llsd, __u64 seq)
1196 {
1197         struct lfsck_layout_seq *lls;
1198
1199         list_for_each_entry(lls, &llsd->llsd_seq_list, lls_list) {
1200                 if (lls->lls_seq == seq)
1201                         return lls;
1202
1203                 if (lls->lls_seq > seq)
1204                         return NULL;
1205         }
1206
1207         return NULL;
1208 }
1209
1210 static void
1211 lfsck_layout_seq_insert(struct lfsck_layout_slave_data *llsd,
1212                         struct lfsck_layout_seq *lls)
1213 {
1214         struct lfsck_layout_seq *tmp;
1215         struct list_head        *pos = &llsd->llsd_seq_list;
1216
1217         list_for_each_entry(tmp, &llsd->llsd_seq_list, lls_list) {
1218                 if (lls->lls_seq < tmp->lls_seq) {
1219                         pos = &tmp->lls_list;
1220                         break;
1221                 }
1222         }
1223         list_add_tail(&lls->lls_list, pos);
1224 }
1225
1226 static int
1227 lfsck_layout_lastid_create(const struct lu_env *env,
1228                            struct lfsck_instance *lfsck,
1229                            struct dt_object *obj)
1230 {
1231         struct lfsck_thread_info *info   = lfsck_env_info(env);
1232         struct lu_attr           *la     = &info->lti_la;
1233         struct dt_object_format  *dof    = &info->lti_dof;
1234         struct lfsck_bookmark    *bk     = &lfsck->li_bookmark_ram;
1235         struct dt_device         *dt     = lfsck_obj2dev(obj);
1236         struct thandle           *th;
1237         __u64                     lastid = 0;
1238         loff_t                    pos    = 0;
1239         int                       rc;
1240         ENTRY;
1241
1242         if (bk->lb_param & LPF_DRYRUN)
1243                 return 0;
1244
1245         memset(la, 0, sizeof(*la));
1246         la->la_mode = S_IFREG |  S_IRUGO | S_IWUSR;
1247         la->la_valid = LA_MODE | LA_UID | LA_GID;
1248         memset(dof, 0, sizeof(*dof));
1249         dof->dof_type = dt_mode_to_dft(S_IFREG);
1250
1251         th = dt_trans_create(env, dt);
1252         if (IS_ERR(th))
1253                 GOTO(log, rc = PTR_ERR(th));
1254
1255         rc = dt_declare_create(env, obj, la, NULL, dof, th);
1256         if (rc != 0)
1257                 GOTO(stop, rc);
1258
1259         rc = dt_declare_record_write(env, obj,
1260                                      lfsck_buf_get(env, &lastid,
1261                                                    sizeof(lastid)),
1262                                      pos, th);
1263         if (rc != 0)
1264                 GOTO(stop, rc);
1265
1266         rc = dt_trans_start_local(env, dt, th);
1267         if (rc != 0)
1268                 GOTO(stop, rc);
1269
1270         dt_write_lock(env, obj, 0);
1271         if (likely(dt_object_exists(obj) == 0)) {
1272                 rc = dt_create(env, obj, la, NULL, dof, th);
1273                 if (rc == 0)
1274                         rc = dt_record_write(env, obj,
1275                                 lfsck_buf_get(env, &lastid, sizeof(lastid)),
1276                                 &pos, th);
1277         }
1278         dt_write_unlock(env, obj);
1279
1280         GOTO(stop, rc);
1281
1282 stop:
1283         dt_trans_stop(env, dt, th);
1284
1285 log:
1286         CDEBUG(D_LFSCK, "%s: layout LFSCK will create LAST_ID for <seq> "
1287                "%#llx: rc = %d\n",
1288                lfsck_lfsck2name(lfsck), fid_seq(lfsck_dto2fid(obj)), rc);
1289
1290         return rc;
1291 }
1292
1293 static int
1294 lfsck_layout_lastid_reload(const struct lu_env *env,
1295                            struct lfsck_component *com,
1296                            struct lfsck_layout_seq *lls)
1297 {
1298         __u64   lastid;
1299         loff_t  pos     = 0;
1300         int     rc;
1301
1302         dt_read_lock(env, lls->lls_lastid_obj, 0);
1303         rc = dt_record_read(env, lls->lls_lastid_obj,
1304                             lfsck_buf_get(env, &lastid, sizeof(lastid)), &pos);
1305         dt_read_unlock(env, lls->lls_lastid_obj);
1306         if (unlikely(rc != 0))
1307                 return rc;
1308
1309         lastid = le64_to_cpu(lastid);
1310         if (lastid < lls->lls_lastid_known) {
1311                 struct lfsck_instance   *lfsck  = com->lc_lfsck;
1312                 struct lfsck_layout     *lo     = com->lc_file_ram;
1313
1314                 lls->lls_lastid = lls->lls_lastid_known;
1315                 lls->lls_dirty = 1;
1316                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
1317                         LASSERT(lfsck->li_out_notify != NULL);
1318
1319                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1320                                              LE_LASTID_REBUILDING);
1321                         lo->ll_flags |= LF_CRASHED_LASTID;
1322
1323                         CDEBUG(D_LFSCK, "%s: layout LFSCK finds crashed "
1324                                "LAST_ID file (1) for the sequence %#llx"
1325                                ", old value %llu, known value %llu\n",
1326                                lfsck_lfsck2name(lfsck), lls->lls_seq,
1327                                lastid, lls->lls_lastid);
1328                 }
1329         } else if (lastid >= lls->lls_lastid) {
1330                 lls->lls_lastid = lastid;
1331                 lls->lls_dirty = 0;
1332         }
1333
1334         return 0;
1335 }
1336
1337 static int
1338 lfsck_layout_lastid_store(const struct lu_env *env,
1339                           struct lfsck_component *com)
1340 {
1341         struct lfsck_instance           *lfsck  = com->lc_lfsck;
1342         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
1343         struct dt_device                *dt     = lfsck->li_bottom;
1344         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
1345         struct lfsck_layout_seq         *lls;
1346         struct thandle                  *th;
1347         __u64                            lastid;
1348         int                              rc     = 0;
1349         int                              rc1    = 0;
1350
1351         list_for_each_entry(lls, &llsd->llsd_seq_list, lls_list) {
1352                 loff_t pos = 0;
1353
1354                 if (!lls->lls_dirty)
1355                         continue;
1356
1357                 CDEBUG(D_LFSCK, "%s: layout LFSCK will sync the LAST_ID for "
1358                        "<seq> %#llx as <oid> %llu\n",
1359                        lfsck_lfsck2name(lfsck), lls->lls_seq, lls->lls_lastid);
1360
1361                 if (bk->lb_param & LPF_DRYRUN) {
1362                         lls->lls_dirty = 0;
1363                         continue;
1364                 }
1365
1366                 th = dt_trans_create(env, dt);
1367                 if (IS_ERR(th)) {
1368                         rc1 = PTR_ERR(th);
1369                         CDEBUG(D_LFSCK, "%s: layout LFSCK failed to store "
1370                                "the LAST_ID for <seq> %#llx(1): rc = %d\n",
1371                                lfsck_lfsck2name(com->lc_lfsck),
1372                                lls->lls_seq, rc1);
1373                         continue;
1374                 }
1375
1376                 lastid = cpu_to_le64(lls->lls_lastid);
1377                 rc = dt_declare_record_write(env, lls->lls_lastid_obj,
1378                                              lfsck_buf_get(env, &lastid,
1379                                                            sizeof(lastid)),
1380                                              pos, th);
1381                 if (rc != 0)
1382                         goto stop;
1383
1384                 rc = dt_trans_start_local(env, dt, th);
1385                 if (rc != 0)
1386                         goto stop;
1387
1388                 dt_write_lock(env, lls->lls_lastid_obj, 0);
1389                 rc = dt_record_write(env, lls->lls_lastid_obj,
1390                                      lfsck_buf_get(env, &lastid,
1391                                      sizeof(lastid)), &pos, th);
1392                 dt_write_unlock(env, lls->lls_lastid_obj);
1393                 if (rc == 0)
1394                         lls->lls_dirty = 0;
1395
1396 stop:
1397                 dt_trans_stop(env, dt, th);
1398                 if (rc != 0) {
1399                         rc1 = rc;
1400                         CDEBUG(D_LFSCK, "%s: layout LFSCK failed to store "
1401                                "the LAST_ID for <seq> %#llx(2): rc = %d\n",
1402                                lfsck_lfsck2name(com->lc_lfsck),
1403                                lls->lls_seq, rc1);
1404                 }
1405         }
1406
1407         return rc1;
1408 }
1409
1410 static int
1411 lfsck_layout_lastid_load(const struct lu_env *env,
1412                          struct lfsck_component *com,
1413                          struct lfsck_layout_seq *lls)
1414 {
1415         struct lfsck_instance   *lfsck  = com->lc_lfsck;
1416         struct lfsck_layout     *lo     = com->lc_file_ram;
1417         struct lu_fid           *fid    = &lfsck_env_info(env)->lti_fid;
1418         struct dt_object        *obj;
1419         loff_t                   pos    = 0;
1420         int                      rc;
1421         ENTRY;
1422
1423         lu_last_id_fid(fid, lls->lls_seq, lfsck_dev_idx(lfsck));
1424         obj = dt_locate(env, lfsck->li_bottom, fid);
1425         if (IS_ERR(obj))
1426                 RETURN(PTR_ERR(obj));
1427
1428         /* LAST_ID crashed, to be rebuilt */
1429         if (dt_object_exists(obj) == 0) {
1430                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
1431                         LASSERT(lfsck->li_out_notify != NULL);
1432
1433                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1434                                              LE_LASTID_REBUILDING);
1435                         lo->ll_flags |= LF_CRASHED_LASTID;
1436
1437                         CDEBUG(D_LFSCK, "%s: layout LFSCK cannot find the "
1438                                "LAST_ID file for sequence %#llx\n",
1439                                lfsck_lfsck2name(lfsck), lls->lls_seq);
1440
1441                         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DELAY4) &&
1442                             cfs_fail_val > 0) {
1443                                 struct ptlrpc_thread *thread =
1444                                         &lfsck->li_thread;
1445
1446                                 up_write(&com->lc_sem);
1447                                 wait_event_idle_timeout(
1448                                         thread->t_ctl_waitq,
1449                                         !thread_is_running(thread),
1450                                         cfs_time_seconds(cfs_fail_val));
1451                                 down_write(&com->lc_sem);
1452                         }
1453                 }
1454
1455                 rc = lfsck_layout_lastid_create(env, lfsck, obj);
1456         } else {
1457                 dt_read_lock(env, obj, 0);
1458                 rc = dt_read(env, obj,
1459                         lfsck_buf_get(env, &lls->lls_lastid, sizeof(__u64)),
1460                         &pos);
1461                 dt_read_unlock(env, obj);
1462                 if (rc != 0 && rc != sizeof(__u64))
1463                         GOTO(out, rc = (rc > 0 ? -EFAULT : rc));
1464
1465                 if (rc == 0 && !(lo->ll_flags & LF_CRASHED_LASTID)) {
1466                         LASSERT(lfsck->li_out_notify != NULL);
1467
1468                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
1469                                              LE_LASTID_REBUILDING);
1470                         lo->ll_flags |= LF_CRASHED_LASTID;
1471
1472                         CDEBUG(D_LFSCK, "%s: layout LFSCK finds invalid "
1473                                "LAST_ID file for the sequence %#llx"
1474                                ": rc = %d\n",
1475                                lfsck_lfsck2name(lfsck), lls->lls_seq, rc);
1476                 }
1477
1478                 lls->lls_lastid = le64_to_cpu(lls->lls_lastid);
1479                 rc = 0;
1480         }
1481
1482         GOTO(out, rc);
1483
1484 out:
1485         if (rc != 0)
1486                 lfsck_object_put(env, obj);
1487         else
1488                 lls->lls_lastid_obj = obj;
1489
1490         return rc;
1491 }
1492
1493 static void lfsck_layout_record_failure(const struct lu_env *env,
1494                                         struct lfsck_instance *lfsck,
1495                                         struct lfsck_layout *lo)
1496 {
1497         __u64 cookie;
1498
1499         lo->ll_objs_failed_phase1++;
1500         cookie = lfsck->li_obj_oit->do_index_ops->dio_it.store(env,
1501                                                         lfsck->li_di_oit);
1502         if (lo->ll_pos_first_inconsistent == 0 ||
1503             lo->ll_pos_first_inconsistent < cookie) {
1504                 lo->ll_pos_first_inconsistent = cookie;
1505
1506                 CDEBUG(D_LFSCK, "%s: layout LFSCK hit first non-repaired "
1507                        "inconsistency at the pos [%llu]\n",
1508                        lfsck_lfsck2name(lfsck),
1509                        lo->ll_pos_first_inconsistent);
1510         }
1511 }
1512
1513 static int lfsck_layout_double_scan_result(const struct lu_env *env,
1514                                            struct lfsck_component *com,
1515                                            int rc)
1516 {
1517         struct lfsck_instance   *lfsck = com->lc_lfsck;
1518         struct lfsck_layout     *lo    = com->lc_file_ram;
1519
1520         CDEBUG(D_LFSCK, "%s: layout LFSCK double scan: rc = %d\n",
1521                lfsck_lfsck2name(lfsck), rc);
1522
1523         down_write(&com->lc_sem);
1524         lo->ll_run_time_phase2 += ktime_get_seconds() -
1525                                   com->lc_time_last_checkpoint;
1526         lo->ll_time_last_checkpoint = ktime_get_real_seconds();
1527         lo->ll_objs_checked_phase2 += com->lc_new_checked;
1528
1529         if (rc > 0) {
1530                 if (lo->ll_flags & LF_INCOMPLETE) {
1531                         lo->ll_status = LS_PARTIAL;
1532                 } else {
1533                         if (lfsck->li_master) {
1534                                 struct lfsck_assistant_data *lad = com->lc_data;
1535
1536                                 if (test_bit(LAD_INCOMPLETE, &lad->lad_flags))
1537                                         lo->ll_status = LS_PARTIAL;
1538                                 else
1539                                         lo->ll_status = LS_COMPLETED;
1540                         } else {
1541                                 lo->ll_status = LS_COMPLETED;
1542                         }
1543                 }
1544                 lo->ll_flags &= ~LF_SCANNED_ONCE;
1545                 if (!(lfsck->li_bookmark_ram.lb_param & LPF_DRYRUN))
1546                         lo->ll_flags &= ~LF_INCONSISTENT;
1547                 lo->ll_time_last_complete = lo->ll_time_last_checkpoint;
1548                 lo->ll_success_count++;
1549         } else if (rc == 0) {
1550                 if (lfsck->li_status != 0)
1551                         lo->ll_status = lfsck->li_status;
1552                 else
1553                         lo->ll_status = LS_STOPPED;
1554         } else {
1555                 lo->ll_status = LS_FAILED;
1556         }
1557
1558         rc = lfsck_layout_store(env, com);
1559         up_write(&com->lc_sem);
1560
1561         CDEBUG(D_LFSCK, "%s: layout LFSCK double scan result %u: rc = %d\n",
1562                lfsck_lfsck2name(lfsck), lo->ll_status, rc);
1563
1564         return rc;
1565 }
1566
1567 static int lfsck_layout_trans_stop(const struct lu_env *env,
1568                                    struct dt_device *dev,
1569                                    struct thandle *handle, int result)
1570 {
1571         int rc;
1572
1573         /* XXX: If there is something worng or it needs to repair nothing,
1574          *      then notify the lower to stop the modification. Currently,
1575          *      we use th_result for such purpose, that may be replaced by
1576          *      some rollback mechanism in the future. */
1577         handle->th_result = result;
1578         rc = dt_trans_stop(env, dev, handle);
1579         if (result != 0)
1580                 return result > 0 ? 0 : result;
1581
1582         return rc == 0 ? 1 : rc;
1583 }
1584
1585 static int lfsck_layout_ins_dangling_rec(const struct lu_env *env,
1586                                          struct lfsck_component *com,
1587                                          const struct lu_fid *pfid,
1588                                          const struct lu_fid *cfid,
1589                                          __u32 comp_id, __u32 ea_off,
1590                                          __u32 ost_idx)
1591 {
1592         struct lfsck_layout_dangling_key *key = &lfsck_env_info(env)->lti_lldk;
1593         struct lu_fid *rec = &lfsck_env_info(env)->lti_fid3;
1594         struct dt_device *dev;
1595         struct dt_object *obj;
1596         struct thandle *th = NULL;
1597         int idx;
1598         int rc = 0;
1599         ENTRY;
1600
1601         idx = lfsck_sub_trace_file_fid2idx(pfid);
1602         obj = com->lc_sub_trace_objs[idx].lsto_obj;
1603         dev = lfsck_obj2dev(obj);
1604
1605         fid_cpu_to_be(&key->lldk_fid, pfid);
1606         key->lldk_comp_id = cpu_to_be32(comp_id);
1607         key->lldk_ea_off = cpu_to_be32(ea_off);
1608
1609         fid_cpu_to_be(rec, cfid);
1610         rec->f_ver = cpu_to_be32(ost_idx);
1611
1612         mutex_lock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1613
1614         th = dt_trans_create(env, dev);
1615         if (IS_ERR(th))
1616                 GOTO(unlock, rc = PTR_ERR(th));
1617
1618         rc = dt_declare_insert(env, obj,
1619                                (const struct dt_rec *)rec,
1620                                (const struct dt_key *)key, th);
1621         if (rc)
1622                 GOTO(unlock, rc);
1623
1624         rc = dt_trans_start_local(env, dev, th);
1625         if (rc)
1626                 GOTO(unlock, rc);
1627
1628         rc = dt_insert(env, obj, (const struct dt_rec *)rec,
1629                        (const struct dt_key *)key, th);
1630
1631         GOTO(unlock, rc);
1632
1633 unlock:
1634         if (th && !IS_ERR(th))
1635                 dt_trans_stop(env, dev, th);
1636
1637         mutex_unlock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1638
1639         CDEBUG(D_LFSCK, "%s: insert the paris "DFID" => "DFID", comp_id = %u, "
1640                "ea_off = %u, ost_idx = %u, into the trace file for further "
1641                "dangling check: rc = %d\n", lfsck_lfsck2name(com->lc_lfsck),
1642                PFID(pfid), PFID(cfid), comp_id, ea_off, ost_idx, rc);
1643
1644         return rc;
1645 }
1646
1647 static int lfsck_layout_del_dangling_rec(const struct lu_env *env,
1648                                          struct lfsck_component *com,
1649                                          const struct lu_fid *fid,
1650                                          __u32 comp_id, __u32 ea_off)
1651 {
1652         struct lfsck_layout_dangling_key *key = &lfsck_env_info(env)->lti_lldk;
1653         struct dt_device *dev;
1654         struct dt_object *obj;
1655         struct thandle *th = NULL;
1656         int idx;
1657         int rc = 0;
1658         ENTRY;
1659
1660         idx = lfsck_sub_trace_file_fid2idx(fid);
1661         obj = com->lc_sub_trace_objs[idx].lsto_obj;
1662         dev = lfsck_obj2dev(obj);
1663
1664         fid_cpu_to_be(&key->lldk_fid, fid);
1665         key->lldk_comp_id = cpu_to_be32(comp_id);
1666         key->lldk_ea_off = cpu_to_be32(ea_off);
1667
1668         mutex_lock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1669
1670         th = dt_trans_create(env, dev);
1671         if (IS_ERR(th))
1672                 GOTO(unlock, rc = PTR_ERR(th));
1673
1674         rc = dt_declare_delete(env, obj, (const struct dt_key *)key, th);
1675         if (rc)
1676                 GOTO(unlock, rc);
1677
1678         rc = dt_trans_start_local(env, dev, th);
1679         if (rc)
1680                 GOTO(unlock, rc);
1681
1682         rc = dt_delete(env, obj, (const struct dt_key *)key, th);
1683
1684         GOTO(unlock, rc);
1685
1686 unlock:
1687         if (th && !IS_ERR(th))
1688                 dt_trans_stop(env, dev, th);
1689
1690         mutex_unlock(&com->lc_sub_trace_objs[idx].lsto_mutex);
1691
1692         CDEBUG(D_LFSCK, "%s: delete the dangling record for "DFID
1693                ", comp_id = %u, ea_off = %u from the trace file: rc = %d\n",
1694                lfsck_lfsck2name(com->lc_lfsck), PFID(fid), comp_id, ea_off, rc);
1695
1696         return rc;
1697 }
1698
1699 /**
1700  * Get the system default stripe size.
1701  *
1702  * \param[in] env       pointer to the thread context
1703  * \param[in] lfsck     pointer to the lfsck instance
1704  * \param[out] size     pointer to the default stripe size
1705  *
1706  * \retval              0 for success
1707  * \retval              negative error number on failure
1708  */
1709 static int lfsck_layout_get_def_stripesize(const struct lu_env *env,
1710                                            struct lfsck_instance *lfsck,
1711                                            __u32 *size)
1712 {
1713         struct lov_user_md      *lum = &lfsck_env_info(env)->lti_lum;
1714         struct dt_object        *root;
1715         int                      rc;
1716
1717         root = dt_locate(env, lfsck->li_next, &lfsck->li_local_root_fid);
1718         if (IS_ERR(root))
1719                 return PTR_ERR(root);
1720
1721         /* Get the default stripe size via xattr_get on the backend root. */
1722         rc = dt_xattr_get(env, root, lfsck_buf_get(env, lum, sizeof(*lum)),
1723                           XATTR_NAME_LOV);
1724         if (rc > 0) {
1725                 /* The lum->lmm_stripe_size is LE mode. The *size also
1726                  * should be LE mode. So it is unnecessary to convert. */
1727                 *size = lum->lmm_stripe_size;
1728                 rc = 0;
1729         } else if (unlikely(rc == 0)) {
1730                 rc = -EINVAL;
1731         }
1732
1733         lfsck_object_put(env, root);
1734
1735         return rc;
1736 }
1737
1738 /**
1739  * \retval       +1: repaired
1740  * \retval        0: did nothing
1741  * \retval      -ve: on error
1742  */
1743 static int lfsck_layout_refill_lovea(const struct lu_env *env,
1744                                      struct lfsck_instance *lfsck,
1745                                      struct thandle *handle,
1746                                      struct dt_object *parent,
1747                                      const struct lu_fid *cfid,
1748                                      struct lu_buf *buf,
1749                                      struct lov_mds_md_v1 *lmm,
1750                                      struct lov_ost_data_v1 *slot,
1751                                      int fl, __u32 ost_idx, int size)
1752 {
1753         struct ost_id           *oi     = &lfsck_env_info(env)->lti_oi;
1754         struct lu_buf            ea_buf;
1755         int                      rc;
1756         __u32                    magic;
1757         __u32                    pattern;
1758         __u16                    count;
1759         ENTRY;
1760
1761         magic = le32_to_cpu(lmm->lmm_magic);
1762         pattern = le32_to_cpu(lmm->lmm_pattern);
1763         count = le16_to_cpu(lmm->lmm_stripe_count);
1764
1765         fid_to_ostid(cfid, oi);
1766         ostid_cpu_to_le(oi, &slot->l_ost_oi);
1767         slot->l_ost_gen = cpu_to_le32(0);
1768         slot->l_ost_idx = cpu_to_le32(ost_idx);
1769
1770         if (pattern & LOV_PATTERN_F_HOLE) {
1771                 struct lov_ost_data_v1 *objs;
1772                 int                     i;
1773
1774                 if (magic == LOV_MAGIC_V1)
1775                         objs = &lmm->lmm_objects[0];
1776                 else
1777                         objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
1778                 for (i = 0; i < count; i++, objs++) {
1779                         if (lovea_slot_is_dummy(objs))
1780                                 break;
1781                 }
1782
1783                 /* If the @slot is the last dummy slot to be refilled,
1784                  * then drop LOV_PATTERN_F_HOLE from lmm::lmm_pattern. */
1785                 if (i == count) {
1786                         lmm->lmm_pattern =
1787                                 cpu_to_le32(pattern & ~LOV_PATTERN_F_HOLE);
1788
1789                         CDEBUG(D_LFSCK, "%s: remove layout HOLE for "DFID
1790                                ": parent "DFID"\n", lfsck_lfsck2name(lfsck),
1791                                PFID(cfid), PFID(lfsck_dto2fid(parent)));
1792                 }
1793         }
1794
1795         lfsck_buf_init(&ea_buf, buf->lb_buf, size);
1796         rc = dt_xattr_set(env, parent, &ea_buf, XATTR_NAME_LOV, fl, handle);
1797         if (rc == 0)
1798                 rc = 1;
1799
1800         RETURN(rc);
1801 }
1802
1803 static struct lov_ost_data_v1 *
1804 __lfsck_layout_new_v1_lovea(struct lov_mds_md_v1 *lmm,
1805                             const struct lu_fid *pfid,
1806                             __u32 stripe_size, __u32 ea_off,
1807                             __u32 pattern, __u16 count)
1808 {
1809         lmm->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
1810         lmm->lmm_pattern = cpu_to_le32(pattern);
1811         fid_to_lmm_oi(pfid, &lmm->lmm_oi);
1812         lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
1813         lmm->lmm_stripe_size = cpu_to_le32(stripe_size);
1814         lmm->lmm_stripe_count = cpu_to_le16(count);
1815         lmm->lmm_layout_gen = cpu_to_le16(1);
1816         memset(&lmm->lmm_objects[0], 0,
1817                sizeof(struct lov_ost_data_v1) * count);
1818
1819         return &lmm->lmm_objects[ea_off];
1820 }
1821
1822 static int lfsck_layout_new_v1_lovea(const struct lu_env *env,
1823                                      struct lfsck_instance *lfsck,
1824                                      struct ost_layout *ol,
1825                                      struct dt_object *parent,
1826                                      struct lu_buf *buf, __u32 ea_off,
1827                                      struct lov_mds_md_v1 **lmm,
1828                                      struct lov_ost_data_v1 **objs)
1829 {
1830         int size;
1831         __u32 stripe_size = ol->ol_stripe_size;
1832         __u32 pattern = LOV_PATTERN_RAID0;
1833         __u16 count;
1834
1835         if (ol->ol_stripe_count != 0)
1836                 count = ol->ol_stripe_count;
1837         else
1838                 count = ea_off + 1;
1839
1840         size = lov_mds_md_size(count, LOV_MAGIC_V1);
1841         LASSERTF(buf->lb_len >= size,
1842                  "buffer len %d is less than real size %d\n",
1843                  (int)buf->lb_len, size);
1844
1845         if (stripe_size == 0) {
1846                 int rc;
1847
1848                 rc = lfsck_layout_get_def_stripesize(env, lfsck, &stripe_size);
1849                 if (rc)
1850                         return rc;
1851         }
1852
1853         *lmm = buf->lb_buf;
1854         if (ol->ol_stripe_count > 1 ||
1855             (ol->ol_stripe_count == 0 && ea_off != 0)) {
1856                 pattern |= LOV_PATTERN_F_HOLE;
1857                 memset(&(*lmm)->lmm_objects[0], 0,
1858                        count * sizeof(struct lov_ost_data_v1));
1859         }
1860
1861         *objs = __lfsck_layout_new_v1_lovea(*lmm, lfsck_dto2fid(parent),
1862                                 stripe_size, ea_off, pattern, count);
1863
1864         return size;
1865 }
1866
1867 static int lfsck_layout_new_comp_lovea(const struct lu_env *env,
1868                                        struct lu_orphan_rec_v3 *rec,
1869                                        struct dt_object *parent,
1870                                        struct lu_buf *buf, __u32 ea_off,
1871                                        struct lov_mds_md_v1 **lmm,
1872                                        struct lov_ost_data_v1 **objs)
1873 {
1874         struct ost_layout *ol = &rec->lor_layout;
1875         struct lov_comp_md_v1 *lcm;
1876         struct lov_comp_md_entry_v1 *lcme;
1877         __u32 pattern = LOV_PATTERN_RAID0;
1878         __u32 offset = sizeof(*lcm) + sizeof(*lcme);
1879         int lcme_size = lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
1880         int size = offset + lcme_size;
1881
1882         LASSERTF(buf->lb_len >= size,
1883                  "buffer len %d is less than real size %d\n",
1884                  (int)buf->lb_len, size);
1885
1886         lcm = buf->lb_buf;
1887         lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
1888         lcm->lcm_size = cpu_to_le32(size);
1889         if (rec->lor_range) {
1890                 lcm->lcm_layout_gen = cpu_to_le32(rec->lor_layout_version +
1891                                                   rec->lor_range);
1892                 lcm->lcm_flags = cpu_to_le16(LCM_FL_WRITE_PENDING);
1893         } else if (rec->lor_layout_version) {
1894                 lcm->lcm_layout_gen = cpu_to_le32(rec->lor_layout_version +
1895                                                   rec->lor_range);
1896                 lcm->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1897         } else {
1898                 lcm->lcm_layout_gen = cpu_to_le32(1);
1899                 lcm->lcm_flags = cpu_to_le16(LCM_FL_NONE);
1900         }
1901         lcm->lcm_entry_count = cpu_to_le16(1);
1902         /* Currently, we do not know how many mirrors will be, set it as zero
1903          * at the beginning. It will be updated when more mirrors are found. */
1904         lcm->lcm_mirror_count = 0;
1905
1906         lcme = &lcm->lcm_entries[0];
1907         lcme->lcme_id = cpu_to_le32(ol->ol_comp_id);
1908         lcme->lcme_flags = cpu_to_le32(LCME_FL_INIT);
1909         lcme->lcme_extent.e_start = cpu_to_le64(ol->ol_comp_start);
1910         lcme->lcme_extent.e_end = cpu_to_le64(ol->ol_comp_end);
1911         lcme->lcme_offset = cpu_to_le32(offset);
1912         lcme->lcme_size = cpu_to_le32(lcme_size);
1913         lcme->lcme_layout_gen = lcm->lcm_layout_gen;
1914         if (ol->ol_stripe_count > 1)
1915                 pattern |= LOV_PATTERN_F_HOLE;
1916
1917         *lmm = buf->lb_buf + offset;
1918         *objs = __lfsck_layout_new_v1_lovea(*lmm, lfsck_dto2fid(parent),
1919                                             ol->ol_stripe_size, ea_off,
1920                                             pattern, ol->ol_stripe_count);
1921
1922         return size;
1923 }
1924
1925 static void lfsck_layout_update_lcm(struct lov_comp_md_v1 *lcm,
1926                                     struct lov_comp_md_entry_v1 *lcme,
1927                                     __u32 version, __u32 range)
1928 {
1929         struct lov_comp_md_entry_v1 *tmp;
1930         __u64 start = le64_to_cpu(lcme->lcme_extent.e_start);
1931         __u64 end = le64_to_cpu(lcme->lcme_extent.e_end);
1932         __u32 gen = version + range;
1933         __u32 tmp_gen;
1934         int i;
1935         __u16 count = le16_to_cpu(lcm->lcm_entry_count);
1936         __u16 flags = le16_to_cpu(lcm->lcm_flags);
1937
1938         if (!gen)
1939                 gen = 1;
1940         lcme->lcme_layout_gen = cpu_to_le32(gen);
1941         if (le32_to_cpu(lcm->lcm_layout_gen) < gen)
1942                 lcm->lcm_layout_gen = cpu_to_le32(gen);
1943
1944         if (range)
1945                 lcm->lcm_flags = cpu_to_le16(LCM_FL_WRITE_PENDING);
1946         else if (flags == LCM_FL_NONE && le16_to_cpu(lcm->lcm_mirror_count) > 0)
1947                 lcm->lcm_flags = cpu_to_le16(LCM_FL_RDONLY);
1948
1949         for (i = 0; i < count; i++) {
1950                 tmp = &lcm->lcm_entries[i];
1951                 if (le64_to_cpu(tmp->lcme_extent.e_end) <= start)
1952                         continue;
1953
1954                 if (le64_to_cpu(tmp->lcme_extent.e_start) >= end)
1955                         continue;
1956
1957                 if (le32_to_cpu(tmp->lcme_flags) & LCME_FL_STALE)
1958                         continue;
1959
1960                 tmp_gen = le32_to_cpu(tmp->lcme_layout_gen);
1961                 /* "lcme_layout_gen == 0" but without LCME_FL_STALE flag,
1962                  * then it should be the latest version of all mirrors. */
1963                 if (tmp_gen == 0 || tmp_gen > gen) {
1964                         lcme->lcme_flags = cpu_to_le32(
1965                                 le32_to_cpu(lcme->lcme_flags) | LCME_FL_STALE);
1966                         break;
1967                 }
1968
1969                 if (tmp_gen < gen)
1970                         tmp->lcme_flags = cpu_to_le32(
1971                                 le32_to_cpu(tmp->lcme_flags) | LCME_FL_STALE);
1972         }
1973 }
1974
1975 static int lfsck_layout_add_comp(const struct lu_env *env,
1976                                  struct lfsck_instance *lfsck,
1977                                  struct thandle *handle,
1978                                  struct lu_orphan_rec_v3 *rec,
1979                                  struct dt_object *parent,
1980                                  const struct lu_fid *cfid,
1981                                  struct lu_buf *buf, __u32 ost_idx,
1982                                  __u32 ea_off, int pos, bool new_mirror)
1983 {
1984         struct ost_layout *ol = &rec->lor_layout;
1985         struct lov_comp_md_v1 *lcm = buf->lb_buf;
1986         struct lov_comp_md_entry_v1 *lcme;
1987         struct lov_mds_md_v1 *lmm;
1988         struct lov_ost_data_v1 *objs;
1989         int added = sizeof(*lcme) +
1990                     lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
1991         int size = le32_to_cpu(lcm->lcm_size) + added;
1992         int rc;
1993         int i;
1994         __u32 offset;
1995         __u32 pattern = LOV_PATTERN_RAID0;
1996         __u16 count = le16_to_cpu(lcm->lcm_entry_count);
1997         ENTRY;
1998
1999         lu_buf_check_and_grow(buf, size);
2000         /* set the lcm again because lu_buf_check_and_grow() may
2001          * have reallocated the buf. */
2002         lcm = buf->lb_buf;
2003         lcm->lcm_size = cpu_to_le32(size);
2004         lcm->lcm_entry_count = cpu_to_le16(count + 1);
2005         if (new_mirror)
2006                 le16_add_cpu(&lcm->lcm_mirror_count, 1);
2007
2008         /* 1. Move the component bodies from [pos, count-1] to [pos+1, count]
2009          *    with distance of 'added'. */
2010         if (pos < count) {
2011                 size = 0;
2012                 for (i = pos; i < count; i++) {
2013                         lcme = &lcm->lcm_entries[i];
2014                         size += le32_to_cpu(lcme->lcme_size);
2015                 }
2016
2017                 offset = le32_to_cpu(lcm->lcm_entries[pos].lcme_offset);
2018                 memmove(buf->lb_buf + offset + added,
2019                         buf->lb_buf + offset, size);
2020         }
2021
2022         size = 0;
2023         /* 2. Move the component header [0, pos-1] to [0, pos-1] with distance
2024          *    of 'sizeof(struct lov_comp_md_entry_v1)' */
2025         if (pos > 0) {
2026                 for (i = 0; i < pos; i++) {
2027                         lcme = &lcm->lcm_entries[i];
2028                         size += le32_to_cpu(lcme->lcme_size);
2029                 }
2030
2031                 offset = le32_to_cpu(lcm->lcm_entries[0].lcme_offset);
2032                 memmove(buf->lb_buf + offset + sizeof(*lcme),
2033                         buf->lb_buf + offset, size);
2034         }
2035
2036         /* 3. Recalculate the enter offset for the component [pos, count-1] */
2037         for (i = count - 1; i >= pos; i--) {
2038                 lcm->lcm_entries[i + 1] = lcm->lcm_entries[i];
2039                 lcm->lcm_entries[i + 1].lcme_offset =
2040                         cpu_to_le32(le32_to_cpu(lcm->lcm_entries[i + 1].
2041                                                 lcme_offset) + added);
2042         }
2043
2044         /* 4. Recalculate the enter offset for the component [0, pos) */
2045         for (i = 0; i < pos; i++) {
2046                 lcm->lcm_entries[i].lcme_offset =
2047                         cpu_to_le32(le32_to_cpu(lcm->lcm_entries[i].
2048                                                 lcme_offset) + sizeof(*lcme));
2049         }
2050
2051         offset = sizeof(*lcm) + sizeof(*lcme) * (count + 1) + size;
2052         /* 4. Insert the new component header (entry) at the slot 'pos'. */
2053         lcme = &lcm->lcm_entries[pos];
2054         lcme->lcme_id = cpu_to_le32(ol->ol_comp_id);
2055         lcme->lcme_flags = cpu_to_le32(LCME_FL_INIT);
2056         lcme->lcme_extent.e_start = cpu_to_le64(ol->ol_comp_start);
2057         lcme->lcme_extent.e_end = cpu_to_le64(ol->ol_comp_end);
2058         lcme->lcme_offset = cpu_to_le32(offset);
2059         lcme->lcme_size = cpu_to_le32(lov_mds_md_size(ol->ol_stripe_count,
2060                                                       LOV_MAGIC_V1));
2061
2062         if (ol->ol_stripe_count > 1)
2063                 pattern |= LOV_PATTERN_F_HOLE;
2064
2065         lmm = buf->lb_buf + offset;
2066         /* 5. Insert teh new component body at the 'offset'. */
2067         objs = __lfsck_layout_new_v1_lovea(lmm, lfsck_dto2fid(parent),
2068                                            ol->ol_stripe_size, ea_off,
2069                                            pattern, ol->ol_stripe_count);
2070
2071         /* 6. Update mirror related flags and version. */
2072         lfsck_layout_update_lcm(lcm, lcme, rec->lor_layout_version,
2073                                 rec->lor_range);
2074
2075         rc = lfsck_layout_refill_lovea(env, lfsck, handle, parent, cfid, buf,
2076                                        lmm, objs, LU_XATTR_REPLACE, ost_idx,
2077                                        le32_to_cpu(lcm->lcm_size));
2078
2079         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant add new COMP for "
2080                DFID": parent "DFID", OST-index %u, stripe-index %u, "
2081                "stripe_size %u, stripe_count %u, comp_id %u, comp_start %llu, "
2082                "comp_end %llu, layout version %u, range %u, "
2083                "%s LOV EA hole: rc = %d\n",
2084                lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
2085                ost_idx, ea_off, ol->ol_stripe_size, ol->ol_stripe_count,
2086                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
2087                rec->lor_layout_version, rec->lor_range,
2088                le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE ?
2089                "with" : "without", rc);
2090
2091         RETURN(rc);
2092 }
2093
2094 static int lfsck_layout_extend_v1v3_lovea(const struct lu_env *env,
2095                                           struct lfsck_instance *lfsck,
2096                                           struct thandle *handle,
2097                                           struct ost_layout *ol,
2098                                           struct dt_object *parent,
2099                                           const struct lu_fid *cfid,
2100                                           struct lu_buf *buf, __u32 ost_idx,
2101                                           __u32 ea_off)
2102 {
2103         struct lov_mds_md_v1 *lmm = buf->lb_buf;
2104         struct lov_ost_data_v1 *objs;
2105         __u16 count = le16_to_cpu(lmm->lmm_stripe_count);
2106         __u32 magic = le32_to_cpu(lmm->lmm_magic);
2107         int size;
2108         int gap;
2109         int rc;
2110         ENTRY;
2111
2112         /* The original LOVEA maybe re-generated via old filter_fid, at
2113          * that time, we do not know the stripe count and stripe size. */
2114         if (ol->ol_stripe_count > count)
2115                 count = ol->ol_stripe_count;
2116         if (ol->ol_stripe_size != 0 &&
2117             ol->ol_stripe_size != le32_to_cpu(lmm->lmm_stripe_size))
2118                 lmm->lmm_stripe_size = cpu_to_le32(ol->ol_stripe_size);
2119
2120         if (magic == LOV_MAGIC_V1)
2121                 objs = &lmm->lmm_objects[count];
2122         else
2123                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[count];
2124
2125         gap = ea_off - count;
2126         if (gap >= 0)
2127                 count = ea_off + 1;
2128
2129         size = lov_mds_md_size(count, magic);
2130         LASSERTF(buf->lb_len >= size,
2131                  "buffer len %d is less than real size %d\n",
2132                  (int)buf->lb_len, size);
2133
2134         if (gap > 0) {
2135                 memset(objs, 0, gap * sizeof(*objs));
2136                 lmm->lmm_pattern |= cpu_to_le32(LOV_PATTERN_F_HOLE);
2137         }
2138
2139         lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
2140         lmm->lmm_stripe_count = cpu_to_le16(count);
2141         objs += gap;
2142
2143         rc = lfsck_layout_refill_lovea(env, lfsck, handle, parent, cfid, buf,
2144                                 lmm, objs, LU_XATTR_REPLACE, ost_idx, size);
2145
2146         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant extend layout EA for "
2147                DFID": parent "DFID", OST-index %u, stripe-index %u, "
2148                "stripe_size %u, stripe_count %u, comp_id %u, comp_start %llu, "
2149                "comp_end %llu, %s LOV EA hole: rc = %d\n",
2150                lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
2151                ost_idx, ea_off, ol->ol_stripe_size, ol->ol_stripe_count,
2152                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
2153                le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE ?
2154                "with" : "without", rc);
2155
2156         RETURN(rc);
2157 }
2158
2159 /**
2160  * \retval       +1: repaired
2161  * \retval        0: did nothing
2162  * \retval      -ve: on error
2163  */
2164 static int lfsck_layout_update_lovea(const struct lu_env *env,
2165                                      struct lfsck_instance *lfsck,
2166                                      struct thandle *handle,
2167                                      struct lu_orphan_rec_v3 *rec,
2168                                      struct dt_object *parent,
2169                                      const struct lu_fid *cfid,
2170                                      struct lu_buf *buf, int fl,
2171                                      __u32 ost_idx, __u32 ea_off)
2172 {
2173         struct ost_layout *ol = &rec->lor_layout;
2174         struct lov_mds_md_v1 *lmm = NULL;
2175         struct lov_ost_data_v1 *objs = NULL;
2176         int rc = 0;
2177         ENTRY;
2178
2179         if (ol->ol_comp_id != 0)
2180                 rc = lfsck_layout_new_comp_lovea(env, rec, parent, buf, ea_off,
2181                                                  &lmm, &objs);
2182         else
2183                 rc = lfsck_layout_new_v1_lovea(env, lfsck, &rec->lor_layout,
2184                                                parent, buf, ea_off, &lmm,
2185                                                &objs);
2186         if (rc > 0)
2187                 rc = lfsck_layout_refill_lovea(env, lfsck, handle, parent, cfid,
2188                                                buf, lmm, objs, fl, ost_idx, rc);
2189
2190         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant created layout EA for "
2191                DFID": parent "DFID", OST-index %u, stripe-index %u, "
2192                "stripe_size %u, stripe_count %u, comp_id %u, comp_start %llu, "
2193                "comp_end %llu, layout version %u, range %u, fl %d, "
2194                "%s LOV EA hole: rc = %d\n",
2195                lfsck_lfsck2name(lfsck), PFID(cfid), PFID(lfsck_dto2fid(parent)),
2196                ost_idx, ea_off, ol->ol_stripe_size, ol->ol_stripe_count,
2197                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
2198                rec->lor_layout_version, rec->lor_range, fl,
2199                le32_to_cpu(lmm->lmm_pattern) & LOV_PATTERN_F_HOLE ?
2200                "with" : "without", rc);
2201
2202         RETURN(rc);
2203 }
2204
2205 static int __lfsck_layout_update_pfid(const struct lu_env *env,
2206                                       struct dt_object *child,
2207                                       const struct lu_fid *pfid,
2208                                       const struct ost_layout *ol, __u32 offset,
2209                                       __u32 version, __u32 range)
2210 {
2211         struct dt_device        *dev    = lfsck_obj2dev(child);
2212         struct filter_fid       *ff     = &lfsck_env_info(env)->lti_ff;
2213         struct thandle          *handle;
2214         struct lu_buf            buf    = { NULL };
2215         int                      rc;
2216
2217         ff->ff_parent.f_seq = cpu_to_le64(pfid->f_seq);
2218         ff->ff_parent.f_oid = cpu_to_le32(pfid->f_oid);
2219         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
2220          * MDT-object's FID::f_ver, instead it is the OST-object index in its
2221          * parent MDT-object's layout EA. */
2222         ff->ff_parent.f_stripe_idx = cpu_to_le32(offset);
2223         ost_layout_cpu_to_le(&ff->ff_layout, ol);
2224         ff->ff_layout_version = cpu_to_le32(version);
2225         ff->ff_range = cpu_to_le32(range);
2226         lfsck_buf_init(&buf, ff, sizeof(*ff));
2227
2228         handle = dt_trans_create(env, dev);
2229         if (IS_ERR(handle))
2230                 RETURN(PTR_ERR(handle));
2231
2232         rc = dt_declare_xattr_set(env, child, &buf, XATTR_NAME_FID, 0, handle);
2233         if (rc != 0)
2234                 GOTO(stop, rc);
2235
2236         rc = dt_trans_start_local(env, dev, handle);
2237         if (rc != 0)
2238                 GOTO(stop, rc);
2239
2240         rc = dt_xattr_set(env, child, &buf, XATTR_NAME_FID, 0, handle);
2241
2242         GOTO(stop, rc);
2243
2244 stop:
2245         dt_trans_stop(env, dev, handle);
2246
2247         return rc;
2248 }
2249
2250 /**
2251  * \retval       +1: repaired
2252  * \retval        0: did nothing
2253  * \retval      -ve: on error
2254  */
2255 static int lfsck_layout_update_pfid(const struct lu_env *env,
2256                                     struct lfsck_component *com,
2257                                     struct dt_object *parent,
2258                                     struct lu_fid *cfid,
2259                                     struct dt_device *cdev,
2260                                     struct lu_orphan_rec_v3 *rec, __u32 ea_off)
2261 {
2262         struct dt_object        *child;
2263         int                      rc     = 0;
2264         ENTRY;
2265
2266         child = lfsck_object_find_by_dev(env, cdev, cfid);
2267         if (IS_ERR(child))
2268                 RETURN(PTR_ERR(child));
2269
2270         rc = __lfsck_layout_update_pfid(env, child,
2271                                         lu_object_fid(&parent->do_lu),
2272                                         &rec->lor_layout, ea_off,
2273                                         rec->lor_layout_version,
2274                                         rec->lor_range);
2275         lfsck_object_put(env, child);
2276
2277         RETURN(rc == 0 ? 1 : rc);
2278 }
2279
2280 static int lfsck_lovea_size(struct ost_layout *ol, __u32 ea_off)
2281 {
2282         if (ol->ol_comp_id != 0)
2283                 return sizeof(struct lov_comp_md_v1) +
2284                        sizeof(struct lov_comp_md_entry_v1) +
2285                        lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
2286
2287         if (ol->ol_stripe_count != 0)
2288                 return lov_mds_md_size(ol->ol_stripe_count, LOV_MAGIC_V1);
2289
2290         return lov_mds_md_size(ea_off + 1, LOV_MAGIC_V1);
2291 }
2292
2293 /**
2294  * This function will create the MDT-object with the given (partial) LOV EA.
2295  *
2296  * Under some data corruption cases, the MDT-object of the file may be lost,
2297  * but its OST-objects, or some of them are there. The layout LFSCK needs to
2298  * re-create the MDT-object with the orphan OST-object(s) information.
2299  *
2300  * On the other hand, the LFSCK may has created some OST-object for repairing
2301  * dangling LOV EA reference, but as the LFSCK processing, it may find that
2302  * the old OST-object is there and should replace the former new created OST
2303  * object. Unfortunately, some others have modified such newly created object.
2304  * To keep the data (both new and old), the LFSCK will create MDT-object with
2305  * new FID to reference the original OST-object.
2306  *
2307  * \param[in] env       pointer to the thread context
2308  * \param[in] com       pointer to the lfsck component
2309  * \param[in] ltd       pointer to target device descriptor
2310  * \param[in] rec       pointer to the record for the orphan OST-object
2311  * \param[in] cfid      pointer to FID for the orphan OST-object
2312  * \param[in] infix     additional information, such as the FID for original
2313  *                      MDT-object and the stripe offset in the LOV EA
2314  * \param[in] type      the type for describing why the orphan MDT-object is
2315  *                      created. The rules are as following:
2316  *
2317  *  type "C":           Multiple OST-objects claim the same MDT-object and the
2318  *                      same slot in the layout EA. Then the LFSCK will create
2319  *                      new MDT-object(s) to hold the conflict OST-object(s).
2320  *
2321  *  type "N":           The orphan OST-object does not know which one was the
2322  *                      real parent MDT-object, so the LFSCK uses new FID for
2323  *                      its parent MDT-object.
2324  *
2325  *  type "R":           The orphan OST-object knows its parent MDT-object FID,
2326  *                      but does not know the position (the file name) in the
2327  *                      layout.
2328  *
2329  *  type "D":           The MDT-object is a directory, it may knows its parent
2330  *                      but because there is no valid linkEA, the LFSCK cannot
2331  *                      know where to put it back to the namespace.
2332  *  type "O":           The MDT-object has no linkEA, and there is no name
2333  *                      entry that references the MDT-object.
2334  *
2335  *  type "P":           The orphan object to be created was a parent directory
2336  *                      of some MDT-object which linkEA shows that the @orphan
2337  *                      object is missing.
2338  *
2339  * The orphan name will be like:
2340  * ${FID}-${infix}-${type}-${conflict_version}
2341  *
2342  * \param[in] ea_off    the stripe offset in the LOV EA
2343  *
2344  * \retval              positive on repaired something
2345  * \retval              0 if needs to repair nothing
2346  * \retval              negative error number on failure
2347  */
2348 static int lfsck_layout_recreate_parent(const struct lu_env *env,
2349                                         struct lfsck_component *com,
2350                                         struct lfsck_tgt_desc *ltd,
2351                                         struct lu_orphan_rec_v3 *rec,
2352                                         struct lu_fid *cfid,
2353                                         const char *infix,
2354                                         const char *type,
2355                                         __u32 ea_off)
2356 {
2357         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2358         struct dt_insert_rec            *dtrec  = &info->lti_dt_rec;
2359         char                            *name   = info->lti_key;
2360         struct lu_attr                  *la     = &info->lti_la2;
2361         struct dt_object_format         *dof    = &info->lti_dof;
2362         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2363         struct lu_fid                   *pfid   = &rec->lor_rec.lor_fid;
2364         struct lu_fid                   *tfid   = &info->lti_fid3;
2365         struct dt_device                *dev    = lfsck->li_bottom;
2366         struct dt_object                *lpf    = lfsck->li_lpf_obj;
2367         struct dt_object                *pobj   = NULL;
2368         struct dt_object                *cobj   = NULL;
2369         struct thandle                  *th     = NULL;
2370         struct lu_buf                   *ea_buf = &info->lti_big_buf;
2371         struct lu_buf                    lov_buf;
2372         struct lfsck_lock_handle        *llh    = &info->lti_llh;
2373         struct linkea_data               ldata  = { NULL };
2374         struct lu_buf                    linkea_buf;
2375         const struct lu_name            *pname;
2376         int                              size   = 0;
2377         int                              idx    = 0;
2378         int                              rc     = 0;
2379         ENTRY;
2380
2381         if (lfsck_is_dryrun(lfsck))
2382                 GOTO(log, rc = 0);
2383
2384         if (unlikely(lpf == NULL))
2385                 GOTO(log, rc = -ENXIO);
2386
2387         /* We use two separated transactions to repair the inconsistency.
2388          *
2389          * 1) create the MDT-object locally.
2390          * 2) update the OST-object's PFID EA if necessary.
2391          *
2392          * If 1) succeed, but 2) failed, then the OST-object's PFID EA will be
2393          * updated when the layout LFSCK run next time.
2394          *
2395          * If 1) failed, but 2) succeed, then such MDT-object will be re-created
2396          * when the layout LFSCK run next time. */
2397
2398         if (fid_is_zero(pfid)) {
2399                 rc = lfsck_fid_alloc(env, lfsck, pfid, false);
2400                 if (rc != 0)
2401                         GOTO(log, rc);
2402
2403                 cobj = lfsck_object_find_by_dev(env, ltd->ltd_tgt, cfid);
2404                 if (IS_ERR(cobj))
2405                         GOTO(log, rc = PTR_ERR(cobj));
2406         }
2407
2408         pobj = lfsck_object_find_by_dev(env, dev, pfid);
2409         if (IS_ERR(pobj))
2410                 GOTO(log, rc = PTR_ERR(pobj));
2411
2412         LASSERT(infix != NULL);
2413         LASSERT(type != NULL);
2414
2415         memset(la, 0, sizeof(*la));
2416         la->la_uid = rec->lor_rec.lor_uid;
2417         la->la_gid = rec->lor_rec.lor_gid;
2418         la->la_mode = S_IFREG | S_IRUSR;
2419         la->la_valid = LA_MODE | LA_UID | LA_GID;
2420
2421         memset(dof, 0, sizeof(*dof));
2422         dof->dof_type = dt_mode_to_dft(S_IFREG);
2423         /* Because the dof->dof_reg.striped = 0, the LOD will not create
2424          * the stripe(s). The LFSCK will specify the LOV EA via
2425          * lfsck_layout_update_lovea(). */
2426
2427         size = lfsck_lovea_size(&rec->lor_layout, ea_off);
2428         if (ea_buf->lb_len < size) {
2429                 lu_buf_realloc(ea_buf, size);
2430                 if (ea_buf->lb_buf == NULL)
2431                         GOTO(log, rc = -ENOMEM);
2432         }
2433
2434 again:
2435         do {
2436                 snprintf(name, NAME_MAX, DFID"%s-%s-%d", PFID(pfid), infix,
2437                          type, idx++);
2438                 rc = dt_lookup_dir(env, lfsck->li_lpf_obj, name, tfid);
2439                 if (rc != 0 && rc != -ENOENT)
2440                         GOTO(log, rc);
2441         } while (rc == 0);
2442
2443         rc = lfsck_lock(env, lfsck, lfsck->li_lpf_obj, name, llh,
2444                         MDS_INODELOCK_UPDATE, LCK_PW);
2445         if (rc != 0)
2446                 GOTO(log, rc);
2447
2448         /* Re-check whether the name conflict with othrs after taken
2449          * the ldlm lock. */
2450         rc = dt_lookup_dir(env, lfsck->li_lpf_obj, name, tfid);
2451         if (unlikely(rc == 0)) {
2452                 lfsck_unlock(llh);
2453                 goto again;
2454         }
2455
2456         if (rc != -ENOENT)
2457                 GOTO(unlock, rc);
2458
2459         pname = lfsck_name_get_const(env, name, strlen(name));
2460         rc = linkea_links_new(&ldata, &lfsck_env_info(env)->lti_linkea_buf,
2461                               pname, lfsck_dto2fid(lfsck->li_lpf_obj));
2462         if (rc != 0)
2463                 GOTO(unlock, rc);
2464
2465         /* The 1st transaction. */
2466         th = dt_trans_create(env, dev);
2467         if (IS_ERR(th))
2468                 GOTO(unlock, rc = PTR_ERR(th));
2469
2470         rc = dt_declare_create(env, pobj, la, NULL, dof, th);
2471         if (rc != 0)
2472                 GOTO(stop, rc);
2473
2474         lfsck_buf_init(&lov_buf, ea_buf->lb_buf, size);
2475         rc = dt_declare_xattr_set(env, pobj, &lov_buf, XATTR_NAME_LOV,
2476                                   LU_XATTR_CREATE, th);
2477         if (rc != 0)
2478                 GOTO(stop, rc);
2479
2480         dtrec->rec_fid = pfid;
2481         dtrec->rec_type = S_IFREG;
2482         rc = dt_declare_insert(env, lpf,
2483                                (const struct dt_rec *)dtrec,
2484                                (const struct dt_key *)name, th);
2485         if (rc != 0)
2486                 GOTO(stop, rc);
2487
2488         lfsck_buf_init(&linkea_buf, ldata.ld_buf->lb_buf,
2489                        ldata.ld_leh->leh_len);
2490         rc = dt_declare_xattr_set(env, pobj, &linkea_buf,
2491                                   XATTR_NAME_LINK, 0, th);
2492         if (rc != 0)
2493                 GOTO(stop, rc);
2494
2495         rc = dt_trans_start_local(env, dev, th);
2496         if (rc != 0)
2497                 GOTO(stop, rc);
2498
2499         dt_write_lock(env, pobj, 0);
2500         rc = dt_create(env, pobj, la, NULL, dof, th);
2501         if (rc == 0)
2502                 rc = lfsck_layout_update_lovea(env, lfsck, th, rec, pobj, cfid,
2503                         &lov_buf, LU_XATTR_CREATE, ltd->ltd_index, ea_off);
2504         dt_write_unlock(env, pobj);
2505         if (rc < 0)
2506                 GOTO(stop, rc);
2507
2508         rc = dt_insert(env, lpf, (const struct dt_rec *)dtrec,
2509                        (const struct dt_key *)name, th);
2510         if (rc != 0)
2511                 GOTO(stop, rc);
2512
2513         rc = dt_xattr_set(env, pobj, &linkea_buf, XATTR_NAME_LINK, 0, th);
2514         if (rc == 0 && cobj != NULL) {
2515                 dt_trans_stop(env, dev, th);
2516                 th = NULL;
2517
2518                 /* The 2nd transaction. */
2519                 rc = __lfsck_layout_update_pfid(env, cobj, pfid,
2520                                                 &rec->lor_layout, ea_off,
2521                                                 rec->lor_layout_version,
2522                                                 rec->lor_range);
2523         }
2524
2525         GOTO(stop, rc);
2526
2527 stop:
2528         if (th != NULL)
2529                 dt_trans_stop(env, dev, th);
2530
2531 unlock:
2532         lfsck_unlock(llh);
2533
2534 log:
2535         if (cobj != NULL && !IS_ERR(cobj))
2536                 lfsck_object_put(env, cobj);
2537         if (pobj != NULL && !IS_ERR(pobj))
2538                 lfsck_object_put(env, pobj);
2539
2540         if (rc < 0)
2541                 CDEBUG(D_LFSCK, "%s layout LFSCK assistant failed to "
2542                        "recreate the lost MDT-object: parent "DFID
2543                        ", child "DFID", OST-index %u, stripe-index %u, "
2544                        "infix %s, type %s: rc = %d\n",
2545                        lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
2546                        ltd->ltd_index, ea_off, infix, type, rc);
2547
2548         return rc >= 0 ? 1 : rc;
2549 }
2550
2551 static int lfsck_layout_master_conditional_destroy(const struct lu_env *env,
2552                                                    struct lfsck_component *com,
2553                                                    const struct lu_fid *fid,
2554                                                    __u32 index)
2555 {
2556         struct lfsck_thread_info *info  = lfsck_env_info(env);
2557         struct lfsck_request     *lr    = &info->lti_lr;
2558         struct lfsck_instance    *lfsck = com->lc_lfsck;
2559         struct lfsck_tgt_desc    *ltd;
2560         struct ptlrpc_request    *req;
2561         struct lfsck_request     *tmp;
2562         struct obd_export        *exp;
2563         int                       rc    = 0;
2564         ENTRY;
2565
2566         ltd = lfsck_tgt_get(&lfsck->li_ost_descs, index);
2567         if (unlikely(ltd == NULL))
2568                 RETURN(-ENXIO);
2569
2570         exp = ltd->ltd_exp;
2571         if (!(exp_connect_flags(exp) & OBD_CONNECT_LFSCK))
2572                 GOTO(put, rc = -EOPNOTSUPP);
2573
2574         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
2575         if (req == NULL)
2576                 GOTO(put, rc = -ENOMEM);
2577
2578         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
2579         if (rc != 0) {
2580                 ptlrpc_request_free(req);
2581
2582                 GOTO(put, rc);
2583         }
2584
2585         memset(lr, 0, sizeof(*lr));
2586         lr->lr_event = LE_CONDITIONAL_DESTROY;
2587         lr->lr_active = LFSCK_TYPE_LAYOUT;
2588         lr->lr_fid = *fid;
2589
2590         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
2591         *tmp = *lr;
2592         ptlrpc_request_set_replen(req);
2593
2594         rc = ptlrpc_queue_wait(req);
2595         ptlrpc_req_finished(req);
2596
2597         GOTO(put, rc);
2598
2599 put:
2600         lfsck_tgt_put(ltd);
2601
2602         return rc;
2603 }
2604
2605 static int lfsck_layout_slave_conditional_destroy(const struct lu_env *env,
2606                                                   struct lfsck_component *com,
2607                                                   struct lfsck_request *lr)
2608 {
2609         struct lfsck_thread_info        *info   = lfsck_env_info(env);
2610         struct lu_attr                  *la     = &info->lti_la;
2611         union ldlm_policy_data          *policy = &info->lti_policy;
2612         struct ldlm_res_id              *resid  = &info->lti_resid;
2613         struct lfsck_instance           *lfsck  = com->lc_lfsck;
2614         struct dt_device                *dev    = lfsck->li_bottom;
2615         struct lu_fid                   *fid    = &lr->lr_fid;
2616         struct dt_object                *obj;
2617         struct thandle                  *th     = NULL;
2618         struct lustre_handle             lh     = { 0 };
2619         __u64                            flags  = 0;
2620         int                              rc     = 0;
2621         ENTRY;
2622
2623         obj = lfsck_object_find_by_dev(env, dev, fid);
2624         if (IS_ERR(obj))
2625                 RETURN(PTR_ERR(obj));
2626
2627         dt_read_lock(env, obj, 0);
2628         if (dt_object_exists(obj) == 0 ||
2629             lfsck_is_dead_obj(obj)) {
2630                 dt_read_unlock(env, obj);
2631
2632                 GOTO(put, rc = -ENOENT);
2633         }
2634
2635         /* Get obj's attr without lock firstly. */
2636         rc = dt_attr_get(env, obj, la);
2637         dt_read_unlock(env, obj);
2638         if (rc != 0)
2639                 GOTO(put, rc);
2640
2641         if (likely(la->la_ctime != 0 || la->la_mode & S_ISUID))
2642                 GOTO(put, rc = -ETXTBSY);
2643
2644         /* Acquire extent lock on [0, EOF] to sync with all possible written. */
2645         LASSERT(lfsck->li_namespace != NULL);
2646
2647         memset(policy, 0, sizeof(*policy));
2648         policy->l_extent.end = OBD_OBJECT_EOF;
2649         ost_fid_build_resid(fid, resid);
2650         rc = ldlm_cli_enqueue_local(env, lfsck->li_namespace, resid,
2651                                     LDLM_EXTENT, policy, LCK_EX, &flags,
2652                                     ldlm_blocking_ast, ldlm_completion_ast,
2653                                     NULL, NULL, 0, LVB_T_NONE, NULL, &lh);
2654         if (rc != ELDLM_OK)
2655                 GOTO(put, rc = -EIO);
2656
2657         dt_write_lock(env, obj, 0);
2658         /* Get obj's attr within lock again. */
2659         rc = dt_attr_get(env, obj, la);
2660         if (rc != 0)
2661                 GOTO(unlock, rc);
2662
2663         if (la->la_ctime != 0)
2664                 GOTO(unlock, rc = -ETXTBSY);
2665
2666         th = dt_trans_create(env, dev);
2667         if (IS_ERR(th))
2668                 GOTO(unlock, rc = PTR_ERR(th));
2669
2670         rc = dt_declare_ref_del(env, obj, th);
2671         if (rc != 0)
2672                 GOTO(stop, rc);
2673
2674         rc = dt_declare_destroy(env, obj, th);
2675         if (rc != 0)
2676                 GOTO(stop, rc);
2677
2678         rc = dt_trans_start_local(env, dev, th);
2679         if (rc != 0)
2680                 GOTO(stop, rc);
2681
2682         rc = dt_ref_del(env, obj, th);
2683         if (rc != 0)
2684                 GOTO(stop, rc);
2685
2686         rc = dt_destroy(env, obj, th);
2687         if (rc == 0)
2688                 CDEBUG(D_LFSCK, "%s: layout LFSCK destroyed the empty "
2689                        "OST-object "DFID" that was created for reparing "
2690                        "dangling referenced case. But the original missing "
2691                        "OST-object is found now.\n",
2692                        lfsck_lfsck2name(lfsck), PFID(fid));
2693
2694         GOTO(stop, rc);
2695
2696 stop:
2697         dt_trans_stop(env, dev, th);
2698
2699 unlock:
2700         dt_write_unlock(env, obj);
2701         ldlm_lock_decref(&lh, LCK_EX);
2702
2703 put:
2704         lfsck_object_put(env, obj);
2705
2706         return rc;
2707 }
2708
2709 /**
2710  * Some OST-object has occupied the specified layout EA slot.
2711  * Such OST-object may be generated by the LFSCK when repair
2712  * dangling referenced MDT-object, which can be indicated by
2713  * attr::la_ctime == 0 but without S_ISUID in la_mode. If it
2714  * is true and such OST-object has not been modified yet, we
2715  * will replace it with the orphan OST-object; otherwise the
2716  * LFSCK will create new MDT-object to reference the orphan.
2717  *
2718  * \retval       +1: repaired
2719  * \retval        0: did nothing
2720  * \retval      -ve: on error
2721  */
2722 static int lfsck_layout_conflict_create(const struct lu_env *env,
2723                                         struct lfsck_component *com,
2724                                         struct lfsck_tgt_desc *ltd,
2725                                         struct lu_orphan_rec_v3 *rec,
2726                                         struct dt_object *parent,
2727                                         struct lu_fid *cfid,
2728                                         struct lu_buf *ea_buf,
2729                                         struct lov_mds_md_v1 *lmm,
2730                                         struct lov_ost_data_v1 *slot,
2731                                         __u32 ea_off, int lovea_size)
2732 {
2733         struct lfsck_thread_info *info          = lfsck_env_info(env);
2734         struct lu_fid            *cfid2         = &info->lti_fid2;
2735         struct ost_id            *oi            = &info->lti_oi;
2736         struct dt_device         *dev           = lfsck_obj2dev(parent);
2737         struct thandle           *th            = NULL;
2738         struct lustre_handle      lh            = { 0 };
2739         __u32                     ost_idx2      = le32_to_cpu(slot->l_ost_idx);
2740         int                       rc            = 0;
2741         ENTRY;
2742
2743         while (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val)) {
2744                 if (unlikely(!thread_is_running(&com->lc_lfsck->li_thread)))
2745                         RETURN(0);
2746         }
2747
2748         ostid_le_to_cpu(&slot->l_ost_oi, oi);
2749         rc = ostid_to_fid(cfid2, oi, ost_idx2);
2750         if (rc != 0)
2751                 GOTO(out, rc);
2752
2753         rc = lfsck_ibits_lock(env, com->lc_lfsck, parent, &lh,
2754                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
2755                               LCK_EX);
2756         if (rc != 0)
2757                 GOTO(out, rc);
2758
2759         rc = lfsck_layout_master_conditional_destroy(env, com, cfid2, ost_idx2);
2760
2761         /* If the conflict OST-obejct is not created for fixing dangling
2762          * referenced MDT-object in former LFSCK check/repair, or it has
2763          * been modified by others, then we cannot destroy it. Re-create
2764          * a new MDT-object for the orphan OST-object. */
2765         if (rc == -ETXTBSY) {
2766                 /* No need the layout lock on the original parent. */
2767                 lfsck_ibits_unlock(&lh, LCK_EX);
2768
2769                 fid_zero(&rec->lor_rec.lor_fid);
2770                 snprintf(info->lti_tmpbuf, sizeof(info->lti_tmpbuf),
2771                          "-"DFID"-%x", PFID(lu_object_fid(&parent->do_lu)),
2772                          ea_off);
2773                 rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
2774                                                 info->lti_tmpbuf, "C", ea_off);
2775
2776                 RETURN(rc);
2777         }
2778
2779         if (rc != 0 && rc != -ENOENT)
2780                 GOTO(unlock, rc);
2781
2782         if (lfsck_is_dryrun(com->lc_lfsck))
2783                 GOTO(unlock, rc = 0);
2784
2785         th = dt_trans_create(env, dev);
2786         if (IS_ERR(th))
2787                 GOTO(unlock, rc = PTR_ERR(th));
2788
2789         rc = dt_declare_xattr_set(env, parent, ea_buf, XATTR_NAME_LOV,
2790                                   LU_XATTR_REPLACE, th);
2791         if (rc != 0)
2792                 GOTO(stop, rc);
2793
2794         rc = dt_trans_start_local(env, dev, th);
2795         if (rc != 0)
2796                 GOTO(stop, rc);
2797
2798         dt_write_lock(env, parent, 0);
2799         lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
2800         rc = lfsck_layout_refill_lovea(env, com->lc_lfsck, th, parent, cfid,
2801                                        ea_buf, lmm, slot, LU_XATTR_REPLACE,
2802                                        ltd->ltd_index, lovea_size);
2803         dt_write_unlock(env, parent);
2804
2805         GOTO(stop, rc);
2806
2807 stop:
2808         dt_trans_stop(env, dev, th);
2809
2810 unlock:
2811         lfsck_ibits_unlock(&lh, LCK_EX);
2812
2813 out:
2814         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant replaced the conflict "
2815                "OST-object "DFID" on the OST %x with the orphan "DFID" on "
2816                "the OST %x: parent "DFID", stripe-index %u: rc = %d\n",
2817                lfsck_lfsck2name(com->lc_lfsck), PFID(cfid2), ost_idx2,
2818                PFID(cfid), ltd->ltd_index, PFID(lfsck_dto2fid(parent)),
2819                ea_off, rc);
2820
2821         return rc >= 0 ? 1 : rc;
2822 }
2823
2824 /**
2825  * \retval       +1: repaired
2826  * \retval        0: did nothing
2827  * \retval      -ve: on error
2828  */
2829 static int lfsck_layout_recreate_lovea(const struct lu_env *env,
2830                                        struct lfsck_component *com,
2831                                        struct lfsck_tgt_desc *ltd,
2832                                        struct lu_orphan_rec_v3 *rec,
2833                                        struct dt_object *parent,
2834                                        struct lu_fid *cfid,
2835                                        __u32 ost_idx, __u32 ea_off)
2836 {
2837         struct lfsck_thread_info *info          = lfsck_env_info(env);
2838         struct lu_buf            *buf           = &info->lti_big_buf;
2839         struct lu_fid            *fid           = &info->lti_fid2;
2840         struct ost_id            *oi            = &info->lti_oi;
2841         struct lfsck_instance    *lfsck         = com->lc_lfsck;
2842         struct dt_device         *dt            = lfsck_obj2dev(parent);
2843         struct lfsck_bookmark    *bk            = &lfsck->li_bookmark_ram;
2844         struct ost_layout        *ol            = &rec->lor_layout;
2845         struct lov_comp_md_v1    *lcm           = NULL;
2846         struct lov_comp_md_entry_v1 *lcme       = NULL;
2847         struct thandle           *handle        = NULL;
2848         size_t                    lovea_size;
2849         struct lov_mds_md_v1     *lmm;
2850         struct lov_ost_data_v1   *objs;
2851         struct lustre_handle      lh            = { 0 };
2852         __u32                     magic;
2853         __u32 flags = 0;
2854         int                       fl            = 0;
2855         int                       rc            = 0;
2856         int                       rc1;
2857         int                       i;
2858         int pos = 0;
2859         __u16 count;
2860         bool locked = false;
2861         bool new_mirror = true;
2862         ENTRY;
2863
2864         if (lfsck_is_dryrun(lfsck))
2865                 RETURN(0);
2866
2867         rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
2868                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
2869                               LCK_EX);
2870         if (rc != 0) {
2871                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant failed to recreate "
2872                        "LOV EA for "DFID": parent "DFID", OST-index %u, "
2873                        "stripe-index %u, comp_id %u, comp_start %llu, "
2874                        "comp_end %llu, layout version %u, range %u: rc = %d\n",
2875                        lfsck_lfsck2name(lfsck), PFID(cfid),
2876                        PFID(lfsck_dto2fid(parent)), ost_idx, ea_off,
2877                        ol->ol_comp_id, ol->ol_comp_start,
2878                        ol->ol_comp_end, rec->lor_layout_version,
2879                        rec->lor_range, rc);
2880
2881                 RETURN(rc);
2882         }
2883
2884 again:
2885         if (locked) {
2886                 dt_write_unlock(env, parent);
2887                 locked = false;
2888         }
2889
2890         if (handle != NULL) {
2891                 dt_trans_stop(env, dt, handle);
2892                 handle = NULL;
2893         }
2894
2895         if (rc < 0)
2896                 GOTO(unlock_layout, rc);
2897
2898         lovea_size = rc;
2899         if (buf->lb_len < lovea_size) {
2900                 lu_buf_realloc(buf, lovea_size);
2901                 if (buf->lb_buf == NULL)
2902                         GOTO(unlock_layout, rc = -ENOMEM);
2903         }
2904
2905         if (!(bk->lb_param & LPF_DRYRUN)) {
2906                 handle = dt_trans_create(env, dt);
2907                 if (IS_ERR(handle))
2908                         GOTO(unlock_layout, rc = PTR_ERR(handle));
2909
2910                 rc = dt_declare_xattr_set(env, parent, buf, XATTR_NAME_LOV,
2911                                           fl, handle);
2912                 if (rc != 0)
2913                         GOTO(stop, rc);
2914
2915                 rc = dt_trans_start_local(env, dt, handle);
2916                 if (rc != 0)
2917                         GOTO(stop, rc);
2918         }
2919
2920         dt_write_lock(env, parent, 0);
2921         locked = true;
2922         rc = dt_xattr_get(env, parent, buf, XATTR_NAME_LOV);
2923         if (rc == -ERANGE) {
2924                 rc = dt_xattr_get(env, parent, &LU_BUF_NULL, XATTR_NAME_LOV);
2925                 LASSERT(rc != 0);
2926                 goto again;
2927         } else if (rc == -ENODATA || rc == 0) {
2928                 lovea_size = lfsck_lovea_size(ol, ea_off);
2929                 /* If the declared is not big enough, re-try. */
2930                 if (buf->lb_len < lovea_size) {
2931                         rc = lovea_size;
2932                         goto again;
2933                 }
2934                 fl = LU_XATTR_CREATE;
2935         } else if (rc < 0) {
2936                 GOTO(unlock_parent, rc);
2937         } else if (unlikely(buf->lb_len == 0)) {
2938                 goto again;
2939         } else {
2940                 fl = LU_XATTR_REPLACE;
2941                 lovea_size = rc;
2942         }
2943
2944         if (fl == LU_XATTR_CREATE) {
2945                 if (bk->lb_param & LPF_DRYRUN)
2946                         GOTO(unlock_parent, rc = 1);
2947
2948                 LASSERT(buf->lb_len >= lovea_size);
2949
2950                 rc = lfsck_layout_update_lovea(env, lfsck, handle, rec, parent,
2951                                                cfid, buf, fl, ost_idx, ea_off);
2952
2953                 GOTO(unlock_parent, rc);
2954         }
2955
2956         lmm = buf->lb_buf;
2957         rc1 = lfsck_layout_verify_header(parent, lmm, lovea_size);
2958
2959         /* If the LOV EA crashed, the rebuild it. */
2960         if (rc1 == -EINVAL) {
2961                 if (bk->lb_param & LPF_DRYRUN)
2962                         GOTO(unlock_parent, rc = 1);
2963
2964                 LASSERT(buf->lb_len >= lovea_size);
2965
2966                 rc = lfsck_layout_update_lovea(env, lfsck, handle, rec, parent,
2967                                                cfid, buf, fl, ost_idx, ea_off);
2968
2969                 GOTO(unlock_parent, rc);
2970         }
2971
2972         /* For other unknown magic/pattern, keep the current LOV EA. */
2973         if (rc1 == -EOPNOTSUPP)
2974                 GOTO(unlock_parent, rc1 = 0);
2975
2976         if (rc1)
2977                 GOTO(unlock_parent, rc = rc1);
2978
2979         magic = le32_to_cpu(lmm->lmm_magic);
2980         if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
2981                 __u64 start;
2982                 __u64 end;
2983                 __u16 mirror_id0 = mirror_id_of(ol->ol_comp_id);
2984                 __u16 mirror_id1;
2985
2986                 if (bk->lb_param & LPF_DRYRUN)
2987                         GOTO(unlock_parent, rc = 1);
2988
2989                 lcm = buf->lb_buf;
2990                 count = le16_to_cpu(lcm->lcm_entry_count);
2991                 for (i = 0; i < count; pos = ++i) {
2992                         lcme = &lcm->lcm_entries[i];
2993                         start = le64_to_cpu(lcme->lcme_extent.e_start);
2994                         end = le64_to_cpu(lcme->lcme_extent.e_end);
2995                         mirror_id1 = mirror_id_of(le32_to_cpu(lcme->lcme_id));
2996
2997                         if (mirror_id0 > mirror_id1)
2998                                 continue;
2999
3000                         if (mirror_id0 < mirror_id1)
3001                                 break;
3002
3003                         new_mirror = false;
3004                         if (end <= ol->ol_comp_start)
3005                                 continue;
3006
3007                         if (start >= ol->ol_comp_end)
3008                                 break;
3009
3010                         lmm = buf->lb_buf + le32_to_cpu(lcme->lcme_offset);
3011                         magic = le32_to_cpu(lmm->lmm_magic);
3012                         flags = le32_to_cpu(lcme->lcme_flags);
3013                         goto further;
3014                 }
3015
3016                 rc = lfsck_layout_add_comp(env, lfsck, handle, rec, parent,
3017                                 cfid, buf, ost_idx, ea_off, pos, new_mirror);
3018
3019                 GOTO(unlock_parent, rc);
3020         }
3021
3022 further:
3023         count = le16_to_cpu(lmm->lmm_stripe_count);
3024         if (count == 0)
3025                 GOTO(unlock_parent, rc = -EINVAL);
3026         LASSERT(count > 0);
3027
3028         /* Exceed the current end of MDT-object layout EA. Then extend it. */
3029         if (count <= ea_off) {
3030                 if (bk->lb_param & LPF_DRYRUN)
3031                         GOTO(unlock_parent, rc = 1);
3032
3033                 lovea_size = lov_mds_md_size(ea_off + 1, magic);
3034                 /* If the declared is not big enough, re-try. */
3035                 if (buf->lb_len < lovea_size) {
3036                         rc = lovea_size;
3037                         goto again;
3038                 }
3039
3040                 if (lcm) {
3041                         LASSERT(lcme);
3042
3043                         lcme->lcme_flags = cpu_to_le32(flags | LCME_FL_INIT);
3044                         lfsck_layout_update_lcm(lcm, lcme,
3045                                                 rec->lor_layout_version,
3046                                                 rec->lor_range);
3047                 }
3048
3049                 rc = lfsck_layout_extend_v1v3_lovea(env, lfsck, handle, ol,
3050                                         parent, cfid, buf, ost_idx, ea_off);
3051
3052                 GOTO(unlock_parent, rc);
3053         }
3054
3055         LASSERTF(rc > 0, "invalid rc = %d\n", rc);
3056
3057         if (magic == LOV_MAGIC_V1) {
3058                 objs = &lmm->lmm_objects[0];
3059         } else {
3060                 LASSERT(magic == LOV_MAGIC_V3);
3061                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
3062         }
3063
3064         for (i = 0; i < count; i++, objs++) {
3065                 /* The MDT-object was created via lfsck_layout_recover_create()
3066                  * by others before, and we fill the dummy layout EA. */
3067                 if ((lcme && !(flags & LCME_FL_INIT)) ||
3068                      lovea_slot_is_dummy(objs)) {
3069                         if (i != ea_off)
3070                                 continue;
3071
3072                         if (bk->lb_param & LPF_DRYRUN)
3073                                 GOTO(unlock_parent, rc = 1);
3074
3075                         lmm->lmm_layout_gen =
3076                             cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
3077                         if (lcme) {
3078                                 LASSERT(lcm);
3079
3080                                 if (le32_to_cpu(lmm->lmm_stripe_size) !=
3081                                         ol->ol_stripe_size ||
3082                                     le16_to_cpu(lmm->lmm_stripe_count) !=
3083                                         ol->ol_stripe_count ||
3084                                     le64_to_cpu(lcme->lcme_extent.e_start) !=
3085                                         ol->ol_comp_start ||
3086                                     le64_to_cpu(lcme->lcme_extent.e_end) !=
3087                                         ol->ol_comp_end) {
3088                                         CDEBUG(D_LFSCK, "%s: found invalid "
3089                                         "component for "DFID ": parent "DFID
3090                                         ", stripe-index %u, stripe_size %u, "
3091                                         "stripe_count %u, comp_id %u, "
3092                                         "comp_start %llu, comp_end %llu, "
3093                                         "cur_stripe_size %u, "
3094                                         "cur_stripe_count %u, "
3095                                         "cur_comp_start %llu, "
3096                                         "cur_comp_end %llu\n",
3097                                         lfsck_lfsck2name(lfsck), PFID(cfid),
3098                                         PFID(lfsck_dto2fid(parent)), ea_off,
3099                                         ol->ol_stripe_size,
3100                                         ol->ol_stripe_count, ol->ol_comp_id,
3101                                         ol->ol_comp_start, ol->ol_comp_end,
3102                                         le32_to_cpu(lmm->lmm_stripe_size),
3103                                         le16_to_cpu(lmm->lmm_stripe_count),
3104                                         le64_to_cpu(lcme->lcme_extent.e_start),
3105                                         le64_to_cpu(lcme->lcme_extent.e_end));
3106
3107                                         GOTO(unlock_parent, rc = -EINVAL);
3108                                 }
3109
3110                                 lovea_size = le32_to_cpu(lcm->lcm_size);
3111                                 lcme->lcme_flags = cpu_to_le32(flags |
3112                                                                LCME_FL_INIT);
3113                                 lfsck_layout_update_lcm(lcm, lcme,
3114                                                         rec->lor_layout_version,
3115                                                         rec->lor_range);
3116                         }
3117
3118                         LASSERTF(buf->lb_len >= lovea_size,
3119                                  "buffer len %d is less than real size %d\n",
3120                                  (int)buf->lb_len, (int)lovea_size);
3121
3122                         rc = lfsck_layout_refill_lovea(env, lfsck, handle,
3123                                                 parent, cfid, buf, lmm, objs,
3124                                                 fl, ost_idx, lovea_size);
3125
3126                         CDEBUG(D_LFSCK, "%s layout LFSCK assistant fill "
3127                                "dummy layout slot for "DFID": parent "DFID
3128                                ", OST-index %u, stripe-index %u: rc = %d\n",
3129                                lfsck_lfsck2name(lfsck), PFID(cfid),
3130                                PFID(lfsck_dto2fid(parent)), ost_idx, i, rc);
3131
3132                         GOTO(unlock_parent, rc);
3133                 }
3134
3135                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
3136                 rc = ostid_to_fid(fid, oi, le32_to_cpu(objs->l_ost_idx));
3137                 if (rc != 0) {
3138                         CDEBUG(D_LFSCK, "%s: the parent "DFID" contains "
3139                                "invalid layout EA at the slot %d, index %u\n",
3140                                lfsck_lfsck2name(lfsck),
3141                                PFID(lfsck_dto2fid(parent)), i,
3142                                le32_to_cpu(objs->l_ost_idx));
3143
3144                         GOTO(unlock_parent, rc);
3145                 }
3146
3147                 /* It should be rare case, the slot is there, but the LFSCK
3148                  * does not handle it during the first-phase cycle scanning. */
3149                 if (unlikely(lu_fid_eq(fid, cfid))) {
3150                         if (i == ea_off) {
3151                                 GOTO(unlock_parent, rc = 0);
3152                         } else {
3153                                 /* Rare case that the OST-object index
3154                                  * does not match the parent MDT-object
3155                                  * layout EA. We trust the later one. */
3156                                 if (bk->lb_param & LPF_DRYRUN)
3157                                         GOTO(unlock_parent, rc = 1);
3158
3159                                 dt_write_unlock(env, parent);
3160                                 if (handle != NULL)
3161                                         dt_trans_stop(env, dt, handle);
3162                                 lfsck_ibits_unlock(&lh, LCK_EX);
3163                                 rc = lfsck_layout_update_pfid(env, com, parent,
3164                                                         cfid, ltd->ltd_tgt,
3165                                                         rec, i);
3166
3167                                 CDEBUG(D_LFSCK, "%s layout LFSCK assistant "
3168                                        "updated OST-object's pfid for "DFID
3169                                        ": parent "DFID", OST-index %u, "
3170                                        "stripe-index %u: rc = %d\n",
3171                                        lfsck_lfsck2name(lfsck), PFID(cfid),
3172                                        PFID(lfsck_dto2fid(parent)),
3173                                        ltd->ltd_index, i, rc);
3174
3175                                 RETURN(rc);
3176                         }
3177                 }
3178         }
3179
3180         /* The MDT-object exists, but related layout EA slot is occupied
3181          * by others. */
3182         if (bk->lb_param & LPF_DRYRUN)
3183                 GOTO(unlock_parent, rc = 1);
3184
3185         dt_write_unlock(env, parent);
3186         if (handle != NULL)
3187                 dt_trans_stop(env, dt, handle);
3188         lfsck_ibits_unlock(&lh, LCK_EX);
3189         if (magic == LOV_MAGIC_V1)
3190                 objs = &lmm->lmm_objects[ea_off];
3191         else
3192                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[ea_off];
3193         rc = lfsck_layout_conflict_create(env, com, ltd, rec, parent, cfid,
3194                                           buf, lmm, objs, ea_off, lovea_size);
3195
3196         RETURN(rc);
3197
3198 unlock_parent:
3199         if (locked)
3200                 dt_write_unlock(env, parent);
3201
3202 stop:
3203         if (handle != NULL)
3204                 dt_trans_stop(env, dt, handle);
3205
3206 unlock_layout:
3207         lfsck_ibits_unlock(&lh, LCK_EX);
3208
3209         return rc;
3210 }
3211
3212 static int lfsck_layout_scan_orphan_one(const struct lu_env *env,
3213                                         struct lfsck_component *com,
3214                                         struct lfsck_tgt_desc *ltd,
3215                                         struct lu_orphan_rec_v3 *rec,
3216                                         struct lu_fid *cfid)
3217 {
3218         struct lfsck_layout     *lo     = com->lc_file_ram;
3219         struct lu_fid           *pfid   = &rec->lor_rec.lor_fid;
3220         struct dt_object        *parent = NULL;
3221         __u32                    ea_off = pfid->f_stripe_idx;
3222         int                      rc     = 0;
3223         ENTRY;
3224
3225         if (!fid_is_sane(cfid))
3226                 GOTO(out, rc = -EINVAL);
3227
3228         pfid->f_ver = 0;
3229         if (fid_is_zero(pfid)) {
3230                 rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
3231                                                   "", "N", ea_off);
3232                 GOTO(out, rc);
3233         }
3234
3235         if (!fid_is_sane(pfid))
3236                 GOTO(out, rc = -EINVAL);
3237
3238         parent = lfsck_object_find_by_dev(env, com->lc_lfsck->li_bottom, pfid);
3239         if (IS_ERR(parent))
3240                 GOTO(out, rc = PTR_ERR(parent));
3241
3242         if (unlikely(dt_object_remote(parent) != 0))
3243                 GOTO(put, rc = -EXDEV);
3244
3245         if (dt_object_exists(parent) == 0) {
3246                 lfsck_object_put(env, parent);
3247                 rc = lfsck_layout_recreate_parent(env, com, ltd, rec, cfid,
3248                                                   "", "R", ea_off);
3249                 GOTO(out, rc);
3250         }
3251
3252         if (!S_ISREG(lu_object_attr(&parent->do_lu)))
3253                 GOTO(put, rc = -EISDIR);
3254
3255         /* The orphan OST-object claims to be the parent's stripe, then
3256          * related dangling record in the trace file is meaningless. */
3257         rc = lfsck_layout_del_dangling_rec(env, com, pfid,
3258                                            rec->lor_layout.ol_comp_id, ea_off);
3259         if (rc && rc != -ENOENT)
3260                 GOTO(put, rc);
3261
3262         rc = lfsck_layout_recreate_lovea(env, com, ltd, rec, parent, cfid,
3263                                          ltd->ltd_index, ea_off);
3264
3265         GOTO(put, rc);
3266
3267 put:
3268         if (rc <= 0)
3269                 lfsck_object_put(env, parent);
3270         else
3271                 /* The layout EA is changed, need to be reloaded next time. */
3272                 dt_object_put_nocache(env, parent);
3273
3274 out:
3275         down_write(&com->lc_sem);
3276         com->lc_new_scanned++;
3277         com->lc_new_checked++;
3278         if (rc > 0) {
3279                 lo->ll_objs_repaired[LLIT_ORPHAN - 1]++;
3280                 rc = 0;
3281         } else if (rc < 0) {
3282                 lo->ll_objs_failed_phase2++;
3283         }
3284         up_write(&com->lc_sem);
3285
3286         return rc;
3287 }
3288
3289 static int lfsck_layout_scan_orphan(const struct lu_env *env,
3290                                     struct lfsck_component *com,
3291                                     struct lfsck_tgt_desc *ltd)
3292 {
3293         struct lfsck_assistant_data     *lad    = com->lc_data;
3294         struct lfsck_instance           *lfsck  = com->lc_lfsck;
3295         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
3296         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3297         struct lu_fid                   *fid    = &info->lti_fid;
3298         struct dt_object                *obj;
3299         const struct dt_it_ops          *iops;
3300         struct dt_it                    *di;
3301         int                              rc     = 0;
3302         ENTRY;
3303
3304         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant starts the orphan "
3305                "scanning for OST%04x\n",
3306                lfsck_lfsck2name(lfsck), ltd->ltd_index);
3307
3308         if (cfs_bitmap_check(lad->lad_bitmap, ltd->ltd_index)) {
3309                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant skip the orphan "
3310                        "scanning for OST%04x\n",
3311                        lfsck_lfsck2name(lfsck), ltd->ltd_index);
3312
3313                 RETURN(0);
3314         }
3315
3316         fid->f_seq = fid_idif_seq(0, ltd->ltd_index);
3317         fid->f_oid = fid->f_ver = 0;
3318
3319         obj = lfsck_object_find_by_dev(env, ltd->ltd_tgt, fid);
3320         if (unlikely(IS_ERR(obj)))
3321                 GOTO(log, rc = PTR_ERR(obj));
3322
3323         rc = obj->do_ops->do_index_try(env, obj,
3324                                        &dt_lfsck_layout_orphan_features);
3325         if (rc != 0)
3326                 GOTO(put, rc);
3327
3328         iops = &obj->do_index_ops->dio_it;
3329         di = iops->init(env, obj, 0);
3330         if (IS_ERR(di))
3331                 GOTO(put, rc = PTR_ERR(di));
3332
3333         rc = iops->load(env, di, 0);
3334         if (rc == -ESRCH) {
3335                 /* -ESRCH means that the orphan OST-objects rbtree has been
3336                  * cleanup because of the OSS server restart or other errors. */
3337                 lfsck_lad_set_bitmap(env, com, ltd->ltd_index);
3338                 GOTO(fini, rc);
3339         }
3340
3341         if (rc == 0)
3342                 rc = iops->next(env, di);
3343         else if (rc > 0)
3344                 rc = 0;
3345
3346         if (rc < 0)
3347                 GOTO(fini, rc);
3348
3349         if (rc > 0)
3350                 GOTO(fini, rc = 0);
3351
3352         do {
3353                 struct dt_key           *key;
3354                 struct lu_orphan_rec_v3 *rec = &info->lti_rec;
3355
3356                 if (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val) &&
3357                     unlikely(!thread_is_running(&lfsck->li_thread)))
3358                         break;
3359
3360                 key = iops->key(env, di);
3361                 com->lc_fid_latest_scanned_phase2 = *(struct lu_fid *)key;
3362                 /* Remote target OST may be runnning old LFSCK */
3363                 memset(rec, 0, sizeof(*rec));
3364                 rc = iops->rec(env, di, (struct dt_rec *)rec, 0);
3365                 if (rc == 0)
3366                         rc = lfsck_layout_scan_orphan_one(env, com, ltd, rec,
3367                                         &com->lc_fid_latest_scanned_phase2);
3368                 if (rc != 0 && bk->lb_param & LPF_FAILOUT)
3369                         GOTO(fini, rc);
3370
3371                 lfsck_control_speed_by_self(com);
3372                 do {
3373                         rc = iops->next(env, di);
3374                 } while (rc < 0 && !(bk->lb_param & LPF_FAILOUT));
3375         } while (rc == 0);
3376
3377         GOTO(fini, rc);
3378
3379 fini:
3380         iops->put(env, di);
3381         iops->fini(env, di);
3382 put:
3383         lfsck_object_put(env, obj);
3384
3385 log:
3386         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant finished the orphan "
3387                "scanning for OST%04x: rc = %d\n",
3388                lfsck_lfsck2name(lfsck), ltd->ltd_index, rc);
3389
3390         return rc > 0 ? 0 : rc;
3391 }
3392
3393 static int lfsck_lov2layout(struct lov_mds_md_v1 *lmm, struct filter_fid *ff,
3394                             __u32 comp_id)
3395 {
3396         struct ost_layout *ol = &ff->ff_layout;
3397         __u32 magic = le32_to_cpu(lmm->lmm_magic);
3398         int rc = 0;
3399         ENTRY;
3400
3401         if (magic == LOV_MAGIC_V1 || magic == LOV_MAGIC_V3) {
3402                 ol->ol_stripe_size = lmm->lmm_stripe_size;
3403                 ol->ol_stripe_count = lmm->lmm_stripe_count;
3404                 ol->ol_comp_start = 0;
3405                 ol->ol_comp_end = 0;
3406                 ol->ol_comp_id = 0;
3407                 ff->ff_layout_version = 0;
3408                 ff->ff_range = 0;
3409         } else if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
3410                 struct lov_comp_md_v1 *lcm = (struct lov_comp_md_v1 *)lmm;
3411                 struct lov_comp_md_entry_v1 *lcme = NULL;
3412                 __u16 count = le16_to_cpu(lcm->lcm_entry_count);
3413                 int i;
3414
3415                 for (i = 0; i < count; i++) {
3416                         lcme = &lcm->lcm_entries[i];
3417                         if (le32_to_cpu(lcme->lcme_id) == comp_id) {
3418                                 LASSERT(le32_to_cpu(lcme->lcme_flags) &
3419                                         LCME_FL_INIT);
3420
3421                                 break;
3422                         }
3423                 }
3424
3425                 /* The comp has been removed, do nothing. */
3426                 if (i == count)
3427                         GOTO(out, rc = 1);
3428
3429                 lmm = (void *)lmm + le32_to_cpu(lcme->lcme_offset);
3430                 ol->ol_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
3431                 ol->ol_stripe_count = le32_to_cpu(lmm->lmm_stripe_count);
3432                 ol->ol_comp_start = le64_to_cpu(lcme->lcme_extent.e_start);
3433                 ol->ol_comp_end = le64_to_cpu(lcme->lcme_extent.e_end);
3434                 ol->ol_comp_id = le32_to_cpu(lcme->lcme_id);
3435                 ff->ff_layout_version = le32_to_cpu(lcme->lcme_layout_gen);
3436                 ff->ff_range = 0;
3437         } else {
3438                 GOTO(out, rc = -EINVAL);
3439         }
3440
3441         EXIT;
3442
3443 out:
3444         return rc;
3445 }
3446
3447 /**
3448  * Repair the MDT-object with dangling LOV EA reference.
3449  *
3450  * we need to repair the inconsistency according to the users' requirement:
3451  *
3452  * 1) Keep the inconsistency there and report the inconsistency case,
3453  *    then give the chance to the application to find related issues,
3454  *    and the users can make the decision about how to handle it with
3455  *    more human knownledge. (by default)
3456  *
3457  * 2) Re-create the missing OST-object with the FID/owner information.
3458  *
3459  * \param[in] env       pointer to the thread context
3460  * \param[in] com       the layout LFSCK component
3461  * \param[in] parent    the MDT-object with dangling LOV EA reference
3462  * \param[in] child     the OST-object to be created
3463  * \param[in] comp_id   the component ID of the OST-object in the LOV EA
3464  * \param[in] ea_off    the offset of the OST-object in the LOV EA
3465  * \param[in] ost_idx   the index of OST on which the OST-object resides
3466  *
3467  * \retval              +1 for repair successfully
3468  * \retval              0 for did nothing
3469  * \retval              negative error number on failure
3470  */
3471 static int __lfsck_layout_repair_dangling(const struct lu_env *env,
3472                                           struct lfsck_component *com,
3473                                           struct dt_object *parent,
3474                                           struct dt_object *child,
3475                                           __u32 comp_id, __u32 ea_off,
3476                                           __u32 ost_idx, bool log)
3477 {
3478         struct lfsck_thread_info *info = lfsck_env_info(env);
3479         struct filter_fid *ff = &info->lti_ff;
3480         struct dt_object_format *dof = &info->lti_dof;
3481         struct lu_attr *la = &info->lti_la;
3482         struct lfsck_instance *lfsck = com->lc_lfsck;
3483         struct dt_device *dev = lfsck_obj2dev(child);
3484         const struct lu_fid *pfid = lfsck_dto2fid(parent);
3485         const struct lu_fid *cfid = lfsck_dto2fid(child);
3486         struct lu_buf *tbuf = &info->lti_big_buf;
3487         struct thandle *handle;
3488         struct lu_buf *buf;
3489         struct lustre_handle lh = { 0 };
3490         int rc;
3491         ENTRY;
3492
3493         if (!(lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ))
3494                 GOTO(log, rc = 1);
3495
3496         rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
3497                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
3498                               LCK_EX);
3499         if (rc != 0)
3500                 GOTO(log, rc);
3501
3502         rc = dt_attr_get(env, parent, la);
3503         if (rc != 0)
3504                 GOTO(unlock1, rc);
3505
3506         la->la_mode = S_IFREG | 0666;
3507         la->la_atime = la->la_mtime = la->la_ctime = 0;
3508         la->la_valid = LA_TYPE | LA_MODE | LA_UID | LA_GID |
3509                        LA_ATIME | LA_MTIME | LA_CTIME;
3510         memset(dof, 0, sizeof(*dof));
3511         ff->ff_parent.f_seq = cpu_to_le64(pfid->f_seq);
3512         ff->ff_parent.f_oid = cpu_to_le32(pfid->f_oid);
3513         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
3514          * MDT-object's FID::f_ver, instead it is the OST-object index in its
3515          * parent MDT-object's layout EA. */
3516         ff->ff_parent.f_stripe_idx = cpu_to_le32(ea_off);
3517
3518         rc = lfsck_layout_get_lovea(env, parent, tbuf);
3519         if (unlikely(rc == -ENODATA))
3520                 rc = 0;
3521         if (rc <= 0)
3522                 GOTO(unlock1, rc);
3523
3524         rc = lfsck_lov2layout(tbuf->lb_buf, ff, comp_id);
3525         if (rc)
3526                 GOTO(unlock1, rc);
3527
3528         buf = lfsck_buf_get(env, ff, sizeof(struct filter_fid));
3529         handle = dt_trans_create(env, dev);
3530         if (IS_ERR(handle))
3531                 GOTO(unlock1, rc = PTR_ERR(handle));
3532
3533         rc = dt_declare_create(env, child, la, NULL, dof, handle);
3534         if (rc != 0)
3535                 GOTO(stop, rc);
3536
3537         rc = dt_declare_xattr_set(env, child, buf, XATTR_NAME_FID,
3538                                   LU_XATTR_CREATE, handle);
3539         if (rc != 0)
3540                 GOTO(stop, rc);
3541
3542         rc = dt_trans_start_local(env, dev, handle);
3543         if (rc != 0)
3544                 GOTO(stop, rc);
3545
3546         dt_read_lock(env, parent, 0);
3547         if (unlikely(lfsck_is_dead_obj(parent)))
3548                 GOTO(unlock2, rc = 0);
3549
3550         if (lfsck->li_bookmark_ram.lb_param & LPF_DELAY_CREATE_OSTOBJ) {
3551                 struct ost_id *oi = &info->lti_oi;
3552                 struct lu_fid *tfid = &info->lti_fid2;
3553                 struct lu_buf *lovea = &info->lti_big_buf;
3554                 struct lov_mds_md_v1 *lmm;
3555                 struct lov_ost_data_v1 *objs;
3556                 __u32 magic;
3557                 int count;
3558                 int idx2;
3559
3560                 rc = lfsck_layout_get_lovea(env, parent, lovea);
3561                 if (unlikely(rc == -ENODATA))
3562                         rc = 0;
3563                 if (rc <= 0)
3564                         GOTO(unlock2, rc);
3565
3566                 lmm = lovea->lb_buf;
3567                 magic = le32_to_cpu(lmm->lmm_magic);
3568                 if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
3569                         struct lov_comp_md_v1 *lcm = buf->lb_buf;
3570                         struct lov_comp_md_entry_v1 *lcme;
3571                         __u16 count = le16_to_cpu(lcm->lcm_entry_count);
3572                         int i;
3573
3574                         for (i = 0; i < count; i++) {
3575                                 lcme = &lcm->lcm_entries[i];
3576                                 if (le32_to_cpu(lcme->lcme_id) == comp_id) {
3577                                         LASSERT(le32_to_cpu(lcme->lcme_flags) &
3578                                                 LCME_FL_INIT);
3579
3580                                         lmm = lovea->lb_buf +
3581                                                 le32_to_cpu(lcme->lcme_offset);
3582                                         magic = le32_to_cpu(lmm->lmm_magic);
3583                                         goto check;
3584                                 }
3585                         }
3586
3587                         /* Someone removed the component, do nothing. */
3588                         GOTO(unlock2, rc = 0);
3589                 }
3590
3591 check:
3592                 count = le16_to_cpu(lmm->lmm_stripe_count);
3593                 /* Someone changed the LOV EA, do nothing. */
3594                 if (count <= ea_off)
3595                         GOTO(unlock2, rc = 0);
3596
3597                 if (magic == LOV_MAGIC_V1) {
3598                         objs = &lmm->lmm_objects[ea_off];
3599                 } else {
3600                         LASSERT(magic == LOV_MAGIC_V3);
3601
3602                         objs = &((struct lov_mds_md_v3 *)lmm)->\
3603                                                         lmm_objects[ea_off];
3604                 }
3605
3606                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
3607                 idx2 = le32_to_cpu(objs->l_ost_idx);
3608                 rc = ostid_to_fid(tfid, oi, idx2);
3609                 /* Someone changed the LOV EA, do nothing. */
3610                 if (rc != 0 || !lu_fid_eq(tfid, cfid))
3611                         GOTO(unlock2, rc);
3612         }
3613
3614         rc = dt_create(env, child, la, NULL, dof, handle);
3615         if (rc != 0)
3616                 GOTO(unlock2, rc);
3617
3618         rc = dt_xattr_set(env, child, buf, XATTR_NAME_FID, LU_XATTR_CREATE,
3619                           handle);
3620
3621         GOTO(unlock2, rc);
3622
3623 unlock2:
3624         dt_read_unlock(env, parent);
3625
3626 stop:
3627         rc = lfsck_layout_trans_stop(env, dev, handle, rc);
3628
3629 unlock1:
3630         lfsck_ibits_unlock(&lh, LCK_EX);
3631
3632 log:
3633         if (rc && log)
3634                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant found "
3635                        "dangling reference for: parent "DFID", child "
3636                        DFID", comp_id %u, ea_off %u, ost_idx %u, %s: "
3637                        "rc = %d\n",
3638                        lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
3639                        comp_id, ea_off, ost_idx,
3640                        (lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ) ?
3641                                 "Create the lost OST-object as required" :
3642                                 "Keep the MDT-object there by default", rc);
3643
3644         return rc;
3645 }
3646
3647 /**
3648  * Repair the MDT-object with dangling LOV EA reference.
3649  *
3650  * Prepare parameters and call __lfsck_layout_repair_dangling()
3651  * to repair the dangling LOV EA reference.
3652  *
3653  * \param[in] env       pointer to the thread context
3654  * \param[in] com       the layout LFSCK component
3655  * \param[in] pfid      the MDT-object's FID
3656  * \param[in] cfid      the FID for the OST-object to be created
3657  * \param[in] comp_id   the component ID of the OST-object in the LOV EA
3658  * \param[in] ea_off    the offset of the OST-object in the LOV EA
3659  * \param[in] ost_idx   the index of OST on which the OST-object resides
3660  *
3661  * \retval              +1 for repair successfully
3662  * \retval              0 for did nothing
3663  * \retval              negative error number on failure
3664  */
3665 static int lfsck_layout_repair_dangling(const struct lu_env *env,
3666                                         struct lfsck_component *com,
3667                                         const struct lu_fid *pfid,
3668                                         const struct lu_fid *cfid,
3669                                         __u32 comp_id, __u32 ea_off,
3670                                         __u32 ost_idx)
3671 {
3672         struct lfsck_instance *lfsck = com->lc_lfsck;
3673         struct dt_object *parent = NULL;
3674         struct dt_object *child = NULL;
3675         struct lfsck_tgt_desc *ltd;
3676         int rc;
3677         ENTRY;
3678
3679         parent = lfsck_object_find_bottom(env, lfsck, pfid);
3680         if (IS_ERR(parent))
3681                 GOTO(log, rc = PTR_ERR(parent));
3682
3683         /* The MDT-object has been removed. */
3684         if (dt_object_exists(parent) == 0)
3685                 GOTO(log, rc = 0);
3686
3687         ltd = lfsck_ltd2tgt(&lfsck->li_ost_descs, ost_idx);
3688         if (unlikely(ltd == NULL))
3689                 GOTO(log, rc = -ENODEV);
3690
3691         child = lfsck_object_find_by_dev(env, ltd->ltd_tgt, cfid);
3692         if (IS_ERR(child))
3693                 GOTO(log, rc = PTR_ERR(child));
3694
3695         /* The OST-object has been created. */
3696         if (unlikely(dt_object_exists(child) != 0))
3697                 GOTO(log, rc = 0);
3698
3699         rc = __lfsck_layout_repair_dangling(env, com, parent, child,
3700                                             comp_id, ea_off, ost_idx, false);
3701
3702         GOTO(log, rc);
3703
3704 log:
3705         if (child != NULL && !IS_ERR(child))
3706                 lfsck_object_put(env, child);
3707
3708         if (parent != NULL && !IS_ERR(parent))
3709                 lfsck_object_put(env, parent);
3710
3711         if (rc)
3712                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant found "
3713                        "dangling reference for: parent "DFID", child "
3714                        DFID", comp_id %u, ea_off %u, ost_idx %u, %s: rc = %d\n",
3715                        lfsck_lfsck2name(lfsck), PFID(pfid), PFID(cfid),
3716                        comp_id, ea_off, ost_idx,
3717                        (lfsck->li_bookmark_ram.lb_param & LPF_CREATE_OSTOBJ) ?
3718                                 "Create the lost OST-object as required" :
3719                                 "Keep the MDT-object there by default", rc);
3720
3721         return rc;
3722 }
3723
3724 /* If the OST-object does not recognize the MDT-object as its parent, and
3725  * there is no other MDT-object claims as its parent, then just trust the
3726  * given MDT-object as its parent. So update the OST-object filter_fid. */
3727 static int lfsck_layout_repair_unmatched_pair(const struct lu_env *env,
3728                                               struct lfsck_component *com,
3729                                               struct dt_object *parent,
3730                                               struct lfsck_layout_req *llr,
3731                                               struct lu_attr *la)
3732 {
3733         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3734         struct filter_fid               *ff     = &info->lti_ff;
3735         struct dt_object                *child  = llr->llr_child;
3736         struct dt_device                *dev    = lfsck_obj2dev(child);
3737         const struct lu_fid             *tfid   = lu_object_fid(&parent->do_lu);
3738         struct lu_buf                   *tbuf   = &info->lti_big_buf;
3739         struct thandle                  *handle;
3740         struct lu_buf                   *buf;
3741         struct lustre_handle             lh     = { 0 };
3742         int                              rc;
3743         ENTRY;
3744
3745         rc = lfsck_ibits_lock(env, com->lc_lfsck, parent, &lh,
3746                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
3747                               LCK_EX);
3748         if (rc != 0)
3749                 GOTO(log, rc);
3750
3751         ff->ff_parent.f_seq = cpu_to_le64(tfid->f_seq);
3752         ff->ff_parent.f_oid = cpu_to_le32(tfid->f_oid);
3753         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
3754          * MDT-object's FID::f_ver, instead it is the OST-object index in its
3755          * parent MDT-object's layout EA. */
3756         ff->ff_parent.f_stripe_idx = cpu_to_le32(llr->llr_lov_idx);
3757
3758         rc = lfsck_layout_get_lovea(env, parent, tbuf);
3759         if (unlikely(rc == -ENODATA))
3760                 rc = 0;
3761         if (rc <= 0)
3762                 GOTO(unlock1, rc);
3763
3764         rc = lfsck_lov2layout(tbuf->lb_buf, ff, llr->llr_comp_id);
3765         if (rc)
3766                 GOTO(unlock1, rc);
3767
3768         buf = lfsck_buf_get(env, ff, sizeof(*ff));
3769
3770         handle = dt_trans_create(env, dev);
3771         if (IS_ERR(handle))
3772                 GOTO(unlock1, rc = PTR_ERR(handle));
3773
3774         rc = dt_declare_xattr_set(env, child, buf, XATTR_NAME_FID, 0, handle);
3775         if (rc != 0)
3776                 GOTO(stop, rc);
3777
3778         rc = dt_attr_get(env, parent, la);
3779         if (rc != 0)
3780                 GOTO(stop, rc);
3781
3782         la->la_valid = LA_UID | LA_GID;
3783         rc = dt_declare_attr_set(env, child, la, handle);
3784         if (rc != 0)
3785                 GOTO(stop, rc);
3786
3787         rc = dt_trans_start_local(env, dev, handle);
3788         if (rc != 0)
3789                 GOTO(stop, rc);
3790
3791         dt_write_lock(env, parent, 0);
3792         if (unlikely(lfsck_is_dead_obj(parent)))
3793                 GOTO(unlock2, rc = 1);
3794
3795         rc = dt_xattr_set(env, child, buf, XATTR_NAME_FID, 0, handle);
3796         if (rc != 0)
3797                 GOTO(unlock2, rc);
3798
3799         /* Get the latest parent's owner. */
3800         rc = dt_attr_get(env, parent, la);
3801         if (rc != 0)
3802                 GOTO(unlock2, rc);
3803
3804         la->la_valid = LA_UID | LA_GID;
3805         rc = dt_attr_set(env, child, la, handle);
3806
3807         GOTO(unlock2, rc);
3808
3809 unlock2:
3810         dt_write_unlock(env, parent);
3811
3812 stop:
3813         rc = lfsck_layout_trans_stop(env, dev, handle, rc);
3814
3815 unlock1:
3816         lfsck_ibits_unlock(&lh, LCK_EX);
3817
3818 log:
3819         if (rc)
3820                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired "
3821                        "unmatched MDT-OST pair for: parent "DFID
3822                        ", child "DFID", comp_id %u, OST-index %u, "
3823                        "stripe-index %u, owner %u/%u: rc = %d\n",
3824                        lfsck_lfsck2name(com->lc_lfsck),
3825                        PFID(lfsck_dto2fid(parent)),
3826                        PFID(lfsck_dto2fid(child)),
3827                        llr->llr_comp_id, llr->llr_ost_idx, llr->llr_lov_idx,
3828                        la->la_uid, la->la_gid, rc);
3829
3830         return rc;
3831 }
3832
3833 /* If there are more than one MDT-objects claim as the OST-object's parent,
3834  * and the OST-object only recognizes one of them, then we need to generate
3835  * new OST-object(s) with new fid(s) for the non-recognized MDT-object(s). */
3836 static int lfsck_layout_repair_multiple_references(const struct lu_env *env,
3837                                                    struct lfsck_component *com,
3838                                                    struct dt_object *parent,
3839                                                    struct lfsck_layout_req *llr,
3840                                                    struct lu_attr *la)
3841 {
3842         struct lfsck_thread_info        *info   = lfsck_env_info(env);
3843         struct dt_allocation_hint       *hint   = &info->lti_hint;
3844         struct dt_object_format         *dof    = &info->lti_dof;
3845         struct ost_id                   *oi     = &info->lti_oi;
3846         struct lu_buf                   *buf    = &info->lti_big_buf;
3847         struct lfsck_instance           *lfsck  = com->lc_lfsck;
3848         struct dt_device                *dev;
3849         struct lu_device                *d      =
3850                                 &lfsck_obj2dev(llr->llr_child)->dd_lu_dev;
3851         struct lu_object                *o;
3852         struct lu_object                *n;
3853         struct dt_object                *child  = NULL;
3854         struct thandle                  *handle = NULL;
3855         struct lov_mds_md_v1            *lmm;
3856         struct lov_ost_data_v1          *objs;
3857         const struct lu_fid             *pfid   = lfsck_dto2fid(parent);
3858         struct lu_fid                    tfid;
3859         struct lustre_handle             lh     = { 0 };
3860         __u32                            magic;
3861         __u32                            index;
3862         int                              rc;
3863         ENTRY;
3864
3865         /* We use two separated transactions to repair the inconsistency.
3866          *
3867          * 1) create the child (OST-object).
3868          * 2) update the parent LOV EA according to the child's FID.
3869          *
3870          * If 1) succeed, but 2) failed or aborted, then such OST-object will be
3871          * handled as orphan when the layout LFSCK run next time.
3872          *
3873          * If 1) failed, but 2) succeed, then such OST-object will be re-created
3874          * as dangling referened case when the layout LFSCK run next time. */
3875
3876         /* The 1st transaction. */
3877         o = lu_object_anon(env, d, NULL);
3878         if (IS_ERR(o))
3879                 GOTO(log, rc = PTR_ERR(o));
3880
3881         n = lu_object_locate(o->lo_header, d->ld_type);
3882         if (unlikely(n == NULL)) {
3883                 lu_object_put_nocache(env, o);
3884
3885                 GOTO(log, rc = -EINVAL);
3886         }
3887
3888         child = container_of(n, struct dt_object, do_lu);
3889         memset(hint, 0, sizeof(*hint));
3890         rc = dt_attr_get(env, parent, la);
3891         if (rc != 0)
3892                 GOTO(log, rc);
3893
3894         la->la_valid = LA_UID | LA_GID;
3895         memset(dof, 0, sizeof(*dof));
3896
3897         dev = lfsck_obj2dev(child);
3898         handle = dt_trans_create(env, dev);
3899         if (IS_ERR(handle))
3900                 GOTO(log, rc = PTR_ERR(handle));
3901
3902         rc = dt_declare_create(env, child, la, hint, dof, handle);
3903         if (rc != 0)
3904                 GOTO(stop, rc);
3905
3906         rc = dt_trans_start_local(env, dev, handle);
3907         if (rc != 0)
3908                 GOTO(stop, rc);
3909
3910         rc = dt_create(env, child, la, hint, dof, handle);
3911         dt_trans_stop(env, dev, handle);
3912         handle = NULL;
3913         if (rc != 0)
3914                 GOTO(log, rc);
3915
3916         rc = lfsck_ibits_lock(env, lfsck, parent, &lh,
3917                               MDS_INODELOCK_LAYOUT | MDS_INODELOCK_XATTR,
3918                               LCK_EX);
3919         if (rc != 0)
3920                 GOTO(log, rc);
3921
3922         /* The 2nd transaction. */
3923
3924         /* XXX: Generally, we should use bottom device (OSD) to update parent
3925          *      LOV EA. But because the LOD-object still references the wrong
3926          *      OSP-object that should be detached after the parent's LOV EA
3927          *      refreshed. Unfortunately, there is no suitable API for that.
3928          *      So we have to make the LOD to re-load the OSP-object(s) via
3929          *      replacing the LOV EA against the LOD-object.
3930          *
3931          *      Once the DNE2 patches have been landed, we can replace the
3932          *      LOD device with the OSD device. LU-6230. */
3933
3934         dev = lfsck->li_next;
3935         parent = lfsck_object_locate(dev, parent);
3936         if (IS_ERR(parent))
3937                 GOTO(log, rc = PTR_ERR(parent));
3938
3939         handle = dt_trans_create(env, dev);
3940         if (IS_ERR(handle))
3941                 GOTO(log, rc = PTR_ERR(handle));
3942
3943         rc = dt_declare_xattr_set(env, parent, buf, XATTR_NAME_LOV,
3944                                   LU_XATTR_REPLACE, handle);
3945         if (rc != 0)
3946                 GOTO(stop, rc);
3947
3948         rc = dt_trans_start_local(env, dev, handle);
3949         if (rc != 0)
3950                 GOTO(stop, rc);
3951
3952         dt_write_lock(env, parent, 0);
3953         if (unlikely(lfsck_is_dead_obj(parent)))
3954                 GOTO(unlock, rc = 0);
3955
3956         rc = lfsck_layout_get_lovea(env, parent, buf);
3957         if (unlikely(rc == -ENODATA))
3958                 rc = 0;
3959         if (rc <= 0)
3960                 GOTO(unlock, rc);
3961
3962         lmm = buf->lb_buf;
3963         magic = le32_to_cpu(lmm->lmm_magic);
3964         if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
3965                 struct lov_comp_md_v1 *lcm = buf->lb_buf;
3966                 struct lov_comp_md_entry_v1 *lcme;
3967                 __u16 count = le16_to_cpu(lcm->lcm_entry_count);
3968                 int i;
3969
3970                 LASSERT(llr->llr_comp_id != 0);
3971
3972                 for (i = 0; i < count; i++) {
3973                         lcme = &lcm->lcm_entries[i];
3974                         if (le32_to_cpu(lcme->lcme_id) == llr->llr_comp_id) {
3975                                 LASSERT(le32_to_cpu(lcme->lcme_flags) &
3976                                         LCME_FL_INIT);
3977
3978                                 le32_add_cpu(&lcm->lcm_layout_gen, 1);
3979                                 lmm = buf->lb_buf +
3980                                         le32_to_cpu(lcme->lcme_offset);
3981                                 magic = le32_to_cpu(lmm->lmm_magic);
3982                                 goto set;
3983                         }
3984                 }
3985
3986                 GOTO(unlock, rc = 0);
3987         }
3988
3989 set:
3990         if (magic == LOV_MAGIC_V1) {
3991                 objs = &lmm->lmm_objects[llr->llr_lov_idx];
3992         } else {
3993                 LASSERT(magic == LOV_MAGIC_V3);
3994                 objs =
3995                 &((struct lov_mds_md_v3 *)lmm)->lmm_objects[llr->llr_lov_idx];
3996         }
3997
3998         ostid_le_to_cpu(&objs->l_ost_oi, oi);
3999         index = le32_to_cpu(objs->l_ost_idx);
4000         rc = ostid_to_fid(&tfid, oi, index);
4001         /* Someone changed layout during the LFSCK, no need to repair then. */
4002         if (rc == 0 && !lu_fid_eq(&tfid, lu_object_fid(&llr->llr_child->do_lu)))
4003                 GOTO(unlock, rc = 0);
4004
4005         lmm->lmm_layout_gen = cpu_to_le16(le16_to_cpu(lmm->lmm_layout_gen) + 1);
4006         fid_to_ostid(lu_object_fid(&child->do_lu), oi);
4007         ostid_cpu_to_le(oi, &objs->l_ost_oi);
4008         objs->l_ost_gen = cpu_to_le32(0);
4009         objs->l_ost_idx = cpu_to_le32(llr->llr_ost_idx);
4010         rc = dt_xattr_set(env, parent, buf, XATTR_NAME_LOV,
4011                           LU_XATTR_REPLACE, handle);
4012
4013         GOTO(unlock, rc = (rc == 0 ? 1 : rc));
4014
4015 unlock:
4016         dt_write_unlock(env, parent);
4017
4018 stop:
4019         if (handle != NULL)
4020                 dt_trans_stop(env, dev, handle);
4021
4022 log:
4023         lfsck_ibits_unlock(&lh, LCK_EX);
4024         if (child != NULL)
4025                 lfsck_object_put(env, child);
4026
4027         if (rc)
4028                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired "
4029                        "multiple references for: parent "DFID", comp_id %u, "
4030                        "OST-index %u, stripe-index %u, owner %u/%u: rc = %d\n",
4031                        lfsck_lfsck2name(lfsck), PFID(pfid),
4032                        llr->llr_comp_id, llr->llr_ost_idx, llr->llr_lov_idx,
4033                        la->la_uid, la->la_gid, rc);
4034
4035         return rc;
4036 }
4037
4038 /* If the MDT-object and the OST-object have different owner information,
4039  * then trust the MDT-object, because the normal chown/chgrp handle order
4040  * is from MDT to OST, and it is possible that some chown/chgrp operation
4041  * is partly done. */
4042 static int lfsck_layout_repair_owner(const struct lu_env *env,
4043                                      struct lfsck_component *com,
4044                                      struct dt_object *parent,
4045                                      struct lfsck_layout_req *llr,
4046                                      struct lu_attr *pla,
4047                                      const struct lu_attr *cla)
4048 {
4049         struct lfsck_thread_info        *info   = lfsck_env_info(env);
4050         struct lu_attr                  *tla    = &info->lti_la2;
4051         struct dt_object                *child  = llr->llr_child;
4052         struct dt_device                *dev    = lfsck_obj2dev(child);
4053         struct thandle                  *handle;
4054         int                              rc;
4055         dt_obj_version_t                 version;
4056         ENTRY;
4057
4058         tla->la_uid = pla->la_uid;
4059         tla->la_gid = pla->la_gid;
4060         tla->la_valid = LA_UID | LA_GID;
4061         handle = dt_trans_create(env, dev);
4062         if (IS_ERR(handle))
4063                 GOTO(log, rc = PTR_ERR(handle));
4064
4065         rc = dt_declare_attr_set(env, child, tla, handle);
4066         if (rc != 0)
4067                 GOTO(stop, rc);
4068
4069         rc = dt_trans_start_local(env, dev, handle);
4070         if (rc != 0)
4071                 GOTO(stop, rc);
4072
4073         /* Use the dt_object lock to serialize with destroy and attr_set. */
4074         dt_read_lock(env, parent, 0);
4075         if (unlikely(lfsck_is_dead_obj(parent)))
4076                 GOTO(unlock, rc = 1);
4077
4078         version = dt_version_get(env, child);
4079         if (version == -EOPNOTSUPP)
4080                 version = 0;
4081
4082         /* Get the latest parent's owner. */
4083         rc = dt_attr_get(env, parent, pla);
4084         if (rc != 0)
4085                 GOTO(unlock, rc);
4086
4087         /* Some others chown/chgrp during the LFSCK, needs to do nothing. */
4088         if (unlikely((!version && tla->la_ctime == 0) ||
4089                      tla->la_uid != pla->la_uid || tla->la_gid != pla->la_gid))
4090                 rc = 1;
4091         else
4092                 rc = dt_attr_set(env, child, tla, handle);
4093
4094         GOTO(unlock, rc);
4095
4096 unlock:
4097         dt_read_unlock(env, parent);
4098
4099 stop:
4100         rc = lfsck_layout_trans_stop(env, dev, handle, rc);
4101
4102 log:
4103         if (rc != 0)
4104                 CDEBUG(D_LFSCK, "%s: layout LFSCK assistant repaired "
4105                        "inconsistent file owner for: parent "DFID", child "DFID
4106                        ", OST-index %u, stripe-index %u, old owner %u/%u, "
4107                        "new owner %u/%u: rc = %d\n",
4108                        lfsck_lfsck2name(com->lc_lfsck),
4109                        PFID(lfsck_dto2fid(parent)), PFID(lfsck_dto2fid(child)),
4110                        llr->llr_ost_idx, llr->llr_lov_idx,
4111                        cla->la_uid, cla->la_gid, tla->la_uid, tla->la_gid, rc);
4112
4113         return rc;
4114 }
4115
4116 #define CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid, cfid, msg)               \
4117         CDEBUG(D_LFSCK, "%s:("DFID"|"DFID")/"DFID":XATTR %s: %s\n",      \
4118                lfsck_lfsck2name(lfsck), PFID(&lso->lso_fid), PFID(pfid), \
4119                PFID(cfid), XATTR_NAME_FID, msg);
4120
4121 /* Check whether the OST-object correctly back points to the
4122  * MDT-object (@parent) via the XATTR_NAME_FID xattr (@pfid). */
4123 static int lfsck_layout_check_parent(const struct lu_env *env,
4124                                      struct lfsck_component *com,
4125                                      struct lfsck_assistant_object *lso,
4126                                      struct filter_fid *ff,
4127                                      const struct lu_fid *cfid,
4128                                      const struct lu_attr *cla,
4129                                      struct lfsck_layout_req *llr)
4130 {
4131         struct lfsck_thread_info        *info   = lfsck_env_info(env);
4132         struct lu_buf                   *buf    = &info->lti_big_buf;
4133         struct lu_fid                   *pfid   = &info->lti_fid;
4134         struct dt_object                *tobj;
4135         struct lov_mds_md_v1            *lmm;
4136         struct lov_ost_data_v1          *objs;
4137         struct lustre_handle             lh     = { 0 };
4138         struct lfsck_instance           *lfsck  = com->lc_lfsck;
4139         int                              rc;
4140         int                              i;
4141         __u32                            magic;
4142         __u32                            idx;
4143         __u16                            count;
4144         ENTRY;
4145
4146         *pfid = ff->ff_parent;
4147         idx = pfid->f_stripe_idx;
4148         pfid->f_ver = 0;
4149
4150         if (unlikely(!fid_is_sane(pfid))) {
4151                 CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid, cfid,
4152                                       "the parent FID is invalid");
4153
4154                 RETURN(LLIT_UNMATCHED_PAIR);
4155         }
4156
4157         if (lu_fid_eq(pfid, &lso->lso_fid)) {
4158                 if (likely(llr->llr_lov_idx == idx))
4159                         RETURN(0);
4160
4161                 CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid, cfid,
4162                                       "the stripe index is unmatched");
4163
4164                 RETURN(LLIT_UNMATCHED_PAIR);
4165         }
4166
4167         tobj = lfsck_object_find_bottom(env, com->lc_lfsck, pfid);
4168         if (IS_ERR(tobj))
4169                 RETURN(PTR_ERR(tobj));
4170
4171         if (dt_object_exists(tobj) == 0) {
4172                 CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid, cfid,
4173                                       "the parent is nonexistent");
4174
4175                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4176         }
4177
4178         if (lfsck_is_dead_obj(tobj)) {
4179                 CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid, cfid,
4180                                       "the parent is dead object");
4181
4182                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4183         }
4184
4185         if (!S_ISREG(lfsck_object_type(tobj))) {
4186                 CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid, cfid,
4187                                       "the parent is not a regular file");
4188
4189                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4190         }
4191
4192         /* Load the tobj's layout EA, in spite of it is a local MDT-object or
4193          * remote one on another MDT. Then check whether the given OST-object
4194          * is in such layout. If yes, it is multiple referenced, otherwise it
4195          * is unmatched referenced case. */
4196         rc = lfsck_layout_get_lovea(env, tobj, buf);
4197         if (rc == 0 || rc == -ENODATA || rc == -ENOENT) {
4198                 CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid, cfid,
4199                                       "the parent has no stripe data");
4200
4201                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4202         }
4203
4204         if (unlikely(rc == -EOPNOTSUPP))
4205                 GOTO(out, rc = LLIT_NONE);
4206
4207         if (rc < 0)
4208                 GOTO(out, rc);
4209
4210         lmm = buf->lb_buf;
4211         magic = le32_to_cpu(lmm->lmm_magic);
4212         if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
4213                 struct lov_comp_md_v1 *lcm = buf->lb_buf;
4214                 struct lov_comp_md_entry_v1 *lcme;
4215
4216                 if (ff->ff_layout.ol_comp_id == 0) {
4217                         CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid, cfid,
4218                                               "the parent has incorrect comp_id");
4219
4220                         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4221                 }
4222
4223                 count = le16_to_cpu(lcm->lcm_entry_count);
4224                 for (i = 0; i < count; i++) {
4225                         lcme = &lcm->lcm_entries[i];
4226                         if (le32_to_cpu(lcme->lcme_id) ==
4227                             ff->ff_layout.ol_comp_id) {
4228                                 lmm = buf->lb_buf +
4229                                         le32_to_cpu(lcme->lcme_offset);
4230                                 magic = le32_to_cpu(lmm->lmm_magic);
4231                                 if (!(le32_to_cpu(lcme->lcme_flags) &
4232                                       LCME_FL_INIT)) {
4233                                         CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid,
4234                                                               cfid,
4235                                                               "the parent has uninitialized component");
4236
4237                                         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4238                                 }
4239
4240                                 goto further;
4241                         }
4242                 }
4243
4244                 CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid, cfid,
4245                                       "the parent has no matched comp_id");
4246
4247                 GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4248         }
4249
4250 further:
4251         if (magic == LOV_MAGIC_V1) {
4252                 objs = &lmm->lmm_objects[0];
4253         } else {
4254                 LASSERT(magic == LOV_MAGIC_V3);
4255                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
4256         }
4257
4258         count = le16_to_cpu(lmm->lmm_stripe_count);
4259         for (i = 0; i < count; i++, objs++) {
4260                 struct lu_fid           *tfid   = &info->lti_fid2;
4261                 struct ost_id           *oi     = &info->lti_oi;
4262                 __u32                    idx2;
4263
4264                 if (lovea_slot_is_dummy(objs))
4265                         continue;
4266
4267                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
4268                 idx2 = le32_to_cpu(objs->l_ost_idx);
4269                 rc = ostid_to_fid(tfid, oi, idx2);
4270                 if (rc != 0) {
4271                         CDEBUG(D_LFSCK, "%s: the parent "DFID" contains "
4272                                "invalid layout EA at the slot %d, index %u\n",
4273                                lfsck_lfsck2name(com->lc_lfsck),
4274                                PFID(pfid), i, idx2);
4275
4276                         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4277                 }
4278
4279                 if (lu_fid_eq(cfid, tfid)) {
4280                         rc = lfsck_ibits_lock(env, com->lc_lfsck, tobj, &lh,
4281                                               MDS_INODELOCK_UPDATE |
4282                                               MDS_INODELOCK_LAYOUT |
4283                                               MDS_INODELOCK_XATTR,
4284                                               LCK_EX);
4285                         if (rc != 0)
4286                                 GOTO(out, rc);
4287
4288                         dt_read_lock(env, tobj, 0);
4289
4290                         /* For local MDT-object, re-check existence
4291                          * after taken the lock. */
4292                         if (!dt_object_remote(tobj)) {
4293                                 if (dt_object_exists(tobj) == 0 ||
4294                                     lfsck_is_dead_obj(tobj)) {
4295                                         CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid,
4296                                                               cfid,
4297                                                               "the parent doesn't exist anymore after lock");
4298
4299                                         rc = LLIT_UNMATCHED_PAIR;
4300                                 } else {
4301                                         rc = LLIT_MULTIPLE_REFERENCED;
4302                                 }
4303
4304                                 GOTO(unlock, rc);
4305                         }
4306
4307                         /* For migration case, the new MDT-object and old
4308                          * MDT-object may reference the same OST-object at
4309                          * some migration internal time.
4310                          *
4311                          * For remote MDT-object, the local MDT may not know
4312                          * whether it has been removed or not.  Try checking
4313                          * for a non-existent xattr to check if this object
4314                          * has been been removed or not. */
4315                         rc = dt_xattr_get(env, tobj, &LU_BUF_NULL,
4316                                           XATTR_NAME_DUMMY);
4317                         if (unlikely(rc == -ENOENT || rc >= 0)) {
4318                                 CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid, cfid,
4319                                                       "the parent is remote object and nonexistent after lock");
4320
4321                                 rc = LLIT_UNMATCHED_PAIR;
4322                         } else if (rc == -ENODATA) {
4323                                 rc = LLIT_MULTIPLE_REFERENCED;
4324                         }
4325
4326                         GOTO(unlock, rc);
4327                 }
4328         }
4329
4330         CDEBUG_UNMATCHED_PAIR(lfsck, lso, pfid, cfid,
4331                               "the parent has no matched stripe");
4332
4333         GOTO(out, rc = LLIT_UNMATCHED_PAIR);
4334
4335 unlock:
4336         if (lustre_handle_is_used(&lh)) {
4337                 dt_read_unlock(env, tobj);
4338                 lfsck_ibits_unlock(&lh, LCK_EX);
4339         }
4340
4341 out:
4342         lfsck_object_put(env, tobj);
4343
4344         return rc;
4345 }
4346
4347 static int lfsck_layout_assistant_handler_p1(const struct lu_env *env,
4348                                              struct lfsck_component *com,
4349                                              struct lfsck_assistant_req *lar)
4350 {
4351         struct lfsck_layout_req              *llr    =
4352                 container_of(lar, struct lfsck_layout_req, llr_lar);
4353         struct lfsck_assistant_object        *lso    = lar->lar_parent;
4354         struct lfsck_layout                  *lo     = com->lc_file_ram;
4355         struct lfsck_thread_info             *info   = lfsck_env_info(env);
4356         struct filter_fid                    *ff     = &info->lti_ff;
4357         struct lu_buf buf = { .lb_buf = ff,
4358                               .lb_len = sizeof(*ff) };
4359         struct dt_object                     *parent = NULL;
4360         struct dt_object                     *child  = llr->llr_child;
4361         struct lu_attr                       *pla    = &lso->lso_attr;
4362         struct lu_attr                       *cla    = &info->lti_la;
4363         struct lfsck_instance                *lfsck  = com->lc_lfsck;
4364         struct lfsck_bookmark                *bk     = &lfsck->li_bookmark_ram;
4365         enum lfsck_layout_inconsistency_type  type   = LLIT_NONE;
4366         int                                   rc;
4367         ENTRY;
4368
4369         if (lso->lso_dead)
4370                 RETURN(0);
4371
4372         CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_ENGINE_DELAY, cfs_fail_val);
4373
4374         rc = dt_attr_get(env, child, cla);
4375         if (rc == -ENOENT) {
4376                 parent = lfsck_assistant_object_load(env, lfsck, lso);
4377                 if (IS_ERR(parent)) {
4378                         rc = PTR_ERR(parent);
4379
4380                         RETURN(rc == -ENOENT ? 0 : rc);
4381                 }
4382
4383                 type = LLIT_DANGLING;
4384                 goto repair;
4385         }
4386
4387         if (rc != 0)
4388                 GOTO(out, rc);
4389
4390         lfsck_buf_init(&buf, ff, sizeof(*ff));
4391         rc = dt_xattr_get(env, child, &buf, XATTR_NAME_FID);
4392         if (unlikely(rc > 0 && rc < sizeof(struct lu_fid))) {
4393                 CDEBUG(D_LFSCK, "%s:"DFID"/"DFID": "
4394                        "the child object's %s is corrupted\n",
4395                        lfsck_lfsck2name(lfsck), PFID(&lso->lso_fid),
4396                        PFID(lu_object_fid(&child->do_lu)),
4397                        XATTR_NAME_FID);
4398
4399                 type = LLIT_UNMATCHED_PAIR;
4400                 goto repair;
4401         }
4402
4403         if (rc < 0 && rc != -ENODATA)
4404                 GOTO(out, rc);
4405
4406         if (rc == 0 || rc == -ENODATA)
4407                 GOTO(check_owner, rc = 0);
4408
4409         filter_fid_le_to_cpu(ff, ff, sizeof(*ff));
4410         rc = lfsck_layout_check_parent(env, com, lso, ff,
4411                                        lu_object_fid(&child->do_lu), cla, llr);
4412         if (rc > 0) {
4413                 type = rc;
4414                 goto repair;
4415         }
4416
4417         if (rc < 0)
4418                 GOTO(out, rc);
4419
4420 check_owner:
4421         /* Someone may has changed the owner after the parent attr pre-loaded.
4422          * It can be handled later inside the lfsck_layout_repair_owner(). */
4423         if (unlikely(cla->la_uid != pla->la_uid ||
4424                      cla->la_gid != pla->la_gid)) {
4425                 type = LLIT_INCONSISTENT_OWNER;
4426                 goto repair;
4427         }
4428
4429 repair:
4430         if (type == LLIT_NONE)
4431                 GOTO(out, rc = 0);
4432
4433         if (bk->lb_param & LPF_DRYRUN)
4434                 GOTO(out, rc = 1);
4435
4436         if (parent == NULL) {
4437                 parent = lfsck_assistant_object_load(env, lfsck, lso);
4438                 if (IS_ERR(parent)) {
4439                         rc = PTR_ERR(parent);
4440
4441                         if (rc == -ENOENT)
4442                                 RETURN(0);
4443
4444                         GOTO(out, rc);
4445                 }
4446         }
4447
4448         switch (type) {
4449         case LLIT_DANGLING:
4450                 if (bk->lb_param & LPF_DELAY_CREATE_OSTOBJ)
4451                         rc = lfsck_layout_ins_dangling_rec(env, com,
4452                                 lfsck_dto2fid(parent), lfsck_dto2fid(child),
4453                                 llr->llr_comp_id, llr->llr_lov_idx,
4454                                 llr->llr_ost_idx);
4455                 else
4456                         rc = __lfsck_layout_repair_dangling(env, com, parent,
4457                                                             llr->llr_child,
4458                                                             llr->llr_comp_id,
4459                                                             llr->llr_lov_idx,
4460                                                             llr->llr_ost_idx,
4461                                                             true);
4462                 break;
4463         case LLIT_UNMATCHED_PAIR:
4464                 rc = lfsck_layout_repair_unmatched_pair(env, com, parent,
4465                                                         llr, pla);
4466                 break;
4467         case LLIT_MULTIPLE_REFERENCED:
4468                 rc = lfsck_layout_repair_multiple_references(env, com, parent,
4469                                                              llr, pla);
4470                 break;
4471         case LLIT_INCONSISTENT_OWNER:
4472                 rc = lfsck_layout_repair_owner(env, com, parent, llr, pla, cla);
4473                 break;
4474         default:
4475                 rc = 0;
4476                 break;
4477         }
4478
4479         GOTO(out, rc);
4480
4481 out:
4482         down_write(&com->lc_sem);
4483         if (rc < 0) {
4484                 struct lfsck_assistant_data *lad = com->lc_data;
4485
4486                 if (unlikely(test_bit(LAD_EXIT, &lad->lad_flags))) {
4487                         rc = 0;
4488                 } else if (rc == -ENOTCONN || rc == -ESHUTDOWN ||
4489                            rc == -ETIMEDOUT || rc == -EHOSTDOWN ||
4490                            rc == -EHOSTUNREACH) {
4491                         /* If cannot touch the target server,
4492                          * mark the LFSCK as INCOMPLETE. */
4493                         CDEBUG(D_LFSCK, "%s: layout LFSCK assistant fail to "
4494                                "talk with OST %x: rc = %d\n",
4495                                lfsck_lfsck2name(lfsck), llr->llr_ost_idx, rc);
4496                         lfsck_lad_set_bitmap(env, com, llr->llr_ost_idx);
4497                         lo->ll_objs_skipped++;
4498                         rc = 0;
4499                 } else {
4500                         lfsck_layout_record_failure(env, lfsck, lo);
4501                 }
4502         } else if (rc > 0 && (type != LLIT_DANGLING ||
4503                               !(bk->lb_param & LPF_DELAY_CREATE_OSTOBJ))) {
4504                 LASSERTF(type > LLIT_NONE && type <= LLIT_MAX,
4505                          "unknown type = %d\n", type);
4506
4507                 lo->ll_objs_repaired[type - 1]++;
4508                 if (bk->lb_param & LPF_DRYRUN &&
4509                     unlikely(lo->ll_pos_first_inconsistent == 0))
4510                         lo->ll_pos_first_inconsistent =
4511                         lfsck->li_obj_oit->do_index_ops->dio_it.store(env,
4512                                                         lfsck->li_di_oit);
4513         }
4514         up_write(&com->lc_sem);
4515
4516         if (parent != NULL && !IS_ERR(parent))
4517                 lfsck_object_put(env, parent);
4518
4519         return rc;
4520 }
4521
4522 static int
4523 lfsck_layout_double_scan_one_trace_file(const struct lu_env *env,
4524                                         struct lfsck_component *com,
4525                                         struct dt_object *obj, bool first)
4526 {
4527         struct lfsck_instance *lfsck = com->lc_lfsck;
4528         struct ptlrpc_thread *thread = &lfsck->li_thread;
4529         struct lfsck_bookmark *bk = &lfsck->li_bookmark_ram;
4530         struct lfsck_layout *lo = com->lc_file_ram;
4531         const struct dt_it_ops *iops = &obj->do_index_ops->dio_it;
4532         struct dt_it *di;
4533         struct dt_key *key;
4534         struct lfsck_layout_dangling_key *parent =
4535                                         &lfsck_env_info(env)->lti_lldk;
4536         struct lu_fid *cfid = &lfsck_env_info(env)->lti_fid3;
4537         __u32 ost_idx;
4538         int rc;
4539         ENTRY;
4540
4541         di = iops->init(env, obj, 0);
4542         if (IS_ERR(di))
4543                 RETURN(PTR_ERR(di));
4544
4545         if (first)
4546                 lldk_cpu_to_be(parent, &lo->ll_lldk_latest_scanned_phase2);
4547         else
4548                 memset(parent, 0, sizeof(*parent));
4549         rc = iops->get(env, di, (const struct dt_key *)parent);
4550         if (rc < 0)
4551                 GOTO(fini, rc);
4552
4553         if (first) {
4554                 /* The start one either has been processed or does not exist,
4555                  * skip it. */
4556                 rc = iops->next(env, di);
4557                 if (rc != 0)
4558                         GOTO(put, rc);
4559         }
4560
4561         do {
4562                 if (CFS_FAIL_TIMEOUT(OBD_FAIL_LFSCK_DELAY3, cfs_fail_val) &&
4563                     unlikely(!thread_is_running(thread)))
4564                         GOTO(put, rc = 0);
4565
4566                 key = iops->key(env, di);
4567                 if (IS_ERR(key)) {
4568                         rc = PTR_ERR(key);
4569                         if (rc == -ENOENT)
4570                                 GOTO(put, rc = 1);
4571
4572                         goto checkpoint;
4573                 }
4574
4575                 lldk_be_to_cpu(parent,
4576                                 (const struct lfsck_layout_dangling_key *)key);
4577                 if (!fid_is_sane(&parent->lldk_fid)) {
4578                         rc = 0;
4579                         goto checkpoint;
4580                 }
4581
4582                 rc = iops->rec(env, di, (struct dt_rec *)cfid, 0);
4583                 if (rc == 0) {
4584                         fid_be_to_cpu(cfid, cfid);
4585                         ost_idx = cfid->f_ver;
4586                         cfid->f_ver = 0;
4587                         if (!fid_is_sane(cfid)) {
4588                                 rc = 0;
4589                                 goto checkpoint;
4590                         }
4591
4592                         rc = lfsck_layout_repair_dangling(env, com,
4593                                         &parent->lldk_fid, cfid,
4594                                         parent->lldk_comp_id,
4595                                         parent->lldk_ea_off, ost_idx);
4596                 }
4597
4598 checkpoint:
4599                 down_write(&com->lc_sem);
4600                 com->lc_new_checked++;
4601                 com->lc_new_scanned++;
4602                 if (rc >= 0)
4603                         lo->ll_lldk_latest_scanned_phase2 = *parent;
4604
4605                 if (rc > 0)
4606                         lo->ll_objs_repaired[LLIT_DANGLING - 1]++;
4607                 else if (rc < 0)
4608                         lo->ll_objs_failed_phase2++;
4609                 up_write(&com->lc_sem);
4610
4611                 if (rc < 0 && bk->lb_param & LPF_FAILOUT)
4612                         GOTO(put, rc);
4613
4614                 if (unlikely(com->lc_time_next_checkpoint <=
4615                              ktime_get_seconds()) &&
4616                     com->lc_new_checked != 0) {
4617                         down_write(&com->lc_sem);
4618                         lo->ll_run_time_phase2 += ktime_get_seconds() -
4619                                                   com->lc_time_last_checkpoint;
4620                         lo->ll_time_last_checkpoint = ktime_get_real_seconds();
4621                         lo->ll_objs_checked_phase2 += com->lc_new_checked;
4622                         com->lc_new_checked = 0;
4623                         lfsck_layout_store(env, com);
4624                         up_write(&com->lc_sem);
4625
4626                         com->lc_time_last_checkpoint = ktime_get_seconds();
4627                         com->lc_time_next_checkpoint =
4628                                 com->lc_time_last_checkpoint +
4629                                 LFSCK_CHECKPOINT_INTERVAL;
4630                 }
4631
4632                 lfsck_control_speed_by_self(com);
4633                 if (unlikely(!thread_is_running(thread)))
4634                         GOTO(put, rc = 0);
4635
4636                 rc = iops->next(env, di);
4637         } while (rc == 0);
4638
4639         GOTO(put, rc);
4640
4641 put:
4642         iops->put(env, di);
4643
4644 fini:
4645         iops->fini(env, di);
4646
4647         return rc;
4648 }
4649
4650 static int lfsck_layout_assistant_handler_p2(const struct lu_env *env,
4651                                              struct lfsck_component *com)
4652 {
4653         struct lfsck_assistant_data     *lad    = com->lc_data;
4654         struct lfsck_instance           *lfsck  = com->lc_lfsck;
4655         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
4656         struct lfsck_tgt_descs          *ltds   = &lfsck->li_ost_descs;
4657         struct lfsck_tgt_desc           *ltd;
4658         int                              rc     = 0;
4659         ENTRY;
4660
4661         CDEBUG(D_LFSCK, "%s: layout LFSCK phase2 scan start\n",
4662                lfsck_lfsck2name(lfsck));
4663
4664         spin_lock(&ltds->ltd_lock);
4665         while (!list_empty(&lad->lad_ost_phase2_list)) {
4666                 ltd = list_entry(lad->lad_ost_phase2_list.next,
4667                                  struct lfsck_tgt_desc,
4668                                  ltd_layout_phase_list);
4669                 list_del_init(&ltd->ltd_layout_phase_list);
4670                 if (bk->lb_param & LPF_OST_ORPHAN) {
4671                         spin_unlock(&ltds->ltd_lock);
4672                         rc = lfsck_layout_scan_orphan(env, com, ltd);
4673                         if (rc != 0 && bk->lb_param & LPF_FAILOUT)
4674                                 RETURN(rc);
4675
4676                         if (unlikely(test_bit(LAD_EXIT, &lad->lad_flags) ||
4677                                      !thread_is_running(&lfsck->li_thread)))
4678                                 RETURN(0);
4679                         spin_lock(&ltds->ltd_lock);
4680                 }
4681         }
4682
4683         if (list_empty(&lad->lad_ost_phase1_list))
4684                 rc = 1;
4685         else
4686                 rc = 0;
4687         spin_unlock(&ltds->ltd_lock);
4688
4689         if (rc == 1 && bk->lb_param & LPF_OST_ORPHAN) {
4690                 struct lfsck_layout *lo = com->lc_file_ram;
4691                 int i;
4692
4693                 com->lc_new_checked = 0;
4694                 com->lc_new_scanned = 0;
4695                 com->lc_time_last_checkpoint = ktime_get_seconds();
4696                 com->lc_time_next_checkpoint = com->lc_time_last_checkpoint +
4697                                                LFSCK_CHECKPOINT_INTERVAL;
4698
4699                 i = lfsck_sub_trace_file_fid2idx(
4700                                 &lo->ll_lldk_latest_scanned_phase2.lldk_fid);
4701                 rc = lfsck_layout_double_scan_one_trace_file(env, com,
4702                                 com->lc_sub_trace_objs[i].lsto_obj, true);
4703                 while (rc > 0 && ++i < LFSCK_STF_COUNT)
4704                         rc = lfsck_layout_double_scan_one_trace_file(env, com,
4705                                 com->lc_sub_trace_objs[i].lsto_obj, false);
4706
4707                 CDEBUG(D_LFSCK, "%s: layout LFSCK phase2 scan dangling stop "
4708                        "at the No. %d trace file: rc = %d\n",
4709                        lfsck_lfsck2name(lfsck), i, rc);
4710         }
4711
4712         CDEBUG(D_LFSCK, "%s: layout LFSCK phase2 scan stop: rc = %d\n",
4713                lfsck_lfsck2name(lfsck), rc);
4714
4715         RETURN(rc);
4716 }
4717
4718 static int
4719 lfsck_layout_slave_async_interpret(const struct lu_env *env,
4720                                    struct ptlrpc_request *req,
4721                                    void *args, int rc)
4722 {
4723         struct lfsck_layout_slave_async_args *llsaa = args;
4724         struct obd_export *exp = llsaa->llsaa_exp;
4725         struct lfsck_component *com = llsaa->llsaa_com;
4726         struct lfsck_layout_slave_target *llst = llsaa->llsaa_llst;
4727         struct lfsck_layout_slave_data *llsd = com->lc_data;
4728         struct lfsck_reply *lr = NULL;
4729         bool done = false;
4730
4731         if (rc != 0) {
4732                 /* It is probably caused by network trouble, or target crash,
4733                  * it will try several times (depends on the obd_timeout, and
4734                  * will not less than 3 times). But to make the LFSCK can go
4735                  * ahead, we should not try for ever. After some try but still
4736                  * hit failure, it will assume that the target exit the LFSCK
4737                  * prcoessing and stop try. */
4738                 if (rc == -ENOTCONN || rc == -ESHUTDOWN) {
4739                         int max_try = max_t(int, obd_timeout / 30, 3);
4740
4741                         if (++(llst->llst_failures) > max_try)
4742                                 done = true;
4743                 } else {
4744                         done = true;
4745                 }
4746         } else {
4747                 llst->llst_failures = 0;
4748                 lr = req_capsule_server_get(&req->rq_pill, &RMF_LFSCK_REPLY);
4749                 if (lr->lr_status != LS_SCANNING_PHASE1 &&
4750                     lr->lr_status != LS_SCANNING_PHASE2)
4751                         done = true;
4752         }
4753
4754         if (done) {
4755                 CDEBUG(D_LFSCK, "%s: layout LFSCK slave gets the MDT %x "
4756                        "status %d, failures_try %d\n", lfsck_lfsck2name(com->lc_lfsck),
4757                        llst->llst_index, lr != NULL ? lr->lr_status : rc,
4758                        llst->llst_failures);
4759
4760                 lfsck_layout_llst_del(llsd, llst);
4761         }
4762
4763         lfsck_layout_llst_put(llst);
4764         lfsck_component_put(env, com);
4765         class_export_put(exp);
4766
4767         return 0;
4768 }
4769
4770 static int lfsck_layout_async_query(const struct lu_env *env,
4771                                     struct lfsck_component *com,
4772                                     struct obd_export *exp,
4773                                     struct lfsck_layout_slave_target *llst,
4774                                     struct lfsck_request *lr,
4775                                     struct ptlrpc_request_set *set)
4776 {
4777         struct lfsck_layout_slave_async_args *llsaa;
4778         struct ptlrpc_request                *req;
4779         struct lfsck_request                 *tmp;
4780         int                                   rc;
4781         ENTRY;
4782
4783         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_QUERY);
4784         if (req == NULL)
4785                 RETURN(-ENOMEM);
4786
4787         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_QUERY);
4788         if (rc != 0) {
4789                 ptlrpc_request_free(req);
4790                 RETURN(rc);
4791         }
4792
4793         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
4794         *tmp = *lr;
4795         ptlrpc_request_set_replen(req);
4796
4797         llsaa = ptlrpc_req_async_args(llsaa, req);
4798         llsaa->llsaa_exp = exp;
4799         llsaa->llsaa_com = lfsck_component_get(com);
4800         llsaa->llsaa_llst = llst;
4801         req->rq_interpret_reply = lfsck_layout_slave_async_interpret;
4802         req->rq_allow_intr = 1;
4803         req->rq_no_delay = 1;
4804         ptlrpc_set_add_req(set, req);
4805
4806         RETURN(0);
4807 }
4808
4809 static int lfsck_layout_async_notify(const struct lu_env *env,
4810                                      struct obd_export *exp,
4811                                      struct lfsck_request *lr,
4812                                      struct ptlrpc_request_set *set)
4813 {
4814         struct ptlrpc_request   *req;
4815         struct lfsck_request    *tmp;
4816         int                      rc;
4817         ENTRY;
4818
4819         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
4820         if (req == NULL)
4821                 RETURN(-ENOMEM);
4822
4823         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
4824         if (rc != 0) {
4825                 ptlrpc_request_free(req);
4826                 RETURN(rc);
4827         }
4828
4829         tmp = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
4830         *tmp = *lr;
4831         ptlrpc_request_set_replen(req);
4832         req->rq_allow_intr = 1;
4833         req->rq_no_delay = 1;
4834         ptlrpc_set_add_req(set, req);
4835
4836         RETURN(0);
4837 }
4838
4839 static int
4840 lfsck_layout_slave_query_master(const struct lu_env *env,
4841                                 struct lfsck_component *com)
4842 {
4843         struct lfsck_request             *lr    = &lfsck_env_info(env)->lti_lr;
4844         struct lfsck_instance            *lfsck = com->lc_lfsck;
4845         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
4846         struct lfsck_layout_slave_target *llst;
4847         struct obd_export                *exp;
4848         struct ptlrpc_request_set        *set;
4849         int                               rc    = 0;
4850         int                               rc1   = 0;
4851         ENTRY;
4852
4853         set = ptlrpc_prep_set();
4854         if (set == NULL)
4855                 GOTO(log, rc = -ENOMEM);
4856
4857         memset(lr, 0, sizeof(*lr));
4858         lr->lr_event = LE_QUERY;
4859         lr->lr_active = LFSCK_TYPE_LAYOUT;
4860
4861         llsd->llsd_touch_gen++;
4862         spin_lock(&llsd->llsd_lock);
4863         while (!list_empty(&llsd->llsd_master_list)) {
4864                 llst = list_entry(llsd->llsd_master_list.next,
4865                                   struct lfsck_layout_slave_target,
4866                                   llst_list);
4867                 if (llst->llst_gen == llsd->llsd_touch_gen)
4868                         break;
4869
4870                 llst->llst_gen = llsd->llsd_touch_gen;
4871                 list_move_tail(&llst->llst_list,
4872                                &llsd->llsd_master_list);
4873                 atomic_inc(&llst->llst_ref);
4874                 spin_unlock(&llsd->llsd_lock);
4875
4876                 exp = lustre_find_lwp_by_index(lfsck->li_obd->obd_name,
4877                                                llst->llst_index);
4878                 if (exp == NULL) {
4879                         lfsck_layout_llst_del(llsd, llst);
4880                         lfsck_layout_llst_put(llst);
4881                         spin_lock(&llsd->llsd_lock);
4882                         continue;
4883                 }
4884
4885                 rc = lfsck_layout_async_query(env, com, exp, llst, lr, set);
4886                 if (rc != 0) {
4887                         CDEBUG(D_LFSCK, "%s: layout LFSCK slave fail to "
4888                                "query %s for layout: rc = %d\n",
4889                                lfsck_lfsck2name(lfsck),
4890                                exp->exp_obd->obd_name, rc);
4891
4892                         rc1 = rc;
4893                         lfsck_layout_llst_put(llst);
4894                         class_export_put(exp);
4895                 }
4896                 spin_lock(&llsd->llsd_lock);
4897         }
4898         spin_unlock(&llsd->llsd_lock);
4899
4900         rc = ptlrpc_set_wait(env, set);
4901         ptlrpc_set_destroy(set);
4902
4903         GOTO(log, rc = (rc1 != 0 ? rc1 : rc));
4904
4905 log:
4906         CDEBUG(D_LFSCK, "%s: layout LFSCK slave queries master: rc = %d\n",
4907                lfsck_lfsck2name(com->lc_lfsck), rc);
4908
4909         return rc;
4910 }
4911
4912 static void
4913 lfsck_layout_slave_notify_master(const struct lu_env *env,
4914                                  struct lfsck_component *com,
4915                                  enum lfsck_events event, int result)
4916 {
4917         struct lfsck_layout              *lo    = com->lc_file_ram;
4918         struct lfsck_instance            *lfsck = com->lc_lfsck;
4919         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
4920         struct lfsck_request             *lr    = &lfsck_env_info(env)->lti_lr;
4921         struct lfsck_layout_slave_target *llst;
4922         struct obd_export                *exp;
4923         struct ptlrpc_request_set        *set;
4924         int                               rc;
4925         ENTRY;
4926
4927         CDEBUG(D_LFSCK, "%s: layout LFSCK slave notifies master\n",
4928                lfsck_lfsck2name(com->lc_lfsck));
4929
4930         set = ptlrpc_prep_set();
4931         if (set == NULL)
4932                 RETURN_EXIT;
4933
4934         memset(lr, 0, sizeof(*lr));
4935         lr->lr_event = event;
4936         lr->lr_flags = LEF_FROM_OST;
4937         lr->lr_status = result;
4938         lr->lr_index = lfsck_dev_idx(lfsck);
4939         lr->lr_active = LFSCK_TYPE_LAYOUT;
4940         lr->lr_flags2 = lo->ll_flags;
4941         llsd->llsd_touch_gen++;
4942         spin_lock(&llsd->llsd_lock);
4943         while (!list_empty(&llsd->llsd_master_list)) {
4944                 llst = list_entry(llsd->llsd_master_list.next,
4945                                   struct lfsck_layout_slave_target,
4946                                   llst_list);
4947                 if (llst->llst_gen == llsd->llsd_touch_gen)
4948                         break;
4949
4950                 llst->llst_gen = llsd->llsd_touch_gen;
4951                 list_move_tail(&llst->llst_list,
4952                                &llsd->llsd_master_list);
4953                 atomic_inc(&llst->llst_ref);
4954                 spin_unlock(&llsd->llsd_lock);
4955
4956                 exp = lustre_find_lwp_by_index(lfsck->li_obd->obd_name,
4957                                                llst->llst_index);
4958                 if (exp == NULL) {
4959                         lfsck_layout_llst_del(llsd, llst);
4960                         lfsck_layout_llst_put(llst);
4961                         spin_lock(&llsd->llsd_lock);
4962                         continue;
4963                 }
4964
4965                 rc = lfsck_layout_async_notify(env, exp, lr, set);
4966                 if (rc != 0)
4967                         CDEBUG(D_LFSCK, "%s: layout LFSCK slave fail to "
4968                                "notify %s for layout: rc = %d\n",
4969                                lfsck_lfsck2name(lfsck),
4970                                exp->exp_obd->obd_name, rc);
4971
4972                 lfsck_layout_llst_put(llst);
4973                 class_export_put(exp);
4974                 spin_lock(&llsd->llsd_lock);
4975         }
4976         spin_unlock(&llsd->llsd_lock);
4977
4978         ptlrpc_set_wait(env, set);
4979         ptlrpc_set_destroy(set);
4980
4981         RETURN_EXIT;
4982 }
4983
4984 /*
4985  * \ret -ENODATA: unrecognized stripe
4986  * \ret = 0     : recognized stripe
4987  * \ret < 0     : other failures
4988  */
4989 static int lfsck_layout_master_check_pairs(const struct lu_env *env,
4990                                            struct lfsck_component *com,
4991                                            struct lu_fid *cfid,
4992                                            struct lu_fid *pfid, __u32 comp_id)
4993 {
4994         struct lfsck_thread_info        *info   = lfsck_env_info(env);
4995         struct lu_buf                   *buf    = &info->lti_big_buf;
4996         struct ost_id                   *oi     = &info->lti_oi;
4997         struct dt_object                *obj;
4998         struct lov_mds_md_v1            *lmm;
4999         struct lov_ost_data_v1          *objs;
5000         __u32                            idx    = pfid->f_stripe_idx;
5001         __u32                            magic;
5002         int                              rc     = 0;
5003         int                              i;
5004         __u16                            count;
5005         ENTRY;
5006
5007         pfid->f_ver = 0;
5008         obj = lfsck_object_find_bottom(env, com->lc_lfsck, pfid);
5009         if (IS_ERR(obj))
5010                 RETURN(PTR_ERR(obj));
5011
5012         dt_read_lock(env, obj, 0);
5013         if (unlikely(dt_object_exists(obj) == 0 ||
5014                      lfsck_is_dead_obj(obj)))
5015                 GOTO(unlock, rc = -ENOENT);
5016
5017         if (!S_ISREG(lfsck_object_type(obj)))
5018                 GOTO(unlock, rc = -ENODATA);
5019
5020         rc = lfsck_layout_get_lovea(env, obj, buf);
5021         if (rc < 0)
5022                 GOTO(unlock, rc);
5023
5024         lmm = buf->lb_buf;
5025         magic = le32_to_cpu(lmm->lmm_magic);
5026         if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
5027                 struct lov_comp_md_v1 *lcm = buf->lb_buf;
5028                 struct lov_comp_md_entry_v1 *lcme;
5029
5030                 if (comp_id == 0)
5031                         GOTO(unlock, rc = -ENODATA);
5032
5033                 count = le16_to_cpu(lcm->lcm_entry_count);
5034                 for (i = 0; i < count; i++) {
5035                         lcme = &lcm->lcm_entries[i];
5036                         if (le32_to_cpu(lcme->lcme_id) == comp_id) {
5037                                 lmm = buf->lb_buf +
5038                                         le32_to_cpu(lcme->lcme_offset);
5039                                 magic = le32_to_cpu(lmm->lmm_magic);
5040                                 if (!(le32_to_cpu(lcme->lcme_flags) &
5041                                       LCME_FL_INIT))
5042                                         GOTO(unlock, rc = -ENODATA);
5043
5044                                 goto further;
5045                         }
5046                 }
5047
5048                 GOTO(unlock, rc = -ENODATA);
5049         }
5050
5051 further:
5052         if (magic == LOV_MAGIC_V1) {
5053                 objs = &lmm->lmm_objects[0];
5054         } else {
5055                 LASSERT(magic == LOV_MAGIC_V3);
5056                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
5057         }
5058
5059         fid_to_ostid(cfid, oi);
5060         count = le16_to_cpu(lmm->lmm_stripe_count);
5061         for (i = 0; i < count; i++, objs++) {
5062                 struct ost_id oi2;
5063
5064                 ostid_le_to_cpu(&objs->l_ost_oi, &oi2);
5065                 if (memcmp(oi, &oi2, sizeof(*oi)) == 0)
5066                         GOTO(unlock, rc = (i != idx ? -ENODATA : 0));
5067         }
5068
5069         GOTO(unlock, rc = -ENODATA);
5070
5071 unlock:
5072         dt_read_unlock(env, obj);
5073         lfsck_object_put(env, obj);
5074
5075         return rc;
5076 }
5077
5078 /*
5079  * The LFSCK-on-OST will ask the LFSCK-on-MDT to check whether the given
5080  * MDT-object/OST-object pairs match or not to aviod transfer MDT-object
5081  * layout EA from MDT to OST. On one hand, the OST no need to understand
5082  * the layout EA structure; on the other hand, it may cause trouble when
5083  * transfer large layout EA from MDT to OST via normal OUT RPC.
5084  *
5085  * \ret > 0: unrecognized stripe
5086  * \ret = 0: recognized stripe
5087  * \ret < 0: other failures
5088  */
5089 static int lfsck_layout_slave_check_pairs(const struct lu_env *env,
5090                                           struct lfsck_component *com,
5091                                           struct lu_fid *cfid,
5092                                           struct lu_fid *pfid, __u32 comp_id)
5093 {
5094         struct lfsck_instance    *lfsck  = com->lc_lfsck;
5095         struct obd_device        *obd    = lfsck->li_obd;
5096         struct seq_server_site   *ss     = lfsck_dev_site(lfsck);
5097         struct obd_export        *exp    = NULL;
5098         struct ptlrpc_request    *req    = NULL;
5099         struct lfsck_request     *lr;
5100         struct lu_seq_range      *range  = &lfsck_env_info(env)->lti_range;
5101         int                       rc     = 0;
5102         ENTRY;
5103
5104         if (unlikely(fid_is_idif(pfid)))
5105                 RETURN(1);
5106
5107         fld_range_set_any(range);
5108         rc = fld_server_lookup(env, ss->ss_server_fld, fid_seq(pfid), range);
5109         if (rc != 0)
5110                 RETURN(rc == -ENOENT ? 1 : rc);
5111
5112         if (unlikely(!fld_range_is_mdt(range)))
5113                 RETURN(1);
5114
5115         exp = lustre_find_lwp_by_index(obd->obd_name, range->lsr_index);
5116         if (unlikely(exp == NULL))
5117                 RETURN(1);
5118
5119         if (!(exp_connect_flags(exp) & OBD_CONNECT_LFSCK))
5120                 GOTO(out, rc = -EOPNOTSUPP);
5121
5122         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_LFSCK_NOTIFY);
5123         if (req == NULL)
5124                 GOTO(out, rc = -ENOMEM);
5125
5126         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, LFSCK_NOTIFY);
5127         if (rc != 0) {
5128                 ptlrpc_request_free(req);
5129
5130                 GOTO(out, rc);
5131         }
5132
5133         lr = req_capsule_client_get(&req->rq_pill, &RMF_LFSCK_REQUEST);
5134         memset(lr, 0, sizeof(*lr));
5135         lr->lr_event = LE_PAIRS_VERIFY;
5136         lr->lr_active = LFSCK_TYPE_LAYOUT;
5137         lr->lr_fid = *cfid; /* OST-object itself FID. */
5138         lr->lr_fid2 = *pfid; /* The claimed parent FID. */
5139         lr->lr_comp_id = comp_id;
5140
5141         ptlrpc_request_set_replen(req);
5142         rc = ptlrpc_queue_wait(req);
5143         ptlrpc_req_finished(req);
5144
5145         if (rc == -ENOENT || rc == -ENODATA)
5146                 rc = 1;
5147
5148         GOTO(out, rc);
5149
5150 out:
5151         if (exp != NULL)
5152                 class_export_put(exp);
5153
5154         return rc;
5155 }
5156
5157 static int lfsck_layout_slave_repair_pfid(const struct lu_env *env,
5158                                           struct lfsck_component *com,
5159                                           struct lfsck_req_local *lrl)
5160 {
5161         struct dt_object        *obj;
5162         int                      rc     = 0;
5163         ENTRY;
5164
5165         obj = lfsck_object_find_bottom(env, com->lc_lfsck, &lrl->lrl_fid);
5166         if (IS_ERR(obj))
5167                 GOTO(log, rc = PTR_ERR(obj));
5168
5169         dt_write_lock(env, obj, 0);
5170         if (unlikely(dt_object_exists(obj) == 0 ||
5171                      lfsck_is_dead_obj(obj)))
5172                 GOTO(unlock, rc = 0);
5173
5174         rc = __lfsck_layout_update_pfid(env, obj, &lrl->lrl_ff_client.ff_parent,
5175                                         &lrl->lrl_ff_client.ff_layout,
5176                                         lrl->lrl_ff_client.ff_layout_version,
5177                                         lrl->lrl_ff_client.ff_range,
5178                                         lrl->lrl_ff_client.ff_parent.f_ver);
5179
5180         GOTO(unlock, rc);
5181
5182 unlock:
5183         dt_write_unlock(env, obj);
5184         lfsck_object_put(env, obj);
5185
5186 log:
5187         CDEBUG(D_LFSCK, "%s: layout LFSCK slave repaired pfid for "DFID
5188                ", parent "DFID": rc = %d\n", lfsck_lfsck2name(com->lc_lfsck),
5189                PFID(&lrl->lrl_fid), PFID(&lrl->lrl_ff_client.ff_parent), rc);
5190
5191         return rc;
5192 }
5193
5194 /* layout APIs */
5195
5196 static void lfsck_layout_slave_quit(const struct lu_env *env,
5197                                     struct lfsck_component *com);
5198
5199 static int lfsck_layout_reset(const struct lu_env *env,
5200                               struct lfsck_component *com, bool init)
5201 {
5202         struct lfsck_layout     *lo    = com->lc_file_ram;
5203         int                      rc;
5204
5205         down_write(&com->lc_sem);
5206         if (init) {
5207                 memset(lo, 0, com->lc_file_size);
5208         } else {
5209                 __u32 count = lo->ll_success_count;
5210                 time64_t last_time = lo->ll_time_last_complete;
5211
5212                 memset(lo, 0, com->lc_file_size);
5213                 lo->ll_success_count = count;
5214                 lo->ll_time_last_complete = last_time;
5215         }
5216
5217         lo->ll_magic = LFSCK_LAYOUT_MAGIC;
5218         lo->ll_status = LS_INIT;
5219
5220         if (com->lc_lfsck->li_master) {
5221                 struct lfsck_assistant_data *lad = com->lc_data;
5222
5223                 clear_bit(LAD_INCOMPLETE, &lad->lad_flags);
5224                 CFS_RESET_BITMAP(lad->lad_bitmap);
5225         }
5226
5227         rc = lfsck_layout_store(env, com);
5228         if (rc == 0 && com->lc_lfsck->li_master)
5229                 rc = lfsck_load_sub_trace_files(env, com,
5230                         &dt_lfsck_layout_dangling_features, LFSCK_LAYOUT, true);
5231         up_write(&com->lc_sem);
5232
5233         CDEBUG(D_LFSCK, "%s: layout LFSCK reset: rc = %d\n",
5234                lfsck_lfsck2name(com->lc_lfsck), rc);
5235
5236         return rc;
5237 }
5238
5239 static void lfsck_layout_fail(const struct lu_env *env,
5240                               struct lfsck_component *com, bool new_checked)
5241 {
5242         struct lfsck_layout *lo = com->lc_file_ram;
5243
5244         down_write(&com->lc_sem);
5245         if (new_checked)
5246                 com->lc_new_checked++;
5247         lfsck_layout_record_failure(env, com->lc_lfsck, lo);
5248         up_write(&com->lc_sem);
5249 }
5250
5251 static int lfsck_layout_master_checkpoint(const struct lu_env *env,
5252                                           struct lfsck_component *com, bool init)
5253 {
5254         struct lfsck_instance   *lfsck   = com->lc_lfsck;
5255         struct lfsck_layout     *lo      = com->lc_file_ram;
5256         int                      rc;
5257
5258         if (!init) {
5259                 rc = lfsck_checkpoint_generic(env, com);
5260                 if (rc != 0)
5261                         return rc > 0 ? 0 : rc;
5262         }
5263
5264         down_write(&com->lc_sem);
5265         if (init) {
5266                 lo->ll_pos_latest_start =
5267                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5268         } else {
5269                 lo->ll_pos_last_checkpoint =
5270                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5271                 lo->ll_run_time_phase1 += ktime_get_seconds() -
5272                                           lfsck->li_time_last_checkpoint;
5273                 lo->ll_time_last_checkpoint = ktime_get_real_seconds();
5274                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5275                 com->lc_new_checked = 0;
5276         }
5277
5278         rc = lfsck_layout_store(env, com);
5279         up_write(&com->lc_sem);
5280
5281         CDEBUG(D_LFSCK, "%s: layout LFSCK master checkpoint at the pos ["
5282                "%llu], status = %d: rc = %d\n", lfsck_lfsck2name(lfsck),
5283                lfsck->li_pos_current.lp_oit_cookie, lo->ll_status, rc);
5284
5285         return rc;
5286 }
5287
5288 static int lfsck_layout_slave_checkpoint(const struct lu_env *env,
5289                                          struct lfsck_component *com, bool init)
5290 {
5291         struct lfsck_instance   *lfsck = com->lc_lfsck;
5292         struct lfsck_layout     *lo    = com->lc_file_ram;
5293         int                      rc;
5294
5295         if (com->lc_new_checked == 0 && !init)
5296                 return 0;
5297
5298         down_write(&com->lc_sem);
5299         if (init) {
5300                 lo->ll_pos_latest_start =
5301                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5302         } else {
5303                 lo->ll_pos_last_checkpoint =
5304                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5305                 lo->ll_run_time_phase1 += ktime_get_seconds() -
5306                                           lfsck->li_time_last_checkpoint;
5307                 lo->ll_time_last_checkpoint = ktime_get_real_seconds();
5308                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
5309                 com->lc_new_checked = 0;
5310         }
5311
5312         rc = lfsck_layout_store(env, com);
5313         up_write(&com->lc_sem);
5314
5315         CDEBUG(D_LFSCK, "%s: layout LFSCK slave checkpoint at the pos ["
5316                "%llu], status = %d: rc = %d\n", lfsck_lfsck2name(lfsck),
5317                lfsck->li_pos_current.lp_oit_cookie, lo->ll_status, rc);
5318
5319         return rc;
5320 }
5321
5322 static int lfsck_layout_prep(const struct lu_env *env,
5323                              struct lfsck_component *com,
5324                              struct lfsck_start *start)
5325 {
5326         struct lfsck_instance   *lfsck  = com->lc_lfsck;
5327         struct lfsck_layout     *lo     = com->lc_file_ram;
5328         struct lfsck_position   *pos    = &com->lc_pos_start;
5329
5330         fid_zero(&pos->lp_dir_parent);
5331         pos->lp_dir_cookie = 0;
5332         if (lo->ll_status == LS_COMPLETED ||
5333             lo->ll_status == LS_PARTIAL ||
5334             /* To handle orphan, must scan from the beginning. */
5335             (start != NULL && start->ls_flags & LPF_OST_ORPHAN)) {
5336                 int rc;
5337
5338                 rc = lfsck_layout_reset(env, com, false);
5339                 if (rc == 0)
5340                         rc = lfsck_set_param(env, lfsck, start, true);
5341
5342                 if (rc != 0) {
5343                         CDEBUG(D_LFSCK, "%s: layout LFSCK prep failed: "
5344                                "rc = %d\n", lfsck_lfsck2name(lfsck), rc);
5345
5346                         return rc;
5347                 }
5348         }
5349
5350         down_write(&com->lc_sem);
5351         lo->ll_time_latest_start = ktime_get_real_seconds();
5352         spin_lock(&lfsck->li_lock);
5353         if (lo->ll_flags & LF_SCANNED_ONCE) {
5354                 if (!lfsck->li_drop_dryrun ||
5355                     lo->ll_pos_first_inconsistent == 0) {
5356                         lo->ll_status = LS_SCANNING_PHASE2;
5357                         list_move_tail(&com->lc_link,
5358                                        &lfsck->li_list_double_scan);
5359                         pos->lp_oit_cookie = 0;
5360                 } else {
5361                         int i;
5362
5363                         lo->ll_status = LS_SCANNING_PHASE1;
5364                         lo->ll_run_time_phase1 = 0;
5365                         lo->ll_run_time_phase2 = 0;
5366                         lo->ll_objs_checked_phase1 = 0;
5367                         lo->ll_objs_checked_phase2 = 0;
5368                         lo->ll_objs_failed_phase1 = 0;
5369                         lo->ll_objs_failed_phase2 = 0;
5370                         for (i = 0; i < LLIT_MAX; i++)
5371                                 lo->ll_objs_repaired[i] = 0;
5372
5373                         pos->lp_oit_cookie = lo->ll_pos_first_inconsistent;
5374                         fid_zero(&com->lc_fid_latest_scanned_phase2);
5375                 }
5376         } else {
5377                 lo->ll_status = LS_SCANNING_PHASE1;
5378                 if (!lfsck->li_drop_dryrun ||
5379                     lo->ll_pos_first_inconsistent == 0)
5380                         pos->lp_oit_cookie = lo->ll_pos_last_checkpoint + 1;
5381                 else
5382                         pos->lp_oit_cookie = lo->ll_pos_first_inconsistent;
5383         }
5384         spin_unlock(&lfsck->li_lock);
5385         up_write(&com->lc_sem);
5386
5387         return 0;
5388 }
5389
5390 static int lfsck_layout_slave_prep(const struct lu_env *env,
5391                                    struct lfsck_component *com,
5392                                    struct lfsck_start_param *lsp)
5393 {
5394         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
5395         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5396         struct lfsck_layout             *lo     = com->lc_file_ram;
5397         struct lfsck_start              *start  = lsp->lsp_start;
5398         int                              rc;
5399
5400         rc = lfsck_layout_prep(env, com, start);
5401         if (rc != 0)
5402                 return rc;
5403
5404         if (lo->ll_flags & LF_CRASHED_LASTID &&
5405             list_empty(&llsd->llsd_master_list)) {
5406                 LASSERT(lfsck->li_out_notify != NULL);
5407
5408                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
5409                                      LE_LASTID_REBUILDING);
5410         }
5411
5412         if (!lsp->lsp_index_valid)
5413                 return 0;
5414
5415         rc = lfsck_layout_llst_add(llsd, lsp->lsp_index);
5416         if (rc == 0 && start != NULL && start->ls_flags & LPF_OST_ORPHAN) {
5417                 LASSERT(!llsd->llsd_rbtree_valid);
5418
5419                 down_write(&llsd->llsd_rb_rwsem);
5420                 rc = lfsck_rbtree_setup(env, com);
5421                 up_write(&llsd->llsd_rb_rwsem);
5422         }
5423
5424         CDEBUG(D_LFSCK, "%s: layout LFSCK slave prep done, start pos ["
5425                "%llu]\n", lfsck_lfsck2name(lfsck),
5426                com->lc_pos_start.lp_oit_cookie);
5427
5428         return rc;
5429 }
5430
5431 static int lfsck_layout_master_prep(const struct lu_env *env,
5432                                     struct lfsck_component *com,
5433                                     struct lfsck_start_param *lsp)
5434 {
5435         int rc;
5436         ENTRY;
5437
5438         rc = lfsck_layout_load_bitmap(env, com);
5439         if (rc != 0) {
5440                 rc = lfsck_layout_reset(env, com, false);
5441                 if (rc == 0)
5442                         rc = lfsck_set_param(env, com->lc_lfsck,
5443                                              lsp->lsp_start, true);
5444
5445                 if (rc != 0)
5446                         GOTO(log, rc);
5447         }
5448
5449         rc = lfsck_layout_prep(env, com, lsp->lsp_start);
5450         if (rc != 0)
5451                 RETURN(rc);
5452
5453         rc = lfsck_start_assistant(env, com, lsp);
5454
5455         GOTO(log, rc);
5456
5457 log:
5458         CDEBUG(D_LFSCK, "%s: layout LFSCK master prep done, start pos ["
5459                "%llu]\n", lfsck_lfsck2name(com->lc_lfsck),
5460                com->lc_pos_start.lp_oit_cookie);
5461
5462         return 0;
5463 }
5464
5465 /* Pre-fetch the attribute for each stripe in the given layout EA. */
5466 static int lfsck_layout_scan_stripes(const struct lu_env *env,
5467                                      struct lfsck_component *com,
5468                                      struct dt_object *parent,
5469                                      struct lov_mds_md_v1 *lmm, __u32 comp_id)
5470 {
5471         struct lfsck_thread_info        *info    = lfsck_env_info(env);
5472         struct lfsck_instance           *lfsck   = com->lc_lfsck;
5473         struct lfsck_bookmark           *bk      = &lfsck->li_bookmark_ram;
5474         struct lfsck_layout             *lo      = com->lc_file_ram;
5475         struct lfsck_assistant_data     *lad     = com->lc_data;
5476         struct lfsck_assistant_object   *lso     = NULL;
5477         struct lov_ost_data_v1          *objs;
5478         struct lfsck_tgt_descs          *ltds    = &lfsck->li_ost_descs;
5479         struct ptlrpc_thread            *mthread = &lfsck->li_thread;
5480         struct ptlrpc_thread            *athread = &lad->lad_thread;
5481         struct lu_buf                    buf;
5482         int                              rc      = 0;
5483         int                              i;
5484         __u32                            magic;
5485         __u16                            count;
5486         ENTRY;
5487
5488         lfsck_buf_init(&buf, &info->lti_ff, sizeof(struct filter_fid));
5489         magic = le32_to_cpu(lmm->lmm_magic);
5490         if (magic == LOV_MAGIC_V1) {
5491                 objs = &lmm->lmm_objects[0];
5492         } else {
5493                 LASSERT(magic == LOV_MAGIC_V3);
5494                 objs = &((struct lov_mds_md_v3 *)lmm)->lmm_objects[0];
5495         }
5496
5497         count = le16_to_cpu(lmm->lmm_stripe_count);
5498         for (i = 0; i < count; i++, objs++) {
5499                 struct lu_fid           *fid    = &info->lti_fid;
5500                 struct ost_id           *oi     = &info->lti_oi;
5501                 struct lfsck_layout_req *llr;
5502                 struct lfsck_tgt_desc   *tgt    = NULL;
5503                 struct dt_object        *cobj   = NULL;
5504                 __u32                    index;
5505                 bool                     wakeup = false;
5506
5507                 if (unlikely(lovea_slot_is_dummy(objs)))
5508                         continue;
5509
5510                 wait_event_idle(mthread->t_ctl_waitq,
5511                                 lad->lad_prefetched < bk->lb_async_windows ||
5512                                 !thread_is_running(mthread) ||
5513                                 thread_is_stopped(athread));
5514
5515                 if (unlikely(!thread_is_running(mthread)) ||
5516                              thread_is_stopped(athread))
5517                         GOTO(out, rc = 0);
5518
5519                 if (unlikely(lfsck_is_dead_obj(parent)))
5520                         GOTO(out, rc = 0);
5521
5522                 ostid_le_to_cpu(&objs->l_ost_oi, oi);
5523                 index = le32_to_cpu(objs->l_ost_idx);
5524                 rc = ostid_to_fid(fid, oi, index);
5525                 if (rc != 0) {
5526                         CDEBUG(D_LFSCK, "%s: get invalid layout EA for "DFID
5527                                ": "DOSTID", idx %u, comp_id %u\n",
5528                                lfsck_lfsck2name(lfsck),
5529                                PFID(lfsck_dto2fid(parent)), POSTID(oi),
5530                                index, comp_id);
5531                         goto next;
5532                 }
5533
5534                 tgt = lfsck_tgt_get(ltds, index);
5535                 if (unlikely(tgt == NULL)) {
5536                         CDEBUG(D_LFSCK, "%s: cannot talk with OST %x which "
5537                                "did not join the layout LFSCK, comp_id %u\n",
5538                                lfsck_lfsck2name(lfsck), index, comp_id);
5539                         lfsck_lad_set_bitmap(env, com, index);
5540                         goto next;
5541                 }
5542
5543                 /* There is potential deadlock race condition between object
5544                  * destroy and layout LFSCK. Consider the following scenario:
5545                  *
5546                  * 1) The LFSCK thread obtained the parent object firstly, at
5547                  *    that time, the parent object has not been destroyed yet.
5548                  *
5549                  * 2) One RPC service thread destroyed the parent and all its
5550                  *    children objects. Because the LFSCK is referencing the
5551                  *    parent object, then the parent object will be marked as
5552                  *    dying in RAM. On the other hand, the parent object is
5553                  *    referencing all its children objects, then all children
5554                  *    objects will be marked as dying in RAM also.
5555                  *
5556                  * 3) The LFSCK thread tries to find some child object with
5557                  *    the parent object referenced. Then it will find that the
5558                  *    child object is dying. According to the object visibility
5559                  *    rules: the object with dying flag cannot be returned to
5560                  *    others. So the LFSCK thread has to wait until the dying
5561                  *    object has been purged from RAM, then it can allocate a
5562                  *    new object (with the same FID) in RAM. Unfortunately, the
5563                  *    LFSCK thread itself is referencing the parent object, and
5564                  *    cause the parent object cannot be purged, then cause the
5565                  *    child object cannot be purged also. So the LFSCK thread
5566                  *    will fall into deadlock.
5567                  */
5568                 cobj = lfsck_object_find_by_dev(env, tgt->ltd_tgt, fid);
5569                 if (IS_ERR(cobj)) {
5570                         if (lfsck_is_dead_obj(parent)) {
5571                                 lfsck_tgt_put(tgt);
5572
5573                                 GOTO(out, rc = 0);
5574                         }
5575
5576                         rc = PTR_ERR(cobj);
5577                         goto next;
5578                 }
5579
5580                 rc = dt_declare_attr_get(env, cobj);
5581                 if (rc)
5582                         goto next;
5583
5584                 rc = dt_declare_xattr_get(env, cobj, &buf, XATTR_NAME_FID);
5585                 if (rc)
5586                         goto next;
5587
5588                 if (lso == NULL) {
5589                         struct lu_attr *attr = &info->lti_la;
5590
5591                         rc = dt_attr_get(env, parent, attr);
5592                         if (rc != 0)
5593                                 goto next;
5594
5595                         lso = lfsck_assistant_object_init(env,
5596                                 lfsck_dto2fid(parent), attr,
5597                                 lfsck->li_pos_current.lp_oit_cookie, false);
5598                         if (IS_ERR(lso)) {
5599                                 rc = PTR_ERR(lso);
5600                                 lso = NULL;
5601
5602                                 goto next;
5603                         }
5604                 }
5605
5606                 llr = lfsck_layout_assistant_req_init(lso, cobj, comp_id,
5607                                                       index, i);
5608                 if (IS_ERR(llr)) {
5609                         rc = PTR_ERR(llr);
5610                         goto next;
5611                 }
5612
5613                 cobj = NULL;
5614                 spin_lock(&lad->lad_lock);
5615                 if (lad->lad_assistant_status < 0) {
5616                         spin_unlock(&lad->lad_lock);
5617                         lfsck_layout_assistant_req_fini(env, &llr->llr_lar);
5618                         lfsck_tgt_put(tgt);
5619                         RETURN(lad->lad_assistant_status);
5620                 }
5621
5622                 list_add_tail(&llr->llr_lar.lar_list, &lad->lad_req_list);
5623                 if (lad->lad_prefetched == 0)
5624                         wakeup = true;
5625
5626                 lad->lad_prefetched++;
5627                 spin_unlock(&lad->lad_lock);
5628                 if (wakeup)
5629                         wake_up(&athread->t_ctl_waitq);
5630
5631 next:
5632                 down_write(&com->lc_sem);
5633                 com->lc_new_checked++;
5634                 if (rc < 0)
5635                         lfsck_layout_record_failure(env, lfsck, lo);
5636                 up_write(&com->lc_sem);
5637
5638                 if (cobj != NULL && !IS_ERR(cobj))
5639                         lfsck_object_put(env, cobj);
5640
5641                 if (likely(tgt != NULL))
5642                         lfsck_tgt_put(tgt);
5643
5644                 if (rc < 0 && bk->lb_param & LPF_FAILOUT)
5645                         GOTO(out, rc);
5646         }
5647
5648         GOTO(out, rc = 0);
5649
5650 out:
5651         if (lso != NULL)
5652                 lfsck_assistant_object_put(env, lso);
5653
5654         return rc;
5655 }
5656
5657 /* For the given object, read its layout EA locally. For each stripe, pre-fetch
5658  * the OST-object's attribute and generate an structure lfsck_layout_req on the
5659  * list ::lad_req_list.
5660  *
5661  * For each request on above list, the lfsck_layout_assistant thread compares
5662  * the OST side attribute with local attribute, if inconsistent, then repair it.
5663  *
5664  * All above processing is async mode with pipeline. */
5665 static int lfsck_layout_master_exec_oit(const struct lu_env *env,
5666                                         struct lfsck_component *com,
5667                                         struct dt_object *obj)
5668 {
5669         struct lfsck_thread_info        *info   = lfsck_env_info(env);
5670         struct ost_id                   *oi     = &info->lti_oi;
5671         struct lfsck_layout             *lo     = com->lc_file_ram;
5672         struct lfsck_assistant_data     *lad    = com->lc_data;
5673         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5674         struct lfsck_bookmark           *bk     = &lfsck->li_bookmark_ram;
5675         struct thandle                  *handle = NULL;
5676         struct lu_buf                   *buf    = &info->lti_big_buf;
5677         struct lov_mds_md_v1            *lmm    = NULL;
5678         struct dt_device                *dev    = lfsck_obj2dev(obj);
5679         struct lustre_handle             lh     = { 0 };
5680         struct lu_buf                    ea_buf = { NULL };
5681         struct lov_comp_md_v1           *lcm    = NULL;
5682         struct lov_comp_md_entry_v1     *lcme   = NULL;
5683         int                              rc     = 0;
5684         int                              size   = 0;
5685         __u32                            magic  = 0;
5686         __u16                            count  = 0;
5687         bool                             locked = false;
5688         bool                             stripe = false;
5689         bool                             bad_oi = false;
5690         ENTRY;
5691
5692         if (!S_ISREG(lfsck_object_type(obj)))
5693                 GOTO(out, rc = 0);
5694
5695         if (lad->lad_assistant_status < 0)
5696                 GOTO(out, rc = -ESRCH);
5697
5698         fid_to_lmm_oi(lfsck_dto2fid(obj), oi);
5699         lmm_oi_cpu_to_le(oi, oi);
5700         dt_read_lock(env, obj, 0);
5701         locked = true;
5702
5703 again:
5704         bad_oi = false;
5705         if (dt_object_exists(obj) == 0 ||
5706             lfsck_is_dead_obj(obj))
5707                 GOTO(out, rc = 0);
5708
5709         rc = lfsck_layout_get_lovea(env, obj, buf);
5710         if (rc == -EINVAL || rc == -ENODATA || rc == -EOPNOTSUPP)
5711                 /* Skip bad lov EA during the 1st cycle scanning, and
5712                  * try to recover it via orphan in the 2nd scanning. */
5713                 rc = 0;
5714         if (rc <= 0)
5715                 GOTO(out, rc);
5716
5717         size = rc;
5718         lmm = buf->lb_buf;
5719         magic = le32_to_cpu(lmm->lmm_magic);
5720         if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
5721                 struct lov_mds_md_v1 *v1;
5722                 int i;
5723
5724                 lcm = buf->lb_buf;
5725                 count = le16_to_cpu(lcm->lcm_entry_count);
5726                 for (i = 0; i < count; i++) {
5727                         lcme = &lcm->lcm_entries[i];
5728                         v1 = buf->lb_buf + le32_to_cpu(lcme->lcme_offset);
5729                         if (memcmp(oi, &v1->lmm_oi, sizeof(*oi)) != 0)
5730                                 goto fix;
5731                 }
5732
5733                 GOTO(out, stripe = true);
5734         } else if (memcmp(oi, &lmm->lmm_oi, sizeof(*oi)) == 0) {
5735                 GOTO(out, stripe = true);
5736         }
5737
5738 fix:
5739         /* Inconsistent lmm_oi, should be repaired. */
5740         bad_oi = true;
5741
5742         if (bk->lb_param & LPF_DRYRUN) {
5743                 lo->ll_objs_repaired[LLIT_OTHERS - 1]++;
5744
5745                 GOTO(out, stripe = true);
5746         }
5747
5748         if (!lustre_handle_is_used(&lh)) {
5749                 dt_read_unlock(env, obj);
5750                 locked = false;
5751                 rc = lfsck_ibits_lock(env, lfsck, obj, &lh,
5752                                       MDS_INODELOCK_LAYOUT |
5753                                       MDS_INODELOCK_XATTR, LCK_EX);
5754                 if (rc != 0)
5755                         GOTO(out, rc);
5756
5757                 handle = dt_trans_create(env, dev);
5758                 if (IS_ERR(handle))
5759                         GOTO(out, rc = PTR_ERR(handle));
5760
5761                 lfsck_buf_init(&ea_buf, buf->lb_buf, size);
5762                 rc = dt_declare_xattr_set(env, obj, &ea_buf, XATTR_NAME_LOV,
5763                                           LU_XATTR_REPLACE, handle);
5764                 if (rc != 0)
5765                         GOTO(out, rc);
5766
5767                 rc = dt_trans_start_local(env, dev, handle);
5768                 if (rc != 0)
5769                         GOTO(out, rc);
5770
5771                 dt_write_lock(env, obj, 0);
5772                 locked = true;
5773
5774                 goto again;
5775         }
5776
5777         if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
5778                 struct lov_mds_md_v1 *v1;
5779                 int i;
5780
5781                 for (i = 0; i < count; i++) {
5782                         lcme = &lcm->lcm_entries[i];
5783                         v1 = buf->lb_buf + le32_to_cpu(lcme->lcme_offset);
5784                         v1->lmm_oi = *oi;
5785                 }
5786         } else {
5787                 lmm->lmm_oi = *oi;
5788         }
5789
5790         rc = dt_xattr_set(env, obj, &ea_buf, XATTR_NAME_LOV,
5791                           LU_XATTR_REPLACE, handle);
5792         if (rc != 0)
5793                 GOTO(out, rc);
5794
5795         lo->ll_objs_repaired[LLIT_OTHERS - 1]++;
5796
5797         GOTO(out, stripe = true);
5798
5799 out:
5800         if (locked) {
5801                 if (lustre_handle_is_used(&lh))
5802                         dt_write_unlock(env, obj);
5803                 else
5804                         dt_read_unlock(env, obj);
5805         }
5806
5807         if (handle != NULL && !IS_ERR(handle))
5808                 dt_trans_stop(env, dev, handle);
5809
5810         lfsck_ibits_unlock(&lh, LCK_EX);
5811
5812         if (bad_oi)
5813                 CDEBUG(D_LFSCK, "%s: layout LFSCK master %s bad lmm_oi for "
5814                        DFID": rc = %d\n", lfsck_lfsck2name(lfsck),
5815                        bk->lb_param & LPF_DRYRUN ? "found" : "repaired",
5816                        PFID(lfsck_dto2fid(obj)), rc);
5817
5818         if (stripe) {
5819                 if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
5820                         int i;
5821
5822                         for (i = 0; i < count; i++) {
5823                                 lcme = &lcm->lcm_entries[i];
5824                                 if (!(le32_to_cpu(lcme->lcme_flags) &
5825                                       LCME_FL_INIT))
5826                                         continue;
5827
5828                                 rc = lfsck_layout_scan_stripes(env, com, obj,
5829                                         (struct lov_mds_md_v1 *)(buf->lb_buf +
5830                                         le32_to_cpu(lcme->lcme_offset)),
5831                                         le32_to_cpu(lcme->lcme_id));
5832                         }
5833                 } else {
5834                         rc = lfsck_layout_scan_stripes(env, com, obj, lmm, 0);
5835                 }
5836         } else {
5837                 down_write(&com->lc_sem);
5838                 com->lc_new_checked++;
5839                 if (rc < 0)
5840                         lfsck_layout_record_failure(env, lfsck, lo);
5841                 up_write(&com->lc_sem);
5842         }
5843
5844         return rc;
5845 }
5846
5847 static int lfsck_layout_slave_exec_oit(const struct lu_env *env,
5848                                        struct lfsck_component *com,
5849                                        struct dt_object *obj)
5850 {
5851         struct lfsck_instance           *lfsck  = com->lc_lfsck;
5852         struct lfsck_layout             *lo     = com->lc_file_ram;
5853         const struct lu_fid             *fid    = lfsck_dto2fid(obj);
5854         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
5855         struct lfsck_layout_seq         *lls;
5856         __u64                            seq;
5857         __u64                            oid;
5858         int                              rc;
5859         ENTRY;
5860
5861         LASSERT(llsd != NULL);
5862
5863         if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DELAY5) &&
5864             cfs_fail_val == lfsck_dev_idx(lfsck)) {
5865                 struct ptlrpc_thread    *thread = &lfsck->li_thread;
5866
5867                 wait_event_idle_timeout(thread->t_ctl_waitq,
5868                                         !thread_is_running(thread),
5869                                         cfs_time_seconds(1));
5870         }
5871
5872         lfsck_rbtree_update_bitmap(env, com, fid, false);
5873
5874         down_write(&com->lc_sem);
5875         if (fid_is_idif(fid))
5876                 seq = 0;
5877         else if (!fid_is_norm(fid) ||
5878                  !fid_is_for_ostobj(env, lfsck, obj, fid))
5879                 GOTO(unlock, rc = 0);
5880         else
5881                 seq = fid_seq(fid);
5882         com->lc_new_checked++;
5883
5884         lls = lfsck_layout_seq_lookup(llsd, seq);
5885         if (lls == NULL) {
5886                 OBD_ALLOC_PTR(lls);
5887                 if (unlikely(lls == NULL))
5888                         GOTO(unlock, rc = -ENOMEM);
5889
5890                 INIT_LIST_HEAD(&lls->lls_list);
5891                 lls->lls_seq = seq;
5892                 rc = lfsck_layout_lastid_load(env, com, lls);
5893                 if (rc != 0) {
5894                         CDEBUG(D_LFSCK, "%s: layout LFSCK failed to "
5895                               "load LAST_ID for %#llx: rc = %d\n",
5896                               lfsck_lfsck2name(com->lc_lfsck), seq, rc);
5897                         lo->ll_objs_failed_phase1++;
5898                         OBD_FREE_PTR(lls);
5899                         GOTO(unlock, rc);
5900                 }
5901
5902                 lfsck_layout_seq_insert(llsd, lls);
5903         }
5904
5905         if (unlikely(fid_is_last_id(fid)))
5906                 GOTO(unlock, rc = 0);
5907
5908         if (fid_is_idif(fid))
5909                 oid = fid_idif_id(fid_seq(fid), fid_oid(fid), fid_ver(fid));
5910         else
5911                 oid = fid_oid(fid);
5912
5913         if (oid > lls->lls_lastid_known)
5914                 lls->lls_lastid_known = oid;
5915
5916         if (oid > lls->lls_lastid) {
5917                 if (!(lo->ll_flags & LF_CRASHED_LASTID)) {
5918                         /* OFD may create new objects during LFSCK scanning. */
5919                         rc = lfsck_layout_lastid_reload(env, com, lls);
5920                         if (unlikely(rc != 0)) {
5921                                 CDEBUG(D_LFSCK, "%s: layout LFSCK failed to "
5922                                       "reload LAST_ID for %#llx: rc = %d\n",
5923                                       lfsck_lfsck2name(com->lc_lfsck),
5924                                       lls->lls_seq, rc);
5925
5926                                 GOTO(unlock, rc);
5927                         }
5928
5929                         if (oid <= lls->lls_lastid ||
5930                             lo->ll_flags & LF_CRASHED_LASTID)
5931                                 GOTO(unlock, rc = 0);
5932
5933                         LASSERT(lfsck->li_out_notify != NULL);
5934
5935                         lfsck->li_out_notify(env, lfsck->li_out_notify_data,
5936                                              LE_LASTID_REBUILDING);
5937                         lo->ll_flags |= LF_CRASHED_LASTID;
5938
5939                         CDEBUG(D_LFSCK, "%s: layout LFSCK finds crashed "
5940                                "LAST_ID file (2) for the sequence %#llx"
5941                                ", old value %llu, known value %llu\n",
5942                                lfsck_lfsck2name(lfsck), lls->lls_seq,
5943                                lls->lls_lastid, oid);
5944                 }
5945
5946                 lls->lls_lastid = oid;
5947                 lls->lls_dirty = 1;
5948         }
5949
5950         GOTO(unlock, rc = 0);
5951
5952 unlock:
5953         up_write(&com->lc_sem);
5954
5955         return rc;
5956 }
5957
5958 static int lfsck_layout_exec_dir(const struct lu_env *env,
5959                                  struct lfsck_component *com,
5960                                  struct lfsck_assistant_object *lso,
5961                                  struct lu_dirent *ent, __u16 type)
5962 {
5963         return 0;
5964 }
5965
5966 static int lfsck_layout_master_post(const struct lu_env *env,
5967                                     struct lfsck_component *com,
5968                                     int result, bool init)
5969 {
5970         struct lfsck_instance   *lfsck  = com->lc_lfsck;
5971         struct lfsck_layout     *lo     = com->lc_file_ram;
5972         int                      rc;
5973         ENTRY;
5974
5975         lfsck_post_generic(env, com, &result);
5976
5977         down_write(&com->lc_sem);
5978         spin_lock(&lfsck->li_lock);
5979         if (!init)
5980                 lo->ll_pos_last_checkpoint =
5981                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
5982
5983         if (result > 0) {
5984                 if (lo->ll_flags & LF_INCOMPLETE)
5985                         lo->ll_status = LS_PARTIAL;
5986                 else
5987                         lo->ll_status = LS_SCANNING_PHASE2;
5988                 lo->ll_flags |= LF_SCANNED_ONCE;
5989                 lo->ll_flags &= ~LF_UPGRADE;
5990                 list_move_tail(&com->lc_link, &lfsck->li_list_double_scan);
5991         } else if (result == 0) {
5992                 if (lfsck->li_status != 0)
5993                         lo->ll_status = lfsck->li_status;
5994                 else
5995                         lo->ll_status = LS_STOPPED;
5996                 if (lo->ll_status != LS_PAUSED)
5997                         list_move_tail(&com->lc_link, &lfsck->li_list_idle);
5998         } else {
5999                 lo->ll_status = LS_FAILED;
6000                 list_move_tail(&com->lc_link, &lfsck->li_list_idle);
6001         }
6002         spin_unlock(&lfsck->li_lock);
6003
6004         if (!init) {
6005                 lo->ll_run_time_phase1 += ktime_get_seconds() -
6006                                           lfsck->li_time_last_checkpoint;
6007                 lo->ll_time_last_checkpoint = ktime_get_real_seconds();
6008                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
6009                 com->lc_new_checked = 0;
6010         }
6011
6012         rc = lfsck_layout_store(env, com);
6013         up_write(&com->lc_sem);
6014
6015         CDEBUG(D_LFSCK, "%s: layout LFSCK master post done: rc = %d\n",
6016                lfsck_lfsck2name(lfsck), rc);
6017
6018         RETURN(rc);
6019 }
6020
6021 static int lfsck_layout_slave_post(const struct lu_env *env,
6022                                    struct lfsck_component *com,
6023                                    int result, bool init)
6024 {
6025         struct lfsck_instance   *lfsck = com->lc_lfsck;
6026         struct lfsck_layout     *lo    = com->lc_file_ram;
6027         int                      rc;
6028         bool                     done  = false;
6029
6030         down_write(&com->lc_sem);
6031         rc = lfsck_layout_lastid_store(env, com);
6032         if (rc != 0)
6033                 result = rc;
6034
6035         LASSERT(lfsck->li_out_notify != NULL);
6036
6037         spin_lock(&lfsck->li_lock);
6038         if (!init)
6039                 lo->ll_pos_last_checkpoint =
6040                                 lfsck->li_pos_checkpoint.lp_oit_cookie;
6041
6042         if (result > 0) {
6043                 lo->ll_status = LS_SCANNING_PHASE2;
6044                 lo->ll_flags |= LF_SCANNED_ONCE;
6045                 if (lo->ll_flags & LF_CRASHED_LASTID) {
6046                         done = true;
6047                         lo->ll_flags &= ~LF_CRASHED_LASTID;
6048
6049                         CDEBUG(D_LFSCK, "%s: layout LFSCK has rebuilt "
6050                                "crashed LAST_ID files successfully\n",
6051                                lfsck_lfsck2name(lfsck));
6052                 }
6053                 lo->ll_flags &= ~LF_UPGRADE;
6054                 list_move_tail(&com->lc_link, &lfsck->li_list_double_scan);
6055         } else if (result == 0) {
6056                 if (lfsck->li_status != 0)
6057                         lo->ll_status = lfsck->li_status;
6058                 else
6059                         lo->ll_status = LS_STOPPED;
6060                 if (lo->ll_status != LS_PAUSED)
6061                         list_move_tail(&com->lc_link, &lfsck->li_list_idle);
6062         } else {
6063                 lo->ll_status = LS_FAILED;
6064                 list_move_tail(&com->lc_link, &lfsck->li_list_idle);
6065         }
6066         spin_unlock(&lfsck->li_lock);
6067
6068         if (done)
6069                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
6070                                      LE_LASTID_REBUILT);
6071
6072         if (!init) {
6073                 lo->ll_run_time_phase1 += ktime_get_seconds() -
6074                                           lfsck->li_time_last_checkpoint;
6075                 lo->ll_time_last_checkpoint = ktime_get_real_seconds();
6076                 lo->ll_objs_checked_phase1 += com->lc_new_checked;
6077                 com->lc_new_checked = 0;
6078         }
6079
6080         rc = lfsck_layout_store(env, com);
6081         up_write(&com->lc_sem);
6082
6083         lfsck_layout_slave_notify_master(env, com, LE_PHASE1_DONE, result);
6084
6085         CDEBUG(D_LFSCK, "%s: layout LFSCK slave post done: rc = %d\n",
6086                lfsck_lfsck2name(lfsck), rc);
6087
6088         return rc;
6089 }
6090
6091 static void lfsck_layout_dump(const struct lu_env *env,
6092                               struct lfsck_component *com, struct seq_file *m)
6093 {
6094         struct lfsck_instance   *lfsck = com->lc_lfsck;
6095         struct lfsck_bookmark   *bk    = &lfsck->li_bookmark_ram;
6096         struct lfsck_layout     *lo    = com->lc_file_ram;
6097         const char *prefix;
6098
6099         down_read(&com->lc_sem);
6100         if (bk->lb_param & LPF_DRYRUN)
6101                 prefix = "inconsistent";
6102         else
6103                 prefix = "repaired";
6104
6105         seq_printf(m, "name: lfsck_layout\n"
6106                    "magic: %#x\n"
6107                    "version: %d\n"
6108                    "status: %s\n",
6109                    lo->ll_magic,
6110                    bk->lb_version,
6111                    lfsck_status2name(lo->ll_status));
6112
6113         lfsck_bits_dump(m, lo->ll_flags, lfsck_flags_names, "flags");
6114
6115         lfsck_bits_dump(m, bk->lb_param, lfsck_param_names, "param");
6116
6117         lfsck_time_dump(m, lo->ll_time_last_complete, "last_completed");
6118
6119         lfsck_time_dump(m, lo->ll_time_latest_start, "latest_start");
6120
6121         lfsck_time_dump(m, lo->ll_time_last_checkpoint, "last_checkpoint");
6122
6123         seq_printf(m, "latest_start_position: %llu\n"
6124                    "last_checkpoint_position: %llu\n"
6125                    "first_failure_position: %llu\n",
6126                    lo->ll_pos_latest_start,
6127                    lo->ll_pos_last_checkpoint,
6128                    lo->ll_pos_first_inconsistent);
6129
6130         seq_printf(m, "success_count: %u\n"
6131                    "%s_dangling: %llu\n"
6132                    "%s_unmatched_pair: %llu\n"
6133                    "%s_multiple_referenced: %llu\n"
6134                    "%s_orphan: %llu\n"
6135                    "%s_inconsistent_owner: %llu\n"
6136                    "%s_others: %llu\n"
6137                    "skipped: %llu\n"
6138                    "failed_phase1: %llu\n"
6139                    "failed_phase2: %llu\n",
6140                    lo->ll_success_count,
6141                    prefix, lo->ll_objs_repaired[LLIT_DANGLING - 1],
6142                    prefix, lo->ll_objs_repaired[LLIT_UNMATCHED_PAIR - 1],
6143                    prefix, lo->ll_objs_repaired[LLIT_MULTIPLE_REFERENCED - 1],
6144                    prefix, lo->ll_objs_repaired[LLIT_ORPHAN - 1],
6145                    prefix, lo->ll_objs_repaired[LLIT_INCONSISTENT_OWNER - 1],
6146                    prefix, lo->ll_objs_repaired[LLIT_OTHERS - 1],
6147                    lo->ll_objs_skipped,
6148                    lo->ll_objs_failed_phase1,
6149                    lo->ll_objs_failed_phase2);
6150
6151         if (lo->ll_status == LS_SCANNING_PHASE1) {
6152                 time64_t duration = ktime_get_seconds() -
6153                                     lfsck->li_time_last_checkpoint;
6154                 u64 checked = lo->ll_objs_checked_phase1 +
6155                               com->lc_new_checked;
6156                 u64 speed = checked;
6157                 u64 new_checked = com->lc_new_checked;
6158                 time64_t rtime = lo->ll_run_time_phase1 + duration;
6159                 u64 pos;
6160
6161                 if (duration != 0)
6162                         new_checked = div64_s64(new_checked, duration);
6163                 if (rtime != 0)
6164                         speed = div64_s64(speed, rtime);
6165                 seq_printf(m, "checked_phase1: %llu\n"
6166                            "checked_phase2: %llu\n"
6167                            "run_time_phase1: %lld seconds\n"
6168                            "run_time_phase2: %lld seconds\n"
6169                            "average_speed_phase1: %llu items/sec\n"
6170                            "average_speed_phase2: N/A\n"
6171                            "real_time_speed_phase1: %llu items/sec\n"
6172                            "real_time_speed_phase2: N/A\n",
6173                            checked,
6174                            lo->ll_objs_checked_phase2,
6175                            rtime,
6176                            lo->ll_run_time_phase2,
6177                            speed,
6178                            new_checked);
6179
6180                 if (likely(lfsck->li_di_oit)) {
6181                         const struct dt_it_ops *iops =
6182                                 &lfsck->li_obj_oit->do_index_ops->dio_it;
6183
6184                         /* The low layer otable-based iteration position may NOT
6185                          * exactly match the layout-based directory traversal
6186                          * cookie. Generally, it is not a serious issue. But the
6187                          * caller should NOT make assumption on that. */
6188                         pos = iops->store(env, lfsck->li_di_oit);
6189                         if (!lfsck->li_current_oit_processed)
6190                                 pos--;
6191                 } else {
6192                         pos = lo->ll_pos_last_checkpoint;
6193                 }
6194
6195                 seq_printf(m, "current_position: %llu\n", pos);
6196         } else if (lo->ll_status == LS_SCANNING_PHASE2) {
6197                 time64_t duration = ktime_get_seconds() -
6198                                     com->lc_time_last_checkpoint;
6199                 u64 checked = lo->ll_objs_checked_phase2 +
6200                               com->lc_new_checked;
6201                 u64 speed1 = lo->ll_objs_checked_phase1;
6202                 u64 speed2 = checked;
6203                 u64 new_checked = com->lc_new_checked;
6204                 time64_t rtime = lo->ll_run_time_phase2 + duration;
6205
6206                 if (duration != 0)
6207                         new_checked = div64_s64(new_checked, duration);
6208                 if (lo->ll_run_time_phase1 != 0)
6209                         speed1 = div64_s64(speed1, lo->ll_run_time_phase1);
6210                 if (rtime != 0)
6211                         speed2 = div64_s64(speed2, rtime);
6212                 seq_printf(m, "checked_phase1: %llu\n"
6213                            "checked_phase2: %llu\n"
6214                            "run_time_phase1: %lld seconds\n"
6215                            "run_time_phase2: %lld seconds\n"
6216                            "average_speed_phase1: %llu items/sec\n"
6217                            "average_speed_phase2: %llu items/sec\n"
6218                            "real_time_speed_phase1: N/A\n"
6219                            "real_time_speed_phase2: %llu items/sec\n"
6220                            "current_position: "DFID"\n",
6221                            lo->ll_objs_checked_phase1,
6222                            checked,
6223                            lo->ll_run_time_phase1,
6224                            rtime,
6225                            speed1,
6226                            speed2,
6227                            new_checked,
6228                            PFID(&com->lc_fid_latest_scanned_phase2));
6229         } else {
6230                 __u64 speed1 = lo->ll_objs_checked_phase1;
6231                 __u64 speed2 = lo->ll_objs_checked_phase2;
6232
6233                 if (lo->ll_run_time_phase1 != 0)
6234                         speed1 = div64_s64(speed1, lo->ll_run_time_phase1);
6235                 if (lo->ll_run_time_phase2 != 0)
6236                         speed2 = div64_s64(speed2, lo->ll_run_time_phase2);
6237                 seq_printf(m, "checked_phase1: %llu\n"
6238                            "checked_phase2: %llu\n"
6239                            "run_time_phase1: %lld seconds\n"
6240                            "run_time_phase2: %lld seconds\n"
6241                            "average_speed_phase1: %llu items/sec\n"
6242                            "average_speed_phase2: %llu objs/sec\n"
6243                            "real_time_speed_phase1: N/A\n"
6244                            "real_time_speed_phase2: N/A\n"
6245                            "current_position: N/A\n",
6246                            lo->ll_objs_checked_phase1,
6247                            lo->ll_objs_checked_phase2,
6248                            lo->ll_run_time_phase1,
6249                            lo->ll_run_time_phase2,
6250                            speed1,
6251                            speed2);
6252         }
6253
6254         up_read(&com->lc_sem);
6255 }
6256
6257 static int lfsck_layout_master_double_scan(const struct lu_env *env,
6258                                            struct lfsck_component *com)
6259 {
6260         struct lfsck_layout             *lo     = com->lc_file_ram;
6261         struct lfsck_assistant_data     *lad    = com->lc_data;
6262         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6263         struct lfsck_tgt_descs          *ltds;
6264         struct lfsck_tgt_desc           *ltd;
6265         struct lfsck_tgt_desc           *next;
6266         int                              rc;
6267
6268         rc = lfsck_double_scan_generic(env, com, lo->ll_status);
6269
6270         if (thread_is_stopped(&lad->lad_thread)) {
6271                 LASSERT(list_empty(&lad->lad_req_list));
6272                 LASSERT(list_empty(&lad->lad_ost_phase1_list));
6273                 LASSERT(list_empty(&lad->lad_mdt_phase1_list));
6274
6275                 ltds = &lfsck->li_ost_descs;
6276                 spin_lock(&ltds->ltd_lock);
6277                 list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase2_list,
6278                                          ltd_layout_phase_list) {
6279                         list_del_init(&ltd->ltd_layout_phase_list);
6280                 }
6281                 spin_unlock(&ltds->ltd_lock);
6282
6283                 ltds = &lfsck->li_mdt_descs;
6284                 spin_lock(&ltds->ltd_lock);
6285                 list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase2_list,
6286                                          ltd_layout_phase_list) {
6287                         list_del_init(&ltd->ltd_layout_phase_list);
6288                 }
6289                 spin_unlock(&ltds->ltd_lock);
6290         }
6291
6292         return rc;
6293 }
6294
6295 static int lfsck_layout_slave_double_scan(const struct lu_env *env,
6296                                           struct lfsck_component *com)
6297 {
6298         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6299         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
6300         struct lfsck_layout             *lo     = com->lc_file_ram;
6301         struct ptlrpc_thread            *thread = &lfsck->li_thread;
6302         int                              rc;
6303         ENTRY;
6304
6305         CDEBUG(D_LFSCK, "%s: layout LFSCK slave phase2 scan start\n",
6306                lfsck_lfsck2name(lfsck));
6307
6308         atomic_inc(&lfsck->li_double_scan_count);
6309
6310         if (lo->ll_flags & LF_INCOMPLETE)
6311                 GOTO(done, rc = 1);
6312
6313         com->lc_new_checked = 0;
6314         com->lc_new_scanned = 0;
6315         com->lc_time_last_checkpoint = ktime_get_seconds();
6316         com->lc_time_next_checkpoint = com->lc_time_last_checkpoint +
6317                                        LFSCK_CHECKPOINT_INTERVAL;
6318
6319         while (1) {
6320                 rc = lfsck_layout_slave_query_master(env, com);
6321                 if (list_empty(&llsd->llsd_master_list)) {
6322                         if (unlikely(!thread_is_running(thread)))
6323                                 rc = 0;
6324                         else
6325                                 rc = 1;
6326
6327                         GOTO(done, rc);
6328                 }
6329
6330                 if (rc < 0)
6331                         GOTO(done, rc);
6332
6333                 rc = wait_event_idle_timeout(
6334                         thread->t_ctl_waitq,
6335                         !thread_is_running(thread) ||
6336                         lo->ll_flags & LF_INCOMPLETE ||
6337                         list_empty(&llsd->llsd_master_list),
6338                         cfs_time_seconds(30));
6339                 if (unlikely(!thread_is_running(thread)))
6340                         GOTO(done, rc = 0);
6341
6342                 if (lo->ll_flags & LF_INCOMPLETE)
6343                         GOTO(done, rc = 1);
6344
6345                 if (rc == 0)
6346                         continue;
6347
6348                 GOTO(done, rc = 1);
6349         }
6350
6351 done:
6352         rc = lfsck_layout_double_scan_result(env, com, rc);
6353         lfsck_layout_slave_notify_master(env, com, LE_PHASE2_DONE,
6354                         (rc > 0 && lo->ll_flags & LF_INCOMPLETE) ? 0 : rc);
6355         lfsck_layout_slave_quit(env, com);
6356         if (atomic_dec_and_test(&lfsck->li_double_scan_count))
6357                 wake_up(&lfsck->li_thread.t_ctl_waitq);
6358
6359         CDEBUG(D_LFSCK, "%s: layout LFSCK slave phase2 scan finished, "
6360                "status %d: rc = %d\n",
6361                lfsck_lfsck2name(lfsck), lo->ll_status, rc);
6362
6363         return rc;
6364 }
6365
6366 static void lfsck_layout_master_data_release(const struct lu_env *env,
6367                                              struct lfsck_component *com)
6368 {
6369         struct lfsck_assistant_data     *lad    = com->lc_data;
6370         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6371         struct lfsck_tgt_descs          *ltds;
6372         struct lfsck_tgt_desc           *ltd;
6373         struct lfsck_tgt_desc           *next;
6374
6375         LASSERT(lad != NULL);
6376         LASSERT(thread_is_init(&lad->lad_thread) ||
6377                 thread_is_stopped(&lad->lad_thread));
6378         LASSERT(list_empty(&lad->lad_req_list));
6379
6380         com->lc_data = NULL;
6381
6382         ltds = &lfsck->li_ost_descs;
6383         spin_lock(&ltds->ltd_lock);
6384         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase1_list,
6385                                  ltd_layout_phase_list) {
6386                 list_del_init(&ltd->ltd_layout_phase_list);
6387         }
6388         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase2_list,
6389                                  ltd_layout_phase_list) {
6390                 list_del_init(&ltd->ltd_layout_phase_list);
6391         }
6392         list_for_each_entry_safe(ltd, next, &lad->lad_ost_list,
6393                                  ltd_layout_list) {
6394                 list_del_init(&ltd->ltd_layout_list);
6395         }
6396         spin_unlock(&ltds->ltd_lock);
6397
6398         ltds = &lfsck->li_mdt_descs;
6399         spin_lock(&ltds->ltd_lock);
6400         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase1_list,
6401                                  ltd_layout_phase_list) {
6402                 list_del_init(&ltd->ltd_layout_phase_list);
6403         }
6404         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase2_list,
6405                                  ltd_layout_phase_list) {
6406                 list_del_init(&ltd->ltd_layout_phase_list);
6407         }
6408         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_list,
6409                                  ltd_layout_list) {
6410                 list_del_init(&ltd->ltd_layout_list);
6411         }
6412         spin_unlock(&ltds->ltd_lock);
6413
6414         if (likely(lad->lad_bitmap != NULL))
6415                 CFS_FREE_BITMAP(lad->lad_bitmap);
6416
6417         OBD_FREE_PTR(lad);
6418 }
6419
6420 static void lfsck_layout_slave_data_release(const struct lu_env *env,
6421                                             struct lfsck_component *com)
6422 {
6423         struct lfsck_layout_slave_data *llsd = com->lc_data;
6424
6425         lfsck_layout_slave_quit(env, com);
6426         com->lc_data = NULL;
6427         OBD_FREE_PTR(llsd);
6428 }
6429
6430 static void lfsck_layout_master_quit(const struct lu_env *env,
6431                                      struct lfsck_component *com)
6432 {
6433         struct lfsck_assistant_data     *lad    = com->lc_data;
6434         struct lfsck_instance           *lfsck  = com->lc_lfsck;
6435         struct lfsck_tgt_descs          *ltds;
6436         struct lfsck_tgt_desc           *ltd;
6437         struct lfsck_tgt_desc           *next;
6438
6439         LASSERT(lad != NULL);
6440
6441         lfsck_quit_generic(env, com);
6442
6443         LASSERT(thread_is_init(&lad->lad_thread) ||
6444                 thread_is_stopped(&lad->lad_thread));
6445         LASSERT(list_empty(&lad->lad_req_list));
6446
6447         ltds = &lfsck->li_ost_descs;
6448         spin_lock(&ltds->ltd_lock);
6449         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase1_list,
6450                                  ltd_layout_phase_list) {
6451                 list_del_init(&ltd->ltd_layout_phase_list);
6452         }
6453         list_for_each_entry_safe(ltd, next, &lad->lad_ost_phase2_list,
6454                                  ltd_layout_phase_list) {
6455                 list_del_init(&ltd->ltd_layout_phase_list);
6456         }
6457         spin_unlock(&ltds->ltd_lock);
6458
6459         ltds = &lfsck->li_mdt_descs;
6460         spin_lock(&ltds->ltd_lock);
6461         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase1_list,
6462                                  ltd_layout_phase_list) {
6463                 list_del_init(&ltd->ltd_layout_phase_list);
6464         }
6465         list_for_each_entry_safe(ltd, next, &lad->lad_mdt_phase2_list,
6466                                  ltd_layout_phase_list) {
6467                 list_del_init(&ltd->ltd_layout_phase_list);
6468         }
6469         spin_unlock(&ltds->ltd_lock);
6470 }
6471
6472 static void lfsck_layout_slave_quit(const struct lu_env *env,
6473                                     struct lfsck_component *com)
6474 {
6475         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
6476         struct lfsck_layout_seq          *lls;
6477         struct lfsck_layout_seq          *next;
6478         struct lfsck_layout_slave_target *llst;
6479
6480         LASSERT(llsd != NULL);
6481
6482         down_write(&com->lc_sem);
6483         list_for_each_entry_safe(lls, next, &llsd->llsd_seq_list,
6484                                  lls_list) {
6485                 list_del_init(&lls->lls_list);
6486                 lfsck_object_put(env, lls->lls_lastid_obj);
6487                 OBD_FREE_PTR(lls);
6488         }
6489         up_write(&com->lc_sem);
6490
6491         spin_lock(&llsd->llsd_lock);
6492         while (!list_empty(&llsd->llsd_master_list)) {
6493                 llst = list_entry(llsd->llsd_master_list.next,
6494                                   struct lfsck_layout_slave_target, llst_list);
6495                 list_del_init(&llst->llst_list);
6496                 spin_unlock(&llsd->llsd_lock);
6497                 lfsck_layout_llst_put(llst);
6498                 spin_lock(&llsd->llsd_lock);
6499         }
6500         spin_unlock(&llsd->llsd_lock);
6501
6502         lfsck_rbtree_cleanup(env, com);
6503 }
6504
6505 static int lfsck_layout_master_in_notify(const struct lu_env *env,
6506                                          struct lfsck_component *com,
6507                                          struct lfsck_request *lr)
6508 {
6509         struct lfsck_instance           *lfsck = com->lc_lfsck;
6510         struct lfsck_layout             *lo    = com->lc_file_ram;
6511         struct lfsck_assistant_data     *lad   = com->lc_data;
6512         struct lfsck_tgt_descs          *ltds;
6513         struct lfsck_tgt_desc           *ltd;
6514         bool                             fail  = false;
6515         ENTRY;
6516
6517         if (lr->lr_event == LE_PAIRS_VERIFY) {
6518                 int rc;
6519
6520                 rc = lfsck_layout_master_check_pairs(env, com, &lr->lr_fid,
6521                                                      &lr->lr_fid2,
6522                                                      lr->lr_comp_id);
6523
6524                 RETURN(rc);
6525         }
6526
6527         CDEBUG(D_LFSCK, "%s: layout LFSCK master handles notify %u "
6528                "from %s %x, status %d, flags %x, flags2 %x\n",
6529                lfsck_lfsck2name(lfsck), lr->lr_event,
6530                (lr->lr_flags & LEF_FROM_OST) ? "OST" : "MDT",
6531                lr->lr_index, lr->lr_status, lr->lr_flags, lr->lr_flags2);
6532
6533         if (lr->lr_event != LE_PHASE1_DONE &&
6534             lr->lr_event != LE_PHASE2_DONE &&
6535             lr->lr_event != LE_PEER_EXIT)
6536                 RETURN(-EINVAL);
6537
6538         if (lr->lr_flags & LEF_FROM_OST)
6539                 ltds = &lfsck->li_ost_descs;
6540         else
6541                 ltds = &lfsck->li_mdt_descs;
6542         spin_lock(&ltds->ltd_lock);
6543         ltd = lfsck_ltd2tgt(ltds, lr->lr_index);
6544         if (ltd == NULL) {
6545                 spin_unlock(&ltds->ltd_lock);
6546
6547                 RETURN(-ENXIO);
6548         }
6549
6550         list_del_init(&ltd->ltd_layout_phase_list);
6551         switch (lr->lr_event) {
6552         case LE_PHASE1_DONE:
6553                 if (lr->lr_status <= 0 || lr->lr_flags2 & LF_INCOMPLETE) {
6554                         if (lr->lr_flags2 & LF_INCOMPLETE) {
6555                                 if (lr->lr_flags & LEF_FROM_OST)
6556                                         lfsck_lad_set_bitmap(env, com,
6557                                                              ltd->ltd_index);
6558                                 else
6559                                         lo->ll_flags |= LF_INCOMPLETE;
6560                         }
6561                         ltd->ltd_layout_done = 1;
6562                         list_del_init(&ltd->ltd_layout_list);
6563                         fail = true;
6564                         break;
6565                 }
6566
6567                 if (lr->lr_flags & LEF_FROM_OST) {
6568                         if (list_empty(&ltd->ltd_layout_list))
6569                                 list_add_tail(&ltd->ltd_layout_list,
6570                                               &lad->lad_ost_list);
6571                         list_add_tail(&ltd->ltd_layout_phase_list,
6572                                       &lad->lad_ost_phase2_list);
6573                 } else {
6574                         if (list_empty(&ltd->ltd_layout_list))
6575                                 list_add_tail(&ltd->ltd_layout_list,
6576                                               &lad->lad_mdt_list);
6577                         list_add_tail(&ltd->ltd_layout_phase_list,
6578                                       &lad->lad_mdt_phase2_list);
6579                 }
6580                 break;
6581         case LE_PHASE2_DONE:
6582                 ltd->ltd_layout_done = 1;
6583                 if (!list_empty(&ltd->ltd_layout_list))
6584                         list_del_init(&ltd->ltd_layout_list);
6585
6586                 if (lr->lr_flags2 & LF_INCOMPLETE) {
6587                         lfsck_lad_set_bitmap(env, com, ltd->ltd_index);
6588                         fail = true;
6589                 }
6590
6591                 break;
6592         case LE_PEER_EXIT:
6593                 fail = true;
6594                 ltd->ltd_layout_done = 1;
6595                 list_del_init(&ltd->ltd_layout_list);
6596                 if (!(lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT) &&
6597                     !(lr->lr_flags & LEF_FROM_OST))
6598                                 lo->ll_flags |= LF_INCOMPLETE;
6599                 break;
6600         default:
6601                 break;
6602         }
6603         spin_unlock(&ltds->ltd_lock);
6604
6605         if (fail && lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT) {
6606                 struct lfsck_stop *stop = &lfsck_env_info(env)->lti_stop;
6607
6608                 memset(stop, 0, sizeof(*stop));
6609                 stop->ls_status = lr->lr_status;
6610                 stop->ls_flags = lr->lr_param & ~LPF_BROADCAST;
6611                 lfsck_stop(env, lfsck->li_bottom, stop);
6612         } else if (lfsck_phase2_next_ready(lad)) {
6613                 wake_up(&lad->lad_thread.t_ctl_waitq);
6614         }
6615
6616         RETURN(0);
6617 }
6618
6619 static int lfsck_layout_slave_in_notify_local(const struct lu_env *env,
6620                                               struct lfsck_component *com,
6621                                               struct lfsck_req_local *lrl,
6622                                               struct thandle *th)
6623 {
6624         ENTRY;
6625
6626         switch (lrl->lrl_event) {
6627         case LEL_FID_ACCESSED:
6628                 lfsck_rbtree_update_bitmap(env, com, &lrl->lrl_fid, true);
6629                 RETURN(0);
6630         case LEL_PAIRS_VERIFY_LOCAL: {
6631                 int rc;
6632
6633                 lrl->lrl_status = LPVS_INIT;
6634                 /* Firstly, if the MDT-object which is claimed via OST-object
6635                  * local stored PFID xattr recognizes the OST-object, then it
6636                  * must be that the client given PFID is wrong. */
6637                 rc = lfsck_layout_slave_check_pairs(env, com, &lrl->lrl_fid,
6638                                 &lrl->lrl_ff_local.ff_parent,
6639                                 lrl->lrl_ff_local.ff_layout.ol_comp_id);
6640                 if (rc <= 0)
6641                         RETURN(0);
6642
6643                 lrl->lrl_status = LPVS_INCONSISTENT;
6644                 /* The OST-object local stored PFID xattr is stale. We need to
6645                  * check whether the MDT-object that is claimed via the client
6646                  * given PFID information recognizes the OST-object or not. If
6647                  * matches, then need to update the OST-object's PFID xattr. */
6648                 rc = lfsck_layout_slave_check_pairs(env, com, &lrl->lrl_fid,
6649                                 &lrl->lrl_ff_client.ff_parent,
6650                                 lrl->lrl_ff_client.ff_layout.ol_comp_id);
6651                 /* For rc < 0 case:
6652                  * We are not sure whether the client given PFID information
6653                  * is correct or not, do nothing to avoid improper fixing.
6654                  *
6655                  * For rc > 0 case:
6656                  * The client given PFID information is also invalid, we can
6657                  * NOT fix the OST-object inconsistency.
6658                  */
6659                 if (!rc) {
6660                         lrl->lrl_status = LPVS_INCONSISTENT_TOFIX;
6661                         rc = lfsck_layout_slave_repair_pfid(env, com, lrl);
6662                 }
6663
6664                 RETURN(rc);
6665         }
6666         default:
6667                 break;
6668         }
6669
6670         RETURN(-EOPNOTSUPP);
6671 }
6672
6673 static int lfsck_layout_slave_in_notify(const struct lu_env *env,
6674                                         struct lfsck_component *com,
6675                                         struct lfsck_request *lr)
6676 {
6677         struct lfsck_instance *lfsck = com->lc_lfsck;
6678         struct lfsck_layout_slave_data *llsd = com->lc_data;
6679         struct lfsck_layout_slave_target *llst;
6680         int rc;
6681         ENTRY;
6682
6683         switch (lr->lr_event) {
6684         case LE_CONDITIONAL_DESTROY:
6685                 rc = lfsck_layout_slave_conditional_destroy(env, com, lr);
6686                 RETURN(rc);
6687         case LE_PHASE1_DONE: {
6688                 if (lr->lr_flags2 & LF_INCOMPLETE) {
6689                         struct lfsck_layout *lo = com->lc_file_ram;
6690
6691                         lo->ll_flags |= LF_INCOMPLETE;
6692                         llst = lfsck_layout_llst_find_and_del(llsd,
6693                                                               lr->lr_index,
6694                                                               true);
6695                         if (llst != NULL) {
6696                                 lfsck_layout_llst_put(llst);
6697                                 wake_up(&lfsck->li_thread.t_ctl_waitq);
6698                         }
6699                 }
6700
6701                 RETURN(0);
6702         }
6703         case LE_PHASE2_DONE:
6704         case LE_PEER_EXIT:
6705                 CDEBUG(D_LFSCK, "%s: layout LFSCK slave handle notify %u "
6706                        "from MDT %x, status %d\n", lfsck_lfsck2name(lfsck),
6707                        lr->lr_event, lr->lr_index, lr->lr_status);
6708                 break;
6709         default:
6710                 RETURN(-EINVAL);
6711         }
6712
6713         llst = lfsck_layout_llst_find_and_del(llsd, lr->lr_index, true);
6714         if (llst == NULL)
6715                 RETURN(0);
6716
6717         lfsck_layout_llst_put(llst);
6718         if (list_empty(&llsd->llsd_master_list))
6719                 wake_up(&lfsck->li_thread.t_ctl_waitq);
6720
6721         if (lr->lr_event == LE_PEER_EXIT &&
6722             (lfsck->li_bookmark_ram.lb_param & LPF_FAILOUT ||
6723              (list_empty(&llsd->llsd_master_list) &&
6724               (lr->lr_status == LS_STOPPED ||
6725                lr->lr_status == LS_CO_STOPPED)))) {
6726                 struct lfsck_stop *stop = &lfsck_env_info(env)->lti_stop;
6727
6728                 memset(stop, 0, sizeof(*stop));
6729                 stop->ls_status = lr->lr_status;
6730                 stop->ls_flags = lr->lr_param & ~LPF_BROADCAST;
6731                 lfsck_stop(env, lfsck->li_bottom, stop);
6732         }
6733
6734         RETURN(0);
6735 }
6736
6737 static void lfsck_layout_repaired(struct lfsck_layout *lo, __u64 *count)
6738 {
6739         int i;
6740
6741         for (i = 0; i < LLIT_MAX; i++)
6742                 *count += lo->ll_objs_repaired[i];
6743 }
6744
6745 static int lfsck_layout_query_all(const struct lu_env *env,
6746                                   struct lfsck_component *com,
6747                                   __u32 *mdts_count, __u32 *osts_count,
6748                                   __u64 *repaired)
6749 {
6750         struct lfsck_layout *lo = com->lc_file_ram;
6751         struct lfsck_tgt_descs *ltds;
6752         struct lfsck_tgt_desc *ltd;
6753         int idx;
6754         int rc;
6755         ENTRY;
6756
6757         rc = lfsck_query_all(env, com);
6758         if (rc != 0)
6759                 RETURN(rc);
6760
6761         ltds = &com->lc_lfsck->li_mdt_descs;
6762         down_read(&ltds->ltd_rw_sem);
6763         cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
6764                 ltd = lfsck_ltd2tgt(ltds, idx);
6765                 LASSERT(ltd != NULL);
6766
6767                 mdts_count[ltd->ltd_layout_status]++;
6768                 *repaired += ltd->ltd_layout_repaired;
6769         }
6770         up_read(&ltds->ltd_rw_sem);
6771
6772         ltds = &com->lc_lfsck->li_ost_descs;
6773         down_read(&ltds->ltd_rw_sem);
6774         cfs_foreach_bit(ltds->ltd_tgts_bitmap, idx) {
6775                 ltd = lfsck_ltd2tgt(ltds, idx);
6776                 LASSERT(ltd != NULL);
6777
6778                 osts_count[ltd->ltd_layout_status]++;
6779                 *repaired += ltd->ltd_layout_repaired;
6780         }
6781         up_read(&ltds->ltd_rw_sem);
6782
6783         down_read(&com->lc_sem);
6784         mdts_count[lo->ll_status]++;
6785         lfsck_layout_repaired(lo, repaired);
6786         up_read(&com->lc_sem);
6787
6788         RETURN(0);
6789 }
6790
6791 static int lfsck_layout_query(const struct lu_env *env,
6792                               struct lfsck_component *com,
6793                               struct lfsck_request *req,
6794                               struct lfsck_reply *rep,
6795                               struct lfsck_query *que, int idx)
6796 {
6797         struct lfsck_layout *lo = com->lc_file_ram;
6798         int rc = 0;
6799
6800         if (que != NULL) {
6801                 LASSERT(com->lc_lfsck->li_master);
6802
6803                 rc = lfsck_layout_query_all(env, com,
6804                                             que->lu_mdts_count[idx],
6805                                             que->lu_osts_count[idx],
6806                                             &que->lu_repaired[idx]);
6807         } else {
6808                 down_read(&com->lc_sem);
6809                 rep->lr_status = lo->ll_status;
6810                 if (req->lr_flags & LEF_QUERY_ALL)
6811                         lfsck_layout_repaired(lo, &rep->lr_repaired);
6812                 up_read(&com->lc_sem);
6813         }
6814
6815         return rc;
6816 }
6817
6818 /* with lfsck::li_lock held */
6819 static int lfsck_layout_slave_join(const struct lu_env *env,
6820                                    struct lfsck_component *com,
6821                                    struct lfsck_start_param *lsp)
6822 {
6823         struct lfsck_instance            *lfsck = com->lc_lfsck;
6824         struct lfsck_layout_slave_data   *llsd  = com->lc_data;
6825         struct lfsck_layout_slave_target *llst;
6826         struct lfsck_start               *start = lsp->lsp_start;
6827         int                               rc    = 0;
6828         ENTRY;
6829
6830         if (start == NULL || !(start->ls_flags & LPF_OST_ORPHAN))
6831                 RETURN(0);
6832
6833         if (!lsp->lsp_index_valid)
6834                 RETURN(-EINVAL);
6835
6836         /* If someone is running the LFSCK without orphan handling,
6837          * it will not maintain the object accessing rbtree. So we
6838          * cannot join it for orphan handling. */
6839         if (!llsd->llsd_rbtree_valid)
6840                 RETURN(-EBUSY);
6841
6842         spin_unlock(&lfsck->li_lock);
6843         rc = lfsck_layout_llst_add(llsd, lsp->lsp_index);
6844         spin_lock(&lfsck->li_lock);
6845         if (rc == 0 && !thread_is_running(&lfsck->li_thread)) {
6846                 spin_unlock(&lfsck->li_lock);
6847                 llst = lfsck_layout_llst_find_and_del(llsd, lsp->lsp_index,
6848                                                       true);
6849                 if (llst != NULL)
6850                         lfsck_layout_llst_put(llst);
6851                 spin_lock(&lfsck->li_lock);
6852                 rc = -EAGAIN;
6853         }
6854
6855         RETURN(rc);
6856 }
6857
6858 static const struct lfsck_operations lfsck_layout_master_ops = {
6859         .lfsck_reset            = lfsck_layout_reset,
6860         .lfsck_fail             = lfsck_layout_fail,
6861         .lfsck_checkpoint       = lfsck_layout_master_checkpoint,
6862         .lfsck_prep             = lfsck_layout_master_prep,
6863         .lfsck_exec_oit         = lfsck_layout_master_exec_oit,
6864         .lfsck_exec_dir         = lfsck_layout_exec_dir,
6865         .lfsck_post             = lfsck_layout_master_post,
6866         .lfsck_dump             = lfsck_layout_dump,
6867         .lfsck_double_scan      = lfsck_layout_master_double_scan,
6868         .lfsck_data_release     = lfsck_layout_master_data_release,
6869         .lfsck_quit             = lfsck_layout_master_quit,
6870         .lfsck_in_notify        = lfsck_layout_master_in_notify,
6871         .lfsck_query            = lfsck_layout_query,
6872 };
6873
6874 static const struct lfsck_operations lfsck_layout_slave_ops = {
6875         .lfsck_reset            = lfsck_layout_reset,
6876         .lfsck_fail             = lfsck_layout_fail,
6877         .lfsck_checkpoint       = lfsck_layout_slave_checkpoint,
6878         .lfsck_prep             = lfsck_layout_slave_prep,
6879         .lfsck_exec_oit         = lfsck_layout_slave_exec_oit,
6880         .lfsck_exec_dir         = lfsck_layout_exec_dir,
6881         .lfsck_post             = lfsck_layout_slave_post,
6882         .lfsck_dump             = lfsck_layout_dump,
6883         .lfsck_double_scan      = lfsck_layout_slave_double_scan,
6884         .lfsck_data_release     = lfsck_layout_slave_data_release,
6885         .lfsck_quit             = lfsck_layout_slave_quit,
6886         .lfsck_in_notify_local  = lfsck_layout_slave_in_notify_local,
6887         .lfsck_in_notify        = lfsck_layout_slave_in_notify,
6888         .lfsck_query            = lfsck_layout_query,
6889         .lfsck_join             = lfsck_layout_slave_join,
6890 };
6891
6892 static void lfsck_layout_assistant_fill_pos(const struct lu_env *env,
6893                                             struct lfsck_component *com,
6894                                             struct lfsck_position *pos)
6895 {
6896         struct lfsck_assistant_data     *lad = com->lc_data;
6897         struct lfsck_layout_req         *llr;
6898
6899         if (((struct lfsck_layout *)(com->lc_file_ram))->ll_status !=
6900             LS_SCANNING_PHASE1)
6901                 return;
6902
6903         if (list_empty(&lad->lad_req_list))
6904                 return;
6905
6906         llr = list_entry(lad->lad_req_list.next,
6907                          struct lfsck_layout_req,
6908                          llr_lar.lar_list);
6909         pos->lp_oit_cookie = llr->llr_lar.lar_parent->lso_oit_cookie - 1;
6910 }
6911
6912 const struct lfsck_assistant_operations lfsck_layout_assistant_ops = {
6913         .la_handler_p1          = lfsck_layout_assistant_handler_p1,
6914         .la_handler_p2          = lfsck_layout_assistant_handler_p2,
6915         .la_fill_pos            = lfsck_layout_assistant_fill_pos,
6916         .la_double_scan_result  = lfsck_layout_double_scan_result,
6917         .la_req_fini            = lfsck_layout_assistant_req_fini,
6918         .la_sync_failures       = lfsck_layout_assistant_sync_failures,
6919 };
6920
6921 int lfsck_layout_setup(const struct lu_env *env, struct lfsck_instance *lfsck)
6922 {
6923         struct lfsck_component  *com;
6924         struct lfsck_layout     *lo;
6925         struct dt_object        *root = NULL;
6926         struct dt_object        *obj;
6927         int                      i;
6928         int                      rc;
6929         ENTRY;
6930
6931         OBD_ALLOC_PTR(com);
6932         if (com == NULL)
6933                 RETURN(-ENOMEM);
6934
6935         INIT_LIST_HEAD(&com->lc_link);
6936         INIT_LIST_HEAD(&com->lc_link_dir);
6937         init_rwsem(&com->lc_sem);
6938         atomic_set(&com->lc_ref, 1);
6939         com->lc_lfsck = lfsck;
6940         com->lc_type = LFSCK_TYPE_LAYOUT;
6941         if (lfsck->li_master) {
6942                 com->lc_ops = &lfsck_layout_master_ops;
6943                 com->lc_data = lfsck_assistant_data_init(
6944                                 &lfsck_layout_assistant_ops,
6945                                 LFSCK_LAYOUT);
6946                 if (com->lc_data == NULL)
6947                         GOTO(out, rc = -ENOMEM);
6948
6949                 for (i = 0; i < LFSCK_STF_COUNT; i++)
6950                         mutex_init(&com->lc_sub_trace_objs[i].lsto_mutex);
6951         } else {
6952                 struct lfsck_layout_slave_data *llsd;
6953
6954                 com->lc_ops = &lfsck_layout_slave_ops;
6955                 OBD_ALLOC_PTR(llsd);
6956                 if (llsd == NULL)
6957                         GOTO(out, rc = -ENOMEM);
6958
6959                 INIT_LIST_HEAD(&llsd->llsd_seq_list);
6960                 INIT_LIST_HEAD(&llsd->llsd_master_list);
6961                 spin_lock_init(&llsd->llsd_lock);
6962                 llsd->llsd_rb_root = RB_ROOT;
6963                 init_rwsem(&llsd->llsd_rb_rwsem);
6964                 com->lc_data = llsd;
6965         }
6966         com->lc_file_size = sizeof(*lo);
6967         OBD_ALLOC(com->lc_file_ram, com->lc_file_size);
6968         if (com->lc_file_ram == NULL)
6969                 GOTO(out, rc = -ENOMEM);
6970
6971         OBD_ALLOC(com->lc_file_disk, com->lc_file_size);
6972         if (com->lc_file_disk == NULL)
6973                 GOTO(out, rc = -ENOMEM);
6974
6975         root = dt_locate(env, lfsck->li_bottom, &lfsck->li_local_root_fid);
6976         if (IS_ERR(root))
6977                 GOTO(out, rc = PTR_ERR(root));
6978
6979         if (unlikely(!dt_try_as_dir(env, root)))
6980                 GOTO(out, rc = -ENOTDIR);
6981
6982         obj = local_file_find_or_create(env, lfsck->li_los, root,
6983                                         LFSCK_LAYOUT,
6984                                         S_IFREG | S_IRUGO | S_IWUSR);
6985         if (IS_ERR(obj))
6986                 GOTO(out, rc = PTR_ERR(obj));
6987
6988         com->lc_obj = obj;
6989         rc = lfsck_layout_load(env, com);
6990         if (rc > 0) {
6991                 rc = lfsck_layout_reset(env, com, true);
6992         } else if (rc == -ENOENT) {
6993                 rc = lfsck_layout_init(env, com);
6994         } else if (lfsck->li_master) {
6995                 rc = lfsck_load_sub_trace_files(env, com,
6996                                 &dt_lfsck_layout_dangling_features,
6997                                 LFSCK_LAYOUT, false);
6998                 if (rc)
6999                         rc = lfsck_layout_reset(env, com, true);
7000         }
7001
7002         if (rc != 0)
7003                 GOTO(out, rc);
7004
7005         lo = com->lc_file_ram;
7006         switch (lo->ll_status) {
7007         case LS_INIT:
7008         case LS_COMPLETED:
7009         case LS_FAILED:
7010         case LS_STOPPED:
7011         case LS_PARTIAL:
7012                 spin_lock(&lfsck->li_lock);
7013                 list_add_tail(&com->lc_link, &lfsck->li_list_idle);
7014                 spin_unlock(&lfsck->li_lock);
7015                 break;
7016         default:
7017                 CERROR("%s: unknown lfsck_layout status %d\n",
7018                        lfsck_lfsck2name(lfsck), lo->ll_status);
7019                 /* fall through */
7020         case LS_SCANNING_PHASE1:
7021         case LS_SCANNING_PHASE2:
7022                 /* No need to store the status to disk right now.
7023                  * If the system crashed before the status stored,
7024                  * it will be loaded back when next time. */
7025                 lo->ll_status = LS_CRASHED;
7026                 if (!lfsck->li_master)
7027                         lo->ll_flags |= LF_INCOMPLETE;
7028                 /* fall through */
7029         case LS_PAUSED:
7030         case LS_CRASHED:
7031         case LS_CO_FAILED:
7032         case LS_CO_STOPPED:
7033         case LS_CO_PAUSED:
7034                 spin_lock(&lfsck->li_lock);
7035                 list_add_tail(&com->lc_link, &lfsck->li_list_scan);
7036                 spin_unlock(&lfsck->li_lock);
7037                 break;
7038         }
7039
7040         if (lo->ll_flags & LF_CRASHED_LASTID) {
7041                 LASSERT(lfsck->li_out_notify != NULL);
7042
7043                 lfsck->li_out_notify(env, lfsck->li_out_notify_data,
7044                                      LE_LASTID_REBUILDING);
7045         }
7046
7047         GOTO(out, rc = 0);
7048
7049 out:
7050         if (root != NULL && !IS_ERR(root))
7051                 lfsck_object_put(env, root);
7052
7053         if (rc != 0) {
7054                 lfsck_component_cleanup(env, com);
7055                 CERROR("%s: fail to init layout LFSCK component: rc = %d\n",
7056                        lfsck_lfsck2name(lfsck), rc);
7057         }
7058
7059         return rc;
7060 }
7061
7062 struct lfsck_orphan_it {
7063         struct lfsck_component           *loi_com;
7064         struct lfsck_rbtree_node         *loi_lrn;
7065         struct lfsck_layout_slave_target *loi_llst;
7066         struct lu_fid                     loi_key;
7067         struct lu_orphan_rec_v3           loi_rec;
7068         __u64                             loi_hash;
7069         unsigned int                      loi_over:1;
7070 };
7071
7072 static int lfsck_fid_match_idx(const struct lu_env *env,
7073                                struct lfsck_instance *lfsck,
7074                                const struct lu_fid *fid, int idx)
7075 {
7076         struct seq_server_site  *ss;
7077         struct lu_server_fld    *sf;
7078         struct lu_seq_range     *range = &lfsck_env_info(env)->lti_range;
7079         int                      rc;
7080
7081         /* All abnormal cases will be returned to MDT0. */
7082         if (!fid_is_norm(fid)) {
7083                 if (idx == 0)
7084                         return 1;
7085
7086                 return 0;
7087         }
7088
7089         ss = lfsck_dev_site(lfsck);
7090         if (unlikely(ss == NULL))
7091                 return -ENOTCONN;
7092
7093         sf = ss->ss_server_fld;
7094         LASSERT(sf != NULL);
7095
7096         fld_range_set_any(range);
7097         rc = fld_server_lookup(env, sf, fid_seq(fid), range);
7098         if (rc != 0)
7099                 return rc;
7100
7101         if (!fld_range_is_mdt(range))
7102                 return -EINVAL;
7103
7104         if (range->lsr_index == idx)
7105                 return 1;
7106
7107         return 0;
7108 }
7109
7110 static void lfsck_layout_destroy_orphan(const struct lu_env *env,
7111                                         struct dt_object *obj)
7112 {
7113         struct dt_device        *dev    = lfsck_obj2dev(obj);
7114         struct thandle          *handle;
7115         int                      rc;
7116         ENTRY;
7117
7118         handle = dt_trans_create(env, dev);
7119         if (IS_ERR(handle))
7120                 RETURN_EXIT;
7121
7122         rc = dt_declare_ref_del(env, obj, handle);
7123         if (rc != 0)
7124                 GOTO(stop, rc);
7125
7126         rc = dt_declare_destroy(env, obj, handle);
7127         if (rc != 0)
7128                 GOTO(stop, rc);
7129
7130         rc = dt_trans_start_local(env, dev, handle);
7131         if (rc != 0)
7132                 GOTO(stop, rc);
7133
7134         dt_write_lock(env, obj, 0);
7135         rc = dt_ref_del(env, obj, handle);
7136         if (rc == 0)
7137                 rc = dt_destroy(env, obj, handle);
7138         dt_write_unlock(env, obj);
7139
7140         GOTO(stop, rc);
7141
7142 stop:
7143         dt_trans_stop(env, dev, handle);
7144
7145         CDEBUG(D_LFSCK, "destroy orphan OST-object "DFID": rc = %d\n",
7146                PFID(lfsck_dto2fid(obj)), rc);
7147
7148         RETURN_EXIT;
7149 }
7150
7151 static int lfsck_orphan_index_lookup(const struct lu_env *env,
7152                                      struct dt_object *dt,
7153                                      struct dt_rec *rec,
7154                                      const struct dt_key *key)
7155 {
7156         return -EOPNOTSUPP;
7157 }
7158
7159 static int lfsck_orphan_index_declare_insert(const struct lu_env *env,
7160                                              struct dt_object *dt,
7161                                              const struct dt_rec *rec,
7162                                              const struct dt_key *key,
7163                                              struct thandle *handle)
7164 {
7165         return -EOPNOTSUPP;
7166 }
7167
7168 static int lfsck_orphan_index_insert(const struct lu_env *env,
7169                                      struct dt_object *dt,
7170                                      const struct dt_rec *rec,
7171                                      const struct dt_key *key,
7172                                      struct thandle *handle)
7173 {
7174         return -EOPNOTSUPP;
7175 }
7176
7177 static int lfsck_orphan_index_declare_delete(const struct lu_env *env,
7178                                              struct dt_object *dt,
7179                                              const struct dt_key *key,
7180                                              struct thandle *handle)
7181 {
7182         return -EOPNOTSUPP;
7183 }
7184
7185 static int lfsck_orphan_index_delete(const struct lu_env *env,
7186                                      struct dt_object *dt,
7187                                      const struct dt_key *key,
7188                                      struct thandle *handle)
7189 {
7190         return -EOPNOTSUPP;
7191 }
7192
7193 static struct dt_it *lfsck_orphan_it_init(const struct lu_env *env,
7194                                           struct dt_object *dt,
7195                                           __u32 attr)
7196 {
7197         struct dt_device                *dev    = lu2dt_dev(dt->do_lu.lo_dev);
7198         struct lfsck_instance           *lfsck;
7199         struct lfsck_component          *com    = NULL;
7200         struct lfsck_layout_slave_data  *llsd;
7201         struct lfsck_orphan_it          *it     = NULL;
7202         struct lfsck_layout             *lo;
7203         int                              rc     = 0;
7204         ENTRY;
7205
7206         lfsck = lfsck_instance_find(dev, true, false);
7207         if (unlikely(lfsck == NULL))
7208                 RETURN(ERR_PTR(-ENXIO));
7209
7210         com = lfsck_component_find(lfsck, LFSCK_TYPE_LAYOUT);
7211         if (unlikely(com == NULL))
7212                 GOTO(out, rc = -ENOENT);
7213
7214         lo = com->lc_file_ram;
7215         if (lo->ll_flags & LF_INCOMPLETE)
7216                 GOTO(out, rc = -ESRCH);
7217
7218         llsd = com->lc_data;
7219         if (!llsd->llsd_rbtree_valid)
7220                 GOTO(out, rc = -ESRCH);
7221
7222         OBD_ALLOC_PTR(it);
7223         if (it == NULL)
7224                 GOTO(out, rc = -ENOMEM);
7225
7226         it->loi_llst = lfsck_layout_llst_find_and_del(llsd, attr, false);
7227         if (it->loi_llst == NULL)
7228                 GOTO(out, rc = -ENXIO);
7229
7230         if (dev->dd_record_fid_accessed) {
7231                 /* The first iteration against the rbtree, scan the whole rbtree
7232                  * to remove the nodes which do NOT need to be handled. */
7233                 down_write(&llsd->llsd_rb_rwsem);
7234                 if (dev->dd_record_fid_accessed) {
7235                         struct rb_node                  *node;
7236                         struct rb_node                  *next;
7237                         struct lfsck_rbtree_node        *lrn;
7238
7239                         /* No need to record the fid accessing anymore. */
7240                         dev->dd_record_fid_accessed = 0;
7241
7242                         node = rb_first(&llsd->llsd_rb_root);
7243                         while (node != NULL) {
7244                                 next = rb_next(node);
7245                                 lrn = rb_entry(node, struct lfsck_rbtree_node,
7246                                                lrn_node);
7247                                 if (atomic_read(&lrn->lrn_known_count) <=
7248                                     atomic_read(&lrn->lrn_accessed_count)) {
7249                                         rb_erase(node, &llsd->llsd_rb_root);
7250                                         lfsck_rbtree_free(lrn);
7251                                 }
7252                                 node = next;
7253                         }
7254                 }
7255                 up_write(&llsd->llsd_rb_rwsem);
7256         }
7257
7258         /* read lock the rbtree when init, and unlock when fini */
7259         down_read(&llsd->llsd_rb_rwsem);
7260         it->loi_com = com;
7261         com = NULL;
7262
7263         GOTO(out, rc = 0);
7264
7265 out:
7266         if (com != NULL)
7267                 lfsck_component_put(env, com);
7268
7269         CDEBUG(D_LFSCK, "%s: init the orphan iteration: rc = %d\n",
7270                lfsck_lfsck2name(lfsck), rc);
7271
7272         lfsck_instance_put(env, lfsck);
7273         if (rc != 0) {
7274                 if (it != NULL)
7275                         OBD_FREE_PTR(it);
7276
7277                 it = (struct lfsck_orphan_it *)ERR_PTR(rc);
7278         }
7279
7280         return (struct dt_it *)it;
7281 }
7282
7283 static void lfsck_orphan_it_fini(const struct lu_env *env,
7284                                  struct dt_it *di)
7285 {
7286         struct lfsck_orphan_it           *it    = (struct lfsck_orphan_it *)di;
7287         struct lfsck_component           *com   = it->loi_com;
7288         struct lfsck_layout_slave_data   *llsd;
7289         struct lfsck_layout_slave_target *llst;
7290
7291         if (com != NULL) {
7292                 CDEBUG(D_LFSCK, "%s: fini the orphan iteration\n",
7293                        lfsck_lfsck2name(com->lc_lfsck));
7294
7295                 llsd = com->lc_data;
7296                 up_read(&llsd->llsd_rb_rwsem);
7297                 llst = it->loi_llst;
7298                 LASSERT(llst != NULL);
7299
7300                 /* Save the key and hash for iterate next. */
7301                 llst->llst_fid = it->loi_key;
7302                 llst->llst_hash = it->loi_hash;
7303                 lfsck_layout_llst_put(llst);
7304                 lfsck_component_put(env, com);
7305         }
7306         OBD_FREE_PTR(it);
7307 }
7308
7309 /**
7310  * \retval       +1: the iteration finished
7311  * \retval        0: on success, not finished
7312  * \retval      -ve: on error
7313  */
7314 static int lfsck_orphan_it_next(const struct lu_env *env,
7315                                 struct dt_it *di)
7316 {
7317         struct lfsck_thread_info        *info   = lfsck_env_info(env);
7318         struct filter_fid               *ff     = &info->lti_ff;
7319         struct lu_attr                  *la     = &info->lti_la;
7320         struct lfsck_orphan_it          *it     = (struct lfsck_orphan_it *)di;
7321         struct lu_fid                   *key    = &it->loi_key;
7322         struct lu_orphan_rec_v3         *rec    = &it->loi_rec;
7323         struct ost_layout               *ol     = &rec->lor_layout;
7324         struct lfsck_component          *com    = it->loi_com;
7325         struct lfsck_instance           *lfsck  = com->lc_lfsck;
7326         struct lfsck_layout_slave_data  *llsd   = com->lc_data;
7327         struct dt_object                *obj;
7328         struct lfsck_rbtree_node        *lrn;
7329         int                              pos;
7330         int                              rc;
7331         __u32                            save;
7332         __u32                            idx    = it->loi_llst->llst_index;
7333         bool                             exact  = false;
7334         ENTRY;
7335
7336         if (it->loi_over)
7337                 RETURN(1);
7338
7339 again0:
7340         lrn = it->loi_lrn;
7341         if (lrn == NULL) {
7342                 lrn = lfsck_rbtree_search(llsd, key, &exact);
7343                 if (lrn == NULL) {
7344                         it->loi_over = 1;
7345                         RETURN(1);
7346                 }
7347
7348                 it->loi_lrn = lrn;
7349                 if (!exact) {
7350                         key->f_seq = lrn->lrn_seq;
7351                         key->f_oid = lrn->lrn_first_oid;
7352                         key->f_ver = 0;
7353                 }
7354         } else {
7355                 key->f_oid++;
7356                 if (unlikely(key->f_oid == 0)) {
7357                         key->f_seq++;
7358                         it->loi_lrn = NULL;
7359                         goto again0;
7360                 }
7361
7362                 if (key->f_oid >=
7363                     lrn->lrn_first_oid + LFSCK_RBTREE_BITMAP_WIDTH) {
7364                         it->loi_lrn = NULL;
7365                         goto again0;
7366                 }
7367         }
7368
7369         if (unlikely(atomic_read(&lrn->lrn_known_count) <=
7370                      atomic_read(&lrn->lrn_accessed_count))) {
7371                 struct rb_node *next = rb_next(&lrn->lrn_node);
7372
7373                 while (next != NULL) {
7374                         lrn = rb_entry(next, struct lfsck_rbtree_node,
7375                                        lrn_node);
7376                         if (atomic_read(&lrn->lrn_known_count) >
7377                             atomic_read(&lrn->lrn_accessed_count))
7378                                 break;
7379                         next = rb_next(next);
7380                 }
7381
7382                 if (next == NULL) {
7383                         it->loi_over = 1;
7384                         RETURN(1);
7385                 }
7386
7387                 it->loi_lrn = lrn;
7388                 key->f_seq = lrn->lrn_seq;
7389                 key->f_oid = lrn->lrn_first_oid;
7390                 key->f_ver = 0;
7391         }
7392
7393         pos = key->f_oid - lrn->lrn_first_oid;
7394
7395 again1:
7396         pos = find_next_bit(lrn->lrn_known_bitmap,
7397                             LFSCK_RBTREE_BITMAP_WIDTH, pos);
7398         if (pos >= LFSCK_RBTREE_BITMAP_WIDTH) {
7399                 key->f_oid = lrn->lrn_first_oid + pos;
7400                 if (unlikely(key->f_oid < lrn->lrn_first_oid)) {
7401                         key->f_seq++;
7402                         key->f_oid = 0;
7403                 }
7404                 it->loi_lrn = NULL;
7405                 goto again0;
7406         }
7407
7408         if (test_bit(pos, lrn->lrn_accessed_bitmap)) {
7409                 pos++;
7410                 goto again1;
7411         }
7412
7413         key->f_oid = lrn->lrn_first_oid + pos;
7414         obj = lfsck_object_find_bottom(env, lfsck, key);
7415         if (IS_ERR(obj)) {
7416                 rc = PTR_ERR(obj);
7417                 if (rc == -ENOENT) {
7418                         pos++;
7419                         goto again1;
7420                 }
7421                 RETURN(rc);
7422         }
7423
7424         dt_read_lock(env, obj, 0);
7425         if (dt_object_exists(obj) == 0 ||
7426             lfsck_is_dead_obj(obj)) {
7427                 dt_read_unlock(env, obj);
7428                 lfsck_object_put(env, obj);
7429                 pos++;
7430                 goto again1;
7431         }
7432
7433         rc = dt_attr_get(env, obj, la);
7434         if (rc != 0)
7435                 GOTO(out, rc);
7436
7437         rc = dt_xattr_get(env, obj, lfsck_buf_get(env, ff, sizeof(*ff)),
7438                           XATTR_NAME_FID);
7439         if (rc == -ENODATA) {
7440                 /* For the pre-created OST-object, update the bitmap to avoid
7441                  * others LFSCK (second phase) iteration to touch it again. */
7442                 if (la->la_ctime == 0) {
7443                         if (!test_and_set_bit(pos, lrn->lrn_accessed_bitmap))
7444                                 atomic_inc(&lrn->lrn_accessed_count);
7445
7446                         /* For the race between repairing dangling referenced
7447                          * MDT-object and unlink the file, it may left orphan
7448                          * OST-object there. Destroy it now! */
7449                         if (unlikely(!(la->la_mode & S_ISUID))) {
7450                                 dt_read_unlock(env, obj);
7451                                 lfsck_layout_destroy_orphan(env, obj);
7452                                 lfsck_object_put(env, obj);
7453                                 pos++;
7454                                 goto again1;
7455                         }
7456                 } else if (idx == 0) {
7457                         /* If the orphan OST-object has no parent information,
7458                          * regard it as referenced by the MDT-object on MDT0. */
7459                         fid_zero(&rec->lor_rec.lor_fid);
7460                         rec->lor_rec.lor_uid = la->la_uid;
7461                         rec->lor_rec.lor_gid = la->la_gid;
7462                         memset(ol, 0, sizeof(*ol));
7463                         rec->lor_layout_version = 0;
7464                         rec->lor_range = 0;
7465
7466                         GOTO(out, rc = 0);
7467                 }
7468
7469                 dt_read_unlock(env, obj);
7470                 lfsck_object_put(env, obj);
7471                 pos++;
7472                 goto again1;
7473         }
7474
7475         if (rc < sizeof(struct lu_fid))
7476                 GOTO(out, rc = (rc < 0 ? rc : -EINVAL));
7477
7478         fid_le_to_cpu(&rec->lor_rec.lor_fid, &ff->ff_parent);
7479         /* Currently, the filter_fid::ff_parent::f_ver is not the real parent
7480          * MDT-object's FID::f_ver, instead it is the OST-object index in its
7481          * parent MDT-object's layout EA. */
7482         save = rec->lor_rec.lor_fid.f_stripe_idx;
7483         rec->lor_rec.lor_fid.f_ver = 0;
7484         rc = lfsck_fid_match_idx(env, lfsck, &rec->lor_rec.lor_fid, idx);
7485         /* If the orphan OST-object does not claim the MDT, then next.
7486          *
7487          * If we do not know whether it matches or not, then return it
7488          * to the MDT for further check. */
7489         if (rc == 0) {
7490                 dt_read_unlock(env, obj);
7491                 lfsck_object_put(env, obj);
7492                 pos++;
7493                 goto again1;
7494         }
7495
7496         rec->lor_rec.lor_fid.f_stripe_idx = save;
7497         rec->lor_rec.lor_uid = la->la_uid;
7498         rec->lor_rec.lor_gid = la->la_gid;
7499         ost_layout_le_to_cpu(ol, &ff->ff_layout);
7500         rec->lor_layout_version =
7501                 le32_to_cpu(ff->ff_layout_version & ~LU_LAYOUT_RESYNC);
7502         rec->lor_range = le32_to_cpu(ff->ff_range);
7503
7504         CDEBUG(D_LFSCK, "%s: return orphan "DFID", PFID "DFID", owner %u:%u, "
7505                "stripe size %u, stripe count %u, COMP id %u, COMP start %llu, "
7506                "COMP end %llu, layout version %u, range %u\n",
7507                lfsck_lfsck2name(com->lc_lfsck), PFID(key),
7508                PFID(&rec->lor_rec.lor_fid), rec->lor_rec.lor_uid,
7509                rec->lor_rec.lor_gid, ol->ol_stripe_size, ol->ol_stripe_count,
7510                ol->ol_comp_id, ol->ol_comp_start, ol->ol_comp_end,
7511                rec->lor_layout_version, rec->lor_range);
7512
7513         GOTO(out, rc = 0);
7514
7515 out:
7516         dt_read_unlock(env, obj);
7517         lfsck_object_put(env, obj);
7518         if (rc == 0)
7519                 it->loi_hash++;
7520
7521         return rc;
7522 }
7523
7524 /**
7525  * \retval       +1: locate to the exactly position
7526  * \retval        0: cannot locate to the exactly position,
7527  *                   call next() to move to a valid position.
7528  * \retval      -ve: on error
7529  */
7530 static int lfsck_orphan_it_get(const struct lu_env *env,
7531                                struct dt_it *di,
7532                                const struct dt_key *key)
7533 {
7534         struct lfsck_orphan_it  *it   = (struct lfsck_orphan_it *)di;
7535         int                      rc;
7536
7537         it->loi_key = *(struct lu_fid *)key;
7538         rc = lfsck_orphan_it_next(env, di);
7539         if (rc == 1)
7540                 return 0;
7541
7542         if (rc == 0)
7543                 return 1;
7544
7545         return rc;
7546 }
7547
7548 static void lfsck_orphan_it_put(const struct lu_env *env,
7549                                 struct dt_it *di)
7550 {
7551 }
7552
7553 static struct dt_key *lfsck_orphan_it_key(const struct lu_env *env,
7554                                           const struct dt_it *di)
7555 {
7556         struct lfsck_orphan_it *it = (struct lfsck_orphan_it *)di;
7557
7558         return (struct dt_key *)&it->loi_key;
7559 }
7560
7561 static int lfsck_orphan_it_key_size(const struct lu_env *env,
7562                                     const struct dt_it *di)
7563 {
7564         return sizeof(struct lu_fid);
7565 }
7566
7567 static int lfsck_orphan_it_rec(const struct lu_env *env,
7568                                const struct dt_it *di,
7569                                struct dt_rec *rec,
7570                                __u32 attr)
7571 {
7572         struct lfsck_orphan_it *it = (struct lfsck_orphan_it *)di;
7573
7574         *(struct lu_orphan_rec_v3 *)rec = it->loi_rec;
7575
7576         return 0;
7577 }
7578
7579 static __u64 lfsck_orphan_it_store(const struct lu_env *env,
7580                                    const struct dt_it *di)
7581 {
7582         struct lfsck_orphan_it  *it   = (struct lfsck_orphan_it *)di;
7583
7584         return it->loi_hash;
7585 }
7586
7587 /**
7588  * \retval       +1: locate to the exactly position
7589  * \retval        0: cannot locate to the exactly position,
7590  *                   call next() to move to a valid position.
7591  * \retval      -ve: on error
7592  */
7593 static int lfsck_orphan_it_load(const struct lu_env *env,
7594                                 const struct dt_it *di,
7595                                 __u64 hash)
7596 {
7597         struct lfsck_orphan_it           *it   = (struct lfsck_orphan_it *)di;
7598         struct lfsck_layout_slave_target *llst = it->loi_llst;
7599         int                               rc;
7600
7601         LASSERT(llst != NULL);
7602
7603         if (hash != llst->llst_hash) {
7604                 CDEBUG(D_LFSCK, "%s: the given hash %llu for orphan "
7605                        "iteration does not match the one when fini "
7606                        "%llu, to be reset.\n",
7607                        lfsck_lfsck2name(it->loi_com->lc_lfsck), hash,
7608                        llst->llst_hash);
7609                 fid_zero(&llst->llst_fid);
7610                 llst->llst_hash = 0;
7611         }
7612
7613         it->loi_key = llst->llst_fid;
7614         it->loi_hash = llst->llst_hash;
7615         rc = lfsck_orphan_it_next(env, (struct dt_it *)di);
7616         if (rc == 1)
7617                 return 0;
7618
7619         if (rc == 0)
7620                 return 1;
7621
7622         return rc;
7623 }
7624
7625 static int lfsck_orphan_it_key_rec(const struct lu_env *env,
7626                                    const struct dt_it *di,
7627                                    void *key_rec)
7628 {
7629         return 0;
7630 }
7631
7632 static const struct dt_index_operations lfsck_orphan_index_ops = {
7633         .dio_lookup             = lfsck_orphan_index_lookup,
7634         .dio_declare_insert     = lfsck_orphan_index_declare_insert,
7635         .dio_insert             = lfsck_orphan_index_insert,
7636         .dio_declare_delete     = lfsck_orphan_index_declare_delete,
7637         .dio_delete             = lfsck_orphan_index_delete,
7638         .dio_it = {
7639                 .init           = lfsck_orphan_it_init,
7640                 .fini           = lfsck_orphan_it_fini,
7641                 .get            = lfsck_orphan_it_get,
7642                 .put            = lfsck_orphan_it_put,
7643                 .next           = lfsck_orphan_it_next,
7644                 .key            = lfsck_orphan_it_key,
7645                 .key_size       = lfsck_orphan_it_key_size,
7646                 .rec            = lfsck_orphan_it_rec,
7647                 .store          = lfsck_orphan_it_store,
7648                 .load           = lfsck_orphan_it_load,
7649                 .key_rec        = lfsck_orphan_it_key_rec,
7650         }
7651 };